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 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1004 * use the balance_callback list if you want balancing.
1006 * this means any call to check_class_changed() must be followed by a call to
1007 * balance_callback().
1009 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1010 const struct sched_class *prev_class,
1013 if (prev_class != p->sched_class) {
1014 if (prev_class->switched_from)
1015 prev_class->switched_from(rq, p);
1017 p->sched_class->switched_to(rq, p);
1018 } else if (oldprio != p->prio || dl_task(p))
1019 p->sched_class->prio_changed(rq, p, oldprio);
1022 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1024 const struct sched_class *class;
1026 if (p->sched_class == rq->curr->sched_class) {
1027 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1029 for_each_class(class) {
1030 if (class == rq->curr->sched_class)
1032 if (class == p->sched_class) {
1040 * A queue event has occurred, and we're going to schedule. In
1041 * this case, we can save a useless back to back clock update.
1043 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1044 rq_clock_skip_update(rq, true);
1049 * This is how migration works:
1051 * 1) we invoke migration_cpu_stop() on the target CPU using
1053 * 2) stopper starts to run (implicitly forcing the migrated thread
1055 * 3) it checks whether the migrated task is still in the wrong runqueue.
1056 * 4) if it's in the wrong runqueue then the migration thread removes
1057 * it and puts it into the right queue.
1058 * 5) stopper completes and stop_one_cpu() returns and the migration
1063 * move_queued_task - move a queued task to new rq.
1065 * Returns (locked) new rq. Old rq's lock is released.
1067 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1069 lockdep_assert_held(&rq->lock);
1071 dequeue_task(rq, p, 0);
1072 p->on_rq = TASK_ON_RQ_MIGRATING;
1073 set_task_cpu(p, new_cpu);
1074 raw_spin_unlock(&rq->lock);
1076 rq = cpu_rq(new_cpu);
1078 raw_spin_lock(&rq->lock);
1079 BUG_ON(task_cpu(p) != new_cpu);
1080 p->on_rq = TASK_ON_RQ_QUEUED;
1081 enqueue_task(rq, p, 0);
1082 check_preempt_curr(rq, p, 0);
1087 struct migration_arg {
1088 struct task_struct *task;
1093 * Move (not current) task off this cpu, onto dest cpu. We're doing
1094 * this because either it can't run here any more (set_cpus_allowed()
1095 * away from this CPU, or CPU going down), or because we're
1096 * attempting to rebalance this task on exec (sched_exec).
1098 * So we race with normal scheduler movements, but that's OK, as long
1099 * as the task is no longer on this CPU.
1101 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1103 if (unlikely(!cpu_active(dest_cpu)))
1106 /* Affinity changed (again). */
1107 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1110 rq = move_queued_task(rq, p, dest_cpu);
1116 * migration_cpu_stop - this will be executed by a highprio stopper thread
1117 * and performs thread migration by bumping thread off CPU then
1118 * 'pushing' onto another runqueue.
1120 static int migration_cpu_stop(void *data)
1122 struct migration_arg *arg = data;
1123 struct task_struct *p = arg->task;
1124 struct rq *rq = this_rq();
1127 * The original target cpu might have gone down and we might
1128 * be on another cpu but it doesn't matter.
1130 local_irq_disable();
1132 * We need to explicitly wake pending tasks before running
1133 * __migrate_task() such that we will not miss enforcing cpus_allowed
1134 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1136 sched_ttwu_pending();
1138 raw_spin_lock(&p->pi_lock);
1139 raw_spin_lock(&rq->lock);
1141 * If task_rq(p) != rq, it cannot be migrated here, because we're
1142 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1143 * we're holding p->pi_lock.
1145 if (task_rq(p) == rq && task_on_rq_queued(p))
1146 rq = __migrate_task(rq, p, arg->dest_cpu);
1147 raw_spin_unlock(&rq->lock);
1148 raw_spin_unlock(&p->pi_lock);
1154 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1156 if (p->sched_class->set_cpus_allowed)
1157 p->sched_class->set_cpus_allowed(p, new_mask);
1159 cpumask_copy(&p->cpus_allowed, new_mask);
1160 p->nr_cpus_allowed = cpumask_weight(new_mask);
1164 * Change a given task's CPU affinity. Migrate the thread to a
1165 * proper CPU and schedule it away if the CPU it's executing on
1166 * is removed from the allowed bitmask.
1168 * NOTE: the caller must have a valid reference to the task, the
1169 * task must not exit() & deallocate itself prematurely. The
1170 * call is not atomic; no spinlocks may be held.
1172 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1174 unsigned long flags;
1176 unsigned int dest_cpu;
1179 rq = task_rq_lock(p, &flags);
1181 if (cpumask_equal(&p->cpus_allowed, new_mask))
1184 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1189 do_set_cpus_allowed(p, new_mask);
1191 /* Can the task run on the task's current CPU? If so, we're done */
1192 if (cpumask_test_cpu(task_cpu(p), new_mask))
1195 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1196 if (task_running(rq, p) || p->state == TASK_WAKING) {
1197 struct migration_arg arg = { p, dest_cpu };
1198 /* Need help from migration thread: drop lock and wait. */
1199 task_rq_unlock(rq, p, &flags);
1200 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1201 tlb_migrate_finish(p->mm);
1203 } else if (task_on_rq_queued(p)) {
1205 * OK, since we're going to drop the lock immediately
1206 * afterwards anyway.
1208 lockdep_unpin_lock(&rq->lock);
1209 rq = move_queued_task(rq, p, dest_cpu);
1210 lockdep_pin_lock(&rq->lock);
1213 task_rq_unlock(rq, p, &flags);
1217 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1219 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1221 #ifdef CONFIG_SCHED_DEBUG
1223 * We should never call set_task_cpu() on a blocked task,
1224 * ttwu() will sort out the placement.
1226 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1229 #ifdef CONFIG_LOCKDEP
1231 * The caller should hold either p->pi_lock or rq->lock, when changing
1232 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1234 * sched_move_task() holds both and thus holding either pins the cgroup,
1237 * Furthermore, all task_rq users should acquire both locks, see
1240 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1241 lockdep_is_held(&task_rq(p)->lock)));
1245 trace_sched_migrate_task(p, new_cpu);
1247 if (task_cpu(p) != new_cpu) {
1248 if (p->sched_class->migrate_task_rq)
1249 p->sched_class->migrate_task_rq(p, new_cpu);
1250 p->se.nr_migrations++;
1251 perf_event_task_migrate(p);
1254 __set_task_cpu(p, new_cpu);
1257 static void __migrate_swap_task(struct task_struct *p, int cpu)
1259 if (task_on_rq_queued(p)) {
1260 struct rq *src_rq, *dst_rq;
1262 src_rq = task_rq(p);
1263 dst_rq = cpu_rq(cpu);
1265 deactivate_task(src_rq, p, 0);
1266 set_task_cpu(p, cpu);
1267 activate_task(dst_rq, p, 0);
1268 check_preempt_curr(dst_rq, p, 0);
1271 * Task isn't running anymore; make it appear like we migrated
1272 * it before it went to sleep. This means on wakeup we make the
1273 * previous cpu our targer instead of where it really is.
1279 struct migration_swap_arg {
1280 struct task_struct *src_task, *dst_task;
1281 int src_cpu, dst_cpu;
1284 static int migrate_swap_stop(void *data)
1286 struct migration_swap_arg *arg = data;
1287 struct rq *src_rq, *dst_rq;
1290 src_rq = cpu_rq(arg->src_cpu);
1291 dst_rq = cpu_rq(arg->dst_cpu);
1293 double_raw_lock(&arg->src_task->pi_lock,
1294 &arg->dst_task->pi_lock);
1295 double_rq_lock(src_rq, dst_rq);
1296 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1299 if (task_cpu(arg->src_task) != arg->src_cpu)
1302 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1305 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1308 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1309 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1314 double_rq_unlock(src_rq, dst_rq);
1315 raw_spin_unlock(&arg->dst_task->pi_lock);
1316 raw_spin_unlock(&arg->src_task->pi_lock);
1322 * Cross migrate two tasks
1324 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1326 struct migration_swap_arg arg;
1329 arg = (struct migration_swap_arg){
1331 .src_cpu = task_cpu(cur),
1333 .dst_cpu = task_cpu(p),
1336 if (arg.src_cpu == arg.dst_cpu)
1340 * These three tests are all lockless; this is OK since all of them
1341 * will be re-checked with proper locks held further down the line.
1343 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1346 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1349 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1352 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1353 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1360 * wait_task_inactive - wait for a thread to unschedule.
1362 * If @match_state is nonzero, it's the @p->state value just checked and
1363 * not expected to change. If it changes, i.e. @p might have woken up,
1364 * then return zero. When we succeed in waiting for @p to be off its CPU,
1365 * we return a positive number (its total switch count). If a second call
1366 * a short while later returns the same number, the caller can be sure that
1367 * @p has remained unscheduled the whole time.
1369 * The caller must ensure that the task *will* unschedule sometime soon,
1370 * else this function might spin for a *long* time. This function can't
1371 * be called with interrupts off, or it may introduce deadlock with
1372 * smp_call_function() if an IPI is sent by the same process we are
1373 * waiting to become inactive.
1375 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1377 unsigned long flags;
1378 int running, queued;
1384 * We do the initial early heuristics without holding
1385 * any task-queue locks at all. We'll only try to get
1386 * the runqueue lock when things look like they will
1392 * If the task is actively running on another CPU
1393 * still, just relax and busy-wait without holding
1396 * NOTE! Since we don't hold any locks, it's not
1397 * even sure that "rq" stays as the right runqueue!
1398 * But we don't care, since "task_running()" will
1399 * return false if the runqueue has changed and p
1400 * is actually now running somewhere else!
1402 while (task_running(rq, p)) {
1403 if (match_state && unlikely(p->state != match_state))
1409 * Ok, time to look more closely! We need the rq
1410 * lock now, to be *sure*. If we're wrong, we'll
1411 * just go back and repeat.
