4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
96 static void update_rq_clock_task(struct rq *rq, s64 delta);
98 void update_rq_clock(struct rq *rq)
102 lockdep_assert_held(&rq->lock);
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
111 update_rq_clock_task(rq, delta);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug unsigned int sysctl_sched_features =
122 #include "features.h"
127 #ifdef CONFIG_SCHED_DEBUG
128 #define SCHED_FEAT(name, enabled) \
131 static const char * const sched_feat_names[] = {
132 #include "features.h"
137 static int sched_feat_show(struct seq_file *m, void *v)
141 for (i = 0; i < __SCHED_FEAT_NR; i++) {
142 if (!(sysctl_sched_features & (1UL << i)))
144 seq_printf(m, "%s ", sched_feat_names[i]);
151 #ifdef HAVE_JUMP_LABEL
153 #define jump_label_key__true STATIC_KEY_INIT_TRUE
154 #define jump_label_key__false STATIC_KEY_INIT_FALSE
156 #define SCHED_FEAT(name, enabled) \
157 jump_label_key__##enabled ,
159 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
160 #include "features.h"
165 static void sched_feat_disable(int i)
167 if (static_key_enabled(&sched_feat_keys[i]))
168 static_key_slow_dec(&sched_feat_keys[i]);
171 static void sched_feat_enable(int i)
173 if (!static_key_enabled(&sched_feat_keys[i]))
174 static_key_slow_inc(&sched_feat_keys[i]);
177 static void sched_feat_disable(int i) { };
178 static void sched_feat_enable(int i) { };
179 #endif /* HAVE_JUMP_LABEL */
181 static int sched_feat_set(char *cmp)
186 if (strncmp(cmp, "NO_", 3) == 0) {
191 for (i = 0; i < __SCHED_FEAT_NR; i++) {
192 if (strcmp(cmp, sched_feat_names[i]) == 0) {
194 sysctl_sched_features &= ~(1UL << i);
195 sched_feat_disable(i);
197 sysctl_sched_features |= (1UL << i);
198 sched_feat_enable(i);
208 sched_feat_write(struct file *filp, const char __user *ubuf,
209 size_t cnt, loff_t *ppos)
219 if (copy_from_user(&buf, ubuf, cnt))
225 /* Ensure the static_key remains in a consistent state */
226 inode = file_inode(filp);
227 mutex_lock(&inode->i_mutex);
228 i = sched_feat_set(cmp);
229 mutex_unlock(&inode->i_mutex);
230 if (i == __SCHED_FEAT_NR)
238 static int sched_feat_open(struct inode *inode, struct file *filp)
240 return single_open(filp, sched_feat_show, NULL);
243 static const struct file_operations sched_feat_fops = {
244 .open = sched_feat_open,
245 .write = sched_feat_write,
248 .release = single_release,
251 static __init int sched_init_debug(void)
253 debugfs_create_file("sched_features", 0644, NULL, NULL,
258 late_initcall(sched_init_debug);
259 #endif /* CONFIG_SCHED_DEBUG */
262 * Number of tasks to iterate in a single balance run.
263 * Limited because this is done with IRQs disabled.
265 const_debug unsigned int sysctl_sched_nr_migrate = 32;
268 * period over which we average the RT time consumption, measured
273 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
276 * period over which we measure -rt task cpu usage in us.
279 unsigned int sysctl_sched_rt_period = 1000000;
281 __read_mostly int scheduler_running;
284 * part of the period that we allow rt tasks to run in us.
287 int sysctl_sched_rt_runtime = 950000;
289 /* cpus with isolated domains */
290 cpumask_var_t cpu_isolated_map;
293 * this_rq_lock - lock this runqueue and disable interrupts.
295 static struct rq *this_rq_lock(void)
302 raw_spin_lock(&rq->lock);
307 #ifdef CONFIG_SCHED_HRTICK
309 * Use HR-timers to deliver accurate preemption points.
312 static void hrtick_clear(struct rq *rq)
314 if (hrtimer_active(&rq->hrtick_timer))
315 hrtimer_cancel(&rq->hrtick_timer);
319 * High-resolution timer tick.
320 * Runs from hardirq context with interrupts disabled.
322 static enum hrtimer_restart hrtick(struct hrtimer *timer)
324 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
326 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
328 raw_spin_lock(&rq->lock);
330 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
331 raw_spin_unlock(&rq->lock);
333 return HRTIMER_NORESTART;
338 static void __hrtick_restart(struct rq *rq)
340 struct hrtimer *timer = &rq->hrtick_timer;
342 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
346 * called from hardirq (IPI) context
348 static void __hrtick_start(void *arg)
352 raw_spin_lock(&rq->lock);
353 __hrtick_restart(rq);
354 rq->hrtick_csd_pending = 0;
355 raw_spin_unlock(&rq->lock);
359 * Called to set the hrtick timer state.
361 * called with rq->lock held and irqs disabled
363 void hrtick_start(struct rq *rq, u64 delay)
365 struct hrtimer *timer = &rq->hrtick_timer;
370 * Don't schedule slices shorter than 10000ns, that just
371 * doesn't make sense and can cause timer DoS.
373 delta = max_t(s64, delay, 10000LL);
374 time = ktime_add_ns(timer->base->get_time(), delta);
376 hrtimer_set_expires(timer, time);
378 if (rq == this_rq()) {
379 __hrtick_restart(rq);
380 } else if (!rq->hrtick_csd_pending) {
381 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
382 rq->hrtick_csd_pending = 1;
387 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
389 int cpu = (int)(long)hcpu;
392 case CPU_UP_CANCELED:
393 case CPU_UP_CANCELED_FROZEN:
394 case CPU_DOWN_PREPARE:
395 case CPU_DOWN_PREPARE_FROZEN:
397 case CPU_DEAD_FROZEN:
398 hrtick_clear(cpu_rq(cpu));
405 static __init void init_hrtick(void)
407 hotcpu_notifier(hotplug_hrtick, 0);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq *rq, u64 delay)
418 * Don't schedule slices shorter than 10000ns, that just
419 * doesn't make sense. Rely on vruntime for fairness.
421 delay = max_t(u64, delay, 10000LL);
422 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
423 HRTIMER_MODE_REL_PINNED);
426 static inline void init_hrtick(void)
429 #endif /* CONFIG_SMP */
431 static void init_rq_hrtick(struct rq *rq)
434 rq->hrtick_csd_pending = 0;
436 rq->hrtick_csd.flags = 0;
437 rq->hrtick_csd.func = __hrtick_start;
438 rq->hrtick_csd.info = rq;
441 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
442 rq->hrtick_timer.function = hrtick;
444 #else /* CONFIG_SCHED_HRTICK */
445 static inline void hrtick_clear(struct rq *rq)
449 static inline void init_rq_hrtick(struct rq *rq)
453 static inline void init_hrtick(void)
456 #endif /* CONFIG_SCHED_HRTICK */
459 * cmpxchg based fetch_or, macro so it works for different integer types
461 #define fetch_or(ptr, val) \
462 ({ typeof(*(ptr)) __old, __val = *(ptr); \
464 __old = cmpxchg((ptr), __val, __val | (val)); \
465 if (__old == __val) \
472 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
474 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
475 * this avoids any races wrt polling state changes and thereby avoids
478 static bool set_nr_and_not_polling(struct task_struct *p)
480 struct thread_info *ti = task_thread_info(p);
481 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
485 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
487 * If this returns true, then the idle task promises to call
488 * sched_ttwu_pending() and reschedule soon.
490 static bool set_nr_if_polling(struct task_struct *p)
492 struct thread_info *ti = task_thread_info(p);
493 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
496 if (!(val & _TIF_POLLING_NRFLAG))
498 if (val & _TIF_NEED_RESCHED)
500 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
509 static bool set_nr_and_not_polling(struct task_struct *p)
511 set_tsk_need_resched(p);
516 static bool set_nr_if_polling(struct task_struct *p)
523 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
525 struct wake_q_node *node = &task->wake_q;
528 * Atomically grab the task, if ->wake_q is !nil already it means
529 * its already queued (either by us or someone else) and will get the
530 * wakeup due to that.
532 * This cmpxchg() implies a full barrier, which pairs with the write
533 * barrier implied by the wakeup in wake_up_list().
535 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
538 get_task_struct(task);
541 * The head is context local, there can be no concurrency.
544 head->lastp = &node->next;
547 void wake_up_q(struct wake_q_head *head)
549 struct wake_q_node *node = head->first;
551 while (node != WAKE_Q_TAIL) {
552 struct task_struct *task;
554 task = container_of(node, struct task_struct, wake_q);
556 /* task can safely be re-inserted now */
558 task->wake_q.next = NULL;
561 * wake_up_process() implies a wmb() to pair with the queueing
562 * in wake_q_add() so as not to miss wakeups.
564 wake_up_process(task);
565 put_task_struct(task);
570 * resched_curr - mark rq's current task 'to be rescheduled now'.
572 * On UP this means the setting of the need_resched flag, on SMP it
573 * might also involve a cross-CPU call to trigger the scheduler on
576 void resched_curr(struct rq *rq)
578 struct task_struct *curr = rq->curr;
581 lockdep_assert_held(&rq->lock);
583 if (test_tsk_need_resched(curr))
588 if (cpu == smp_processor_id()) {
589 set_tsk_need_resched(curr);
590 set_preempt_need_resched();
594 if (set_nr_and_not_polling(curr))
595 smp_send_reschedule(cpu);
597 trace_sched_wake_idle_without_ipi(cpu);
600 void resched_cpu(int cpu)
602 struct rq *rq = cpu_rq(cpu);
605 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
608 raw_spin_unlock_irqrestore(&rq->lock, flags);
612 #ifdef CONFIG_NO_HZ_COMMON
614 * In the semi idle case, use the nearest busy cpu for migrating timers
615 * from an idle cpu. This is good for power-savings.
617 * We don't do similar optimization for completely idle system, as
618 * selecting an idle cpu will add more delays to the timers than intended
619 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
621 int get_nohz_timer_target(void)
623 int i, cpu = smp_processor_id();
624 struct sched_domain *sd;
630 for_each_domain(cpu, sd) {
631 for_each_cpu(i, sched_domain_span(sd)) {
643 * When add_timer_on() enqueues a timer into the timer wheel of an
644 * idle CPU then this timer might expire before the next timer event
645 * which is scheduled to wake up that CPU. In case of a completely
646 * idle system the next event might even be infinite time into the
647 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
648 * leaves the inner idle loop so the newly added timer is taken into
649 * account when the CPU goes back to idle and evaluates the timer
650 * wheel for the next timer event.
652 static void wake_up_idle_cpu(int cpu)
654 struct rq *rq = cpu_rq(cpu);
656 if (cpu == smp_processor_id())
659 if (set_nr_and_not_polling(rq->idle))
660 smp_send_reschedule(cpu);
662 trace_sched_wake_idle_without_ipi(cpu);
665 static bool wake_up_full_nohz_cpu(int cpu)
668 * We just need the target to call irq_exit() and re-evaluate
669 * the next tick. The nohz full kick at least implies that.
670 * If needed we can still optimize that later with an
673 if (tick_nohz_full_cpu(cpu)) {
674 if (cpu != smp_processor_id() ||
675 tick_nohz_tick_stopped())
676 tick_nohz_full_kick_cpu(cpu);
683 void wake_up_nohz_cpu(int cpu)
685 if (!wake_up_full_nohz_cpu(cpu))
686 wake_up_idle_cpu(cpu);
689 static inline bool got_nohz_idle_kick(void)
691 int cpu = smp_processor_id();
693 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
696 if (idle_cpu(cpu) && !need_resched())
700 * We can't run Idle Load Balance on this CPU for this time so we
701 * cancel it and clear NOHZ_BALANCE_KICK
703 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
707 #else /* CONFIG_NO_HZ_COMMON */
709 static inline bool got_nohz_idle_kick(void)
714 #endif /* CONFIG_NO_HZ_COMMON */
716 #ifdef CONFIG_NO_HZ_FULL
717 bool sched_can_stop_tick(void)
720 * FIFO realtime policy runs the highest priority task. Other runnable
721 * tasks are of a lower priority. The scheduler tick does nothing.
723 if (current->policy == SCHED_FIFO)
727 * Round-robin realtime tasks time slice with other tasks at the same
728 * realtime priority. Is this task the only one at this priority?
730 if (current->policy == SCHED_RR) {
731 struct sched_rt_entity *rt_se = ¤t->rt;
733 return rt_se->run_list.prev == rt_se->run_list.next;
737 * More than one running task need preemption.
738 * nr_running update is assumed to be visible
739 * after IPI is sent from wakers.
741 if (this_rq()->nr_running > 1)
746 #endif /* CONFIG_NO_HZ_FULL */
748 void sched_avg_update(struct rq *rq)
750 s64 period = sched_avg_period();
752 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
754 * Inline assembly required to prevent the compiler
755 * optimising this loop into a divmod call.
756 * See __iter_div_u64_rem() for another example of this.
758 asm("" : "+rm" (rq->age_stamp));
759 rq->age_stamp += period;
764 #endif /* CONFIG_SMP */
766 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
767 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
769 * Iterate task_group tree rooted at *from, calling @down when first entering a
770 * node and @up when leaving it for the final time.
772 * Caller must hold rcu_lock or sufficient equivalent.
774 int walk_tg_tree_from(struct task_group *from,
775 tg_visitor down, tg_visitor up, void *data)
777 struct task_group *parent, *child;
783 ret = (*down)(parent, data);
786 list_for_each_entry_rcu(child, &parent->children, siblings) {
793 ret = (*up)(parent, data);
794 if (ret || parent == from)
798 parent = parent->parent;
805 int tg_nop(struct task_group *tg, void *data)
811 static void set_load_weight(struct task_struct *p)
813 int prio = p->static_prio - MAX_RT_PRIO;
814 struct load_weight *load = &p->se.load;
817 * SCHED_IDLE tasks get minimal weight:
819 if (p->policy == SCHED_IDLE) {
820 load->weight = scale_load(WEIGHT_IDLEPRIO);
821 load->inv_weight = WMULT_IDLEPRIO;
825 load->weight = scale_load(prio_to_weight[prio]);
826 load->inv_weight = prio_to_wmult[prio];
829 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
832 sched_info_queued(rq, p);
833 p->sched_class->enqueue_task(rq, p, flags);
836 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
839 sched_info_dequeued(rq, p);
840 p->sched_class->dequeue_task(rq, p, flags);
843 void activate_task(struct rq *rq, struct task_struct *p, int flags)
845 if (task_contributes_to_load(p))
846 rq->nr_uninterruptible--;
848 enqueue_task(rq, p, flags);
851 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
853 if (task_contributes_to_load(p))
854 rq->nr_uninterruptible++;
856 dequeue_task(rq, p, flags);
859 static void update_rq_clock_task(struct rq *rq, s64 delta)
862 * In theory, the compile should just see 0 here, and optimize out the call
863 * to sched_rt_avg_update. But I don't trust it...
865 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
866 s64 steal = 0, irq_delta = 0;
868 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
869 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
872 * Since irq_time is only updated on {soft,}irq_exit, we might run into
873 * this case when a previous update_rq_clock() happened inside a
876 * When this happens, we stop ->clock_task and only update the
877 * prev_irq_time stamp to account for the part that fit, so that a next
878 * update will consume the rest. This ensures ->clock_task is
881 * It does however cause some slight miss-attribution of {soft,}irq
882 * time, a more accurate solution would be to update the irq_time using
883 * the current rq->clock timestamp, except that would require using
886 if (irq_delta > delta)
889 rq->prev_irq_time += irq_delta;
892 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
893 if (static_key_false((¶virt_steal_rq_enabled))) {
894 steal = paravirt_steal_clock(cpu_of(rq));
895 steal -= rq->prev_steal_time_rq;
897 if (unlikely(steal > delta))
900 rq->prev_steal_time_rq += steal;
905 rq->clock_task += delta;
907 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
908 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
909 sched_rt_avg_update(rq, irq_delta + steal);
913 void sched_set_stop_task(int cpu, struct task_struct *stop)
915 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
916 struct task_struct *old_stop = cpu_rq(cpu)->stop;
920 * Make it appear like a SCHED_FIFO task, its something
921 * userspace knows about and won't get confused about.
923 * Also, it will make PI more or less work without too
924 * much confusion -- but then, stop work should not
925 * rely on PI working anyway.
927 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
929 stop->sched_class = &stop_sched_class;
932 cpu_rq(cpu)->stop = stop;
936 * Reset it back to a normal scheduling class so that
937 * it can die in pieces.
939 old_stop->sched_class = &rt_sched_class;
944 * __normal_prio - return the priority that is based on the static prio
946 static inline int __normal_prio(struct task_struct *p)
948 return p->static_prio;
952 * Calculate the expected normal priority: i.e. priority
953 * without taking RT-inheritance into account. Might be
954 * boosted by interactivity modifiers. Changes upon fork,
955 * setprio syscalls, and whenever the interactivity
956 * estimator recalculates.
958 static inline int normal_prio(struct task_struct *p)
962 if (task_has_dl_policy(p))
963 prio = MAX_DL_PRIO-1;
964 else if (task_has_rt_policy(p))
965 prio = MAX_RT_PRIO-1 - p->rt_priority;
967 prio = __normal_prio(p);
972 * Calculate the current priority, i.e. the priority
973 * taken into account by the scheduler. This value might
974 * be boosted by RT tasks, or might be boosted by
975 * interactivity modifiers. Will be RT if the task got
976 * RT-boosted. If not then it returns p->normal_prio.
978 static int effective_prio(struct task_struct *p)
980 p->normal_prio = normal_prio(p);
982 * If we are RT tasks or we were boosted to RT priority,
983 * keep the priority unchanged. Otherwise, update priority
984 * to the normal priority:
986 if (!rt_prio(p->prio))
987 return p->normal_prio;
992 * task_curr - is this task currently executing on a CPU?
993 * @p: the task in question.
995 * Return: 1 if the task is currently executing. 0 otherwise.
997 inline int task_curr(const struct task_struct *p)
999 return cpu_curr(task_cpu(p)) == p;
1003 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
1005 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1006 const struct sched_class *prev_class,
1009 if (prev_class != p->sched_class) {
1010 if (prev_class->switched_from)
1011 prev_class->switched_from(rq, p);
1012 /* Possble rq->lock 'hole'. */
1013 p->sched_class->switched_to(rq, p);
1014 } else if (oldprio != p->prio || dl_task(p))
1015 p->sched_class->prio_changed(rq, p, oldprio);
1018 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1020 const struct sched_class *class;
1022 if (p->sched_class == rq->curr->sched_class) {
1023 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1025 for_each_class(class) {
1026 if (class == rq->curr->sched_class)
1028 if (class == p->sched_class) {
1036 * A queue event has occurred, and we're going to schedule. In
1037 * this case, we can save a useless back to back clock update.
1039 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1040 rq_clock_skip_update(rq, true);
1044 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1046 #ifdef CONFIG_SCHED_DEBUG
1048 * We should never call set_task_cpu() on a blocked task,
1049 * ttwu() will sort out the placement.
1051 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1054 #ifdef CONFIG_LOCKDEP
1056 * The caller should hold either p->pi_lock or rq->lock, when changing
1057 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1059 * sched_move_task() holds both and thus holding either pins the cgroup,
1062 * Furthermore, all task_rq users should acquire both locks, see
1065 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1066 lockdep_is_held(&task_rq(p)->lock)));
1070 trace_sched_migrate_task(p, new_cpu);
1072 if (task_cpu(p) != new_cpu) {
1073 if (p->sched_class->migrate_task_rq)
1074 p->sched_class->migrate_task_rq(p, new_cpu);
1075 p->se.nr_migrations++;
1076 perf_event_task_migrate(p);
1079 __set_task_cpu(p, new_cpu);
1082 static void __migrate_swap_task(struct task_struct *p, int cpu)
1084 if (task_on_rq_queued(p)) {
1085 struct rq *src_rq, *dst_rq;
1087 src_rq = task_rq(p);
1088 dst_rq = cpu_rq(cpu);
1090 deactivate_task(src_rq, p, 0);
1091 set_task_cpu(p, cpu);
1092 activate_task(dst_rq, p, 0);
1093 check_preempt_curr(dst_rq, p, 0);
1096 * Task isn't running anymore; make it appear like we migrated
1097 * it before it went to sleep. This means on wakeup we make the
1098 * previous cpu our targer instead of where it really is.
1104 struct migration_swap_arg {
1105 struct task_struct *src_task, *dst_task;
1106 int src_cpu, dst_cpu;
1109 static int migrate_swap_stop(void *data)
1111 struct migration_swap_arg *arg = data;
1112 struct rq *src_rq, *dst_rq;
1115 src_rq = cpu_rq(arg->src_cpu);
1116 dst_rq = cpu_rq(arg->dst_cpu);
1118 double_raw_lock(&arg->src_task->pi_lock,
1119 &arg->dst_task->pi_lock);
1120 double_rq_lock(src_rq, dst_rq);
1121 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1124 if (task_cpu(arg->src_task) != arg->src_cpu)
1127 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1130 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1133 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1134 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1139 double_rq_unlock(src_rq, dst_rq);
1140 raw_spin_unlock(&arg->dst_task->pi_lock);
1141 raw_spin_unlock(&arg->src_task->pi_lock);
1147 * Cross migrate two tasks
1149 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1151 struct migration_swap_arg arg;
1154 arg = (struct migration_swap_arg){
1156 .src_cpu = task_cpu(cur),
1158 .dst_cpu = task_cpu(p),
1161 if (arg.src_cpu == arg.dst_cpu)
1165 * These three tests are all lockless; this is OK since all of them
1166 * will be re-checked with proper locks held further down the line.
1168 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1171 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1174 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1177 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1178 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1184 struct migration_arg {
1185 struct task_struct *task;
1189 static int migration_cpu_stop(void *data);
1192 * wait_task_inactive - wait for a thread to unschedule.
1194 * If @match_state is nonzero, it's the @p->state value just checked and
1195 * not expected to change. If it changes, i.e. @p might have woken up,
1196 * then return zero. When we succeed in waiting for @p to be off its CPU,
1197 * we return a positive number (its total switch count). If a second call
1198 * a short while later returns the same number, the caller can be sure that
1199 * @p has remained unscheduled the whole time.
1201 * The caller must ensure that the task *will* unschedule sometime soon,
1202 * else this function might spin for a *long* time. This function can't
1203 * be called with interrupts off, or it may introduce deadlock with
1204 * smp_call_function() if an IPI is sent by the same process we are
1205 * waiting to become inactive.
