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
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
74 #include <linux/init_task.h>
75 #include <linux/binfmts.h>
76 #include <linux/context_tracking.h>
77 #include <linux/compiler.h>
79 #include <asm/switch_to.h>
81 #include <asm/irq_regs.h>
82 #include <asm/mutex.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
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);
227 i = sched_feat_set(cmp);
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 * this_rq_lock - lock this runqueue and disable interrupts.
294 static struct rq *this_rq_lock(void)
301 raw_spin_lock(&rq->lock);
306 #ifdef CONFIG_SCHED_HRTICK
308 * Use HR-timers to deliver accurate preemption points.
311 static void hrtick_clear(struct rq *rq)
313 if (hrtimer_active(&rq->hrtick_timer))
314 hrtimer_cancel(&rq->hrtick_timer);
318 * High-resolution timer tick.
319 * Runs from hardirq context with interrupts disabled.
321 static enum hrtimer_restart hrtick(struct hrtimer *timer)
323 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
325 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
327 raw_spin_lock(&rq->lock);
329 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
330 raw_spin_unlock(&rq->lock);
332 return HRTIMER_NORESTART;
337 static void __hrtick_restart(struct rq *rq)
339 struct hrtimer *timer = &rq->hrtick_timer;
341 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
345 * called from hardirq (IPI) context
347 static void __hrtick_start(void *arg)
351 raw_spin_lock(&rq->lock);
352 __hrtick_restart(rq);
353 rq->hrtick_csd_pending = 0;
354 raw_spin_unlock(&rq->lock);
358 * Called to set the hrtick timer state.
360 * called with rq->lock held and irqs disabled
362 void hrtick_start(struct rq *rq, u64 delay)
364 struct hrtimer *timer = &rq->hrtick_timer;
369 * Don't schedule slices shorter than 10000ns, that just
370 * doesn't make sense and can cause timer DoS.
372 delta = max_t(s64, delay, 10000LL);
373 time = ktime_add_ns(timer->base->get_time(), delta);
375 hrtimer_set_expires(timer, time);
377 if (rq == this_rq()) {
378 __hrtick_restart(rq);
379 } else if (!rq->hrtick_csd_pending) {
380 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
381 rq->hrtick_csd_pending = 1;
386 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
388 int cpu = (int)(long)hcpu;
391 case CPU_UP_CANCELED:
392 case CPU_UP_CANCELED_FROZEN:
393 case CPU_DOWN_PREPARE:
394 case CPU_DOWN_PREPARE_FROZEN:
396 case CPU_DEAD_FROZEN:
397 hrtick_clear(cpu_rq(cpu));
404 static __init void init_hrtick(void)
406 hotcpu_notifier(hotplug_hrtick, 0);
410 * Called to set the hrtick timer state.
412 * called with rq->lock held and irqs disabled
414 void hrtick_start(struct rq *rq, u64 delay)
417 * Don't schedule slices shorter than 10000ns, that just
418 * doesn't make sense. Rely on vruntime for fairness.
420 delay = max_t(u64, delay, 10000LL);
421 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
422 HRTIMER_MODE_REL_PINNED);
425 static inline void init_hrtick(void)
428 #endif /* CONFIG_SMP */
430 static void init_rq_hrtick(struct rq *rq)
433 rq->hrtick_csd_pending = 0;
435 rq->hrtick_csd.flags = 0;
436 rq->hrtick_csd.func = __hrtick_start;
437 rq->hrtick_csd.info = rq;
440 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
441 rq->hrtick_timer.function = hrtick;
443 #else /* CONFIG_SCHED_HRTICK */
444 static inline void hrtick_clear(struct rq *rq)
448 static inline void init_rq_hrtick(struct rq *rq)
452 static inline void init_hrtick(void)
455 #endif /* CONFIG_SCHED_HRTICK */
458 * cmpxchg based fetch_or, macro so it works for different integer types
460 #define fetch_or(ptr, val) \
461 ({ typeof(*(ptr)) __old, __val = *(ptr); \
463 __old = cmpxchg((ptr), __val, __val | (val)); \
464 if (__old == __val) \
471 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
473 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
474 * this avoids any races wrt polling state changes and thereby avoids
477 static bool set_nr_and_not_polling(struct task_struct *p)
479 struct thread_info *ti = task_thread_info(p);
480 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
484 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
486 * If this returns true, then the idle task promises to call
487 * sched_ttwu_pending() and reschedule soon.
489 static bool set_nr_if_polling(struct task_struct *p)
491 struct thread_info *ti = task_thread_info(p);
492 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
495 if (!(val & _TIF_POLLING_NRFLAG))
497 if (val & _TIF_NEED_RESCHED)
499 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
508 static bool set_nr_and_not_polling(struct task_struct *p)
510 set_tsk_need_resched(p);
515 static bool set_nr_if_polling(struct task_struct *p)
522 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
524 struct wake_q_node *node = &task->wake_q;
527 * Atomically grab the task, if ->wake_q is !nil already it means
528 * its already queued (either by us or someone else) and will get the
529 * wakeup due to that.
531 * This cmpxchg() implies a full barrier, which pairs with the write
532 * barrier implied by the wakeup in wake_up_list().
534 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
537 get_task_struct(task);
540 * The head is context local, there can be no concurrency.
543 head->lastp = &node->next;
546 void wake_up_q(struct wake_q_head *head)
548 struct wake_q_node *node = head->first;
550 while (node != WAKE_Q_TAIL) {
551 struct task_struct *task;
553 task = container_of(node, struct task_struct, wake_q);
555 /* task can safely be re-inserted now */
557 task->wake_q.next = NULL;
560 * wake_up_process() implies a wmb() to pair with the queueing
561 * in wake_q_add() so as not to miss wakeups.
563 wake_up_process(task);
564 put_task_struct(task);
569 * resched_curr - mark rq's current task 'to be rescheduled now'.
571 * On UP this means the setting of the need_resched flag, on SMP it
572 * might also involve a cross-CPU call to trigger the scheduler on
575 void resched_curr(struct rq *rq)
577 struct task_struct *curr = rq->curr;
580 lockdep_assert_held(&rq->lock);
582 if (test_tsk_need_resched(curr))
587 if (cpu == smp_processor_id()) {
588 set_tsk_need_resched(curr);
589 set_preempt_need_resched();
593 if (set_nr_and_not_polling(curr))
594 smp_send_reschedule(cpu);
596 trace_sched_wake_idle_without_ipi(cpu);
599 void resched_cpu(int cpu)
601 struct rq *rq = cpu_rq(cpu);
604 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
607 raw_spin_unlock_irqrestore(&rq->lock, flags);
611 #ifdef CONFIG_NO_HZ_COMMON
613 * In the semi idle case, use the nearest busy cpu for migrating timers
614 * from an idle cpu. This is good for power-savings.
616 * We don't do similar optimization for completely idle system, as
617 * selecting an idle cpu will add more delays to the timers than intended
618 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
620 int get_nohz_timer_target(void)
622 int i, cpu = smp_processor_id();
623 struct sched_domain *sd;
625 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
629 for_each_domain(cpu, sd) {
630 for_each_cpu(i, sched_domain_span(sd)) {
631 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
638 if (!is_housekeeping_cpu(cpu))
639 cpu = housekeeping_any_cpu();
645 * When add_timer_on() enqueues a timer into the timer wheel of an
646 * idle CPU then this timer might expire before the next timer event
647 * which is scheduled to wake up that CPU. In case of a completely
648 * idle system the next event might even be infinite time into the
649 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
650 * leaves the inner idle loop so the newly added timer is taken into
651 * account when the CPU goes back to idle and evaluates the timer
652 * wheel for the next timer event.
654 static void wake_up_idle_cpu(int cpu)
656 struct rq *rq = cpu_rq(cpu);
658 if (cpu == smp_processor_id())
661 if (set_nr_and_not_polling(rq->idle))
662 smp_send_reschedule(cpu);
664 trace_sched_wake_idle_without_ipi(cpu);
667 static bool wake_up_full_nohz_cpu(int cpu)
670 * We just need the target to call irq_exit() and re-evaluate
671 * the next tick. The nohz full kick at least implies that.
672 * If needed we can still optimize that later with an
675 if (tick_nohz_full_cpu(cpu)) {
676 if (cpu != smp_processor_id() ||
677 tick_nohz_tick_stopped())
678 tick_nohz_full_kick_cpu(cpu);
685 void wake_up_nohz_cpu(int cpu)
687 if (!wake_up_full_nohz_cpu(cpu))
688 wake_up_idle_cpu(cpu);
691 static inline bool got_nohz_idle_kick(void)
693 int cpu = smp_processor_id();
695 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
698 if (idle_cpu(cpu) && !need_resched())
702 * We can't run Idle Load Balance on this CPU for this time so we
703 * cancel it and clear NOHZ_BALANCE_KICK
705 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
709 #else /* CONFIG_NO_HZ_COMMON */
711 static inline bool got_nohz_idle_kick(void)
716 #endif /* CONFIG_NO_HZ_COMMON */
718 #ifdef CONFIG_NO_HZ_FULL
719 bool sched_can_stop_tick(void)
722 * FIFO realtime policy runs the highest priority task. Other runnable
723 * tasks are of a lower priority. The scheduler tick does nothing.
725 if (current->policy == SCHED_FIFO)
729 * Round-robin realtime tasks time slice with other tasks at the same
730 * realtime priority. Is this task the only one at this priority?
732 if (current->policy == SCHED_RR) {
733 struct sched_rt_entity *rt_se = ¤t->rt;
735 return list_is_singular(&rt_se->run_list);
739 * More than one running task need preemption.
740 * nr_running update is assumed to be visible
741 * after IPI is sent from wakers.
743 if (this_rq()->nr_running > 1)
748 #endif /* CONFIG_NO_HZ_FULL */
750 void sched_avg_update(struct rq *rq)
752 s64 period = sched_avg_period();
754 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
756 * Inline assembly required to prevent the compiler
757 * optimising this loop into a divmod call.
758 * See __iter_div_u64_rem() for another example of this.
760 asm("" : "+rm" (rq->age_stamp));
761 rq->age_stamp += period;
766 #endif /* CONFIG_SMP */
768 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
769 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
771 * Iterate task_group tree rooted at *from, calling @down when first entering a
772 * node and @up when leaving it for the final time.
774 * Caller must hold rcu_lock or sufficient equivalent.
776 int walk_tg_tree_from(struct task_group *from,
777 tg_visitor down, tg_visitor up, void *data)
779 struct task_group *parent, *child;
785 ret = (*down)(parent, data);
788 list_for_each_entry_rcu(child, &parent->children, siblings) {
795 ret = (*up)(parent, data);
796 if (ret || parent == from)
800 parent = parent->parent;
807 int tg_nop(struct task_group *tg, void *data)
813 static void set_load_weight(struct task_struct *p)
815 int prio = p->static_prio - MAX_RT_PRIO;
816 struct load_weight *load = &p->se.load;
819 * SCHED_IDLE tasks get minimal weight:
821 if (idle_policy(p->policy)) {
822 load->weight = scale_load(WEIGHT_IDLEPRIO);
823 load->inv_weight = WMULT_IDLEPRIO;
827 load->weight = scale_load(sched_prio_to_weight[prio]);
828 load->inv_weight = sched_prio_to_wmult[prio];
831 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
834 if (!(flags & ENQUEUE_RESTORE))
835 sched_info_queued(rq, p);
836 p->sched_class->enqueue_task(rq, p, flags);
839 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
842 if (!(flags & DEQUEUE_SAVE))
843 sched_info_dequeued(rq, p);
844 p->sched_class->dequeue_task(rq, p, flags);
847 void activate_task(struct rq *rq, struct task_struct *p, int flags)
849 if (task_contributes_to_load(p))
850 rq->nr_uninterruptible--;
852 enqueue_task(rq, p, flags);
855 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
857 if (task_contributes_to_load(p))
858 rq->nr_uninterruptible++;
860 dequeue_task(rq, p, flags);
863 static void update_rq_clock_task(struct rq *rq, s64 delta)
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
869 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 s64 steal = 0, irq_delta = 0;
872 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
876 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 * this case when a previous update_rq_clock() happened inside a
880 * When this happens, we stop ->clock_task and only update the
881 * prev_irq_time stamp to account for the part that fit, so that a next
882 * update will consume the rest. This ensures ->clock_task is
885 * It does however cause some slight miss-attribution of {soft,}irq
886 * time, a more accurate solution would be to update the irq_time using
887 * the current rq->clock timestamp, except that would require using
890 if (irq_delta > delta)
893 rq->prev_irq_time += irq_delta;
896 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 if (static_key_false((¶virt_steal_rq_enabled))) {
898 steal = paravirt_steal_clock(cpu_of(rq));
899 steal -= rq->prev_steal_time_rq;
901 if (unlikely(steal > delta))
904 rq->prev_steal_time_rq += steal;
909 rq->clock_task += delta;
911 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
912 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
913 sched_rt_avg_update(rq, irq_delta + steal);
917 void sched_set_stop_task(int cpu, struct task_struct *stop)
919 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
920 struct task_struct *old_stop = cpu_rq(cpu)->stop;
924 * Make it appear like a SCHED_FIFO task, its something
925 * userspace knows about and won't get confused about.
927 * Also, it will make PI more or less work without too
928 * much confusion -- but then, stop work should not
929 * rely on PI working anyway.
931 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
933 stop->sched_class = &stop_sched_class;
936 cpu_rq(cpu)->stop = stop;
940 * Reset it back to a normal scheduling class so that
941 * it can die in pieces.
943 old_stop->sched_class = &rt_sched_class;
948 * __normal_prio - return the priority that is based on the static prio
950 static inline int __normal_prio(struct task_struct *p)
952 return p->static_prio;
956 * Calculate the expected normal priority: i.e. priority
957 * without taking RT-inheritance into account. Might be
958 * boosted by interactivity modifiers. Changes upon fork,
959 * setprio syscalls, and whenever the interactivity
960 * estimator recalculates.
962 static inline int normal_prio(struct task_struct *p)
966 if (task_has_dl_policy(p))
967 prio = MAX_DL_PRIO-1;
968 else if (task_has_rt_policy(p))
969 prio = MAX_RT_PRIO-1 - p->rt_priority;
971 prio = __normal_prio(p);
976 * Calculate the current priority, i.e. the priority
977 * taken into account by the scheduler. This value might
978 * be boosted by RT tasks, or might be boosted by
979 * interactivity modifiers. Will be RT if the task got
980 * RT-boosted. If not then it returns p->normal_prio.
982 static int effective_prio(struct task_struct *p)
984 p->normal_prio = normal_prio(p);
986 * If we are RT tasks or we were boosted to RT priority,
987 * keep the priority unchanged. Otherwise, update priority
988 * to the normal priority:
990 if (!rt_prio(p->prio))
991 return p->normal_prio;
996 * task_curr - is this task currently executing on a CPU?
997 * @p: the task in question.
999 * Return: 1 if the task is currently executing. 0 otherwise.
1001 inline int task_curr(const struct task_struct *p)
1003 return cpu_curr(task_cpu(p)) == p;
1007 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1008 * use the balance_callback list if you want balancing.
1010 * this means any call to check_class_changed() must be followed by a call to
1011 * balance_callback().
1013 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1014 const struct sched_class *prev_class,
1017 if (prev_class != p->sched_class) {
1018 if (prev_class->switched_from)
1019 prev_class->switched_from(rq, p);
1021 p->sched_class->switched_to(rq, p);
1022 } else if (oldprio != p->prio || dl_task(p))
1023 p->sched_class->prio_changed(rq, p, oldprio);
1026 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1028 const struct sched_class *class;
1030 if (p->sched_class == rq->curr->sched_class) {
1031 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1033 for_each_class(class) {
1034 if (class == rq->curr->sched_class)
1036 if (class == p->sched_class) {
1044 * A queue event has occurred, and we're going to schedule. In
1045 * this case, we can save a useless back to back clock update.
1047 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1048 rq_clock_skip_update(rq, true);
1053 * This is how migration works:
1055 * 1) we invoke migration_cpu_stop() on the target CPU using
1057 * 2) stopper starts to run (implicitly forcing the migrated thread
1059 * 3) it checks whether the migrated task is still in the wrong runqueue.
1060 * 4) if it's in the wrong runqueue then the migration thread removes
1061 * it and puts it into the right queue.
1062 * 5) stopper completes and stop_one_cpu() returns and the migration
1067 * move_queued_task - move a queued task to new rq.
1069 * Returns (locked) new rq. Old rq's lock is released.
1071 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1073 lockdep_assert_held(&rq->lock);
1075 p->on_rq = TASK_ON_RQ_MIGRATING;
1076 dequeue_task(rq, p, 0);
1077 set_task_cpu(p, new_cpu);
1078 raw_spin_unlock(&rq->lock);
1080 rq = cpu_rq(new_cpu);
1082 raw_spin_lock(&rq->lock);
1083 BUG_ON(task_cpu(p) != new_cpu);
1084 enqueue_task(rq, p, 0);
1085 p->on_rq = TASK_ON_RQ_QUEUED;
1086 check_preempt_curr(rq, p, 0);
1091 struct migration_arg {
1092 struct task_struct *task;
1097 * Move (not current) task off this cpu, onto dest cpu. We're doing
1098 * this because either it can't run here any more (set_cpus_allowed()
1099 * away from this CPU, or CPU going down), or because we're
1100 * attempting to rebalance this task on exec (sched_exec).
1102 * So we race with normal scheduler movements, but that's OK, as long
1103 * as the task is no longer on this CPU.
1105 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1107 if (unlikely(!cpu_active(dest_cpu)))
1110 /* Affinity changed (again). */
1111 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1114 rq = move_queued_task(rq, p, dest_cpu);
1120 * migration_cpu_stop - this will be executed by a highprio stopper thread
1121 * and performs thread migration by bumping thread off CPU then
1122 * 'pushing' onto another runqueue.
1124 static int migration_cpu_stop(void *data)
1126 struct migration_arg *arg = data;
1127 struct task_struct *p = arg->task;
1128 struct rq *rq = this_rq();
1131 * The original target cpu might have gone down and we might
1132 * be on another cpu but it doesn't matter.
1134 local_irq_disable();
1136 * We need to explicitly wake pending tasks before running
1137 * __migrate_task() such that we will not miss enforcing cpus_allowed
1138 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1140 sched_ttwu_pending();
1142 raw_spin_lock(&p->pi_lock);
1143 raw_spin_lock(&rq->lock);
1145 * If task_rq(p) != rq, it cannot be migrated here, because we're
1146 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1147 * we're holding p->pi_lock.
1149 if (task_rq(p) == rq && task_on_rq_queued(p))
1150 rq = __migrate_task(rq, p, arg->dest_cpu);
1151 raw_spin_unlock(&rq->lock);
1152 raw_spin_unlock(&p->pi_lock);
1159 * sched_class::set_cpus_allowed must do the below, but is not required to
1160 * actually call this function.
1162 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1164 cpumask_copy(&p->cpus_allowed, new_mask);
1165 p->nr_cpus_allowed = cpumask_weight(new_mask);
1168 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1170 struct rq *rq = task_rq(p);
1171 bool queued, running;
1173 lockdep_assert_held(&p->pi_lock);
1175 queued = task_on_rq_queued(p);
1176 running = task_current(rq, p);
1180 * Because __kthread_bind() calls this on blocked tasks without
1183 lockdep_assert_held(&rq->lock);
1184 dequeue_task(rq, p, DEQUEUE_SAVE);
1187 put_prev_task(rq, p);
1189 p->sched_class->set_cpus_allowed(p, new_mask);
1192 p->sched_class->set_curr_task(rq);
1194 enqueue_task(rq, p, ENQUEUE_RESTORE);
1198 * Change a given task's CPU affinity. Migrate the thread to a
1199 * proper CPU and schedule it away if the CPU it's executing on
1200 * is removed from the allowed bitmask.
1202 * NOTE: the caller must have a valid reference to the task, the
1203 * task must not exit() & deallocate itself prematurely. The
1204 * call is not atomic; no spinlocks may be held.
1206 static int __set_cpus_allowed_ptr(struct task_struct *p,
1207 const struct cpumask *new_mask, bool check)
1209 unsigned long flags;
1211 unsigned int dest_cpu;
1214 rq = task_rq_lock(p, &flags);
1217 * Must re-check here, to close a race against __kthread_bind(),
1218 * sched_setaffinity() is not guaranteed to observe the flag.
1220 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1225 if (cpumask_equal(&p->cpus_allowed, new_mask))
1228 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1233 do_set_cpus_allowed(p, new_mask);
1235 /* Can the task run on the task's current CPU? If so, we're done */
1236 if (cpumask_test_cpu(task_cpu(p), new_mask))
1239 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1240 if (task_running(rq, p) || p->state == TASK_WAKING) {
1241 struct migration_arg arg = { p, dest_cpu };
1242 /* Need help from migration thread: drop lock and wait. */
1243 task_rq_unlock(rq, p, &flags);
1244 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1245 tlb_migrate_finish(p->mm);
1247 } else if (task_on_rq_queued(p)) {
1249 * OK, since we're going to drop the lock immediately
1250 * afterwards anyway.
1252 lockdep_unpin_lock(&rq->lock);
1253 rq = move_queued_task(rq, p, dest_cpu);
1254 lockdep_pin_lock(&rq->lock);
1257 task_rq_unlock(rq, p, &flags);
1262 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1264 return __set_cpus_allowed_ptr(p, new_mask, false);
1266 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1268 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1270 #ifdef CONFIG_SCHED_DEBUG
1272 * We should never call set_task_cpu() on a blocked task,
1273 * ttwu() will sort out the placement.
1275 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1279 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1280 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1281 * time relying on p->on_rq.
1283 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1284 p->sched_class == &fair_sched_class &&
1285 (p->on_rq && !task_on_rq_migrating(p)));
1287 #ifdef CONFIG_LOCKDEP
1289 * The caller should hold either p->pi_lock or rq->lock, when changing
1290 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1292 * sched_move_task() holds both and thus holding either pins the cgroup,
1295 * Furthermore, all task_rq users should acquire both locks, see
1298 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1299 lockdep_is_held(&task_rq(p)->lock)));
1303 trace_sched_migrate_task(p, new_cpu);
1305 if (task_cpu(p) != new_cpu) {
1306 if (p->sched_class->migrate_task_rq)
1307 p->sched_class->migrate_task_rq(p);
1308 p->se.nr_migrations++;
1309 perf_event_task_migrate(p);
1312 __set_task_cpu(p, new_cpu);
1315 static void __migrate_swap_task(struct task_struct *p, int cpu)
1317 if (task_on_rq_queued(p)) {
1318 struct rq *src_rq, *dst_rq;
1320 src_rq = task_rq(p);
1321 dst_rq = cpu_rq(cpu);
1323 p->on_rq = TASK_ON_RQ_MIGRATING;
1324 deactivate_task(src_rq, p, 0);
1325 set_task_cpu(p, cpu);
1326 activate_task(dst_rq, p, 0);
1327 p->on_rq = TASK_ON_RQ_QUEUED;
1328 check_preempt_curr(dst_rq, p, 0);
1331 * Task isn't running anymore; make it appear like we migrated
1332 * it before it went to sleep. This means on wakeup we make the
1333 * previous cpu our targer instead of where it really is.