1413 rq = task_rq_lock(p, &flags);
1414 trace_sched_wait_task(p);
1415 running = task_running(rq, p);
1416 queued = task_on_rq_queued(p);
1418 if (!match_state || p->state == match_state)
1419 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1420 task_rq_unlock(rq, p, &flags);
1423 * If it changed from the expected state, bail out now.
1425 if (unlikely(!ncsw))
1429 * Was it really running after all now that we
1430 * checked with the proper locks actually held?
1432 * Oops. Go back and try again..
1434 if (unlikely(running)) {
1440 * It's not enough that it's not actively running,
1441 * it must be off the runqueue _entirely_, and not
1444 * So if it was still runnable (but just not actively
1445 * running right now), it's preempted, and we should
1446 * yield - it could be a while.
1448 if (unlikely(queued)) {
1449 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1451 set_current_state(TASK_UNINTERRUPTIBLE);
1452 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1457 * Ahh, all good. It wasn't running, and it wasn't
1458 * runnable, which means that it will never become
1459 * running in the future either. We're all done!
1468 * kick_process - kick a running thread to enter/exit the kernel
1469 * @p: the to-be-kicked thread
1471 * Cause a process which is running on another CPU to enter
1472 * kernel-mode, without any delay. (to get signals handled.)
1474 * NOTE: this function doesn't have to take the runqueue lock,
1475 * because all it wants to ensure is that the remote task enters
1476 * the kernel. If the IPI races and the task has been migrated
1477 * to another CPU then no harm is done and the purpose has been
1480 void kick_process(struct task_struct *p)
1486 if ((cpu != smp_processor_id()) && task_curr(p))
1487 smp_send_reschedule(cpu);
1490 EXPORT_SYMBOL_GPL(kick_process);
1493 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1495 static int select_fallback_rq(int cpu, struct task_struct *p)
1497 int nid = cpu_to_node(cpu);
1498 const struct cpumask *nodemask = NULL;
1499 enum { cpuset, possible, fail } state = cpuset;
1503 * If the node that the cpu is on has been offlined, cpu_to_node()
1504 * will return -1. There is no cpu on the node, and we should
1505 * select the cpu on the other node.
1508 nodemask = cpumask_of_node(nid);
1510 /* Look for allowed, online CPU in same node. */
1511 for_each_cpu(dest_cpu, nodemask) {
1512 if (!cpu_online(dest_cpu))
1514 if (!cpu_active(dest_cpu))
1516 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1522 /* Any allowed, online CPU? */
1523 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1524 if (!cpu_online(dest_cpu))
1526 if (!cpu_active(dest_cpu))
1533 /* No more Mr. Nice Guy. */
1534 cpuset_cpus_allowed_fallback(p);
1539 do_set_cpus_allowed(p, cpu_possible_mask);
1550 if (state != cpuset) {
1552 * Don't tell them about moving exiting tasks or
1553 * kernel threads (both mm NULL), since they never
1556 if (p->mm && printk_ratelimit()) {
1557 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1558 task_pid_nr(p), p->comm, cpu);
1566 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1569 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1571 lockdep_assert_held(&p->pi_lock);
1573 if (p->nr_cpus_allowed > 1)
1574 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1577 * In order not to call set_task_cpu() on a blocking task we need
1578 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1581 * Since this is common to all placement strategies, this lives here.
1583 * [ this allows ->select_task() to simply return task_cpu(p) and
1584 * not worry about this generic constraint ]
1586 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1588 cpu = select_fallback_rq(task_cpu(p), p);
1593 static void update_avg(u64 *avg, u64 sample)
1595 s64 diff = sample - *avg;
1598 #endif /* CONFIG_SMP */
1601 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1603 #ifdef CONFIG_SCHEDSTATS
1604 struct rq *rq = this_rq();
1607 int this_cpu = smp_processor_id();
1609 if (cpu == this_cpu) {
1610 schedstat_inc(rq, ttwu_local);
1611 schedstat_inc(p, se.statistics.nr_wakeups_local);
1613 struct sched_domain *sd;
1615 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1617 for_each_domain(this_cpu, sd) {
1618 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1619 schedstat_inc(sd, ttwu_wake_remote);
1626 if (wake_flags & WF_MIGRATED)
1627 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1629 #endif /* CONFIG_SMP */
1631 schedstat_inc(rq, ttwu_count);
1632 schedstat_inc(p, se.statistics.nr_wakeups);
1634 if (wake_flags & WF_SYNC)
1635 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1637 #endif /* CONFIG_SCHEDSTATS */
1640 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1642 activate_task(rq, p, en_flags);
1643 p->on_rq = TASK_ON_RQ_QUEUED;
1645 /* if a worker is waking up, notify workqueue */
1646 if (p->flags & PF_WQ_WORKER)
1647 wq_worker_waking_up(p, cpu_of(rq));
1651 * Mark the task runnable and perform wakeup-preemption.
1654 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1656 check_preempt_curr(rq, p, wake_flags);
1657 trace_sched_wakeup(p, true);
1659 p->state = TASK_RUNNING;
1661 if (p->sched_class->task_woken) {
1663 * Our task @p is fully woken up and running; so its safe to
1664 * drop the rq->lock, hereafter rq is only used for statistics.
1666 lockdep_unpin_lock(&rq->lock);
1667 p->sched_class->task_woken(rq, p);
1668 lockdep_pin_lock(&rq->lock);
1671 if (rq->idle_stamp) {
1672 u64 delta = rq_clock(rq) - rq->idle_stamp;
1673 u64 max = 2*rq->max_idle_balance_cost;
1675 update_avg(&rq->avg_idle, delta);
1677 if (rq->avg_idle > max)
1686 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1688 lockdep_assert_held(&rq->lock);
1691 if (p->sched_contributes_to_load)
1692 rq->nr_uninterruptible--;
1695 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1696 ttwu_do_wakeup(rq, p, wake_flags);
1700 * Called in case the task @p isn't fully descheduled from its runqueue,
1701 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1702 * since all we need to do is flip p->state to TASK_RUNNING, since
1703 * the task is still ->on_rq.
1705 static int ttwu_remote(struct task_struct *p, int wake_flags)
1710 rq = __task_rq_lock(p);
1711 if (task_on_rq_queued(p)) {
1712 /* check_preempt_curr() may use rq clock */
1713 update_rq_clock(rq);
1714 ttwu_do_wakeup(rq, p, wake_flags);
1717 __task_rq_unlock(rq);
1723 void sched_ttwu_pending(void)
1725 struct rq *rq = this_rq();
1726 struct llist_node *llist = llist_del_all(&rq->wake_list);
1727 struct task_struct *p;
1728 unsigned long flags;
1733 raw_spin_lock_irqsave(&rq->lock, flags);
1734 lockdep_pin_lock(&rq->lock);
1737 p = llist_entry(llist, struct task_struct, wake_entry);
1738 llist = llist_next(llist);
1739 ttwu_do_activate(rq, p, 0);
1742 lockdep_unpin_lock(&rq->lock);
1743 raw_spin_unlock_irqrestore(&rq->lock, flags);
1746 void scheduler_ipi(void)
1749 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1750 * TIF_NEED_RESCHED remotely (for the first time) will also send
1753 preempt_fold_need_resched();
1755 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1759 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1760 * traditionally all their work was done from the interrupt return
1761 * path. Now that we actually do some work, we need to make sure
1764 * Some archs already do call them, luckily irq_enter/exit nest
1767 * Arguably we should visit all archs and update all handlers,
1768 * however a fair share of IPIs are still resched only so this would
1769 * somewhat pessimize the simple resched case.
1772 sched_ttwu_pending();
1775 * Check if someone kicked us for doing the nohz idle load balance.
1777 if (unlikely(got_nohz_idle_kick())) {
1778 this_rq()->idle_balance = 1;
1779 raise_softirq_irqoff(SCHED_SOFTIRQ);
1784 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1786 struct rq *rq = cpu_rq(cpu);
1788 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1789 if (!set_nr_if_polling(rq->idle))
1790 smp_send_reschedule(cpu);
1792 trace_sched_wake_idle_without_ipi(cpu);
1796 void wake_up_if_idle(int cpu)
1798 struct rq *rq = cpu_rq(cpu);
1799 unsigned long flags;
1803 if (!is_idle_task(rcu_dereference(rq->curr)))
1806 if (set_nr_if_polling(rq->idle)) {
1807 trace_sched_wake_idle_without_ipi(cpu);
1809 raw_spin_lock_irqsave(&rq->lock, flags);
1810 if (is_idle_task(rq->curr))
1811 smp_send_reschedule(cpu);
1812 /* Else cpu is not in idle, do nothing here */
1813 raw_spin_unlock_irqrestore(&rq->lock, flags);
1820 bool cpus_share_cache(int this_cpu, int that_cpu)
1822 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1824 #endif /* CONFIG_SMP */
1826 static void ttwu_queue(struct task_struct *p, int cpu)
1828 struct rq *rq = cpu_rq(cpu);
1830 #if defined(CONFIG_SMP)
1831 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1832 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1833 ttwu_queue_remote(p, cpu);
1838 raw_spin_lock(&rq->lock);
1839 lockdep_pin_lock(&rq->lock);
1840 ttwu_do_activate(rq, p, 0);
1841 lockdep_unpin_lock(&rq->lock);
1842 raw_spin_unlock(&rq->lock);
1846 * try_to_wake_up - wake up a thread
1847 * @p: the thread to be awakened
1848 * @state: the mask of task states that can be woken
1849 * @wake_flags: wake modifier flags (WF_*)
1851 * Put it on the run-queue if it's not already there. The "current"
1852 * thread is always on the run-queue (except when the actual
1853 * re-schedule is in progress), and as such you're allowed to do
1854 * the simpler "current->state = TASK_RUNNING" to mark yourself
1855 * runnable without the overhead of this.
1857 * Return: %true if @p was woken up, %false if it was already running.
1858 * or @state didn't match @p's state.
