4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
77 #include <linux/frame.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);
3292 STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3294 static inline void sched_submit_work(struct task_struct *tsk)
3296 if (!tsk->state || tsk_is_pi_blocked(tsk))
3299 * If we are going to sleep and we have plugged IO queued,
3300 * make sure to submit it to avoid deadlocks.
3302 if (blk_needs_flush_plug(tsk))
3303 blk_schedule_flush_plug(tsk);
3306 asmlinkage __visible void __sched schedule(void)
3308 struct task_struct *tsk = current;
3310 sched_submit_work(tsk);
3314 sched_preempt_enable_no_resched();
3315 } while (need_resched());
3317 EXPORT_SYMBOL(schedule);
3319 #ifdef CONFIG_CONTEXT_TRACKING
3320 asmlinkage __visible void __sched schedule_user(void)
3323 * If we come here after a random call to set_need_resched(),
3324 * or we have been woken up remotely but the IPI has not yet arrived,
3325 * we haven't yet exited the RCU idle mode. Do it here manually until
3326 * we find a better solution.
3328 * NB: There are buggy callers of this function. Ideally we
3329 * should warn if prev_state != CONTEXT_USER, but that will trigger
3330 * too frequently to make sense yet.
3332 enum ctx_state prev_state = exception_enter();
3334 exception_exit(prev_state);
3339 * schedule_preempt_disabled - called with preemption disabled
3341 * Returns with preemption disabled. Note: preempt_count must be 1
3343 void __sched schedule_preempt_disabled(void)
3345 sched_preempt_enable_no_resched();
3350 static void __sched notrace preempt_schedule_common(void)
3353 preempt_disable_notrace();
3355 preempt_enable_no_resched_notrace();
3358 * Check again in case we missed a preemption opportunity
3359 * between schedule and now.
3361 } while (need_resched());
3364 #ifdef CONFIG_PREEMPT
3366 * this is the entry point to schedule() from in-kernel preemption
3367 * off of preempt_enable. Kernel preemptions off return from interrupt
3368 * occur there and call schedule directly.
3370 asmlinkage __visible void __sched notrace preempt_schedule(void)
3373 * If there is a non-zero preempt_count or interrupts are disabled,
3374 * we do not want to preempt the current task. Just return..
3376 if (likely(!preemptible()))
3379 preempt_schedule_common();
3381 NOKPROBE_SYMBOL(preempt_schedule);
3382 EXPORT_SYMBOL(preempt_schedule);
3385 * preempt_schedule_notrace - preempt_schedule called by tracing
3387 * The tracing infrastructure uses preempt_enable_notrace to prevent
3388 * recursion and tracing preempt enabling caused by the tracing
3389 * infrastructure itself. But as tracing can happen in areas coming
3390 * from userspace or just about to enter userspace, a preempt enable
3391 * can occur before user_exit() is called. This will cause the scheduler
3392 * to be called when the system is still in usermode.
3394 * To prevent this, the preempt_enable_notrace will use this function
3395 * instead of preempt_schedule() to exit user context if needed before
3396 * calling the scheduler.
3398 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3400 enum ctx_state prev_ctx;
3402 if (likely(!preemptible()))
3406 preempt_disable_notrace();
3408 * Needs preempt disabled in case user_exit() is traced
3409 * and the tracer calls preempt_enable_notrace() causing
3410 * an infinite recursion.
3412 prev_ctx = exception_enter();
3414 exception_exit(prev_ctx);
3416 preempt_enable_no_resched_notrace();
3417 } while (need_resched());
3419 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3421 #endif /* CONFIG_PREEMPT */
3424 * this is the entry point to schedule() from kernel preemption
3425 * off of irq context.
3426 * Note, that this is called and return with irqs disabled. This will
3427 * protect us against recursive calling from irq.
3429 asmlinkage __visible void __sched preempt_schedule_irq(void)
3431 enum ctx_state prev_state;
3433 /* Catch callers which need to be fixed */
3434 BUG_ON(preempt_count() || !irqs_disabled());
3436 prev_state = exception_enter();
3442 local_irq_disable();
3443 sched_preempt_enable_no_resched();
3444 } while (need_resched());
3446 exception_exit(prev_state);
3449 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3452 return try_to_wake_up(curr->private, mode, wake_flags);
3454 EXPORT_SYMBOL(default_wake_function);
3456 #ifdef CONFIG_RT_MUTEXES
3459 * rt_mutex_setprio - set the current priority of a task
3461 * @prio: prio value (kernel-internal form)
3463 * This function changes the 'effective' priority of a task. It does
3464 * not touch ->normal_prio like __setscheduler().
3466 * Used by the rt_mutex code to implement priority inheritance
3467 * logic. Call site only calls if the priority of the task changed.
3469 void rt_mutex_setprio(struct task_struct *p, int prio)
3471 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3473 const struct sched_class *prev_class;
3475 BUG_ON(prio > MAX_PRIO);
3477 rq = __task_rq_lock(p);
3480 * Idle task boosting is a nono in general. There is one
3481 * exception, when PREEMPT_RT and NOHZ is active:
3483 * The idle task calls get_next_timer_interrupt() and holds
3484 * the timer wheel base->lock on the CPU and another CPU wants
3485 * to access the timer (probably to cancel it). We can safely
3486 * ignore the boosting request, as the idle CPU runs this code
3487 * with interrupts disabled and will complete the lock
3488 * protected section without being interrupted. So there is no
3489 * real need to boost.
3491 if (unlikely(p == rq->idle)) {
3492 WARN_ON(p != rq->curr);
3493 WARN_ON(p->pi_blocked_on);
3497 trace_sched_pi_setprio(p, prio);
3499 prev_class = p->sched_class;
3500 queued = task_on_rq_queued(p);
3501 running = task_current(rq, p);
3503 dequeue_task(rq, p, DEQUEUE_SAVE);
3505 put_prev_task(rq, p);
3508 * Boosting condition are:
3509 * 1. -rt task is running and holds mutex A
3510 * --> -dl task blocks on mutex A
3512 * 2. -dl task is running and holds mutex A
3513 * --> -dl task blocks on mutex A and could preempt the
3516 if (dl_prio(prio)) {
3517 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3518 if (!dl_prio(p->normal_prio) ||
3519 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3520 p->dl.dl_boosted = 1;
3521 enqueue_flag |= ENQUEUE_REPLENISH;
3523 p->dl.dl_boosted = 0;
3524 p->sched_class = &dl_sched_class;
3525 } else if (rt_prio(prio)) {
3526 if (dl_prio(oldprio))
3527 p->dl.dl_boosted = 0;
3529 enqueue_flag |= ENQUEUE_HEAD;
3530 p->sched_class = &rt_sched_class;
3532 if (dl_prio(oldprio))
3533 p->dl.dl_boosted = 0;
3534 if (rt_prio(oldprio))
3536 p->sched_class = &fair_sched_class;
3542 p->sched_class->set_curr_task(rq);
3544 enqueue_task(rq, p, enqueue_flag);
3546 check_class_changed(rq, p, prev_class, oldprio);
3548 preempt_disable(); /* avoid rq from going away on us */
3549 __task_rq_unlock(rq);
3551 balance_callback(rq);
3556 void set_user_nice(struct task_struct *p, long nice)
3558 int old_prio, delta, queued;
3559 unsigned long flags;
3562 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3565 * We have to be careful, if called from sys_setpriority(),
3566 * the task might be in the middle of scheduling on another CPU.
3568 rq = task_rq_lock(p, &flags);
3570 * The RT priorities are set via sched_setscheduler(), but we still
3571 * allow the 'normal' nice value to be set - but as expected
3572 * it wont have any effect on scheduling until the task is
3573 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3575 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3576 p->static_prio = NICE_TO_PRIO(nice);
3579 queued = task_on_rq_queued(p);
3581 dequeue_task(rq, p, DEQUEUE_SAVE);
3583 p->static_prio = NICE_TO_PRIO(nice);
3586 p->prio = effective_prio(p);
3587 delta = p->prio - old_prio;
3590 enqueue_task(rq, p, ENQUEUE_RESTORE);
3592 * If the task increased its priority or is running and
3593 * lowered its priority, then reschedule its CPU:
3595 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3599 task_rq_unlock(rq, p, &flags);
3601 EXPORT_SYMBOL(set_user_nice);
3604 * can_nice - check if a task can reduce its nice value
3608 int can_nice(const struct task_struct *p, const int nice)
3610 /* convert nice value [19,-20] to rlimit style value [1,40] */
3611 int nice_rlim = nice_to_rlimit(nice);
3613 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3614 capable(CAP_SYS_NICE));
3617 #ifdef __ARCH_WANT_SYS_NICE
3620 * sys_nice - change the priority of the current process.
3621 * @increment: priority increment
3623 * sys_setpriority is a more generic, but much slower function that
3624 * does similar things.
3626 SYSCALL_DEFINE1(nice, int, increment)
3631 * Setpriority might change our priority at the same moment.
3632 * We don't have to worry. Conceptually one call occurs first
3633 * and we have a single winner.
3635 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3636 nice = task_nice(current) + increment;
3638 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3639 if (increment < 0 && !can_nice(current, nice))
3642 retval = security_task_setnice(current, nice);
3646 set_user_nice(current, nice);
3653 * task_prio - return the priority value of a given task.
3654 * @p: the task in question.
3656 * Return: The priority value as seen by users in /proc.
3657 * RT tasks are offset by -200. Normal tasks are centered
3658 * around 0, value goes from -16 to +15.
3660 int task_prio(const struct task_struct *p)
3662 return p->prio - MAX_RT_PRIO;
3666 * idle_cpu - is a given cpu idle currently?
3667 * @cpu: the processor in question.
3669 * Return: 1 if the CPU is currently idle. 0 otherwise.
3671 int idle_cpu(int cpu)
3673 struct rq *rq = cpu_rq(cpu);
3675 if (rq->curr != rq->idle)
3682 if (!llist_empty(&rq->wake_list))
3690 * idle_task - return the idle task for a given cpu.
3691 * @cpu: the processor in question.
3693 * Return: The idle task for the cpu @cpu.
3695 struct task_struct *idle_task(int cpu)
3697 return cpu_rq(cpu)->idle;
3701 * find_process_by_pid - find a process with a matching PID value.
3702 * @pid: the pid in question.
3704 * The task of @pid, if found. %NULL otherwise.
3706 static struct task_struct *find_process_by_pid(pid_t pid)
3708 return pid ? find_task_by_vpid(pid) : current;
3712 * This function initializes the sched_dl_entity of a newly becoming
3713 * SCHED_DEADLINE task.
3715 * Only the static values are considered here, the actual runtime and the
3716 * absolute deadline will be properly calculated when the task is enqueued
3717 * for the first time with its new policy.
3720 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3722 struct sched_dl_entity *dl_se = &p->dl;
3724 dl_se->dl_runtime = attr->sched_runtime;
3725 dl_se->dl_deadline = attr->sched_deadline;
3726 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3727 dl_se->flags = attr->sched_flags;
3728 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3731 * Changing the parameters of a task is 'tricky' and we're not doing
3732 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3734 * What we SHOULD do is delay the bandwidth release until the 0-lag
3735 * point. This would include retaining the task_struct until that time
3736 * and change dl_overflow() to not immediately decrement the current
3739 * Instead we retain the current runtime/deadline and let the new
3740 * parameters take effect after the current reservation period lapses.
3741 * This is safe (albeit pessimistic) because the 0-lag point is always
3742 * before the current scheduling deadline.
3744 * We can still have temporary overloads because we do not delay the
3745 * change in bandwidth until that time; so admission control is
3746 * not on the safe side. It does however guarantee tasks will never
3747 * consume more than promised.
3752 * sched_setparam() passes in -1 for its policy, to let the functions
3753 * it calls know not to change it.
3755 #define SETPARAM_POLICY -1
3757 static void __setscheduler_params(struct task_struct *p,
3758 const struct sched_attr *attr)
3760 int policy = attr->sched_policy;
3762 if (policy == SETPARAM_POLICY)
3767 if (dl_policy(policy))
3768 __setparam_dl(p, attr);
3769 else if (fair_policy(policy))
3770 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3773 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3774 * !rt_policy. Always setting this ensures that things like
3775 * getparam()/getattr() don't report silly values for !rt tasks.
3777 p->rt_priority = attr->sched_priority;
3778 p->normal_prio = normal_prio(p);
3782 /* Actually do priority change: must hold pi & rq lock. */
3783 static void __setscheduler(struct rq *rq, struct task_struct *p,
3784 const struct sched_attr *attr, bool keep_boost)
3786 __setscheduler_params(p, attr);
3789 * Keep a potential priority boosting if called from
3790 * sched_setscheduler().
3793 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3795 p->prio = normal_prio(p);
3797 if (dl_prio(p->prio))
3798 p->sched_class = &dl_sched_class;
3799 else if (rt_prio(p->prio))
3800 p->sched_class = &rt_sched_class;
3802 p->sched_class = &fair_sched_class;
3806 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3808 struct sched_dl_entity *dl_se = &p->dl;
3810 attr->sched_priority = p->rt_priority;
3811 attr->sched_runtime = dl_se->dl_runtime;
3812 attr->sched_deadline = dl_se->dl_deadline;
3813 attr->sched_period = dl_se->dl_period;
3814 attr->sched_flags = dl_se->flags;
3818 * This function validates the new parameters of a -deadline task.
3819 * We ask for the deadline not being zero, and greater or equal
3820 * than the runtime, as well as the period of being zero or
3821 * greater than deadline. Furthermore, we have to be sure that
3822 * user parameters are above the internal resolution of 1us (we
3823 * check sched_runtime only since it is always the smaller one) and
3824 * below 2^63 ns (we have to check both sched_deadline and
3825 * sched_period, as the latter can be zero).
3828 __checkparam_dl(const struct sched_attr *attr)
3831 if (attr->sched_deadline == 0)
3835 * Since we truncate DL_SCALE bits, make sure we're at least
3838 if (attr->sched_runtime < (1ULL << DL_SCALE))
3842 * Since we use the MSB for wrap-around and sign issues, make
3843 * sure it's not set (mind that period can be equal to zero).