1339 struct migration_swap_arg {
1340 struct task_struct *src_task, *dst_task;
1341 int src_cpu, dst_cpu;
1344 static int migrate_swap_stop(void *data)
1346 struct migration_swap_arg *arg = data;
1347 struct rq *src_rq, *dst_rq;
1350 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1353 src_rq = cpu_rq(arg->src_cpu);
1354 dst_rq = cpu_rq(arg->dst_cpu);
1356 double_raw_lock(&arg->src_task->pi_lock,
1357 &arg->dst_task->pi_lock);
1358 double_rq_lock(src_rq, dst_rq);
1360 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1363 if (task_cpu(arg->src_task) != arg->src_cpu)
1366 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1369 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1372 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1373 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1378 double_rq_unlock(src_rq, dst_rq);
1379 raw_spin_unlock(&arg->dst_task->pi_lock);
1380 raw_spin_unlock(&arg->src_task->pi_lock);
1386 * Cross migrate two tasks
1388 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1390 struct migration_swap_arg arg;
1393 arg = (struct migration_swap_arg){
1395 .src_cpu = task_cpu(cur),
1397 .dst_cpu = task_cpu(p),
1400 if (arg.src_cpu == arg.dst_cpu)
1404 * These three tests are all lockless; this is OK since all of them
1405 * will be re-checked with proper locks held further down the line.
1407 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1410 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1413 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1416 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1417 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1424 * wait_task_inactive - wait for a thread to unschedule.
1426 * If @match_state is nonzero, it's the @p->state value just checked and
1427 * not expected to change. If it changes, i.e. @p might have woken up,
1428 * then return zero. When we succeed in waiting for @p to be off its CPU,
1429 * we return a positive number (its total switch count). If a second call
1430 * a short while later returns the same number, the caller can be sure that
1431 * @p has remained unscheduled the whole time.
1433 * The caller must ensure that the task *will* unschedule sometime soon,
1434 * else this function might spin for a *long* time. This function can't
1435 * be called with interrupts off, or it may introduce deadlock with
1436 * smp_call_function() if an IPI is sent by the same process we are
1437 * waiting to become inactive.
1439 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1441 unsigned long flags;
1442 int running, queued;
1448 * We do the initial early heuristics without holding
1449 * any task-queue locks at all. We'll only try to get
1450 * the runqueue lock when things look like they will
1456 * If the task is actively running on another CPU
1457 * still, just relax and busy-wait without holding
1460 * NOTE! Since we don't hold any locks, it's not
1461 * even sure that "rq" stays as the right runqueue!
1462 * But we don't care, since "task_running()" will
1463 * return false if the runqueue has changed and p
1464 * is actually now running somewhere else!
1466 while (task_running(rq, p)) {
1467 if (match_state && unlikely(p->state != match_state))
1473 * Ok, time to look more closely! We need the rq
1474 * lock now, to be *sure*. If we're wrong, we'll
1475 * just go back and repeat.
1477 rq = task_rq_lock(p, &flags);
1478 trace_sched_wait_task(p);
1479 running = task_running(rq, p);
1480 queued = task_on_rq_queued(p);
1482 if (!match_state || p->state == match_state)
1483 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1484 task_rq_unlock(rq, p, &flags);
1487 * If it changed from the expected state, bail out now.
1489 if (unlikely(!ncsw))
1493 * Was it really running after all now that we
1494 * checked with the proper locks actually held?
1496 * Oops. Go back and try again..
1498 if (unlikely(running)) {
1504 * It's not enough that it's not actively running,
1505 * it must be off the runqueue _entirely_, and not
1508 * So if it was still runnable (but just not actively
1509 * running right now), it's preempted, and we should
1510 * yield - it could be a while.
1512 if (unlikely(queued)) {
1513 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1515 set_current_state(TASK_UNINTERRUPTIBLE);
1516 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1521 * Ahh, all good. It wasn't running, and it wasn't
1522 * runnable, which means that it will never become
1523 * running in the future either. We're all done!
1532 * kick_process - kick a running thread to enter/exit the kernel
1533 * @p: the to-be-kicked thread
1535 * Cause a process which is running on another CPU to enter
1536 * kernel-mode, without any delay. (to get signals handled.)
1538 * NOTE: this function doesn't have to take the runqueue lock,
1539 * because all it wants to ensure is that the remote task enters
1540 * the kernel. If the IPI races and the task has been migrated
1541 * to another CPU then no harm is done and the purpose has been
1544 void kick_process(struct task_struct *p)
1550 if ((cpu != smp_processor_id()) && task_curr(p))
1551 smp_send_reschedule(cpu);
1554 EXPORT_SYMBOL_GPL(kick_process);
1557 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1559 static int select_fallback_rq(int cpu, struct task_struct *p)
1561 int nid = cpu_to_node(cpu);
1562 const struct cpumask *nodemask = NULL;
1563 enum { cpuset, possible, fail } state = cpuset;
1567 * If the node that the cpu is on has been offlined, cpu_to_node()
1568 * will return -1. There is no cpu on the node, and we should
1569 * select the cpu on the other node.
1572 nodemask = cpumask_of_node(nid);
1574 /* Look for allowed, online CPU in same node. */
1575 for_each_cpu(dest_cpu, nodemask) {
1576 if (!cpu_online(dest_cpu))
1578 if (!cpu_active(dest_cpu))
1580 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1586 /* Any allowed, online CPU? */
1587 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1588 if (!cpu_online(dest_cpu))
1590 if (!cpu_active(dest_cpu))
1595 /* No more Mr. Nice Guy. */
1598 if (IS_ENABLED(CONFIG_CPUSETS)) {
1599 cpuset_cpus_allowed_fallback(p);
1605 do_set_cpus_allowed(p, cpu_possible_mask);
1616 if (state != cpuset) {
1618 * Don't tell them about moving exiting tasks or
1619 * kernel threads (both mm NULL), since they never
1622 if (p->mm && printk_ratelimit()) {
1623 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1624 task_pid_nr(p), p->comm, cpu);
1632 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1635 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1637 lockdep_assert_held(&p->pi_lock);
1639 if (p->nr_cpus_allowed > 1)
1640 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1643 * In order not to call set_task_cpu() on a blocking task we need
1644 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1647 * Since this is common to all placement strategies, this lives here.
1649 * [ this allows ->select_task() to simply return task_cpu(p) and
1650 * not worry about this generic constraint ]
1652 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1654 cpu = select_fallback_rq(task_cpu(p), p);
1659 static void update_avg(u64 *avg, u64 sample)
1661 s64 diff = sample - *avg;
1667 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1668 const struct cpumask *new_mask, bool check)
1670 return set_cpus_allowed_ptr(p, new_mask);
1673 #endif /* CONFIG_SMP */
1676 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1678 #ifdef CONFIG_SCHEDSTATS
1679 struct rq *rq = this_rq();
1682 int this_cpu = smp_processor_id();
1684 if (cpu == this_cpu) {
1685 schedstat_inc(rq, ttwu_local);
1686 schedstat_inc(p, se.statistics.nr_wakeups_local);
1688 struct sched_domain *sd;
1690 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1692 for_each_domain(this_cpu, sd) {
1693 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1694 schedstat_inc(sd, ttwu_wake_remote);
1701 if (wake_flags & WF_MIGRATED)
1702 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1704 #endif /* CONFIG_SMP */
1706 schedstat_inc(rq, ttwu_count);
1707 schedstat_inc(p, se.statistics.nr_wakeups);
1709 if (wake_flags & WF_SYNC)
1710 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1712 #endif /* CONFIG_SCHEDSTATS */
1715 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1717 activate_task(rq, p, en_flags);
1718 p->on_rq = TASK_ON_RQ_QUEUED;
1720 /* if a worker is waking up, notify workqueue */
1721 if (p->flags & PF_WQ_WORKER)
1722 wq_worker_waking_up(p, cpu_of(rq));
1726 * Mark the task runnable and perform wakeup-preemption.
1729 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1731 check_preempt_curr(rq, p, wake_flags);
1732 p->state = TASK_RUNNING;
1733 trace_sched_wakeup(p);
1736 if (p->sched_class->task_woken) {
1738 * Our task @p is fully woken up and running; so its safe to
1739 * drop the rq->lock, hereafter rq is only used for statistics.
1741 lockdep_unpin_lock(&rq->lock);
1742 p->sched_class->task_woken(rq, p);
1743 lockdep_pin_lock(&rq->lock);
1746 if (rq->idle_stamp) {
1747 u64 delta = rq_clock(rq) - rq->idle_stamp;
1748 u64 max = 2*rq->max_idle_balance_cost;
1750 update_avg(&rq->avg_idle, delta);
1752 if (rq->avg_idle > max)
1761 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1763 lockdep_assert_held(&rq->lock);
1766 if (p->sched_contributes_to_load)
1767 rq->nr_uninterruptible--;
1770 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1771 ttwu_do_wakeup(rq, p, wake_flags);
1775 * Called in case the task @p isn't fully descheduled from its runqueue,
1776 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1777 * since all we need to do is flip p->state to TASK_RUNNING, since
1778 * the task is still ->on_rq.
1780 static int ttwu_remote(struct task_struct *p, int wake_flags)
1785 rq = __task_rq_lock(p);
1786 if (task_on_rq_queued(p)) {
1787 /* check_preempt_curr() may use rq clock */
1788 update_rq_clock(rq);
1789 ttwu_do_wakeup(rq, p, wake_flags);
1792 __task_rq_unlock(rq);
1798 void sched_ttwu_pending(void)
1800 struct rq *rq = this_rq();
1801 struct llist_node *llist = llist_del_all(&rq->wake_list);
1802 struct task_struct *p;
1803 unsigned long flags;
1808 raw_spin_lock_irqsave(&rq->lock, flags);
1809 lockdep_pin_lock(&rq->lock);
1812 p = llist_entry(llist, struct task_struct, wake_entry);
1813 llist = llist_next(llist);
1814 ttwu_do_activate(rq, p, 0);
1817 lockdep_unpin_lock(&rq->lock);
1818 raw_spin_unlock_irqrestore(&rq->lock, flags);
1821 void scheduler_ipi(void)
1824 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1825 * TIF_NEED_RESCHED remotely (for the first time) will also send
1828 preempt_fold_need_resched();
1830 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1834 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1835 * traditionally all their work was done from the interrupt return
1836 * path. Now that we actually do some work, we need to make sure
1839 * Some archs already do call them, luckily irq_enter/exit nest
1842 * Arguably we should visit all archs and update all handlers,
1843 * however a fair share of IPIs are still resched only so this would
1844 * somewhat pessimize the simple resched case.
1847 sched_ttwu_pending();
1850 * Check if someone kicked us for doing the nohz idle load balance.
1852 if (unlikely(got_nohz_idle_kick())) {
1853 this_rq()->idle_balance = 1;
1854 raise_softirq_irqoff(SCHED_SOFTIRQ);
1859 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1861 struct rq *rq = cpu_rq(cpu);
1863 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1864 if (!set_nr_if_polling(rq->idle))
1865 smp_send_reschedule(cpu);
1867 trace_sched_wake_idle_without_ipi(cpu);
1871 void wake_up_if_idle(int cpu)
1873 struct rq *rq = cpu_rq(cpu);
1874 unsigned long flags;
1878 if (!is_idle_task(rcu_dereference(rq->curr)))
1881 if (set_nr_if_polling(rq->idle)) {
1882 trace_sched_wake_idle_without_ipi(cpu);
1884 raw_spin_lock_irqsave(&rq->lock, flags);
1885 if (is_idle_task(rq->curr))
1886 smp_send_reschedule(cpu);
1887 /* Else cpu is not in idle, do nothing here */
1888 raw_spin_unlock_irqrestore(&rq->lock, flags);
1895 bool cpus_share_cache(int this_cpu, int that_cpu)
1897 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1899 #endif /* CONFIG_SMP */
1901 static void ttwu_queue(struct task_struct *p, int cpu)
1903 struct rq *rq = cpu_rq(cpu);
1905 #if defined(CONFIG_SMP)
1906 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1907 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1908 ttwu_queue_remote(p, cpu);
1913 raw_spin_lock(&rq->lock);
1914 lockdep_pin_lock(&rq->lock);
1915 ttwu_do_activate(rq, p, 0);
1916 lockdep_unpin_lock(&rq->lock);
1917 raw_spin_unlock(&rq->lock);
1921 * Notes on Program-Order guarantees on SMP systems.
1925 * The basic program-order guarantee on SMP systems is that when a task [t]
1926 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1927 * execution on its new cpu [c1].
1929 * For migration (of runnable tasks) this is provided by the following means:
1931 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1932 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1933 * rq(c1)->lock (if not at the same time, then in that order).
1934 * C) LOCK of the rq(c1)->lock scheduling in task
1936 * Transitivity guarantees that B happens after A and C after B.
1937 * Note: we only require RCpc transitivity.
1938 * Note: the cpu doing B need not be c0 or c1
1947 * UNLOCK rq(0)->lock
1949 * LOCK rq(0)->lock // orders against CPU0
1951 * UNLOCK rq(0)->lock
1955 * UNLOCK rq(1)->lock
1957 * LOCK rq(1)->lock // orders against CPU2
1960 * UNLOCK rq(1)->lock
1963 * BLOCKING -- aka. SLEEP + WAKEUP
1965 * For blocking we (obviously) need to provide the same guarantee as for
1966 * migration. However the means are completely different as there is no lock
1967 * chain to provide order. Instead we do:
1969 * 1) smp_store_release(X->on_cpu, 0)
1970 * 2) smp_cond_acquire(!X->on_cpu)
1974 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1976 * LOCK rq(0)->lock LOCK X->pi_lock
1979 * smp_store_release(X->on_cpu, 0);
1981 * smp_cond_acquire(!X->on_cpu);
1987 * X->state = RUNNING
1988 * UNLOCK rq(2)->lock
1990 * LOCK rq(2)->lock // orders against CPU1
1993 * UNLOCK rq(2)->lock
1996 * UNLOCK rq(0)->lock
1999 * However; for wakeups there is a second guarantee we must provide, namely we
2000 * must observe the state that lead to our wakeup. That is, not only must our
2001 * task observe its own prior state, it must also observe the stores prior to
2004 * This means that any means of doing remote wakeups must order the CPU doing
2005 * the wakeup against the CPU the task is going to end up running on. This,
2006 * however, is already required for the regular Program-Order guarantee above,
2007 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
2012 * try_to_wake_up - wake up a thread
2013 * @p: the thread to be awakened
2014 * @state: the mask of task states that can be woken
2015 * @wake_flags: wake modifier flags (WF_*)
2017 * Put it on the run-queue if it's not already there. The "current"
2018 * thread is always on the run-queue (except when the actual
2019 * re-schedule is in progress), and as such you're allowed to do
2020 * the simpler "current->state = TASK_RUNNING" to mark yourself
2021 * runnable without the overhead of this.
2023 * Return: %true if @p was woken up, %false if it was already running.
2024 * or @state didn't match @p's state.
2027 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2029 unsigned long flags;
2030 int cpu, success = 0;
2033 * If we are going to wake up a thread waiting for CONDITION we
2034 * need to ensure that CONDITION=1 done by the caller can not be
2035 * reordered with p->state check below. This pairs with mb() in
2036 * set_current_state() the waiting thread does.
2038 smp_mb__before_spinlock();
2039 raw_spin_lock_irqsave(&p->pi_lock, flags);
2040 if (!(p->state & state))
2043 trace_sched_waking(p);
2045 success = 1; /* we're going to change ->state */
2048 if (p->on_rq && ttwu_remote(p, wake_flags))
2053 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2054 * possible to, falsely, observe p->on_cpu == 0.
2056 * One must be running (->on_cpu == 1) in order to remove oneself
2057 * from the runqueue.
2059 * [S] ->on_cpu = 1; [L] ->on_rq
2063 * [S] ->on_rq = 0; [L] ->on_cpu
2065 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2066 * from the consecutive calls to schedule(); the first switching to our
2067 * task, the second putting it to sleep.
2072 * If the owning (remote) cpu is still in the middle of schedule() with
2073 * this task as prev, wait until its done referencing the task.
2075 * Pairs with the smp_store_release() in finish_lock_switch().
2077 * This ensures that tasks getting woken will be fully ordered against
2078 * their previous state and preserve Program Order.
2080 smp_cond_acquire(!p->on_cpu);
2082 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2083 p->state = TASK_WAKING;
2085 if (p->sched_class->task_waking)
2086 p->sched_class->task_waking(p);
2088 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2089 if (task_cpu(p) != cpu) {
2090 wake_flags |= WF_MIGRATED;
2091 set_task_cpu(p, cpu);
2093 #endif /* CONFIG_SMP */
2097 ttwu_stat(p, cpu, wake_flags);
2099 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2105 * try_to_wake_up_local - try to wake up a local task with rq lock held
2106 * @p: the thread to be awakened
2108 * Put @p on the run-queue if it's not already there. The caller must
2109 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2112 static void try_to_wake_up_local(struct task_struct *p)
2114 struct rq *rq = task_rq(p);
2116 if (WARN_ON_ONCE(rq != this_rq()) ||
2117 WARN_ON_ONCE(p == current))
2120 lockdep_assert_held(&rq->lock);
2122 if (!raw_spin_trylock(&p->pi_lock)) {
2124 * This is OK, because current is on_cpu, which avoids it being
2125 * picked for load-balance and preemption/IRQs are still
2126 * disabled avoiding further scheduler activity on it and we've
2127 * not yet picked a replacement task.
2129 lockdep_unpin_lock(&rq->lock);
2130 raw_spin_unlock(&rq->lock);
2131 raw_spin_lock(&p->pi_lock);
2132 raw_spin_lock(&rq->lock);
2133 lockdep_pin_lock(&rq->lock);
2136 if (!(p->state & TASK_NORMAL))
2139 trace_sched_waking(p);
2141 if (!task_on_rq_queued(p))
2142 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2144 ttwu_do_wakeup(rq, p, 0);
2145 ttwu_stat(p, smp_processor_id(), 0);
2147 raw_spin_unlock(&p->pi_lock);
2151 * wake_up_process - Wake up a specific process
2152 * @p: The process to be woken up.
2154 * Attempt to wake up the nominated process and move it to the set of runnable
2157 * Return: 1 if the process was woken up, 0 if it was already running.
2159 * It may be assumed that this function implies a write memory barrier before
2160 * changing the task state if and only if any tasks are woken up.
2162 int wake_up_process(struct task_struct *p)
2164 return try_to_wake_up(p, TASK_NORMAL, 0);
2166 EXPORT_SYMBOL(wake_up_process);
2168 int wake_up_state(struct task_struct *p, unsigned int state)
2170 return try_to_wake_up(p, state, 0);
2174 * This function clears the sched_dl_entity static params.
2176 void __dl_clear_params(struct task_struct *p)
2178 struct sched_dl_entity *dl_se = &p->dl;
2180 dl_se->dl_runtime = 0;
2181 dl_se->dl_deadline = 0;
2182 dl_se->dl_period = 0;
2186 dl_se->dl_throttled = 0;
2188 dl_se->dl_yielded = 0;
2192 * Perform scheduler related setup for a newly forked process p.
2193 * p is forked by current.
2195 * __sched_fork() is basic setup used by init_idle() too:
2197 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2202 p->se.exec_start = 0;
2203 p->se.sum_exec_runtime = 0;
2204 p->se.prev_sum_exec_runtime = 0;
2205 p->se.nr_migrations = 0;
2207 INIT_LIST_HEAD(&p->se.group_node);
2209 #ifdef CONFIG_FAIR_GROUP_SCHED
2210 p->se.cfs_rq = NULL;
2213 #ifdef CONFIG_SCHEDSTATS
2214 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2217 RB_CLEAR_NODE(&p->dl.rb_node);
2218 init_dl_task_timer(&p->dl);
2219 __dl_clear_params(p);
2221 INIT_LIST_HEAD(&p->rt.run_list);
2223 #ifdef CONFIG_PREEMPT_NOTIFIERS
2224 INIT_HLIST_HEAD(&p->preempt_notifiers);
2227 #ifdef CONFIG_NUMA_BALANCING
2228 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2229 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2230 p->mm->numa_scan_seq = 0;
2233 if (clone_flags & CLONE_VM)
2234 p->numa_preferred_nid = current->numa_preferred_nid;
2236 p->numa_preferred_nid = -1;
2238 p->node_stamp = 0ULL;
2239 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2240 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2241 p->numa_work.next = &p->numa_work;
2242 p->numa_faults = NULL;
2243 p->last_task_numa_placement = 0;
2244 p->last_sum_exec_runtime = 0;
2246 p->numa_group = NULL;
2247 #endif /* CONFIG_NUMA_BALANCING */
2250 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2252 #ifdef CONFIG_NUMA_BALANCING
2254 void set_numabalancing_state(bool enabled)
2257 static_branch_enable(&sched_numa_balancing);
2259 static_branch_disable(&sched_numa_balancing);
2262 #ifdef CONFIG_PROC_SYSCTL
2263 int sysctl_numa_balancing(struct ctl_table *table, int write,
2264 void __user *buffer, size_t *lenp, loff_t *ppos)
2268 int state = static_branch_likely(&sched_numa_balancing);
2270 if (write && !capable(CAP_SYS_ADMIN))
2275 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2279 set_numabalancing_state(state);
2286 * fork()/clone()-time setup:
2288 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2290 unsigned long flags;
2291 int cpu = get_cpu();
2293 __sched_fork(clone_flags, p);
2295 * We mark the process as running here. This guarantees that
2296 * nobody will actually run it, and a signal or other external
2297 * event cannot wake it up and insert it on the runqueue either.
2299 p->state = TASK_RUNNING;
2302 * Make sure we do not leak PI boosting priority to the child.
2304 p->prio = current->normal_prio;
2307 * Revert to default priority/policy on fork if requested.
2309 if (unlikely(p->sched_reset_on_fork)) {
2310 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2311 p->policy = SCHED_NORMAL;
2312 p->static_prio = NICE_TO_PRIO(0);
2314 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2315 p->static_prio = NICE_TO_PRIO(0);
2317 p->prio = p->normal_prio = __normal_prio(p);
2321 * We don't need the reset flag anymore after the fork. It has
2322 * fulfilled its duty:
2324 p->sched_reset_on_fork = 0;
2327 if (dl_prio(p->prio)) {
2330 } else if (rt_prio(p->prio)) {
2331 p->sched_class = &rt_sched_class;
2333 p->sched_class = &fair_sched_class;
2336 if (p->sched_class->task_fork)
2337 p->sched_class->task_fork(p);
2340 * The child is not yet in the pid-hash so no cgroup attach races,
2341 * and the cgroup is pinned to this child due to cgroup_fork()
2342 * is ran before sched_fork().
2344 * Silence PROVE_RCU.