1207 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1209 unsigned long flags;
1210 int running, queued;
1216 * We do the initial early heuristics without holding
1217 * any task-queue locks at all. We'll only try to get
1218 * the runqueue lock when things look like they will
1224 * If the task is actively running on another CPU
1225 * still, just relax and busy-wait without holding
1228 * NOTE! Since we don't hold any locks, it's not
1229 * even sure that "rq" stays as the right runqueue!
1230 * But we don't care, since "task_running()" will
1231 * return false if the runqueue has changed and p
1232 * is actually now running somewhere else!
1234 while (task_running(rq, p)) {
1235 if (match_state && unlikely(p->state != match_state))
1241 * Ok, time to look more closely! We need the rq
1242 * lock now, to be *sure*. If we're wrong, we'll
1243 * just go back and repeat.
1245 rq = task_rq_lock(p, &flags);
1246 trace_sched_wait_task(p);
1247 running = task_running(rq, p);
1248 queued = task_on_rq_queued(p);
1250 if (!match_state || p->state == match_state)
1251 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1252 task_rq_unlock(rq, p, &flags);
1255 * If it changed from the expected state, bail out now.
1257 if (unlikely(!ncsw))
1261 * Was it really running after all now that we
1262 * checked with the proper locks actually held?
1264 * Oops. Go back and try again..
1266 if (unlikely(running)) {
1272 * It's not enough that it's not actively running,
1273 * it must be off the runqueue _entirely_, and not
1276 * So if it was still runnable (but just not actively
1277 * running right now), it's preempted, and we should
1278 * yield - it could be a while.
1280 if (unlikely(queued)) {
1281 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1283 set_current_state(TASK_UNINTERRUPTIBLE);
1284 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1289 * Ahh, all good. It wasn't running, and it wasn't
1290 * runnable, which means that it will never become
1291 * running in the future either. We're all done!
1300 * kick_process - kick a running thread to enter/exit the kernel
1301 * @p: the to-be-kicked thread
1303 * Cause a process which is running on another CPU to enter
1304 * kernel-mode, without any delay. (to get signals handled.)
1306 * NOTE: this function doesn't have to take the runqueue lock,
1307 * because all it wants to ensure is that the remote task enters
1308 * the kernel. If the IPI races and the task has been migrated
1309 * to another CPU then no harm is done and the purpose has been
1312 void kick_process(struct task_struct *p)
1318 if ((cpu != smp_processor_id()) && task_curr(p))
1319 smp_send_reschedule(cpu);
1322 EXPORT_SYMBOL_GPL(kick_process);
1323 #endif /* CONFIG_SMP */
1327 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1329 static int select_fallback_rq(int cpu, struct task_struct *p)
1331 int nid = cpu_to_node(cpu);
1332 const struct cpumask *nodemask = NULL;
1333 enum { cpuset, possible, fail } state = cpuset;
1337 * If the node that the cpu is on has been offlined, cpu_to_node()
1338 * will return -1. There is no cpu on the node, and we should
1339 * select the cpu on the other node.
1342 nodemask = cpumask_of_node(nid);
1344 /* Look for allowed, online CPU in same node. */
1345 for_each_cpu(dest_cpu, nodemask) {
1346 if (!cpu_online(dest_cpu))
1348 if (!cpu_active(dest_cpu))
1350 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1356 /* Any allowed, online CPU? */
1357 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1358 if (!cpu_online(dest_cpu))
1360 if (!cpu_active(dest_cpu))
1367 /* No more Mr. Nice Guy. */
1368 cpuset_cpus_allowed_fallback(p);
1373 do_set_cpus_allowed(p, cpu_possible_mask);
1384 if (state != cpuset) {
1386 * Don't tell them about moving exiting tasks or
1387 * kernel threads (both mm NULL), since they never
1390 if (p->mm && printk_ratelimit()) {
1391 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1392 task_pid_nr(p), p->comm, cpu);
1400 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1403 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1405 if (p->nr_cpus_allowed > 1)
1406 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1409 * In order not to call set_task_cpu() on a blocking task we need
1410 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1413 * Since this is common to all placement strategies, this lives here.
1415 * [ this allows ->select_task() to simply return task_cpu(p) and
1416 * not worry about this generic constraint ]
1418 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1420 cpu = select_fallback_rq(task_cpu(p), p);
1425 static void update_avg(u64 *avg, u64 sample)
1427 s64 diff = sample - *avg;
1433 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1435 #ifdef CONFIG_SCHEDSTATS
1436 struct rq *rq = this_rq();
1439 int this_cpu = smp_processor_id();
1441 if (cpu == this_cpu) {
1442 schedstat_inc(rq, ttwu_local);
1443 schedstat_inc(p, se.statistics.nr_wakeups_local);
1445 struct sched_domain *sd;
1447 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1449 for_each_domain(this_cpu, sd) {
1450 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1451 schedstat_inc(sd, ttwu_wake_remote);
1458 if (wake_flags & WF_MIGRATED)
1459 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1461 #endif /* CONFIG_SMP */
1463 schedstat_inc(rq, ttwu_count);
1464 schedstat_inc(p, se.statistics.nr_wakeups);
1466 if (wake_flags & WF_SYNC)
1467 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1469 #endif /* CONFIG_SCHEDSTATS */
1472 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1474 activate_task(rq, p, en_flags);
1475 p->on_rq = TASK_ON_RQ_QUEUED;
1477 /* if a worker is waking up, notify workqueue */
1478 if (p->flags & PF_WQ_WORKER)
1479 wq_worker_waking_up(p, cpu_of(rq));
1483 * Mark the task runnable and perform wakeup-preemption.
1486 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1488 check_preempt_curr(rq, p, wake_flags);
1489 trace_sched_wakeup(p, true);
1491 p->state = TASK_RUNNING;
1493 if (p->sched_class->task_woken)
1494 p->sched_class->task_woken(rq, p);
1496 if (rq->idle_stamp) {
1497 u64 delta = rq_clock(rq) - rq->idle_stamp;
1498 u64 max = 2*rq->max_idle_balance_cost;
1500 update_avg(&rq->avg_idle, delta);
1502 if (rq->avg_idle > max)
1511 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1514 if (p->sched_contributes_to_load)
1515 rq->nr_uninterruptible--;
1518 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1519 ttwu_do_wakeup(rq, p, wake_flags);
1523 * Called in case the task @p isn't fully descheduled from its runqueue,
1524 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1525 * since all we need to do is flip p->state to TASK_RUNNING, since
1526 * the task is still ->on_rq.
1528 static int ttwu_remote(struct task_struct *p, int wake_flags)
1533 rq = __task_rq_lock(p);
1534 if (task_on_rq_queued(p)) {
1535 /* check_preempt_curr() may use rq clock */
1536 update_rq_clock(rq);
1537 ttwu_do_wakeup(rq, p, wake_flags);
1540 __task_rq_unlock(rq);
1546 void sched_ttwu_pending(void)
1548 struct rq *rq = this_rq();
1549 struct llist_node *llist = llist_del_all(&rq->wake_list);
1550 struct task_struct *p;
1551 unsigned long flags;
1556 raw_spin_lock_irqsave(&rq->lock, flags);
1559 p = llist_entry(llist, struct task_struct, wake_entry);
1560 llist = llist_next(llist);
1561 ttwu_do_activate(rq, p, 0);
1564 raw_spin_unlock_irqrestore(&rq->lock, flags);
1567 void scheduler_ipi(void)
1570 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1571 * TIF_NEED_RESCHED remotely (for the first time) will also send
1574 preempt_fold_need_resched();
1576 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1580 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1581 * traditionally all their work was done from the interrupt return
1582 * path. Now that we actually do some work, we need to make sure
1585 * Some archs already do call them, luckily irq_enter/exit nest
1588 * Arguably we should visit all archs and update all handlers,
1589 * however a fair share of IPIs are still resched only so this would
1590 * somewhat pessimize the simple resched case.
1593 sched_ttwu_pending();
1596 * Check if someone kicked us for doing the nohz idle load balance.
1598 if (unlikely(got_nohz_idle_kick())) {
1599 this_rq()->idle_balance = 1;
1600 raise_softirq_irqoff(SCHED_SOFTIRQ);
1605 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1607 struct rq *rq = cpu_rq(cpu);
1609 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1610 if (!set_nr_if_polling(rq->idle))
1611 smp_send_reschedule(cpu);
1613 trace_sched_wake_idle_without_ipi(cpu);
1617 void wake_up_if_idle(int cpu)
1619 struct rq *rq = cpu_rq(cpu);
1620 unsigned long flags;
1624 if (!is_idle_task(rcu_dereference(rq->curr)))
1627 if (set_nr_if_polling(rq->idle)) {
1628 trace_sched_wake_idle_without_ipi(cpu);
1630 raw_spin_lock_irqsave(&rq->lock, flags);
1631 if (is_idle_task(rq->curr))
1632 smp_send_reschedule(cpu);
1633 /* Else cpu is not in idle, do nothing here */
1634 raw_spin_unlock_irqrestore(&rq->lock, flags);
1641 bool cpus_share_cache(int this_cpu, int that_cpu)
1643 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1645 #endif /* CONFIG_SMP */
1647 static void ttwu_queue(struct task_struct *p, int cpu)
1649 struct rq *rq = cpu_rq(cpu);
1651 #if defined(CONFIG_SMP)
1652 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1653 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1654 ttwu_queue_remote(p, cpu);
1659 raw_spin_lock(&rq->lock);
1660 ttwu_do_activate(rq, p, 0);
1661 raw_spin_unlock(&rq->lock);
1665 * try_to_wake_up - wake up a thread
1666 * @p: the thread to be awakened
1667 * @state: the mask of task states that can be woken
1668 * @wake_flags: wake modifier flags (WF_*)
1670 * Put it on the run-queue if it's not already there. The "current"
1671 * thread is always on the run-queue (except when the actual
1672 * re-schedule is in progress), and as such you're allowed to do
1673 * the simpler "current->state = TASK_RUNNING" to mark yourself
1674 * runnable without the overhead of this.
1676 * Return: %true if @p was woken up, %false if it was already running.
1677 * or @state didn't match @p's state.
1680 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1682 unsigned long flags;
1683 int cpu, success = 0;
1686 * If we are going to wake up a thread waiting for CONDITION we
1687 * need to ensure that CONDITION=1 done by the caller can not be
1688 * reordered with p->state check below. This pairs with mb() in
1689 * set_current_state() the waiting thread does.
1691 smp_mb__before_spinlock();
1692 raw_spin_lock_irqsave(&p->pi_lock, flags);
1693 if (!(p->state & state))
1696 success = 1; /* we're going to change ->state */
1699 if (p->on_rq && ttwu_remote(p, wake_flags))
1704 * If the owning (remote) cpu is still in the middle of schedule() with
1705 * this task as prev, wait until its done referencing the task.
1710 * Pairs with the smp_wmb() in finish_lock_switch().
1714 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1715 p->state = TASK_WAKING;
1717 if (p->sched_class->task_waking)
1718 p->sched_class->task_waking(p);
1720 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1721 if (task_cpu(p) != cpu) {
1722 wake_flags |= WF_MIGRATED;
1723 set_task_cpu(p, cpu);
1725 #endif /* CONFIG_SMP */
1729 ttwu_stat(p, cpu, wake_flags);
1731 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1737 * try_to_wake_up_local - try to wake up a local task with rq lock held
1738 * @p: the thread to be awakened
1740 * Put @p on the run-queue if it's not already there. The caller must
1741 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1744 static void try_to_wake_up_local(struct task_struct *p)
1746 struct rq *rq = task_rq(p);
1748 if (WARN_ON_ONCE(rq != this_rq()) ||
1749 WARN_ON_ONCE(p == current))
1752 lockdep_assert_held(&rq->lock);
1754 if (!raw_spin_trylock(&p->pi_lock)) {
1755 raw_spin_unlock(&rq->lock);
1756 raw_spin_lock(&p->pi_lock);
1757 raw_spin_lock(&rq->lock);
1760 if (!(p->state & TASK_NORMAL))
1763 if (!task_on_rq_queued(p))
1764 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1766 ttwu_do_wakeup(rq, p, 0);
1767 ttwu_stat(p, smp_processor_id(), 0);
1769 raw_spin_unlock(&p->pi_lock);
1773 * wake_up_process - Wake up a specific process
1774 * @p: The process to be woken up.
1776 * Attempt to wake up the nominated process and move it to the set of runnable
1779 * Return: 1 if the process was woken up, 0 if it was already running.
1781 * It may be assumed that this function implies a write memory barrier before
1782 * changing the task state if and only if any tasks are woken up.
1784 int wake_up_process(struct task_struct *p)
1786 WARN_ON(task_is_stopped_or_traced(p));
1787 return try_to_wake_up(p, TASK_NORMAL, 0);
1789 EXPORT_SYMBOL(wake_up_process);
1791 int wake_up_state(struct task_struct *p, unsigned int state)
1793 return try_to_wake_up(p, state, 0);
1797 * This function clears the sched_dl_entity static params.
1799 void __dl_clear_params(struct task_struct *p)
1801 struct sched_dl_entity *dl_se = &p->dl;
1803 dl_se->dl_runtime = 0;
1804 dl_se->dl_deadline = 0;
1805 dl_se->dl_period = 0;
1809 dl_se->dl_throttled = 0;
1811 dl_se->dl_yielded = 0;
1815 * Perform scheduler related setup for a newly forked process p.
1816 * p is forked by current.
1818 * __sched_fork() is basic setup used by init_idle() too:
1820 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1825 p->se.exec_start = 0;
1826 p->se.sum_exec_runtime = 0;
1827 p->se.prev_sum_exec_runtime = 0;
1828 p->se.nr_migrations = 0;
1831 p->se.avg.decay_count = 0;
1833 INIT_LIST_HEAD(&p->se.group_node);
1835 #ifdef CONFIG_SCHEDSTATS
1836 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1839 RB_CLEAR_NODE(&p->dl.rb_node);
1840 init_dl_task_timer(&p->dl);
1841 __dl_clear_params(p);
1843 INIT_LIST_HEAD(&p->rt.run_list);
1845 #ifdef CONFIG_PREEMPT_NOTIFIERS
1846 INIT_HLIST_HEAD(&p->preempt_notifiers);
1849 #ifdef CONFIG_NUMA_BALANCING
1850 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1851 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1852 p->mm->numa_scan_seq = 0;
1855 if (clone_flags & CLONE_VM)
1856 p->numa_preferred_nid = current->numa_preferred_nid;
1858 p->numa_preferred_nid = -1;
1860 p->node_stamp = 0ULL;
1861 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1862 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1863 p->numa_work.next = &p->numa_work;
1864 p->numa_faults = NULL;
1865 p->last_task_numa_placement = 0;
1866 p->last_sum_exec_runtime = 0;
1868 p->numa_group = NULL;
1869 #endif /* CONFIG_NUMA_BALANCING */
1872 #ifdef CONFIG_NUMA_BALANCING
1873 #ifdef CONFIG_SCHED_DEBUG
1874 void set_numabalancing_state(bool enabled)
1877 sched_feat_set("NUMA");
1879 sched_feat_set("NO_NUMA");
1882 __read_mostly bool numabalancing_enabled;
1884 void set_numabalancing_state(bool enabled)
1886 numabalancing_enabled = enabled;
1888 #endif /* CONFIG_SCHED_DEBUG */
1890 #ifdef CONFIG_PROC_SYSCTL
1891 int sysctl_numa_balancing(struct ctl_table *table, int write,
1892 void __user *buffer, size_t *lenp, loff_t *ppos)
1896 int state = numabalancing_enabled;
1898 if (write && !capable(CAP_SYS_ADMIN))
1903 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1907 set_numabalancing_state(state);
1914 * fork()/clone()-time setup:
1916 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1918 unsigned long flags;
1919 int cpu = get_cpu();
1921 __sched_fork(clone_flags, p);
1923 * We mark the process as running here. This guarantees that
1924 * nobody will actually run it, and a signal or other external
1925 * event cannot wake it up and insert it on the runqueue either.
1927 p->state = TASK_RUNNING;
1930 * Make sure we do not leak PI boosting priority to the child.
1932 p->prio = current->normal_prio;
1935 * Revert to default priority/policy on fork if requested.
1937 if (unlikely(p->sched_reset_on_fork)) {
1938 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1939 p->policy = SCHED_NORMAL;
1940 p->static_prio = NICE_TO_PRIO(0);
1942 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1943 p->static_prio = NICE_TO_PRIO(0);
1945 p->prio = p->normal_prio = __normal_prio(p);
1949 * We don't need the reset flag anymore after the fork. It has
1950 * fulfilled its duty:
1952 p->sched_reset_on_fork = 0;
1955 if (dl_prio(p->prio)) {
1958 } else if (rt_prio(p->prio)) {
1959 p->sched_class = &rt_sched_class;
1961 p->sched_class = &fair_sched_class;
1964 if (p->sched_class->task_fork)
1965 p->sched_class->task_fork(p);
1968 * The child is not yet in the pid-hash so no cgroup attach races,
1969 * and the cgroup is pinned to this child due to cgroup_fork()
1970 * is ran before sched_fork().
1972 * Silence PROVE_RCU.
1974 raw_spin_lock_irqsave(&p->pi_lock, flags);
1975 set_task_cpu(p, cpu);
1976 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1978 #ifdef CONFIG_SCHED_INFO
1979 if (likely(sched_info_on()))
1980 memset(&p->sched_info, 0, sizeof(p->sched_info));
1982 #if defined(CONFIG_SMP)
1985 init_task_preempt_count(p);
1987 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1988 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1995 unsigned long to_ratio(u64 period, u64 runtime)
1997 if (runtime == RUNTIME_INF)
2001 * Doing this here saves a lot of checks in all
2002 * the calling paths, and returning zero seems
2003 * safe for them anyway.
2008 return div64_u64(runtime << 20, period);
2012 inline struct dl_bw *dl_bw_of(int i)
2014 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2015 "sched RCU must be held");
2016 return &cpu_rq(i)->rd->dl_bw;
2019 static inline int dl_bw_cpus(int i)
2021 struct root_domain *rd = cpu_rq(i)->rd;
2024 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2025 "sched RCU must be held");
2026 for_each_cpu_and(i, rd->span, cpu_active_mask)
2032 inline struct dl_bw *dl_bw_of(int i)
2034 return &cpu_rq(i)->dl.dl_bw;
2037 static inline int dl_bw_cpus(int i)
2044 * We must be sure that accepting a new task (or allowing changing the
2045 * parameters of an existing one) is consistent with the bandwidth
2046 * constraints. If yes, this function also accordingly updates the currently
2047 * allocated bandwidth to reflect the new situation.
2049 * This function is called while holding p's rq->lock.
2051 * XXX we should delay bw change until the task's 0-lag point, see
2054 static int dl_overflow(struct task_struct *p, int policy,
2055 const struct sched_attr *attr)
2058 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2059 u64 period = attr->sched_period ?: attr->sched_deadline;
2060 u64 runtime = attr->sched_runtime;
2061 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2064 if (new_bw == p->dl.dl_bw)
2068 * Either if a task, enters, leave, or stays -deadline but changes
2069 * its parameters, we may need to update accordingly the total
2070 * allocated bandwidth of the container.
2072 raw_spin_lock(&dl_b->lock);
2073 cpus = dl_bw_cpus(task_cpu(p));
2074 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2075 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2076 __dl_add(dl_b, new_bw);
2078 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2079 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2080 __dl_clear(dl_b, p->dl.dl_bw);
2081 __dl_add(dl_b, new_bw);
2083 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2084 __dl_clear(dl_b, p->dl.dl_bw);
2087 raw_spin_unlock(&dl_b->lock);
2092 extern void init_dl_bw(struct dl_bw *dl_b);
2095 * wake_up_new_task - wake up a newly created task for the first time.
2097 * This function will do some initial scheduler statistics housekeeping
2098 * that must be done for every newly created context, then puts the task
2099 * on the runqueue and wakes it.
2101 void wake_up_new_task(struct task_struct *p)
2103 unsigned long flags;
2106 raw_spin_lock_irqsave(&p->pi_lock, flags);
2109 * Fork balancing, do it here and not earlier because:
2110 * - cpus_allowed can change in the fork path
2111 * - any previously selected cpu might disappear through hotplug
2113 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2116 /* Initialize new task's runnable average */
2117 init_task_runnable_average(p);
2118 rq = __task_rq_lock(p);
2119 activate_task(rq, p, 0);
2120 p->on_rq = TASK_ON_RQ_QUEUED;
2121 trace_sched_wakeup_new(p, true);
2122 check_preempt_curr(rq, p, WF_FORK);
2124 if (p->sched_class->task_woken)
2125 p->sched_class->task_woken(rq, p);
2127 task_rq_unlock(rq, p, &flags);
2130 #ifdef CONFIG_PREEMPT_NOTIFIERS
2132 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2135 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2136 * @notifier: notifier struct to register
2138 void preempt_notifier_register(struct preempt_notifier *notifier)
2140 static_key_slow_inc(&preempt_notifier_key);
2141 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2143 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2146 * preempt_notifier_unregister - no longer interested in preemption notifications
2147 * @notifier: notifier struct to unregister
2149 * This is *not* safe to call from within a preemption notifier.
2151 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2153 hlist_del(¬ifier->link);
2154 static_key_slow_dec(&preempt_notifier_key);
2156 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2158 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2160 struct preempt_notifier *notifier;
2162 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2163 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2166 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2168 if (static_key_false(&preempt_notifier_key))
2169 __fire_sched_in_preempt_notifiers(curr);
2173 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2174 struct task_struct *next)
2176 struct preempt_notifier *notifier;
2178 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2179 notifier->ops->sched_out(notifier, next);
2182 static __always_inline void
2183 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2184 struct task_struct *next)
2186 if (static_key_false(&preempt_notifier_key))
2187 __fire_sched_out_preempt_notifiers(curr, next);
2190 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2192 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2197 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2198 struct task_struct *next)
2202 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2205 * prepare_task_switch - prepare to switch tasks
2206 * @rq: the runqueue preparing to switch
2207 * @prev: the current task that is being switched out
2208 * @next: the task we are going to switch to.