3845 if (attr->sched_deadline & (1ULL << 63) ||
3846 attr->sched_period & (1ULL << 63))
3849 /* runtime <= deadline <= period (if period != 0) */
3850 if ((attr->sched_period != 0 &&
3851 attr->sched_period < attr->sched_deadline) ||
3852 attr->sched_deadline < attr->sched_runtime)
3859 * check the target process has a UID that matches the current process's
3861 static bool check_same_owner(struct task_struct *p)
3863 const struct cred *cred = current_cred(), *pcred;
3867 pcred = __task_cred(p);
3868 match = (uid_eq(cred->euid, pcred->euid) ||
3869 uid_eq(cred->euid, pcred->uid));
3874 static bool dl_param_changed(struct task_struct *p,
3875 const struct sched_attr *attr)
3877 struct sched_dl_entity *dl_se = &p->dl;
3879 if (dl_se->dl_runtime != attr->sched_runtime ||
3880 dl_se->dl_deadline != attr->sched_deadline ||
3881 dl_se->dl_period != attr->sched_period ||
3882 dl_se->flags != attr->sched_flags)
3888 static int __sched_setscheduler(struct task_struct *p,
3889 const struct sched_attr *attr,
3892 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3893 MAX_RT_PRIO - 1 - attr->sched_priority;
3894 int retval, oldprio, oldpolicy = -1, queued, running;
3895 int new_effective_prio, policy = attr->sched_policy;
3896 unsigned long flags;
3897 const struct sched_class *prev_class;
3901 /* may grab non-irq protected spin_locks */
3902 BUG_ON(in_interrupt());
3904 /* double check policy once rq lock held */
3906 reset_on_fork = p->sched_reset_on_fork;
3907 policy = oldpolicy = p->policy;
3909 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3911 if (!valid_policy(policy))
3915 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3919 * Valid priorities for SCHED_FIFO and SCHED_RR are
3920 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3921 * SCHED_BATCH and SCHED_IDLE is 0.
3923 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3924 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3926 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3927 (rt_policy(policy) != (attr->sched_priority != 0)))
3931 * Allow unprivileged RT tasks to decrease priority:
3933 if (user && !capable(CAP_SYS_NICE)) {
3934 if (fair_policy(policy)) {
3935 if (attr->sched_nice < task_nice(p) &&
3936 !can_nice(p, attr->sched_nice))
3940 if (rt_policy(policy)) {
3941 unsigned long rlim_rtprio =
3942 task_rlimit(p, RLIMIT_RTPRIO);
3944 /* can't set/change the rt policy */
3945 if (policy != p->policy && !rlim_rtprio)
3948 /* can't increase priority */
3949 if (attr->sched_priority > p->rt_priority &&
3950 attr->sched_priority > rlim_rtprio)
3955 * Can't set/change SCHED_DEADLINE policy at all for now
3956 * (safest behavior); in the future we would like to allow
3957 * unprivileged DL tasks to increase their relative deadline
3958 * or reduce their runtime (both ways reducing utilization)
3960 if (dl_policy(policy))
3964 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3965 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3967 if (idle_policy(p->policy) && !idle_policy(policy)) {
3968 if (!can_nice(p, task_nice(p)))
3972 /* can't change other user's priorities */
3973 if (!check_same_owner(p))
3976 /* Normal users shall not reset the sched_reset_on_fork flag */
3977 if (p->sched_reset_on_fork && !reset_on_fork)
3982 retval = security_task_setscheduler(p);
3988 * make sure no PI-waiters arrive (or leave) while we are
3989 * changing the priority of the task:
3991 * To be able to change p->policy safely, the appropriate
3992 * runqueue lock must be held.
3994 rq = task_rq_lock(p, &flags);
3997 * Changing the policy of the stop threads its a very bad idea
3999 if (p == rq->stop) {
4000 task_rq_unlock(rq, p, &flags);
4005 * If not changing anything there's no need to proceed further,
4006 * but store a possible modification of reset_on_fork.
4008 if (unlikely(policy == p->policy)) {
4009 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4011 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4013 if (dl_policy(policy) && dl_param_changed(p, attr))
4016 p->sched_reset_on_fork = reset_on_fork;
4017 task_rq_unlock(rq, p, &flags);
4023 #ifdef CONFIG_RT_GROUP_SCHED
4025 * Do not allow realtime tasks into groups that have no runtime
4028 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4029 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4030 !task_group_is_autogroup(task_group(p))) {
4031 task_rq_unlock(rq, p, &flags);
4036 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4037 cpumask_t *span = rq->rd->span;
4040 * Don't allow tasks with an affinity mask smaller than
4041 * the entire root_domain to become SCHED_DEADLINE. We
4042 * will also fail if there's no bandwidth available.
4044 if (!cpumask_subset(span, &p->cpus_allowed) ||
4045 rq->rd->dl_bw.bw == 0) {
4046 task_rq_unlock(rq, p, &flags);
4053 /* recheck policy now with rq lock held */
4054 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4055 policy = oldpolicy = -1;
4056 task_rq_unlock(rq, p, &flags);
4061 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4062 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4065 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4066 task_rq_unlock(rq, p, &flags);
4070 p->sched_reset_on_fork = reset_on_fork;
4075 * Take priority boosted tasks into account. If the new
4076 * effective priority is unchanged, we just store the new
4077 * normal parameters and do not touch the scheduler class and
4078 * the runqueue. This will be done when the task deboost
4081 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4082 if (new_effective_prio == oldprio) {
4083 __setscheduler_params(p, attr);
4084 task_rq_unlock(rq, p, &flags);
4089 queued = task_on_rq_queued(p);
4090 running = task_current(rq, p);
4092 dequeue_task(rq, p, DEQUEUE_SAVE);
4094 put_prev_task(rq, p);
4096 prev_class = p->sched_class;
4097 __setscheduler(rq, p, attr, pi);
4100 p->sched_class->set_curr_task(rq);
4102 int enqueue_flags = ENQUEUE_RESTORE;
4104 * We enqueue to tail when the priority of a task is
4105 * increased (user space view).
4107 if (oldprio <= p->prio)
4108 enqueue_flags |= ENQUEUE_HEAD;
4110 enqueue_task(rq, p, enqueue_flags);
4113 check_class_changed(rq, p, prev_class, oldprio);
4114 preempt_disable(); /* avoid rq from going away on us */
4115 task_rq_unlock(rq, p, &flags);
4118 rt_mutex_adjust_pi(p);
4121 * Run balance callbacks after we've adjusted the PI chain.
4123 balance_callback(rq);
4129 static int _sched_setscheduler(struct task_struct *p, int policy,
4130 const struct sched_param *param, bool check)
4132 struct sched_attr attr = {
4133 .sched_policy = policy,
4134 .sched_priority = param->sched_priority,
4135 .sched_nice = PRIO_TO_NICE(p->static_prio),
4138 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4139 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4140 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4141 policy &= ~SCHED_RESET_ON_FORK;
4142 attr.sched_policy = policy;
4145 return __sched_setscheduler(p, &attr, check, true);
4148 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4149 * @p: the task in question.
4150 * @policy: new policy.
4151 * @param: structure containing the new RT priority.
4153 * Return: 0 on success. An error code otherwise.
4155 * NOTE that the task may be already dead.
4157 int sched_setscheduler(struct task_struct *p, int policy,
4158 const struct sched_param *param)
4160 return _sched_setscheduler(p, policy, param, true);
4162 EXPORT_SYMBOL_GPL(sched_setscheduler);
4164 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4166 return __sched_setscheduler(p, attr, true, true);
4168 EXPORT_SYMBOL_GPL(sched_setattr);
4171 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4172 * @p: the task in question.
4173 * @policy: new policy.
4174 * @param: structure containing the new RT priority.
4176 * Just like sched_setscheduler, only don't bother checking if the
4177 * current context has permission. For example, this is needed in
4178 * stop_machine(): we create temporary high priority worker threads,
4179 * but our caller might not have that capability.
4181 * Return: 0 on success. An error code otherwise.
4183 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4184 const struct sched_param *param)
4186 return _sched_setscheduler(p, policy, param, false);
4188 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4191 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4193 struct sched_param lparam;
4194 struct task_struct *p;
4197 if (!param || pid < 0)
4199 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4204 p = find_process_by_pid(pid);
4206 retval = sched_setscheduler(p, policy, &lparam);
4213 * Mimics kernel/events/core.c perf_copy_attr().
4215 static int sched_copy_attr(struct sched_attr __user *uattr,
4216 struct sched_attr *attr)
4221 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4225 * zero the full structure, so that a short copy will be nice.
4227 memset(attr, 0, sizeof(*attr));
4229 ret = get_user(size, &uattr->size);
4233 if (size > PAGE_SIZE) /* silly large */
4236 if (!size) /* abi compat */
4237 size = SCHED_ATTR_SIZE_VER0;
4239 if (size < SCHED_ATTR_SIZE_VER0)
4243 * If we're handed a bigger struct than we know of,
4244 * ensure all the unknown bits are 0 - i.e. new
4245 * user-space does not rely on any kernel feature
4246 * extensions we dont know about yet.
4248 if (size > sizeof(*attr)) {
4249 unsigned char __user *addr;
4250 unsigned char __user *end;
4253 addr = (void __user *)uattr + sizeof(*attr);
4254 end = (void __user *)uattr + size;
4256 for (; addr < end; addr++) {
4257 ret = get_user(val, addr);
4263 size = sizeof(*attr);
4266 ret = copy_from_user(attr, uattr, size);
4271 * XXX: do we want to be lenient like existing syscalls; or do we want
4272 * to be strict and return an error on out-of-bounds values?
4274 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4279 put_user(sizeof(*attr), &uattr->size);
4284 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4285 * @pid: the pid in question.
4286 * @policy: new policy.
4287 * @param: structure containing the new RT priority.
4289 * Return: 0 on success. An error code otherwise.
4291 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4292 struct sched_param __user *, param)
4294 /* negative values for policy are not valid */
4298 return do_sched_setscheduler(pid, policy, param);
4302 * sys_sched_setparam - set/change the RT priority of a thread
4303 * @pid: the pid in question.
4304 * @param: structure containing the new RT priority.
4306 * Return: 0 on success. An error code otherwise.
4308 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4310 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4314 * sys_sched_setattr - same as above, but with extended sched_attr
4315 * @pid: the pid in question.
4316 * @uattr: structure containing the extended parameters.
4317 * @flags: for future extension.
4319 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4320 unsigned int, flags)
4322 struct sched_attr attr;
4323 struct task_struct *p;
4326 if (!uattr || pid < 0 || flags)
4329 retval = sched_copy_attr(uattr, &attr);
4333 if ((int)attr.sched_policy < 0)
4338 p = find_process_by_pid(pid);
4340 retval = sched_setattr(p, &attr);
4347 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4348 * @pid: the pid in question.
4350 * Return: On success, the policy of the thread. Otherwise, a negative error
4353 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4355 struct task_struct *p;
4363 p = find_process_by_pid(pid);
4365 retval = security_task_getscheduler(p);
4368 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4375 * sys_sched_getparam - get the RT priority of a thread
4376 * @pid: the pid in question.
4377 * @param: structure containing the RT priority.
4379 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4382 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4384 struct sched_param lp = { .sched_priority = 0 };
4385 struct task_struct *p;
4388 if (!param || pid < 0)
4392 p = find_process_by_pid(pid);
4397 retval = security_task_getscheduler(p);
4401 if (task_has_rt_policy(p))
4402 lp.sched_priority = p->rt_priority;
4406 * This one might sleep, we cannot do it with a spinlock held ...
4408 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4417 static int sched_read_attr(struct sched_attr __user *uattr,
4418 struct sched_attr *attr,
4423 if (!access_ok(VERIFY_WRITE, uattr, usize))
4427 * If we're handed a smaller struct than we know of,
4428 * ensure all the unknown bits are 0 - i.e. old
4429 * user-space does not get uncomplete information.
4431 if (usize < sizeof(*attr)) {
4432 unsigned char *addr;
4435 addr = (void *)attr + usize;
4436 end = (void *)attr + sizeof(*attr);
4438 for (; addr < end; addr++) {
4446 ret = copy_to_user(uattr, attr, attr->size);
4454 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4455 * @pid: the pid in question.
4456 * @uattr: structure containing the extended parameters.
4457 * @size: sizeof(attr) for fwd/bwd comp.
4458 * @flags: for future extension.
4460 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4461 unsigned int, size, unsigned int, flags)
4463 struct sched_attr attr = {
4464 .size = sizeof(struct sched_attr),
4466 struct task_struct *p;
4469 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4470 size < SCHED_ATTR_SIZE_VER0 || flags)
4474 p = find_process_by_pid(pid);
4479 retval = security_task_getscheduler(p);
4483 attr.sched_policy = p->policy;
4484 if (p->sched_reset_on_fork)
4485 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4486 if (task_has_dl_policy(p))
4487 __getparam_dl(p, &attr);
4488 else if (task_has_rt_policy(p))
4489 attr.sched_priority = p->rt_priority;
4491 attr.sched_nice = task_nice(p);
4495 retval = sched_read_attr(uattr, &attr, size);
4503 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4505 cpumask_var_t cpus_allowed, new_mask;
4506 struct task_struct *p;
4511 p = find_process_by_pid(pid);
4517 /* Prevent p going away */
4521 if (p->flags & PF_NO_SETAFFINITY) {
4525 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4529 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4531 goto out_free_cpus_allowed;
4534 if (!check_same_owner(p)) {
4536 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4538 goto out_free_new_mask;
4543 retval = security_task_setscheduler(p);
4545 goto out_free_new_mask;
4548 cpuset_cpus_allowed(p, cpus_allowed);
4549 cpumask_and(new_mask, in_mask, cpus_allowed);
4552 * Since bandwidth control happens on root_domain basis,
4553 * if admission test is enabled, we only admit -deadline
4554 * tasks allowed to run on all the CPUs in the task's
4558 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4560 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4563 goto out_free_new_mask;
4569 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4572 cpuset_cpus_allowed(p, cpus_allowed);
4573 if (!cpumask_subset(new_mask, cpus_allowed)) {
4575 * We must have raced with a concurrent cpuset
4576 * update. Just reset the cpus_allowed to the
4577 * cpuset's cpus_allowed
4579 cpumask_copy(new_mask, cpus_allowed);
4584 free_cpumask_var(new_mask);
4585 out_free_cpus_allowed:
4586 free_cpumask_var(cpus_allowed);
4592 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4593 struct cpumask *new_mask)
4595 if (len < cpumask_size())
4596 cpumask_clear(new_mask);
4597 else if (len > cpumask_size())
4598 len = cpumask_size();
4600 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4604 * sys_sched_setaffinity - set the cpu affinity of a process
4605 * @pid: pid of the process
4606 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4607 * @user_mask_ptr: user-space pointer to the new cpu mask
4609 * Return: 0 on success. An error code otherwise.
4611 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4612 unsigned long __user *, user_mask_ptr)
4614 cpumask_var_t new_mask;
4617 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4620 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4622 retval = sched_setaffinity(pid, new_mask);
4623 free_cpumask_var(new_mask);
4627 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4629 struct task_struct *p;
4630 unsigned long flags;
4636 p = find_process_by_pid(pid);
4640 retval = security_task_getscheduler(p);
4644 raw_spin_lock_irqsave(&p->pi_lock, flags);
4645 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4646 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4655 * sys_sched_getaffinity - get the cpu affinity of a process
4656 * @pid: pid of the process
4657 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4658 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4660 * Return: 0 on success. An error code otherwise.