2346 raw_spin_lock_irqsave(&p->pi_lock, flags);
2347 set_task_cpu(p, cpu);
2348 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2350 #ifdef CONFIG_SCHED_INFO
2351 if (likely(sched_info_on()))
2352 memset(&p->sched_info, 0, sizeof(p->sched_info));
2354 #if defined(CONFIG_SMP)
2357 init_task_preempt_count(p);
2359 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2360 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2367 unsigned long to_ratio(u64 period, u64 runtime)
2369 if (runtime == RUNTIME_INF)
2373 * Doing this here saves a lot of checks in all
2374 * the calling paths, and returning zero seems
2375 * safe for them anyway.
2380 return div64_u64(runtime << 20, period);
2384 inline struct dl_bw *dl_bw_of(int i)
2386 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2387 "sched RCU must be held");
2388 return &cpu_rq(i)->rd->dl_bw;
2391 static inline int dl_bw_cpus(int i)
2393 struct root_domain *rd = cpu_rq(i)->rd;
2396 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2397 "sched RCU must be held");
2398 for_each_cpu_and(i, rd->span, cpu_active_mask)
2404 inline struct dl_bw *dl_bw_of(int i)
2406 return &cpu_rq(i)->dl.dl_bw;
2409 static inline int dl_bw_cpus(int i)
2416 * We must be sure that accepting a new task (or allowing changing the
2417 * parameters of an existing one) is consistent with the bandwidth
2418 * constraints. If yes, this function also accordingly updates the currently
2419 * allocated bandwidth to reflect the new situation.
2421 * This function is called while holding p's rq->lock.
2423 * XXX we should delay bw change until the task's 0-lag point, see
2426 static int dl_overflow(struct task_struct *p, int policy,
2427 const struct sched_attr *attr)
2430 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2431 u64 period = attr->sched_period ?: attr->sched_deadline;
2432 u64 runtime = attr->sched_runtime;
2433 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2436 if (new_bw == p->dl.dl_bw)
2440 * Either if a task, enters, leave, or stays -deadline but changes
2441 * its parameters, we may need to update accordingly the total
2442 * allocated bandwidth of the container.
2444 raw_spin_lock(&dl_b->lock);
2445 cpus = dl_bw_cpus(task_cpu(p));
2446 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2447 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2448 __dl_add(dl_b, new_bw);
2450 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2451 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2452 __dl_clear(dl_b, p->dl.dl_bw);
2453 __dl_add(dl_b, new_bw);
2455 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2456 __dl_clear(dl_b, p->dl.dl_bw);
2459 raw_spin_unlock(&dl_b->lock);
2464 extern void init_dl_bw(struct dl_bw *dl_b);
2467 * wake_up_new_task - wake up a newly created task for the first time.
2469 * This function will do some initial scheduler statistics housekeeping
2470 * that must be done for every newly created context, then puts the task
2471 * on the runqueue and wakes it.
2473 void wake_up_new_task(struct task_struct *p)
2475 unsigned long flags;
2478 raw_spin_lock_irqsave(&p->pi_lock, flags);
2479 /* Initialize new task's runnable average */
2480 init_entity_runnable_average(&p->se);
2483 * Fork balancing, do it here and not earlier because:
2484 * - cpus_allowed can change in the fork path
2485 * - any previously selected cpu might disappear through hotplug
2487 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2490 rq = __task_rq_lock(p);
2491 activate_task(rq, p, 0);
2492 p->on_rq = TASK_ON_RQ_QUEUED;
2493 trace_sched_wakeup_new(p);
2494 check_preempt_curr(rq, p, WF_FORK);
2496 if (p->sched_class->task_woken) {
2498 * Nothing relies on rq->lock after this, so its fine to
2501 lockdep_unpin_lock(&rq->lock);
2502 p->sched_class->task_woken(rq, p);
2503 lockdep_pin_lock(&rq->lock);
2506 task_rq_unlock(rq, p, &flags);
2509 #ifdef CONFIG_PREEMPT_NOTIFIERS
2511 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2513 void preempt_notifier_inc(void)
2515 static_key_slow_inc(&preempt_notifier_key);
2517 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2519 void preempt_notifier_dec(void)
2521 static_key_slow_dec(&preempt_notifier_key);
2523 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2526 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2527 * @notifier: notifier struct to register
2529 void preempt_notifier_register(struct preempt_notifier *notifier)
2531 if (!static_key_false(&preempt_notifier_key))
2532 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2534 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2536 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2539 * preempt_notifier_unregister - no longer interested in preemption notifications
2540 * @notifier: notifier struct to unregister
2542 * This is *not* safe to call from within a preemption notifier.
2544 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2546 hlist_del(¬ifier->link);
2548 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2550 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2552 struct preempt_notifier *notifier;
2554 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2555 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2558 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2560 if (static_key_false(&preempt_notifier_key))
2561 __fire_sched_in_preempt_notifiers(curr);
2565 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2566 struct task_struct *next)
2568 struct preempt_notifier *notifier;
2570 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2571 notifier->ops->sched_out(notifier, next);
2574 static __always_inline void
2575 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2576 struct task_struct *next)
2578 if (static_key_false(&preempt_notifier_key))
2579 __fire_sched_out_preempt_notifiers(curr, next);
2582 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2584 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2589 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2590 struct task_struct *next)
2594 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2597 * prepare_task_switch - prepare to switch tasks
2598 * @rq: the runqueue preparing to switch
2599 * @prev: the current task that is being switched out
2600 * @next: the task we are going to switch to.
2602 * This is called with the rq lock held and interrupts off. It must
2603 * be paired with a subsequent finish_task_switch after the context
2606 * prepare_task_switch sets up locking and calls architecture specific
2610 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2611 struct task_struct *next)
2613 sched_info_switch(rq, prev, next);
2614 perf_event_task_sched_out(prev, next);
2615 fire_sched_out_preempt_notifiers(prev, next);
2616 prepare_lock_switch(rq, next);
2617 prepare_arch_switch(next);
2621 * finish_task_switch - clean up after a task-switch
2622 * @prev: the thread we just switched away from.
2624 * finish_task_switch must be called after the context switch, paired
2625 * with a prepare_task_switch call before the context switch.
2626 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2627 * and do any other architecture-specific cleanup actions.
2629 * Note that we may have delayed dropping an mm in context_switch(). If
2630 * so, we finish that here outside of the runqueue lock. (Doing it
2631 * with the lock held can cause deadlocks; see schedule() for
2634 * The context switch have flipped the stack from under us and restored the
2635 * local variables which were saved when this task called schedule() in the
2636 * past. prev == current is still correct but we need to recalculate this_rq
2637 * because prev may have moved to another CPU.
2639 static struct rq *finish_task_switch(struct task_struct *prev)
2640 __releases(rq->lock)
2642 struct rq *rq = this_rq();
2643 struct mm_struct *mm = rq->prev_mm;
2647 * The previous task will have left us with a preempt_count of 2
2648 * because it left us after:
2651 * preempt_disable(); // 1
2653 * raw_spin_lock_irq(&rq->lock) // 2
2655 * Also, see FORK_PREEMPT_COUNT.
2657 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2658 "corrupted preempt_count: %s/%d/0x%x\n",
2659 current->comm, current->pid, preempt_count()))
2660 preempt_count_set(FORK_PREEMPT_COUNT);
2665 * A task struct has one reference for the use as "current".
2666 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2667 * schedule one last time. The schedule call will never return, and
2668 * the scheduled task must drop that reference.
2670 * We must observe prev->state before clearing prev->on_cpu (in
2671 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2672 * running on another CPU and we could rave with its RUNNING -> DEAD
2673 * transition, resulting in a double drop.
2675 prev_state = prev->state;
2676 vtime_task_switch(prev);
2677 perf_event_task_sched_in(prev, current);
2678 finish_lock_switch(rq, prev);
2679 finish_arch_post_lock_switch();
2681 fire_sched_in_preempt_notifiers(current);
2684 if (unlikely(prev_state == TASK_DEAD)) {
2685 if (prev->sched_class->task_dead)
2686 prev->sched_class->task_dead(prev);
2689 * Remove function-return probe instances associated with this
2690 * task and put them back on the free list.
2692 kprobe_flush_task(prev);
2693 put_task_struct(prev);
2696 tick_nohz_task_switch();
2702 /* rq->lock is NOT held, but preemption is disabled */
2703 static void __balance_callback(struct rq *rq)
2705 struct callback_head *head, *next;
2706 void (*func)(struct rq *rq);
2707 unsigned long flags;
2709 raw_spin_lock_irqsave(&rq->lock, flags);
2710 head = rq->balance_callback;
2711 rq->balance_callback = NULL;
2713 func = (void (*)(struct rq *))head->func;
2720 raw_spin_unlock_irqrestore(&rq->lock, flags);
2723 static inline void balance_callback(struct rq *rq)
2725 if (unlikely(rq->balance_callback))
2726 __balance_callback(rq);
2731 static inline void balance_callback(struct rq *rq)
2738 * schedule_tail - first thing a freshly forked thread must call.
2739 * @prev: the thread we just switched away from.
2741 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2742 __releases(rq->lock)
2747 * New tasks start with FORK_PREEMPT_COUNT, see there and
2748 * finish_task_switch() for details.
2750 * finish_task_switch() will drop rq->lock() and lower preempt_count
2751 * and the preempt_enable() will end up enabling preemption (on
2752 * PREEMPT_COUNT kernels).
2755 rq = finish_task_switch(prev);
2756 balance_callback(rq);
2759 if (current->set_child_tid)
2760 put_user(task_pid_vnr(current), current->set_child_tid);
2764 * context_switch - switch to the new MM and the new thread's register state.
2766 static inline struct rq *
2767 context_switch(struct rq *rq, struct task_struct *prev,
2768 struct task_struct *next)
2770 struct mm_struct *mm, *oldmm;
2772 prepare_task_switch(rq, prev, next);
2775 oldmm = prev->active_mm;
2777 * For paravirt, this is coupled with an exit in switch_to to
2778 * combine the page table reload and the switch backend into
2781 arch_start_context_switch(prev);
2784 next->active_mm = oldmm;
2785 atomic_inc(&oldmm->mm_count);
2786 enter_lazy_tlb(oldmm, next);
2788 switch_mm(oldmm, mm, next);
2791 prev->active_mm = NULL;
2792 rq->prev_mm = oldmm;
2795 * Since the runqueue lock will be released by the next
2796 * task (which is an invalid locking op but in the case
2797 * of the scheduler it's an obvious special-case), so we
2798 * do an early lockdep release here:
2800 lockdep_unpin_lock(&rq->lock);
2801 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2803 /* Here we just switch the register state and the stack. */
2804 switch_to(prev, next, prev);
2807 return finish_task_switch(prev);
2811 * nr_running and nr_context_switches:
2813 * externally visible scheduler statistics: current number of runnable
2814 * threads, total number of context switches performed since bootup.
2816 unsigned long nr_running(void)
2818 unsigned long i, sum = 0;
2820 for_each_online_cpu(i)
2821 sum += cpu_rq(i)->nr_running;
2827 * Check if only the current task is running on the cpu.
2829 * Caution: this function does not check that the caller has disabled
2830 * preemption, thus the result might have a time-of-check-to-time-of-use
2831 * race. The caller is responsible to use it correctly, for example:
2833 * - from a non-preemptable section (of course)
2835 * - from a thread that is bound to a single CPU
2837 * - in a loop with very short iterations (e.g. a polling loop)
2839 bool single_task_running(void)
2841 return raw_rq()->nr_running == 1;
2843 EXPORT_SYMBOL(single_task_running);
2845 unsigned long long nr_context_switches(void)
2848 unsigned long long sum = 0;
2850 for_each_possible_cpu(i)
2851 sum += cpu_rq(i)->nr_switches;
2856 unsigned long nr_iowait(void)
2858 unsigned long i, sum = 0;
2860 for_each_possible_cpu(i)
2861 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2866 unsigned long nr_iowait_cpu(int cpu)
2868 struct rq *this = cpu_rq(cpu);
2869 return atomic_read(&this->nr_iowait);
2872 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2874 struct rq *rq = this_rq();
2875 *nr_waiters = atomic_read(&rq->nr_iowait);
2876 *load = rq->load.weight;
2882 * sched_exec - execve() is a valuable balancing opportunity, because at
2883 * this point the task has the smallest effective memory and cache footprint.
2885 void sched_exec(void)
2887 struct task_struct *p = current;
2888 unsigned long flags;
2891 raw_spin_lock_irqsave(&p->pi_lock, flags);
2892 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2893 if (dest_cpu == smp_processor_id())
2896 if (likely(cpu_active(dest_cpu))) {
2897 struct migration_arg arg = { p, dest_cpu };
2899 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2900 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2904 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2909 DEFINE_PER_CPU(struct kernel_stat, kstat);
2910 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2912 EXPORT_PER_CPU_SYMBOL(kstat);
2913 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2916 * Return accounted runtime for the task.
2917 * In case the task is currently running, return the runtime plus current's
2918 * pending runtime that have not been accounted yet.
2920 unsigned long long task_sched_runtime(struct task_struct *p)
2922 unsigned long flags;
2926 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2928 * 64-bit doesn't need locks to atomically read a 64bit value.
2929 * So we have a optimization chance when the task's delta_exec is 0.
2930 * Reading ->on_cpu is racy, but this is ok.
2932 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2933 * If we race with it entering cpu, unaccounted time is 0. This is
2934 * indistinguishable from the read occurring a few cycles earlier.
2935 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2936 * been accounted, so we're correct here as well.
2938 if (!p->on_cpu || !task_on_rq_queued(p))
2939 return p->se.sum_exec_runtime;
2942 rq = task_rq_lock(p, &flags);
2944 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2945 * project cycles that may never be accounted to this
2946 * thread, breaking clock_gettime().
2948 if (task_current(rq, p) && task_on_rq_queued(p)) {
2949 update_rq_clock(rq);
2950 p->sched_class->update_curr(rq);
2952 ns = p->se.sum_exec_runtime;
2953 task_rq_unlock(rq, p, &flags);
2959 * This function gets called by the timer code, with HZ frequency.
2960 * We call it with interrupts disabled.
2962 void scheduler_tick(void)
2964 int cpu = smp_processor_id();
2965 struct rq *rq = cpu_rq(cpu);
2966 struct task_struct *curr = rq->curr;
2970 raw_spin_lock(&rq->lock);
2971 update_rq_clock(rq);
2972 curr->sched_class->task_tick(rq, curr, 0);
2973 update_cpu_load_active(rq);
2974 calc_global_load_tick(rq);
2975 raw_spin_unlock(&rq->lock);
2977 perf_event_task_tick();
2980 rq->idle_balance = idle_cpu(cpu);
2981 trigger_load_balance(rq);
2983 rq_last_tick_reset(rq);
2986 #ifdef CONFIG_NO_HZ_FULL
2988 * scheduler_tick_max_deferment
2990 * Keep at least one tick per second when a single
2991 * active task is running because the scheduler doesn't
2992 * yet completely support full dynticks environment.
2994 * This makes sure that uptime, CFS vruntime, load
2995 * balancing, etc... continue to move forward, even
2996 * with a very low granularity.
2998 * Return: Maximum deferment in nanoseconds.
3000 u64 scheduler_tick_max_deferment(void)
3002 struct rq *rq = this_rq();
3003 unsigned long next, now = READ_ONCE(jiffies);
3005 next = rq->last_sched_tick + HZ;
3007 if (time_before_eq(next, now))
3010 return jiffies_to_nsecs(next - now);
3014 notrace unsigned long get_parent_ip(unsigned long addr)
3016 if (in_lock_functions(addr)) {
3017 addr = CALLER_ADDR2;
3018 if (in_lock_functions(addr))
3019 addr = CALLER_ADDR3;
3024 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3025 defined(CONFIG_PREEMPT_TRACER))
3027 void preempt_count_add(int val)
3029 #ifdef CONFIG_DEBUG_PREEMPT
3033 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3036 __preempt_count_add(val);
3037 #ifdef CONFIG_DEBUG_PREEMPT
3039 * Spinlock count overflowing soon?
3041 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3044 if (preempt_count() == val) {
3045 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3046 #ifdef CONFIG_DEBUG_PREEMPT
3047 current->preempt_disable_ip = ip;
3049 trace_preempt_off(CALLER_ADDR0, ip);
3052 EXPORT_SYMBOL(preempt_count_add);
3053 NOKPROBE_SYMBOL(preempt_count_add);
3055 void preempt_count_sub(int val)
3057 #ifdef CONFIG_DEBUG_PREEMPT
3061 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3064 * Is the spinlock portion underflowing?
3066 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3067 !(preempt_count() & PREEMPT_MASK)))
3071 if (preempt_count() == val)
3072 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3073 __preempt_count_sub(val);
3075 EXPORT_SYMBOL(preempt_count_sub);
3076 NOKPROBE_SYMBOL(preempt_count_sub);
3081 * Print scheduling while atomic bug:
3083 static noinline void __schedule_bug(struct task_struct *prev)
3085 if (oops_in_progress)
3088 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3089 prev->comm, prev->pid, preempt_count());
3091 debug_show_held_locks(prev);
3093 if (irqs_disabled())
3094 print_irqtrace_events(prev);
3095 #ifdef CONFIG_DEBUG_PREEMPT
3096 if (in_atomic_preempt_off()) {
3097 pr_err("Preemption disabled at:");
3098 print_ip_sym(current->preempt_disable_ip);
3103 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3107 * Various schedule()-time debugging checks and statistics:
3109 static inline void schedule_debug(struct task_struct *prev)
3111 #ifdef CONFIG_SCHED_STACK_END_CHECK
3112 BUG_ON(task_stack_end_corrupted(prev));
3115 if (unlikely(in_atomic_preempt_off())) {
3116 __schedule_bug(prev);
3117 preempt_count_set(PREEMPT_DISABLED);
3121 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3123 schedstat_inc(this_rq(), sched_count);
3127 * Pick up the highest-prio task:
3129 static inline struct task_struct *
3130 pick_next_task(struct rq *rq, struct task_struct *prev)
3132 const struct sched_class *class = &fair_sched_class;
3133 struct task_struct *p;
3136 * Optimization: we know that if all tasks are in
3137 * the fair class we can call that function directly:
3139 if (likely(prev->sched_class == class &&
3140 rq->nr_running == rq->cfs.h_nr_running)) {
3141 p = fair_sched_class.pick_next_task(rq, prev);
3142 if (unlikely(p == RETRY_TASK))
3145 /* assumes fair_sched_class->next == idle_sched_class */
3147 p = idle_sched_class.pick_next_task(rq, prev);
3153 for_each_class(class) {
3154 p = class->pick_next_task(rq, prev);
3156 if (unlikely(p == RETRY_TASK))
3162 BUG(); /* the idle class will always have a runnable task */
3166 * __schedule() is the main scheduler function.
3168 * The main means of driving the scheduler and thus entering this function are:
3170 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3172 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3173 * paths. For example, see arch/x86/entry_64.S.
3175 * To drive preemption between tasks, the scheduler sets the flag in timer
3176 * interrupt handler scheduler_tick().
3178 * 3. Wakeups don't really cause entry into schedule(). They add a
3179 * task to the run-queue and that's it.
3181 * Now, if the new task added to the run-queue preempts the current
3182 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3183 * called on the nearest possible occasion:
3185 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3187 * - in syscall or exception context, at the next outmost
3188 * preempt_enable(). (this might be as soon as the wake_up()'s
3191 * - in IRQ context, return from interrupt-handler to
3192 * preemptible context
3194 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3197 * - cond_resched() call
3198 * - explicit schedule() call
3199 * - return from syscall or exception to user-space
3200 * - return from interrupt-handler to user-space
3202 * WARNING: must be called with preemption disabled!
3204 static void __sched notrace __schedule(bool preempt)
3206 struct task_struct *prev, *next;
3207 unsigned long *switch_count;
3211 cpu = smp_processor_id();
3216 * do_exit() calls schedule() with preemption disabled as an exception;
3217 * however we must fix that up, otherwise the next task will see an
3218 * inconsistent (higher) preempt count.
3220 * It also avoids the below schedule_debug() test from complaining
3223 if (unlikely(prev->state == TASK_DEAD))
3224 preempt_enable_no_resched_notrace();
3226 schedule_debug(prev);
3228 if (sched_feat(HRTICK))
3231 local_irq_disable();
3232 rcu_note_context_switch();
3235 * Make sure that signal_pending_state()->signal_pending() below
3236 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3237 * done by the caller to avoid the race with signal_wake_up().
3239 smp_mb__before_spinlock();
3240 raw_spin_lock(&rq->lock);
3241 lockdep_pin_lock(&rq->lock);
3243 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3245 switch_count = &prev->nivcsw;
3246 if (!preempt && prev->state) {
3247 if (unlikely(signal_pending_state(prev->state, prev))) {
3248 prev->state = TASK_RUNNING;
3250 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3254 * If a worker went to sleep, notify and ask workqueue
3255 * whether it wants to wake up a task to maintain
3258 if (prev->flags & PF_WQ_WORKER) {
3259 struct task_struct *to_wakeup;
3261 to_wakeup = wq_worker_sleeping(prev, cpu);
3263 try_to_wake_up_local(to_wakeup);
3266 switch_count = &prev->nvcsw;
3269 if (task_on_rq_queued(prev))
3270 update_rq_clock(rq);
3272 next = pick_next_task(rq, prev);
3273 clear_tsk_need_resched(prev);
3274 clear_preempt_need_resched();
3275 rq->clock_skip_update = 0;
3277 if (likely(prev != next)) {
3282 trace_sched_switch(preempt, prev, next);
3283 rq = context_switch(rq, prev, next); /* unlocks the rq */
3286 lockdep_unpin_lock(&rq->lock);
3287 raw_spin_unlock_irq(&rq->lock);
3290 balance_callback(rq);
3293 static inline void sched_submit_work(struct task_struct *tsk)
3295 if (!tsk->state || tsk_is_pi_blocked(tsk))
3298 * If we are going to sleep and we have plugged IO queued,
3299 * make sure to submit it to avoid deadlocks.
3301 if (blk_needs_flush_plug(tsk))
3302 blk_schedule_flush_plug(tsk);
3305 asmlinkage __visible void __sched schedule(void)
3307 struct task_struct *tsk = current;
3309 sched_submit_work(tsk);
3313 sched_preempt_enable_no_resched();
3314 } while (need_resched());
3316 EXPORT_SYMBOL(schedule);
3318 #ifdef CONFIG_CONTEXT_TRACKING
3319 asmlinkage __visible void __sched schedule_user(void)
3322 * If we come here after a random call to set_need_resched(),
3323 * or we have been woken up remotely but the IPI has not yet arrived,
3324 * we haven't yet exited the RCU idle mode. Do it here manually until
3325 * we find a better solution.
3327 * NB: There are buggy callers of this function. Ideally we
3328 * should warn if prev_state != CONTEXT_USER, but that will trigger
3329 * too frequently to make sense yet.
3331 enum ctx_state prev_state = exception_enter();
3333 exception_exit(prev_state);
3338 * schedule_preempt_disabled - called with preemption disabled
3340 * Returns with preemption disabled. Note: preempt_count must be 1
3342 void __sched schedule_preempt_disabled(void)
3344 sched_preempt_enable_no_resched();
3349 static void __sched notrace preempt_schedule_common(void)
3352 preempt_disable_notrace();
3354 preempt_enable_no_resched_notrace();
3357 * Check again in case we missed a preemption opportunity
3358 * between schedule and now.