2210 * This is called with the rq lock held and interrupts off. It must
2211 * be paired with a subsequent finish_task_switch after the context
2214 * prepare_task_switch sets up locking and calls architecture specific
2218 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2219 struct task_struct *next)
2221 trace_sched_switch(prev, next);
2222 sched_info_switch(rq, prev, next);
2223 perf_event_task_sched_out(prev, next);
2224 fire_sched_out_preempt_notifiers(prev, next);
2225 prepare_lock_switch(rq, next);
2226 prepare_arch_switch(next);
2230 * finish_task_switch - clean up after a task-switch
2231 * @prev: the thread we just switched away from.
2233 * finish_task_switch must be called after the context switch, paired
2234 * with a prepare_task_switch call before the context switch.
2235 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2236 * and do any other architecture-specific cleanup actions.
2238 * Note that we may have delayed dropping an mm in context_switch(). If
2239 * so, we finish that here outside of the runqueue lock. (Doing it
2240 * with the lock held can cause deadlocks; see schedule() for
2243 * The context switch have flipped the stack from under us and restored the
2244 * local variables which were saved when this task called schedule() in the
2245 * past. prev == current is still correct but we need to recalculate this_rq
2246 * because prev may have moved to another CPU.
2248 static struct rq *finish_task_switch(struct task_struct *prev)
2249 __releases(rq->lock)
2251 struct rq *rq = this_rq();
2252 struct mm_struct *mm = rq->prev_mm;
2258 * A task struct has one reference for the use as "current".
2259 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2260 * schedule one last time. The schedule call will never return, and
2261 * the scheduled task must drop that reference.
2262 * The test for TASK_DEAD must occur while the runqueue locks are
2263 * still held, otherwise prev could be scheduled on another cpu, die
2264 * there before we look at prev->state, and then the reference would
2266 * Manfred Spraul <manfred@colorfullife.com>
2268 prev_state = prev->state;
2269 vtime_task_switch(prev);
2270 finish_arch_switch(prev);
2271 perf_event_task_sched_in(prev, current);
2272 finish_lock_switch(rq, prev);
2273 finish_arch_post_lock_switch();
2275 fire_sched_in_preempt_notifiers(current);
2278 if (unlikely(prev_state == TASK_DEAD)) {
2279 if (prev->sched_class->task_dead)
2280 prev->sched_class->task_dead(prev);
2283 * Remove function-return probe instances associated with this
2284 * task and put them back on the free list.
2286 kprobe_flush_task(prev);
2287 put_task_struct(prev);
2290 tick_nohz_task_switch(current);
2296 /* rq->lock is NOT held, but preemption is disabled */
2297 static inline void post_schedule(struct rq *rq)
2299 if (rq->post_schedule) {
2300 unsigned long flags;
2302 raw_spin_lock_irqsave(&rq->lock, flags);
2303 if (rq->curr->sched_class->post_schedule)
2304 rq->curr->sched_class->post_schedule(rq);
2305 raw_spin_unlock_irqrestore(&rq->lock, flags);
2307 rq->post_schedule = 0;
2313 static inline void post_schedule(struct rq *rq)
2320 * schedule_tail - first thing a freshly forked thread must call.
2321 * @prev: the thread we just switched away from.
2323 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2324 __releases(rq->lock)
2328 /* finish_task_switch() drops rq->lock and enables preemtion */
2330 rq = finish_task_switch(prev);
2334 if (current->set_child_tid)
2335 put_user(task_pid_vnr(current), current->set_child_tid);
2339 * context_switch - switch to the new MM and the new thread's register state.
2341 static inline struct rq *
2342 context_switch(struct rq *rq, struct task_struct *prev,
2343 struct task_struct *next)
2345 struct mm_struct *mm, *oldmm;
2347 prepare_task_switch(rq, prev, next);
2350 oldmm = prev->active_mm;
2352 * For paravirt, this is coupled with an exit in switch_to to
2353 * combine the page table reload and the switch backend into
2356 arch_start_context_switch(prev);
2359 next->active_mm = oldmm;
2360 atomic_inc(&oldmm->mm_count);
2361 enter_lazy_tlb(oldmm, next);
2363 switch_mm(oldmm, mm, next);
2366 prev->active_mm = NULL;
2367 rq->prev_mm = oldmm;
2370 * Since the runqueue lock will be released by the next
2371 * task (which is an invalid locking op but in the case
2372 * of the scheduler it's an obvious special-case), so we
2373 * do an early lockdep release here:
2375 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2377 /* Here we just switch the register state and the stack. */
2378 switch_to(prev, next, prev);
2381 return finish_task_switch(prev);
2385 * nr_running and nr_context_switches:
2387 * externally visible scheduler statistics: current number of runnable
2388 * threads, total number of context switches performed since bootup.
2390 unsigned long nr_running(void)
2392 unsigned long i, sum = 0;
2394 for_each_online_cpu(i)
2395 sum += cpu_rq(i)->nr_running;
2401 * Check if only the current task is running on the cpu.
2403 bool single_task_running(void)
2405 if (cpu_rq(smp_processor_id())->nr_running == 1)
2410 EXPORT_SYMBOL(single_task_running);
2412 unsigned long long nr_context_switches(void)
2415 unsigned long long sum = 0;
2417 for_each_possible_cpu(i)
2418 sum += cpu_rq(i)->nr_switches;
2423 unsigned long nr_iowait(void)
2425 unsigned long i, sum = 0;
2427 for_each_possible_cpu(i)
2428 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2433 unsigned long nr_iowait_cpu(int cpu)
2435 struct rq *this = cpu_rq(cpu);
2436 return atomic_read(&this->nr_iowait);
2439 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2441 struct rq *rq = this_rq();
2442 *nr_waiters = atomic_read(&rq->nr_iowait);
2443 *load = rq->load.weight;
2449 * sched_exec - execve() is a valuable balancing opportunity, because at
2450 * this point the task has the smallest effective memory and cache footprint.
2452 void sched_exec(void)
2454 struct task_struct *p = current;
2455 unsigned long flags;
2458 raw_spin_lock_irqsave(&p->pi_lock, flags);
2459 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2460 if (dest_cpu == smp_processor_id())
2463 if (likely(cpu_active(dest_cpu))) {
2464 struct migration_arg arg = { p, dest_cpu };
2466 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2467 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2471 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2476 DEFINE_PER_CPU(struct kernel_stat, kstat);
2477 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2479 EXPORT_PER_CPU_SYMBOL(kstat);
2480 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2483 * Return accounted runtime for the task.
2484 * In case the task is currently running, return the runtime plus current's
2485 * pending runtime that have not been accounted yet.
2487 unsigned long long task_sched_runtime(struct task_struct *p)
2489 unsigned long flags;
2493 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2495 * 64-bit doesn't need locks to atomically read a 64bit value.
2496 * So we have a optimization chance when the task's delta_exec is 0.
2497 * Reading ->on_cpu is racy, but this is ok.
2499 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2500 * If we race with it entering cpu, unaccounted time is 0. This is
2501 * indistinguishable from the read occurring a few cycles earlier.
2502 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2503 * been accounted, so we're correct here as well.
2505 if (!p->on_cpu || !task_on_rq_queued(p))
2506 return p->se.sum_exec_runtime;
2509 rq = task_rq_lock(p, &flags);
2511 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2512 * project cycles that may never be accounted to this
2513 * thread, breaking clock_gettime().
2515 if (task_current(rq, p) && task_on_rq_queued(p)) {
2516 update_rq_clock(rq);
2517 p->sched_class->update_curr(rq);
2519 ns = p->se.sum_exec_runtime;
2520 task_rq_unlock(rq, p, &flags);
2526 * This function gets called by the timer code, with HZ frequency.
2527 * We call it with interrupts disabled.
2529 void scheduler_tick(void)
2531 int cpu = smp_processor_id();
2532 struct rq *rq = cpu_rq(cpu);
2533 struct task_struct *curr = rq->curr;
2537 raw_spin_lock(&rq->lock);
2538 update_rq_clock(rq);
2539 curr->sched_class->task_tick(rq, curr, 0);
2540 update_cpu_load_active(rq);
2541 calc_global_load_tick(rq);
2542 raw_spin_unlock(&rq->lock);
2544 perf_event_task_tick();
2547 rq->idle_balance = idle_cpu(cpu);
2548 trigger_load_balance(rq);
2550 rq_last_tick_reset(rq);
2553 #ifdef CONFIG_NO_HZ_FULL
2555 * scheduler_tick_max_deferment
2557 * Keep at least one tick per second when a single
2558 * active task is running because the scheduler doesn't
2559 * yet completely support full dynticks environment.
2561 * This makes sure that uptime, CFS vruntime, load
2562 * balancing, etc... continue to move forward, even
2563 * with a very low granularity.
2565 * Return: Maximum deferment in nanoseconds.
2567 u64 scheduler_tick_max_deferment(void)
2569 struct rq *rq = this_rq();
2570 unsigned long next, now = READ_ONCE(jiffies);
2572 next = rq->last_sched_tick + HZ;
2574 if (time_before_eq(next, now))
2577 return jiffies_to_nsecs(next - now);
2581 notrace unsigned long get_parent_ip(unsigned long addr)
2583 if (in_lock_functions(addr)) {
2584 addr = CALLER_ADDR2;
2585 if (in_lock_functions(addr))
2586 addr = CALLER_ADDR3;
2591 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2592 defined(CONFIG_PREEMPT_TRACER))
2594 void preempt_count_add(int val)
2596 #ifdef CONFIG_DEBUG_PREEMPT
2600 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2603 __preempt_count_add(val);
2604 #ifdef CONFIG_DEBUG_PREEMPT
2606 * Spinlock count overflowing soon?
2608 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2611 if (preempt_count() == val) {
2612 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2613 #ifdef CONFIG_DEBUG_PREEMPT
2614 current->preempt_disable_ip = ip;
2616 trace_preempt_off(CALLER_ADDR0, ip);
2619 EXPORT_SYMBOL(preempt_count_add);
2620 NOKPROBE_SYMBOL(preempt_count_add);
2622 void preempt_count_sub(int val)
2624 #ifdef CONFIG_DEBUG_PREEMPT
2628 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2631 * Is the spinlock portion underflowing?
2633 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2634 !(preempt_count() & PREEMPT_MASK)))
2638 if (preempt_count() == val)
2639 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2640 __preempt_count_sub(val);
2642 EXPORT_SYMBOL(preempt_count_sub);
2643 NOKPROBE_SYMBOL(preempt_count_sub);
2648 * Print scheduling while atomic bug:
2650 static noinline void __schedule_bug(struct task_struct *prev)
2652 if (oops_in_progress)
2655 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2656 prev->comm, prev->pid, preempt_count());
2658 debug_show_held_locks(prev);
2660 if (irqs_disabled())
2661 print_irqtrace_events(prev);
2662 #ifdef CONFIG_DEBUG_PREEMPT
2663 if (in_atomic_preempt_off()) {
2664 pr_err("Preemption disabled at:");
2665 print_ip_sym(current->preempt_disable_ip);
2670 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2674 * Various schedule()-time debugging checks and statistics:
2676 static inline void schedule_debug(struct task_struct *prev)
2678 #ifdef CONFIG_SCHED_STACK_END_CHECK
2679 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2682 * Test if we are atomic. Since do_exit() needs to call into
2683 * schedule() atomically, we ignore that path. Otherwise whine
2684 * if we are scheduling when we should not.
2686 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2687 __schedule_bug(prev);
2690 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2692 schedstat_inc(this_rq(), sched_count);
2696 * Pick up the highest-prio task:
2698 static inline struct task_struct *
2699 pick_next_task(struct rq *rq, struct task_struct *prev)
2701 const struct sched_class *class = &fair_sched_class;
2702 struct task_struct *p;
2705 * Optimization: we know that if all tasks are in
2706 * the fair class we can call that function directly:
2708 if (likely(prev->sched_class == class &&
2709 rq->nr_running == rq->cfs.h_nr_running)) {
2710 p = fair_sched_class.pick_next_task(rq, prev);
2711 if (unlikely(p == RETRY_TASK))
2714 /* assumes fair_sched_class->next == idle_sched_class */
2716 p = idle_sched_class.pick_next_task(rq, prev);
2722 for_each_class(class) {
2723 p = class->pick_next_task(rq, prev);
2725 if (unlikely(p == RETRY_TASK))
2731 BUG(); /* the idle class will always have a runnable task */
2735 * __schedule() is the main scheduler function.
2737 * The main means of driving the scheduler and thus entering this function are:
2739 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2741 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2742 * paths. For example, see arch/x86/entry_64.S.
2744 * To drive preemption between tasks, the scheduler sets the flag in timer
2745 * interrupt handler scheduler_tick().
2747 * 3. Wakeups don't really cause entry into schedule(). They add a
2748 * task to the run-queue and that's it.
2750 * Now, if the new task added to the run-queue preempts the current
2751 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2752 * called on the nearest possible occasion:
2754 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2756 * - in syscall or exception context, at the next outmost
2757 * preempt_enable(). (this might be as soon as the wake_up()'s
2760 * - in IRQ context, return from interrupt-handler to
2761 * preemptible context
2763 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2766 * - cond_resched() call
2767 * - explicit schedule() call
2768 * - return from syscall or exception to user-space
2769 * - return from interrupt-handler to user-space
2771 * WARNING: must be called with preemption disabled!
2773 static void __sched __schedule(void)
2775 struct task_struct *prev, *next;
2776 unsigned long *switch_count;
2780 cpu = smp_processor_id();
2782 rcu_note_context_switch();
2785 schedule_debug(prev);
2787 if (sched_feat(HRTICK))
2791 * Make sure that signal_pending_state()->signal_pending() below
2792 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2793 * done by the caller to avoid the race with signal_wake_up().
2795 smp_mb__before_spinlock();
2796 raw_spin_lock_irq(&rq->lock);
2798 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2800 switch_count = &prev->nivcsw;
2801 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2802 if (unlikely(signal_pending_state(prev->state, prev))) {
2803 prev->state = TASK_RUNNING;
2805 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2809 * If a worker went to sleep, notify and ask workqueue
2810 * whether it wants to wake up a task to maintain
2813 if (prev->flags & PF_WQ_WORKER) {
2814 struct task_struct *to_wakeup;
2816 to_wakeup = wq_worker_sleeping(prev, cpu);
2818 try_to_wake_up_local(to_wakeup);
2821 switch_count = &prev->nvcsw;
2824 if (task_on_rq_queued(prev))
2825 update_rq_clock(rq);
2827 next = pick_next_task(rq, prev);
2828 clear_tsk_need_resched(prev);
2829 clear_preempt_need_resched();
2830 rq->clock_skip_update = 0;
2832 if (likely(prev != next)) {
2837 rq = context_switch(rq, prev, next); /* unlocks the rq */
2840 raw_spin_unlock_irq(&rq->lock);
2845 static inline void sched_submit_work(struct task_struct *tsk)
2847 if (!tsk->state || tsk_is_pi_blocked(tsk))
2850 * If we are going to sleep and we have plugged IO queued,
2851 * make sure to submit it to avoid deadlocks.
2853 if (blk_needs_flush_plug(tsk))
2854 blk_schedule_flush_plug(tsk);
2857 asmlinkage __visible void __sched schedule(void)
2859 struct task_struct *tsk = current;
2861 sched_submit_work(tsk);
2865 sched_preempt_enable_no_resched();
2866 } while (need_resched());
2868 EXPORT_SYMBOL(schedule);
2870 #ifdef CONFIG_CONTEXT_TRACKING
2871 asmlinkage __visible void __sched schedule_user(void)
2874 * If we come here after a random call to set_need_resched(),
2875 * or we have been woken up remotely but the IPI has not yet arrived,
2876 * we haven't yet exited the RCU idle mode. Do it here manually until
2877 * we find a better solution.
2879 * NB: There are buggy callers of this function. Ideally we
2880 * should warn if prev_state != CONTEXT_USER, but that will trigger
2881 * too frequently to make sense yet.
2883 enum ctx_state prev_state = exception_enter();
2885 exception_exit(prev_state);
2890 * schedule_preempt_disabled - called with preemption disabled
2892 * Returns with preemption disabled. Note: preempt_count must be 1
2894 void __sched schedule_preempt_disabled(void)
2896 sched_preempt_enable_no_resched();
2901 static void __sched notrace preempt_schedule_common(void)
2904 preempt_active_enter();
2906 preempt_active_exit();
2909 * Check again in case we missed a preemption opportunity
2910 * between schedule and now.
2912 } while (need_resched());
2915 #ifdef CONFIG_PREEMPT
2917 * this is the entry point to schedule() from in-kernel preemption
2918 * off of preempt_enable. Kernel preemptions off return from interrupt
2919 * occur there and call schedule directly.
2921 asmlinkage __visible void __sched notrace preempt_schedule(void)
2924 * If there is a non-zero preempt_count or interrupts are disabled,
2925 * we do not want to preempt the current task. Just return..
2927 if (likely(!preemptible()))
2930 preempt_schedule_common();
2932 NOKPROBE_SYMBOL(preempt_schedule);
2933 EXPORT_SYMBOL(preempt_schedule);
2936 * preempt_schedule_notrace - preempt_schedule called by tracing
2938 * The tracing infrastructure uses preempt_enable_notrace to prevent
2939 * recursion and tracing preempt enabling caused by the tracing
2940 * infrastructure itself. But as tracing can happen in areas coming
2941 * from userspace or just about to enter userspace, a preempt enable
2942 * can occur before user_exit() is called. This will cause the scheduler
2943 * to be called when the system is still in usermode.
2945 * To prevent this, the preempt_enable_notrace will use this function
2946 * instead of preempt_schedule() to exit user context if needed before
2947 * calling the scheduler.
2949 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
2951 enum ctx_state prev_ctx;
2953 if (likely(!preemptible()))
2958 * Use raw __prempt_count() ops that don't call function.
2959 * We can't call functions before disabling preemption which
2960 * disarm preemption tracing recursions.
2962 __preempt_count_add(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
2965 * Needs preempt disabled in case user_exit() is traced
2966 * and the tracer calls preempt_enable_notrace() causing
2967 * an infinite recursion.
2969 prev_ctx = exception_enter();
2971 exception_exit(prev_ctx);
2974 __preempt_count_sub(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
2975 } while (need_resched());
2977 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
2979 #endif /* CONFIG_PREEMPT */
2982 * this is the entry point to schedule() from kernel preemption
2983 * off of irq context.
2984 * Note, that this is called and return with irqs disabled. This will
2985 * protect us against recursive calling from irq.
2987 asmlinkage __visible void __sched preempt_schedule_irq(void)
2989 enum ctx_state prev_state;
2991 /* Catch callers which need to be fixed */
2992 BUG_ON(preempt_count() || !irqs_disabled());
2994 prev_state = exception_enter();
2997 preempt_active_enter();
3000 local_irq_disable();
3001 preempt_active_exit();
3002 } while (need_resched());
3004 exception_exit(prev_state);
3007 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3010 return try_to_wake_up(curr->private, mode, wake_flags);
3012 EXPORT_SYMBOL(default_wake_function);
3014 #ifdef CONFIG_RT_MUTEXES
3017 * rt_mutex_setprio - set the current priority of a task
3019 * @prio: prio value (kernel-internal form)
3021 * This function changes the 'effective' priority of a task. It does
3022 * not touch ->normal_prio like __setscheduler().
3024 * Used by the rt_mutex code to implement priority inheritance
3025 * logic. Call site only calls if the priority of the task changed.
3027 void rt_mutex_setprio(struct task_struct *p, int prio)
3029 int oldprio, queued, running, enqueue_flag = 0;
3031 const struct sched_class *prev_class;
3033 BUG_ON(prio > MAX_PRIO);
3035 rq = __task_rq_lock(p);
3038 * Idle task boosting is a nono in general. There is one
3039 * exception, when PREEMPT_RT and NOHZ is active:
3041 * The idle task calls get_next_timer_interrupt() and holds
3042 * the timer wheel base->lock on the CPU and another CPU wants
3043 * to access the timer (probably to cancel it). We can safely
3044 * ignore the boosting request, as the idle CPU runs this code
3045 * with interrupts disabled and will complete the lock
3046 * protected section without being interrupted. So there is no
3047 * real need to boost.
3049 if (unlikely(p == rq->idle)) {
3050 WARN_ON(p != rq->curr);
3051 WARN_ON(p->pi_blocked_on);
3055 trace_sched_pi_setprio(p, prio);
3057 prev_class = p->sched_class;
3058 queued = task_on_rq_queued(p);
3059 running = task_current(rq, p);
3061 dequeue_task(rq, p, 0);
3063 put_prev_task(rq, p);
3066 * Boosting condition are:
3067 * 1. -rt task is running and holds mutex A
3068 * --> -dl task blocks on mutex A
3070 * 2. -dl task is running and holds mutex A
3071 * --> -dl task blocks on mutex A and could preempt the
3074 if (dl_prio(prio)) {
3075 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3076 if (!dl_prio(p->normal_prio) ||
3077 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3078 p->dl.dl_boosted = 1;
3079 enqueue_flag = ENQUEUE_REPLENISH;
3081 p->dl.dl_boosted = 0;
3082 p->sched_class = &dl_sched_class;
3083 } else if (rt_prio(prio)) {
3084 if (dl_prio(oldprio))
3085 p->dl.dl_boosted = 0;
3087 enqueue_flag = ENQUEUE_HEAD;
3088 p->sched_class = &rt_sched_class;
3090 if (dl_prio(oldprio))
3091 p->dl.dl_boosted = 0;
3092 if (rt_prio(oldprio))
3094 p->sched_class = &fair_sched_class;
3100 p->sched_class->set_curr_task(rq);
3102 enqueue_task(rq, p, enqueue_flag);
3104 check_class_changed(rq, p, prev_class, oldprio);
3106 __task_rq_unlock(rq);
3110 void set_user_nice(struct task_struct *p, long nice)
3112 int old_prio, delta, queued;
3113 unsigned long flags;
3116 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3119 * We have to be careful, if called from sys_setpriority(),
3120 * the task might be in the middle of scheduling on another CPU.