4662 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4663 unsigned long __user *, user_mask_ptr)
4668 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4670 if (len & (sizeof(unsigned long)-1))
4673 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4676 ret = sched_getaffinity(pid, mask);
4678 size_t retlen = min_t(size_t, len, cpumask_size());
4680 if (copy_to_user(user_mask_ptr, mask, retlen))
4685 free_cpumask_var(mask);
4691 * sys_sched_yield - yield the current processor to other threads.
4693 * This function yields the current CPU to other tasks. If there are no
4694 * other threads running on this CPU then this function will return.
4698 SYSCALL_DEFINE0(sched_yield)
4700 struct rq *rq = this_rq_lock();
4702 schedstat_inc(rq, yld_count);
4703 current->sched_class->yield_task(rq);
4706 * Since we are going to call schedule() anyway, there's
4707 * no need to preempt or enable interrupts:
4709 __release(rq->lock);
4710 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4711 do_raw_spin_unlock(&rq->lock);
4712 sched_preempt_enable_no_resched();
4719 int __sched _cond_resched(void)
4721 if (should_resched(0)) {
4722 preempt_schedule_common();
4727 EXPORT_SYMBOL(_cond_resched);
4730 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4731 * call schedule, and on return reacquire the lock.
4733 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4734 * operations here to prevent schedule() from being called twice (once via
4735 * spin_unlock(), once by hand).
4737 int __cond_resched_lock(spinlock_t *lock)
4739 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4742 lockdep_assert_held(lock);
4744 if (spin_needbreak(lock) || resched) {
4747 preempt_schedule_common();
4755 EXPORT_SYMBOL(__cond_resched_lock);
4757 int __sched __cond_resched_softirq(void)
4759 BUG_ON(!in_softirq());
4761 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4763 preempt_schedule_common();
4769 EXPORT_SYMBOL(__cond_resched_softirq);
4772 * yield - yield the current processor to other threads.
4774 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4776 * The scheduler is at all times free to pick the calling task as the most
4777 * eligible task to run, if removing the yield() call from your code breaks
4778 * it, its already broken.
4780 * Typical broken usage is:
4785 * where one assumes that yield() will let 'the other' process run that will
4786 * make event true. If the current task is a SCHED_FIFO task that will never
4787 * happen. Never use yield() as a progress guarantee!!
4789 * If you want to use yield() to wait for something, use wait_event().
4790 * If you want to use yield() to be 'nice' for others, use cond_resched().
4791 * If you still want to use yield(), do not!
4793 void __sched yield(void)
4795 set_current_state(TASK_RUNNING);
4798 EXPORT_SYMBOL(yield);
4801 * yield_to - yield the current processor to another thread in
4802 * your thread group, or accelerate that thread toward the
4803 * processor it's on.
4805 * @preempt: whether task preemption is allowed or not
4807 * It's the caller's job to ensure that the target task struct
4808 * can't go away on us before we can do any checks.
4811 * true (>0) if we indeed boosted the target task.
4812 * false (0) if we failed to boost the target.
4813 * -ESRCH if there's no task to yield to.
4815 int __sched yield_to(struct task_struct *p, bool preempt)
4817 struct task_struct *curr = current;
4818 struct rq *rq, *p_rq;
4819 unsigned long flags;
4822 local_irq_save(flags);
4828 * If we're the only runnable task on the rq and target rq also
4829 * has only one task, there's absolutely no point in yielding.
4831 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4836 double_rq_lock(rq, p_rq);
4837 if (task_rq(p) != p_rq) {
4838 double_rq_unlock(rq, p_rq);
4842 if (!curr->sched_class->yield_to_task)
4845 if (curr->sched_class != p->sched_class)
4848 if (task_running(p_rq, p) || p->state)
4851 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4853 schedstat_inc(rq, yld_count);
4855 * Make p's CPU reschedule; pick_next_entity takes care of
4858 if (preempt && rq != p_rq)
4863 double_rq_unlock(rq, p_rq);
4865 local_irq_restore(flags);
4872 EXPORT_SYMBOL_GPL(yield_to);
4875 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4876 * that process accounting knows that this is a task in IO wait state.
4878 long __sched io_schedule_timeout(long timeout)
4880 int old_iowait = current->in_iowait;
4884 current->in_iowait = 1;
4885 blk_schedule_flush_plug(current);
4887 delayacct_blkio_start();
4889 atomic_inc(&rq->nr_iowait);
4890 ret = schedule_timeout(timeout);
4891 current->in_iowait = old_iowait;
4892 atomic_dec(&rq->nr_iowait);
4893 delayacct_blkio_end();
4897 EXPORT_SYMBOL(io_schedule_timeout);
4900 * sys_sched_get_priority_max - return maximum RT priority.
4901 * @policy: scheduling class.
4903 * Return: On success, this syscall returns the maximum
4904 * rt_priority that can be used by a given scheduling class.
4905 * On failure, a negative error code is returned.
4907 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4914 ret = MAX_USER_RT_PRIO-1;
4916 case SCHED_DEADLINE:
4927 * sys_sched_get_priority_min - return minimum RT priority.
4928 * @policy: scheduling class.
4930 * Return: On success, this syscall returns the minimum
4931 * rt_priority that can be used by a given scheduling class.
4932 * On failure, a negative error code is returned.
4934 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4943 case SCHED_DEADLINE:
4953 * sys_sched_rr_get_interval - return the default timeslice of a process.
4954 * @pid: pid of the process.
4955 * @interval: userspace pointer to the timeslice value.
4957 * this syscall writes the default timeslice value of a given process
4958 * into the user-space timespec buffer. A value of '0' means infinity.
4960 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4963 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4964 struct timespec __user *, interval)
4966 struct task_struct *p;
4967 unsigned int time_slice;
4968 unsigned long flags;
4978 p = find_process_by_pid(pid);
4982 retval = security_task_getscheduler(p);
4986 rq = task_rq_lock(p, &flags);
4988 if (p->sched_class->get_rr_interval)
4989 time_slice = p->sched_class->get_rr_interval(rq, p);
4990 task_rq_unlock(rq, p, &flags);
4993 jiffies_to_timespec(time_slice, &t);
4994 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5002 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5004 void sched_show_task(struct task_struct *p)
5006 unsigned long free = 0;
5008 unsigned long state = p->state;
5011 state = __ffs(state) + 1;
5012 printk(KERN_INFO "%-15.15s %c", p->comm,
5013 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5014 #if BITS_PER_LONG == 32
5015 if (state == TASK_RUNNING)
5016 printk(KERN_CONT " running ");
5018 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5020 if (state == TASK_RUNNING)
5021 printk(KERN_CONT " running task ");
5023 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5025 #ifdef CONFIG_DEBUG_STACK_USAGE
5026 free = stack_not_used(p);
5031 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5033 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5034 task_pid_nr(p), ppid,
5035 (unsigned long)task_thread_info(p)->flags);
5037 print_worker_info(KERN_INFO, p);
5038 show_stack(p, NULL);
5041 void show_state_filter(unsigned long state_filter)
5043 struct task_struct *g, *p;
5045 #if BITS_PER_LONG == 32
5047 " task PC stack pid father\n");
5050 " task PC stack pid father\n");
5053 for_each_process_thread(g, p) {
5055 * reset the NMI-timeout, listing all files on a slow
5056 * console might take a lot of time:
5058 touch_nmi_watchdog();
5059 if (!state_filter || (p->state & state_filter))
5063 touch_all_softlockup_watchdogs();
5065 #ifdef CONFIG_SCHED_DEBUG
5066 sysrq_sched_debug_show();
5070 * Only show locks if all tasks are dumped:
5073 debug_show_all_locks();
5076 void init_idle_bootup_task(struct task_struct *idle)
5078 idle->sched_class = &idle_sched_class;
5082 * init_idle - set up an idle thread for a given CPU
5083 * @idle: task in question
5084 * @cpu: cpu the idle task belongs to
5086 * NOTE: this function does not set the idle thread's NEED_RESCHED
5087 * flag, to make booting more robust.
5089 void init_idle(struct task_struct *idle, int cpu)
5091 struct rq *rq = cpu_rq(cpu);
5092 unsigned long flags;
5094 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5095 raw_spin_lock(&rq->lock);
5097 __sched_fork(0, idle);
5098 idle->state = TASK_RUNNING;
5099 idle->se.exec_start = sched_clock();
5103 * Its possible that init_idle() gets called multiple times on a task,
5104 * in that case do_set_cpus_allowed() will not do the right thing.
5106 * And since this is boot we can forgo the serialization.
5108 set_cpus_allowed_common(idle, cpumask_of(cpu));
5111 * We're having a chicken and egg problem, even though we are
5112 * holding rq->lock, the cpu isn't yet set to this cpu so the
5113 * lockdep check in task_group() will fail.
5115 * Similar case to sched_fork(). / Alternatively we could
5116 * use task_rq_lock() here and obtain the other rq->lock.
5121 __set_task_cpu(idle, cpu);
5124 rq->curr = rq->idle = idle;
5125 idle->on_rq = TASK_ON_RQ_QUEUED;
5129 raw_spin_unlock(&rq->lock);
5130 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5132 /* Set the preempt count _outside_ the spinlocks! */
5133 init_idle_preempt_count(idle, cpu);
5136 * The idle tasks have their own, simple scheduling class:
5138 idle->sched_class = &idle_sched_class;
5139 ftrace_graph_init_idle_task(idle, cpu);
5140 vtime_init_idle(idle, cpu);
5142 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5146 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5147 const struct cpumask *trial)
5149 int ret = 1, trial_cpus;
5150 struct dl_bw *cur_dl_b;
5151 unsigned long flags;
5153 if (!cpumask_weight(cur))
5156 rcu_read_lock_sched();
5157 cur_dl_b = dl_bw_of(cpumask_any(cur));
5158 trial_cpus = cpumask_weight(trial);
5160 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5161 if (cur_dl_b->bw != -1 &&
5162 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5164 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5165 rcu_read_unlock_sched();
5170 int task_can_attach(struct task_struct *p,
5171 const struct cpumask *cs_cpus_allowed)
5176 * Kthreads which disallow setaffinity shouldn't be moved
5177 * to a new cpuset; we don't want to change their cpu
5178 * affinity and isolating such threads by their set of
5179 * allowed nodes is unnecessary. Thus, cpusets are not
5180 * applicable for such threads. This prevents checking for
5181 * success of set_cpus_allowed_ptr() on all attached tasks
5182 * before cpus_allowed may be changed.
5184 if (p->flags & PF_NO_SETAFFINITY) {
5190 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5192 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5197 unsigned long flags;
5199 rcu_read_lock_sched();
5200 dl_b = dl_bw_of(dest_cpu);
5201 raw_spin_lock_irqsave(&dl_b->lock, flags);
5202 cpus = dl_bw_cpus(dest_cpu);
5203 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5208 * We reserve space for this task in the destination
5209 * root_domain, as we can't fail after this point.
5210 * We will free resources in the source root_domain
5211 * later on (see set_cpus_allowed_dl()).
5213 __dl_add(dl_b, p->dl.dl_bw);
5215 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5216 rcu_read_unlock_sched();
5226 #ifdef CONFIG_NUMA_BALANCING
5227 /* Migrate current task p to target_cpu */
5228 int migrate_task_to(struct task_struct *p, int target_cpu)
5230 struct migration_arg arg = { p, target_cpu };
5231 int curr_cpu = task_cpu(p);
5233 if (curr_cpu == target_cpu)
5236 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5239 /* TODO: This is not properly updating schedstats */
5241 trace_sched_move_numa(p, curr_cpu, target_cpu);
5242 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5246 * Requeue a task on a given node and accurately track the number of NUMA
5247 * tasks on the runqueues
5249 void sched_setnuma(struct task_struct *p, int nid)
5252 unsigned long flags;
5253 bool queued, running;
5255 rq = task_rq_lock(p, &flags);
5256 queued = task_on_rq_queued(p);
5257 running = task_current(rq, p);
5260 dequeue_task(rq, p, DEQUEUE_SAVE);
5262 put_prev_task(rq, p);
5264 p->numa_preferred_nid = nid;
5267 p->sched_class->set_curr_task(rq);
5269 enqueue_task(rq, p, ENQUEUE_RESTORE);
5270 task_rq_unlock(rq, p, &flags);
5272 #endif /* CONFIG_NUMA_BALANCING */
5274 #ifdef CONFIG_HOTPLUG_CPU
5276 * Ensures that the idle task is using init_mm right before its cpu goes
5279 void idle_task_exit(void)
5281 struct mm_struct *mm = current->active_mm;
5283 BUG_ON(cpu_online(smp_processor_id()));
5285 if (mm != &init_mm) {
5286 switch_mm(mm, &init_mm, current);
5287 finish_arch_post_lock_switch();
5293 * Since this CPU is going 'away' for a while, fold any nr_active delta
5294 * we might have. Assumes we're called after migrate_tasks() so that the
5295 * nr_active count is stable.
5297 * Also see the comment "Global load-average calculations".
5299 static void calc_load_migrate(struct rq *rq)
5301 long delta = calc_load_fold_active(rq);
5303 atomic_long_add(delta, &calc_load_tasks);
5306 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5310 static const struct sched_class fake_sched_class = {
5311 .put_prev_task = put_prev_task_fake,
5314 static struct task_struct fake_task = {
5316 * Avoid pull_{rt,dl}_task()
5318 .prio = MAX_PRIO + 1,
5319 .sched_class = &fake_sched_class,
5323 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5324 * try_to_wake_up()->select_task_rq().
5326 * Called with rq->lock held even though we'er in stop_machine() and
5327 * there's no concurrency possible, we hold the required locks anyway
5328 * because of lock validation efforts.
5330 static void migrate_tasks(struct rq *dead_rq)
5332 struct rq *rq = dead_rq;
5333 struct task_struct *next, *stop = rq->stop;
5337 * Fudge the rq selection such that the below task selection loop
5338 * doesn't get stuck on the currently eligible stop task.
5340 * We're currently inside stop_machine() and the rq is either stuck
5341 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5342 * either way we should never end up calling schedule() until we're
5348 * put_prev_task() and pick_next_task() sched
5349 * class method both need to have an up-to-date
5350 * value of rq->clock[_task]
5352 update_rq_clock(rq);
5356 * There's this thread running, bail when that's the only
5359 if (rq->nr_running == 1)
5363 * pick_next_task assumes pinned rq->lock.
5365 lockdep_pin_lock(&rq->lock);
5366 next = pick_next_task(rq, &fake_task);
5368 next->sched_class->put_prev_task(rq, next);
5371 * Rules for changing task_struct::cpus_allowed are holding
5372 * both pi_lock and rq->lock, such that holding either
5373 * stabilizes the mask.