3360 } while (need_resched());
3363 #ifdef CONFIG_PREEMPT
3365 * this is the entry point to schedule() from in-kernel preemption
3366 * off of preempt_enable. Kernel preemptions off return from interrupt
3367 * occur there and call schedule directly.
3369 asmlinkage __visible void __sched notrace preempt_schedule(void)
3372 * If there is a non-zero preempt_count or interrupts are disabled,
3373 * we do not want to preempt the current task. Just return..
3375 if (likely(!preemptible()))
3378 preempt_schedule_common();
3380 NOKPROBE_SYMBOL(preempt_schedule);
3381 EXPORT_SYMBOL(preempt_schedule);
3384 * preempt_schedule_notrace - preempt_schedule called by tracing
3386 * The tracing infrastructure uses preempt_enable_notrace to prevent
3387 * recursion and tracing preempt enabling caused by the tracing
3388 * infrastructure itself. But as tracing can happen in areas coming
3389 * from userspace or just about to enter userspace, a preempt enable
3390 * can occur before user_exit() is called. This will cause the scheduler
3391 * to be called when the system is still in usermode.
3393 * To prevent this, the preempt_enable_notrace will use this function
3394 * instead of preempt_schedule() to exit user context if needed before
3395 * calling the scheduler.
3397 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3399 enum ctx_state prev_ctx;
3401 if (likely(!preemptible()))
3405 preempt_disable_notrace();
3407 * Needs preempt disabled in case user_exit() is traced
3408 * and the tracer calls preempt_enable_notrace() causing
3409 * an infinite recursion.
3411 prev_ctx = exception_enter();
3413 exception_exit(prev_ctx);
3415 preempt_enable_no_resched_notrace();
3416 } while (need_resched());
3418 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3420 #endif /* CONFIG_PREEMPT */
3423 * this is the entry point to schedule() from kernel preemption
3424 * off of irq context.
3425 * Note, that this is called and return with irqs disabled. This will
3426 * protect us against recursive calling from irq.
3428 asmlinkage __visible void __sched preempt_schedule_irq(void)
3430 enum ctx_state prev_state;
3432 /* Catch callers which need to be fixed */
3433 BUG_ON(preempt_count() || !irqs_disabled());
3435 prev_state = exception_enter();
3441 local_irq_disable();
3442 sched_preempt_enable_no_resched();
3443 } while (need_resched());
3445 exception_exit(prev_state);
3448 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3451 return try_to_wake_up(curr->private, mode, wake_flags);
3453 EXPORT_SYMBOL(default_wake_function);
3455 #ifdef CONFIG_RT_MUTEXES
3458 * rt_mutex_setprio - set the current priority of a task
3460 * @prio: prio value (kernel-internal form)
3462 * This function changes the 'effective' priority of a task. It does
3463 * not touch ->normal_prio like __setscheduler().
3465 * Used by the rt_mutex code to implement priority inheritance
3466 * logic. Call site only calls if the priority of the task changed.
3468 void rt_mutex_setprio(struct task_struct *p, int prio)
3470 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3472 const struct sched_class *prev_class;
3474 BUG_ON(prio > MAX_PRIO);
3476 rq = __task_rq_lock(p);
3479 * Idle task boosting is a nono in general. There is one
3480 * exception, when PREEMPT_RT and NOHZ is active:
3482 * The idle task calls get_next_timer_interrupt() and holds
3483 * the timer wheel base->lock on the CPU and another CPU wants
3484 * to access the timer (probably to cancel it). We can safely
3485 * ignore the boosting request, as the idle CPU runs this code
3486 * with interrupts disabled and will complete the lock
3487 * protected section without being interrupted. So there is no
3488 * real need to boost.
3490 if (unlikely(p == rq->idle)) {
3491 WARN_ON(p != rq->curr);
3492 WARN_ON(p->pi_blocked_on);
3496 trace_sched_pi_setprio(p, prio);
3498 prev_class = p->sched_class;
3499 queued = task_on_rq_queued(p);
3500 running = task_current(rq, p);
3502 dequeue_task(rq, p, DEQUEUE_SAVE);
3504 put_prev_task(rq, p);
3507 * Boosting condition are:
3508 * 1. -rt task is running and holds mutex A
3509 * --> -dl task blocks on mutex A
3511 * 2. -dl task is running and holds mutex A
3512 * --> -dl task blocks on mutex A and could preempt the
3515 if (dl_prio(prio)) {
3516 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3517 if (!dl_prio(p->normal_prio) ||
3518 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3519 p->dl.dl_boosted = 1;
3520 enqueue_flag |= ENQUEUE_REPLENISH;
3522 p->dl.dl_boosted = 0;
3523 p->sched_class = &dl_sched_class;
3524 } else if (rt_prio(prio)) {
3525 if (dl_prio(oldprio))
3526 p->dl.dl_boosted = 0;
3528 enqueue_flag |= ENQUEUE_HEAD;
3529 p->sched_class = &rt_sched_class;
3531 if (dl_prio(oldprio))
3532 p->dl.dl_boosted = 0;
3533 if (rt_prio(oldprio))
3535 p->sched_class = &fair_sched_class;
3541 p->sched_class->set_curr_task(rq);
3543 enqueue_task(rq, p, enqueue_flag);
3545 check_class_changed(rq, p, prev_class, oldprio);
3547 preempt_disable(); /* avoid rq from going away on us */
3548 __task_rq_unlock(rq);
3550 balance_callback(rq);
3555 void set_user_nice(struct task_struct *p, long nice)
3557 int old_prio, delta, queued;
3558 unsigned long flags;
3561 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3564 * We have to be careful, if called from sys_setpriority(),
3565 * the task might be in the middle of scheduling on another CPU.
3567 rq = task_rq_lock(p, &flags);
3569 * The RT priorities are set via sched_setscheduler(), but we still
3570 * allow the 'normal' nice value to be set - but as expected
3571 * it wont have any effect on scheduling until the task is
3572 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3574 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3575 p->static_prio = NICE_TO_PRIO(nice);
3578 queued = task_on_rq_queued(p);
3580 dequeue_task(rq, p, DEQUEUE_SAVE);
3582 p->static_prio = NICE_TO_PRIO(nice);
3585 p->prio = effective_prio(p);
3586 delta = p->prio - old_prio;
3589 enqueue_task(rq, p, ENQUEUE_RESTORE);
3591 * If the task increased its priority or is running and
3592 * lowered its priority, then reschedule its CPU:
3594 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3598 task_rq_unlock(rq, p, &flags);
3600 EXPORT_SYMBOL(set_user_nice);
3603 * can_nice - check if a task can reduce its nice value
3607 int can_nice(const struct task_struct *p, const int nice)
3609 /* convert nice value [19,-20] to rlimit style value [1,40] */
3610 int nice_rlim = nice_to_rlimit(nice);
3612 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3613 capable(CAP_SYS_NICE));
3616 #ifdef __ARCH_WANT_SYS_NICE
3619 * sys_nice - change the priority of the current process.
3620 * @increment: priority increment
3622 * sys_setpriority is a more generic, but much slower function that
3623 * does similar things.
3625 SYSCALL_DEFINE1(nice, int, increment)
3630 * Setpriority might change our priority at the same moment.
3631 * We don't have to worry. Conceptually one call occurs first
3632 * and we have a single winner.
3634 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3635 nice = task_nice(current) + increment;
3637 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3638 if (increment < 0 && !can_nice(current, nice))
3641 retval = security_task_setnice(current, nice);
3645 set_user_nice(current, nice);
3652 * task_prio - return the priority value of a given task.
3653 * @p: the task in question.
3655 * Return: The priority value as seen by users in /proc.
3656 * RT tasks are offset by -200. Normal tasks are centered
3657 * around 0, value goes from -16 to +15.
3659 int task_prio(const struct task_struct *p)
3661 return p->prio - MAX_RT_PRIO;
3665 * idle_cpu - is a given cpu idle currently?
3666 * @cpu: the processor in question.
3668 * Return: 1 if the CPU is currently idle. 0 otherwise.
3670 int idle_cpu(int cpu)
3672 struct rq *rq = cpu_rq(cpu);
3674 if (rq->curr != rq->idle)
3681 if (!llist_empty(&rq->wake_list))
3689 * idle_task - return the idle task for a given cpu.
3690 * @cpu: the processor in question.
3692 * Return: The idle task for the cpu @cpu.
3694 struct task_struct *idle_task(int cpu)
3696 return cpu_rq(cpu)->idle;
3700 * find_process_by_pid - find a process with a matching PID value.
3701 * @pid: the pid in question.
3703 * The task of @pid, if found. %NULL otherwise.
3705 static struct task_struct *find_process_by_pid(pid_t pid)
3707 return pid ? find_task_by_vpid(pid) : current;
3711 * This function initializes the sched_dl_entity of a newly becoming
3712 * SCHED_DEADLINE task.
3714 * Only the static values are considered here, the actual runtime and the
3715 * absolute deadline will be properly calculated when the task is enqueued
3716 * for the first time with its new policy.
3719 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3721 struct sched_dl_entity *dl_se = &p->dl;
3723 dl_se->dl_runtime = attr->sched_runtime;
3724 dl_se->dl_deadline = attr->sched_deadline;
3725 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3726 dl_se->flags = attr->sched_flags;
3727 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3730 * Changing the parameters of a task is 'tricky' and we're not doing
3731 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3733 * What we SHOULD do is delay the bandwidth release until the 0-lag
3734 * point. This would include retaining the task_struct until that time
3735 * and change dl_overflow() to not immediately decrement the current
3738 * Instead we retain the current runtime/deadline and let the new
3739 * parameters take effect after the current reservation period lapses.
3740 * This is safe (albeit pessimistic) because the 0-lag point is always
3741 * before the current scheduling deadline.
3743 * We can still have temporary overloads because we do not delay the
3744 * change in bandwidth until that time; so admission control is
3745 * not on the safe side. It does however guarantee tasks will never
3746 * consume more than promised.
3751 * sched_setparam() passes in -1 for its policy, to let the functions
3752 * it calls know not to change it.
3754 #define SETPARAM_POLICY -1
3756 static void __setscheduler_params(struct task_struct *p,
3757 const struct sched_attr *attr)
3759 int policy = attr->sched_policy;
3761 if (policy == SETPARAM_POLICY)
3766 if (dl_policy(policy))
3767 __setparam_dl(p, attr);
3768 else if (fair_policy(policy))
3769 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3772 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3773 * !rt_policy. Always setting this ensures that things like
3774 * getparam()/getattr() don't report silly values for !rt tasks.
3776 p->rt_priority = attr->sched_priority;
3777 p->normal_prio = normal_prio(p);
3781 /* Actually do priority change: must hold pi & rq lock. */
3782 static void __setscheduler(struct rq *rq, struct task_struct *p,
3783 const struct sched_attr *attr, bool keep_boost)
3785 __setscheduler_params(p, attr);
3788 * Keep a potential priority boosting if called from
3789 * sched_setscheduler().
3792 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3794 p->prio = normal_prio(p);
3796 if (dl_prio(p->prio))
3797 p->sched_class = &dl_sched_class;
3798 else if (rt_prio(p->prio))
3799 p->sched_class = &rt_sched_class;
3801 p->sched_class = &fair_sched_class;
3805 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3807 struct sched_dl_entity *dl_se = &p->dl;
3809 attr->sched_priority = p->rt_priority;
3810 attr->sched_runtime = dl_se->dl_runtime;
3811 attr->sched_deadline = dl_se->dl_deadline;
3812 attr->sched_period = dl_se->dl_period;
3813 attr->sched_flags = dl_se->flags;
3817 * This function validates the new parameters of a -deadline task.
3818 * We ask for the deadline not being zero, and greater or equal
3819 * than the runtime, as well as the period of being zero or
3820 * greater than deadline. Furthermore, we have to be sure that
3821 * user parameters are above the internal resolution of 1us (we
3822 * check sched_runtime only since it is always the smaller one) and
3823 * below 2^63 ns (we have to check both sched_deadline and
3824 * sched_period, as the latter can be zero).
3827 __checkparam_dl(const struct sched_attr *attr)
3830 if (attr->sched_deadline == 0)
3834 * Since we truncate DL_SCALE bits, make sure we're at least
3837 if (attr->sched_runtime < (1ULL << DL_SCALE))
3841 * Since we use the MSB for wrap-around and sign issues, make
3842 * sure it's not set (mind that period can be equal to zero).
3844 if (attr->sched_deadline & (1ULL << 63) ||
3845 attr->sched_period & (1ULL << 63))
3848 /* runtime <= deadline <= period (if period != 0) */
3849 if ((attr->sched_period != 0 &&
3850 attr->sched_period < attr->sched_deadline) ||
3851 attr->sched_deadline < attr->sched_runtime)
3858 * check the target process has a UID that matches the current process's
3860 static bool check_same_owner(struct task_struct *p)
3862 const struct cred *cred = current_cred(), *pcred;
3866 pcred = __task_cred(p);
3867 match = (uid_eq(cred->euid, pcred->euid) ||
3868 uid_eq(cred->euid, pcred->uid));
3873 static bool dl_param_changed(struct task_struct *p,
3874 const struct sched_attr *attr)
3876 struct sched_dl_entity *dl_se = &p->dl;
3878 if (dl_se->dl_runtime != attr->sched_runtime ||
3879 dl_se->dl_deadline != attr->sched_deadline ||
3880 dl_se->dl_period != attr->sched_period ||
3881 dl_se->flags != attr->sched_flags)
3887 static int __sched_setscheduler(struct task_struct *p,
3888 const struct sched_attr *attr,
3891 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3892 MAX_RT_PRIO - 1 - attr->sched_priority;
3893 int retval, oldprio, oldpolicy = -1, queued, running;
3894 int new_effective_prio, policy = attr->sched_policy;
3895 unsigned long flags;
3896 const struct sched_class *prev_class;
3900 /* may grab non-irq protected spin_locks */
3901 BUG_ON(in_interrupt());
3903 /* double check policy once rq lock held */
3905 reset_on_fork = p->sched_reset_on_fork;
3906 policy = oldpolicy = p->policy;
3908 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3910 if (!valid_policy(policy))
3914 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3918 * Valid priorities for SCHED_FIFO and SCHED_RR are
3919 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3920 * SCHED_BATCH and SCHED_IDLE is 0.
3922 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3923 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3925 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3926 (rt_policy(policy) != (attr->sched_priority != 0)))
3930 * Allow unprivileged RT tasks to decrease priority:
3932 if (user && !capable(CAP_SYS_NICE)) {
3933 if (fair_policy(policy)) {
3934 if (attr->sched_nice < task_nice(p) &&
3935 !can_nice(p, attr->sched_nice))
3939 if (rt_policy(policy)) {
3940 unsigned long rlim_rtprio =
3941 task_rlimit(p, RLIMIT_RTPRIO);
3943 /* can't set/change the rt policy */
3944 if (policy != p->policy && !rlim_rtprio)
3947 /* can't increase priority */
3948 if (attr->sched_priority > p->rt_priority &&
3949 attr->sched_priority > rlim_rtprio)
3954 * Can't set/change SCHED_DEADLINE policy at all for now
3955 * (safest behavior); in the future we would like to allow
3956 * unprivileged DL tasks to increase their relative deadline
3957 * or reduce their runtime (both ways reducing utilization)
3959 if (dl_policy(policy))
3963 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3964 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3966 if (idle_policy(p->policy) && !idle_policy(policy)) {
3967 if (!can_nice(p, task_nice(p)))
3971 /* can't change other user's priorities */
3972 if (!check_same_owner(p))
3975 /* Normal users shall not reset the sched_reset_on_fork flag */
3976 if (p->sched_reset_on_fork && !reset_on_fork)
3981 retval = security_task_setscheduler(p);
3987 * make sure no PI-waiters arrive (or leave) while we are
3988 * changing the priority of the task:
3990 * To be able to change p->policy safely, the appropriate
3991 * runqueue lock must be held.
3993 rq = task_rq_lock(p, &flags);
3996 * Changing the policy of the stop threads its a very bad idea
3998 if (p == rq->stop) {
3999 task_rq_unlock(rq, p, &flags);
4004 * If not changing anything there's no need to proceed further,
4005 * but store a possible modification of reset_on_fork.
4007 if (unlikely(policy == p->policy)) {
4008 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4010 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4012 if (dl_policy(policy) && dl_param_changed(p, attr))
4015 p->sched_reset_on_fork = reset_on_fork;
4016 task_rq_unlock(rq, p, &flags);
4022 #ifdef CONFIG_RT_GROUP_SCHED
4024 * Do not allow realtime tasks into groups that have no runtime
4027 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4028 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4029 !task_group_is_autogroup(task_group(p))) {
4030 task_rq_unlock(rq, p, &flags);
4035 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4036 cpumask_t *span = rq->rd->span;
4039 * Don't allow tasks with an affinity mask smaller than
4040 * the entire root_domain to become SCHED_DEADLINE. We
4041 * will also fail if there's no bandwidth available.
4043 if (!cpumask_subset(span, &p->cpus_allowed) ||
4044 rq->rd->dl_bw.bw == 0) {
4045 task_rq_unlock(rq, p, &flags);
4052 /* recheck policy now with rq lock held */
4053 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4054 policy = oldpolicy = -1;
4055 task_rq_unlock(rq, p, &flags);
4060 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4061 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4064 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4065 task_rq_unlock(rq, p, &flags);
4069 p->sched_reset_on_fork = reset_on_fork;
4074 * Take priority boosted tasks into account. If the new
4075 * effective priority is unchanged, we just store the new
4076 * normal parameters and do not touch the scheduler class and
4077 * the runqueue. This will be done when the task deboost
4080 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4081 if (new_effective_prio == oldprio) {
4082 __setscheduler_params(p, attr);
4083 task_rq_unlock(rq, p, &flags);
4088 queued = task_on_rq_queued(p);
4089 running = task_current(rq, p);
4091 dequeue_task(rq, p, DEQUEUE_SAVE);
4093 put_prev_task(rq, p);
4095 prev_class = p->sched_class;
4096 __setscheduler(rq, p, attr, pi);
4099 p->sched_class->set_curr_task(rq);
4101 int enqueue_flags = ENQUEUE_RESTORE;
4103 * We enqueue to tail when the priority of a task is
4104 * increased (user space view).
4106 if (oldprio <= p->prio)
4107 enqueue_flags |= ENQUEUE_HEAD;
4109 enqueue_task(rq, p, enqueue_flags);
4112 check_class_changed(rq, p, prev_class, oldprio);
4113 preempt_disable(); /* avoid rq from going away on us */
4114 task_rq_unlock(rq, p, &flags);
4117 rt_mutex_adjust_pi(p);
4120 * Run balance callbacks after we've adjusted the PI chain.
4122 balance_callback(rq);
4128 static int _sched_setscheduler(struct task_struct *p, int policy,
4129 const struct sched_param *param, bool check)
4131 struct sched_attr attr = {
4132 .sched_policy = policy,
4133 .sched_priority = param->sched_priority,
4134 .sched_nice = PRIO_TO_NICE(p->static_prio),
4137 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4138 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4139 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4140 policy &= ~SCHED_RESET_ON_FORK;
4141 attr.sched_policy = policy;
4144 return __sched_setscheduler(p, &attr, check, true);
4147 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4148 * @p: the task in question.
4149 * @policy: new policy.
4150 * @param: structure containing the new RT priority.
4152 * Return: 0 on success. An error code otherwise.
4154 * NOTE that the task may be already dead.
4156 int sched_setscheduler(struct task_struct *p, int policy,
4157 const struct sched_param *param)
4159 return _sched_setscheduler(p, policy, param, true);
4161 EXPORT_SYMBOL_GPL(sched_setscheduler);
4163 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4165 return __sched_setscheduler(p, attr, true, true);
4167 EXPORT_SYMBOL_GPL(sched_setattr);
4170 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4171 * @p: the task in question.
4172 * @policy: new policy.
4173 * @param: structure containing the new RT priority.
4175 * Just like sched_setscheduler, only don't bother checking if the
4176 * current context has permission. For example, this is needed in
4177 * stop_machine(): we create temporary high priority worker threads,
4178 * but our caller might not have that capability.
4180 * Return: 0 on success. An error code otherwise.
4182 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4183 const struct sched_param *param)
4185 return _sched_setscheduler(p, policy, param, false);
4187 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4190 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4192 struct sched_param lparam;
4193 struct task_struct *p;
4196 if (!param || pid < 0)
4198 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4203 p = find_process_by_pid(pid);
4205 retval = sched_setscheduler(p, policy, &lparam);
4212 * Mimics kernel/events/core.c perf_copy_attr().
4214 static int sched_copy_attr(struct sched_attr __user *uattr,
4215 struct sched_attr *attr)
4220 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4224 * zero the full structure, so that a short copy will be nice.
4226 memset(attr, 0, sizeof(*attr));
4228 ret = get_user(size, &uattr->size);
4232 if (size > PAGE_SIZE) /* silly large */
4235 if (!size) /* abi compat */
4236 size = SCHED_ATTR_SIZE_VER0;
4238 if (size < SCHED_ATTR_SIZE_VER0)
4242 * If we're handed a bigger struct than we know of,
4243 * ensure all the unknown bits are 0 - i.e. new
4244 * user-space does not rely on any kernel feature
4245 * extensions we dont know about yet.
4247 if (size > sizeof(*attr)) {
4248 unsigned char __user *addr;
4249 unsigned char __user *end;
4252 addr = (void __user *)uattr + sizeof(*attr);
4253 end = (void __user *)uattr + size;
4255 for (; addr < end; addr++) {
4256 ret = get_user(val, addr);
4262 size = sizeof(*attr);
4265 ret = copy_from_user(attr, uattr, size);
4270 * XXX: do we want to be lenient like existing syscalls; or do we want
4271 * to be strict and return an error on out-of-bounds values?
4273 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4278 put_user(sizeof(*attr), &uattr->size);
4283 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4284 * @pid: the pid in question.
4285 * @policy: new policy.
4286 * @param: structure containing the new RT priority.
4288 * Return: 0 on success. An error code otherwise.
4290 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4291 struct sched_param __user *, param)
4293 /* negative values for policy are not valid */
4297 return do_sched_setscheduler(pid, policy, param);
4301 * sys_sched_setparam - set/change the RT priority of a thread
4302 * @pid: the pid in question.
4303 * @param: structure containing the new RT priority.
4305 * Return: 0 on success. An error code otherwise.
4307 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4309 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4313 * sys_sched_setattr - same as above, but with extended sched_attr
4314 * @pid: the pid in question.
4315 * @uattr: structure containing the extended parameters.
4316 * @flags: for future extension.
4318 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4319 unsigned int, flags)
4321 struct sched_attr attr;
4322 struct task_struct *p;
4325 if (!uattr || pid < 0 || flags)
4328 retval = sched_copy_attr(uattr, &attr);
4332 if ((int)attr.sched_policy < 0)
4337 p = find_process_by_pid(pid);
4339 retval = sched_setattr(p, &attr);
4346 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4347 * @pid: the pid in question.