3122 rq = task_rq_lock(p, &flags);
3124 * The RT priorities are set via sched_setscheduler(), but we still
3125 * allow the 'normal' nice value to be set - but as expected
3126 * it wont have any effect on scheduling until the task is
3127 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3129 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3130 p->static_prio = NICE_TO_PRIO(nice);
3133 queued = task_on_rq_queued(p);
3135 dequeue_task(rq, p, 0);
3137 p->static_prio = NICE_TO_PRIO(nice);
3140 p->prio = effective_prio(p);
3141 delta = p->prio - old_prio;
3144 enqueue_task(rq, p, 0);
3146 * If the task increased its priority or is running and
3147 * lowered its priority, then reschedule its CPU:
3149 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3153 task_rq_unlock(rq, p, &flags);
3155 EXPORT_SYMBOL(set_user_nice);
3158 * can_nice - check if a task can reduce its nice value
3162 int can_nice(const struct task_struct *p, const int nice)
3164 /* convert nice value [19,-20] to rlimit style value [1,40] */
3165 int nice_rlim = nice_to_rlimit(nice);
3167 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3168 capable(CAP_SYS_NICE));
3171 #ifdef __ARCH_WANT_SYS_NICE
3174 * sys_nice - change the priority of the current process.
3175 * @increment: priority increment
3177 * sys_setpriority is a more generic, but much slower function that
3178 * does similar things.
3180 SYSCALL_DEFINE1(nice, int, increment)
3185 * Setpriority might change our priority at the same moment.
3186 * We don't have to worry. Conceptually one call occurs first
3187 * and we have a single winner.
3189 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3190 nice = task_nice(current) + increment;
3192 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3193 if (increment < 0 && !can_nice(current, nice))
3196 retval = security_task_setnice(current, nice);
3200 set_user_nice(current, nice);
3207 * task_prio - return the priority value of a given task.
3208 * @p: the task in question.
3210 * Return: The priority value as seen by users in /proc.
3211 * RT tasks are offset by -200. Normal tasks are centered
3212 * around 0, value goes from -16 to +15.
3214 int task_prio(const struct task_struct *p)
3216 return p->prio - MAX_RT_PRIO;
3220 * idle_cpu - is a given cpu idle currently?
3221 * @cpu: the processor in question.
3223 * Return: 1 if the CPU is currently idle. 0 otherwise.
3225 int idle_cpu(int cpu)
3227 struct rq *rq = cpu_rq(cpu);
3229 if (rq->curr != rq->idle)
3236 if (!llist_empty(&rq->wake_list))
3244 * idle_task - return the idle task for a given cpu.
3245 * @cpu: the processor in question.
3247 * Return: The idle task for the cpu @cpu.
3249 struct task_struct *idle_task(int cpu)
3251 return cpu_rq(cpu)->idle;
3255 * find_process_by_pid - find a process with a matching PID value.
3256 * @pid: the pid in question.
3258 * The task of @pid, if found. %NULL otherwise.
3260 static struct task_struct *find_process_by_pid(pid_t pid)
3262 return pid ? find_task_by_vpid(pid) : current;
3266 * This function initializes the sched_dl_entity of a newly becoming
3267 * SCHED_DEADLINE task.
3269 * Only the static values are considered here, the actual runtime and the
3270 * absolute deadline will be properly calculated when the task is enqueued
3271 * for the first time with its new policy.
3274 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3276 struct sched_dl_entity *dl_se = &p->dl;
3278 dl_se->dl_runtime = attr->sched_runtime;
3279 dl_se->dl_deadline = attr->sched_deadline;
3280 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3281 dl_se->flags = attr->sched_flags;
3282 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3285 * Changing the parameters of a task is 'tricky' and we're not doing
3286 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3288 * What we SHOULD do is delay the bandwidth release until the 0-lag
3289 * point. This would include retaining the task_struct until that time
3290 * and change dl_overflow() to not immediately decrement the current
3293 * Instead we retain the current runtime/deadline and let the new
3294 * parameters take effect after the current reservation period lapses.
3295 * This is safe (albeit pessimistic) because the 0-lag point is always
3296 * before the current scheduling deadline.
3298 * We can still have temporary overloads because we do not delay the
3299 * change in bandwidth until that time; so admission control is
3300 * not on the safe side. It does however guarantee tasks will never
3301 * consume more than promised.
3306 * sched_setparam() passes in -1 for its policy, to let the functions
3307 * it calls know not to change it.
3309 #define SETPARAM_POLICY -1
3311 static void __setscheduler_params(struct task_struct *p,
3312 const struct sched_attr *attr)
3314 int policy = attr->sched_policy;
3316 if (policy == SETPARAM_POLICY)
3321 if (dl_policy(policy))
3322 __setparam_dl(p, attr);
3323 else if (fair_policy(policy))
3324 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3327 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3328 * !rt_policy. Always setting this ensures that things like
3329 * getparam()/getattr() don't report silly values for !rt tasks.
3331 p->rt_priority = attr->sched_priority;
3332 p->normal_prio = normal_prio(p);
3336 /* Actually do priority change: must hold pi & rq lock. */
3337 static void __setscheduler(struct rq *rq, struct task_struct *p,
3338 const struct sched_attr *attr, bool keep_boost)
3340 __setscheduler_params(p, attr);
3343 * Keep a potential priority boosting if called from
3344 * sched_setscheduler().
3347 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3349 p->prio = normal_prio(p);
3351 if (dl_prio(p->prio))
3352 p->sched_class = &dl_sched_class;
3353 else if (rt_prio(p->prio))
3354 p->sched_class = &rt_sched_class;
3356 p->sched_class = &fair_sched_class;
3360 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3362 struct sched_dl_entity *dl_se = &p->dl;
3364 attr->sched_priority = p->rt_priority;
3365 attr->sched_runtime = dl_se->dl_runtime;
3366 attr->sched_deadline = dl_se->dl_deadline;
3367 attr->sched_period = dl_se->dl_period;
3368 attr->sched_flags = dl_se->flags;
3372 * This function validates the new parameters of a -deadline task.
3373 * We ask for the deadline not being zero, and greater or equal
3374 * than the runtime, as well as the period of being zero or
3375 * greater than deadline. Furthermore, we have to be sure that
3376 * user parameters are above the internal resolution of 1us (we
3377 * check sched_runtime only since it is always the smaller one) and
3378 * below 2^63 ns (we have to check both sched_deadline and
3379 * sched_period, as the latter can be zero).
3382 __checkparam_dl(const struct sched_attr *attr)
3385 if (attr->sched_deadline == 0)
3389 * Since we truncate DL_SCALE bits, make sure we're at least
3392 if (attr->sched_runtime < (1ULL << DL_SCALE))
3396 * Since we use the MSB for wrap-around and sign issues, make
3397 * sure it's not set (mind that period can be equal to zero).
3399 if (attr->sched_deadline & (1ULL << 63) ||
3400 attr->sched_period & (1ULL << 63))
3403 /* runtime <= deadline <= period (if period != 0) */
3404 if ((attr->sched_period != 0 &&
3405 attr->sched_period < attr->sched_deadline) ||
3406 attr->sched_deadline < attr->sched_runtime)
3413 * check the target process has a UID that matches the current process's
3415 static bool check_same_owner(struct task_struct *p)
3417 const struct cred *cred = current_cred(), *pcred;
3421 pcred = __task_cred(p);
3422 match = (uid_eq(cred->euid, pcred->euid) ||
3423 uid_eq(cred->euid, pcred->uid));
3428 static bool dl_param_changed(struct task_struct *p,
3429 const struct sched_attr *attr)
3431 struct sched_dl_entity *dl_se = &p->dl;
3433 if (dl_se->dl_runtime != attr->sched_runtime ||
3434 dl_se->dl_deadline != attr->sched_deadline ||
3435 dl_se->dl_period != attr->sched_period ||
3436 dl_se->flags != attr->sched_flags)
3442 static int __sched_setscheduler(struct task_struct *p,
3443 const struct sched_attr *attr,
3446 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3447 MAX_RT_PRIO - 1 - attr->sched_priority;
3448 int retval, oldprio, oldpolicy = -1, queued, running;
3449 int new_effective_prio, policy = attr->sched_policy;
3450 unsigned long flags;
3451 const struct sched_class *prev_class;
3455 /* may grab non-irq protected spin_locks */
3456 BUG_ON(in_interrupt());
3458 /* double check policy once rq lock held */
3460 reset_on_fork = p->sched_reset_on_fork;
3461 policy = oldpolicy = p->policy;
3463 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3465 if (policy != SCHED_DEADLINE &&
3466 policy != SCHED_FIFO && policy != SCHED_RR &&
3467 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3468 policy != SCHED_IDLE)
3472 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3476 * Valid priorities for SCHED_FIFO and SCHED_RR are
3477 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3478 * SCHED_BATCH and SCHED_IDLE is 0.
3480 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3481 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3483 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3484 (rt_policy(policy) != (attr->sched_priority != 0)))
3488 * Allow unprivileged RT tasks to decrease priority:
3490 if (user && !capable(CAP_SYS_NICE)) {
3491 if (fair_policy(policy)) {
3492 if (attr->sched_nice < task_nice(p) &&
3493 !can_nice(p, attr->sched_nice))
3497 if (rt_policy(policy)) {
3498 unsigned long rlim_rtprio =
3499 task_rlimit(p, RLIMIT_RTPRIO);
3501 /* can't set/change the rt policy */
3502 if (policy != p->policy && !rlim_rtprio)
3505 /* can't increase priority */
3506 if (attr->sched_priority > p->rt_priority &&
3507 attr->sched_priority > rlim_rtprio)
3512 * Can't set/change SCHED_DEADLINE policy at all for now
3513 * (safest behavior); in the future we would like to allow
3514 * unprivileged DL tasks to increase their relative deadline
3515 * or reduce their runtime (both ways reducing utilization)
3517 if (dl_policy(policy))
3521 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3522 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3524 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3525 if (!can_nice(p, task_nice(p)))
3529 /* can't change other user's priorities */
3530 if (!check_same_owner(p))
3533 /* Normal users shall not reset the sched_reset_on_fork flag */
3534 if (p->sched_reset_on_fork && !reset_on_fork)
3539 retval = security_task_setscheduler(p);
3545 * make sure no PI-waiters arrive (or leave) while we are
3546 * changing the priority of the task:
3548 * To be able to change p->policy safely, the appropriate
3549 * runqueue lock must be held.
3551 rq = task_rq_lock(p, &flags);
3554 * Changing the policy of the stop threads its a very bad idea
3556 if (p == rq->stop) {
3557 task_rq_unlock(rq, p, &flags);
3562 * If not changing anything there's no need to proceed further,
3563 * but store a possible modification of reset_on_fork.
3565 if (unlikely(policy == p->policy)) {
3566 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3568 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3570 if (dl_policy(policy) && dl_param_changed(p, attr))
3573 p->sched_reset_on_fork = reset_on_fork;
3574 task_rq_unlock(rq, p, &flags);
3580 #ifdef CONFIG_RT_GROUP_SCHED
3582 * Do not allow realtime tasks into groups that have no runtime
3585 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3586 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3587 !task_group_is_autogroup(task_group(p))) {
3588 task_rq_unlock(rq, p, &flags);
3593 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3594 cpumask_t *span = rq->rd->span;
3597 * Don't allow tasks with an affinity mask smaller than
3598 * the entire root_domain to become SCHED_DEADLINE. We
3599 * will also fail if there's no bandwidth available.
3601 if (!cpumask_subset(span, &p->cpus_allowed) ||
3602 rq->rd->dl_bw.bw == 0) {
3603 task_rq_unlock(rq, p, &flags);
3610 /* recheck policy now with rq lock held */
3611 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3612 policy = oldpolicy = -1;
3613 task_rq_unlock(rq, p, &flags);
3618 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3619 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3622 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3623 task_rq_unlock(rq, p, &flags);
3627 p->sched_reset_on_fork = reset_on_fork;
3631 * Take priority boosted tasks into account. If the new
3632 * effective priority is unchanged, we just store the new
3633 * normal parameters and do not touch the scheduler class and
3634 * the runqueue. This will be done when the task deboost
3637 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3638 if (new_effective_prio == oldprio) {
3639 __setscheduler_params(p, attr);
3640 task_rq_unlock(rq, p, &flags);
3644 queued = task_on_rq_queued(p);
3645 running = task_current(rq, p);
3647 dequeue_task(rq, p, 0);
3649 put_prev_task(rq, p);
3651 prev_class = p->sched_class;
3652 __setscheduler(rq, p, attr, true);
3655 p->sched_class->set_curr_task(rq);
3658 * We enqueue to tail when the priority of a task is
3659 * increased (user space view).
3661 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3664 check_class_changed(rq, p, prev_class, oldprio);
3665 task_rq_unlock(rq, p, &flags);
3667 rt_mutex_adjust_pi(p);
3672 static int _sched_setscheduler(struct task_struct *p, int policy,
3673 const struct sched_param *param, bool check)
3675 struct sched_attr attr = {
3676 .sched_policy = policy,
3677 .sched_priority = param->sched_priority,
3678 .sched_nice = PRIO_TO_NICE(p->static_prio),
3681 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3682 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3683 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3684 policy &= ~SCHED_RESET_ON_FORK;
3685 attr.sched_policy = policy;
3688 return __sched_setscheduler(p, &attr, check);
3691 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3692 * @p: the task in question.
3693 * @policy: new policy.
3694 * @param: structure containing the new RT priority.
3696 * Return: 0 on success. An error code otherwise.
3698 * NOTE that the task may be already dead.
3700 int sched_setscheduler(struct task_struct *p, int policy,
3701 const struct sched_param *param)
3703 return _sched_setscheduler(p, policy, param, true);
3705 EXPORT_SYMBOL_GPL(sched_setscheduler);
3707 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3709 return __sched_setscheduler(p, attr, true);
3711 EXPORT_SYMBOL_GPL(sched_setattr);
3714 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3715 * @p: the task in question.
3716 * @policy: new policy.
3717 * @param: structure containing the new RT priority.
3719 * Just like sched_setscheduler, only don't bother checking if the
3720 * current context has permission. For example, this is needed in
3721 * stop_machine(): we create temporary high priority worker threads,
3722 * but our caller might not have that capability.
3724 * Return: 0 on success. An error code otherwise.
3726 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3727 const struct sched_param *param)
3729 return _sched_setscheduler(p, policy, param, false);
3733 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3735 struct sched_param lparam;
3736 struct task_struct *p;
3739 if (!param || pid < 0)
3741 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3746 p = find_process_by_pid(pid);
3748 retval = sched_setscheduler(p, policy, &lparam);
3755 * Mimics kernel/events/core.c perf_copy_attr().
3757 static int sched_copy_attr(struct sched_attr __user *uattr,
3758 struct sched_attr *attr)
3763 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3767 * zero the full structure, so that a short copy will be nice.
3769 memset(attr, 0, sizeof(*attr));
3771 ret = get_user(size, &uattr->size);
3775 if (size > PAGE_SIZE) /* silly large */
3778 if (!size) /* abi compat */
3779 size = SCHED_ATTR_SIZE_VER0;
3781 if (size < SCHED_ATTR_SIZE_VER0)
3785 * If we're handed a bigger struct than we know of,
3786 * ensure all the unknown bits are 0 - i.e. new
3787 * user-space does not rely on any kernel feature
3788 * extensions we dont know about yet.
3790 if (size > sizeof(*attr)) {
3791 unsigned char __user *addr;
3792 unsigned char __user *end;
3795 addr = (void __user *)uattr + sizeof(*attr);
3796 end = (void __user *)uattr + size;
3798 for (; addr < end; addr++) {
3799 ret = get_user(val, addr);
3805 size = sizeof(*attr);
3808 ret = copy_from_user(attr, uattr, size);
3813 * XXX: do we want to be lenient like existing syscalls; or do we want
3814 * to be strict and return an error on out-of-bounds values?
3816 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3821 put_user(sizeof(*attr), &uattr->size);
3826 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3827 * @pid: the pid in question.
3828 * @policy: new policy.
3829 * @param: structure containing the new RT priority.
3831 * Return: 0 on success. An error code otherwise.
3833 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3834 struct sched_param __user *, param)
3836 /* negative values for policy are not valid */
3840 return do_sched_setscheduler(pid, policy, param);
3844 * sys_sched_setparam - set/change the RT priority of a thread
3845 * @pid: the pid in question.
3846 * @param: structure containing the new RT priority.
3848 * Return: 0 on success. An error code otherwise.
3850 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3852 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3856 * sys_sched_setattr - same as above, but with extended sched_attr
3857 * @pid: the pid in question.
3858 * @uattr: structure containing the extended parameters.
3859 * @flags: for future extension.
3861 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3862 unsigned int, flags)
3864 struct sched_attr attr;
3865 struct task_struct *p;
3868 if (!uattr || pid < 0 || flags)
3871 retval = sched_copy_attr(uattr, &attr);
3875 if ((int)attr.sched_policy < 0)
3880 p = find_process_by_pid(pid);
3882 retval = sched_setattr(p, &attr);
3889 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3890 * @pid: the pid in question.
3892 * Return: On success, the policy of the thread. Otherwise, a negative error
3895 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3897 struct task_struct *p;
3905 p = find_process_by_pid(pid);
3907 retval = security_task_getscheduler(p);
3910 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3917 * sys_sched_getparam - get the RT priority of a thread
3918 * @pid: the pid in question.
3919 * @param: structure containing the RT priority.
3921 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3924 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3926 struct sched_param lp = { .sched_priority = 0 };
3927 struct task_struct *p;
3930 if (!param || pid < 0)
3934 p = find_process_by_pid(pid);
3939 retval = security_task_getscheduler(p);
3943 if (task_has_rt_policy(p))
3944 lp.sched_priority = p->rt_priority;
3948 * This one might sleep, we cannot do it with a spinlock held ...
3950 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3959 static int sched_read_attr(struct sched_attr __user *uattr,
3960 struct sched_attr *attr,
3965 if (!access_ok(VERIFY_WRITE, uattr, usize))
3969 * If we're handed a smaller struct than we know of,
3970 * ensure all the unknown bits are 0 - i.e. old
3971 * user-space does not get uncomplete information.
3973 if (usize < sizeof(*attr)) {
3974 unsigned char *addr;
3977 addr = (void *)attr + usize;
3978 end = (void *)attr + sizeof(*attr);
3980 for (; addr < end; addr++) {
3988 ret = copy_to_user(uattr, attr, attr->size);
3996 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3997 * @pid: the pid in question.
3998 * @uattr: structure containing the extended parameters.
3999 * @size: sizeof(attr) for fwd/bwd comp.
4000 * @flags: for future extension.
4002 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4003 unsigned int, size, unsigned int, flags)
4005 struct sched_attr attr = {
4006 .size = sizeof(struct sched_attr),
4008 struct task_struct *p;
4011 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4012 size < SCHED_ATTR_SIZE_VER0 || flags)
4016 p = find_process_by_pid(pid);
4021 retval = security_task_getscheduler(p);
4025 attr.sched_policy = p->policy;
4026 if (p->sched_reset_on_fork)
4027 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4028 if (task_has_dl_policy(p))
4029 __getparam_dl(p, &attr);
4030 else if (task_has_rt_policy(p))
4031 attr.sched_priority = p->rt_priority;
4033 attr.sched_nice = task_nice(p);
4037 retval = sched_read_attr(uattr, &attr, size);
4045 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4047 cpumask_var_t cpus_allowed, new_mask;
4048 struct task_struct *p;
4053 p = find_process_by_pid(pid);
4059 /* Prevent p going away */
4063 if (p->flags & PF_NO_SETAFFINITY) {
4067 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4071 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4073 goto out_free_cpus_allowed;
4076 if (!check_same_owner(p)) {
4078 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4080 goto out_free_new_mask;
4085 retval = security_task_setscheduler(p);
4087 goto out_free_new_mask;
4090 cpuset_cpus_allowed(p, cpus_allowed);
4091 cpumask_and(new_mask, in_mask, cpus_allowed);
4094 * Since bandwidth control happens on root_domain basis,
4095 * if admission test is enabled, we only admit -deadline
4096 * tasks allowed to run on all the CPUs in the task's
4100 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4102 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4105 goto out_free_new_mask;
4111 retval = set_cpus_allowed_ptr(p, new_mask);
4114 cpuset_cpus_allowed(p, cpus_allowed);
4115 if (!cpumask_subset(new_mask, cpus_allowed)) {
4117 * We must have raced with a concurrent cpuset
4118 * update. Just reset the cpus_allowed to the
4119 * cpuset's cpus_allowed
4121 cpumask_copy(new_mask, cpus_allowed);
4126 free_cpumask_var(new_mask);
4127 out_free_cpus_allowed:
4128 free_cpumask_var(cpus_allowed);
4134 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4135 struct cpumask *new_mask)
4137 if (len < cpumask_size())
4138 cpumask_clear(new_mask);
4139 else if (len > cpumask_size())
4140 len = cpumask_size();
4142 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4146 * sys_sched_setaffinity - set the cpu affinity of a process
4147 * @pid: pid of the process
4148 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4149 * @user_mask_ptr: user-space pointer to the new cpu mask
4151 * Return: 0 on success. An error code otherwise.
4153 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4154 unsigned long __user *, user_mask_ptr)
4156 cpumask_var_t new_mask;
4159 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4162 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4164 retval = sched_setaffinity(pid, new_mask);
4165 free_cpumask_var(new_mask);
4169 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4171 struct task_struct *p;
4172 unsigned long flags;
4178 p = find_process_by_pid(pid);
4182 retval = security_task_getscheduler(p);
4186 raw_spin_lock_irqsave(&p->pi_lock, flags);
4187 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4188 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4197 * sys_sched_getaffinity - get the cpu affinity of a process
4198 * @pid: pid of the process
4199 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4200 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4202 * Return: 0 on success. An error code otherwise.
4204 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4205 unsigned long __user *, user_mask_ptr)
4210 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4212 if (len & (sizeof(unsigned long)-1))
4215 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4218 ret = sched_getaffinity(pid, mask);
4220 size_t retlen = min_t(size_t, len, cpumask_size());
4222 if (copy_to_user(user_mask_ptr, mask, retlen))
4227 free_cpumask_var(mask);
4233 * sys_sched_yield - yield the current processor to other threads.
4235 * This function yields the current CPU to other tasks. If there are no
4236 * other threads running on this CPU then this function will return.
4240 SYSCALL_DEFINE0(sched_yield)
4242 struct rq *rq = this_rq_lock();
4244 schedstat_inc(rq, yld_count);
4245 current->sched_class->yield_task(rq);
4248 * Since we are going to call schedule() anyway, there's
4249 * no need to preempt or enable interrupts:
4251 __release(rq->lock);
4252 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4253 do_raw_spin_unlock(&rq->lock);
4254 sched_preempt_enable_no_resched();
4261 int __sched _cond_resched(void)
4263 if (should_resched()) {
4264 preempt_schedule_common();
4269 EXPORT_SYMBOL(_cond_resched);
4272 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4273 * call schedule, and on return reacquire the lock.