5375 * Drop rq->lock is not quite as disastrous as it usually is
5376 * because !cpu_active at this point, which means load-balance
5377 * will not interfere. Also, stop-machine.
5379 lockdep_unpin_lock(&rq->lock);
5380 raw_spin_unlock(&rq->lock);
5381 raw_spin_lock(&next->pi_lock);
5382 raw_spin_lock(&rq->lock);
5385 * Since we're inside stop-machine, _nothing_ should have
5386 * changed the task, WARN if weird stuff happened, because in
5387 * that case the above rq->lock drop is a fail too.
5389 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5390 raw_spin_unlock(&next->pi_lock);
5394 /* Find suitable destination for @next, with force if needed. */
5395 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5397 rq = __migrate_task(rq, next, dest_cpu);
5398 if (rq != dead_rq) {
5399 raw_spin_unlock(&rq->lock);
5401 raw_spin_lock(&rq->lock);
5403 raw_spin_unlock(&next->pi_lock);
5408 #endif /* CONFIG_HOTPLUG_CPU */
5410 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5412 static struct ctl_table sd_ctl_dir[] = {
5414 .procname = "sched_domain",
5420 static struct ctl_table sd_ctl_root[] = {
5422 .procname = "kernel",
5424 .child = sd_ctl_dir,
5429 static struct ctl_table *sd_alloc_ctl_entry(int n)
5431 struct ctl_table *entry =
5432 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5437 static void sd_free_ctl_entry(struct ctl_table **tablep)
5439 struct ctl_table *entry;
5442 * In the intermediate directories, both the child directory and
5443 * procname are dynamically allocated and could fail but the mode
5444 * will always be set. In the lowest directory the names are
5445 * static strings and all have proc handlers.
5447 for (entry = *tablep; entry->mode; entry++) {
5449 sd_free_ctl_entry(&entry->child);
5450 if (entry->proc_handler == NULL)
5451 kfree(entry->procname);
5458 static int min_load_idx = 0;
5459 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5462 set_table_entry(struct ctl_table *entry,
5463 const char *procname, void *data, int maxlen,
5464 umode_t mode, proc_handler *proc_handler,
5467 entry->procname = procname;
5469 entry->maxlen = maxlen;
5471 entry->proc_handler = proc_handler;
5474 entry->extra1 = &min_load_idx;
5475 entry->extra2 = &max_load_idx;
5479 static struct ctl_table *
5480 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5482 struct ctl_table *table = sd_alloc_ctl_entry(14);
5487 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5488 sizeof(long), 0644, proc_doulongvec_minmax, false);
5489 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5490 sizeof(long), 0644, proc_doulongvec_minmax, false);
5491 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5492 sizeof(int), 0644, proc_dointvec_minmax, true);
5493 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5494 sizeof(int), 0644, proc_dointvec_minmax, true);
5495 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5496 sizeof(int), 0644, proc_dointvec_minmax, true);
5497 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5498 sizeof(int), 0644, proc_dointvec_minmax, true);
5499 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5500 sizeof(int), 0644, proc_dointvec_minmax, true);
5501 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5502 sizeof(int), 0644, proc_dointvec_minmax, false);
5503 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5504 sizeof(int), 0644, proc_dointvec_minmax, false);
5505 set_table_entry(&table[9], "cache_nice_tries",
5506 &sd->cache_nice_tries,
5507 sizeof(int), 0644, proc_dointvec_minmax, false);
5508 set_table_entry(&table[10], "flags", &sd->flags,
5509 sizeof(int), 0644, proc_dointvec_minmax, false);
5510 set_table_entry(&table[11], "max_newidle_lb_cost",
5511 &sd->max_newidle_lb_cost,
5512 sizeof(long), 0644, proc_doulongvec_minmax, false);
5513 set_table_entry(&table[12], "name", sd->name,
5514 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5515 /* &table[13] is terminator */
5520 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5522 struct ctl_table *entry, *table;
5523 struct sched_domain *sd;
5524 int domain_num = 0, i;
5527 for_each_domain(cpu, sd)
5529 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5534 for_each_domain(cpu, sd) {
5535 snprintf(buf, 32, "domain%d", i);
5536 entry->procname = kstrdup(buf, GFP_KERNEL);
5538 entry->child = sd_alloc_ctl_domain_table(sd);
5545 static struct ctl_table_header *sd_sysctl_header;
5546 static void register_sched_domain_sysctl(void)
5548 int i, cpu_num = num_possible_cpus();
5549 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5552 WARN_ON(sd_ctl_dir[0].child);
5553 sd_ctl_dir[0].child = entry;
5558 for_each_possible_cpu(i) {
5559 snprintf(buf, 32, "cpu%d", i);
5560 entry->procname = kstrdup(buf, GFP_KERNEL);
5562 entry->child = sd_alloc_ctl_cpu_table(i);
5566 WARN_ON(sd_sysctl_header);
5567 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5570 /* may be called multiple times per register */
5571 static void unregister_sched_domain_sysctl(void)
5573 unregister_sysctl_table(sd_sysctl_header);
5574 sd_sysctl_header = NULL;
5575 if (sd_ctl_dir[0].child)
5576 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5579 static void register_sched_domain_sysctl(void)
5582 static void unregister_sched_domain_sysctl(void)
5585 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5587 static void set_rq_online(struct rq *rq)
5590 const struct sched_class *class;
5592 cpumask_set_cpu(rq->cpu, rq->rd->online);
5595 for_each_class(class) {
5596 if (class->rq_online)
5597 class->rq_online(rq);
5602 static void set_rq_offline(struct rq *rq)
5605 const struct sched_class *class;
5607 for_each_class(class) {
5608 if (class->rq_offline)
5609 class->rq_offline(rq);
5612 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5618 * migration_call - callback that gets triggered when a CPU is added.
5619 * Here we can start up the necessary migration thread for the new CPU.
5622 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5624 int cpu = (long)hcpu;
5625 unsigned long flags;
5626 struct rq *rq = cpu_rq(cpu);
5628 switch (action & ~CPU_TASKS_FROZEN) {
5630 case CPU_UP_PREPARE:
5631 rq->calc_load_update = calc_load_update;
5635 /* Update our root-domain */
5636 raw_spin_lock_irqsave(&rq->lock, flags);
5638 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5642 raw_spin_unlock_irqrestore(&rq->lock, flags);
5645 #ifdef CONFIG_HOTPLUG_CPU
5647 sched_ttwu_pending();
5648 /* Update our root-domain */
5649 raw_spin_lock_irqsave(&rq->lock, flags);
5651 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5655 BUG_ON(rq->nr_running != 1); /* the migration thread */
5656 raw_spin_unlock_irqrestore(&rq->lock, flags);
5660 calc_load_migrate(rq);
5665 update_max_interval();
5671 * Register at high priority so that task migration (migrate_all_tasks)
5672 * happens before everything else. This has to be lower priority than
5673 * the notifier in the perf_event subsystem, though.
5675 static struct notifier_block migration_notifier = {
5676 .notifier_call = migration_call,
5677 .priority = CPU_PRI_MIGRATION,
5680 static void set_cpu_rq_start_time(void)
5682 int cpu = smp_processor_id();
5683 struct rq *rq = cpu_rq(cpu);
5684 rq->age_stamp = sched_clock_cpu(cpu);
5687 static int sched_cpu_active(struct notifier_block *nfb,
5688 unsigned long action, void *hcpu)
5690 int cpu = (long)hcpu;
5692 switch (action & ~CPU_TASKS_FROZEN) {
5694 set_cpu_rq_start_time();
5699 * At this point a starting CPU has marked itself as online via
5700 * set_cpu_online(). But it might not yet have marked itself
5701 * as active, which is essential from here on.
5703 set_cpu_active(cpu, true);
5704 stop_machine_unpark(cpu);
5707 case CPU_DOWN_FAILED:
5708 set_cpu_active(cpu, true);
5716 static int sched_cpu_inactive(struct notifier_block *nfb,
5717 unsigned long action, void *hcpu)
5719 switch (action & ~CPU_TASKS_FROZEN) {
5720 case CPU_DOWN_PREPARE:
5721 set_cpu_active((long)hcpu, false);
5728 static int __init migration_init(void)
5730 void *cpu = (void *)(long)smp_processor_id();
5733 /* Initialize migration for the boot CPU */
5734 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5735 BUG_ON(err == NOTIFY_BAD);
5736 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5737 register_cpu_notifier(&migration_notifier);
5739 /* Register cpu active notifiers */
5740 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5741 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5745 early_initcall(migration_init);
5747 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5749 #ifdef CONFIG_SCHED_DEBUG
5751 static __read_mostly int sched_debug_enabled;
5753 static int __init sched_debug_setup(char *str)
5755 sched_debug_enabled = 1;
5759 early_param("sched_debug", sched_debug_setup);
5761 static inline bool sched_debug(void)
5763 return sched_debug_enabled;
5766 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5767 struct cpumask *groupmask)
5769 struct sched_group *group = sd->groups;
5771 cpumask_clear(groupmask);
5773 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5775 if (!(sd->flags & SD_LOAD_BALANCE)) {
5776 printk("does not load-balance\n");
5778 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5783 printk(KERN_CONT "span %*pbl level %s\n",
5784 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5786 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5787 printk(KERN_ERR "ERROR: domain->span does not contain "
5790 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5791 printk(KERN_ERR "ERROR: domain->groups does not contain"
5795 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5799 printk(KERN_ERR "ERROR: group is NULL\n");
5803 if (!cpumask_weight(sched_group_cpus(group))) {
5804 printk(KERN_CONT "\n");
5805 printk(KERN_ERR "ERROR: empty group\n");
5809 if (!(sd->flags & SD_OVERLAP) &&
5810 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5811 printk(KERN_CONT "\n");
5812 printk(KERN_ERR "ERROR: repeated CPUs\n");
5816 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5818 printk(KERN_CONT " %*pbl",
5819 cpumask_pr_args(sched_group_cpus(group)));
5820 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5821 printk(KERN_CONT " (cpu_capacity = %d)",
5822 group->sgc->capacity);
5825 group = group->next;
5826 } while (group != sd->groups);
5827 printk(KERN_CONT "\n");
5829 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5830 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5833 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5834 printk(KERN_ERR "ERROR: parent span is not a superset "
5835 "of domain->span\n");
5839 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5843 if (!sched_debug_enabled)
5847 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5851 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5854 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5862 #else /* !CONFIG_SCHED_DEBUG */
5863 # define sched_domain_debug(sd, cpu) do { } while (0)
5864 static inline bool sched_debug(void)
5868 #endif /* CONFIG_SCHED_DEBUG */
5870 static int sd_degenerate(struct sched_domain *sd)
5872 if (cpumask_weight(sched_domain_span(sd)) == 1)
5875 /* Following flags need at least 2 groups */
5876 if (sd->flags & (SD_LOAD_BALANCE |
5877 SD_BALANCE_NEWIDLE |
5880 SD_SHARE_CPUCAPACITY |
5881 SD_SHARE_PKG_RESOURCES |
5882 SD_SHARE_POWERDOMAIN)) {
5883 if (sd->groups != sd->groups->next)
5887 /* Following flags don't use groups */
5888 if (sd->flags & (SD_WAKE_AFFINE))
5895 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5897 unsigned long cflags = sd->flags, pflags = parent->flags;
5899 if (sd_degenerate(parent))
5902 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5905 /* Flags needing groups don't count if only 1 group in parent */
5906 if (parent->groups == parent->groups->next) {
5907 pflags &= ~(SD_LOAD_BALANCE |
5908 SD_BALANCE_NEWIDLE |
5911 SD_SHARE_CPUCAPACITY |
5912 SD_SHARE_PKG_RESOURCES |
5914 SD_SHARE_POWERDOMAIN);
5915 if (nr_node_ids == 1)
5916 pflags &= ~SD_SERIALIZE;
5918 if (~cflags & pflags)
5924 static void free_rootdomain(struct rcu_head *rcu)
5926 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5928 cpupri_cleanup(&rd->cpupri);
5929 cpudl_cleanup(&rd->cpudl);
5930 free_cpumask_var(rd->dlo_mask);
5931 free_cpumask_var(rd->rto_mask);
5932 free_cpumask_var(rd->online);
5933 free_cpumask_var(rd->span);
5937 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5939 struct root_domain *old_rd = NULL;
5940 unsigned long flags;
5942 raw_spin_lock_irqsave(&rq->lock, flags);
5947 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5950 cpumask_clear_cpu(rq->cpu, old_rd->span);
5953 * If we dont want to free the old_rd yet then
5954 * set old_rd to NULL to skip the freeing later
5957 if (!atomic_dec_and_test(&old_rd->refcount))
5961 atomic_inc(&rd->refcount);
5964 cpumask_set_cpu(rq->cpu, rd->span);
5965 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5968 raw_spin_unlock_irqrestore(&rq->lock, flags);
5971 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5974 static int init_rootdomain(struct root_domain *rd)
5976 memset(rd, 0, sizeof(*rd));
5978 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5980 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5982 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5984 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5987 init_dl_bw(&rd->dl_bw);
5988 if (cpudl_init(&rd->cpudl) != 0)
5991 if (cpupri_init(&rd->cpupri) != 0)
5996 free_cpumask_var(rd->rto_mask);
5998 free_cpumask_var(rd->dlo_mask);
6000 free_cpumask_var(rd->online);
6002 free_cpumask_var(rd->span);
6008 * By default the system creates a single root-domain with all cpus as
6009 * members (mimicking the global state we have today).
6011 struct root_domain def_root_domain;
6013 static void init_defrootdomain(void)
6015 init_rootdomain(&def_root_domain);
6017 atomic_set(&def_root_domain.refcount, 1);
6020 static struct root_domain *alloc_rootdomain(void)
6022 struct root_domain *rd;
6024 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6028 if (init_rootdomain(rd) != 0) {
6036 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6038 struct sched_group *tmp, *first;
6047 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6052 } while (sg != first);
6055 static void free_sched_domain(struct rcu_head *rcu)
6057 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6060 * If its an overlapping domain it has private groups, iterate and
6063 if (sd->flags & SD_OVERLAP) {
6064 free_sched_groups(sd->groups, 1);
6065 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6066 kfree(sd->groups->sgc);
6072 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6074 call_rcu(&sd->rcu, free_sched_domain);
6077 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6079 for (; sd; sd = sd->parent)
6080 destroy_sched_domain(sd, cpu);
6084 * Keep a special pointer to the highest sched_domain that has
6085 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6086 * allows us to avoid some pointer chasing select_idle_sibling().
6088 * Also keep a unique ID per domain (we use the first cpu number in
6089 * the cpumask of the domain), this allows us to quickly tell if
6090 * two cpus are in the same cache domain, see cpus_share_cache().