4349 * Return: On success, the policy of the thread. Otherwise, a negative error
4352 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4354 struct task_struct *p;
4362 p = find_process_by_pid(pid);
4364 retval = security_task_getscheduler(p);
4367 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4374 * sys_sched_getparam - get the RT priority of a thread
4375 * @pid: the pid in question.
4376 * @param: structure containing the RT priority.
4378 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4381 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4383 struct sched_param lp = { .sched_priority = 0 };
4384 struct task_struct *p;
4387 if (!param || pid < 0)
4391 p = find_process_by_pid(pid);
4396 retval = security_task_getscheduler(p);
4400 if (task_has_rt_policy(p))
4401 lp.sched_priority = p->rt_priority;
4405 * This one might sleep, we cannot do it with a spinlock held ...
4407 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4416 static int sched_read_attr(struct sched_attr __user *uattr,
4417 struct sched_attr *attr,
4422 if (!access_ok(VERIFY_WRITE, uattr, usize))
4426 * If we're handed a smaller struct than we know of,
4427 * ensure all the unknown bits are 0 - i.e. old
4428 * user-space does not get uncomplete information.
4430 if (usize < sizeof(*attr)) {
4431 unsigned char *addr;
4434 addr = (void *)attr + usize;
4435 end = (void *)attr + sizeof(*attr);
4437 for (; addr < end; addr++) {
4445 ret = copy_to_user(uattr, attr, attr->size);
4453 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4454 * @pid: the pid in question.
4455 * @uattr: structure containing the extended parameters.
4456 * @size: sizeof(attr) for fwd/bwd comp.
4457 * @flags: for future extension.
4459 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4460 unsigned int, size, unsigned int, flags)
4462 struct sched_attr attr = {
4463 .size = sizeof(struct sched_attr),
4465 struct task_struct *p;
4468 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4469 size < SCHED_ATTR_SIZE_VER0 || flags)
4473 p = find_process_by_pid(pid);
4478 retval = security_task_getscheduler(p);
4482 attr.sched_policy = p->policy;
4483 if (p->sched_reset_on_fork)
4484 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4485 if (task_has_dl_policy(p))
4486 __getparam_dl(p, &attr);
4487 else if (task_has_rt_policy(p))
4488 attr.sched_priority = p->rt_priority;
4490 attr.sched_nice = task_nice(p);
4494 retval = sched_read_attr(uattr, &attr, size);
4502 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4504 cpumask_var_t cpus_allowed, new_mask;
4505 struct task_struct *p;
4510 p = find_process_by_pid(pid);
4516 /* Prevent p going away */
4520 if (p->flags & PF_NO_SETAFFINITY) {
4524 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4528 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4530 goto out_free_cpus_allowed;
4533 if (!check_same_owner(p)) {
4535 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4537 goto out_free_new_mask;
4542 retval = security_task_setscheduler(p);
4544 goto out_free_new_mask;
4547 cpuset_cpus_allowed(p, cpus_allowed);
4548 cpumask_and(new_mask, in_mask, cpus_allowed);
4551 * Since bandwidth control happens on root_domain basis,
4552 * if admission test is enabled, we only admit -deadline
4553 * tasks allowed to run on all the CPUs in the task's
4557 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4559 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4562 goto out_free_new_mask;
4568 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4571 cpuset_cpus_allowed(p, cpus_allowed);
4572 if (!cpumask_subset(new_mask, cpus_allowed)) {
4574 * We must have raced with a concurrent cpuset
4575 * update. Just reset the cpus_allowed to the
4576 * cpuset's cpus_allowed
4578 cpumask_copy(new_mask, cpus_allowed);
4583 free_cpumask_var(new_mask);
4584 out_free_cpus_allowed:
4585 free_cpumask_var(cpus_allowed);
4591 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4592 struct cpumask *new_mask)
4594 if (len < cpumask_size())
4595 cpumask_clear(new_mask);
4596 else if (len > cpumask_size())
4597 len = cpumask_size();
4599 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4603 * sys_sched_setaffinity - set the cpu affinity of a process
4604 * @pid: pid of the process
4605 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4606 * @user_mask_ptr: user-space pointer to the new cpu mask
4608 * Return: 0 on success. An error code otherwise.
4610 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4611 unsigned long __user *, user_mask_ptr)
4613 cpumask_var_t new_mask;
4616 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4619 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4621 retval = sched_setaffinity(pid, new_mask);
4622 free_cpumask_var(new_mask);
4626 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4628 struct task_struct *p;
4629 unsigned long flags;
4635 p = find_process_by_pid(pid);
4639 retval = security_task_getscheduler(p);
4643 raw_spin_lock_irqsave(&p->pi_lock, flags);
4644 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4645 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4654 * sys_sched_getaffinity - get the cpu affinity of a process
4655 * @pid: pid of the process
4656 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4657 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4659 * Return: 0 on success. An error code otherwise.
4661 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4662 unsigned long __user *, user_mask_ptr)
4667 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4669 if (len & (sizeof(unsigned long)-1))
4672 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4675 ret = sched_getaffinity(pid, mask);
4677 size_t retlen = min_t(size_t, len, cpumask_size());
4679 if (copy_to_user(user_mask_ptr, mask, retlen))
4684 free_cpumask_var(mask);
4690 * sys_sched_yield - yield the current processor to other threads.
4692 * This function yields the current CPU to other tasks. If there are no
4693 * other threads running on this CPU then this function will return.
4697 SYSCALL_DEFINE0(sched_yield)
4699 struct rq *rq = this_rq_lock();
4701 schedstat_inc(rq, yld_count);
4702 current->sched_class->yield_task(rq);
4705 * Since we are going to call schedule() anyway, there's
4706 * no need to preempt or enable interrupts:
4708 __release(rq->lock);
4709 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4710 do_raw_spin_unlock(&rq->lock);
4711 sched_preempt_enable_no_resched();
4718 int __sched _cond_resched(void)
4720 if (should_resched(0)) {
4721 preempt_schedule_common();
4726 EXPORT_SYMBOL(_cond_resched);
4729 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4730 * call schedule, and on return reacquire the lock.
4732 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4733 * operations here to prevent schedule() from being called twice (once via
4734 * spin_unlock(), once by hand).
4736 int __cond_resched_lock(spinlock_t *lock)
4738 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4741 lockdep_assert_held(lock);
4743 if (spin_needbreak(lock) || resched) {
4746 preempt_schedule_common();
4754 EXPORT_SYMBOL(__cond_resched_lock);
4756 int __sched __cond_resched_softirq(void)
4758 BUG_ON(!in_softirq());
4760 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4762 preempt_schedule_common();
4768 EXPORT_SYMBOL(__cond_resched_softirq);
4771 * yield - yield the current processor to other threads.
4773 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4775 * The scheduler is at all times free to pick the calling task as the most
4776 * eligible task to run, if removing the yield() call from your code breaks
4777 * it, its already broken.
4779 * Typical broken usage is:
4784 * where one assumes that yield() will let 'the other' process run that will
4785 * make event true. If the current task is a SCHED_FIFO task that will never
4786 * happen. Never use yield() as a progress guarantee!!
4788 * If you want to use yield() to wait for something, use wait_event().
4789 * If you want to use yield() to be 'nice' for others, use cond_resched().
4790 * If you still want to use yield(), do not!
4792 void __sched yield(void)
4794 set_current_state(TASK_RUNNING);
4797 EXPORT_SYMBOL(yield);
4800 * yield_to - yield the current processor to another thread in
4801 * your thread group, or accelerate that thread toward the
4802 * processor it's on.
4804 * @preempt: whether task preemption is allowed or not
4806 * It's the caller's job to ensure that the target task struct
4807 * can't go away on us before we can do any checks.
4810 * true (>0) if we indeed boosted the target task.
4811 * false (0) if we failed to boost the target.
4812 * -ESRCH if there's no task to yield to.
4814 int __sched yield_to(struct task_struct *p, bool preempt)
4816 struct task_struct *curr = current;
4817 struct rq *rq, *p_rq;
4818 unsigned long flags;
4821 local_irq_save(flags);
4827 * If we're the only runnable task on the rq and target rq also
4828 * has only one task, there's absolutely no point in yielding.
4830 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4835 double_rq_lock(rq, p_rq);
4836 if (task_rq(p) != p_rq) {
4837 double_rq_unlock(rq, p_rq);
4841 if (!curr->sched_class->yield_to_task)
4844 if (curr->sched_class != p->sched_class)
4847 if (task_running(p_rq, p) || p->state)
4850 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4852 schedstat_inc(rq, yld_count);
4854 * Make p's CPU reschedule; pick_next_entity takes care of
4857 if (preempt && rq != p_rq)
4862 double_rq_unlock(rq, p_rq);
4864 local_irq_restore(flags);
4871 EXPORT_SYMBOL_GPL(yield_to);
4874 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4875 * that process accounting knows that this is a task in IO wait state.
4877 long __sched io_schedule_timeout(long timeout)
4879 int old_iowait = current->in_iowait;
4883 current->in_iowait = 1;
4884 blk_schedule_flush_plug(current);
4886 delayacct_blkio_start();
4888 atomic_inc(&rq->nr_iowait);
4889 ret = schedule_timeout(timeout);
4890 current->in_iowait = old_iowait;
4891 atomic_dec(&rq->nr_iowait);
4892 delayacct_blkio_end();
4896 EXPORT_SYMBOL(io_schedule_timeout);
4899 * sys_sched_get_priority_max - return maximum RT priority.
4900 * @policy: scheduling class.
4902 * Return: On success, this syscall returns the maximum
4903 * rt_priority that can be used by a given scheduling class.
4904 * On failure, a negative error code is returned.
4906 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4913 ret = MAX_USER_RT_PRIO-1;
4915 case SCHED_DEADLINE:
4926 * sys_sched_get_priority_min - return minimum RT priority.
4927 * @policy: scheduling class.
4929 * Return: On success, this syscall returns the minimum
4930 * rt_priority that can be used by a given scheduling class.
4931 * On failure, a negative error code is returned.
4933 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4942 case SCHED_DEADLINE:
4952 * sys_sched_rr_get_interval - return the default timeslice of a process.
4953 * @pid: pid of the process.
4954 * @interval: userspace pointer to the timeslice value.
4956 * this syscall writes the default timeslice value of a given process
4957 * into the user-space timespec buffer. A value of '0' means infinity.
4959 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4962 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4963 struct timespec __user *, interval)
4965 struct task_struct *p;
4966 unsigned int time_slice;
4967 unsigned long flags;
4977 p = find_process_by_pid(pid);
4981 retval = security_task_getscheduler(p);
4985 rq = task_rq_lock(p, &flags);
4987 if (p->sched_class->get_rr_interval)
4988 time_slice = p->sched_class->get_rr_interval(rq, p);
4989 task_rq_unlock(rq, p, &flags);
4992 jiffies_to_timespec(time_slice, &t);
4993 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5001 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5003 void sched_show_task(struct task_struct *p)
5005 unsigned long free = 0;
5007 unsigned long state = p->state;
5010 state = __ffs(state) + 1;
5011 printk(KERN_INFO "%-15.15s %c", p->comm,
5012 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5013 #if BITS_PER_LONG == 32
5014 if (state == TASK_RUNNING)
5015 printk(KERN_CONT " running ");
5017 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5019 if (state == TASK_RUNNING)
5020 printk(KERN_CONT " running task ");
5022 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5024 #ifdef CONFIG_DEBUG_STACK_USAGE
5025 free = stack_not_used(p);
5030 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5032 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5033 task_pid_nr(p), ppid,
5034 (unsigned long)task_thread_info(p)->flags);
5036 print_worker_info(KERN_INFO, p);
5037 show_stack(p, NULL);
5040 void show_state_filter(unsigned long state_filter)
5042 struct task_struct *g, *p;
5044 #if BITS_PER_LONG == 32
5046 " task PC stack pid father\n");
5049 " task PC stack pid father\n");
5052 for_each_process_thread(g, p) {
5054 * reset the NMI-timeout, listing all files on a slow
5055 * console might take a lot of time:
5057 touch_nmi_watchdog();
5058 if (!state_filter || (p->state & state_filter))
5062 touch_all_softlockup_watchdogs();
5064 #ifdef CONFIG_SCHED_DEBUG
5065 sysrq_sched_debug_show();
5069 * Only show locks if all tasks are dumped:
5072 debug_show_all_locks();
5075 void init_idle_bootup_task(struct task_struct *idle)
5077 idle->sched_class = &idle_sched_class;
5081 * init_idle - set up an idle thread for a given CPU
5082 * @idle: task in question
5083 * @cpu: cpu the idle task belongs to
5085 * NOTE: this function does not set the idle thread's NEED_RESCHED
5086 * flag, to make booting more robust.
5088 void init_idle(struct task_struct *idle, int cpu)
5090 struct rq *rq = cpu_rq(cpu);
5091 unsigned long flags;
5093 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5094 raw_spin_lock(&rq->lock);
5096 __sched_fork(0, idle);
5097 idle->state = TASK_RUNNING;
5098 idle->se.exec_start = sched_clock();
5100 kasan_unpoison_task_stack(idle);
5104 * Its possible that init_idle() gets called multiple times on a task,
5105 * in that case do_set_cpus_allowed() will not do the right thing.
5107 * And since this is boot we can forgo the serialization.
5109 set_cpus_allowed_common(idle, cpumask_of(cpu));
5112 * We're having a chicken and egg problem, even though we are
5113 * holding rq->lock, the cpu isn't yet set to this cpu so the
5114 * lockdep check in task_group() will fail.
5116 * Similar case to sched_fork(). / Alternatively we could
5117 * use task_rq_lock() here and obtain the other rq->lock.
5122 __set_task_cpu(idle, cpu);
5125 rq->curr = rq->idle = idle;
5126 idle->on_rq = TASK_ON_RQ_QUEUED;
5130 raw_spin_unlock(&rq->lock);
5131 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5133 /* Set the preempt count _outside_ the spinlocks! */
5134 init_idle_preempt_count(idle, cpu);
5137 * The idle tasks have their own, simple scheduling class:
5139 idle->sched_class = &idle_sched_class;
5140 ftrace_graph_init_idle_task(idle, cpu);
5141 vtime_init_idle(idle, cpu);
5143 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5147 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5148 const struct cpumask *trial)
5150 int ret = 1, trial_cpus;
5151 struct dl_bw *cur_dl_b;
5152 unsigned long flags;
5154 if (!cpumask_weight(cur))
5157 rcu_read_lock_sched();
5158 cur_dl_b = dl_bw_of(cpumask_any(cur));
5159 trial_cpus = cpumask_weight(trial);
5161 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5162 if (cur_dl_b->bw != -1 &&
5163 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5165 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5166 rcu_read_unlock_sched();
5171 int task_can_attach(struct task_struct *p,
5172 const struct cpumask *cs_cpus_allowed)
5177 * Kthreads which disallow setaffinity shouldn't be moved
5178 * to a new cpuset; we don't want to change their cpu
5179 * affinity and isolating such threads by their set of
5180 * allowed nodes is unnecessary. Thus, cpusets are not
5181 * applicable for such threads. This prevents checking for
5182 * success of set_cpus_allowed_ptr() on all attached tasks
5183 * before cpus_allowed may be changed.
5185 if (p->flags & PF_NO_SETAFFINITY) {
5191 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5193 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5198 unsigned long flags;
5200 rcu_read_lock_sched();
5201 dl_b = dl_bw_of(dest_cpu);
5202 raw_spin_lock_irqsave(&dl_b->lock, flags);
5203 cpus = dl_bw_cpus(dest_cpu);
5204 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5209 * We reserve space for this task in the destination
5210 * root_domain, as we can't fail after this point.
5211 * We will free resources in the source root_domain
5212 * later on (see set_cpus_allowed_dl()).
5214 __dl_add(dl_b, p->dl.dl_bw);
5216 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5217 rcu_read_unlock_sched();
5227 #ifdef CONFIG_NUMA_BALANCING
5228 /* Migrate current task p to target_cpu */
5229 int migrate_task_to(struct task_struct *p, int target_cpu)
5231 struct migration_arg arg = { p, target_cpu };
5232 int curr_cpu = task_cpu(p);
5234 if (curr_cpu == target_cpu)
5237 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5240 /* TODO: This is not properly updating schedstats */
5242 trace_sched_move_numa(p, curr_cpu, target_cpu);
5243 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5247 * Requeue a task on a given node and accurately track the number of NUMA
5248 * tasks on the runqueues
5250 void sched_setnuma(struct task_struct *p, int nid)
5253 unsigned long flags;
5254 bool queued, running;
5256 rq = task_rq_lock(p, &flags);
5257 queued = task_on_rq_queued(p);
5258 running = task_current(rq, p);
5261 dequeue_task(rq, p, DEQUEUE_SAVE);
5263 put_prev_task(rq, p);
5265 p->numa_preferred_nid = nid;
5268 p->sched_class->set_curr_task(rq);
5270 enqueue_task(rq, p, ENQUEUE_RESTORE);
5271 task_rq_unlock(rq, p, &flags);
5273 #endif /* CONFIG_NUMA_BALANCING */
5275 #ifdef CONFIG_HOTPLUG_CPU
5277 * Ensures that the idle task is using init_mm right before its cpu goes
5280 void idle_task_exit(void)
5282 struct mm_struct *mm = current->active_mm;
5284 BUG_ON(cpu_online(smp_processor_id()));
5286 if (mm != &init_mm) {
5287 switch_mm(mm, &init_mm, current);
5288 finish_arch_post_lock_switch();
5294 * Since this CPU is going 'away' for a while, fold any nr_active delta
5295 * we might have. Assumes we're called after migrate_tasks() so that the
5296 * nr_active count is stable.
5298 * Also see the comment "Global load-average calculations".
5300 static void calc_load_migrate(struct rq *rq)
5302 long delta = calc_load_fold_active(rq);
5304 atomic_long_add(delta, &calc_load_tasks);
5307 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5311 static const struct sched_class fake_sched_class = {
5312 .put_prev_task = put_prev_task_fake,
5315 static struct task_struct fake_task = {
5317 * Avoid pull_{rt,dl}_task()
5319 .prio = MAX_PRIO + 1,
5320 .sched_class = &fake_sched_class,
5324 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5325 * try_to_wake_up()->select_task_rq().
5327 * Called with rq->lock held even though we'er in stop_machine() and
5328 * there's no concurrency possible, we hold the required locks anyway
5329 * because of lock validation efforts.
5331 static void migrate_tasks(struct rq *dead_rq)
5333 struct rq *rq = dead_rq;
5334 struct task_struct *next, *stop = rq->stop;
5338 * Fudge the rq selection such that the below task selection loop
5339 * doesn't get stuck on the currently eligible stop task.
5341 * We're currently inside stop_machine() and the rq is either stuck
5342 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5343 * either way we should never end up calling schedule() until we're
5349 * put_prev_task() and pick_next_task() sched
5350 * class method both need to have an up-to-date
5351 * value of rq->clock[_task]
5353 update_rq_clock(rq);
5357 * There's this thread running, bail when that's the only
5360 if (rq->nr_running == 1)
5364 * pick_next_task assumes pinned rq->lock.
5366 lockdep_pin_lock(&rq->lock);
5367 next = pick_next_task(rq, &fake_task);
5369 next->sched_class->put_prev_task(rq, next);
5372 * Rules for changing task_struct::cpus_allowed are holding
5373 * both pi_lock and rq->lock, such that holding either
5374 * stabilizes the mask.
5376 * Drop rq->lock is not quite as disastrous as it usually is
5377 * because !cpu_active at this point, which means load-balance
5378 * will not interfere. Also, stop-machine.
5380 lockdep_unpin_lock(&rq->lock);
5381 raw_spin_unlock(&rq->lock);
5382 raw_spin_lock(&next->pi_lock);
5383 raw_spin_lock(&rq->lock);
5386 * Since we're inside stop-machine, _nothing_ should have
5387 * changed the task, WARN if weird stuff happened, because in
5388 * that case the above rq->lock drop is a fail too.
5390 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5391 raw_spin_unlock(&next->pi_lock);
5395 /* Find suitable destination for @next, with force if needed. */
5396 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5398 rq = __migrate_task(rq, next, dest_cpu);
5399 if (rq != dead_rq) {
5400 raw_spin_unlock(&rq->lock);
5402 raw_spin_lock(&rq->lock);
5404 raw_spin_unlock(&next->pi_lock);
5409 #endif /* CONFIG_HOTPLUG_CPU */
5411 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5413 static struct ctl_table sd_ctl_dir[] = {
5415 .procname = "sched_domain",
5421 static struct ctl_table sd_ctl_root[] = {
5423 .procname = "kernel",
5425 .child = sd_ctl_dir,
5430 static struct ctl_table *sd_alloc_ctl_entry(int n)
5432 struct ctl_table *entry =
5433 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5438 static void sd_free_ctl_entry(struct ctl_table **tablep)
5440 struct ctl_table *entry;
5443 * In the intermediate directories, both the child directory and
5444 * procname are dynamically allocated and could fail but the mode
5445 * will always be set. In the lowest directory the names are
5446 * static strings and all have proc handlers.
5448 for (entry = *tablep; entry->mode; entry++) {
5450 sd_free_ctl_entry(&entry->child);
5451 if (entry->proc_handler == NULL)
5452 kfree(entry->procname);
5459 static int min_load_idx = 0;
5460 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5463 set_table_entry(struct ctl_table *entry,
5464 const char *procname, void *data, int maxlen,
5465 umode_t mode, proc_handler *proc_handler,
5468 entry->procname = procname;
5470 entry->maxlen = maxlen;
5472 entry->proc_handler = proc_handler;
5475 entry->extra1 = &min_load_idx;
5476 entry->extra2 = &max_load_idx;
5480 static struct ctl_table *
5481 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5483 struct ctl_table *table = sd_alloc_ctl_entry(14);
5488 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5489 sizeof(long), 0644, proc_doulongvec_minmax, false);
5490 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5491 sizeof(long), 0644, proc_doulongvec_minmax, false);
5492 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5493 sizeof(int), 0644, proc_dointvec_minmax, true);
5494 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5495 sizeof(int), 0644, proc_dointvec_minmax, true);
5496 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5497 sizeof(int), 0644, proc_dointvec_minmax, true);
5498 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5499 sizeof(int), 0644, proc_dointvec_minmax, true);
5500 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5501 sizeof(int), 0644, proc_dointvec_minmax, true);
5502 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5503 sizeof(int), 0644, proc_dointvec_minmax, false);
5504 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5505 sizeof(int), 0644, proc_dointvec_minmax, false);
5506 set_table_entry(&table[9], "cache_nice_tries",
5507 &sd->cache_nice_tries,
5508 sizeof(int), 0644, proc_dointvec_minmax, false);
5509 set_table_entry(&table[10], "flags", &sd->flags,
5510 sizeof(int), 0644, proc_dointvec_minmax, false);
5511 set_table_entry(&table[11], "max_newidle_lb_cost",
5512 &sd->max_newidle_lb_cost,
5513 sizeof(long), 0644, proc_doulongvec_minmax, false);
5514 set_table_entry(&table[12], "name", sd->name,
5515 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5516 /* &table[13] is terminator */
5521 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5523 struct ctl_table *entry, *table;
5524 struct sched_domain *sd;
5525 int domain_num = 0, i;
5528 for_each_domain(cpu, sd)
5530 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5535 for_each_domain(cpu, sd) {
5536 snprintf(buf, 32, "domain%d", i);
5537 entry->procname = kstrdup(buf, GFP_KERNEL);
5539 entry->child = sd_alloc_ctl_domain_table(sd);
5546 static struct ctl_table_header *sd_sysctl_header;
5547 static void register_sched_domain_sysctl(void)
5549 int i, cpu_num = num_possible_cpus();
5550 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5553 WARN_ON(sd_ctl_dir[0].child);
5554 sd_ctl_dir[0].child = entry;
5559 for_each_possible_cpu(i) {
5560 snprintf(buf, 32, "cpu%d", i);
5561 entry->procname = kstrdup(buf, GFP_KERNEL);
5563 entry->child = sd_alloc_ctl_cpu_table(i);
5567 WARN_ON(sd_sysctl_header);
5568 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5571 /* may be called multiple times per register */
5572 static void unregister_sched_domain_sysctl(void)
5574 unregister_sysctl_table(sd_sysctl_header);
5575 sd_sysctl_header = NULL;
5576 if (sd_ctl_dir[0].child)
5577 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5580 static void register_sched_domain_sysctl(void)
5583 static void unregister_sched_domain_sysctl(void)
5586 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5588 static void set_rq_online(struct rq *rq)
5591 const struct sched_class *class;
5593 cpumask_set_cpu(rq->cpu, rq->rd->online);
5596 for_each_class(class) {
5597 if (class->rq_online)
5598 class->rq_online(rq);
5603 static void set_rq_offline(struct rq *rq)
5606 const struct sched_class *class;
5608 for_each_class(class) {
5609 if (class->rq_offline)
5610 class->rq_offline(rq);
5613 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5619 * migration_call - callback that gets triggered when a CPU is added.