4275 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4276 * operations here to prevent schedule() from being called twice (once via
4277 * spin_unlock(), once by hand).
4279 int __cond_resched_lock(spinlock_t *lock)
4281 int resched = should_resched();
4284 lockdep_assert_held(lock);
4286 if (spin_needbreak(lock) || resched) {
4289 preempt_schedule_common();
4297 EXPORT_SYMBOL(__cond_resched_lock);
4299 int __sched __cond_resched_softirq(void)
4301 BUG_ON(!in_softirq());
4303 if (should_resched()) {
4305 preempt_schedule_common();
4311 EXPORT_SYMBOL(__cond_resched_softirq);
4314 * yield - yield the current processor to other threads.
4316 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4318 * The scheduler is at all times free to pick the calling task as the most
4319 * eligible task to run, if removing the yield() call from your code breaks
4320 * it, its already broken.
4322 * Typical broken usage is:
4327 * where one assumes that yield() will let 'the other' process run that will
4328 * make event true. If the current task is a SCHED_FIFO task that will never
4329 * happen. Never use yield() as a progress guarantee!!
4331 * If you want to use yield() to wait for something, use wait_event().
4332 * If you want to use yield() to be 'nice' for others, use cond_resched().
4333 * If you still want to use yield(), do not!
4335 void __sched yield(void)
4337 set_current_state(TASK_RUNNING);
4340 EXPORT_SYMBOL(yield);
4343 * yield_to - yield the current processor to another thread in
4344 * your thread group, or accelerate that thread toward the
4345 * processor it's on.
4347 * @preempt: whether task preemption is allowed or not
4349 * It's the caller's job to ensure that the target task struct
4350 * can't go away on us before we can do any checks.
4353 * true (>0) if we indeed boosted the target task.
4354 * false (0) if we failed to boost the target.
4355 * -ESRCH if there's no task to yield to.
4357 int __sched yield_to(struct task_struct *p, bool preempt)
4359 struct task_struct *curr = current;
4360 struct rq *rq, *p_rq;
4361 unsigned long flags;
4364 local_irq_save(flags);
4370 * If we're the only runnable task on the rq and target rq also
4371 * has only one task, there's absolutely no point in yielding.
4373 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4378 double_rq_lock(rq, p_rq);
4379 if (task_rq(p) != p_rq) {
4380 double_rq_unlock(rq, p_rq);
4384 if (!curr->sched_class->yield_to_task)
4387 if (curr->sched_class != p->sched_class)
4390 if (task_running(p_rq, p) || p->state)
4393 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4395 schedstat_inc(rq, yld_count);
4397 * Make p's CPU reschedule; pick_next_entity takes care of
4400 if (preempt && rq != p_rq)
4405 double_rq_unlock(rq, p_rq);
4407 local_irq_restore(flags);
4414 EXPORT_SYMBOL_GPL(yield_to);
4417 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4418 * that process accounting knows that this is a task in IO wait state.
4420 long __sched io_schedule_timeout(long timeout)
4422 int old_iowait = current->in_iowait;
4426 current->in_iowait = 1;
4427 blk_schedule_flush_plug(current);
4429 delayacct_blkio_start();
4431 atomic_inc(&rq->nr_iowait);
4432 ret = schedule_timeout(timeout);
4433 current->in_iowait = old_iowait;
4434 atomic_dec(&rq->nr_iowait);
4435 delayacct_blkio_end();
4439 EXPORT_SYMBOL(io_schedule_timeout);
4442 * sys_sched_get_priority_max - return maximum RT priority.
4443 * @policy: scheduling class.
4445 * Return: On success, this syscall returns the maximum
4446 * rt_priority that can be used by a given scheduling class.
4447 * On failure, a negative error code is returned.
4449 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4456 ret = MAX_USER_RT_PRIO-1;
4458 case SCHED_DEADLINE:
4469 * sys_sched_get_priority_min - return minimum RT priority.
4470 * @policy: scheduling class.
4472 * Return: On success, this syscall returns the minimum
4473 * rt_priority that can be used by a given scheduling class.
4474 * On failure, a negative error code is returned.
4476 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4485 case SCHED_DEADLINE:
4495 * sys_sched_rr_get_interval - return the default timeslice of a process.
4496 * @pid: pid of the process.
4497 * @interval: userspace pointer to the timeslice value.
4499 * this syscall writes the default timeslice value of a given process
4500 * into the user-space timespec buffer. A value of '0' means infinity.
4502 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4505 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4506 struct timespec __user *, interval)
4508 struct task_struct *p;
4509 unsigned int time_slice;
4510 unsigned long flags;
4520 p = find_process_by_pid(pid);
4524 retval = security_task_getscheduler(p);
4528 rq = task_rq_lock(p, &flags);
4530 if (p->sched_class->get_rr_interval)
4531 time_slice = p->sched_class->get_rr_interval(rq, p);
4532 task_rq_unlock(rq, p, &flags);
4535 jiffies_to_timespec(time_slice, &t);
4536 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4544 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4546 void sched_show_task(struct task_struct *p)
4548 unsigned long free = 0;
4550 unsigned long state = p->state;
4553 state = __ffs(state) + 1;
4554 printk(KERN_INFO "%-15.15s %c", p->comm,
4555 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4556 #if BITS_PER_LONG == 32
4557 if (state == TASK_RUNNING)
4558 printk(KERN_CONT " running ");
4560 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4562 if (state == TASK_RUNNING)
4563 printk(KERN_CONT " running task ");
4565 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4567 #ifdef CONFIG_DEBUG_STACK_USAGE
4568 free = stack_not_used(p);
4573 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4575 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4576 task_pid_nr(p), ppid,
4577 (unsigned long)task_thread_info(p)->flags);
4579 print_worker_info(KERN_INFO, p);
4580 show_stack(p, NULL);
4583 void show_state_filter(unsigned long state_filter)
4585 struct task_struct *g, *p;
4587 #if BITS_PER_LONG == 32
4589 " task PC stack pid father\n");
4592 " task PC stack pid father\n");
4595 for_each_process_thread(g, p) {
4597 * reset the NMI-timeout, listing all files on a slow
4598 * console might take a lot of time:
4600 touch_nmi_watchdog();
4601 if (!state_filter || (p->state & state_filter))
4605 touch_all_softlockup_watchdogs();
4607 #ifdef CONFIG_SCHED_DEBUG
4608 sysrq_sched_debug_show();
4612 * Only show locks if all tasks are dumped:
4615 debug_show_all_locks();
4618 void init_idle_bootup_task(struct task_struct *idle)
4620 idle->sched_class = &idle_sched_class;
4624 * init_idle - set up an idle thread for a given CPU
4625 * @idle: task in question
4626 * @cpu: cpu the idle task belongs to
4628 * NOTE: this function does not set the idle thread's NEED_RESCHED
4629 * flag, to make booting more robust.
4631 void init_idle(struct task_struct *idle, int cpu)
4633 struct rq *rq = cpu_rq(cpu);
4634 unsigned long flags;
4636 raw_spin_lock_irqsave(&rq->lock, flags);
4638 __sched_fork(0, idle);
4639 idle->state = TASK_RUNNING;
4640 idle->se.exec_start = sched_clock();
4642 do_set_cpus_allowed(idle, cpumask_of(cpu));
4644 * We're having a chicken and egg problem, even though we are
4645 * holding rq->lock, the cpu isn't yet set to this cpu so the
4646 * lockdep check in task_group() will fail.
4648 * Similar case to sched_fork(). / Alternatively we could
4649 * use task_rq_lock() here and obtain the other rq->lock.
4654 __set_task_cpu(idle, cpu);
4657 rq->curr = rq->idle = idle;
4658 idle->on_rq = TASK_ON_RQ_QUEUED;
4659 #if defined(CONFIG_SMP)
4662 raw_spin_unlock_irqrestore(&rq->lock, flags);
4664 /* Set the preempt count _outside_ the spinlocks! */
4665 init_idle_preempt_count(idle, cpu);
4668 * The idle tasks have their own, simple scheduling class:
4670 idle->sched_class = &idle_sched_class;
4671 ftrace_graph_init_idle_task(idle, cpu);
4672 vtime_init_idle(idle, cpu);
4673 #if defined(CONFIG_SMP)
4674 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4678 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4679 const struct cpumask *trial)
4681 int ret = 1, trial_cpus;
4682 struct dl_bw *cur_dl_b;
4683 unsigned long flags;
4685 if (!cpumask_weight(cur))
4688 rcu_read_lock_sched();
4689 cur_dl_b = dl_bw_of(cpumask_any(cur));
4690 trial_cpus = cpumask_weight(trial);
4692 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4693 if (cur_dl_b->bw != -1 &&
4694 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4696 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4697 rcu_read_unlock_sched();
4702 int task_can_attach(struct task_struct *p,
4703 const struct cpumask *cs_cpus_allowed)
4708 * Kthreads which disallow setaffinity shouldn't be moved
4709 * to a new cpuset; we don't want to change their cpu
4710 * affinity and isolating such threads by their set of
4711 * allowed nodes is unnecessary. Thus, cpusets are not
4712 * applicable for such threads. This prevents checking for
4713 * success of set_cpus_allowed_ptr() on all attached tasks
4714 * before cpus_allowed may be changed.
4716 if (p->flags & PF_NO_SETAFFINITY) {
4722 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4724 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4729 unsigned long flags;
4731 rcu_read_lock_sched();
4732 dl_b = dl_bw_of(dest_cpu);
4733 raw_spin_lock_irqsave(&dl_b->lock, flags);
4734 cpus = dl_bw_cpus(dest_cpu);
4735 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4740 * We reserve space for this task in the destination
4741 * root_domain, as we can't fail after this point.
4742 * We will free resources in the source root_domain
4743 * later on (see set_cpus_allowed_dl()).
4745 __dl_add(dl_b, p->dl.dl_bw);
4747 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4748 rcu_read_unlock_sched();
4758 * move_queued_task - move a queued task to new rq.
4760 * Returns (locked) new rq. Old rq's lock is released.
4762 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4764 struct rq *rq = task_rq(p);
4766 lockdep_assert_held(&rq->lock);
4768 dequeue_task(rq, p, 0);
4769 p->on_rq = TASK_ON_RQ_MIGRATING;
4770 set_task_cpu(p, new_cpu);
4771 raw_spin_unlock(&rq->lock);
4773 rq = cpu_rq(new_cpu);
4775 raw_spin_lock(&rq->lock);
4776 BUG_ON(task_cpu(p) != new_cpu);
4777 p->on_rq = TASK_ON_RQ_QUEUED;
4778 enqueue_task(rq, p, 0);
4779 check_preempt_curr(rq, p, 0);
4784 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4786 if (p->sched_class->set_cpus_allowed)
4787 p->sched_class->set_cpus_allowed(p, new_mask);
4789 cpumask_copy(&p->cpus_allowed, new_mask);
4790 p->nr_cpus_allowed = cpumask_weight(new_mask);
4794 * This is how migration works:
4796 * 1) we invoke migration_cpu_stop() on the target CPU using
4798 * 2) stopper starts to run (implicitly forcing the migrated thread
4800 * 3) it checks whether the migrated task is still in the wrong runqueue.
4801 * 4) if it's in the wrong runqueue then the migration thread removes
4802 * it and puts it into the right queue.
4803 * 5) stopper completes and stop_one_cpu() returns and the migration
4808 * Change a given task's CPU affinity. Migrate the thread to a
4809 * proper CPU and schedule it away if the CPU it's executing on
4810 * is removed from the allowed bitmask.
4812 * NOTE: the caller must have a valid reference to the task, the
4813 * task must not exit() & deallocate itself prematurely. The
4814 * call is not atomic; no spinlocks may be held.
4816 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4818 unsigned long flags;
4820 unsigned int dest_cpu;
4823 rq = task_rq_lock(p, &flags);
4825 if (cpumask_equal(&p->cpus_allowed, new_mask))
4828 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4833 do_set_cpus_allowed(p, new_mask);
4835 /* Can the task run on the task's current CPU? If so, we're done */
4836 if (cpumask_test_cpu(task_cpu(p), new_mask))
4839 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4840 if (task_running(rq, p) || p->state == TASK_WAKING) {
4841 struct migration_arg arg = { p, dest_cpu };
4842 /* Need help from migration thread: drop lock and wait. */
4843 task_rq_unlock(rq, p, &flags);
4844 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4845 tlb_migrate_finish(p->mm);
4847 } else if (task_on_rq_queued(p))
4848 rq = move_queued_task(p, dest_cpu);
4850 task_rq_unlock(rq, p, &flags);
4854 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4857 * Move (not current) task off this cpu, onto dest cpu. We're doing
4858 * this because either it can't run here any more (set_cpus_allowed()
4859 * away from this CPU, or CPU going down), or because we're
4860 * attempting to rebalance this task on exec (sched_exec).
4862 * So we race with normal scheduler movements, but that's OK, as long
4863 * as the task is no longer on this CPU.
4865 * Returns non-zero if task was successfully migrated.
4867 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4872 if (unlikely(!cpu_active(dest_cpu)))
4875 rq = cpu_rq(src_cpu);
4877 raw_spin_lock(&p->pi_lock);
4878 raw_spin_lock(&rq->lock);
4879 /* Already moved. */
4880 if (task_cpu(p) != src_cpu)
4883 /* Affinity changed (again). */
4884 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4888 * If we're not on a rq, the next wake-up will ensure we're
4891 if (task_on_rq_queued(p))
4892 rq = move_queued_task(p, dest_cpu);
4896 raw_spin_unlock(&rq->lock);
4897 raw_spin_unlock(&p->pi_lock);
4901 #ifdef CONFIG_NUMA_BALANCING
4902 /* Migrate current task p to target_cpu */
4903 int migrate_task_to(struct task_struct *p, int target_cpu)
4905 struct migration_arg arg = { p, target_cpu };
4906 int curr_cpu = task_cpu(p);
4908 if (curr_cpu == target_cpu)
4911 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4914 /* TODO: This is not properly updating schedstats */
4916 trace_sched_move_numa(p, curr_cpu, target_cpu);
4917 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4921 * Requeue a task on a given node and accurately track the number of NUMA
4922 * tasks on the runqueues
4924 void sched_setnuma(struct task_struct *p, int nid)
4927 unsigned long flags;
4928 bool queued, running;
4930 rq = task_rq_lock(p, &flags);
4931 queued = task_on_rq_queued(p);
4932 running = task_current(rq, p);
4935 dequeue_task(rq, p, 0);
4937 put_prev_task(rq, p);
4939 p->numa_preferred_nid = nid;
4942 p->sched_class->set_curr_task(rq);
4944 enqueue_task(rq, p, 0);
4945 task_rq_unlock(rq, p, &flags);
4950 * migration_cpu_stop - this will be executed by a highprio stopper thread
4951 * and performs thread migration by bumping thread off CPU then
4952 * 'pushing' onto another runqueue.
4954 static int migration_cpu_stop(void *data)
4956 struct migration_arg *arg = data;
4959 * The original target cpu might have gone down and we might
4960 * be on another cpu but it doesn't matter.
4962 local_irq_disable();
4964 * We need to explicitly wake pending tasks before running
4965 * __migrate_task() such that we will not miss enforcing cpus_allowed
4966 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4968 sched_ttwu_pending();
4969 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4974 #ifdef CONFIG_HOTPLUG_CPU
4977 * Ensures that the idle task is using init_mm right before its cpu goes
4980 void idle_task_exit(void)
4982 struct mm_struct *mm = current->active_mm;
4984 BUG_ON(cpu_online(smp_processor_id()));
4986 if (mm != &init_mm) {
4987 switch_mm(mm, &init_mm, current);
4988 finish_arch_post_lock_switch();
4994 * Since this CPU is going 'away' for a while, fold any nr_active delta
4995 * we might have. Assumes we're called after migrate_tasks() so that the
4996 * nr_active count is stable.
4998 * Also see the comment "Global load-average calculations".
5000 static void calc_load_migrate(struct rq *rq)
5002 long delta = calc_load_fold_active(rq);
5004 atomic_long_add(delta, &calc_load_tasks);
5007 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5011 static const struct sched_class fake_sched_class = {
5012 .put_prev_task = put_prev_task_fake,
5015 static struct task_struct fake_task = {
5017 * Avoid pull_{rt,dl}_task()
5019 .prio = MAX_PRIO + 1,
5020 .sched_class = &fake_sched_class,
5024 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5025 * try_to_wake_up()->select_task_rq().
5027 * Called with rq->lock held even though we'er in stop_machine() and
5028 * there's no concurrency possible, we hold the required locks anyway
5029 * because of lock validation efforts.
5031 static void migrate_tasks(unsigned int dead_cpu)
5033 struct rq *rq = cpu_rq(dead_cpu);
5034 struct task_struct *next, *stop = rq->stop;
5038 * Fudge the rq selection such that the below task selection loop
5039 * doesn't get stuck on the currently eligible stop task.
5041 * We're currently inside stop_machine() and the rq is either stuck
5042 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5043 * either way we should never end up calling schedule() until we're
5049 * put_prev_task() and pick_next_task() sched
5050 * class method both need to have an up-to-date
5051 * value of rq->clock[_task]
5053 update_rq_clock(rq);
5057 * There's this thread running, bail when that's the only
5060 if (rq->nr_running == 1)
5063 next = pick_next_task(rq, &fake_task);
5065 next->sched_class->put_prev_task(rq, next);
5067 /* Find suitable destination for @next, with force if needed. */
5068 dest_cpu = select_fallback_rq(dead_cpu, next);
5069 raw_spin_unlock(&rq->lock);
5071 __migrate_task(next, dead_cpu, dest_cpu);
5073 raw_spin_lock(&rq->lock);
5079 #endif /* CONFIG_HOTPLUG_CPU */
5081 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5083 static struct ctl_table sd_ctl_dir[] = {
5085 .procname = "sched_domain",
5091 static struct ctl_table sd_ctl_root[] = {
5093 .procname = "kernel",
5095 .child = sd_ctl_dir,
5100 static struct ctl_table *sd_alloc_ctl_entry(int n)
5102 struct ctl_table *entry =
5103 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5108 static void sd_free_ctl_entry(struct ctl_table **tablep)
5110 struct ctl_table *entry;
5113 * In the intermediate directories, both the child directory and
5114 * procname are dynamically allocated and could fail but the mode
5115 * will always be set. In the lowest directory the names are
5116 * static strings and all have proc handlers.
5118 for (entry = *tablep; entry->mode; entry++) {
5120 sd_free_ctl_entry(&entry->child);
5121 if (entry->proc_handler == NULL)
5122 kfree(entry->procname);
5129 static int min_load_idx = 0;
5130 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5133 set_table_entry(struct ctl_table *entry,
5134 const char *procname, void *data, int maxlen,
5135 umode_t mode, proc_handler *proc_handler,
5138 entry->procname = procname;
5140 entry->maxlen = maxlen;
5142 entry->proc_handler = proc_handler;
5145 entry->extra1 = &min_load_idx;
5146 entry->extra2 = &max_load_idx;
5150 static struct ctl_table *
5151 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5153 struct ctl_table *table = sd_alloc_ctl_entry(14);
5158 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5159 sizeof(long), 0644, proc_doulongvec_minmax, false);
5160 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5161 sizeof(long), 0644, proc_doulongvec_minmax, false);
5162 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5163 sizeof(int), 0644, proc_dointvec_minmax, true);
5164 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5165 sizeof(int), 0644, proc_dointvec_minmax, true);
5166 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5167 sizeof(int), 0644, proc_dointvec_minmax, true);
5168 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5169 sizeof(int), 0644, proc_dointvec_minmax, true);
5170 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5171 sizeof(int), 0644, proc_dointvec_minmax, true);
5172 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5173 sizeof(int), 0644, proc_dointvec_minmax, false);
5174 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5175 sizeof(int), 0644, proc_dointvec_minmax, false);
5176 set_table_entry(&table[9], "cache_nice_tries",
5177 &sd->cache_nice_tries,
5178 sizeof(int), 0644, proc_dointvec_minmax, false);
5179 set_table_entry(&table[10], "flags", &sd->flags,
5180 sizeof(int), 0644, proc_dointvec_minmax, false);
5181 set_table_entry(&table[11], "max_newidle_lb_cost",
5182 &sd->max_newidle_lb_cost,
5183 sizeof(long), 0644, proc_doulongvec_minmax, false);
5184 set_table_entry(&table[12], "name", sd->name,
5185 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5186 /* &table[13] is terminator */
5191 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5193 struct ctl_table *entry, *table;
5194 struct sched_domain *sd;
5195 int domain_num = 0, i;
5198 for_each_domain(cpu, sd)
5200 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5205 for_each_domain(cpu, sd) {
5206 snprintf(buf, 32, "domain%d", i);
5207 entry->procname = kstrdup(buf, GFP_KERNEL);
5209 entry->child = sd_alloc_ctl_domain_table(sd);
5216 static struct ctl_table_header *sd_sysctl_header;
5217 static void register_sched_domain_sysctl(void)
5219 int i, cpu_num = num_possible_cpus();
5220 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5223 WARN_ON(sd_ctl_dir[0].child);
5224 sd_ctl_dir[0].child = entry;
5229 for_each_possible_cpu(i) {
5230 snprintf(buf, 32, "cpu%d", i);
5231 entry->procname = kstrdup(buf, GFP_KERNEL);
5233 entry->child = sd_alloc_ctl_cpu_table(i);
5237 WARN_ON(sd_sysctl_header);
5238 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5241 /* may be called multiple times per register */
5242 static void unregister_sched_domain_sysctl(void)
5244 if (sd_sysctl_header)
5245 unregister_sysctl_table(sd_sysctl_header);
5246 sd_sysctl_header = NULL;
5247 if (sd_ctl_dir[0].child)
5248 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5251 static void register_sched_domain_sysctl(void)
5254 static void unregister_sched_domain_sysctl(void)
5259 static void set_rq_online(struct rq *rq)
5262 const struct sched_class *class;
5264 cpumask_set_cpu(rq->cpu, rq->rd->online);
5267 for_each_class(class) {
5268 if (class->rq_online)
5269 class->rq_online(rq);
5274 static void set_rq_offline(struct rq *rq)
5277 const struct sched_class *class;
5279 for_each_class(class) {
5280 if (class->rq_offline)
5281 class->rq_offline(rq);
5284 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5290 * migration_call - callback that gets triggered when a CPU is added.