6092 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6093 DEFINE_PER_CPU(int, sd_llc_size);
6094 DEFINE_PER_CPU(int, sd_llc_id);
6095 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6096 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6097 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6099 static void update_top_cache_domain(int cpu)
6101 struct sched_domain *sd;
6102 struct sched_domain *busy_sd = NULL;
6106 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6108 id = cpumask_first(sched_domain_span(sd));
6109 size = cpumask_weight(sched_domain_span(sd));
6110 busy_sd = sd->parent; /* sd_busy */
6112 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6114 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6115 per_cpu(sd_llc_size, cpu) = size;
6116 per_cpu(sd_llc_id, cpu) = id;
6118 sd = lowest_flag_domain(cpu, SD_NUMA);
6119 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6121 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6122 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6126 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6127 * hold the hotplug lock.
6130 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6132 struct rq *rq = cpu_rq(cpu);
6133 struct sched_domain *tmp;
6135 /* Remove the sched domains which do not contribute to scheduling. */
6136 for (tmp = sd; tmp; ) {
6137 struct sched_domain *parent = tmp->parent;
6141 if (sd_parent_degenerate(tmp, parent)) {
6142 tmp->parent = parent->parent;
6144 parent->parent->child = tmp;
6146 * Transfer SD_PREFER_SIBLING down in case of a
6147 * degenerate parent; the spans match for this
6148 * so the property transfers.
6150 if (parent->flags & SD_PREFER_SIBLING)
6151 tmp->flags |= SD_PREFER_SIBLING;
6152 destroy_sched_domain(parent, cpu);
6157 if (sd && sd_degenerate(sd)) {
6160 destroy_sched_domain(tmp, cpu);
6165 sched_domain_debug(sd, cpu);
6167 rq_attach_root(rq, rd);
6169 rcu_assign_pointer(rq->sd, sd);
6170 destroy_sched_domains(tmp, cpu);
6172 update_top_cache_domain(cpu);
6175 /* Setup the mask of cpus configured for isolated domains */
6176 static int __init isolated_cpu_setup(char *str)
6178 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6179 cpulist_parse(str, cpu_isolated_map);
6183 __setup("isolcpus=", isolated_cpu_setup);
6186 struct sched_domain ** __percpu sd;
6187 struct root_domain *rd;
6198 * Build an iteration mask that can exclude certain CPUs from the upwards
6201 * Asymmetric node setups can result in situations where the domain tree is of
6202 * unequal depth, make sure to skip domains that already cover the entire
6205 * In that case build_sched_domains() will have terminated the iteration early
6206 * and our sibling sd spans will be empty. Domains should always include the
6207 * cpu they're built on, so check that.
6210 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6212 const struct cpumask *span = sched_domain_span(sd);
6213 struct sd_data *sdd = sd->private;
6214 struct sched_domain *sibling;
6217 for_each_cpu(i, span) {
6218 sibling = *per_cpu_ptr(sdd->sd, i);
6219 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6222 cpumask_set_cpu(i, sched_group_mask(sg));
6227 * Return the canonical balance cpu for this group, this is the first cpu
6228 * of this group that's also in the iteration mask.
6230 int group_balance_cpu(struct sched_group *sg)
6232 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6236 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6238 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6239 const struct cpumask *span = sched_domain_span(sd);
6240 struct cpumask *covered = sched_domains_tmpmask;
6241 struct sd_data *sdd = sd->private;
6242 struct sched_domain *sibling;
6245 cpumask_clear(covered);
6247 for_each_cpu(i, span) {
6248 struct cpumask *sg_span;
6250 if (cpumask_test_cpu(i, covered))
6253 sibling = *per_cpu_ptr(sdd->sd, i);
6255 /* See the comment near build_group_mask(). */
6256 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6259 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6260 GFP_KERNEL, cpu_to_node(cpu));
6265 sg_span = sched_group_cpus(sg);
6267 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6269 cpumask_set_cpu(i, sg_span);
6271 cpumask_or(covered, covered, sg_span);
6273 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6274 if (atomic_inc_return(&sg->sgc->ref) == 1)
6275 build_group_mask(sd, sg);
6278 * Initialize sgc->capacity such that even if we mess up the
6279 * domains and no possible iteration will get us here, we won't
6282 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6285 * Make sure the first group of this domain contains the
6286 * canonical balance cpu. Otherwise the sched_domain iteration
6287 * breaks. See update_sg_lb_stats().
6289 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6290 group_balance_cpu(sg) == cpu)
6300 sd->groups = groups;
6305 free_sched_groups(first, 0);
6310 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6312 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6313 struct sched_domain *child = sd->child;
6316 cpu = cpumask_first(sched_domain_span(child));
6319 *sg = *per_cpu_ptr(sdd->sg, cpu);
6320 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6321 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6328 * build_sched_groups will build a circular linked list of the groups
6329 * covered by the given span, and will set each group's ->cpumask correctly,
6330 * and ->cpu_capacity to 0.
6332 * Assumes the sched_domain tree is fully constructed
6335 build_sched_groups(struct sched_domain *sd, int cpu)
6337 struct sched_group *first = NULL, *last = NULL;
6338 struct sd_data *sdd = sd->private;
6339 const struct cpumask *span = sched_domain_span(sd);
6340 struct cpumask *covered;
6343 get_group(cpu, sdd, &sd->groups);
6344 atomic_inc(&sd->groups->ref);
6346 if (cpu != cpumask_first(span))
6349 lockdep_assert_held(&sched_domains_mutex);
6350 covered = sched_domains_tmpmask;
6352 cpumask_clear(covered);
6354 for_each_cpu(i, span) {
6355 struct sched_group *sg;
6358 if (cpumask_test_cpu(i, covered))
6361 group = get_group(i, sdd, &sg);
6362 cpumask_setall(sched_group_mask(sg));
6364 for_each_cpu(j, span) {
6365 if (get_group(j, sdd, NULL) != group)
6368 cpumask_set_cpu(j, covered);
6369 cpumask_set_cpu(j, sched_group_cpus(sg));
6384 * Initialize sched groups cpu_capacity.
6386 * cpu_capacity indicates the capacity of sched group, which is used while
6387 * distributing the load between different sched groups in a sched domain.
6388 * Typically cpu_capacity for all the groups in a sched domain will be same
6389 * unless there are asymmetries in the topology. If there are asymmetries,
6390 * group having more cpu_capacity will pickup more load compared to the
6391 * group having less cpu_capacity.
6393 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6395 struct sched_group *sg = sd->groups;
6400 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6402 } while (sg != sd->groups);
6404 if (cpu != group_balance_cpu(sg))
6407 update_group_capacity(sd, cpu);
6408 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6412 * Initializers for schedule domains
6413 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6416 static int default_relax_domain_level = -1;
6417 int sched_domain_level_max;
6419 static int __init setup_relax_domain_level(char *str)
6421 if (kstrtoint(str, 0, &default_relax_domain_level))
6422 pr_warn("Unable to set relax_domain_level\n");
6426 __setup("relax_domain_level=", setup_relax_domain_level);
6428 static void set_domain_attribute(struct sched_domain *sd,
6429 struct sched_domain_attr *attr)
6433 if (!attr || attr->relax_domain_level < 0) {
6434 if (default_relax_domain_level < 0)
6437 request = default_relax_domain_level;
6439 request = attr->relax_domain_level;
6440 if (request < sd->level) {
6441 /* turn off idle balance on this domain */
6442 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6444 /* turn on idle balance on this domain */
6445 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6449 static void __sdt_free(const struct cpumask *cpu_map);
6450 static int __sdt_alloc(const struct cpumask *cpu_map);
6452 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6453 const struct cpumask *cpu_map)
6457 if (!atomic_read(&d->rd->refcount))
6458 free_rootdomain(&d->rd->rcu); /* fall through */
6460 free_percpu(d->sd); /* fall through */
6462 __sdt_free(cpu_map); /* fall through */
6468 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6469 const struct cpumask *cpu_map)
6471 memset(d, 0, sizeof(*d));
6473 if (__sdt_alloc(cpu_map))
6474 return sa_sd_storage;
6475 d->sd = alloc_percpu(struct sched_domain *);
6477 return sa_sd_storage;
6478 d->rd = alloc_rootdomain();
6481 return sa_rootdomain;
6485 * NULL the sd_data elements we've used to build the sched_domain and
6486 * sched_group structure so that the subsequent __free_domain_allocs()
6487 * will not free the data we're using.
6489 static void claim_allocations(int cpu, struct sched_domain *sd)
6491 struct sd_data *sdd = sd->private;
6493 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6494 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6496 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6497 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6499 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6500 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6504 static int sched_domains_numa_levels;
6505 enum numa_topology_type sched_numa_topology_type;
6506 static int *sched_domains_numa_distance;
6507 int sched_max_numa_distance;
6508 static struct cpumask ***sched_domains_numa_masks;
6509 static int sched_domains_curr_level;
6513 * SD_flags allowed in topology descriptions.
6515 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6516 * SD_SHARE_PKG_RESOURCES - describes shared caches
6517 * SD_NUMA - describes NUMA topologies
6518 * SD_SHARE_POWERDOMAIN - describes shared power domain
6521 * SD_ASYM_PACKING - describes SMT quirks
6523 #define TOPOLOGY_SD_FLAGS \
6524 (SD_SHARE_CPUCAPACITY | \
6525 SD_SHARE_PKG_RESOURCES | \
6528 SD_SHARE_POWERDOMAIN)
6530 static struct sched_domain *
6531 sd_init(struct sched_domain_topology_level *tl, int cpu)
6533 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6534 int sd_weight, sd_flags = 0;
6538 * Ugly hack to pass state to sd_numa_mask()...
6540 sched_domains_curr_level = tl->numa_level;
6543 sd_weight = cpumask_weight(tl->mask(cpu));
6546 sd_flags = (*tl->sd_flags)();
6547 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6548 "wrong sd_flags in topology description\n"))
6549 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6551 *sd = (struct sched_domain){
6552 .min_interval = sd_weight,
6553 .max_interval = 2*sd_weight,
6555 .imbalance_pct = 125,
6557 .cache_nice_tries = 0,
6564 .flags = 1*SD_LOAD_BALANCE
6565 | 1*SD_BALANCE_NEWIDLE
6570 | 0*SD_SHARE_CPUCAPACITY
6571 | 0*SD_SHARE_PKG_RESOURCES
6573 | 0*SD_PREFER_SIBLING
6578 .last_balance = jiffies,
6579 .balance_interval = sd_weight,
6581 .max_newidle_lb_cost = 0,
6582 .next_decay_max_lb_cost = jiffies,
6583 #ifdef CONFIG_SCHED_DEBUG
6589 * Convert topological properties into behaviour.
6592 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6593 sd->flags |= SD_PREFER_SIBLING;
6594 sd->imbalance_pct = 110;
6595 sd->smt_gain = 1178; /* ~15% */
6597 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6598 sd->imbalance_pct = 117;
6599 sd->cache_nice_tries = 1;
6603 } else if (sd->flags & SD_NUMA) {
6604 sd->cache_nice_tries = 2;
6608 sd->flags |= SD_SERIALIZE;
6609 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6610 sd->flags &= ~(SD_BALANCE_EXEC |
6617 sd->flags |= SD_PREFER_SIBLING;
6618 sd->cache_nice_tries = 1;
6623 sd->private = &tl->data;
6629 * Topology list, bottom-up.
6631 static struct sched_domain_topology_level default_topology[] = {
6632 #ifdef CONFIG_SCHED_SMT
6633 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6635 #ifdef CONFIG_SCHED_MC
6636 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6638 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6642 static struct sched_domain_topology_level *sched_domain_topology =
6645 #define for_each_sd_topology(tl) \
6646 for (tl = sched_domain_topology; tl->mask; tl++)
6648 void set_sched_topology(struct sched_domain_topology_level *tl)
6650 sched_domain_topology = tl;
6655 static const struct cpumask *sd_numa_mask(int cpu)
6657 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6660 static void sched_numa_warn(const char *str)
6662 static int done = false;
6670 printk(KERN_WARNING "ERROR: %s\n\n", str);
6672 for (i = 0; i < nr_node_ids; i++) {
6673 printk(KERN_WARNING " ");
6674 for (j = 0; j < nr_node_ids; j++)
6675 printk(KERN_CONT "%02d ", node_distance(i,j));
6676 printk(KERN_CONT "\n");
6678 printk(KERN_WARNING "\n");
6681 bool find_numa_distance(int distance)
6685 if (distance == node_distance(0, 0))
6688 for (i = 0; i < sched_domains_numa_levels; i++) {
6689 if (sched_domains_numa_distance[i] == distance)
6697 * A system can have three types of NUMA topology:
6698 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6699 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6700 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6702 * The difference between a glueless mesh topology and a backplane
6703 * topology lies in whether communication between not directly
6704 * connected nodes goes through intermediary nodes (where programs
6705 * could run), or through backplane controllers. This affects
6706 * placement of programs.
6708 * The type of topology can be discerned with the following tests:
6709 * - If the maximum distance between any nodes is 1 hop, the system
6710 * is directly connected.
6711 * - If for two nodes A and B, located N > 1 hops away from each other,
6712 * there is an intermediary node C, which is < N hops away from both
6713 * nodes A and B, the system is a glueless mesh.
6715 static void init_numa_topology_type(void)
6719 n = sched_max_numa_distance;
6721 if (sched_domains_numa_levels <= 1) {
6722 sched_numa_topology_type = NUMA_DIRECT;
6726 for_each_online_node(a) {
6727 for_each_online_node(b) {
6728 /* Find two nodes furthest removed from each other. */
6729 if (node_distance(a, b) < n)
6732 /* Is there an intermediary node between a and b? */
6733 for_each_online_node(c) {
6734 if (node_distance(a, c) < n &&
6735 node_distance(b, c) < n) {
6736 sched_numa_topology_type =
6742 sched_numa_topology_type = NUMA_BACKPLANE;
6748 static void sched_init_numa(void)
6750 int next_distance, curr_distance = node_distance(0, 0);
6751 struct sched_domain_topology_level *tl;
6755 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6756 if (!sched_domains_numa_distance)
6760 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6761 * unique distances in the node_distance() table.
6763 * Assumes node_distance(0,j) includes all distances in
6764 * node_distance(i,j) in order to avoid cubic time.
6766 next_distance = curr_distance;
6767 for (i = 0; i < nr_node_ids; i++) {
6768 for (j = 0; j < nr_node_ids; j++) {
6769 for (k = 0; k < nr_node_ids; k++) {
6770 int distance = node_distance(i, k);
6772 if (distance > curr_distance &&
6773 (distance < next_distance ||
6774 next_distance == curr_distance))
6775 next_distance = distance;
6778 * While not a strong assumption it would be nice to know
6779 * about cases where if node A is connected to B, B is not
6780 * equally connected to A.