1861 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1863 unsigned long flags;
1864 int cpu, success = 0;
1867 * If we are going to wake up a thread waiting for CONDITION we
1868 * need to ensure that CONDITION=1 done by the caller can not be
1869 * reordered with p->state check below. This pairs with mb() in
1870 * set_current_state() the waiting thread does.
1872 smp_mb__before_spinlock();
1873 raw_spin_lock_irqsave(&p->pi_lock, flags);
1874 if (!(p->state & state))
1877 success = 1; /* we're going to change ->state */
1880 if (p->on_rq && ttwu_remote(p, wake_flags))
1885 * If the owning (remote) cpu is still in the middle of schedule() with
1886 * this task as prev, wait until its done referencing the task.
1891 * Pairs with the smp_wmb() in finish_lock_switch().
1895 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1896 p->state = TASK_WAKING;
1898 if (p->sched_class->task_waking)
1899 p->sched_class->task_waking(p);
1901 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1902 if (task_cpu(p) != cpu) {
1903 wake_flags |= WF_MIGRATED;
1904 set_task_cpu(p, cpu);
1906 #endif /* CONFIG_SMP */
1910 ttwu_stat(p, cpu, wake_flags);
1912 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1918 * try_to_wake_up_local - try to wake up a local task with rq lock held
1919 * @p: the thread to be awakened
1921 * Put @p on the run-queue if it's not already there. The caller must
1922 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1925 static void try_to_wake_up_local(struct task_struct *p)
1927 struct rq *rq = task_rq(p);
1929 if (WARN_ON_ONCE(rq != this_rq()) ||
1930 WARN_ON_ONCE(p == current))
1933 lockdep_assert_held(&rq->lock);
1935 if (!raw_spin_trylock(&p->pi_lock)) {
1937 * This is OK, because current is on_cpu, which avoids it being
1938 * picked for load-balance and preemption/IRQs are still
1939 * disabled avoiding further scheduler activity on it and we've
1940 * not yet picked a replacement task.
1942 lockdep_unpin_lock(&rq->lock);
1943 raw_spin_unlock(&rq->lock);
1944 raw_spin_lock(&p->pi_lock);
1945 raw_spin_lock(&rq->lock);
1946 lockdep_pin_lock(&rq->lock);
1949 if (!(p->state & TASK_NORMAL))
1952 if (!task_on_rq_queued(p))
1953 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1955 ttwu_do_wakeup(rq, p, 0);
1956 ttwu_stat(p, smp_processor_id(), 0);
1958 raw_spin_unlock(&p->pi_lock);
1962 * wake_up_process - Wake up a specific process
1963 * @p: The process to be woken up.
1965 * Attempt to wake up the nominated process and move it to the set of runnable
1968 * Return: 1 if the process was woken up, 0 if it was already running.
1970 * It may be assumed that this function implies a write memory barrier before
1971 * changing the task state if and only if any tasks are woken up.
1973 int wake_up_process(struct task_struct *p)
1975 WARN_ON(task_is_stopped_or_traced(p));
1976 return try_to_wake_up(p, TASK_NORMAL, 0);
1978 EXPORT_SYMBOL(wake_up_process);
1980 int wake_up_state(struct task_struct *p, unsigned int state)
1982 return try_to_wake_up(p, state, 0);
1986 * This function clears the sched_dl_entity static params.
1988 void __dl_clear_params(struct task_struct *p)
1990 struct sched_dl_entity *dl_se = &p->dl;
1992 dl_se->dl_runtime = 0;
1993 dl_se->dl_deadline = 0;
1994 dl_se->dl_period = 0;
1998 dl_se->dl_throttled = 0;
2000 dl_se->dl_yielded = 0;
2004 * Perform scheduler related setup for a newly forked process p.
2005 * p is forked by current.
2007 * __sched_fork() is basic setup used by init_idle() too:
2009 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2014 p->se.exec_start = 0;
2015 p->se.sum_exec_runtime = 0;
2016 p->se.prev_sum_exec_runtime = 0;
2017 p->se.nr_migrations = 0;
2020 p->se.avg.decay_count = 0;
2022 INIT_LIST_HEAD(&p->se.group_node);
2024 #ifdef CONFIG_SCHEDSTATS
2025 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2028 RB_CLEAR_NODE(&p->dl.rb_node);
2029 init_dl_task_timer(&p->dl);
2030 __dl_clear_params(p);
2032 INIT_LIST_HEAD(&p->rt.run_list);
2034 #ifdef CONFIG_PREEMPT_NOTIFIERS
2035 INIT_HLIST_HEAD(&p->preempt_notifiers);
2038 #ifdef CONFIG_NUMA_BALANCING
2039 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2040 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2041 p->mm->numa_scan_seq = 0;
2044 if (clone_flags & CLONE_VM)
2045 p->numa_preferred_nid = current->numa_preferred_nid;
2047 p->numa_preferred_nid = -1;
2049 p->node_stamp = 0ULL;
2050 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2051 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2052 p->numa_work.next = &p->numa_work;
2053 p->numa_faults = NULL;
2054 p->last_task_numa_placement = 0;
2055 p->last_sum_exec_runtime = 0;
2057 p->numa_group = NULL;
2058 #endif /* CONFIG_NUMA_BALANCING */
2061 #ifdef CONFIG_NUMA_BALANCING
2062 #ifdef CONFIG_SCHED_DEBUG
2063 void set_numabalancing_state(bool enabled)
2066 sched_feat_set("NUMA");
2068 sched_feat_set("NO_NUMA");
2071 __read_mostly bool numabalancing_enabled;
2073 void set_numabalancing_state(bool enabled)
2075 numabalancing_enabled = enabled;
2077 #endif /* CONFIG_SCHED_DEBUG */
2079 #ifdef CONFIG_PROC_SYSCTL
2080 int sysctl_numa_balancing(struct ctl_table *table, int write,
2081 void __user *buffer, size_t *lenp, loff_t *ppos)
2085 int state = numabalancing_enabled;
2087 if (write && !capable(CAP_SYS_ADMIN))
2092 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2096 set_numabalancing_state(state);
2103 * fork()/clone()-time setup:
2105 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2107 unsigned long flags;
2108 int cpu = get_cpu();
2110 __sched_fork(clone_flags, p);
2112 * We mark the process as running here. This guarantees that
2113 * nobody will actually run it, and a signal or other external
2114 * event cannot wake it up and insert it on the runqueue either.
2116 p->state = TASK_RUNNING;
2119 * Make sure we do not leak PI boosting priority to the child.
2121 p->prio = current->normal_prio;
2124 * Revert to default priority/policy on fork if requested.
2126 if (unlikely(p->sched_reset_on_fork)) {
2127 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2128 p->policy = SCHED_NORMAL;
2129 p->static_prio = NICE_TO_PRIO(0);
2131 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2132 p->static_prio = NICE_TO_PRIO(0);
2134 p->prio = p->normal_prio = __normal_prio(p);
2138 * We don't need the reset flag anymore after the fork. It has
2139 * fulfilled its duty:
2141 p->sched_reset_on_fork = 0;
2144 if (dl_prio(p->prio)) {
2147 } else if (rt_prio(p->prio)) {
2148 p->sched_class = &rt_sched_class;
2150 p->sched_class = &fair_sched_class;
2153 if (p->sched_class->task_fork)
2154 p->sched_class->task_fork(p);
2157 * The child is not yet in the pid-hash so no cgroup attach races,
2158 * and the cgroup is pinned to this child due to cgroup_fork()
2159 * is ran before sched_fork().
2161 * Silence PROVE_RCU.
2163 raw_spin_lock_irqsave(&p->pi_lock, flags);
2164 set_task_cpu(p, cpu);
2165 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2167 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2168 if (likely(sched_info_on()))
2169 memset(&p->sched_info, 0, sizeof(p->sched_info));
2171 #if defined(CONFIG_SMP)
2174 init_task_preempt_count(p);
2176 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2177 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2184 unsigned long to_ratio(u64 period, u64 runtime)
2186 if (runtime == RUNTIME_INF)
2190 * Doing this here saves a lot of checks in all
2191 * the calling paths, and returning zero seems
2192 * safe for them anyway.
2197 return div64_u64(runtime << 20, period);
2201 inline struct dl_bw *dl_bw_of(int i)
2203 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2204 "sched RCU must be held");
2205 return &cpu_rq(i)->rd->dl_bw;
2208 static inline int dl_bw_cpus(int i)
2210 struct root_domain *rd = cpu_rq(i)->rd;
2213 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2214 "sched RCU must be held");
2215 for_each_cpu_and(i, rd->span, cpu_active_mask)
2221 inline struct dl_bw *dl_bw_of(int i)
2223 return &cpu_rq(i)->dl.dl_bw;
2226 static inline int dl_bw_cpus(int i)
2233 * We must be sure that accepting a new task (or allowing changing the
2234 * parameters of an existing one) is consistent with the bandwidth
2235 * constraints. If yes, this function also accordingly updates the currently
2236 * allocated bandwidth to reflect the new situation.
2238 * This function is called while holding p's rq->lock.
2240 * XXX we should delay bw change until the task's 0-lag point, see
2243 static int dl_overflow(struct task_struct *p, int policy,
2244 const struct sched_attr *attr)
2247 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2248 u64 period = attr->sched_period ?: attr->sched_deadline;
2249 u64 runtime = attr->sched_runtime;
2250 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2253 if (new_bw == p->dl.dl_bw)
2257 * Either if a task, enters, leave, or stays -deadline but changes
2258 * its parameters, we may need to update accordingly the total
2259 * allocated bandwidth of the container.
2261 raw_spin_lock(&dl_b->lock);
2262 cpus = dl_bw_cpus(task_cpu(p));
2263 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2264 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2265 __dl_add(dl_b, new_bw);
2267 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2268 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2269 __dl_clear(dl_b, p->dl.dl_bw);
2270 __dl_add(dl_b, new_bw);
2272 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2273 __dl_clear(dl_b, p->dl.dl_bw);
2276 raw_spin_unlock(&dl_b->lock);
2281 extern void init_dl_bw(struct dl_bw *dl_b);
2284 * wake_up_new_task - wake up a newly created task for the first time.