5291 * Here we can start up the necessary migration thread for the new CPU.
5294 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5296 int cpu = (long)hcpu;
5297 unsigned long flags;
5298 struct rq *rq = cpu_rq(cpu);
5300 switch (action & ~CPU_TASKS_FROZEN) {
5302 case CPU_UP_PREPARE:
5303 rq->calc_load_update = calc_load_update;
5307 /* Update our root-domain */
5308 raw_spin_lock_irqsave(&rq->lock, flags);
5310 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5314 raw_spin_unlock_irqrestore(&rq->lock, flags);
5317 #ifdef CONFIG_HOTPLUG_CPU
5319 sched_ttwu_pending();
5320 /* Update our root-domain */
5321 raw_spin_lock_irqsave(&rq->lock, flags);
5323 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5327 BUG_ON(rq->nr_running != 1); /* the migration thread */
5328 raw_spin_unlock_irqrestore(&rq->lock, flags);
5332 calc_load_migrate(rq);
5337 update_max_interval();
5343 * Register at high priority so that task migration (migrate_all_tasks)
5344 * happens before everything else. This has to be lower priority than
5345 * the notifier in the perf_event subsystem, though.
5347 static struct notifier_block migration_notifier = {
5348 .notifier_call = migration_call,
5349 .priority = CPU_PRI_MIGRATION,
5352 static void set_cpu_rq_start_time(void)
5354 int cpu = smp_processor_id();
5355 struct rq *rq = cpu_rq(cpu);
5356 rq->age_stamp = sched_clock_cpu(cpu);
5359 static int sched_cpu_active(struct notifier_block *nfb,
5360 unsigned long action, void *hcpu)
5362 switch (action & ~CPU_TASKS_FROZEN) {
5364 set_cpu_rq_start_time();
5366 case CPU_DOWN_FAILED:
5367 set_cpu_active((long)hcpu, true);
5374 static int sched_cpu_inactive(struct notifier_block *nfb,
5375 unsigned long action, void *hcpu)
5377 switch (action & ~CPU_TASKS_FROZEN) {
5378 case CPU_DOWN_PREPARE:
5379 set_cpu_active((long)hcpu, false);
5386 static int __init migration_init(void)
5388 void *cpu = (void *)(long)smp_processor_id();
5391 /* Initialize migration for the boot CPU */
5392 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5393 BUG_ON(err == NOTIFY_BAD);
5394 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5395 register_cpu_notifier(&migration_notifier);
5397 /* Register cpu active notifiers */
5398 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5399 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5403 early_initcall(migration_init);
5408 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5410 #ifdef CONFIG_SCHED_DEBUG
5412 static __read_mostly int sched_debug_enabled;
5414 static int __init sched_debug_setup(char *str)
5416 sched_debug_enabled = 1;
5420 early_param("sched_debug", sched_debug_setup);
5422 static inline bool sched_debug(void)
5424 return sched_debug_enabled;
5427 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5428 struct cpumask *groupmask)
5430 struct sched_group *group = sd->groups;
5432 cpumask_clear(groupmask);
5434 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5436 if (!(sd->flags & SD_LOAD_BALANCE)) {
5437 printk("does not load-balance\n");
5439 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5444 printk(KERN_CONT "span %*pbl level %s\n",
5445 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5447 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5448 printk(KERN_ERR "ERROR: domain->span does not contain "
5451 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5452 printk(KERN_ERR "ERROR: domain->groups does not contain"
5456 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5460 printk(KERN_ERR "ERROR: group is NULL\n");
5464 if (!cpumask_weight(sched_group_cpus(group))) {
5465 printk(KERN_CONT "\n");
5466 printk(KERN_ERR "ERROR: empty group\n");
5470 if (!(sd->flags & SD_OVERLAP) &&
5471 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5472 printk(KERN_CONT "\n");
5473 printk(KERN_ERR "ERROR: repeated CPUs\n");
5477 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5479 printk(KERN_CONT " %*pbl",
5480 cpumask_pr_args(sched_group_cpus(group)));
5481 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5482 printk(KERN_CONT " (cpu_capacity = %d)",
5483 group->sgc->capacity);
5486 group = group->next;
5487 } while (group != sd->groups);
5488 printk(KERN_CONT "\n");
5490 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5491 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5494 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5495 printk(KERN_ERR "ERROR: parent span is not a superset "
5496 "of domain->span\n");
5500 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5504 if (!sched_debug_enabled)
5508 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5512 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5515 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5523 #else /* !CONFIG_SCHED_DEBUG */
5524 # define sched_domain_debug(sd, cpu) do { } while (0)
5525 static inline bool sched_debug(void)
5529 #endif /* CONFIG_SCHED_DEBUG */
5531 static int sd_degenerate(struct sched_domain *sd)
5533 if (cpumask_weight(sched_domain_span(sd)) == 1)
5536 /* Following flags need at least 2 groups */
5537 if (sd->flags & (SD_LOAD_BALANCE |
5538 SD_BALANCE_NEWIDLE |
5541 SD_SHARE_CPUCAPACITY |
5542 SD_SHARE_PKG_RESOURCES |
5543 SD_SHARE_POWERDOMAIN)) {
5544 if (sd->groups != sd->groups->next)
5548 /* Following flags don't use groups */
5549 if (sd->flags & (SD_WAKE_AFFINE))
5556 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5558 unsigned long cflags = sd->flags, pflags = parent->flags;
5560 if (sd_degenerate(parent))
5563 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5566 /* Flags needing groups don't count if only 1 group in parent */
5567 if (parent->groups == parent->groups->next) {
5568 pflags &= ~(SD_LOAD_BALANCE |
5569 SD_BALANCE_NEWIDLE |
5572 SD_SHARE_CPUCAPACITY |
5573 SD_SHARE_PKG_RESOURCES |
5575 SD_SHARE_POWERDOMAIN);
5576 if (nr_node_ids == 1)
5577 pflags &= ~SD_SERIALIZE;
5579 if (~cflags & pflags)
5585 static void free_rootdomain(struct rcu_head *rcu)
5587 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5589 cpupri_cleanup(&rd->cpupri);
5590 cpudl_cleanup(&rd->cpudl);
5591 free_cpumask_var(rd->dlo_mask);
5592 free_cpumask_var(rd->rto_mask);
5593 free_cpumask_var(rd->online);
5594 free_cpumask_var(rd->span);
5598 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5600 struct root_domain *old_rd = NULL;
5601 unsigned long flags;
5603 raw_spin_lock_irqsave(&rq->lock, flags);
5608 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5611 cpumask_clear_cpu(rq->cpu, old_rd->span);
5614 * If we dont want to free the old_rd yet then
5615 * set old_rd to NULL to skip the freeing later
5618 if (!atomic_dec_and_test(&old_rd->refcount))
5622 atomic_inc(&rd->refcount);
5625 cpumask_set_cpu(rq->cpu, rd->span);
5626 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5629 raw_spin_unlock_irqrestore(&rq->lock, flags);
5632 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5635 static int init_rootdomain(struct root_domain *rd)
5637 memset(rd, 0, sizeof(*rd));
5639 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5641 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5643 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5645 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5648 init_dl_bw(&rd->dl_bw);
5649 if (cpudl_init(&rd->cpudl) != 0)
5652 if (cpupri_init(&rd->cpupri) != 0)
5657 free_cpumask_var(rd->rto_mask);
5659 free_cpumask_var(rd->dlo_mask);
5661 free_cpumask_var(rd->online);
5663 free_cpumask_var(rd->span);
5669 * By default the system creates a single root-domain with all cpus as
5670 * members (mimicking the global state we have today).
5672 struct root_domain def_root_domain;
5674 static void init_defrootdomain(void)
5676 init_rootdomain(&def_root_domain);
5678 atomic_set(&def_root_domain.refcount, 1);
5681 static struct root_domain *alloc_rootdomain(void)
5683 struct root_domain *rd;
5685 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5689 if (init_rootdomain(rd) != 0) {
5697 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5699 struct sched_group *tmp, *first;
5708 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5713 } while (sg != first);
5716 static void free_sched_domain(struct rcu_head *rcu)
5718 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5721 * If its an overlapping domain it has private groups, iterate and
5724 if (sd->flags & SD_OVERLAP) {
5725 free_sched_groups(sd->groups, 1);
5726 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5727 kfree(sd->groups->sgc);
5733 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5735 call_rcu(&sd->rcu, free_sched_domain);
5738 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5740 for (; sd; sd = sd->parent)
5741 destroy_sched_domain(sd, cpu);
5745 * Keep a special pointer to the highest sched_domain that has
5746 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5747 * allows us to avoid some pointer chasing select_idle_sibling().
5749 * Also keep a unique ID per domain (we use the first cpu number in
5750 * the cpumask of the domain), this allows us to quickly tell if
5751 * two cpus are in the same cache domain, see cpus_share_cache().
5753 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5754 DEFINE_PER_CPU(int, sd_llc_size);
5755 DEFINE_PER_CPU(int, sd_llc_id);
5756 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5757 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5758 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5760 static void update_top_cache_domain(int cpu)
5762 struct sched_domain *sd;
5763 struct sched_domain *busy_sd = NULL;
5767 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5769 id = cpumask_first(sched_domain_span(sd));
5770 size = cpumask_weight(sched_domain_span(sd));
5771 busy_sd = sd->parent; /* sd_busy */
5773 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5775 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5776 per_cpu(sd_llc_size, cpu) = size;
5777 per_cpu(sd_llc_id, cpu) = id;
5779 sd = lowest_flag_domain(cpu, SD_NUMA);
5780 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5782 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5783 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5787 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5788 * hold the hotplug lock.
5791 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5793 struct rq *rq = cpu_rq(cpu);
5794 struct sched_domain *tmp;
5796 /* Remove the sched domains which do not contribute to scheduling. */
5797 for (tmp = sd; tmp; ) {
5798 struct sched_domain *parent = tmp->parent;
5802 if (sd_parent_degenerate(tmp, parent)) {
5803 tmp->parent = parent->parent;
5805 parent->parent->child = tmp;
5807 * Transfer SD_PREFER_SIBLING down in case of a
5808 * degenerate parent; the spans match for this
5809 * so the property transfers.
5811 if (parent->flags & SD_PREFER_SIBLING)
5812 tmp->flags |= SD_PREFER_SIBLING;
5813 destroy_sched_domain(parent, cpu);
5818 if (sd && sd_degenerate(sd)) {
5821 destroy_sched_domain(tmp, cpu);
5826 sched_domain_debug(sd, cpu);
5828 rq_attach_root(rq, rd);
5830 rcu_assign_pointer(rq->sd, sd);
5831 destroy_sched_domains(tmp, cpu);
5833 update_top_cache_domain(cpu);
5836 /* Setup the mask of cpus configured for isolated domains */
5837 static int __init isolated_cpu_setup(char *str)
5839 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5840 cpulist_parse(str, cpu_isolated_map);
5844 __setup("isolcpus=", isolated_cpu_setup);
5847 struct sched_domain ** __percpu sd;
5848 struct root_domain *rd;
5859 * Build an iteration mask that can exclude certain CPUs from the upwards
5862 * Asymmetric node setups can result in situations where the domain tree is of
5863 * unequal depth, make sure to skip domains that already cover the entire
5866 * In that case build_sched_domains() will have terminated the iteration early
5867 * and our sibling sd spans will be empty. Domains should always include the
5868 * cpu they're built on, so check that.
5871 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5873 const struct cpumask *span = sched_domain_span(sd);
5874 struct sd_data *sdd = sd->private;
5875 struct sched_domain *sibling;
5878 for_each_cpu(i, span) {
5879 sibling = *per_cpu_ptr(sdd->sd, i);
5880 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5883 cpumask_set_cpu(i, sched_group_mask(sg));
5888 * Return the canonical balance cpu for this group, this is the first cpu
5889 * of this group that's also in the iteration mask.
5891 int group_balance_cpu(struct sched_group *sg)
5893 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5897 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5899 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5900 const struct cpumask *span = sched_domain_span(sd);
5901 struct cpumask *covered = sched_domains_tmpmask;
5902 struct sd_data *sdd = sd->private;
5903 struct sched_domain *sibling;
5906 cpumask_clear(covered);
5908 for_each_cpu(i, span) {
5909 struct cpumask *sg_span;
5911 if (cpumask_test_cpu(i, covered))
5914 sibling = *per_cpu_ptr(sdd->sd, i);
5916 /* See the comment near build_group_mask(). */
5917 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5920 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5921 GFP_KERNEL, cpu_to_node(cpu));
5926 sg_span = sched_group_cpus(sg);
5928 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5930 cpumask_set_cpu(i, sg_span);
5932 cpumask_or(covered, covered, sg_span);
5934 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5935 if (atomic_inc_return(&sg->sgc->ref) == 1)
5936 build_group_mask(sd, sg);
5939 * Initialize sgc->capacity such that even if we mess up the
5940 * domains and no possible iteration will get us here, we won't
5943 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5946 * Make sure the first group of this domain contains the
5947 * canonical balance cpu. Otherwise the sched_domain iteration
5948 * breaks. See update_sg_lb_stats().
5950 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5951 group_balance_cpu(sg) == cpu)
5961 sd->groups = groups;
5966 free_sched_groups(first, 0);
5971 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5973 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5974 struct sched_domain *child = sd->child;
5977 cpu = cpumask_first(sched_domain_span(child));
5980 *sg = *per_cpu_ptr(sdd->sg, cpu);
5981 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5982 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5989 * build_sched_groups will build a circular linked list of the groups
5990 * covered by the given span, and will set each group's ->cpumask correctly,
5991 * and ->cpu_capacity to 0.
5993 * Assumes the sched_domain tree is fully constructed
5996 build_sched_groups(struct sched_domain *sd, int cpu)
5998 struct sched_group *first = NULL, *last = NULL;
5999 struct sd_data *sdd = sd->private;
6000 const struct cpumask *span = sched_domain_span(sd);
6001 struct cpumask *covered;
6004 get_group(cpu, sdd, &sd->groups);
6005 atomic_inc(&sd->groups->ref);
6007 if (cpu != cpumask_first(span))
6010 lockdep_assert_held(&sched_domains_mutex);
6011 covered = sched_domains_tmpmask;
6013 cpumask_clear(covered);
6015 for_each_cpu(i, span) {
6016 struct sched_group *sg;
6019 if (cpumask_test_cpu(i, covered))
6022 group = get_group(i, sdd, &sg);
6023 cpumask_setall(sched_group_mask(sg));
6025 for_each_cpu(j, span) {
6026 if (get_group(j, sdd, NULL) != group)
6029 cpumask_set_cpu(j, covered);
6030 cpumask_set_cpu(j, sched_group_cpus(sg));
6045 * Initialize sched groups cpu_capacity.
6047 * cpu_capacity indicates the capacity of sched group, which is used while
6048 * distributing the load between different sched groups in a sched domain.
6049 * Typically cpu_capacity for all the groups in a sched domain will be same
6050 * unless there are asymmetries in the topology. If there are asymmetries,
6051 * group having more cpu_capacity will pickup more load compared to the
6052 * group having less cpu_capacity.
6054 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6056 struct sched_group *sg = sd->groups;
6061 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6063 } while (sg != sd->groups);
6065 if (cpu != group_balance_cpu(sg))
6068 update_group_capacity(sd, cpu);
6069 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6073 * Initializers for schedule domains
6074 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6077 static int default_relax_domain_level = -1;
6078 int sched_domain_level_max;
6080 static int __init setup_relax_domain_level(char *str)
6082 if (kstrtoint(str, 0, &default_relax_domain_level))
6083 pr_warn("Unable to set relax_domain_level\n");
6087 __setup("relax_domain_level=", setup_relax_domain_level);
6089 static void set_domain_attribute(struct sched_domain *sd,
6090 struct sched_domain_attr *attr)
6094 if (!attr || attr->relax_domain_level < 0) {
6095 if (default_relax_domain_level < 0)
6098 request = default_relax_domain_level;
6100 request = attr->relax_domain_level;
6101 if (request < sd->level) {
6102 /* turn off idle balance on this domain */
6103 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6105 /* turn on idle balance on this domain */
6106 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6110 static void __sdt_free(const struct cpumask *cpu_map);
6111 static int __sdt_alloc(const struct cpumask *cpu_map);
6113 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6114 const struct cpumask *cpu_map)
6118 if (!atomic_read(&d->rd->refcount))
6119 free_rootdomain(&d->rd->rcu); /* fall through */
6121 free_percpu(d->sd); /* fall through */
6123 __sdt_free(cpu_map); /* fall through */
6129 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6130 const struct cpumask *cpu_map)
6132 memset(d, 0, sizeof(*d));
6134 if (__sdt_alloc(cpu_map))
6135 return sa_sd_storage;
6136 d->sd = alloc_percpu(struct sched_domain *);
6138 return sa_sd_storage;
6139 d->rd = alloc_rootdomain();
6142 return sa_rootdomain;
6146 * NULL the sd_data elements we've used to build the sched_domain and
6147 * sched_group structure so that the subsequent __free_domain_allocs()
6148 * will not free the data we're using.
6150 static void claim_allocations(int cpu, struct sched_domain *sd)
6152 struct sd_data *sdd = sd->private;
6154 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6155 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6157 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6158 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6160 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6161 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6165 static int sched_domains_numa_levels;
6166 enum numa_topology_type sched_numa_topology_type;
6167 static int *sched_domains_numa_distance;
6168 int sched_max_numa_distance;
6169 static struct cpumask ***sched_domains_numa_masks;
6170 static int sched_domains_curr_level;
6174 * SD_flags allowed in topology descriptions.
6176 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6177 * SD_SHARE_PKG_RESOURCES - describes shared caches
6178 * SD_NUMA - describes NUMA topologies
6179 * SD_SHARE_POWERDOMAIN - describes shared power domain
6182 * SD_ASYM_PACKING - describes SMT quirks
6184 #define TOPOLOGY_SD_FLAGS \
6185 (SD_SHARE_CPUCAPACITY | \
6186 SD_SHARE_PKG_RESOURCES | \
6189 SD_SHARE_POWERDOMAIN)
6191 static struct sched_domain *
6192 sd_init(struct sched_domain_topology_level *tl, int cpu)
6194 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6195 int sd_weight, sd_flags = 0;
6199 * Ugly hack to pass state to sd_numa_mask()...
6201 sched_domains_curr_level = tl->numa_level;
6204 sd_weight = cpumask_weight(tl->mask(cpu));
6207 sd_flags = (*tl->sd_flags)();
6208 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6209 "wrong sd_flags in topology description\n"))
6210 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6212 *sd = (struct sched_domain){
6213 .min_interval = sd_weight,
6214 .max_interval = 2*sd_weight,
6216 .imbalance_pct = 125,
6218 .cache_nice_tries = 0,
6225 .flags = 1*SD_LOAD_BALANCE
6226 | 1*SD_BALANCE_NEWIDLE
6231 | 0*SD_SHARE_CPUCAPACITY
6232 | 0*SD_SHARE_PKG_RESOURCES
6234 | 0*SD_PREFER_SIBLING
6239 .last_balance = jiffies,
6240 .balance_interval = sd_weight,
6242 .max_newidle_lb_cost = 0,
6243 .next_decay_max_lb_cost = jiffies,
6244 #ifdef CONFIG_SCHED_DEBUG
6250 * Convert topological properties into behaviour.
6253 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6254 sd->flags |= SD_PREFER_SIBLING;
6255 sd->imbalance_pct = 110;
6256 sd->smt_gain = 1178; /* ~15% */
6258 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6259 sd->imbalance_pct = 117;
6260 sd->cache_nice_tries = 1;
6264 } else if (sd->flags & SD_NUMA) {
6265 sd->cache_nice_tries = 2;
6269 sd->flags |= SD_SERIALIZE;
6270 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6271 sd->flags &= ~(SD_BALANCE_EXEC |
6278 sd->flags |= SD_PREFER_SIBLING;
6279 sd->cache_nice_tries = 1;
6284 sd->private = &tl->data;
6290 * Topology list, bottom-up.
6292 static struct sched_domain_topology_level default_topology[] = {
6293 #ifdef CONFIG_SCHED_SMT
6294 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6296 #ifdef CONFIG_SCHED_MC
6297 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6299 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6303 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6305 #define for_each_sd_topology(tl) \
6306 for (tl = sched_domain_topology; tl->mask; tl++)
6308 void set_sched_topology(struct sched_domain_topology_level *tl)
6310 sched_domain_topology = tl;
6315 static const struct cpumask *sd_numa_mask(int cpu)
6317 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6320 static void sched_numa_warn(const char *str)
6322 static int done = false;
6330 printk(KERN_WARNING "ERROR: %s\n\n", str);
6332 for (i = 0; i < nr_node_ids; i++) {
6333 printk(KERN_WARNING " ");
6334 for (j = 0; j < nr_node_ids; j++)
6335 printk(KERN_CONT "%02d ", node_distance(i,j));
6336 printk(KERN_CONT "\n");
6338 printk(KERN_WARNING "\n");
6341 bool find_numa_distance(int distance)
6345 if (distance == node_distance(0, 0))
6348 for (i = 0; i < sched_domains_numa_levels; i++) {
6349 if (sched_domains_numa_distance[i] == distance)
6357 * A system can have three types of NUMA topology:
6358 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6359 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6360 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6362 * The difference between a glueless mesh topology and a backplane
6363 * topology lies in whether communication between not directly
6364 * connected nodes goes through intermediary nodes (where programs
6365 * could run), or through backplane controllers. This affects
6366 * placement of programs.
6368 * The type of topology can be discerned with the following tests:
6369 * - If the maximum distance between any nodes is 1 hop, the system
6370 * is directly connected.
6371 * - If for two nodes A and B, located N > 1 hops away from each other,
6372 * there is an intermediary node C, which is < N hops away from both
6373 * nodes A and B, the system is a glueless mesh.
6375 static void init_numa_topology_type(void)
6379 n = sched_max_numa_distance;
6382 sched_numa_topology_type = NUMA_DIRECT;
6384 for_each_online_node(a) {
6385 for_each_online_node(b) {
6386 /* Find two nodes furthest removed from each other. */
6387 if (node_distance(a, b) < n)
6390 /* Is there an intermediary node between a and b? */
6391 for_each_online_node(c) {
6392 if (node_distance(a, c) < n &&
6393 node_distance(b, c) < n) {
6394 sched_numa_topology_type =
6400 sched_numa_topology_type = NUMA_BACKPLANE;
6406 static void sched_init_numa(void)
6408 int next_distance, curr_distance = node_distance(0, 0);
6409 struct sched_domain_topology_level *tl;
6413 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6414 if (!sched_domains_numa_distance)
6418 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6419 * unique distances in the node_distance() table.