5620 * Here we can start up the necessary migration thread for the new CPU.
5623 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5625 int cpu = (long)hcpu;
5626 unsigned long flags;
5627 struct rq *rq = cpu_rq(cpu);
5629 switch (action & ~CPU_TASKS_FROZEN) {
5631 case CPU_UP_PREPARE:
5632 rq->calc_load_update = calc_load_update;
5636 /* Update our root-domain */
5637 raw_spin_lock_irqsave(&rq->lock, flags);
5639 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5643 raw_spin_unlock_irqrestore(&rq->lock, flags);
5646 #ifdef CONFIG_HOTPLUG_CPU
5648 sched_ttwu_pending();
5649 /* Update our root-domain */
5650 raw_spin_lock_irqsave(&rq->lock, flags);
5652 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5656 BUG_ON(rq->nr_running != 1); /* the migration thread */
5657 raw_spin_unlock_irqrestore(&rq->lock, flags);
5661 calc_load_migrate(rq);
5666 update_max_interval();
5672 * Register at high priority so that task migration (migrate_all_tasks)
5673 * happens before everything else. This has to be lower priority than
5674 * the notifier in the perf_event subsystem, though.
5676 static struct notifier_block migration_notifier = {
5677 .notifier_call = migration_call,
5678 .priority = CPU_PRI_MIGRATION,
5681 static void set_cpu_rq_start_time(void)
5683 int cpu = smp_processor_id();
5684 struct rq *rq = cpu_rq(cpu);
5685 rq->age_stamp = sched_clock_cpu(cpu);
5688 static int sched_cpu_active(struct notifier_block *nfb,
5689 unsigned long action, void *hcpu)
5691 int cpu = (long)hcpu;
5693 switch (action & ~CPU_TASKS_FROZEN) {
5695 set_cpu_rq_start_time();
5700 * At this point a starting CPU has marked itself as online via
5701 * set_cpu_online(). But it might not yet have marked itself
5702 * as active, which is essential from here on.
5704 set_cpu_active(cpu, true);
5705 stop_machine_unpark(cpu);
5708 case CPU_DOWN_FAILED:
5709 set_cpu_active(cpu, true);
5717 static int sched_cpu_inactive(struct notifier_block *nfb,
5718 unsigned long action, void *hcpu)
5720 switch (action & ~CPU_TASKS_FROZEN) {
5721 case CPU_DOWN_PREPARE:
5722 set_cpu_active((long)hcpu, false);
5729 static int __init migration_init(void)
5731 void *cpu = (void *)(long)smp_processor_id();
5734 /* Initialize migration for the boot CPU */
5735 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5736 BUG_ON(err == NOTIFY_BAD);
5737 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5738 register_cpu_notifier(&migration_notifier);
5740 /* Register cpu active notifiers */
5741 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5742 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5746 early_initcall(migration_init);
5748 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5750 #ifdef CONFIG_SCHED_DEBUG
5752 static __read_mostly int sched_debug_enabled;
5754 static int __init sched_debug_setup(char *str)
5756 sched_debug_enabled = 1;
5760 early_param("sched_debug", sched_debug_setup);
5762 static inline bool sched_debug(void)
5764 return sched_debug_enabled;
5767 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5768 struct cpumask *groupmask)
5770 struct sched_group *group = sd->groups;
5772 cpumask_clear(groupmask);
5774 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5776 if (!(sd->flags & SD_LOAD_BALANCE)) {
5777 printk("does not load-balance\n");
5779 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5784 printk(KERN_CONT "span %*pbl level %s\n",
5785 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5787 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5788 printk(KERN_ERR "ERROR: domain->span does not contain "
5791 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5792 printk(KERN_ERR "ERROR: domain->groups does not contain"
5796 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5800 printk(KERN_ERR "ERROR: group is NULL\n");
5804 if (!cpumask_weight(sched_group_cpus(group))) {
5805 printk(KERN_CONT "\n");
5806 printk(KERN_ERR "ERROR: empty group\n");
5810 if (!(sd->flags & SD_OVERLAP) &&
5811 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5812 printk(KERN_CONT "\n");
5813 printk(KERN_ERR "ERROR: repeated CPUs\n");
5817 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5819 printk(KERN_CONT " %*pbl",
5820 cpumask_pr_args(sched_group_cpus(group)));
5821 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5822 printk(KERN_CONT " (cpu_capacity = %d)",
5823 group->sgc->capacity);
5826 group = group->next;
5827 } while (group != sd->groups);
5828 printk(KERN_CONT "\n");
5830 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5831 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5834 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5835 printk(KERN_ERR "ERROR: parent span is not a superset "
5836 "of domain->span\n");
5840 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5844 if (!sched_debug_enabled)
5848 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5852 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5855 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5863 #else /* !CONFIG_SCHED_DEBUG */
5864 # define sched_domain_debug(sd, cpu) do { } while (0)
5865 static inline bool sched_debug(void)
5869 #endif /* CONFIG_SCHED_DEBUG */
5871 static int sd_degenerate(struct sched_domain *sd)
5873 if (cpumask_weight(sched_domain_span(sd)) == 1)
5876 /* Following flags need at least 2 groups */
5877 if (sd->flags & (SD_LOAD_BALANCE |
5878 SD_BALANCE_NEWIDLE |
5881 SD_SHARE_CPUCAPACITY |
5882 SD_SHARE_PKG_RESOURCES |
5883 SD_SHARE_POWERDOMAIN)) {
5884 if (sd->groups != sd->groups->next)
5888 /* Following flags don't use groups */
5889 if (sd->flags & (SD_WAKE_AFFINE))
5896 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5898 unsigned long cflags = sd->flags, pflags = parent->flags;
5900 if (sd_degenerate(parent))
5903 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5906 /* Flags needing groups don't count if only 1 group in parent */
5907 if (parent->groups == parent->groups->next) {
5908 pflags &= ~(SD_LOAD_BALANCE |
5909 SD_BALANCE_NEWIDLE |
5912 SD_SHARE_CPUCAPACITY |
5913 SD_SHARE_PKG_RESOURCES |
5915 SD_SHARE_POWERDOMAIN);
5916 if (nr_node_ids == 1)
5917 pflags &= ~SD_SERIALIZE;
5919 if (~cflags & pflags)
5925 static void free_rootdomain(struct rcu_head *rcu)
5927 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5929 cpupri_cleanup(&rd->cpupri);
5930 cpudl_cleanup(&rd->cpudl);
5931 free_cpumask_var(rd->dlo_mask);
5932 free_cpumask_var(rd->rto_mask);
5933 free_cpumask_var(rd->online);
5934 free_cpumask_var(rd->span);
5938 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5940 struct root_domain *old_rd = NULL;
5941 unsigned long flags;
5943 raw_spin_lock_irqsave(&rq->lock, flags);
5948 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5951 cpumask_clear_cpu(rq->cpu, old_rd->span);
5954 * If we dont want to free the old_rd yet then
5955 * set old_rd to NULL to skip the freeing later
5958 if (!atomic_dec_and_test(&old_rd->refcount))
5962 atomic_inc(&rd->refcount);
5965 cpumask_set_cpu(rq->cpu, rd->span);
5966 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5969 raw_spin_unlock_irqrestore(&rq->lock, flags);
5972 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5975 static int init_rootdomain(struct root_domain *rd)
5977 memset(rd, 0, sizeof(*rd));
5979 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5981 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5983 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5985 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5988 init_dl_bw(&rd->dl_bw);
5989 if (cpudl_init(&rd->cpudl) != 0)
5992 if (cpupri_init(&rd->cpupri) != 0)
5997 free_cpumask_var(rd->rto_mask);
5999 free_cpumask_var(rd->dlo_mask);
6001 free_cpumask_var(rd->online);
6003 free_cpumask_var(rd->span);
6009 * By default the system creates a single root-domain with all cpus as
6010 * members (mimicking the global state we have today).
6012 struct root_domain def_root_domain;
6014 static void init_defrootdomain(void)
6016 init_rootdomain(&def_root_domain);
6018 atomic_set(&def_root_domain.refcount, 1);
6021 static struct root_domain *alloc_rootdomain(void)
6023 struct root_domain *rd;
6025 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6029 if (init_rootdomain(rd) != 0) {
6037 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6039 struct sched_group *tmp, *first;
6048 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6053 } while (sg != first);
6056 static void free_sched_domain(struct rcu_head *rcu)
6058 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6061 * If its an overlapping domain it has private groups, iterate and
6064 if (sd->flags & SD_OVERLAP) {
6065 free_sched_groups(sd->groups, 1);
6066 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6067 kfree(sd->groups->sgc);
6073 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6075 call_rcu(&sd->rcu, free_sched_domain);
6078 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6080 for (; sd; sd = sd->parent)
6081 destroy_sched_domain(sd, cpu);
6085 * Keep a special pointer to the highest sched_domain that has
6086 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6087 * allows us to avoid some pointer chasing select_idle_sibling().
6089 * Also keep a unique ID per domain (we use the first cpu number in
6090 * the cpumask of the domain), this allows us to quickly tell if
6091 * two cpus are in the same cache domain, see cpus_share_cache().
6093 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6094 DEFINE_PER_CPU(int, sd_llc_size);
6095 DEFINE_PER_CPU(int, sd_llc_id);
6096 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6097 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6098 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6100 static void update_top_cache_domain(int cpu)
6102 struct sched_domain *sd;
6103 struct sched_domain *busy_sd = NULL;
6107 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6109 id = cpumask_first(sched_domain_span(sd));
6110 size = cpumask_weight(sched_domain_span(sd));
6111 busy_sd = sd->parent; /* sd_busy */
6113 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6115 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6116 per_cpu(sd_llc_size, cpu) = size;
6117 per_cpu(sd_llc_id, cpu) = id;
6119 sd = lowest_flag_domain(cpu, SD_NUMA);
6120 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6122 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6123 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6127 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6128 * hold the hotplug lock.
6131 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6133 struct rq *rq = cpu_rq(cpu);
6134 struct sched_domain *tmp;
6136 /* Remove the sched domains which do not contribute to scheduling. */
6137 for (tmp = sd; tmp; ) {
6138 struct sched_domain *parent = tmp->parent;
6142 if (sd_parent_degenerate(tmp, parent)) {
6143 tmp->parent = parent->parent;
6145 parent->parent->child = tmp;
6147 * Transfer SD_PREFER_SIBLING down in case of a
6148 * degenerate parent; the spans match for this
6149 * so the property transfers.
6151 if (parent->flags & SD_PREFER_SIBLING)
6152 tmp->flags |= SD_PREFER_SIBLING;
6153 destroy_sched_domain(parent, cpu);
6158 if (sd && sd_degenerate(sd)) {
6161 destroy_sched_domain(tmp, cpu);
6166 sched_domain_debug(sd, cpu);
6168 rq_attach_root(rq, rd);
6170 rcu_assign_pointer(rq->sd, sd);
6171 destroy_sched_domains(tmp, cpu);
6173 update_top_cache_domain(cpu);
6176 /* Setup the mask of cpus configured for isolated domains */
6177 static int __init isolated_cpu_setup(char *str)
6179 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6180 cpulist_parse(str, cpu_isolated_map);
6184 __setup("isolcpus=", isolated_cpu_setup);
6187 struct sched_domain ** __percpu sd;
6188 struct root_domain *rd;
6199 * Build an iteration mask that can exclude certain CPUs from the upwards
6202 * Asymmetric node setups can result in situations where the domain tree is of
6203 * unequal depth, make sure to skip domains that already cover the entire
6206 * In that case build_sched_domains() will have terminated the iteration early
6207 * and our sibling sd spans will be empty. Domains should always include the
6208 * cpu they're built on, so check that.
6211 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6213 const struct cpumask *span = sched_domain_span(sd);
6214 struct sd_data *sdd = sd->private;
6215 struct sched_domain *sibling;
6218 for_each_cpu(i, span) {
6219 sibling = *per_cpu_ptr(sdd->sd, i);
6220 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6223 cpumask_set_cpu(i, sched_group_mask(sg));
6228 * Return the canonical balance cpu for this group, this is the first cpu
6229 * of this group that's also in the iteration mask.
6231 int group_balance_cpu(struct sched_group *sg)
6233 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6237 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6239 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6240 const struct cpumask *span = sched_domain_span(sd);
6241 struct cpumask *covered = sched_domains_tmpmask;
6242 struct sd_data *sdd = sd->private;
6243 struct sched_domain *sibling;
6246 cpumask_clear(covered);
6248 for_each_cpu(i, span) {
6249 struct cpumask *sg_span;
6251 if (cpumask_test_cpu(i, covered))
6254 sibling = *per_cpu_ptr(sdd->sd, i);
6256 /* See the comment near build_group_mask(). */
6257 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6260 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6261 GFP_KERNEL, cpu_to_node(cpu));
6266 sg_span = sched_group_cpus(sg);
6268 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6270 cpumask_set_cpu(i, sg_span);
6272 cpumask_or(covered, covered, sg_span);
6274 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6275 if (atomic_inc_return(&sg->sgc->ref) == 1)
6276 build_group_mask(sd, sg);
6279 * Initialize sgc->capacity such that even if we mess up the
6280 * domains and no possible iteration will get us here, we won't
6283 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6286 * Make sure the first group of this domain contains the
6287 * canonical balance cpu. Otherwise the sched_domain iteration
6288 * breaks. See update_sg_lb_stats().
6290 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6291 group_balance_cpu(sg) == cpu)
6301 sd->groups = groups;
6306 free_sched_groups(first, 0);
6311 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6313 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6314 struct sched_domain *child = sd->child;
6317 cpu = cpumask_first(sched_domain_span(child));
6320 *sg = *per_cpu_ptr(sdd->sg, cpu);
6321 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6322 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6329 * build_sched_groups will build a circular linked list of the groups
6330 * covered by the given span, and will set each group's ->cpumask correctly,
6331 * and ->cpu_capacity to 0.
6333 * Assumes the sched_domain tree is fully constructed
6336 build_sched_groups(struct sched_domain *sd, int cpu)
6338 struct sched_group *first = NULL, *last = NULL;
6339 struct sd_data *sdd = sd->private;
6340 const struct cpumask *span = sched_domain_span(sd);
6341 struct cpumask *covered;
6344 get_group(cpu, sdd, &sd->groups);
6345 atomic_inc(&sd->groups->ref);
6347 if (cpu != cpumask_first(span))
6350 lockdep_assert_held(&sched_domains_mutex);
6351 covered = sched_domains_tmpmask;
6353 cpumask_clear(covered);
6355 for_each_cpu(i, span) {
6356 struct sched_group *sg;
6359 if (cpumask_test_cpu(i, covered))
6362 group = get_group(i, sdd, &sg);
6363 cpumask_setall(sched_group_mask(sg));
6365 for_each_cpu(j, span) {
6366 if (get_group(j, sdd, NULL) != group)
6369 cpumask_set_cpu(j, covered);
6370 cpumask_set_cpu(j, sched_group_cpus(sg));
6385 * Initialize sched groups cpu_capacity.
6387 * cpu_capacity indicates the capacity of sched group, which is used while
6388 * distributing the load between different sched groups in a sched domain.
6389 * Typically cpu_capacity for all the groups in a sched domain will be same
6390 * unless there are asymmetries in the topology. If there are asymmetries,
6391 * group having more cpu_capacity will pickup more load compared to the
6392 * group having less cpu_capacity.
6394 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6396 struct sched_group *sg = sd->groups;
6401 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6403 } while (sg != sd->groups);
6405 if (cpu != group_balance_cpu(sg))
6408 update_group_capacity(sd, cpu);
6409 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6413 * Initializers for schedule domains
6414 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6417 static int default_relax_domain_level = -1;
6418 int sched_domain_level_max;
6420 static int __init setup_relax_domain_level(char *str)
6422 if (kstrtoint(str, 0, &default_relax_domain_level))
6423 pr_warn("Unable to set relax_domain_level\n");
6427 __setup("relax_domain_level=", setup_relax_domain_level);
6429 static void set_domain_attribute(struct sched_domain *sd,
6430 struct sched_domain_attr *attr)
6434 if (!attr || attr->relax_domain_level < 0) {
6435 if (default_relax_domain_level < 0)
6438 request = default_relax_domain_level;
6440 request = attr->relax_domain_level;
6441 if (request < sd->level) {
6442 /* turn off idle balance on this domain */
6443 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6445 /* turn on idle balance on this domain */
6446 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6450 static void __sdt_free(const struct cpumask *cpu_map);
6451 static int __sdt_alloc(const struct cpumask *cpu_map);
6453 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6454 const struct cpumask *cpu_map)
6458 if (!atomic_read(&d->rd->refcount))
6459 free_rootdomain(&d->rd->rcu); /* fall through */
6461 free_percpu(d->sd); /* fall through */
6463 __sdt_free(cpu_map); /* fall through */
6469 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6470 const struct cpumask *cpu_map)
6472 memset(d, 0, sizeof(*d));
6474 if (__sdt_alloc(cpu_map))
6475 return sa_sd_storage;
6476 d->sd = alloc_percpu(struct sched_domain *);
6478 return sa_sd_storage;
6479 d->rd = alloc_rootdomain();
6482 return sa_rootdomain;
6486 * NULL the sd_data elements we've used to build the sched_domain and
6487 * sched_group structure so that the subsequent __free_domain_allocs()
6488 * will not free the data we're using.
6490 static void claim_allocations(int cpu, struct sched_domain *sd)
6492 struct sd_data *sdd = sd->private;
6494 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6495 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6497 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6498 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6500 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6501 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6505 static int sched_domains_numa_levels;
6506 enum numa_topology_type sched_numa_topology_type;
6507 static int *sched_domains_numa_distance;
6508 int sched_max_numa_distance;
6509 static struct cpumask ***sched_domains_numa_masks;
6510 static int sched_domains_curr_level;
6514 * SD_flags allowed in topology descriptions.
6516 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6517 * SD_SHARE_PKG_RESOURCES - describes shared caches
6518 * SD_NUMA - describes NUMA topologies
6519 * SD_SHARE_POWERDOMAIN - describes shared power domain
6522 * SD_ASYM_PACKING - describes SMT quirks
6524 #define TOPOLOGY_SD_FLAGS \
6525 (SD_SHARE_CPUCAPACITY | \
6526 SD_SHARE_PKG_RESOURCES | \
6529 SD_SHARE_POWERDOMAIN)
6531 static struct sched_domain *
6532 sd_init(struct sched_domain_topology_level *tl, int cpu)
6534 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6535 int sd_weight, sd_flags = 0;
6539 * Ugly hack to pass state to sd_numa_mask()...
6541 sched_domains_curr_level = tl->numa_level;
6544 sd_weight = cpumask_weight(tl->mask(cpu));
6547 sd_flags = (*tl->sd_flags)();
6548 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6549 "wrong sd_flags in topology description\n"))
6550 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6552 *sd = (struct sched_domain){
6553 .min_interval = sd_weight,
6554 .max_interval = 2*sd_weight,
6556 .imbalance_pct = 125,
6558 .cache_nice_tries = 0,
6565 .flags = 1*SD_LOAD_BALANCE
6566 | 1*SD_BALANCE_NEWIDLE
6571 | 0*SD_SHARE_CPUCAPACITY
6572 | 0*SD_SHARE_PKG_RESOURCES
6574 | 0*SD_PREFER_SIBLING
6579 .last_balance = jiffies,
6580 .balance_interval = sd_weight,
6582 .max_newidle_lb_cost = 0,
6583 .next_decay_max_lb_cost = jiffies,
6584 #ifdef CONFIG_SCHED_DEBUG
6590 * Convert topological properties into behaviour.
6593 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6594 sd->flags |= SD_PREFER_SIBLING;
6595 sd->imbalance_pct = 110;
6596 sd->smt_gain = 1178; /* ~15% */
6598 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6599 sd->imbalance_pct = 117;
6600 sd->cache_nice_tries = 1;
6604 } else if (sd->flags & SD_NUMA) {
6605 sd->cache_nice_tries = 2;
6609 sd->flags |= SD_SERIALIZE;
6610 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6611 sd->flags &= ~(SD_BALANCE_EXEC |
6618 sd->flags |= SD_PREFER_SIBLING;
6619 sd->cache_nice_tries = 1;
6624 sd->private = &tl->data;
6630 * Topology list, bottom-up.
6632 static struct sched_domain_topology_level default_topology[] = {
6633 #ifdef CONFIG_SCHED_SMT
6634 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6636 #ifdef CONFIG_SCHED_MC
6637 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6639 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6643 static struct sched_domain_topology_level *sched_domain_topology =
6646 #define for_each_sd_topology(tl) \
6647 for (tl = sched_domain_topology; tl->mask; tl++)
6649 void set_sched_topology(struct sched_domain_topology_level *tl)
6651 sched_domain_topology = tl;
6656 static const struct cpumask *sd_numa_mask(int cpu)
6658 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6661 static void sched_numa_warn(const char *str)
6663 static int done = false;
6671 printk(KERN_WARNING "ERROR: %s\n\n", str);
6673 for (i = 0; i < nr_node_ids; i++) {
6674 printk(KERN_WARNING " ");
6675 for (j = 0; j < nr_node_ids; j++)
6676 printk(KERN_CONT "%02d ", node_distance(i,j));
6677 printk(KERN_CONT "\n");
6679 printk(KERN_WARNING "\n");
6682 bool find_numa_distance(int distance)
6686 if (distance == node_distance(0, 0))
6689 for (i = 0; i < sched_domains_numa_levels; i++) {
6690 if (sched_domains_numa_distance[i] == distance)
6698 * A system can have three types of NUMA topology:
6699 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6700 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6701 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6703 * The difference between a glueless mesh topology and a backplane
6704 * topology lies in whether communication between not directly
6705 * connected nodes goes through intermediary nodes (where programs
6706 * could run), or through backplane controllers. This affects
6707 * placement of programs.