6782 if (sched_debug() && node_distance(k, i) != distance)
6783 sched_numa_warn("Node-distance not symmetric");
6785 if (sched_debug() && i && !find_numa_distance(distance))
6786 sched_numa_warn("Node-0 not representative");
6788 if (next_distance != curr_distance) {
6789 sched_domains_numa_distance[level++] = next_distance;
6790 sched_domains_numa_levels = level;
6791 curr_distance = next_distance;
6796 * In case of sched_debug() we verify the above assumption.
6806 * 'level' contains the number of unique distances, excluding the
6807 * identity distance node_distance(i,i).
6809 * The sched_domains_numa_distance[] array includes the actual distance
6814 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6815 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6816 * the array will contain less then 'level' members. This could be
6817 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6818 * in other functions.
6820 * We reset it to 'level' at the end of this function.
6822 sched_domains_numa_levels = 0;
6824 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6825 if (!sched_domains_numa_masks)
6829 * Now for each level, construct a mask per node which contains all
6830 * cpus of nodes that are that many hops away from us.
6832 for (i = 0; i < level; i++) {
6833 sched_domains_numa_masks[i] =
6834 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6835 if (!sched_domains_numa_masks[i])
6838 for (j = 0; j < nr_node_ids; j++) {
6839 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6843 sched_domains_numa_masks[i][j] = mask;
6846 if (node_distance(j, k) > sched_domains_numa_distance[i])
6849 cpumask_or(mask, mask, cpumask_of_node(k));
6854 /* Compute default topology size */
6855 for (i = 0; sched_domain_topology[i].mask; i++);
6857 tl = kzalloc((i + level + 1) *
6858 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6863 * Copy the default topology bits..
6865 for (i = 0; sched_domain_topology[i].mask; i++)
6866 tl[i] = sched_domain_topology[i];
6869 * .. and append 'j' levels of NUMA goodness.
6871 for (j = 0; j < level; i++, j++) {
6872 tl[i] = (struct sched_domain_topology_level){
6873 .mask = sd_numa_mask,
6874 .sd_flags = cpu_numa_flags,
6875 .flags = SDTL_OVERLAP,
6881 sched_domain_topology = tl;
6883 sched_domains_numa_levels = level;
6884 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6886 init_numa_topology_type();
6889 static void sched_domains_numa_masks_set(int cpu)
6892 int node = cpu_to_node(cpu);
6894 for (i = 0; i < sched_domains_numa_levels; i++) {
6895 for (j = 0; j < nr_node_ids; j++) {
6896 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6897 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6902 static void sched_domains_numa_masks_clear(int cpu)
6905 for (i = 0; i < sched_domains_numa_levels; i++) {
6906 for (j = 0; j < nr_node_ids; j++)
6907 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6912 * Update sched_domains_numa_masks[level][node] array when new cpus
6915 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6916 unsigned long action,
6919 int cpu = (long)hcpu;
6921 switch (action & ~CPU_TASKS_FROZEN) {
6923 sched_domains_numa_masks_set(cpu);
6927 sched_domains_numa_masks_clear(cpu);
6937 static inline void sched_init_numa(void)
6941 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6942 unsigned long action,
6947 #endif /* CONFIG_NUMA */
6949 static int __sdt_alloc(const struct cpumask *cpu_map)
6951 struct sched_domain_topology_level *tl;
6954 for_each_sd_topology(tl) {
6955 struct sd_data *sdd = &tl->data;
6957 sdd->sd = alloc_percpu(struct sched_domain *);
6961 sdd->sg = alloc_percpu(struct sched_group *);
6965 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6969 for_each_cpu(j, cpu_map) {
6970 struct sched_domain *sd;
6971 struct sched_group *sg;
6972 struct sched_group_capacity *sgc;
6974 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6975 GFP_KERNEL, cpu_to_node(j));
6979 *per_cpu_ptr(sdd->sd, j) = sd;
6981 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6982 GFP_KERNEL, cpu_to_node(j));
6988 *per_cpu_ptr(sdd->sg, j) = sg;
6990 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6991 GFP_KERNEL, cpu_to_node(j));
6995 *per_cpu_ptr(sdd->sgc, j) = sgc;
7002 static void __sdt_free(const struct cpumask *cpu_map)
7004 struct sched_domain_topology_level *tl;
7007 for_each_sd_topology(tl) {
7008 struct sd_data *sdd = &tl->data;
7010 for_each_cpu(j, cpu_map) {
7011 struct sched_domain *sd;
7014 sd = *per_cpu_ptr(sdd->sd, j);
7015 if (sd && (sd->flags & SD_OVERLAP))
7016 free_sched_groups(sd->groups, 0);
7017 kfree(*per_cpu_ptr(sdd->sd, j));
7021 kfree(*per_cpu_ptr(sdd->sg, j));
7023 kfree(*per_cpu_ptr(sdd->sgc, j));
7025 free_percpu(sdd->sd);
7027 free_percpu(sdd->sg);
7029 free_percpu(sdd->sgc);
7034 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7035 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7036 struct sched_domain *child, int cpu)
7038 struct sched_domain *sd = sd_init(tl, cpu);
7042 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7044 sd->level = child->level + 1;
7045 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7049 if (!cpumask_subset(sched_domain_span(child),
7050 sched_domain_span(sd))) {
7051 pr_err("BUG: arch topology borken\n");
7052 #ifdef CONFIG_SCHED_DEBUG
7053 pr_err(" the %s domain not a subset of the %s domain\n",
7054 child->name, sd->name);
7056 /* Fixup, ensure @sd has at least @child cpus. */
7057 cpumask_or(sched_domain_span(sd),
7058 sched_domain_span(sd),
7059 sched_domain_span(child));
7063 set_domain_attribute(sd, attr);
7069 * Build sched domains for a given set of cpus and attach the sched domains
7070 * to the individual cpus
7072 static int build_sched_domains(const struct cpumask *cpu_map,
7073 struct sched_domain_attr *attr)
7075 enum s_alloc alloc_state;
7076 struct sched_domain *sd;
7078 int i, ret = -ENOMEM;
7080 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7081 if (alloc_state != sa_rootdomain)
7084 /* Set up domains for cpus specified by the cpu_map. */
7085 for_each_cpu(i, cpu_map) {
7086 struct sched_domain_topology_level *tl;
7089 for_each_sd_topology(tl) {
7090 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7091 if (tl == sched_domain_topology)
7092 *per_cpu_ptr(d.sd, i) = sd;
7093 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7094 sd->flags |= SD_OVERLAP;
7095 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7100 /* Build the groups for the domains */
7101 for_each_cpu(i, cpu_map) {
7102 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7103 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7104 if (sd->flags & SD_OVERLAP) {
7105 if (build_overlap_sched_groups(sd, i))
7108 if (build_sched_groups(sd, i))
7114 /* Calculate CPU capacity for physical packages and nodes */
7115 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7116 if (!cpumask_test_cpu(i, cpu_map))
7119 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7120 claim_allocations(i, sd);
7121 init_sched_groups_capacity(i, sd);
7125 /* Attach the domains */
7127 for_each_cpu(i, cpu_map) {
7128 sd = *per_cpu_ptr(d.sd, i);
7129 cpu_attach_domain(sd, d.rd, i);
7135 __free_domain_allocs(&d, alloc_state, cpu_map);
7139 static cpumask_var_t *doms_cur; /* current sched domains */
7140 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7141 static struct sched_domain_attr *dattr_cur;
7142 /* attribues of custom domains in 'doms_cur' */
7145 * Special case: If a kmalloc of a doms_cur partition (array of
7146 * cpumask) fails, then fallback to a single sched domain,
7147 * as determined by the single cpumask fallback_doms.
7149 static cpumask_var_t fallback_doms;
7152 * arch_update_cpu_topology lets virtualized architectures update the
7153 * cpu core maps. It is supposed to return 1 if the topology changed
7154 * or 0 if it stayed the same.
7156 int __weak arch_update_cpu_topology(void)
7161 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7164 cpumask_var_t *doms;
7166 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7169 for (i = 0; i < ndoms; i++) {
7170 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7171 free_sched_domains(doms, i);
7178 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7181 for (i = 0; i < ndoms; i++)
7182 free_cpumask_var(doms[i]);
7187 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7188 * For now this just excludes isolated cpus, but could be used to
7189 * exclude other special cases in the future.
7191 static int init_sched_domains(const struct cpumask *cpu_map)
7195 arch_update_cpu_topology();
7197 doms_cur = alloc_sched_domains(ndoms_cur);
7199 doms_cur = &fallback_doms;
7200 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7201 err = build_sched_domains(doms_cur[0], NULL);
7202 register_sched_domain_sysctl();
7208 * Detach sched domains from a group of cpus specified in cpu_map
7209 * These cpus will now be attached to the NULL domain
7211 static void detach_destroy_domains(const struct cpumask *cpu_map)
7216 for_each_cpu(i, cpu_map)
7217 cpu_attach_domain(NULL, &def_root_domain, i);
7221 /* handle null as "default" */
7222 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7223 struct sched_domain_attr *new, int idx_new)
7225 struct sched_domain_attr tmp;
7232 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7233 new ? (new + idx_new) : &tmp,
7234 sizeof(struct sched_domain_attr));
7238 * Partition sched domains as specified by the 'ndoms_new'
7239 * cpumasks in the array doms_new[] of cpumasks. This compares
7240 * doms_new[] to the current sched domain partitioning, doms_cur[].
7241 * It destroys each deleted domain and builds each new domain.
7243 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7244 * The masks don't intersect (don't overlap.) We should setup one
7245 * sched domain for each mask. CPUs not in any of the cpumasks will
7246 * not be load balanced. If the same cpumask appears both in the
7247 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7250 * The passed in 'doms_new' should be allocated using
7251 * alloc_sched_domains. This routine takes ownership of it and will
7252 * free_sched_domains it when done with it. If the caller failed the
7253 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7254 * and partition_sched_domains() will fallback to the single partition
7255 * 'fallback_doms', it also forces the domains to be rebuilt.
7257 * If doms_new == NULL it will be replaced with cpu_online_mask.
7258 * ndoms_new == 0 is a special case for destroying existing domains,
7259 * and it will not create the default domain.
7261 * Call with hotplug lock held
7263 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7264 struct sched_domain_attr *dattr_new)
7269 mutex_lock(&sched_domains_mutex);
7271 /* always unregister in case we don't destroy any domains */
7272 unregister_sched_domain_sysctl();
7274 /* Let architecture update cpu core mappings. */
7275 new_topology = arch_update_cpu_topology();
7277 n = doms_new ? ndoms_new : 0;
7279 /* Destroy deleted domains */
7280 for (i = 0; i < ndoms_cur; i++) {
7281 for (j = 0; j < n && !new_topology; j++) {
7282 if (cpumask_equal(doms_cur[i], doms_new[j])
7283 && dattrs_equal(dattr_cur, i, dattr_new, j))
7286 /* no match - a current sched domain not in new doms_new[] */
7287 detach_destroy_domains(doms_cur[i]);
7293 if (doms_new == NULL) {
7295 doms_new = &fallback_doms;
7296 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7297 WARN_ON_ONCE(dattr_new);
7300 /* Build new domains */
7301 for (i = 0; i < ndoms_new; i++) {
7302 for (j = 0; j < n && !new_topology; j++) {
7303 if (cpumask_equal(doms_new[i], doms_cur[j])
7304 && dattrs_equal(dattr_new, i, dattr_cur, j))
7307 /* no match - add a new doms_new */
7308 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7313 /* Remember the new sched domains */
7314 if (doms_cur != &fallback_doms)
7315 free_sched_domains(doms_cur, ndoms_cur);
7316 kfree(dattr_cur); /* kfree(NULL) is safe */
7317 doms_cur = doms_new;
7318 dattr_cur = dattr_new;
7319 ndoms_cur = ndoms_new;
7321 register_sched_domain_sysctl();
7323 mutex_unlock(&sched_domains_mutex);
7326 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7329 * Update cpusets according to cpu_active mask. If cpusets are
7330 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7331 * around partition_sched_domains().
7333 * If we come here as part of a suspend/resume, don't touch cpusets because we
7334 * want to restore it back to its original state upon resume anyway.
7336 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7340 case CPU_ONLINE_FROZEN:
7341 case CPU_DOWN_FAILED_FROZEN:
7344 * num_cpus_frozen tracks how many CPUs are involved in suspend
7345 * resume sequence. As long as this is not the last online
7346 * operation in the resume sequence, just build a single sched
7347 * domain, ignoring cpusets.
7350 if (likely(num_cpus_frozen)) {
7351 partition_sched_domains(1, NULL, NULL);
7356 * This is the last CPU online operation. So fall through and
7357 * restore the original sched domains by considering the
7358 * cpuset configurations.
7362 cpuset_update_active_cpus(true);
7370 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7373 unsigned long flags;
7374 long cpu = (long)hcpu;
7380 case CPU_DOWN_PREPARE:
7381 rcu_read_lock_sched();
7382 dl_b = dl_bw_of(cpu);
7384 raw_spin_lock_irqsave(&dl_b->lock, flags);
7385 cpus = dl_bw_cpus(cpu);
7386 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7387 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7389 rcu_read_unlock_sched();
7392 return notifier_from_errno(-EBUSY);
7393 cpuset_update_active_cpus(false);
7395 case CPU_DOWN_PREPARE_FROZEN:
7397 partition_sched_domains(1, NULL, NULL);
7405 void __init sched_init_smp(void)
7407 cpumask_var_t non_isolated_cpus;
7409 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7410 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7415 * There's no userspace yet to cause hotplug operations; hence all the
7416 * cpu masks are stable and all blatant races in the below code cannot
7419 mutex_lock(&sched_domains_mutex);
7420 init_sched_domains(cpu_active_mask);
7421 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7422 if (cpumask_empty(non_isolated_cpus))
7423 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7424 mutex_unlock(&sched_domains_mutex);
7426 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7427 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7428 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7432 /* Move init over to a non-isolated CPU */
7433 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7435 sched_init_granularity();
7436 free_cpumask_var(non_isolated_cpus);
7438 init_sched_rt_class();
7439 init_sched_dl_class();
7442 void __init sched_init_smp(void)
7444 sched_init_granularity();
7446 #endif /* CONFIG_SMP */
7448 int in_sched_functions(unsigned long addr)
7450 return in_lock_functions(addr) ||
7451 (addr >= (unsigned long)__sched_text_start
7452 && addr < (unsigned long)__sched_text_end);
7455 #ifdef CONFIG_CGROUP_SCHED
7457 * Default task group.
7458 * Every task in system belongs to this group at bootup.