2286 * This function will do some initial scheduler statistics housekeeping
2287 * that must be done for every newly created context, then puts the task
2288 * on the runqueue and wakes it.
2290 void wake_up_new_task(struct task_struct *p)
2292 unsigned long flags;
2295 raw_spin_lock_irqsave(&p->pi_lock, flags);
2298 * Fork balancing, do it here and not earlier because:
2299 * - cpus_allowed can change in the fork path
2300 * - any previously selected cpu might disappear through hotplug
2302 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2305 /* Initialize new task's runnable average */
2306 init_task_runnable_average(p);
2307 rq = __task_rq_lock(p);
2308 activate_task(rq, p, 0);
2309 p->on_rq = TASK_ON_RQ_QUEUED;
2310 trace_sched_wakeup_new(p, true);
2311 check_preempt_curr(rq, p, WF_FORK);
2313 if (p->sched_class->task_woken)
2314 p->sched_class->task_woken(rq, p);
2316 task_rq_unlock(rq, p, &flags);
2319 #ifdef CONFIG_PREEMPT_NOTIFIERS
2321 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2324 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2325 * @notifier: notifier struct to register
2327 void preempt_notifier_register(struct preempt_notifier *notifier)
2329 static_key_slow_inc(&preempt_notifier_key);
2330 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2332 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2335 * preempt_notifier_unregister - no longer interested in preemption notifications
2336 * @notifier: notifier struct to unregister
2338 * This is *not* safe to call from within a preemption notifier.
2340 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2342 hlist_del(¬ifier->link);
2343 static_key_slow_dec(&preempt_notifier_key);
2345 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2347 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2349 struct preempt_notifier *notifier;
2351 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2352 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2355 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2357 if (static_key_false(&preempt_notifier_key))
2358 __fire_sched_in_preempt_notifiers(curr);
2362 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2363 struct task_struct *next)
2365 struct preempt_notifier *notifier;
2367 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2368 notifier->ops->sched_out(notifier, next);
2371 static __always_inline void
2372 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2373 struct task_struct *next)
2375 if (static_key_false(&preempt_notifier_key))
2376 __fire_sched_out_preempt_notifiers(curr, next);
2379 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2381 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2386 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2387 struct task_struct *next)
2391 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2394 * prepare_task_switch - prepare to switch tasks
2395 * @rq: the runqueue preparing to switch
2396 * @prev: the current task that is being switched out
2397 * @next: the task we are going to switch to.
2399 * This is called with the rq lock held and interrupts off. It must
2400 * be paired with a subsequent finish_task_switch after the context
2403 * prepare_task_switch sets up locking and calls architecture specific
2407 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2408 struct task_struct *next)
2410 trace_sched_switch(prev, next);
2411 sched_info_switch(rq, prev, next);
2412 perf_event_task_sched_out(prev, next);
2413 fire_sched_out_preempt_notifiers(prev, next);
2414 prepare_lock_switch(rq, next);
2415 prepare_arch_switch(next);
2419 * finish_task_switch - clean up after a task-switch
2420 * @prev: the thread we just switched away from.
2422 * finish_task_switch must be called after the context switch, paired
2423 * with a prepare_task_switch call before the context switch.
2424 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2425 * and do any other architecture-specific cleanup actions.
2427 * Note that we may have delayed dropping an mm in context_switch(). If
2428 * so, we finish that here outside of the runqueue lock. (Doing it
2429 * with the lock held can cause deadlocks; see schedule() for
2432 * The context switch have flipped the stack from under us and restored the
2433 * local variables which were saved when this task called schedule() in the
2434 * past. prev == current is still correct but we need to recalculate this_rq
2435 * because prev may have moved to another CPU.
2437 static struct rq *finish_task_switch(struct task_struct *prev)
2438 __releases(rq->lock)
2440 struct rq *rq = this_rq();
2441 struct mm_struct *mm = rq->prev_mm;
2447 * A task struct has one reference for the use as "current".
2448 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2449 * schedule one last time. The schedule call will never return, and
2450 * the scheduled task must drop that reference.
2451 * The test for TASK_DEAD must occur while the runqueue locks are
2452 * still held, otherwise prev could be scheduled on another cpu, die
2453 * there before we look at prev->state, and then the reference would
2455 * Manfred Spraul <manfred@colorfullife.com>
2457 prev_state = prev->state;
2458 vtime_task_switch(prev);
2459 finish_arch_switch(prev);
2460 perf_event_task_sched_in(prev, current);
2461 finish_lock_switch(rq, prev);
2462 finish_arch_post_lock_switch();
2464 fire_sched_in_preempt_notifiers(current);
2467 if (unlikely(prev_state == TASK_DEAD)) {
2468 if (prev->sched_class->task_dead)
2469 prev->sched_class->task_dead(prev);
2472 * Remove function-return probe instances associated with this
2473 * task and put them back on the free list.
2475 kprobe_flush_task(prev);
2476 put_task_struct(prev);
2479 tick_nohz_task_switch(current);
2485 /* rq->lock is NOT held, but preemption is disabled */
2486 static void __balance_callback(struct rq *rq)
2488 struct callback_head *head, *next;
2489 void (*func)(struct rq *rq);
2490 unsigned long flags;
2492 raw_spin_lock_irqsave(&rq->lock, flags);
2493 head = rq->balance_callback;
2494 rq->balance_callback = NULL;
2496 func = (void (*)(struct rq *))head->func;
2503 raw_spin_unlock_irqrestore(&rq->lock, flags);
2506 static inline void balance_callback(struct rq *rq)
2508 if (unlikely(rq->balance_callback))
2509 __balance_callback(rq);
2514 static inline void balance_callback(struct rq *rq)
2521 * schedule_tail - first thing a freshly forked thread must call.
2522 * @prev: the thread we just switched away from.
2524 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2525 __releases(rq->lock)
2529 /* finish_task_switch() drops rq->lock and enables preemtion */
2531 rq = finish_task_switch(prev);
2532 balance_callback(rq);
2535 if (current->set_child_tid)
2536 put_user(task_pid_vnr(current), current->set_child_tid);
2540 * context_switch - switch to the new MM and the new thread's register state.
2542 static inline struct rq *
2543 context_switch(struct rq *rq, struct task_struct *prev,
2544 struct task_struct *next)
2546 struct mm_struct *mm, *oldmm;
2548 prepare_task_switch(rq, prev, next);
2551 oldmm = prev->active_mm;
2553 * For paravirt, this is coupled with an exit in switch_to to
2554 * combine the page table reload and the switch backend into
2557 arch_start_context_switch(prev);
2560 next->active_mm = oldmm;
2561 atomic_inc(&oldmm->mm_count);
2562 enter_lazy_tlb(oldmm, next);
2564 switch_mm(oldmm, mm, next);
2567 prev->active_mm = NULL;
2568 rq->prev_mm = oldmm;
2571 * Since the runqueue lock will be released by the next
2572 * task (which is an invalid locking op but in the case
2573 * of the scheduler it's an obvious special-case), so we
2574 * do an early lockdep release here:
2576 lockdep_unpin_lock(&rq->lock);
2577 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2579 /* Here we just switch the register state and the stack. */
2580 switch_to(prev, next, prev);
2583 return finish_task_switch(prev);
2587 * nr_running and nr_context_switches:
2589 * externally visible scheduler statistics: current number of runnable
2590 * threads, total number of context switches performed since bootup.
2592 unsigned long nr_running(void)
2594 unsigned long i, sum = 0;
2596 for_each_online_cpu(i)
2597 sum += cpu_rq(i)->nr_running;
2603 * Check if only the current task is running on the cpu.
2605 bool single_task_running(void)
2607 if (cpu_rq(smp_processor_id())->nr_running == 1)
2612 EXPORT_SYMBOL(single_task_running);
2614 unsigned long long nr_context_switches(void)
2617 unsigned long long sum = 0;
2619 for_each_possible_cpu(i)
2620 sum += cpu_rq(i)->nr_switches;
2625 unsigned long nr_iowait(void)
2627 unsigned long i, sum = 0;
2629 for_each_possible_cpu(i)
2630 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2635 unsigned long nr_iowait_cpu(int cpu)
2637 struct rq *this = cpu_rq(cpu);
2638 return atomic_read(&this->nr_iowait);
2641 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2643 struct rq *rq = this_rq();
2644 *nr_waiters = atomic_read(&rq->nr_iowait);
2645 *load = rq->load.weight;
2651 * sched_exec - execve() is a valuable balancing opportunity, because at
2652 * this point the task has the smallest effective memory and cache footprint.
2654 void sched_exec(void)
2656 struct task_struct *p = current;
2657 unsigned long flags;
2660 raw_spin_lock_irqsave(&p->pi_lock, flags);
2661 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2662 if (dest_cpu == smp_processor_id())
2665 if (likely(cpu_active(dest_cpu))) {
2666 struct migration_arg arg = { p, dest_cpu };
2668 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2669 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2673 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2678 DEFINE_PER_CPU(struct kernel_stat, kstat);
2679 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2681 EXPORT_PER_CPU_SYMBOL(kstat);
2682 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2685 * Return accounted runtime for the task.
2686 * In case the task is currently running, return the runtime plus current's
2687 * pending runtime that have not been accounted yet.
2689 unsigned long long task_sched_runtime(struct task_struct *p)
2691 unsigned long flags;
2695 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2697 * 64-bit doesn't need locks to atomically read a 64bit value.
2698 * So we have a optimization chance when the task's delta_exec is 0.
2699 * Reading ->on_cpu is racy, but this is ok.
2701 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2702 * If we race with it entering cpu, unaccounted time is 0. This is
2703 * indistinguishable from the read occurring a few cycles earlier.
2704 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2705 * been accounted, so we're correct here as well.
2707 if (!p->on_cpu || !task_on_rq_queued(p))
2708 return p->se.sum_exec_runtime;
2711 rq = task_rq_lock(p, &flags);
2713 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2714 * project cycles that may never be accounted to this
2715 * thread, breaking clock_gettime().