6709 * The type of topology can be discerned with the following tests:
6710 * - If the maximum distance between any nodes is 1 hop, the system
6711 * is directly connected.
6712 * - If for two nodes A and B, located N > 1 hops away from each other,
6713 * there is an intermediary node C, which is < N hops away from both
6714 * nodes A and B, the system is a glueless mesh.
6716 static void init_numa_topology_type(void)
6720 n = sched_max_numa_distance;
6722 if (sched_domains_numa_levels <= 1) {
6723 sched_numa_topology_type = NUMA_DIRECT;
6727 for_each_online_node(a) {
6728 for_each_online_node(b) {
6729 /* Find two nodes furthest removed from each other. */
6730 if (node_distance(a, b) < n)
6733 /* Is there an intermediary node between a and b? */
6734 for_each_online_node(c) {
6735 if (node_distance(a, c) < n &&
6736 node_distance(b, c) < n) {
6737 sched_numa_topology_type =
6743 sched_numa_topology_type = NUMA_BACKPLANE;
6749 static void sched_init_numa(void)
6751 int next_distance, curr_distance = node_distance(0, 0);
6752 struct sched_domain_topology_level *tl;
6756 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6757 if (!sched_domains_numa_distance)
6761 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6762 * unique distances in the node_distance() table.
6764 * Assumes node_distance(0,j) includes all distances in
6765 * node_distance(i,j) in order to avoid cubic time.
6767 next_distance = curr_distance;
6768 for (i = 0; i < nr_node_ids; i++) {
6769 for (j = 0; j < nr_node_ids; j++) {
6770 for (k = 0; k < nr_node_ids; k++) {
6771 int distance = node_distance(i, k);
6773 if (distance > curr_distance &&
6774 (distance < next_distance ||
6775 next_distance == curr_distance))
6776 next_distance = distance;
6779 * While not a strong assumption it would be nice to know
6780 * about cases where if node A is connected to B, B is not
6781 * equally connected to A.
6783 if (sched_debug() && node_distance(k, i) != distance)
6784 sched_numa_warn("Node-distance not symmetric");
6786 if (sched_debug() && i && !find_numa_distance(distance))
6787 sched_numa_warn("Node-0 not representative");
6789 if (next_distance != curr_distance) {
6790 sched_domains_numa_distance[level++] = next_distance;
6791 sched_domains_numa_levels = level;
6792 curr_distance = next_distance;
6797 * In case of sched_debug() we verify the above assumption.
6807 * 'level' contains the number of unique distances, excluding the
6808 * identity distance node_distance(i,i).
6810 * The sched_domains_numa_distance[] array includes the actual distance
6815 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6816 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6817 * the array will contain less then 'level' members. This could be
6818 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6819 * in other functions.
6821 * We reset it to 'level' at the end of this function.
6823 sched_domains_numa_levels = 0;
6825 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6826 if (!sched_domains_numa_masks)
6830 * Now for each level, construct a mask per node which contains all
6831 * cpus of nodes that are that many hops away from us.
6833 for (i = 0; i < level; i++) {
6834 sched_domains_numa_masks[i] =
6835 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6836 if (!sched_domains_numa_masks[i])
6839 for (j = 0; j < nr_node_ids; j++) {
6840 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6844 sched_domains_numa_masks[i][j] = mask;
6847 if (node_distance(j, k) > sched_domains_numa_distance[i])
6850 cpumask_or(mask, mask, cpumask_of_node(k));
6855 /* Compute default topology size */
6856 for (i = 0; sched_domain_topology[i].mask; i++);
6858 tl = kzalloc((i + level + 1) *
6859 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6864 * Copy the default topology bits..
6866 for (i = 0; sched_domain_topology[i].mask; i++)
6867 tl[i] = sched_domain_topology[i];
6870 * .. and append 'j' levels of NUMA goodness.
6872 for (j = 0; j < level; i++, j++) {
6873 tl[i] = (struct sched_domain_topology_level){
6874 .mask = sd_numa_mask,
6875 .sd_flags = cpu_numa_flags,
6876 .flags = SDTL_OVERLAP,
6882 sched_domain_topology = tl;
6884 sched_domains_numa_levels = level;
6885 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6887 init_numa_topology_type();
6890 static void sched_domains_numa_masks_set(int cpu)
6893 int node = cpu_to_node(cpu);
6895 for (i = 0; i < sched_domains_numa_levels; i++) {
6896 for (j = 0; j < nr_node_ids; j++) {
6897 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6898 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6903 static void sched_domains_numa_masks_clear(int cpu)
6906 for (i = 0; i < sched_domains_numa_levels; i++) {
6907 for (j = 0; j < nr_node_ids; j++)
6908 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6913 * Update sched_domains_numa_masks[level][node] array when new cpus
6916 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6917 unsigned long action,
6920 int cpu = (long)hcpu;
6922 switch (action & ~CPU_TASKS_FROZEN) {
6924 sched_domains_numa_masks_set(cpu);
6928 sched_domains_numa_masks_clear(cpu);
6938 static inline void sched_init_numa(void)
6942 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6943 unsigned long action,
6948 #endif /* CONFIG_NUMA */
6950 static int __sdt_alloc(const struct cpumask *cpu_map)
6952 struct sched_domain_topology_level *tl;
6955 for_each_sd_topology(tl) {
6956 struct sd_data *sdd = &tl->data;
6958 sdd->sd = alloc_percpu(struct sched_domain *);
6962 sdd->sg = alloc_percpu(struct sched_group *);
6966 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6970 for_each_cpu(j, cpu_map) {
6971 struct sched_domain *sd;
6972 struct sched_group *sg;
6973 struct sched_group_capacity *sgc;
6975 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6976 GFP_KERNEL, cpu_to_node(j));
6980 *per_cpu_ptr(sdd->sd, j) = sd;
6982 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6983 GFP_KERNEL, cpu_to_node(j));
6989 *per_cpu_ptr(sdd->sg, j) = sg;
6991 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6992 GFP_KERNEL, cpu_to_node(j));
6996 *per_cpu_ptr(sdd->sgc, j) = sgc;
7003 static void __sdt_free(const struct cpumask *cpu_map)
7005 struct sched_domain_topology_level *tl;
7008 for_each_sd_topology(tl) {
7009 struct sd_data *sdd = &tl->data;
7011 for_each_cpu(j, cpu_map) {
7012 struct sched_domain *sd;
7015 sd = *per_cpu_ptr(sdd->sd, j);
7016 if (sd && (sd->flags & SD_OVERLAP))
7017 free_sched_groups(sd->groups, 0);
7018 kfree(*per_cpu_ptr(sdd->sd, j));
7022 kfree(*per_cpu_ptr(sdd->sg, j));
7024 kfree(*per_cpu_ptr(sdd->sgc, j));
7026 free_percpu(sdd->sd);
7028 free_percpu(sdd->sg);
7030 free_percpu(sdd->sgc);
7035 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7036 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7037 struct sched_domain *child, int cpu)
7039 struct sched_domain *sd = sd_init(tl, cpu);
7043 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7045 sd->level = child->level + 1;
7046 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7050 if (!cpumask_subset(sched_domain_span(child),
7051 sched_domain_span(sd))) {
7052 pr_err("BUG: arch topology borken\n");
7053 #ifdef CONFIG_SCHED_DEBUG
7054 pr_err(" the %s domain not a subset of the %s domain\n",
7055 child->name, sd->name);
7057 /* Fixup, ensure @sd has at least @child cpus. */
7058 cpumask_or(sched_domain_span(sd),
7059 sched_domain_span(sd),
7060 sched_domain_span(child));
7064 set_domain_attribute(sd, attr);
7070 * Build sched domains for a given set of cpus and attach the sched domains
7071 * to the individual cpus
7073 static int build_sched_domains(const struct cpumask *cpu_map,
7074 struct sched_domain_attr *attr)
7076 enum s_alloc alloc_state;
7077 struct sched_domain *sd;
7079 int i, ret = -ENOMEM;
7081 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7082 if (alloc_state != sa_rootdomain)
7085 /* Set up domains for cpus specified by the cpu_map. */
7086 for_each_cpu(i, cpu_map) {
7087 struct sched_domain_topology_level *tl;
7090 for_each_sd_topology(tl) {
7091 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7092 if (tl == sched_domain_topology)
7093 *per_cpu_ptr(d.sd, i) = sd;
7094 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7095 sd->flags |= SD_OVERLAP;
7096 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7101 /* Build the groups for the domains */
7102 for_each_cpu(i, cpu_map) {
7103 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7104 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7105 if (sd->flags & SD_OVERLAP) {
7106 if (build_overlap_sched_groups(sd, i))
7109 if (build_sched_groups(sd, i))
7115 /* Calculate CPU capacity for physical packages and nodes */
7116 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7117 if (!cpumask_test_cpu(i, cpu_map))
7120 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7121 claim_allocations(i, sd);
7122 init_sched_groups_capacity(i, sd);
7126 /* Attach the domains */
7128 for_each_cpu(i, cpu_map) {
7129 sd = *per_cpu_ptr(d.sd, i);
7130 cpu_attach_domain(sd, d.rd, i);
7136 __free_domain_allocs(&d, alloc_state, cpu_map);
7140 static cpumask_var_t *doms_cur; /* current sched domains */
7141 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7142 static struct sched_domain_attr *dattr_cur;
7143 /* attribues of custom domains in 'doms_cur' */
7146 * Special case: If a kmalloc of a doms_cur partition (array of
7147 * cpumask) fails, then fallback to a single sched domain,
7148 * as determined by the single cpumask fallback_doms.
7150 static cpumask_var_t fallback_doms;
7153 * arch_update_cpu_topology lets virtualized architectures update the
7154 * cpu core maps. It is supposed to return 1 if the topology changed
7155 * or 0 if it stayed the same.
7157 int __weak arch_update_cpu_topology(void)
7162 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7165 cpumask_var_t *doms;
7167 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7170 for (i = 0; i < ndoms; i++) {
7171 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7172 free_sched_domains(doms, i);
7179 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7182 for (i = 0; i < ndoms; i++)
7183 free_cpumask_var(doms[i]);
7188 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7189 * For now this just excludes isolated cpus, but could be used to
7190 * exclude other special cases in the future.
7192 static int init_sched_domains(const struct cpumask *cpu_map)
7196 arch_update_cpu_topology();
7198 doms_cur = alloc_sched_domains(ndoms_cur);
7200 doms_cur = &fallback_doms;
7201 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7202 err = build_sched_domains(doms_cur[0], NULL);
7203 register_sched_domain_sysctl();
7209 * Detach sched domains from a group of cpus specified in cpu_map
7210 * These cpus will now be attached to the NULL domain
7212 static void detach_destroy_domains(const struct cpumask *cpu_map)
7217 for_each_cpu(i, cpu_map)
7218 cpu_attach_domain(NULL, &def_root_domain, i);
7222 /* handle null as "default" */
7223 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7224 struct sched_domain_attr *new, int idx_new)
7226 struct sched_domain_attr tmp;
7233 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7234 new ? (new + idx_new) : &tmp,
7235 sizeof(struct sched_domain_attr));
7239 * Partition sched domains as specified by the 'ndoms_new'
7240 * cpumasks in the array doms_new[] of cpumasks. This compares
7241 * doms_new[] to the current sched domain partitioning, doms_cur[].
7242 * It destroys each deleted domain and builds each new domain.
7244 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7245 * The masks don't intersect (don't overlap.) We should setup one
7246 * sched domain for each mask. CPUs not in any of the cpumasks will
7247 * not be load balanced. If the same cpumask appears both in the
7248 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7251 * The passed in 'doms_new' should be allocated using
7252 * alloc_sched_domains. This routine takes ownership of it and will
7253 * free_sched_domains it when done with it. If the caller failed the
7254 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7255 * and partition_sched_domains() will fallback to the single partition
7256 * 'fallback_doms', it also forces the domains to be rebuilt.
7258 * If doms_new == NULL it will be replaced with cpu_online_mask.
7259 * ndoms_new == 0 is a special case for destroying existing domains,
7260 * and it will not create the default domain.
7262 * Call with hotplug lock held
7264 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7265 struct sched_domain_attr *dattr_new)
7270 mutex_lock(&sched_domains_mutex);
7272 /* always unregister in case we don't destroy any domains */
7273 unregister_sched_domain_sysctl();
7275 /* Let architecture update cpu core mappings. */
7276 new_topology = arch_update_cpu_topology();
7278 n = doms_new ? ndoms_new : 0;
7280 /* Destroy deleted domains */
7281 for (i = 0; i < ndoms_cur; i++) {
7282 for (j = 0; j < n && !new_topology; j++) {
7283 if (cpumask_equal(doms_cur[i], doms_new[j])
7284 && dattrs_equal(dattr_cur, i, dattr_new, j))
7287 /* no match - a current sched domain not in new doms_new[] */
7288 detach_destroy_domains(doms_cur[i]);
7294 if (doms_new == NULL) {
7296 doms_new = &fallback_doms;
7297 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7298 WARN_ON_ONCE(dattr_new);
7301 /* Build new domains */
7302 for (i = 0; i < ndoms_new; i++) {
7303 for (j = 0; j < n && !new_topology; j++) {
7304 if (cpumask_equal(doms_new[i], doms_cur[j])
7305 && dattrs_equal(dattr_new, i, dattr_cur, j))
7308 /* no match - add a new doms_new */
7309 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7314 /* Remember the new sched domains */
7315 if (doms_cur != &fallback_doms)
7316 free_sched_domains(doms_cur, ndoms_cur);
7317 kfree(dattr_cur); /* kfree(NULL) is safe */
7318 doms_cur = doms_new;
7319 dattr_cur = dattr_new;
7320 ndoms_cur = ndoms_new;
7322 register_sched_domain_sysctl();
7324 mutex_unlock(&sched_domains_mutex);
7327 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7330 * Update cpusets according to cpu_active mask. If cpusets are
7331 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7332 * around partition_sched_domains().
7334 * If we come here as part of a suspend/resume, don't touch cpusets because we
7335 * want to restore it back to its original state upon resume anyway.
7337 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7341 case CPU_ONLINE_FROZEN:
7342 case CPU_DOWN_FAILED_FROZEN:
7345 * num_cpus_frozen tracks how many CPUs are involved in suspend
7346 * resume sequence. As long as this is not the last online
7347 * operation in the resume sequence, just build a single sched
7348 * domain, ignoring cpusets.
7351 if (likely(num_cpus_frozen)) {
7352 partition_sched_domains(1, NULL, NULL);
7357 * This is the last CPU online operation. So fall through and
7358 * restore the original sched domains by considering the
7359 * cpuset configurations.
7363 cpuset_update_active_cpus(true);
7371 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7374 unsigned long flags;
7375 long cpu = (long)hcpu;
7381 case CPU_DOWN_PREPARE:
7382 rcu_read_lock_sched();
7383 dl_b = dl_bw_of(cpu);
7385 raw_spin_lock_irqsave(&dl_b->lock, flags);
7386 cpus = dl_bw_cpus(cpu);
7387 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7388 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7390 rcu_read_unlock_sched();
7393 return notifier_from_errno(-EBUSY);
7394 cpuset_update_active_cpus(false);
7396 case CPU_DOWN_PREPARE_FROZEN:
7398 partition_sched_domains(1, NULL, NULL);
7406 void __init sched_init_smp(void)
7408 cpumask_var_t non_isolated_cpus;
7410 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7411 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7416 * There's no userspace yet to cause hotplug operations; hence all the
7417 * cpu masks are stable and all blatant races in the below code cannot
7420 mutex_lock(&sched_domains_mutex);
7421 init_sched_domains(cpu_active_mask);
7422 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7423 if (cpumask_empty(non_isolated_cpus))
7424 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7425 mutex_unlock(&sched_domains_mutex);
7427 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7428 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7429 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7433 /* Move init over to a non-isolated CPU */
7434 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7436 sched_init_granularity();
7437 free_cpumask_var(non_isolated_cpus);
7439 init_sched_rt_class();
7440 init_sched_dl_class();
7443 void __init sched_init_smp(void)
7445 sched_init_granularity();
7447 #endif /* CONFIG_SMP */
7449 int in_sched_functions(unsigned long addr)
7451 return in_lock_functions(addr) ||
7452 (addr >= (unsigned long)__sched_text_start
7453 && addr < (unsigned long)__sched_text_end);
7456 #ifdef CONFIG_CGROUP_SCHED
7458 * Default task group.
7459 * Every task in system belongs to this group at bootup.
7461 struct task_group root_task_group;
7462 LIST_HEAD(task_groups);
7464 /* Cacheline aligned slab cache for task_group */
7465 static struct kmem_cache *task_group_cache __read_mostly;
7468 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7470 void __init sched_init(void)
7473 unsigned long alloc_size = 0, ptr;
7475 #ifdef CONFIG_FAIR_GROUP_SCHED
7476 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7478 #ifdef CONFIG_RT_GROUP_SCHED
7479 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7482 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7484 #ifdef CONFIG_FAIR_GROUP_SCHED
7485 root_task_group.se = (struct sched_entity **)ptr;
7486 ptr += nr_cpu_ids * sizeof(void **);
7488 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7489 ptr += nr_cpu_ids * sizeof(void **);
7491 #endif /* CONFIG_FAIR_GROUP_SCHED */
7492 #ifdef CONFIG_RT_GROUP_SCHED
7493 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7494 ptr += nr_cpu_ids * sizeof(void **);
7496 root_task_group.rt_rq = (struct rt_rq **)ptr;
7497 ptr += nr_cpu_ids * sizeof(void **);
7499 #endif /* CONFIG_RT_GROUP_SCHED */
7501 #ifdef CONFIG_CPUMASK_OFFSTACK
7502 for_each_possible_cpu(i) {
7503 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7504 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7506 #endif /* CONFIG_CPUMASK_OFFSTACK */
7508 init_rt_bandwidth(&def_rt_bandwidth,
7509 global_rt_period(), global_rt_runtime());
7510 init_dl_bandwidth(&def_dl_bandwidth,
7511 global_rt_period(), global_rt_runtime());
7514 init_defrootdomain();
7517 #ifdef CONFIG_RT_GROUP_SCHED
7518 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7519 global_rt_period(), global_rt_runtime());
7520 #endif /* CONFIG_RT_GROUP_SCHED */
7522 #ifdef CONFIG_CGROUP_SCHED
7523 task_group_cache = KMEM_CACHE(task_group, 0);
7525 list_add(&root_task_group.list, &task_groups);
7526 INIT_LIST_HEAD(&root_task_group.children);
7527 INIT_LIST_HEAD(&root_task_group.siblings);
7528 autogroup_init(&init_task);
7529 #endif /* CONFIG_CGROUP_SCHED */
7531 for_each_possible_cpu(i) {
7535 raw_spin_lock_init(&rq->lock);
7537 rq->calc_load_active = 0;
7538 rq->calc_load_update = jiffies + LOAD_FREQ;
7539 init_cfs_rq(&rq->cfs);
7540 init_rt_rq(&rq->rt);
7541 init_dl_rq(&rq->dl);
7542 #ifdef CONFIG_FAIR_GROUP_SCHED
7543 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7544 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7546 * How much cpu bandwidth does root_task_group get?
7548 * In case of task-groups formed thr' the cgroup filesystem, it
7549 * gets 100% of the cpu resources in the system. This overall
7550 * system cpu resource is divided among the tasks of
7551 * root_task_group and its child task-groups in a fair manner,
7552 * based on each entity's (task or task-group's) weight
7553 * (se->load.weight).
7555 * In other words, if root_task_group has 10 tasks of weight
7556 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7557 * then A0's share of the cpu resource is:
7559 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7561 * We achieve this by letting root_task_group's tasks sit
7562 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7564 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7565 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7566 #endif /* CONFIG_FAIR_GROUP_SCHED */
7568 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7569 #ifdef CONFIG_RT_GROUP_SCHED
7570 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7573 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7574 rq->cpu_load[j] = 0;
7576 rq->last_load_update_tick = jiffies;
7581 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7582 rq->balance_callback = NULL;
7583 rq->active_balance = 0;
7584 rq->next_balance = jiffies;
7589 rq->avg_idle = 2*sysctl_sched_migration_cost;
7590 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7592 INIT_LIST_HEAD(&rq->cfs_tasks);
7594 rq_attach_root(rq, &def_root_domain);
7595 #ifdef CONFIG_NO_HZ_COMMON
7598 #ifdef CONFIG_NO_HZ_FULL
7599 rq->last_sched_tick = 0;
7603 atomic_set(&rq->nr_iowait, 0);
7606 set_load_weight(&init_task);
7608 #ifdef CONFIG_PREEMPT_NOTIFIERS
7609 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7613 * The boot idle thread does lazy MMU switching as well:
7615 atomic_inc(&init_mm.mm_count);
7616 enter_lazy_tlb(&init_mm, current);
7619 * During early bootup we pretend to be a normal task:
7621 current->sched_class = &fair_sched_class;
7624 * Make us the idle thread. Technically, schedule() should not be
7625 * called from this thread, however somewhere below it might be,
7626 * but because we are the idle thread, we just pick up running again
7627 * when this runqueue becomes "idle".
7629 init_idle(current, smp_processor_id());
7631 calc_load_update = jiffies + LOAD_FREQ;
7634 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7635 /* May be allocated at isolcpus cmdline parse time */
7636 if (cpu_isolated_map == NULL)
7637 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7638 idle_thread_set_boot_cpu();
7639 set_cpu_rq_start_time();
7641 init_sched_fair_class();
7643 scheduler_running = 1;
7646 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7647 static inline int preempt_count_equals(int preempt_offset)
7649 int nested = preempt_count() + rcu_preempt_depth();
7651 return (nested == preempt_offset);
7654 void __might_sleep(const char *file, int line, int preempt_offset)
7657 * Blocking primitives will set (and therefore destroy) current->state,
7658 * since we will exit with TASK_RUNNING make sure we enter with it,
7659 * otherwise we will destroy state.