7460 struct task_group root_task_group;
7461 LIST_HEAD(task_groups);
7463 /* Cacheline aligned slab cache for task_group */
7464 static struct kmem_cache *task_group_cache __read_mostly;
7467 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7469 void __init sched_init(void)
7472 unsigned long alloc_size = 0, ptr;
7474 #ifdef CONFIG_FAIR_GROUP_SCHED
7475 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7477 #ifdef CONFIG_RT_GROUP_SCHED
7478 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7481 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7483 #ifdef CONFIG_FAIR_GROUP_SCHED
7484 root_task_group.se = (struct sched_entity **)ptr;
7485 ptr += nr_cpu_ids * sizeof(void **);
7487 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7488 ptr += nr_cpu_ids * sizeof(void **);
7490 #endif /* CONFIG_FAIR_GROUP_SCHED */
7491 #ifdef CONFIG_RT_GROUP_SCHED
7492 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7493 ptr += nr_cpu_ids * sizeof(void **);
7495 root_task_group.rt_rq = (struct rt_rq **)ptr;
7496 ptr += nr_cpu_ids * sizeof(void **);
7498 #endif /* CONFIG_RT_GROUP_SCHED */
7500 #ifdef CONFIG_CPUMASK_OFFSTACK
7501 for_each_possible_cpu(i) {
7502 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7503 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7505 #endif /* CONFIG_CPUMASK_OFFSTACK */
7507 init_rt_bandwidth(&def_rt_bandwidth,
7508 global_rt_period(), global_rt_runtime());
7509 init_dl_bandwidth(&def_dl_bandwidth,
7510 global_rt_period(), global_rt_runtime());
7513 init_defrootdomain();
7516 #ifdef CONFIG_RT_GROUP_SCHED
7517 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7518 global_rt_period(), global_rt_runtime());
7519 #endif /* CONFIG_RT_GROUP_SCHED */
7521 #ifdef CONFIG_CGROUP_SCHED
7522 task_group_cache = KMEM_CACHE(task_group, 0);
7524 list_add(&root_task_group.list, &task_groups);
7525 INIT_LIST_HEAD(&root_task_group.children);
7526 INIT_LIST_HEAD(&root_task_group.siblings);
7527 autogroup_init(&init_task);
7528 #endif /* CONFIG_CGROUP_SCHED */
7530 for_each_possible_cpu(i) {
7534 raw_spin_lock_init(&rq->lock);
7536 rq->calc_load_active = 0;
7537 rq->calc_load_update = jiffies + LOAD_FREQ;
7538 init_cfs_rq(&rq->cfs);
7539 init_rt_rq(&rq->rt);
7540 init_dl_rq(&rq->dl);
7541 #ifdef CONFIG_FAIR_GROUP_SCHED
7542 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7543 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7545 * How much cpu bandwidth does root_task_group get?
7547 * In case of task-groups formed thr' the cgroup filesystem, it
7548 * gets 100% of the cpu resources in the system. This overall
7549 * system cpu resource is divided among the tasks of
7550 * root_task_group and its child task-groups in a fair manner,
7551 * based on each entity's (task or task-group's) weight
7552 * (se->load.weight).
7554 * In other words, if root_task_group has 10 tasks of weight
7555 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7556 * then A0's share of the cpu resource is:
7558 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7560 * We achieve this by letting root_task_group's tasks sit
7561 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7563 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7564 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7565 #endif /* CONFIG_FAIR_GROUP_SCHED */
7567 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7568 #ifdef CONFIG_RT_GROUP_SCHED
7569 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7572 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7573 rq->cpu_load[j] = 0;
7575 rq->last_load_update_tick = jiffies;
7580 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7581 rq->balance_callback = NULL;
7582 rq->active_balance = 0;
7583 rq->next_balance = jiffies;
7588 rq->avg_idle = 2*sysctl_sched_migration_cost;
7589 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7591 INIT_LIST_HEAD(&rq->cfs_tasks);
7593 rq_attach_root(rq, &def_root_domain);
7594 #ifdef CONFIG_NO_HZ_COMMON
7597 #ifdef CONFIG_NO_HZ_FULL
7598 rq->last_sched_tick = 0;
7602 atomic_set(&rq->nr_iowait, 0);
7605 set_load_weight(&init_task);
7607 #ifdef CONFIG_PREEMPT_NOTIFIERS
7608 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7612 * The boot idle thread does lazy MMU switching as well:
7614 atomic_inc(&init_mm.mm_count);
7615 enter_lazy_tlb(&init_mm, current);
7618 * During early bootup we pretend to be a normal task:
7620 current->sched_class = &fair_sched_class;
7623 * Make us the idle thread. Technically, schedule() should not be
7624 * called from this thread, however somewhere below it might be,
7625 * but because we are the idle thread, we just pick up running again
7626 * when this runqueue becomes "idle".
7628 init_idle(current, smp_processor_id());
7630 calc_load_update = jiffies + LOAD_FREQ;
7633 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7634 /* May be allocated at isolcpus cmdline parse time */
7635 if (cpu_isolated_map == NULL)
7636 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7637 idle_thread_set_boot_cpu();
7638 set_cpu_rq_start_time();
7640 init_sched_fair_class();
7642 scheduler_running = 1;
7645 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7646 static inline int preempt_count_equals(int preempt_offset)
7648 int nested = preempt_count() + rcu_preempt_depth();
7650 return (nested == preempt_offset);
7653 void __might_sleep(const char *file, int line, int preempt_offset)
7656 * Blocking primitives will set (and therefore destroy) current->state,
7657 * since we will exit with TASK_RUNNING make sure we enter with it,
7658 * otherwise we will destroy state.
7660 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7661 "do not call blocking ops when !TASK_RUNNING; "
7662 "state=%lx set at [<%p>] %pS\n",
7664 (void *)current->task_state_change,
7665 (void *)current->task_state_change);
7667 ___might_sleep(file, line, preempt_offset);
7669 EXPORT_SYMBOL(__might_sleep);
7671 void ___might_sleep(const char *file, int line, int preempt_offset)
7673 static unsigned long prev_jiffy; /* ratelimiting */
7675 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7676 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7677 !is_idle_task(current)) ||
7678 system_state != SYSTEM_RUNNING || oops_in_progress)
7680 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7682 prev_jiffy = jiffies;
7685 "BUG: sleeping function called from invalid context at %s:%d\n",
7688 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7689 in_atomic(), irqs_disabled(),
7690 current->pid, current->comm);
7692 if (task_stack_end_corrupted(current))
7693 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7695 debug_show_held_locks(current);
7696 if (irqs_disabled())
7697 print_irqtrace_events(current);
7698 #ifdef CONFIG_DEBUG_PREEMPT
7699 if (!preempt_count_equals(preempt_offset)) {
7700 pr_err("Preemption disabled at:");
7701 print_ip_sym(current->preempt_disable_ip);
7707 EXPORT_SYMBOL(___might_sleep);
7710 #ifdef CONFIG_MAGIC_SYSRQ
7711 void normalize_rt_tasks(void)
7713 struct task_struct *g, *p;
7714 struct sched_attr attr = {
7715 .sched_policy = SCHED_NORMAL,
7718 read_lock(&tasklist_lock);
7719 for_each_process_thread(g, p) {
7721 * Only normalize user tasks:
7723 if (p->flags & PF_KTHREAD)
7726 p->se.exec_start = 0;
7727 #ifdef CONFIG_SCHEDSTATS
7728 p->se.statistics.wait_start = 0;
7729 p->se.statistics.sleep_start = 0;
7730 p->se.statistics.block_start = 0;
7733 if (!dl_task(p) && !rt_task(p)) {
7735 * Renice negative nice level userspace
7738 if (task_nice(p) < 0)
7739 set_user_nice(p, 0);
7743 __sched_setscheduler(p, &attr, false, false);
7745 read_unlock(&tasklist_lock);
7748 #endif /* CONFIG_MAGIC_SYSRQ */
7750 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7752 * These functions are only useful for the IA64 MCA handling, or kdb.
7754 * They can only be called when the whole system has been
7755 * stopped - every CPU needs to be quiescent, and no scheduling
7756 * activity can take place. Using them for anything else would
7757 * be a serious bug, and as a result, they aren't even visible
7758 * under any other configuration.
7762 * curr_task - return the current task for a given cpu.
7763 * @cpu: the processor in question.
7765 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7767 * Return: The current task for @cpu.
7769 struct task_struct *curr_task(int cpu)
7771 return cpu_curr(cpu);
7774 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7778 * set_curr_task - set the current task for a given cpu.
7779 * @cpu: the processor in question.
7780 * @p: the task pointer to set.
7782 * Description: This function must only be used when non-maskable interrupts
7783 * are serviced on a separate stack. It allows the architecture to switch the
7784 * notion of the current task on a cpu in a non-blocking manner. This function
7785 * must be called with all CPU's synchronized, and interrupts disabled, the
7786 * and caller must save the original value of the current task (see
7787 * curr_task() above) and restore that value before reenabling interrupts and
7788 * re-starting the system.
7790 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7792 void set_curr_task(int cpu, struct task_struct *p)
7799 #ifdef CONFIG_CGROUP_SCHED
7800 /* task_group_lock serializes the addition/removal of task groups */
7801 static DEFINE_SPINLOCK(task_group_lock);
7803 static void free_sched_group(struct task_group *tg)
7805 free_fair_sched_group(tg);
7806 free_rt_sched_group(tg);
7808 kmem_cache_free(task_group_cache, tg);
7811 /* allocate runqueue etc for a new task group */
7812 struct task_group *sched_create_group(struct task_group *parent)
7814 struct task_group *tg;
7816 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7818 return ERR_PTR(-ENOMEM);
7820 if (!alloc_fair_sched_group(tg, parent))
7823 if (!alloc_rt_sched_group(tg, parent))
7829 free_sched_group(tg);
7830 return ERR_PTR(-ENOMEM);
7833 void sched_online_group(struct task_group *tg, struct task_group *parent)
7835 unsigned long flags;
7837 spin_lock_irqsave(&task_group_lock, flags);
7838 list_add_rcu(&tg->list, &task_groups);
7840 WARN_ON(!parent); /* root should already exist */
7842 tg->parent = parent;
7843 INIT_LIST_HEAD(&tg->children);
7844 list_add_rcu(&tg->siblings, &parent->children);
7845 spin_unlock_irqrestore(&task_group_lock, flags);
7848 /* rcu callback to free various structures associated with a task group */
7849 static void free_sched_group_rcu(struct rcu_head *rhp)
7851 /* now it should be safe to free those cfs_rqs */
7852 free_sched_group(container_of(rhp, struct task_group, rcu));
7855 /* Destroy runqueue etc associated with a task group */
7856 void sched_destroy_group(struct task_group *tg)
7858 /* wait for possible concurrent references to cfs_rqs complete */
7859 call_rcu(&tg->rcu, free_sched_group_rcu);
7862 void sched_offline_group(struct task_group *tg)
7864 unsigned long flags;
7867 /* end participation in shares distribution */
7868 for_each_possible_cpu(i)
7869 unregister_fair_sched_group(tg, i);
7871 spin_lock_irqsave(&task_group_lock, flags);
7872 list_del_rcu(&tg->list);
7873 list_del_rcu(&tg->siblings);
7874 spin_unlock_irqrestore(&task_group_lock, flags);
7877 /* change task's runqueue when it moves between groups.
7878 * The caller of this function should have put the task in its new group
7879 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7880 * reflect its new group.
7882 void sched_move_task(struct task_struct *tsk)
7884 struct task_group *tg;
7885 int queued, running;
7886 unsigned long flags;
7889 rq = task_rq_lock(tsk, &flags);
7891 running = task_current(rq, tsk);
7892 queued = task_on_rq_queued(tsk);
7895 dequeue_task(rq, tsk, DEQUEUE_SAVE);
7896 if (unlikely(running))
7897 put_prev_task(rq, tsk);
7900 * All callers are synchronized by task_rq_lock(); we do not use RCU
7901 * which is pointless here. Thus, we pass "true" to task_css_check()
7902 * to prevent lockdep warnings.
7904 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7905 struct task_group, css);
7906 tg = autogroup_task_group(tsk, tg);
7907 tsk->sched_task_group = tg;
7909 #ifdef CONFIG_FAIR_GROUP_SCHED
7910 if (tsk->sched_class->task_move_group)
7911 tsk->sched_class->task_move_group(tsk);
7914 set_task_rq(tsk, task_cpu(tsk));
7916 if (unlikely(running))
7917 tsk->sched_class->set_curr_task(rq);
7919 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
7921 task_rq_unlock(rq, tsk, &flags);
7923 #endif /* CONFIG_CGROUP_SCHED */
7925 #ifdef CONFIG_RT_GROUP_SCHED
7927 * Ensure that the real time constraints are schedulable.
7929 static DEFINE_MUTEX(rt_constraints_mutex);
7931 /* Must be called with tasklist_lock held */
7932 static inline int tg_has_rt_tasks(struct task_group *tg)
7934 struct task_struct *g, *p;
7937 * Autogroups do not have RT tasks; see autogroup_create().
7939 if (task_group_is_autogroup(tg))
7942 for_each_process_thread(g, p) {
7943 if (rt_task(p) && task_group(p) == tg)
7950 struct rt_schedulable_data {
7951 struct task_group *tg;
7956 static int tg_rt_schedulable(struct task_group *tg, void *data)
7958 struct rt_schedulable_data *d = data;
7959 struct task_group *child;
7960 unsigned long total, sum = 0;
7961 u64 period, runtime;
7963 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7964 runtime = tg->rt_bandwidth.rt_runtime;
7967 period = d->rt_period;
7968 runtime = d->rt_runtime;
7972 * Cannot have more runtime than the period.
7974 if (runtime > period && runtime != RUNTIME_INF)
7978 * Ensure we don't starve existing RT tasks.
7980 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7983 total = to_ratio(period, runtime);
7986 * Nobody can have more than the global setting allows.
7988 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7992 * The sum of our children's runtime should not exceed our own.
7994 list_for_each_entry_rcu(child, &tg->children, siblings) {
7995 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7996 runtime = child->rt_bandwidth.rt_runtime;
7998 if (child == d->tg) {
7999 period = d->rt_period;
8000 runtime = d->rt_runtime;
8003 sum += to_ratio(period, runtime);
8012 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8016 struct rt_schedulable_data data = {
8018 .rt_period = period,
8019 .rt_runtime = runtime,
8023 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8029 static int tg_set_rt_bandwidth(struct task_group *tg,
8030 u64 rt_period, u64 rt_runtime)
8035 * Disallowing the root group RT runtime is BAD, it would disallow the
8036 * kernel creating (and or operating) RT threads.