6421 * Assumes node_distance(0,j) includes all distances in
6422 * node_distance(i,j) in order to avoid cubic time.
6424 next_distance = curr_distance;
6425 for (i = 0; i < nr_node_ids; i++) {
6426 for (j = 0; j < nr_node_ids; j++) {
6427 for (k = 0; k < nr_node_ids; k++) {
6428 int distance = node_distance(i, k);
6430 if (distance > curr_distance &&
6431 (distance < next_distance ||
6432 next_distance == curr_distance))
6433 next_distance = distance;
6436 * While not a strong assumption it would be nice to know
6437 * about cases where if node A is connected to B, B is not
6438 * equally connected to A.
6440 if (sched_debug() && node_distance(k, i) != distance)
6441 sched_numa_warn("Node-distance not symmetric");
6443 if (sched_debug() && i && !find_numa_distance(distance))
6444 sched_numa_warn("Node-0 not representative");
6446 if (next_distance != curr_distance) {
6447 sched_domains_numa_distance[level++] = next_distance;
6448 sched_domains_numa_levels = level;
6449 curr_distance = next_distance;
6454 * In case of sched_debug() we verify the above assumption.
6464 * 'level' contains the number of unique distances, excluding the
6465 * identity distance node_distance(i,i).
6467 * The sched_domains_numa_distance[] array includes the actual distance
6472 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6473 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6474 * the array will contain less then 'level' members. This could be
6475 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6476 * in other functions.
6478 * We reset it to 'level' at the end of this function.
6480 sched_domains_numa_levels = 0;
6482 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6483 if (!sched_domains_numa_masks)
6487 * Now for each level, construct a mask per node which contains all
6488 * cpus of nodes that are that many hops away from us.
6490 for (i = 0; i < level; i++) {
6491 sched_domains_numa_masks[i] =
6492 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6493 if (!sched_domains_numa_masks[i])
6496 for (j = 0; j < nr_node_ids; j++) {
6497 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6501 sched_domains_numa_masks[i][j] = mask;
6503 for (k = 0; k < nr_node_ids; k++) {
6504 if (node_distance(j, k) > sched_domains_numa_distance[i])
6507 cpumask_or(mask, mask, cpumask_of_node(k));
6512 /* Compute default topology size */
6513 for (i = 0; sched_domain_topology[i].mask; i++);
6515 tl = kzalloc((i + level + 1) *
6516 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6521 * Copy the default topology bits..
6523 for (i = 0; sched_domain_topology[i].mask; i++)
6524 tl[i] = sched_domain_topology[i];
6527 * .. and append 'j' levels of NUMA goodness.
6529 for (j = 0; j < level; i++, j++) {
6530 tl[i] = (struct sched_domain_topology_level){
6531 .mask = sd_numa_mask,
6532 .sd_flags = cpu_numa_flags,
6533 .flags = SDTL_OVERLAP,
6539 sched_domain_topology = tl;
6541 sched_domains_numa_levels = level;
6542 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6544 init_numa_topology_type();
6547 static void sched_domains_numa_masks_set(int cpu)
6550 int node = cpu_to_node(cpu);
6552 for (i = 0; i < sched_domains_numa_levels; i++) {
6553 for (j = 0; j < nr_node_ids; j++) {
6554 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6555 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6560 static void sched_domains_numa_masks_clear(int cpu)
6563 for (i = 0; i < sched_domains_numa_levels; i++) {
6564 for (j = 0; j < nr_node_ids; j++)
6565 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6570 * Update sched_domains_numa_masks[level][node] array when new cpus
6573 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6574 unsigned long action,
6577 int cpu = (long)hcpu;
6579 switch (action & ~CPU_TASKS_FROZEN) {
6581 sched_domains_numa_masks_set(cpu);
6585 sched_domains_numa_masks_clear(cpu);
6595 static inline void sched_init_numa(void)
6599 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6600 unsigned long action,
6605 #endif /* CONFIG_NUMA */
6607 static int __sdt_alloc(const struct cpumask *cpu_map)
6609 struct sched_domain_topology_level *tl;
6612 for_each_sd_topology(tl) {
6613 struct sd_data *sdd = &tl->data;
6615 sdd->sd = alloc_percpu(struct sched_domain *);
6619 sdd->sg = alloc_percpu(struct sched_group *);
6623 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6627 for_each_cpu(j, cpu_map) {
6628 struct sched_domain *sd;
6629 struct sched_group *sg;
6630 struct sched_group_capacity *sgc;
6632 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6633 GFP_KERNEL, cpu_to_node(j));
6637 *per_cpu_ptr(sdd->sd, j) = sd;
6639 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6640 GFP_KERNEL, cpu_to_node(j));
6646 *per_cpu_ptr(sdd->sg, j) = sg;
6648 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6649 GFP_KERNEL, cpu_to_node(j));
6653 *per_cpu_ptr(sdd->sgc, j) = sgc;
6660 static void __sdt_free(const struct cpumask *cpu_map)
6662 struct sched_domain_topology_level *tl;
6665 for_each_sd_topology(tl) {
6666 struct sd_data *sdd = &tl->data;
6668 for_each_cpu(j, cpu_map) {
6669 struct sched_domain *sd;
6672 sd = *per_cpu_ptr(sdd->sd, j);
6673 if (sd && (sd->flags & SD_OVERLAP))
6674 free_sched_groups(sd->groups, 0);
6675 kfree(*per_cpu_ptr(sdd->sd, j));
6679 kfree(*per_cpu_ptr(sdd->sg, j));
6681 kfree(*per_cpu_ptr(sdd->sgc, j));
6683 free_percpu(sdd->sd);
6685 free_percpu(sdd->sg);
6687 free_percpu(sdd->sgc);
6692 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6693 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6694 struct sched_domain *child, int cpu)
6696 struct sched_domain *sd = sd_init(tl, cpu);
6700 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6702 sd->level = child->level + 1;
6703 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6707 if (!cpumask_subset(sched_domain_span(child),
6708 sched_domain_span(sd))) {
6709 pr_err("BUG: arch topology borken\n");
6710 #ifdef CONFIG_SCHED_DEBUG
6711 pr_err(" the %s domain not a subset of the %s domain\n",
6712 child->name, sd->name);
6714 /* Fixup, ensure @sd has at least @child cpus. */
6715 cpumask_or(sched_domain_span(sd),
6716 sched_domain_span(sd),
6717 sched_domain_span(child));
6721 set_domain_attribute(sd, attr);
6727 * Build sched domains for a given set of cpus and attach the sched domains
6728 * to the individual cpus
6730 static int build_sched_domains(const struct cpumask *cpu_map,
6731 struct sched_domain_attr *attr)
6733 enum s_alloc alloc_state;
6734 struct sched_domain *sd;
6736 int i, ret = -ENOMEM;
6738 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6739 if (alloc_state != sa_rootdomain)
6742 /* Set up domains for cpus specified by the cpu_map. */
6743 for_each_cpu(i, cpu_map) {
6744 struct sched_domain_topology_level *tl;
6747 for_each_sd_topology(tl) {
6748 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6749 if (tl == sched_domain_topology)
6750 *per_cpu_ptr(d.sd, i) = sd;
6751 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6752 sd->flags |= SD_OVERLAP;
6753 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6758 /* Build the groups for the domains */
6759 for_each_cpu(i, cpu_map) {
6760 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6761 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6762 if (sd->flags & SD_OVERLAP) {
6763 if (build_overlap_sched_groups(sd, i))
6766 if (build_sched_groups(sd, i))
6772 /* Calculate CPU capacity for physical packages and nodes */
6773 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6774 if (!cpumask_test_cpu(i, cpu_map))
6777 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6778 claim_allocations(i, sd);
6779 init_sched_groups_capacity(i, sd);
6783 /* Attach the domains */
6785 for_each_cpu(i, cpu_map) {
6786 sd = *per_cpu_ptr(d.sd, i);
6787 cpu_attach_domain(sd, d.rd, i);
6793 __free_domain_allocs(&d, alloc_state, cpu_map);
6797 static cpumask_var_t *doms_cur; /* current sched domains */
6798 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6799 static struct sched_domain_attr *dattr_cur;
6800 /* attribues of custom domains in 'doms_cur' */
6803 * Special case: If a kmalloc of a doms_cur partition (array of
6804 * cpumask) fails, then fallback to a single sched domain,
6805 * as determined by the single cpumask fallback_doms.
6807 static cpumask_var_t fallback_doms;
6810 * arch_update_cpu_topology lets virtualized architectures update the
6811 * cpu core maps. It is supposed to return 1 if the topology changed
6812 * or 0 if it stayed the same.
6814 int __weak arch_update_cpu_topology(void)
6819 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6822 cpumask_var_t *doms;
6824 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6827 for (i = 0; i < ndoms; i++) {
6828 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6829 free_sched_domains(doms, i);
6836 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6839 for (i = 0; i < ndoms; i++)
6840 free_cpumask_var(doms[i]);
6845 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6846 * For now this just excludes isolated cpus, but could be used to
6847 * exclude other special cases in the future.
6849 static int init_sched_domains(const struct cpumask *cpu_map)
6853 arch_update_cpu_topology();
6855 doms_cur = alloc_sched_domains(ndoms_cur);
6857 doms_cur = &fallback_doms;
6858 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6859 err = build_sched_domains(doms_cur[0], NULL);
6860 register_sched_domain_sysctl();
6866 * Detach sched domains from a group of cpus specified in cpu_map
6867 * These cpus will now be attached to the NULL domain
6869 static void detach_destroy_domains(const struct cpumask *cpu_map)
6874 for_each_cpu(i, cpu_map)
6875 cpu_attach_domain(NULL, &def_root_domain, i);
6879 /* handle null as "default" */
6880 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6881 struct sched_domain_attr *new, int idx_new)
6883 struct sched_domain_attr tmp;
6890 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6891 new ? (new + idx_new) : &tmp,
6892 sizeof(struct sched_domain_attr));
6896 * Partition sched domains as specified by the 'ndoms_new'
6897 * cpumasks in the array doms_new[] of cpumasks. This compares
6898 * doms_new[] to the current sched domain partitioning, doms_cur[].
6899 * It destroys each deleted domain and builds each new domain.
6901 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6902 * The masks don't intersect (don't overlap.) We should setup one
6903 * sched domain for each mask. CPUs not in any of the cpumasks will
6904 * not be load balanced. If the same cpumask appears both in the
6905 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6908 * The passed in 'doms_new' should be allocated using
6909 * alloc_sched_domains. This routine takes ownership of it and will
6910 * free_sched_domains it when done with it. If the caller failed the
6911 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6912 * and partition_sched_domains() will fallback to the single partition
6913 * 'fallback_doms', it also forces the domains to be rebuilt.
6915 * If doms_new == NULL it will be replaced with cpu_online_mask.
6916 * ndoms_new == 0 is a special case for destroying existing domains,
6917 * and it will not create the default domain.
6919 * Call with hotplug lock held
6921 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6922 struct sched_domain_attr *dattr_new)
6927 mutex_lock(&sched_domains_mutex);
6929 /* always unregister in case we don't destroy any domains */
6930 unregister_sched_domain_sysctl();
6932 /* Let architecture update cpu core mappings. */
6933 new_topology = arch_update_cpu_topology();
6935 n = doms_new ? ndoms_new : 0;
6937 /* Destroy deleted domains */
6938 for (i = 0; i < ndoms_cur; i++) {
6939 for (j = 0; j < n && !new_topology; j++) {
6940 if (cpumask_equal(doms_cur[i], doms_new[j])
6941 && dattrs_equal(dattr_cur, i, dattr_new, j))
6944 /* no match - a current sched domain not in new doms_new[] */
6945 detach_destroy_domains(doms_cur[i]);
6951 if (doms_new == NULL) {
6953 doms_new = &fallback_doms;
6954 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6955 WARN_ON_ONCE(dattr_new);
6958 /* Build new domains */
6959 for (i = 0; i < ndoms_new; i++) {
6960 for (j = 0; j < n && !new_topology; j++) {
6961 if (cpumask_equal(doms_new[i], doms_cur[j])
6962 && dattrs_equal(dattr_new, i, dattr_cur, j))
6965 /* no match - add a new doms_new */
6966 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6971 /* Remember the new sched domains */
6972 if (doms_cur != &fallback_doms)
6973 free_sched_domains(doms_cur, ndoms_cur);
6974 kfree(dattr_cur); /* kfree(NULL) is safe */
6975 doms_cur = doms_new;
6976 dattr_cur = dattr_new;
6977 ndoms_cur = ndoms_new;
6979 register_sched_domain_sysctl();
6981 mutex_unlock(&sched_domains_mutex);
6984 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6987 * Update cpusets according to cpu_active mask. If cpusets are
6988 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6989 * around partition_sched_domains().
6991 * If we come here as part of a suspend/resume, don't touch cpusets because we
6992 * want to restore it back to its original state upon resume anyway.
6994 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6998 case CPU_ONLINE_FROZEN:
6999 case CPU_DOWN_FAILED_FROZEN:
7002 * num_cpus_frozen tracks how many CPUs are involved in suspend
7003 * resume sequence. As long as this is not the last online
7004 * operation in the resume sequence, just build a single sched
7005 * domain, ignoring cpusets.
7008 if (likely(num_cpus_frozen)) {
7009 partition_sched_domains(1, NULL, NULL);
7014 * This is the last CPU online operation. So fall through and
7015 * restore the original sched domains by considering the
7016 * cpuset configurations.
7020 cpuset_update_active_cpus(true);
7028 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7031 unsigned long flags;
7032 long cpu = (long)hcpu;
7038 case CPU_DOWN_PREPARE:
7039 rcu_read_lock_sched();
7040 dl_b = dl_bw_of(cpu);
7042 raw_spin_lock_irqsave(&dl_b->lock, flags);
7043 cpus = dl_bw_cpus(cpu);
7044 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7045 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7047 rcu_read_unlock_sched();
7050 return notifier_from_errno(-EBUSY);
7051 cpuset_update_active_cpus(false);
7053 case CPU_DOWN_PREPARE_FROZEN:
7055 partition_sched_domains(1, NULL, NULL);
7063 void __init sched_init_smp(void)
7065 cpumask_var_t non_isolated_cpus;
7067 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7068 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7070 /* nohz_full won't take effect without isolating the cpus. */
7071 tick_nohz_full_add_cpus_to(cpu_isolated_map);
7076 * There's no userspace yet to cause hotplug operations; hence all the
7077 * cpu masks are stable and all blatant races in the below code cannot
7080 mutex_lock(&sched_domains_mutex);
7081 init_sched_domains(cpu_active_mask);
7082 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7083 if (cpumask_empty(non_isolated_cpus))
7084 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7085 mutex_unlock(&sched_domains_mutex);
7087 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7088 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7089 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7093 /* Move init over to a non-isolated CPU */
7094 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7096 sched_init_granularity();
7097 free_cpumask_var(non_isolated_cpus);
7099 init_sched_rt_class();
7100 init_sched_dl_class();
7103 void __init sched_init_smp(void)
7105 sched_init_granularity();
7107 #endif /* CONFIG_SMP */
7109 int in_sched_functions(unsigned long addr)
7111 return in_lock_functions(addr) ||
7112 (addr >= (unsigned long)__sched_text_start
7113 && addr < (unsigned long)__sched_text_end);
7116 #ifdef CONFIG_CGROUP_SCHED
7118 * Default task group.
7119 * Every task in system belongs to this group at bootup.
7121 struct task_group root_task_group;
7122 LIST_HEAD(task_groups);
7125 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7127 void __init sched_init(void)
7130 unsigned long alloc_size = 0, ptr;
7132 #ifdef CONFIG_FAIR_GROUP_SCHED
7133 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7135 #ifdef CONFIG_RT_GROUP_SCHED
7136 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7139 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7141 #ifdef CONFIG_FAIR_GROUP_SCHED
7142 root_task_group.se = (struct sched_entity **)ptr;
7143 ptr += nr_cpu_ids * sizeof(void **);
7145 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7146 ptr += nr_cpu_ids * sizeof(void **);
7148 #endif /* CONFIG_FAIR_GROUP_SCHED */
7149 #ifdef CONFIG_RT_GROUP_SCHED
7150 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7151 ptr += nr_cpu_ids * sizeof(void **);
7153 root_task_group.rt_rq = (struct rt_rq **)ptr;
7154 ptr += nr_cpu_ids * sizeof(void **);
7156 #endif /* CONFIG_RT_GROUP_SCHED */
7158 #ifdef CONFIG_CPUMASK_OFFSTACK
7159 for_each_possible_cpu(i) {
7160 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7161 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7163 #endif /* CONFIG_CPUMASK_OFFSTACK */
7165 init_rt_bandwidth(&def_rt_bandwidth,
7166 global_rt_period(), global_rt_runtime());
7167 init_dl_bandwidth(&def_dl_bandwidth,
7168 global_rt_period(), global_rt_runtime());
7171 init_defrootdomain();
7174 #ifdef CONFIG_RT_GROUP_SCHED
7175 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7176 global_rt_period(), global_rt_runtime());
7177 #endif /* CONFIG_RT_GROUP_SCHED */
7179 #ifdef CONFIG_CGROUP_SCHED
7180 list_add(&root_task_group.list, &task_groups);
7181 INIT_LIST_HEAD(&root_task_group.children);
7182 INIT_LIST_HEAD(&root_task_group.siblings);
7183 autogroup_init(&init_task);
7185 #endif /* CONFIG_CGROUP_SCHED */
7187 for_each_possible_cpu(i) {
7191 raw_spin_lock_init(&rq->lock);
7193 rq->calc_load_active = 0;
7194 rq->calc_load_update = jiffies + LOAD_FREQ;
7195 init_cfs_rq(&rq->cfs);
7196 init_rt_rq(&rq->rt);
7197 init_dl_rq(&rq->dl);
7198 #ifdef CONFIG_FAIR_GROUP_SCHED
7199 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7200 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7202 * How much cpu bandwidth does root_task_group get?
7204 * In case of task-groups formed thr' the cgroup filesystem, it
7205 * gets 100% of the cpu resources in the system. This overall
7206 * system cpu resource is divided among the tasks of
7207 * root_task_group and its child task-groups in a fair manner,
7208 * based on each entity's (task or task-group's) weight
7209 * (se->load.weight).
7211 * In other words, if root_task_group has 10 tasks of weight
7212 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7213 * then A0's share of the cpu resource is:
7215 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7217 * We achieve this by letting root_task_group's tasks sit
7218 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7220 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7221 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7222 #endif /* CONFIG_FAIR_GROUP_SCHED */
7224 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7225 #ifdef CONFIG_RT_GROUP_SCHED
7226 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7229 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7230 rq->cpu_load[j] = 0;
7232 rq->last_load_update_tick = jiffies;
7237 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7238 rq->post_schedule = 0;
7239 rq->active_balance = 0;
7240 rq->next_balance = jiffies;
7245 rq->avg_idle = 2*sysctl_sched_migration_cost;
7246 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7248 INIT_LIST_HEAD(&rq->cfs_tasks);
7250 rq_attach_root(rq, &def_root_domain);
7251 #ifdef CONFIG_NO_HZ_COMMON
7254 #ifdef CONFIG_NO_HZ_FULL
7255 rq->last_sched_tick = 0;
7259 atomic_set(&rq->nr_iowait, 0);
7262 set_load_weight(&init_task);
7264 #ifdef CONFIG_PREEMPT_NOTIFIERS
7265 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7269 * The boot idle thread does lazy MMU switching as well:
7271 atomic_inc(&init_mm.mm_count);
7272 enter_lazy_tlb(&init_mm, current);
7275 * During early bootup we pretend to be a normal task:
7277 current->sched_class = &fair_sched_class;
7280 * Make us the idle thread. Technically, schedule() should not be
7281 * called from this thread, however somewhere below it might be,
7282 * but because we are the idle thread, we just pick up running again
7283 * when this runqueue becomes "idle".
7285 init_idle(current, smp_processor_id());
7287 calc_load_update = jiffies + LOAD_FREQ;
7290 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7291 /* May be allocated at isolcpus cmdline parse time */
7292 if (cpu_isolated_map == NULL)
7293 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7294 idle_thread_set_boot_cpu();
7295 set_cpu_rq_start_time();
7297 init_sched_fair_class();
7299 scheduler_running = 1;
7302 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7303 static inline int preempt_count_equals(int preempt_offset)
7305 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7307 return (nested == preempt_offset);
7310 void __might_sleep(const char *file, int line, int preempt_offset)
7313 * Blocking primitives will set (and therefore destroy) current->state,
7314 * since we will exit with TASK_RUNNING make sure we enter with it,
7315 * otherwise we will destroy state.