2717 if (task_current(rq, p) && task_on_rq_queued(p)) {
2718 update_rq_clock(rq);
2719 p->sched_class->update_curr(rq);
2721 ns = p->se.sum_exec_runtime;
2722 task_rq_unlock(rq, p, &flags);
2728 * This function gets called by the timer code, with HZ frequency.
2729 * We call it with interrupts disabled.
2731 void scheduler_tick(void)
2733 int cpu = smp_processor_id();
2734 struct rq *rq = cpu_rq(cpu);
2735 struct task_struct *curr = rq->curr;
2739 raw_spin_lock(&rq->lock);
2740 update_rq_clock(rq);
2741 curr->sched_class->task_tick(rq, curr, 0);
2742 update_cpu_load_active(rq);
2743 calc_global_load_tick(rq);
2744 raw_spin_unlock(&rq->lock);
2746 perf_event_task_tick();
2749 rq->idle_balance = idle_cpu(cpu);
2750 trigger_load_balance(rq);
2752 rq_last_tick_reset(rq);
2755 #ifdef CONFIG_NO_HZ_FULL
2757 * scheduler_tick_max_deferment
2759 * Keep at least one tick per second when a single
2760 * active task is running because the scheduler doesn't
2761 * yet completely support full dynticks environment.
2763 * This makes sure that uptime, CFS vruntime, load
2764 * balancing, etc... continue to move forward, even
2765 * with a very low granularity.
2767 * Return: Maximum deferment in nanoseconds.
2769 u64 scheduler_tick_max_deferment(void)
2771 struct rq *rq = this_rq();
2772 unsigned long next, now = READ_ONCE(jiffies);
2774 next = rq->last_sched_tick + HZ;
2776 if (time_before_eq(next, now))
2779 return jiffies_to_nsecs(next - now);
2783 notrace unsigned long get_parent_ip(unsigned long addr)
2785 if (in_lock_functions(addr)) {
2786 addr = CALLER_ADDR2;
2787 if (in_lock_functions(addr))
2788 addr = CALLER_ADDR3;
2793 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2794 defined(CONFIG_PREEMPT_TRACER))
2796 void preempt_count_add(int val)
2798 #ifdef CONFIG_DEBUG_PREEMPT
2802 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2805 __preempt_count_add(val);
2806 #ifdef CONFIG_DEBUG_PREEMPT
2808 * Spinlock count overflowing soon?
2810 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2813 if (preempt_count() == val) {
2814 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2815 #ifdef CONFIG_DEBUG_PREEMPT
2816 current->preempt_disable_ip = ip;
2818 trace_preempt_off(CALLER_ADDR0, ip);
2821 EXPORT_SYMBOL(preempt_count_add);
2822 NOKPROBE_SYMBOL(preempt_count_add);
2824 void preempt_count_sub(int val)
2826 #ifdef CONFIG_DEBUG_PREEMPT
2830 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2833 * Is the spinlock portion underflowing?
2835 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2836 !(preempt_count() & PREEMPT_MASK)))
2840 if (preempt_count() == val)
2841 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2842 __preempt_count_sub(val);
2844 EXPORT_SYMBOL(preempt_count_sub);
2845 NOKPROBE_SYMBOL(preempt_count_sub);
2850 * Print scheduling while atomic bug:
2852 static noinline void __schedule_bug(struct task_struct *prev)
2854 if (oops_in_progress)
2857 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2858 prev->comm, prev->pid, preempt_count());
2860 debug_show_held_locks(prev);
2862 if (irqs_disabled())
2863 print_irqtrace_events(prev);
2864 #ifdef CONFIG_DEBUG_PREEMPT
2865 if (in_atomic_preempt_off()) {
2866 pr_err("Preemption disabled at:");
2867 print_ip_sym(current->preempt_disable_ip);
2872 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2876 * Various schedule()-time debugging checks and statistics:
2878 static inline void schedule_debug(struct task_struct *prev)
2880 #ifdef CONFIG_SCHED_STACK_END_CHECK
2881 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2884 * Test if we are atomic. Since do_exit() needs to call into
2885 * schedule() atomically, we ignore that path. Otherwise whine
2886 * if we are scheduling when we should not.
2888 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2889 __schedule_bug(prev);
2892 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2894 schedstat_inc(this_rq(), sched_count);
2898 * Pick up the highest-prio task:
2900 static inline struct task_struct *
2901 pick_next_task(struct rq *rq, struct task_struct *prev)
2903 const struct sched_class *class = &fair_sched_class;
2904 struct task_struct *p;
2907 * Optimization: we know that if all tasks are in
2908 * the fair class we can call that function directly:
2910 if (likely(prev->sched_class == class &&
2911 rq->nr_running == rq->cfs.h_nr_running)) {
2912 p = fair_sched_class.pick_next_task(rq, prev);
2913 if (unlikely(p == RETRY_TASK))
2916 /* assumes fair_sched_class->next == idle_sched_class */
2918 p = idle_sched_class.pick_next_task(rq, prev);
2924 for_each_class(class) {
2925 p = class->pick_next_task(rq, prev);
2927 if (unlikely(p == RETRY_TASK))
2933 BUG(); /* the idle class will always have a runnable task */
2937 * __schedule() is the main scheduler function.
2939 * The main means of driving the scheduler and thus entering this function are:
2941 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2943 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2944 * paths. For example, see arch/x86/entry_64.S.
2946 * To drive preemption between tasks, the scheduler sets the flag in timer
2947 * interrupt handler scheduler_tick().
2949 * 3. Wakeups don't really cause entry into schedule(). They add a
2950 * task to the run-queue and that's it.
2952 * Now, if the new task added to the run-queue preempts the current
2953 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2954 * called on the nearest possible occasion:
2956 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2958 * - in syscall or exception context, at the next outmost
2959 * preempt_enable(). (this might be as soon as the wake_up()'s
2962 * - in IRQ context, return from interrupt-handler to
2963 * preemptible context
2965 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2968 * - cond_resched() call
2969 * - explicit schedule() call
2970 * - return from syscall or exception to user-space
2971 * - return from interrupt-handler to user-space
2973 * WARNING: must be called with preemption disabled!
2975 static void __sched __schedule(void)
2977 struct task_struct *prev, *next;
2978 unsigned long *switch_count;
2982 cpu = smp_processor_id();
2984 rcu_note_context_switch();
2987 schedule_debug(prev);
2989 if (sched_feat(HRTICK))
2993 * Make sure that signal_pending_state()->signal_pending() below
2994 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2995 * done by the caller to avoid the race with signal_wake_up().
2997 smp_mb__before_spinlock();
2998 raw_spin_lock_irq(&rq->lock);
2999 lockdep_pin_lock(&rq->lock);
3001 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3003 switch_count = &prev->nivcsw;
3004 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3005 if (unlikely(signal_pending_state(prev->state, prev))) {
3006 prev->state = TASK_RUNNING;
3008 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3012 * If a worker went to sleep, notify and ask workqueue
3013 * whether it wants to wake up a task to maintain
3016 if (prev->flags & PF_WQ_WORKER) {
3017 struct task_struct *to_wakeup;
3019 to_wakeup = wq_worker_sleeping(prev, cpu);
3021 try_to_wake_up_local(to_wakeup);
3024 switch_count = &prev->nvcsw;
3027 if (task_on_rq_queued(prev))
3028 update_rq_clock(rq);
3030 next = pick_next_task(rq, prev);
3031 clear_tsk_need_resched(prev);
3032 clear_preempt_need_resched();
3033 rq->clock_skip_update = 0;
3035 if (likely(prev != next)) {
3040 rq = context_switch(rq, prev, next); /* unlocks the rq */
3043 lockdep_unpin_lock(&rq->lock);
3044 raw_spin_unlock_irq(&rq->lock);
3047 balance_callback(rq);
3050 static inline void sched_submit_work(struct task_struct *tsk)
3052 if (!tsk->state || tsk_is_pi_blocked(tsk))
3055 * If we are going to sleep and we have plugged IO queued,
3056 * make sure to submit it to avoid deadlocks.
3058 if (blk_needs_flush_plug(tsk))
3059 blk_schedule_flush_plug(tsk);
3062 asmlinkage __visible void __sched schedule(void)
3064 struct task_struct *tsk = current;
3066 sched_submit_work(tsk);
3070 sched_preempt_enable_no_resched();
3071 } while (need_resched());
3073 EXPORT_SYMBOL(schedule);
3075 #ifdef CONFIG_CONTEXT_TRACKING
3076 asmlinkage __visible void __sched schedule_user(void)
3079 * If we come here after a random call to set_need_resched(),
3080 * or we have been woken up remotely but the IPI has not yet arrived,
3081 * we haven't yet exited the RCU idle mode. Do it here manually until
3082 * we find a better solution.
3084 * NB: There are buggy callers of this function. Ideally we
3085 * should warn if prev_state != CONTEXT_USER, but that will trigger
3086 * too frequently to make sense yet.
3088 enum ctx_state prev_state = exception_enter();
3090 exception_exit(prev_state);
3095 * schedule_preempt_disabled - called with preemption disabled
3097 * Returns with preemption disabled. Note: preempt_count must be 1
3099 void __sched schedule_preempt_disabled(void)
3101 sched_preempt_enable_no_resched();
3106 static void __sched notrace preempt_schedule_common(void)
3109 preempt_active_enter();
3111 preempt_active_exit();
3114 * Check again in case we missed a preemption opportunity
3115 * between schedule and now.
3117 } while (need_resched());
3120 #ifdef CONFIG_PREEMPT
3122 * this is the entry point to schedule() from in-kernel preemption
3123 * off of preempt_enable. Kernel preemptions off return from interrupt
3124 * occur there and call schedule directly.
3126 asmlinkage __visible void __sched notrace preempt_schedule(void)
3129 * If there is a non-zero preempt_count or interrupts are disabled,
3130 * we do not want to preempt the current task. Just return..