8038 if (tg == &root_task_group && rt_runtime == 0)
8041 /* No period doesn't make any sense. */
8045 mutex_lock(&rt_constraints_mutex);
8046 read_lock(&tasklist_lock);
8047 err = __rt_schedulable(tg, rt_period, rt_runtime);
8051 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8052 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8053 tg->rt_bandwidth.rt_runtime = rt_runtime;
8055 for_each_possible_cpu(i) {
8056 struct rt_rq *rt_rq = tg->rt_rq[i];
8058 raw_spin_lock(&rt_rq->rt_runtime_lock);
8059 rt_rq->rt_runtime = rt_runtime;
8060 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8062 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8064 read_unlock(&tasklist_lock);
8065 mutex_unlock(&rt_constraints_mutex);
8070 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8072 u64 rt_runtime, rt_period;
8074 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8075 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8076 if (rt_runtime_us < 0)
8077 rt_runtime = RUNTIME_INF;
8079 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8082 static long sched_group_rt_runtime(struct task_group *tg)
8086 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8089 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8090 do_div(rt_runtime_us, NSEC_PER_USEC);
8091 return rt_runtime_us;
8094 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8096 u64 rt_runtime, rt_period;
8098 rt_period = rt_period_us * NSEC_PER_USEC;
8099 rt_runtime = tg->rt_bandwidth.rt_runtime;
8101 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8104 static long sched_group_rt_period(struct task_group *tg)
8108 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8109 do_div(rt_period_us, NSEC_PER_USEC);
8110 return rt_period_us;
8112 #endif /* CONFIG_RT_GROUP_SCHED */
8114 #ifdef CONFIG_RT_GROUP_SCHED
8115 static int sched_rt_global_constraints(void)
8119 mutex_lock(&rt_constraints_mutex);
8120 read_lock(&tasklist_lock);
8121 ret = __rt_schedulable(NULL, 0, 0);
8122 read_unlock(&tasklist_lock);
8123 mutex_unlock(&rt_constraints_mutex);
8128 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8130 /* Don't accept realtime tasks when there is no way for them to run */
8131 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8137 #else /* !CONFIG_RT_GROUP_SCHED */
8138 static int sched_rt_global_constraints(void)
8140 unsigned long flags;
8143 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8144 for_each_possible_cpu(i) {
8145 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8147 raw_spin_lock(&rt_rq->rt_runtime_lock);
8148 rt_rq->rt_runtime = global_rt_runtime();
8149 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8151 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8155 #endif /* CONFIG_RT_GROUP_SCHED */
8157 static int sched_dl_global_validate(void)
8159 u64 runtime = global_rt_runtime();
8160 u64 period = global_rt_period();
8161 u64 new_bw = to_ratio(period, runtime);
8164 unsigned long flags;
8167 * Here we want to check the bandwidth not being set to some
8168 * value smaller than the currently allocated bandwidth in
8169 * any of the root_domains.
8171 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8172 * cycling on root_domains... Discussion on different/better
8173 * solutions is welcome!
8175 for_each_possible_cpu(cpu) {
8176 rcu_read_lock_sched();
8177 dl_b = dl_bw_of(cpu);
8179 raw_spin_lock_irqsave(&dl_b->lock, flags);
8180 if (new_bw < dl_b->total_bw)
8182 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8184 rcu_read_unlock_sched();
8193 static void sched_dl_do_global(void)
8198 unsigned long flags;
8200 def_dl_bandwidth.dl_period = global_rt_period();
8201 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8203 if (global_rt_runtime() != RUNTIME_INF)
8204 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8207 * FIXME: As above...
8209 for_each_possible_cpu(cpu) {
8210 rcu_read_lock_sched();
8211 dl_b = dl_bw_of(cpu);
8213 raw_spin_lock_irqsave(&dl_b->lock, flags);
8215 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8217 rcu_read_unlock_sched();
8221 static int sched_rt_global_validate(void)
8223 if (sysctl_sched_rt_period <= 0)
8226 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8227 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8233 static void sched_rt_do_global(void)
8235 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8236 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8239 int sched_rt_handler(struct ctl_table *table, int write,
8240 void __user *buffer, size_t *lenp,
8243 int old_period, old_runtime;
8244 static DEFINE_MUTEX(mutex);
8248 old_period = sysctl_sched_rt_period;
8249 old_runtime = sysctl_sched_rt_runtime;
8251 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8253 if (!ret && write) {
8254 ret = sched_rt_global_validate();
8258 ret = sched_dl_global_validate();
8262 ret = sched_rt_global_constraints();
8266 sched_rt_do_global();
8267 sched_dl_do_global();
8271 sysctl_sched_rt_period = old_period;
8272 sysctl_sched_rt_runtime = old_runtime;
8274 mutex_unlock(&mutex);
8279 int sched_rr_handler(struct ctl_table *table, int write,
8280 void __user *buffer, size_t *lenp,
8284 static DEFINE_MUTEX(mutex);
8287 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8288 /* make sure that internally we keep jiffies */
8289 /* also, writing zero resets timeslice to default */
8290 if (!ret && write) {
8291 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8292 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8294 mutex_unlock(&mutex);
8298 #ifdef CONFIG_CGROUP_SCHED
8300 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8302 return css ? container_of(css, struct task_group, css) : NULL;
8305 static struct cgroup_subsys_state *
8306 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8308 struct task_group *parent = css_tg(parent_css);
8309 struct task_group *tg;
8312 /* This is early initialization for the top cgroup */
8313 return &root_task_group.css;
8316 tg = sched_create_group(parent);
8318 return ERR_PTR(-ENOMEM);
8323 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8325 struct task_group *tg = css_tg(css);
8326 struct task_group *parent = css_tg(css->parent);
8329 sched_online_group(tg, parent);
8333 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8335 struct task_group *tg = css_tg(css);
8337 sched_destroy_group(tg);
8340 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8342 struct task_group *tg = css_tg(css);
8344 sched_offline_group(tg);
8347 static void cpu_cgroup_fork(struct task_struct *task)
8349 sched_move_task(task);
8352 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8354 struct task_struct *task;
8355 struct cgroup_subsys_state *css;
8357 cgroup_taskset_for_each(task, css, tset) {
8358 #ifdef CONFIG_RT_GROUP_SCHED
8359 if (!sched_rt_can_attach(css_tg(css), task))
8362 /* We don't support RT-tasks being in separate groups */
8363 if (task->sched_class != &fair_sched_class)
8370 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8372 struct task_struct *task;
8373 struct cgroup_subsys_state *css;
8375 cgroup_taskset_for_each(task, css, tset)
8376 sched_move_task(task);
8379 #ifdef CONFIG_FAIR_GROUP_SCHED
8380 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8381 struct cftype *cftype, u64 shareval)
8383 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8386 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8389 struct task_group *tg = css_tg(css);
8391 return (u64) scale_load_down(tg->shares);
8394 #ifdef CONFIG_CFS_BANDWIDTH
8395 static DEFINE_MUTEX(cfs_constraints_mutex);
8397 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8398 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8400 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8402 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8404 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8405 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8407 if (tg == &root_task_group)
8411 * Ensure we have at some amount of bandwidth every period. This is
8412 * to prevent reaching a state of large arrears when throttled via
8413 * entity_tick() resulting in prolonged exit starvation.
8415 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8419 * Likewise, bound things on the otherside by preventing insane quota
8420 * periods. This also allows us to normalize in computing quota
8423 if (period > max_cfs_quota_period)
8427 * Prevent race between setting of cfs_rq->runtime_enabled and
8428 * unthrottle_offline_cfs_rqs().
8431 mutex_lock(&cfs_constraints_mutex);
8432 ret = __cfs_schedulable(tg, period, quota);
8436 runtime_enabled = quota != RUNTIME_INF;
8437 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8439 * If we need to toggle cfs_bandwidth_used, off->on must occur
8440 * before making related changes, and on->off must occur afterwards
8442 if (runtime_enabled && !runtime_was_enabled)
8443 cfs_bandwidth_usage_inc();
8444 raw_spin_lock_irq(&cfs_b->lock);
8445 cfs_b->period = ns_to_ktime(period);
8446 cfs_b->quota = quota;
8448 __refill_cfs_bandwidth_runtime(cfs_b);
8449 /* restart the period timer (if active) to handle new period expiry */
8450 if (runtime_enabled)
8451 start_cfs_bandwidth(cfs_b);
8452 raw_spin_unlock_irq(&cfs_b->lock);
8454 for_each_online_cpu(i) {
8455 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8456 struct rq *rq = cfs_rq->rq;
8458 raw_spin_lock_irq(&rq->lock);
8459 cfs_rq->runtime_enabled = runtime_enabled;
8460 cfs_rq->runtime_remaining = 0;
8462 if (cfs_rq->throttled)
8463 unthrottle_cfs_rq(cfs_rq);
8464 raw_spin_unlock_irq(&rq->lock);
8466 if (runtime_was_enabled && !runtime_enabled)
8467 cfs_bandwidth_usage_dec();
8469 mutex_unlock(&cfs_constraints_mutex);
8475 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8479 period = ktime_to_ns(tg->cfs_bandwidth.period);
8480 if (cfs_quota_us < 0)
8481 quota = RUNTIME_INF;
8483 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8485 return tg_set_cfs_bandwidth(tg, period, quota);
8488 long tg_get_cfs_quota(struct task_group *tg)
8492 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8495 quota_us = tg->cfs_bandwidth.quota;
8496 do_div(quota_us, NSEC_PER_USEC);
8501 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8505 period = (u64)cfs_period_us * NSEC_PER_USEC;
8506 quota = tg->cfs_bandwidth.quota;
8508 return tg_set_cfs_bandwidth(tg, period, quota);
8511 long tg_get_cfs_period(struct task_group *tg)
8515 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8516 do_div(cfs_period_us, NSEC_PER_USEC);
8518 return cfs_period_us;
8521 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8524 return tg_get_cfs_quota(css_tg(css));
8527 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8528 struct cftype *cftype, s64 cfs_quota_us)
8530 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8533 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8536 return tg_get_cfs_period(css_tg(css));
8539 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8540 struct cftype *cftype, u64 cfs_period_us)
8542 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8545 struct cfs_schedulable_data {
8546 struct task_group *tg;
8551 * normalize group quota/period to be quota/max_period
8552 * note: units are usecs
8554 static u64 normalize_cfs_quota(struct task_group *tg,
8555 struct cfs_schedulable_data *d)
8563 period = tg_get_cfs_period(tg);
8564 quota = tg_get_cfs_quota(tg);
8567 /* note: these should typically be equivalent */
8568 if (quota == RUNTIME_INF || quota == -1)
8571 return to_ratio(period, quota);
8574 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8576 struct cfs_schedulable_data *d = data;
8577 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8578 s64 quota = 0, parent_quota = -1;
8581 quota = RUNTIME_INF;
8583 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8585 quota = normalize_cfs_quota(tg, d);
8586 parent_quota = parent_b->hierarchical_quota;
8589 * ensure max(child_quota) <= parent_quota, inherit when no
8592 if (quota == RUNTIME_INF)
8593 quota = parent_quota;
8594 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8597 cfs_b->hierarchical_quota = quota;
8602 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8605 struct cfs_schedulable_data data = {
8611 if (quota != RUNTIME_INF) {
8612 do_div(data.period, NSEC_PER_USEC);
8613 do_div(data.quota, NSEC_PER_USEC);
8617 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8623 static int cpu_stats_show(struct seq_file *sf, void *v)
8625 struct task_group *tg = css_tg(seq_css(sf));
8626 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8628 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8629 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8630 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8634 #endif /* CONFIG_CFS_BANDWIDTH */
8635 #endif /* CONFIG_FAIR_GROUP_SCHED */
8637 #ifdef CONFIG_RT_GROUP_SCHED
8638 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8639 struct cftype *cft, s64 val)
8641 return sched_group_set_rt_runtime(css_tg(css), val);
8644 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8647 return sched_group_rt_runtime(css_tg(css));
8650 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8651 struct cftype *cftype, u64 rt_period_us)
8653 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8656 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8659 return sched_group_rt_period(css_tg(css));
8661 #endif /* CONFIG_RT_GROUP_SCHED */
8663 static struct cftype cpu_files[] = {
8664 #ifdef CONFIG_FAIR_GROUP_SCHED
8667 .read_u64 = cpu_shares_read_u64,
8668 .write_u64 = cpu_shares_write_u64,
8671 #ifdef CONFIG_CFS_BANDWIDTH
8673 .name = "cfs_quota_us",
8674 .read_s64 = cpu_cfs_quota_read_s64,
8675 .write_s64 = cpu_cfs_quota_write_s64,
8678 .name = "cfs_period_us",
8679 .read_u64 = cpu_cfs_period_read_u64,
8680 .write_u64 = cpu_cfs_period_write_u64,
8684 .seq_show = cpu_stats_show,
8687 #ifdef CONFIG_RT_GROUP_SCHED
8689 .name = "rt_runtime_us",
8690 .read_s64 = cpu_rt_runtime_read,
8691 .write_s64 = cpu_rt_runtime_write,
8694 .name = "rt_period_us",
8695 .read_u64 = cpu_rt_period_read_uint,
8696 .write_u64 = cpu_rt_period_write_uint,
8702 struct cgroup_subsys cpu_cgrp_subsys = {
8703 .css_alloc = cpu_cgroup_css_alloc,
8704 .css_free = cpu_cgroup_css_free,
8705 .css_online = cpu_cgroup_css_online,
8706 .css_offline = cpu_cgroup_css_offline,
8707 .fork = cpu_cgroup_fork,
8708 .can_attach = cpu_cgroup_can_attach,
8709 .attach = cpu_cgroup_attach,
8710 .legacy_cftypes = cpu_files,
8714 #endif /* CONFIG_CGROUP_SCHED */
8716 void dump_cpu_task(int cpu)
8718 pr_info("Task dump for CPU %d:\n", cpu);
8719 sched_show_task(cpu_curr(cpu));
8723 * Nice levels are multiplicative, with a gentle 10% change for every
8724 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8725 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8726 * that remained on nice 0.
8728 * The "10% effect" is relative and cumulative: from _any_ nice level,
8729 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8730 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8731 * If a task goes up by ~10% and another task goes down by ~10% then
8732 * the relative distance between them is ~25%.)
8734 const int sched_prio_to_weight[40] = {
8735 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8736 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8737 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8738 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8739 /* 0 */ 1024, 820, 655, 526, 423,
8740 /* 5 */ 335, 272, 215, 172, 137,
8741 /* 10 */ 110, 87, 70, 56, 45,
8742 /* 15 */ 36, 29, 23, 18, 15,
8746 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8748 * In cases where the weight does not change often, we can use the
8749 * precalculated inverse to speed up arithmetics by turning divisions
8750 * into multiplications:
8752 const u32 sched_prio_to_wmult[40] = {
8753 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8754 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8755 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8756 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8757 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8758 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8759 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8760 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,