7317 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7318 "do not call blocking ops when !TASK_RUNNING; "
7319 "state=%lx set at [<%p>] %pS\n",
7321 (void *)current->task_state_change,
7322 (void *)current->task_state_change);
7324 ___might_sleep(file, line, preempt_offset);
7326 EXPORT_SYMBOL(__might_sleep);
7328 void ___might_sleep(const char *file, int line, int preempt_offset)
7330 static unsigned long prev_jiffy; /* ratelimiting */
7332 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7333 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7334 !is_idle_task(current)) ||
7335 system_state != SYSTEM_RUNNING || oops_in_progress)
7337 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7339 prev_jiffy = jiffies;
7342 "BUG: sleeping function called from invalid context at %s:%d\n",
7345 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7346 in_atomic(), irqs_disabled(),
7347 current->pid, current->comm);
7349 if (task_stack_end_corrupted(current))
7350 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7352 debug_show_held_locks(current);
7353 if (irqs_disabled())
7354 print_irqtrace_events(current);
7355 #ifdef CONFIG_DEBUG_PREEMPT
7356 if (!preempt_count_equals(preempt_offset)) {
7357 pr_err("Preemption disabled at:");
7358 print_ip_sym(current->preempt_disable_ip);
7364 EXPORT_SYMBOL(___might_sleep);
7367 #ifdef CONFIG_MAGIC_SYSRQ
7368 static void normalize_task(struct rq *rq, struct task_struct *p)
7370 const struct sched_class *prev_class = p->sched_class;
7371 struct sched_attr attr = {
7372 .sched_policy = SCHED_NORMAL,
7374 int old_prio = p->prio;
7377 queued = task_on_rq_queued(p);
7379 dequeue_task(rq, p, 0);
7380 __setscheduler(rq, p, &attr, false);
7382 enqueue_task(rq, p, 0);
7386 check_class_changed(rq, p, prev_class, old_prio);
7389 void normalize_rt_tasks(void)
7391 struct task_struct *g, *p;
7392 unsigned long flags;
7395 read_lock(&tasklist_lock);
7396 for_each_process_thread(g, p) {
7398 * Only normalize user tasks:
7400 if (p->flags & PF_KTHREAD)
7403 p->se.exec_start = 0;
7404 #ifdef CONFIG_SCHEDSTATS
7405 p->se.statistics.wait_start = 0;
7406 p->se.statistics.sleep_start = 0;
7407 p->se.statistics.block_start = 0;
7410 if (!dl_task(p) && !rt_task(p)) {
7412 * Renice negative nice level userspace
7415 if (task_nice(p) < 0)
7416 set_user_nice(p, 0);
7420 rq = task_rq_lock(p, &flags);
7421 normalize_task(rq, p);
7422 task_rq_unlock(rq, p, &flags);
7424 read_unlock(&tasklist_lock);
7427 #endif /* CONFIG_MAGIC_SYSRQ */
7429 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7431 * These functions are only useful for the IA64 MCA handling, or kdb.
7433 * They can only be called when the whole system has been
7434 * stopped - every CPU needs to be quiescent, and no scheduling
7435 * activity can take place. Using them for anything else would
7436 * be a serious bug, and as a result, they aren't even visible
7437 * under any other configuration.
7441 * curr_task - return the current task for a given cpu.
7442 * @cpu: the processor in question.
7444 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7446 * Return: The current task for @cpu.
7448 struct task_struct *curr_task(int cpu)
7450 return cpu_curr(cpu);
7453 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7457 * set_curr_task - set the current task for a given cpu.
7458 * @cpu: the processor in question.
7459 * @p: the task pointer to set.
7461 * Description: This function must only be used when non-maskable interrupts
7462 * are serviced on a separate stack. It allows the architecture to switch the
7463 * notion of the current task on a cpu in a non-blocking manner. This function
7464 * must be called with all CPU's synchronized, and interrupts disabled, the
7465 * and caller must save the original value of the current task (see
7466 * curr_task() above) and restore that value before reenabling interrupts and
7467 * re-starting the system.
7469 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7471 void set_curr_task(int cpu, struct task_struct *p)
7478 #ifdef CONFIG_CGROUP_SCHED
7479 /* task_group_lock serializes the addition/removal of task groups */
7480 static DEFINE_SPINLOCK(task_group_lock);
7482 static void free_sched_group(struct task_group *tg)
7484 free_fair_sched_group(tg);
7485 free_rt_sched_group(tg);
7490 /* allocate runqueue etc for a new task group */
7491 struct task_group *sched_create_group(struct task_group *parent)
7493 struct task_group *tg;
7495 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7497 return ERR_PTR(-ENOMEM);
7499 if (!alloc_fair_sched_group(tg, parent))
7502 if (!alloc_rt_sched_group(tg, parent))
7508 free_sched_group(tg);
7509 return ERR_PTR(-ENOMEM);
7512 void sched_online_group(struct task_group *tg, struct task_group *parent)
7514 unsigned long flags;
7516 spin_lock_irqsave(&task_group_lock, flags);
7517 list_add_rcu(&tg->list, &task_groups);
7519 WARN_ON(!parent); /* root should already exist */
7521 tg->parent = parent;
7522 INIT_LIST_HEAD(&tg->children);
7523 list_add_rcu(&tg->siblings, &parent->children);
7524 spin_unlock_irqrestore(&task_group_lock, flags);
7527 /* rcu callback to free various structures associated with a task group */
7528 static void free_sched_group_rcu(struct rcu_head *rhp)
7530 /* now it should be safe to free those cfs_rqs */
7531 free_sched_group(container_of(rhp, struct task_group, rcu));
7534 /* Destroy runqueue etc associated with a task group */
7535 void sched_destroy_group(struct task_group *tg)
7537 /* wait for possible concurrent references to cfs_rqs complete */
7538 call_rcu(&tg->rcu, free_sched_group_rcu);
7541 void sched_offline_group(struct task_group *tg)
7543 unsigned long flags;
7546 /* end participation in shares distribution */
7547 for_each_possible_cpu(i)
7548 unregister_fair_sched_group(tg, i);
7550 spin_lock_irqsave(&task_group_lock, flags);
7551 list_del_rcu(&tg->list);
7552 list_del_rcu(&tg->siblings);
7553 spin_unlock_irqrestore(&task_group_lock, flags);
7556 /* change task's runqueue when it moves between groups.
7557 * The caller of this function should have put the task in its new group
7558 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7559 * reflect its new group.
7561 void sched_move_task(struct task_struct *tsk)
7563 struct task_group *tg;
7564 int queued, running;
7565 unsigned long flags;
7568 rq = task_rq_lock(tsk, &flags);
7570 running = task_current(rq, tsk);
7571 queued = task_on_rq_queued(tsk);
7574 dequeue_task(rq, tsk, 0);
7575 if (unlikely(running))
7576 put_prev_task(rq, tsk);
7579 * All callers are synchronized by task_rq_lock(); we do not use RCU
7580 * which is pointless here. Thus, we pass "true" to task_css_check()
7581 * to prevent lockdep warnings.
7583 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7584 struct task_group, css);
7585 tg = autogroup_task_group(tsk, tg);
7586 tsk->sched_task_group = tg;
7588 #ifdef CONFIG_FAIR_GROUP_SCHED
7589 if (tsk->sched_class->task_move_group)
7590 tsk->sched_class->task_move_group(tsk, queued);
7593 set_task_rq(tsk, task_cpu(tsk));
7595 if (unlikely(running))
7596 tsk->sched_class->set_curr_task(rq);
7598 enqueue_task(rq, tsk, 0);
7600 task_rq_unlock(rq, tsk, &flags);
7602 #endif /* CONFIG_CGROUP_SCHED */
7604 #ifdef CONFIG_RT_GROUP_SCHED
7606 * Ensure that the real time constraints are schedulable.
7608 static DEFINE_MUTEX(rt_constraints_mutex);
7610 /* Must be called with tasklist_lock held */
7611 static inline int tg_has_rt_tasks(struct task_group *tg)
7613 struct task_struct *g, *p;
7616 * Autogroups do not have RT tasks; see autogroup_create().
7618 if (task_group_is_autogroup(tg))
7621 for_each_process_thread(g, p) {
7622 if (rt_task(p) && task_group(p) == tg)
7629 struct rt_schedulable_data {
7630 struct task_group *tg;
7635 static int tg_rt_schedulable(struct task_group *tg, void *data)
7637 struct rt_schedulable_data *d = data;
7638 struct task_group *child;
7639 unsigned long total, sum = 0;
7640 u64 period, runtime;
7642 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7643 runtime = tg->rt_bandwidth.rt_runtime;
7646 period = d->rt_period;
7647 runtime = d->rt_runtime;
7651 * Cannot have more runtime than the period.
7653 if (runtime > period && runtime != RUNTIME_INF)
7657 * Ensure we don't starve existing RT tasks.
7659 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7662 total = to_ratio(period, runtime);
7665 * Nobody can have more than the global setting allows.
7667 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7671 * The sum of our children's runtime should not exceed our own.
7673 list_for_each_entry_rcu(child, &tg->children, siblings) {
7674 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7675 runtime = child->rt_bandwidth.rt_runtime;
7677 if (child == d->tg) {
7678 period = d->rt_period;
7679 runtime = d->rt_runtime;
7682 sum += to_ratio(period, runtime);
7691 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7695 struct rt_schedulable_data data = {
7697 .rt_period = period,
7698 .rt_runtime = runtime,
7702 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7708 static int tg_set_rt_bandwidth(struct task_group *tg,
7709 u64 rt_period, u64 rt_runtime)
7714 * Disallowing the root group RT runtime is BAD, it would disallow the
7715 * kernel creating (and or operating) RT threads.
7717 if (tg == &root_task_group && rt_runtime == 0)
7720 /* No period doesn't make any sense. */
7724 mutex_lock(&rt_constraints_mutex);
7725 read_lock(&tasklist_lock);
7726 err = __rt_schedulable(tg, rt_period, rt_runtime);
7730 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7731 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7732 tg->rt_bandwidth.rt_runtime = rt_runtime;
7734 for_each_possible_cpu(i) {
7735 struct rt_rq *rt_rq = tg->rt_rq[i];
7737 raw_spin_lock(&rt_rq->rt_runtime_lock);
7738 rt_rq->rt_runtime = rt_runtime;
7739 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7741 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7743 read_unlock(&tasklist_lock);
7744 mutex_unlock(&rt_constraints_mutex);
7749 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7751 u64 rt_runtime, rt_period;
7753 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7754 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7755 if (rt_runtime_us < 0)
7756 rt_runtime = RUNTIME_INF;
7758 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7761 static long sched_group_rt_runtime(struct task_group *tg)
7765 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7768 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7769 do_div(rt_runtime_us, NSEC_PER_USEC);
7770 return rt_runtime_us;
7773 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7775 u64 rt_runtime, rt_period;
7777 rt_period = rt_period_us * NSEC_PER_USEC;
7778 rt_runtime = tg->rt_bandwidth.rt_runtime;
7780 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7783 static long sched_group_rt_period(struct task_group *tg)
7787 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7788 do_div(rt_period_us, NSEC_PER_USEC);
7789 return rt_period_us;
7791 #endif /* CONFIG_RT_GROUP_SCHED */
7793 #ifdef CONFIG_RT_GROUP_SCHED
7794 static int sched_rt_global_constraints(void)
7798 mutex_lock(&rt_constraints_mutex);
7799 read_lock(&tasklist_lock);
7800 ret = __rt_schedulable(NULL, 0, 0);
7801 read_unlock(&tasklist_lock);
7802 mutex_unlock(&rt_constraints_mutex);
7807 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7809 /* Don't accept realtime tasks when there is no way for them to run */
7810 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7816 #else /* !CONFIG_RT_GROUP_SCHED */
7817 static int sched_rt_global_constraints(void)
7819 unsigned long flags;
7822 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7823 for_each_possible_cpu(i) {
7824 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7826 raw_spin_lock(&rt_rq->rt_runtime_lock);
7827 rt_rq->rt_runtime = global_rt_runtime();
7828 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7830 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7834 #endif /* CONFIG_RT_GROUP_SCHED */
7836 static int sched_dl_global_validate(void)
7838 u64 runtime = global_rt_runtime();
7839 u64 period = global_rt_period();
7840 u64 new_bw = to_ratio(period, runtime);
7843 unsigned long flags;
7846 * Here we want to check the bandwidth not being set to some
7847 * value smaller than the currently allocated bandwidth in
7848 * any of the root_domains.
7850 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7851 * cycling on root_domains... Discussion on different/better
7852 * solutions is welcome!
7854 for_each_possible_cpu(cpu) {
7855 rcu_read_lock_sched();
7856 dl_b = dl_bw_of(cpu);
7858 raw_spin_lock_irqsave(&dl_b->lock, flags);
7859 if (new_bw < dl_b->total_bw)
7861 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7863 rcu_read_unlock_sched();
7872 static void sched_dl_do_global(void)
7877 unsigned long flags;
7879 def_dl_bandwidth.dl_period = global_rt_period();
7880 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7882 if (global_rt_runtime() != RUNTIME_INF)
7883 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7886 * FIXME: As above...
7888 for_each_possible_cpu(cpu) {
7889 rcu_read_lock_sched();
7890 dl_b = dl_bw_of(cpu);
7892 raw_spin_lock_irqsave(&dl_b->lock, flags);
7894 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7896 rcu_read_unlock_sched();
7900 static int sched_rt_global_validate(void)
7902 if (sysctl_sched_rt_period <= 0)
7905 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7906 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7912 static void sched_rt_do_global(void)
7914 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7915 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7918 int sched_rt_handler(struct ctl_table *table, int write,
7919 void __user *buffer, size_t *lenp,
7922 int old_period, old_runtime;
7923 static DEFINE_MUTEX(mutex);
7927 old_period = sysctl_sched_rt_period;
7928 old_runtime = sysctl_sched_rt_runtime;
7930 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7932 if (!ret && write) {
7933 ret = sched_rt_global_validate();
7937 ret = sched_dl_global_validate();
7941 ret = sched_rt_global_constraints();
7945 sched_rt_do_global();
7946 sched_dl_do_global();
7950 sysctl_sched_rt_period = old_period;
7951 sysctl_sched_rt_runtime = old_runtime;
7953 mutex_unlock(&mutex);
7958 int sched_rr_handler(struct ctl_table *table, int write,
7959 void __user *buffer, size_t *lenp,
7963 static DEFINE_MUTEX(mutex);
7966 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7967 /* make sure that internally we keep jiffies */
7968 /* also, writing zero resets timeslice to default */
7969 if (!ret && write) {
7970 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7971 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7973 mutex_unlock(&mutex);
7977 #ifdef CONFIG_CGROUP_SCHED
7979 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7981 return css ? container_of(css, struct task_group, css) : NULL;
7984 static struct cgroup_subsys_state *
7985 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7987 struct task_group *parent = css_tg(parent_css);
7988 struct task_group *tg;
7991 /* This is early initialization for the top cgroup */
7992 return &root_task_group.css;
7995 tg = sched_create_group(parent);
7997 return ERR_PTR(-ENOMEM);
8002 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8004 struct task_group *tg = css_tg(css);
8005 struct task_group *parent = css_tg(css->parent);
8008 sched_online_group(tg, parent);
8012 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8014 struct task_group *tg = css_tg(css);
8016 sched_destroy_group(tg);
8019 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8021 struct task_group *tg = css_tg(css);
8023 sched_offline_group(tg);
8026 static void cpu_cgroup_fork(struct task_struct *task)
8028 sched_move_task(task);
8031 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8032 struct cgroup_taskset *tset)
8034 struct task_struct *task;
8036 cgroup_taskset_for_each(task, tset) {
8037 #ifdef CONFIG_RT_GROUP_SCHED
8038 if (!sched_rt_can_attach(css_tg(css), task))
8041 /* We don't support RT-tasks being in separate groups */
8042 if (task->sched_class != &fair_sched_class)
8049 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8050 struct cgroup_taskset *tset)
8052 struct task_struct *task;
8054 cgroup_taskset_for_each(task, tset)
8055 sched_move_task(task);
8058 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8059 struct cgroup_subsys_state *old_css,
8060 struct task_struct *task)
8063 * cgroup_exit() is called in the copy_process() failure path.
8064 * Ignore this case since the task hasn't ran yet, this avoids
8065 * trying to poke a half freed task state from generic code.
8067 if (!(task->flags & PF_EXITING))
8070 sched_move_task(task);
8073 #ifdef CONFIG_FAIR_GROUP_SCHED
8074 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8075 struct cftype *cftype, u64 shareval)
8077 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8080 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8083 struct task_group *tg = css_tg(css);
8085 return (u64) scale_load_down(tg->shares);
8088 #ifdef CONFIG_CFS_BANDWIDTH
8089 static DEFINE_MUTEX(cfs_constraints_mutex);
8091 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8092 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8094 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8096 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8098 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8099 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8101 if (tg == &root_task_group)
8105 * Ensure we have at some amount of bandwidth every period. This is
8106 * to prevent reaching a state of large arrears when throttled via
8107 * entity_tick() resulting in prolonged exit starvation.
8109 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8113 * Likewise, bound things on the otherside by preventing insane quota
8114 * periods. This also allows us to normalize in computing quota
8117 if (period > max_cfs_quota_period)
8121 * Prevent race between setting of cfs_rq->runtime_enabled and
8122 * unthrottle_offline_cfs_rqs().
8125 mutex_lock(&cfs_constraints_mutex);
8126 ret = __cfs_schedulable(tg, period, quota);
8130 runtime_enabled = quota != RUNTIME_INF;
8131 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8133 * If we need to toggle cfs_bandwidth_used, off->on must occur
8134 * before making related changes, and on->off must occur afterwards
8136 if (runtime_enabled && !runtime_was_enabled)
8137 cfs_bandwidth_usage_inc();
8138 raw_spin_lock_irq(&cfs_b->lock);
8139 cfs_b->period = ns_to_ktime(period);
8140 cfs_b->quota = quota;
8142 __refill_cfs_bandwidth_runtime(cfs_b);
8143 /* restart the period timer (if active) to handle new period expiry */
8144 if (runtime_enabled)
8145 start_cfs_bandwidth(cfs_b);
8146 raw_spin_unlock_irq(&cfs_b->lock);
8148 for_each_online_cpu(i) {
8149 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8150 struct rq *rq = cfs_rq->rq;
8152 raw_spin_lock_irq(&rq->lock);
8153 cfs_rq->runtime_enabled = runtime_enabled;
8154 cfs_rq->runtime_remaining = 0;
8156 if (cfs_rq->throttled)
8157 unthrottle_cfs_rq(cfs_rq);
8158 raw_spin_unlock_irq(&rq->lock);
8160 if (runtime_was_enabled && !runtime_enabled)
8161 cfs_bandwidth_usage_dec();
8163 mutex_unlock(&cfs_constraints_mutex);
8169 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8173 period = ktime_to_ns(tg->cfs_bandwidth.period);
8174 if (cfs_quota_us < 0)
8175 quota = RUNTIME_INF;
8177 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8179 return tg_set_cfs_bandwidth(tg, period, quota);
8182 long tg_get_cfs_quota(struct task_group *tg)
8186 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8189 quota_us = tg->cfs_bandwidth.quota;
8190 do_div(quota_us, NSEC_PER_USEC);
8195 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8199 period = (u64)cfs_period_us * NSEC_PER_USEC;
8200 quota = tg->cfs_bandwidth.quota;
8202 return tg_set_cfs_bandwidth(tg, period, quota);
8205 long tg_get_cfs_period(struct task_group *tg)
8209 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8210 do_div(cfs_period_us, NSEC_PER_USEC);
8212 return cfs_period_us;
8215 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8218 return tg_get_cfs_quota(css_tg(css));
8221 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8222 struct cftype *cftype, s64 cfs_quota_us)
8224 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8227 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8230 return tg_get_cfs_period(css_tg(css));
8233 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8234 struct cftype *cftype, u64 cfs_period_us)
8236 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8239 struct cfs_schedulable_data {
8240 struct task_group *tg;
8245 * normalize group quota/period to be quota/max_period
8246 * note: units are usecs
8248 static u64 normalize_cfs_quota(struct task_group *tg,
8249 struct cfs_schedulable_data *d)
8257 period = tg_get_cfs_period(tg);
8258 quota = tg_get_cfs_quota(tg);
8261 /* note: these should typically be equivalent */
8262 if (quota == RUNTIME_INF || quota == -1)
8265 return to_ratio(period, quota);
8268 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8270 struct cfs_schedulable_data *d = data;
8271 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8272 s64 quota = 0, parent_quota = -1;
8275 quota = RUNTIME_INF;
8277 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8279 quota = normalize_cfs_quota(tg, d);
8280 parent_quota = parent_b->hierarchical_quota;
8283 * ensure max(child_quota) <= parent_quota, inherit when no
8286 if (quota == RUNTIME_INF)
8287 quota = parent_quota;
8288 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8291 cfs_b->hierarchical_quota = quota;
8296 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8299 struct cfs_schedulable_data data = {
8305 if (quota != RUNTIME_INF) {
8306 do_div(data.period, NSEC_PER_USEC);
8307 do_div(data.quota, NSEC_PER_USEC);
8311 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8317 static int cpu_stats_show(struct seq_file *sf, void *v)
8319 struct task_group *tg = css_tg(seq_css(sf));
8320 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8322 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8323 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8324 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8328 #endif /* CONFIG_CFS_BANDWIDTH */
8329 #endif /* CONFIG_FAIR_GROUP_SCHED */
8331 #ifdef CONFIG_RT_GROUP_SCHED
8332 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8333 struct cftype *cft, s64 val)
8335 return sched_group_set_rt_runtime(css_tg(css), val);
8338 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8341 return sched_group_rt_runtime(css_tg(css));
8344 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8345 struct cftype *cftype, u64 rt_period_us)
8347 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8350 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8353 return sched_group_rt_period(css_tg(css));
8355 #endif /* CONFIG_RT_GROUP_SCHED */
8357 static struct cftype cpu_files[] = {
8358 #ifdef CONFIG_FAIR_GROUP_SCHED
8361 .read_u64 = cpu_shares_read_u64,
8362 .write_u64 = cpu_shares_write_u64,
8365 #ifdef CONFIG_CFS_BANDWIDTH
8367 .name = "cfs_quota_us",
8368 .read_s64 = cpu_cfs_quota_read_s64,
8369 .write_s64 = cpu_cfs_quota_write_s64,
8372 .name = "cfs_period_us",
8373 .read_u64 = cpu_cfs_period_read_u64,
8374 .write_u64 = cpu_cfs_period_write_u64,
8378 .seq_show = cpu_stats_show,
8381 #ifdef CONFIG_RT_GROUP_SCHED
8383 .name = "rt_runtime_us",
8384 .read_s64 = cpu_rt_runtime_read,
8385 .write_s64 = cpu_rt_runtime_write,
8388 .name = "rt_period_us",
8389 .read_u64 = cpu_rt_period_read_uint,
8390 .write_u64 = cpu_rt_period_write_uint,
8396 struct cgroup_subsys cpu_cgrp_subsys = {
8397 .css_alloc = cpu_cgroup_css_alloc,
8398 .css_free = cpu_cgroup_css_free,
8399 .css_online = cpu_cgroup_css_online,
8400 .css_offline = cpu_cgroup_css_offline,
8401 .fork = cpu_cgroup_fork,
8402 .can_attach = cpu_cgroup_can_attach,
8403 .attach = cpu_cgroup_attach,
8404 .exit = cpu_cgroup_exit,
8405 .legacy_cftypes = cpu_files,
8409 #endif /* CONFIG_CGROUP_SCHED */
8411 void dump_cpu_task(int cpu)
8413 pr_info("Task dump for CPU %d:\n", cpu);
8414 sched_show_task(cpu_curr(cpu));