3132 if (likely(!preemptible()))
3135 preempt_schedule_common();
3137 NOKPROBE_SYMBOL(preempt_schedule);
3138 EXPORT_SYMBOL(preempt_schedule);
3141 * preempt_schedule_notrace - preempt_schedule called by tracing
3143 * The tracing infrastructure uses preempt_enable_notrace to prevent
3144 * recursion and tracing preempt enabling caused by the tracing
3145 * infrastructure itself. But as tracing can happen in areas coming
3146 * from userspace or just about to enter userspace, a preempt enable
3147 * can occur before user_exit() is called. This will cause the scheduler
3148 * to be called when the system is still in usermode.
3150 * To prevent this, the preempt_enable_notrace will use this function
3151 * instead of preempt_schedule() to exit user context if needed before
3152 * calling the scheduler.
3154 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3156 enum ctx_state prev_ctx;
3158 if (likely(!preemptible()))
3163 * Use raw __prempt_count() ops that don't call function.
3164 * We can't call functions before disabling preemption which
3165 * disarm preemption tracing recursions.
3167 __preempt_count_add(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
3170 * Needs preempt disabled in case user_exit() is traced
3171 * and the tracer calls preempt_enable_notrace() causing
3172 * an infinite recursion.
3174 prev_ctx = exception_enter();
3176 exception_exit(prev_ctx);
3179 __preempt_count_sub(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
3180 } while (need_resched());
3182 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3184 #endif /* CONFIG_PREEMPT */
3187 * this is the entry point to schedule() from kernel preemption
3188 * off of irq context.
3189 * Note, that this is called and return with irqs disabled. This will
3190 * protect us against recursive calling from irq.
3192 asmlinkage __visible void __sched preempt_schedule_irq(void)
3194 enum ctx_state prev_state;
3196 /* Catch callers which need to be fixed */
3197 BUG_ON(preempt_count() || !irqs_disabled());
3199 prev_state = exception_enter();
3202 preempt_active_enter();
3205 local_irq_disable();
3206 preempt_active_exit();
3207 } while (need_resched());
3209 exception_exit(prev_state);
3212 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3215 return try_to_wake_up(curr->private, mode, wake_flags);
3217 EXPORT_SYMBOL(default_wake_function);
3219 #ifdef CONFIG_RT_MUTEXES
3222 * rt_mutex_setprio - set the current priority of a task
3224 * @prio: prio value (kernel-internal form)
3226 * This function changes the 'effective' priority of a task. It does
3227 * not touch ->normal_prio like __setscheduler().
3229 * Used by the rt_mutex code to implement priority inheritance
3230 * logic. Call site only calls if the priority of the task changed.
3232 void rt_mutex_setprio(struct task_struct *p, int prio)
3234 int oldprio, queued, running, enqueue_flag = 0;
3236 const struct sched_class *prev_class;
3238 BUG_ON(prio > MAX_PRIO);
3240 rq = __task_rq_lock(p);
3243 * Idle task boosting is a nono in general. There is one
3244 * exception, when PREEMPT_RT and NOHZ is active:
3246 * The idle task calls get_next_timer_interrupt() and holds
3247 * the timer wheel base->lock on the CPU and another CPU wants
3248 * to access the timer (probably to cancel it). We can safely
3249 * ignore the boosting request, as the idle CPU runs this code
3250 * with interrupts disabled and will complete the lock
3251 * protected section without being interrupted. So there is no
3252 * real need to boost.
3254 if (unlikely(p == rq->idle)) {
3255 WARN_ON(p != rq->curr);
3256 WARN_ON(p->pi_blocked_on);
3260 trace_sched_pi_setprio(p, prio);
3262 prev_class = p->sched_class;
3263 queued = task_on_rq_queued(p);
3264 running = task_current(rq, p);
3266 dequeue_task(rq, p, 0);
3268 put_prev_task(rq, p);
3271 * Boosting condition are:
3272 * 1. -rt task is running and holds mutex A
3273 * --> -dl task blocks on mutex A
3275 * 2. -dl task is running and holds mutex A
3276 * --> -dl task blocks on mutex A and could preempt the
3279 if (dl_prio(prio)) {
3280 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3281 if (!dl_prio(p->normal_prio) ||
3282 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3283 p->dl.dl_boosted = 1;
3284 enqueue_flag = ENQUEUE_REPLENISH;
3286 p->dl.dl_boosted = 0;
3287 p->sched_class = &dl_sched_class;
3288 } else if (rt_prio(prio)) {
3289 if (dl_prio(oldprio))
3290 p->dl.dl_boosted = 0;
3292 enqueue_flag = ENQUEUE_HEAD;
3293 p->sched_class = &rt_sched_class;
3295 if (dl_prio(oldprio))
3296 p->dl.dl_boosted = 0;
3297 if (rt_prio(oldprio))
3299 p->sched_class = &fair_sched_class;
3305 p->sched_class->set_curr_task(rq);
3307 enqueue_task(rq, p, enqueue_flag);
3309 check_class_changed(rq, p, prev_class, oldprio);
3311 preempt_disable(); /* avoid rq from going away on us */
3312 __task_rq_unlock(rq);
3314 balance_callback(rq);
3319 void set_user_nice(struct task_struct *p, long nice)
3321 int old_prio, delta, queued;
3322 unsigned long flags;
3325 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3328 * We have to be careful, if called from sys_setpriority(),
3329 * the task might be in the middle of scheduling on another CPU.
3331 rq = task_rq_lock(p, &flags);
3333 * The RT priorities are set via sched_setscheduler(), but we still
3334 * allow the 'normal' nice value to be set - but as expected
3335 * it wont have any effect on scheduling until the task is
3336 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3338 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3339 p->static_prio = NICE_TO_PRIO(nice);
3342 queued = task_on_rq_queued(p);
3344 dequeue_task(rq, p, 0);
3346 p->static_prio = NICE_TO_PRIO(nice);
3349 p->prio = effective_prio(p);
3350 delta = p->prio - old_prio;
3353 enqueue_task(rq, p, 0);
3355 * If the task increased its priority or is running and
3356 * lowered its priority, then reschedule its CPU:
3358 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3362 task_rq_unlock(rq, p, &flags);
3364 EXPORT_SYMBOL(set_user_nice);
3367 * can_nice - check if a task can reduce its nice value
3371 int can_nice(const struct task_struct *p, const int nice)
3373 /* convert nice value [19,-20] to rlimit style value [1,40] */
3374 int nice_rlim = nice_to_rlimit(nice);
3376 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3377 capable(CAP_SYS_NICE));
3380 #ifdef __ARCH_WANT_SYS_NICE
3383 * sys_nice - change the priority of the current process.
3384 * @increment: priority increment
3386 * sys_setpriority is a more generic, but much slower function that
3387 * does similar things.
3389 SYSCALL_DEFINE1(nice, int, increment)
3394 * Setpriority might change our priority at the same moment.
3395 * We don't have to worry. Conceptually one call occurs first
3396 * and we have a single winner.
3398 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3399 nice = task_nice(current) + increment;
3401 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3402 if (increment < 0 && !can_nice(current, nice))
3405 retval = security_task_setnice(current, nice);
3409 set_user_nice(current, nice);
3416 * task_prio - return the priority value of a given task.
3417 * @p: the task in question.
3419 * Return: The priority value as seen by users in /proc.
3420 * RT tasks are offset by -200. Normal tasks are centered
3421 * around 0, value goes from -16 to +15.
3423 int task_prio(const struct task_struct *p)
3425 return p->prio - MAX_RT_PRIO;
3429 * idle_cpu - is a given cpu idle currently?
3430 * @cpu: the processor in question.
3432 * Return: 1 if the CPU is currently idle. 0 otherwise.
3434 int idle_cpu(int cpu)
3436 struct rq *rq = cpu_rq(cpu);
3438 if (rq->curr != rq->idle)
3445 if (!llist_empty(&rq->wake_list))
3453 * idle_task - return the idle task for a given cpu.
3454 * @cpu: the processor in question.
3456 * Return: The idle task for the cpu @cpu.
3458 struct task_struct *idle_task(int cpu)
3460 return cpu_rq(cpu)->idle;
3464 * find_process_by_pid - find a process with a matching PID value.
3465 * @pid: the pid in question.
3467 * The task of @pid, if found. %NULL otherwise.
3469 static struct task_struct *find_process_by_pid(pid_t pid)
3471 return pid ? find_task_by_vpid(pid) : current;
3475 * This function initializes the sched_dl_entity of a newly becoming
3476 * SCHED_DEADLINE task.
3478 * Only the static values are considered here, the actual runtime and the
3479 * absolute deadline will be properly calculated when the task is enqueued
3480 * for the first time with its new policy.
3483 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3485 struct sched_dl_entity *dl_se = &p->dl;
3487 dl_se->dl_runtime = attr->sched_runtime;
3488 dl_se->dl_deadline = attr->sched_deadline;
3489 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3490 dl_se->flags = attr->sched_flags;
3491 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3494 * Changing the parameters of a task is 'tricky' and we're not doing
3495 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3497 * What we SHOULD do is delay the bandwidth release until the 0-lag
3498 * point. This would include retaining the task_struct until that time
3499 * and change dl_overflow() to not immediately decrement the current
3502 * Instead we retain the current runtime/deadline and let the new
3503 * parameters take effect after the current reservation period lapses.
3504 * This is safe (albeit pessimistic) because the 0-lag point is always
3505 * before the current scheduling deadline.
3507 * We can still have temporary overloads because we do not delay the
3508 * change in bandwidth until that time; so admission control is
3509 * not on the safe side. It does however guarantee tasks will never
3510 * consume more than promised.
3515 * sched_setparam() passes in -1 for its policy, to let the functions
3516 * it calls know not to change it.
3518 #define SETPARAM_POLICY -1
3520 static void __setscheduler_params(struct task_struct *p,
3521 const struct sched_attr *attr)
3523 int policy = attr->sched_policy;
3525 if (policy == SETPARAM_POLICY)
3530 if (dl_policy(policy))
3531 __setparam_dl(p, attr);
3532 else if (fair_policy(policy))
3533 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3536 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3537 * !rt_policy. Always setting this ensures that things like
3538 * getparam()/getattr() don't report silly values for !rt tasks.