7661 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7662 "do not call blocking ops when !TASK_RUNNING; "
7663 "state=%lx set at [<%p>] %pS\n",
7665 (void *)current->task_state_change,
7666 (void *)current->task_state_change);
7668 ___might_sleep(file, line, preempt_offset);
7670 EXPORT_SYMBOL(__might_sleep);
7672 void ___might_sleep(const char *file, int line, int preempt_offset)
7674 static unsigned long prev_jiffy; /* ratelimiting */
7676 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7677 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7678 !is_idle_task(current)) ||
7679 system_state != SYSTEM_RUNNING || oops_in_progress)
7681 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7683 prev_jiffy = jiffies;
7686 "BUG: sleeping function called from invalid context at %s:%d\n",
7689 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7690 in_atomic(), irqs_disabled(),
7691 current->pid, current->comm);
7693 if (task_stack_end_corrupted(current))
7694 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7696 debug_show_held_locks(current);
7697 if (irqs_disabled())
7698 print_irqtrace_events(current);
7699 #ifdef CONFIG_DEBUG_PREEMPT
7700 if (!preempt_count_equals(preempt_offset)) {
7701 pr_err("Preemption disabled at:");
7702 print_ip_sym(current->preempt_disable_ip);
7708 EXPORT_SYMBOL(___might_sleep);
7711 #ifdef CONFIG_MAGIC_SYSRQ
7712 void normalize_rt_tasks(void)
7714 struct task_struct *g, *p;
7715 struct sched_attr attr = {
7716 .sched_policy = SCHED_NORMAL,
7719 read_lock(&tasklist_lock);
7720 for_each_process_thread(g, p) {
7722 * Only normalize user tasks:
7724 if (p->flags & PF_KTHREAD)
7727 p->se.exec_start = 0;
7728 #ifdef CONFIG_SCHEDSTATS
7729 p->se.statistics.wait_start = 0;
7730 p->se.statistics.sleep_start = 0;
7731 p->se.statistics.block_start = 0;
7734 if (!dl_task(p) && !rt_task(p)) {
7736 * Renice negative nice level userspace
7739 if (task_nice(p) < 0)
7740 set_user_nice(p, 0);
7744 __sched_setscheduler(p, &attr, false, false);
7746 read_unlock(&tasklist_lock);
7749 #endif /* CONFIG_MAGIC_SYSRQ */
7751 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7753 * These functions are only useful for the IA64 MCA handling, or kdb.
7755 * They can only be called when the whole system has been
7756 * stopped - every CPU needs to be quiescent, and no scheduling
7757 * activity can take place. Using them for anything else would
7758 * be a serious bug, and as a result, they aren't even visible
7759 * under any other configuration.
7763 * curr_task - return the current task for a given cpu.
7764 * @cpu: the processor in question.
7766 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7768 * Return: The current task for @cpu.
7770 struct task_struct *curr_task(int cpu)
7772 return cpu_curr(cpu);
7775 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7779 * set_curr_task - set the current task for a given cpu.
7780 * @cpu: the processor in question.
7781 * @p: the task pointer to set.
7783 * Description: This function must only be used when non-maskable interrupts
7784 * are serviced on a separate stack. It allows the architecture to switch the
7785 * notion of the current task on a cpu in a non-blocking manner. This function
7786 * must be called with all CPU's synchronized, and interrupts disabled, the
7787 * and caller must save the original value of the current task (see
7788 * curr_task() above) and restore that value before reenabling interrupts and
7789 * re-starting the system.
7791 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7793 void set_curr_task(int cpu, struct task_struct *p)
7800 #ifdef CONFIG_CGROUP_SCHED
7801 /* task_group_lock serializes the addition/removal of task groups */
7802 static DEFINE_SPINLOCK(task_group_lock);
7804 static void free_sched_group(struct task_group *tg)
7806 free_fair_sched_group(tg);
7807 free_rt_sched_group(tg);
7809 kmem_cache_free(task_group_cache, tg);
7812 /* allocate runqueue etc for a new task group */
7813 struct task_group *sched_create_group(struct task_group *parent)
7815 struct task_group *tg;
7817 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7819 return ERR_PTR(-ENOMEM);
7821 if (!alloc_fair_sched_group(tg, parent))
7824 if (!alloc_rt_sched_group(tg, parent))
7830 free_sched_group(tg);
7831 return ERR_PTR(-ENOMEM);
7834 void sched_online_group(struct task_group *tg, struct task_group *parent)
7836 unsigned long flags;
7838 spin_lock_irqsave(&task_group_lock, flags);
7839 list_add_rcu(&tg->list, &task_groups);
7841 WARN_ON(!parent); /* root should already exist */
7843 tg->parent = parent;
7844 INIT_LIST_HEAD(&tg->children);
7845 list_add_rcu(&tg->siblings, &parent->children);
7846 spin_unlock_irqrestore(&task_group_lock, flags);
7849 /* rcu callback to free various structures associated with a task group */
7850 static void free_sched_group_rcu(struct rcu_head *rhp)
7852 /* now it should be safe to free those cfs_rqs */
7853 free_sched_group(container_of(rhp, struct task_group, rcu));
7856 /* Destroy runqueue etc associated with a task group */
7857 void sched_destroy_group(struct task_group *tg)
7859 /* wait for possible concurrent references to cfs_rqs complete */
7860 call_rcu(&tg->rcu, free_sched_group_rcu);
7863 void sched_offline_group(struct task_group *tg)
7865 unsigned long flags;
7868 /* end participation in shares distribution */
7869 for_each_possible_cpu(i)
7870 unregister_fair_sched_group(tg, i);
7872 spin_lock_irqsave(&task_group_lock, flags);
7873 list_del_rcu(&tg->list);
7874 list_del_rcu(&tg->siblings);
7875 spin_unlock_irqrestore(&task_group_lock, flags);
7878 /* change task's runqueue when it moves between groups.
7879 * The caller of this function should have put the task in its new group
7880 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7881 * reflect its new group.
7883 void sched_move_task(struct task_struct *tsk)
7885 struct task_group *tg;
7886 int queued, running;
7887 unsigned long flags;
7890 rq = task_rq_lock(tsk, &flags);
7892 running = task_current(rq, tsk);
7893 queued = task_on_rq_queued(tsk);
7896 dequeue_task(rq, tsk, DEQUEUE_SAVE);
7897 if (unlikely(running))
7898 put_prev_task(rq, tsk);
7901 * All callers are synchronized by task_rq_lock(); we do not use RCU
7902 * which is pointless here. Thus, we pass "true" to task_css_check()
7903 * to prevent lockdep warnings.
7905 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7906 struct task_group, css);
7907 tg = autogroup_task_group(tsk, tg);
7908 tsk->sched_task_group = tg;
7910 #ifdef CONFIG_FAIR_GROUP_SCHED
7911 if (tsk->sched_class->task_move_group)
7912 tsk->sched_class->task_move_group(tsk);
7915 set_task_rq(tsk, task_cpu(tsk));
7917 if (unlikely(running))
7918 tsk->sched_class->set_curr_task(rq);
7920 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
7922 task_rq_unlock(rq, tsk, &flags);
7924 #endif /* CONFIG_CGROUP_SCHED */
7926 #ifdef CONFIG_RT_GROUP_SCHED
7928 * Ensure that the real time constraints are schedulable.
7930 static DEFINE_MUTEX(rt_constraints_mutex);
7932 /* Must be called with tasklist_lock held */
7933 static inline int tg_has_rt_tasks(struct task_group *tg)
7935 struct task_struct *g, *p;
7938 * Autogroups do not have RT tasks; see autogroup_create().
7940 if (task_group_is_autogroup(tg))
7943 for_each_process_thread(g, p) {
7944 if (rt_task(p) && task_group(p) == tg)
7951 struct rt_schedulable_data {
7952 struct task_group *tg;
7957 static int tg_rt_schedulable(struct task_group *tg, void *data)
7959 struct rt_schedulable_data *d = data;
7960 struct task_group *child;
7961 unsigned long total, sum = 0;
7962 u64 period, runtime;
7964 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7965 runtime = tg->rt_bandwidth.rt_runtime;
7968 period = d->rt_period;
7969 runtime = d->rt_runtime;
7973 * Cannot have more runtime than the period.
7975 if (runtime > period && runtime != RUNTIME_INF)
7979 * Ensure we don't starve existing RT tasks.
7981 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7984 total = to_ratio(period, runtime);
7987 * Nobody can have more than the global setting allows.
7989 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7993 * The sum of our children's runtime should not exceed our own.
7995 list_for_each_entry_rcu(child, &tg->children, siblings) {
7996 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7997 runtime = child->rt_bandwidth.rt_runtime;
7999 if (child == d->tg) {
8000 period = d->rt_period;
8001 runtime = d->rt_runtime;
8004 sum += to_ratio(period, runtime);
8013 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8017 struct rt_schedulable_data data = {
8019 .rt_period = period,
8020 .rt_runtime = runtime,
8024 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8030 static int tg_set_rt_bandwidth(struct task_group *tg,
8031 u64 rt_period, u64 rt_runtime)
8036 * Disallowing the root group RT runtime is BAD, it would disallow the
8037 * kernel creating (and or operating) RT threads.
8039 if (tg == &root_task_group && rt_runtime == 0)
8042 /* No period doesn't make any sense. */
8046 mutex_lock(&rt_constraints_mutex);
8047 read_lock(&tasklist_lock);
8048 err = __rt_schedulable(tg, rt_period, rt_runtime);
8052 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8053 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8054 tg->rt_bandwidth.rt_runtime = rt_runtime;
8056 for_each_possible_cpu(i) {
8057 struct rt_rq *rt_rq = tg->rt_rq[i];
8059 raw_spin_lock(&rt_rq->rt_runtime_lock);
8060 rt_rq->rt_runtime = rt_runtime;
8061 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8063 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8065 read_unlock(&tasklist_lock);
8066 mutex_unlock(&rt_constraints_mutex);
8071 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8073 u64 rt_runtime, rt_period;
8075 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8076 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8077 if (rt_runtime_us < 0)
8078 rt_runtime = RUNTIME_INF;
8080 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8083 static long sched_group_rt_runtime(struct task_group *tg)
8087 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8090 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8091 do_div(rt_runtime_us, NSEC_PER_USEC);
8092 return rt_runtime_us;
8095 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8097 u64 rt_runtime, rt_period;
8099 rt_period = rt_period_us * NSEC_PER_USEC;
8100 rt_runtime = tg->rt_bandwidth.rt_runtime;
8102 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8105 static long sched_group_rt_period(struct task_group *tg)
8109 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8110 do_div(rt_period_us, NSEC_PER_USEC);
8111 return rt_period_us;
8113 #endif /* CONFIG_RT_GROUP_SCHED */
8115 #ifdef CONFIG_RT_GROUP_SCHED
8116 static int sched_rt_global_constraints(void)
8120 mutex_lock(&rt_constraints_mutex);
8121 read_lock(&tasklist_lock);
8122 ret = __rt_schedulable(NULL, 0, 0);
8123 read_unlock(&tasklist_lock);
8124 mutex_unlock(&rt_constraints_mutex);
8129 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8131 /* Don't accept realtime tasks when there is no way for them to run */
8132 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8138 #else /* !CONFIG_RT_GROUP_SCHED */
8139 static int sched_rt_global_constraints(void)
8141 unsigned long flags;
8144 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8145 for_each_possible_cpu(i) {
8146 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8148 raw_spin_lock(&rt_rq->rt_runtime_lock);
8149 rt_rq->rt_runtime = global_rt_runtime();
8150 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8152 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8156 #endif /* CONFIG_RT_GROUP_SCHED */
8158 static int sched_dl_global_validate(void)
8160 u64 runtime = global_rt_runtime();
8161 u64 period = global_rt_period();
8162 u64 new_bw = to_ratio(period, runtime);
8165 unsigned long flags;
8168 * Here we want to check the bandwidth not being set to some
8169 * value smaller than the currently allocated bandwidth in
8170 * any of the root_domains.
8172 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8173 * cycling on root_domains... Discussion on different/better
8174 * solutions is welcome!
8176 for_each_possible_cpu(cpu) {
8177 rcu_read_lock_sched();
8178 dl_b = dl_bw_of(cpu);
8180 raw_spin_lock_irqsave(&dl_b->lock, flags);
8181 if (new_bw < dl_b->total_bw)
8183 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8185 rcu_read_unlock_sched();
8194 static void sched_dl_do_global(void)
8199 unsigned long flags;
8201 def_dl_bandwidth.dl_period = global_rt_period();
8202 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8204 if (global_rt_runtime() != RUNTIME_INF)
8205 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8208 * FIXME: As above...
8210 for_each_possible_cpu(cpu) {
8211 rcu_read_lock_sched();
8212 dl_b = dl_bw_of(cpu);
8214 raw_spin_lock_irqsave(&dl_b->lock, flags);
8216 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8218 rcu_read_unlock_sched();
8222 static int sched_rt_global_validate(void)
8224 if (sysctl_sched_rt_period <= 0)
8227 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8228 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8234 static void sched_rt_do_global(void)
8236 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8237 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8240 int sched_rt_handler(struct ctl_table *table, int write,
8241 void __user *buffer, size_t *lenp,
8244 int old_period, old_runtime;
8245 static DEFINE_MUTEX(mutex);
8249 old_period = sysctl_sched_rt_period;
8250 old_runtime = sysctl_sched_rt_runtime;
8252 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8254 if (!ret && write) {
8255 ret = sched_rt_global_validate();
8259 ret = sched_dl_global_validate();
8263 ret = sched_rt_global_constraints();
8267 sched_rt_do_global();
8268 sched_dl_do_global();
8272 sysctl_sched_rt_period = old_period;
8273 sysctl_sched_rt_runtime = old_runtime;
8275 mutex_unlock(&mutex);
8280 int sched_rr_handler(struct ctl_table *table, int write,
8281 void __user *buffer, size_t *lenp,
8285 static DEFINE_MUTEX(mutex);
8288 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8289 /* make sure that internally we keep jiffies */
8290 /* also, writing zero resets timeslice to default */
8291 if (!ret && write) {
8292 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8293 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8295 mutex_unlock(&mutex);
8299 #ifdef CONFIG_CGROUP_SCHED
8301 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8303 return css ? container_of(css, struct task_group, css) : NULL;
8306 static struct cgroup_subsys_state *
8307 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8309 struct task_group *parent = css_tg(parent_css);
8310 struct task_group *tg;
8313 /* This is early initialization for the top cgroup */
8314 return &root_task_group.css;
8317 tg = sched_create_group(parent);
8319 return ERR_PTR(-ENOMEM);
8324 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8326 struct task_group *tg = css_tg(css);
8327 struct task_group *parent = css_tg(css->parent);
8330 sched_online_group(tg, parent);
8334 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8336 struct task_group *tg = css_tg(css);
8338 sched_destroy_group(tg);
8341 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8343 struct task_group *tg = css_tg(css);
8345 sched_offline_group(tg);
8348 static void cpu_cgroup_fork(struct task_struct *task)
8350 sched_move_task(task);
8353 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8355 struct task_struct *task;
8356 struct cgroup_subsys_state *css;
8358 cgroup_taskset_for_each(task, css, tset) {
8359 #ifdef CONFIG_RT_GROUP_SCHED
8360 if (!sched_rt_can_attach(css_tg(css), task))
8363 /* We don't support RT-tasks being in separate groups */
8364 if (task->sched_class != &fair_sched_class)
8371 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8373 struct task_struct *task;
8374 struct cgroup_subsys_state *css;
8376 cgroup_taskset_for_each(task, css, tset)
8377 sched_move_task(task);
8380 #ifdef CONFIG_FAIR_GROUP_SCHED
8381 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8382 struct cftype *cftype, u64 shareval)
8384 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8387 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8390 struct task_group *tg = css_tg(css);
8392 return (u64) scale_load_down(tg->shares);
8395 #ifdef CONFIG_CFS_BANDWIDTH
8396 static DEFINE_MUTEX(cfs_constraints_mutex);
8398 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8399 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8401 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8403 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8405 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8406 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8408 if (tg == &root_task_group)
8412 * Ensure we have at some amount of bandwidth every period. This is
8413 * to prevent reaching a state of large arrears when throttled via
8414 * entity_tick() resulting in prolonged exit starvation.
8416 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8420 * Likewise, bound things on the otherside by preventing insane quota
8421 * periods. This also allows us to normalize in computing quota
8424 if (period > max_cfs_quota_period)
8428 * Prevent race between setting of cfs_rq->runtime_enabled and
8429 * unthrottle_offline_cfs_rqs().
8432 mutex_lock(&cfs_constraints_mutex);
8433 ret = __cfs_schedulable(tg, period, quota);
8437 runtime_enabled = quota != RUNTIME_INF;
8438 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8440 * If we need to toggle cfs_bandwidth_used, off->on must occur
8441 * before making related changes, and on->off must occur afterwards
8443 if (runtime_enabled && !runtime_was_enabled)
8444 cfs_bandwidth_usage_inc();
8445 raw_spin_lock_irq(&cfs_b->lock);
8446 cfs_b->period = ns_to_ktime(period);
8447 cfs_b->quota = quota;
8449 __refill_cfs_bandwidth_runtime(cfs_b);
8450 /* restart the period timer (if active) to handle new period expiry */
8451 if (runtime_enabled)
8452 start_cfs_bandwidth(cfs_b);
8453 raw_spin_unlock_irq(&cfs_b->lock);
8455 for_each_online_cpu(i) {
8456 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8457 struct rq *rq = cfs_rq->rq;
8459 raw_spin_lock_irq(&rq->lock);
8460 cfs_rq->runtime_enabled = runtime_enabled;
8461 cfs_rq->runtime_remaining = 0;
8463 if (cfs_rq->throttled)
8464 unthrottle_cfs_rq(cfs_rq);
8465 raw_spin_unlock_irq(&rq->lock);
8467 if (runtime_was_enabled && !runtime_enabled)
8468 cfs_bandwidth_usage_dec();
8470 mutex_unlock(&cfs_constraints_mutex);
8476 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8480 period = ktime_to_ns(tg->cfs_bandwidth.period);
8481 if (cfs_quota_us < 0)
8482 quota = RUNTIME_INF;
8484 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8486 return tg_set_cfs_bandwidth(tg, period, quota);
8489 long tg_get_cfs_quota(struct task_group *tg)
8493 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8496 quota_us = tg->cfs_bandwidth.quota;
8497 do_div(quota_us, NSEC_PER_USEC);
8502 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8506 period = (u64)cfs_period_us * NSEC_PER_USEC;
8507 quota = tg->cfs_bandwidth.quota;
8509 return tg_set_cfs_bandwidth(tg, period, quota);
8512 long tg_get_cfs_period(struct task_group *tg)
8516 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8517 do_div(cfs_period_us, NSEC_PER_USEC);
8519 return cfs_period_us;
8522 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8525 return tg_get_cfs_quota(css_tg(css));
8528 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8529 struct cftype *cftype, s64 cfs_quota_us)
8531 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8534 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8537 return tg_get_cfs_period(css_tg(css));
8540 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8541 struct cftype *cftype, u64 cfs_period_us)
8543 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8546 struct cfs_schedulable_data {
8547 struct task_group *tg;
8552 * normalize group quota/period to be quota/max_period
8553 * note: units are usecs
8555 static u64 normalize_cfs_quota(struct task_group *tg,
8556 struct cfs_schedulable_data *d)
8564 period = tg_get_cfs_period(tg);
8565 quota = tg_get_cfs_quota(tg);
8568 /* note: these should typically be equivalent */
8569 if (quota == RUNTIME_INF || quota == -1)
8572 return to_ratio(period, quota);
8575 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8577 struct cfs_schedulable_data *d = data;
8578 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8579 s64 quota = 0, parent_quota = -1;
8582 quota = RUNTIME_INF;
8584 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8586 quota = normalize_cfs_quota(tg, d);
8587 parent_quota = parent_b->hierarchical_quota;
8590 * ensure max(child_quota) <= parent_quota, inherit when no
8593 if (quota == RUNTIME_INF)
8594 quota = parent_quota;
8595 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8598 cfs_b->hierarchical_quota = quota;
8603 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8606 struct cfs_schedulable_data data = {
8612 if (quota != RUNTIME_INF) {
8613 do_div(data.period, NSEC_PER_USEC);
8614 do_div(data.quota, NSEC_PER_USEC);
8618 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8624 static int cpu_stats_show(struct seq_file *sf, void *v)
8626 struct task_group *tg = css_tg(seq_css(sf));
8627 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8629 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8630 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8631 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8635 #endif /* CONFIG_CFS_BANDWIDTH */
8636 #endif /* CONFIG_FAIR_GROUP_SCHED */
8638 #ifdef CONFIG_RT_GROUP_SCHED
8639 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8640 struct cftype *cft, s64 val)
8642 return sched_group_set_rt_runtime(css_tg(css), val);
8645 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8648 return sched_group_rt_runtime(css_tg(css));
8651 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8652 struct cftype *cftype, u64 rt_period_us)
8654 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8657 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8660 return sched_group_rt_period(css_tg(css));
8662 #endif /* CONFIG_RT_GROUP_SCHED */
8664 static struct cftype cpu_files[] = {
8665 #ifdef CONFIG_FAIR_GROUP_SCHED
8668 .read_u64 = cpu_shares_read_u64,
8669 .write_u64 = cpu_shares_write_u64,
8672 #ifdef CONFIG_CFS_BANDWIDTH
8674 .name = "cfs_quota_us",
8675 .read_s64 = cpu_cfs_quota_read_s64,
8676 .write_s64 = cpu_cfs_quota_write_s64,
8679 .name = "cfs_period_us",
8680 .read_u64 = cpu_cfs_period_read_u64,
8681 .write_u64 = cpu_cfs_period_write_u64,
8685 .seq_show = cpu_stats_show,
8688 #ifdef CONFIG_RT_GROUP_SCHED
8690 .name = "rt_runtime_us",
8691 .read_s64 = cpu_rt_runtime_read,
8692 .write_s64 = cpu_rt_runtime_write,
8695 .name = "rt_period_us",
8696 .read_u64 = cpu_rt_period_read_uint,
8697 .write_u64 = cpu_rt_period_write_uint,
8703 struct cgroup_subsys cpu_cgrp_subsys = {
8704 .css_alloc = cpu_cgroup_css_alloc,
8705 .css_free = cpu_cgroup_css_free,
8706 .css_online = cpu_cgroup_css_online,
8707 .css_offline = cpu_cgroup_css_offline,
8708 .fork = cpu_cgroup_fork,
8709 .can_attach = cpu_cgroup_can_attach,
8710 .attach = cpu_cgroup_attach,
8711 .legacy_cftypes = cpu_files,
8715 #endif /* CONFIG_CGROUP_SCHED */
8717 void dump_cpu_task(int cpu)
8719 pr_info("Task dump for CPU %d:\n", cpu);
8720 sched_show_task(cpu_curr(cpu));
8724 * Nice levels are multiplicative, with a gentle 10% change for every
8725 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8726 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8727 * that remained on nice 0.
8729 * The "10% effect" is relative and cumulative: from _any_ nice level,
8730 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8731 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8732 * If a task goes up by ~10% and another task goes down by ~10% then
8733 * the relative distance between them is ~25%.)
8735 const int sched_prio_to_weight[40] = {
8736 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8737 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8738 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8739 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8740 /* 0 */ 1024, 820, 655, 526, 423,
8741 /* 5 */ 335, 272, 215, 172, 137,
8742 /* 10 */ 110, 87, 70, 56, 45,
8743 /* 15 */ 36, 29, 23, 18, 15,
8747 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8749 * In cases where the weight does not change often, we can use the
8750 * precalculated inverse to speed up arithmetics by turning divisions
8751 * into multiplications:
8753 const u32 sched_prio_to_wmult[40] = {
8754 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8755 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8756 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8757 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8758 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8759 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8760 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8761 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,