3540 p->rt_priority = attr->sched_priority;
3541 p->normal_prio = normal_prio(p);
3545 /* Actually do priority change: must hold pi & rq lock. */
3546 static void __setscheduler(struct rq *rq, struct task_struct *p,
3547 const struct sched_attr *attr, bool keep_boost)
3549 __setscheduler_params(p, attr);
3552 * Keep a potential priority boosting if called from
3553 * sched_setscheduler().
3556 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3558 p->prio = normal_prio(p);
3560 if (dl_prio(p->prio))
3561 p->sched_class = &dl_sched_class;
3562 else if (rt_prio(p->prio))
3563 p->sched_class = &rt_sched_class;
3565 p->sched_class = &fair_sched_class;
3569 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3571 struct sched_dl_entity *dl_se = &p->dl;
3573 attr->sched_priority = p->rt_priority;
3574 attr->sched_runtime = dl_se->dl_runtime;
3575 attr->sched_deadline = dl_se->dl_deadline;
3576 attr->sched_period = dl_se->dl_period;
3577 attr->sched_flags = dl_se->flags;
3581 * This function validates the new parameters of a -deadline task.
3582 * We ask for the deadline not being zero, and greater or equal
3583 * than the runtime, as well as the period of being zero or
3584 * greater than deadline. Furthermore, we have to be sure that
3585 * user parameters are above the internal resolution of 1us (we
3586 * check sched_runtime only since it is always the smaller one) and
3587 * below 2^63 ns (we have to check both sched_deadline and
3588 * sched_period, as the latter can be zero).
3591 __checkparam_dl(const struct sched_attr *attr)
3594 if (attr->sched_deadline == 0)
3598 * Since we truncate DL_SCALE bits, make sure we're at least
3601 if (attr->sched_runtime < (1ULL << DL_SCALE))
3605 * Since we use the MSB for wrap-around and sign issues, make
3606 * sure it's not set (mind that period can be equal to zero).
3608 if (attr->sched_deadline & (1ULL << 63) ||
3609 attr->sched_period & (1ULL << 63))
3612 /* runtime <= deadline <= period (if period != 0) */
3613 if ((attr->sched_period != 0 &&
3614 attr->sched_period < attr->sched_deadline) ||
3615 attr->sched_deadline < attr->sched_runtime)
3622 * check the target process has a UID that matches the current process's
3624 static bool check_same_owner(struct task_struct *p)
3626 const struct cred *cred = current_cred(), *pcred;
3630 pcred = __task_cred(p);
3631 match = (uid_eq(cred->euid, pcred->euid) ||
3632 uid_eq(cred->euid, pcred->uid));
3637 static bool dl_param_changed(struct task_struct *p,
3638 const struct sched_attr *attr)
3640 struct sched_dl_entity *dl_se = &p->dl;
3642 if (dl_se->dl_runtime != attr->sched_runtime ||
3643 dl_se->dl_deadline != attr->sched_deadline ||
3644 dl_se->dl_period != attr->sched_period ||
3645 dl_se->flags != attr->sched_flags)
3651 static int __sched_setscheduler(struct task_struct *p,
3652 const struct sched_attr *attr,
3655 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3656 MAX_RT_PRIO - 1 - attr->sched_priority;
3657 int retval, oldprio, oldpolicy = -1, queued, running;
3658 int new_effective_prio, policy = attr->sched_policy;
3659 unsigned long flags;
3660 const struct sched_class *prev_class;
3664 /* may grab non-irq protected spin_locks */
3665 BUG_ON(in_interrupt());
3667 /* double check policy once rq lock held */
3669 reset_on_fork = p->sched_reset_on_fork;
3670 policy = oldpolicy = p->policy;
3672 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3674 if (policy != SCHED_DEADLINE &&
3675 policy != SCHED_FIFO && policy != SCHED_RR &&
3676 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3677 policy != SCHED_IDLE)
3681 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3685 * Valid priorities for SCHED_FIFO and SCHED_RR are
3686 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3687 * SCHED_BATCH and SCHED_IDLE is 0.
3689 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3690 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3692 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3693 (rt_policy(policy) != (attr->sched_priority != 0)))
3697 * Allow unprivileged RT tasks to decrease priority:
3699 if (user && !capable(CAP_SYS_NICE)) {
3700 if (fair_policy(policy)) {
3701 if (attr->sched_nice < task_nice(p) &&
3702 !can_nice(p, attr->sched_nice))
3706 if (rt_policy(policy)) {
3707 unsigned long rlim_rtprio =
3708 task_rlimit(p, RLIMIT_RTPRIO);
3710 /* can't set/change the rt policy */
3711 if (policy != p->policy && !rlim_rtprio)
3714 /* can't increase priority */
3715 if (attr->sched_priority > p->rt_priority &&
3716 attr->sched_priority > rlim_rtprio)
3721 * Can't set/change SCHED_DEADLINE policy at all for now
3722 * (safest behavior); in the future we would like to allow
3723 * unprivileged DL tasks to increase their relative deadline
3724 * or reduce their runtime (both ways reducing utilization)
3726 if (dl_policy(policy))
3730 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3731 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3733 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3734 if (!can_nice(p, task_nice(p)))
3738 /* can't change other user's priorities */
3739 if (!check_same_owner(p))
3742 /* Normal users shall not reset the sched_reset_on_fork flag */
3743 if (p->sched_reset_on_fork && !reset_on_fork)
3748 retval = security_task_setscheduler(p);
3754 * make sure no PI-waiters arrive (or leave) while we are
3755 * changing the priority of the task:
3757 * To be able to change p->policy safely, the appropriate
3758 * runqueue lock must be held.
3760 rq = task_rq_lock(p, &flags);
3763 * Changing the policy of the stop threads its a very bad idea
3765 if (p == rq->stop) {
3766 task_rq_unlock(rq, p, &flags);
3771 * If not changing anything there's no need to proceed further,
3772 * but store a possible modification of reset_on_fork.
3774 if (unlikely(policy == p->policy)) {
3775 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3777 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3779 if (dl_policy(policy) && dl_param_changed(p, attr))
3782 p->sched_reset_on_fork = reset_on_fork;
3783 task_rq_unlock(rq, p, &flags);
3789 #ifdef CONFIG_RT_GROUP_SCHED
3791 * Do not allow realtime tasks into groups that have no runtime
3794 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3795 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3796 !task_group_is_autogroup(task_group(p))) {
3797 task_rq_unlock(rq, p, &flags);
3802 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3803 cpumask_t *span = rq->rd->span;
3806 * Don't allow tasks with an affinity mask smaller than
3807 * the entire root_domain to become SCHED_DEADLINE. We
3808 * will also fail if there's no bandwidth available.
3810 if (!cpumask_subset(span, &p->cpus_allowed) ||
3811 rq->rd->dl_bw.bw == 0) {
3812 task_rq_unlock(rq, p, &flags);
3819 /* recheck policy now with rq lock held */
3820 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3821 policy = oldpolicy = -1;
3822 task_rq_unlock(rq, p, &flags);
3827 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3828 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3831 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3832 task_rq_unlock(rq, p, &flags);
3836 p->sched_reset_on_fork = reset_on_fork;
3841 * Take priority boosted tasks into account. If the new
3842 * effective priority is unchanged, we just store the new
3843 * normal parameters and do not touch the scheduler class and
3844 * the runqueue. This will be done when the task deboost
3847 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3848 if (new_effective_prio == oldprio) {
3849 __setscheduler_params(p, attr);
3850 task_rq_unlock(rq, p, &flags);
3855 queued = task_on_rq_queued(p);
3856 running = task_current(rq, p);
3858 dequeue_task(rq, p, 0);
3860 put_prev_task(rq, p);
3862 prev_class = p->sched_class;
3863 __setscheduler(rq, p, attr, pi);
3866 p->sched_class->set_curr_task(rq);
3869 * We enqueue to tail when the priority of a task is
3870 * increased (user space view).
3872 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3875 check_class_changed(rq, p, prev_class, oldprio);
3876 preempt_disable(); /* avoid rq from going away on us */
3877 task_rq_unlock(rq, p, &flags);
3880 rt_mutex_adjust_pi(p);
3883 * Run balance callbacks after we've adjusted the PI chain.
3885 balance_callback(rq);
3891 static int _sched_setscheduler(struct task_struct *p, int policy,
3892 const struct sched_param *param, bool check)
3894 struct sched_attr attr = {
3895 .sched_policy = policy,
3896 .sched_priority = param->sched_priority,
3897 .sched_nice = PRIO_TO_NICE(p->static_prio),
3900 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3901 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3902 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3903 policy &= ~SCHED_RESET_ON_FORK;
3904 attr.sched_policy = policy;
3907 return __sched_setscheduler(p, &attr, check, true);
3910 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3911 * @p: the task in question.
3912 * @policy: new policy.
3913 * @param: structure containing the new RT priority.
3915 * Return: 0 on success. An error code otherwise.
3917 * NOTE that the task may be already dead.
3919 int sched_setscheduler(struct task_struct *p, int policy,
3920 const struct sched_param *param)
3922 return _sched_setscheduler(p, policy, param, true);
3924 EXPORT_SYMBOL_GPL(sched_setscheduler);
3926 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3928 return __sched_setscheduler(p, attr, true, true);
3930 EXPORT_SYMBOL_GPL(sched_setattr);
3933 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3934 * @p: the task in question.
3935 * @policy: new policy.
3936 * @param: structure containing the new RT priority.
3938 * Just like sched_setscheduler, only don't bother checking if the
3939 * current context has permission. For example, this is needed in
3940 * stop_machine(): we create temporary high priority worker threads,
3941 * but our caller might not have that capability.
3943 * Return: 0 on success. An error code otherwise.
3945 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3946 const struct sched_param *param)
3948 return _sched_setscheduler(p, policy, param, false);
3952 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3954 struct sched_param lparam;
3955 struct task_struct *p;
3958 if (!param || pid < 0)
3960 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3965 p = find_process_by_pid(pid);
3967 retval = sched_setscheduler(p, policy, &lparam);
3974 * Mimics kernel/events/core.c perf_copy_attr().
3976 static int sched_copy_attr(struct sched_attr __user *uattr,
3977 struct sched_attr *attr)
3982 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3986 * zero the full structure, so that a short copy will be nice.
3988 memset(attr, 0, sizeof(*attr));
3990 ret = get_user(size, &uattr->size);
3994 if (size > PAGE_SIZE) /* silly large */
3997 if (!size) /* abi compat */
3998 size = SCHED_ATTR_SIZE_VER0;
4000 if (size < SCHED_ATTR_SIZE_VER0)
4004 * If we're handed a bigger struct than we know of,
4005 * ensure all the unknown bits are 0 - i.e. new
4006 * user-space does not rely on any kernel feature
4007 * extensions we dont know about yet.
4009 if (size > sizeof(*attr)) {
4010 unsigned char __user *addr;
4011 unsigned char __user *end;
4014 addr = (void __user *)uattr + sizeof(*attr);
4015 end = (void __user *)uattr + size;
4017 for (; addr < end; addr++) {
4018 ret = get_user(val, addr);
4024 size = sizeof(*attr);
4027 ret = copy_from_user(attr, uattr, size);
4032 * XXX: do we want to be lenient like existing syscalls; or do we want
4033 * to be strict and return an error on out-of-bounds values?
4035 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4040 put_user(sizeof(*attr), &uattr->size);
4045 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4046 * @pid: the pid in question.
4047 * @policy: new policy.
4048 * @param: structure containing the new RT priority.
4050 * Return: 0 on success. An error code otherwise.
4052 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4053 struct sched_param __user *, param)
4055 /* negative values for policy are not valid */
4059 return do_sched_setscheduler(pid, policy, param);
4063 * sys_sched_setparam - set/change the RT priority of a thread
4064 * @pid: the pid in question.
4065 * @param: structure containing the new RT priority.
4067 * Return: 0 on success. An error code otherwise.
4069 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4071 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4075 * sys_sched_setattr - same as above, but with extended sched_attr
4076 * @pid: the pid in question.
4077 * @uattr: structure containing the extended parameters.
4078 * @flags: for future extension.
4080 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4081 unsigned int, flags)
4083 struct sched_attr attr;
4084 struct task_struct *p;
4087 if (!uattr || pid < 0 || flags)
4090 retval = sched_copy_attr(uattr, &attr);
4094 if ((int)attr.sched_policy < 0)
4099 p = find_process_by_pid(pid);
4101 retval = sched_setattr(p, &attr);
4108 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4109 * @pid: the pid in question.
4111 * Return: On success, the policy of the thread. Otherwise, a negative error
4114 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4116 struct task_struct *p;
4124 p = find_process_by_pid(pid);
4126 retval = security_task_getscheduler(p);
4129 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4136 * sys_sched_getparam - get the RT priority of a thread
4137 * @pid: the pid in question.
4138 * @param: structure containing the RT priority.
4140 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4143 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4145 struct sched_param lp = { .sched_priority = 0 };
4146 struct task_struct *p;
4149 if (!param || pid < 0)
4153 p = find_process_by_pid(pid);
4158 retval = security_task_getscheduler(p);
4162 if (task_has_rt_policy(p))
4163 lp.sched_priority = p->rt_priority;
4167 * This one might sleep, we cannot do it with a spinlock held ...
4169 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4178 static int sched_read_attr(struct sched_attr __user *uattr,
4179 struct sched_attr *attr,
4184 if (!access_ok(VERIFY_WRITE, uattr, usize))
4188 * If we're handed a smaller struct than we know of,
4189 * ensure all the unknown bits are 0 - i.e. old
4190 * user-space does not get uncomplete information.
4192 if (usize < sizeof(*attr)) {
4193 unsigned char *addr;
4196 addr = (void *)attr + usize;
4197 end = (void *)attr + sizeof(*attr);
4199 for (; addr < end; addr++) {
4207 ret = copy_to_user(uattr, attr, attr->size);
4215 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4216 * @pid: the pid in question.
4217 * @uattr: structure containing the extended parameters.
4218 * @size: sizeof(attr) for fwd/bwd comp.
4219 * @flags: for future extension.
4221 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4222 unsigned int, size, unsigned int, flags)
4224 struct sched_attr attr = {
4225 .size = sizeof(struct sched_attr),
4227 struct task_struct *p;
4230 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4231 size < SCHED_ATTR_SIZE_VER0 || flags)
4235 p = find_process_by_pid(pid);
4240 retval = security_task_getscheduler(p);
4244 attr.sched_policy = p->policy;
4245 if (p->sched_reset_on_fork)
4246 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4247 if (task_has_dl_policy(p))
4248 __getparam_dl(p, &attr);
4249 else if (task_has_rt_policy(p))
4250 attr.sched_priority = p->rt_priority;
4252 attr.sched_nice = task_nice(p);
4256 retval = sched_read_attr(uattr, &attr, size);
4264 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4266 cpumask_var_t cpus_allowed, new_mask;
4267 struct task_struct *p;
4272 p = find_process_by_pid(pid);
4278 /* Prevent p going away */
4282 if (p->flags & PF_NO_SETAFFINITY) {
4286 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4290 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4292 goto out_free_cpus_allowed;
4295 if (!check_same_owner(p)) {
4297 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4299 goto out_free_new_mask;
4304 retval = security_task_setscheduler(p);
4306 goto out_free_new_mask;
4309 cpuset_cpus_allowed(p, cpus_allowed);
4310 cpumask_and(new_mask, in_mask, cpus_allowed);
4313 * Since bandwidth control happens on root_domain basis,
4314 * if admission test is enabled, we only admit -deadline
4315 * tasks allowed to run on all the CPUs in the task's
4319 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4321 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4324 goto out_free_new_mask;
4330 retval = set_cpus_allowed_ptr(p, new_mask);
4333 cpuset_cpus_allowed(p, cpus_allowed);
4334 if (!cpumask_subset(new_mask, cpus_allowed)) {
4336 * We must have raced with a concurrent cpuset
4337 * update. Just reset the cpus_allowed to the
4338 * cpuset's cpus_allowed
4340 cpumask_copy(new_mask, cpus_allowed);
4345 free_cpumask_var(new_mask);
4346 out_free_cpus_allowed:
4347 free_cpumask_var(cpus_allowed);
4353 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4354 struct cpumask *new_mask)
4356 if (len < cpumask_size())
4357 cpumask_clear(new_mask);
4358 else if (len > cpumask_size())
4359 len = cpumask_size();
4361 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4365 * sys_sched_setaffinity - set the cpu affinity of a process
4366 * @pid: pid of the process
4367 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4368 * @user_mask_ptr: user-space pointer to the new cpu mask
4370 * Return: 0 on success. An error code otherwise.
4372 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4373 unsigned long __user *, user_mask_ptr)
4375 cpumask_var_t new_mask;
4378 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4381 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4383 retval = sched_setaffinity(pid, new_mask);
4384 free_cpumask_var(new_mask);
4388 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4390 struct task_struct *p;
4391 unsigned long flags;
4397 p = find_process_by_pid(pid);
4401 retval = security_task_getscheduler(p);
4405 raw_spin_lock_irqsave(&p->pi_lock, flags);
4406 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4407 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4416 * sys_sched_getaffinity - get the cpu affinity of a process
4417 * @pid: pid of the process
4418 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4419 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4421 * Return: 0 on success. An error code otherwise.
4423 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4424 unsigned long __user *, user_mask_ptr)
4429 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4431 if (len & (sizeof(unsigned long)-1))
4434 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4437 ret = sched_getaffinity(pid, mask);
4439 size_t retlen = min_t(size_t, len, cpumask_size());
4441 if (copy_to_user(user_mask_ptr, mask, retlen))
4446 free_cpumask_var(mask);
4452 * sys_sched_yield - yield the current processor to other threads.
4454 * This function yields the current CPU to other tasks. If there are no
4455 * other threads running on this CPU then this function will return.
4459 SYSCALL_DEFINE0(sched_yield)
4461 struct rq *rq = this_rq_lock();
4463 schedstat_inc(rq, yld_count);
4464 current->sched_class->yield_task(rq);
4467 * Since we are going to call schedule() anyway, there's
4468 * no need to preempt or enable interrupts:
4470 __release(rq->lock);
4471 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4472 do_raw_spin_unlock(&rq->lock);
4473 sched_preempt_enable_no_resched();
4480 int __sched _cond_resched(void)
4482 if (should_resched()) {
4483 preempt_schedule_common();
4488 EXPORT_SYMBOL(_cond_resched);
4491 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4492 * call schedule, and on return reacquire the lock.
4494 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4495 * operations here to prevent schedule() from being called twice (once via
4496 * spin_unlock(), once by hand).
4498 int __cond_resched_lock(spinlock_t *lock)
4500 int resched = should_resched();
4503 lockdep_assert_held(lock);
4505 if (spin_needbreak(lock) || resched) {
4508 preempt_schedule_common();
4516 EXPORT_SYMBOL(__cond_resched_lock);
4518 int __sched __cond_resched_softirq(void)
4520 BUG_ON(!in_softirq());
4522 if (should_resched()) {
4524 preempt_schedule_common();
4530 EXPORT_SYMBOL(__cond_resched_softirq);
4533 * yield - yield the current processor to other threads.
4535 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4537 * The scheduler is at all times free to pick the calling task as the most
4538 * eligible task to run, if removing the yield() call from your code breaks
4539 * it, its already broken.
4541 * Typical broken usage is:
4546 * where one assumes that yield() will let 'the other' process run that will
4547 * make event true. If the current task is a SCHED_FIFO task that will never
4548 * happen. Never use yield() as a progress guarantee!!
4550 * If you want to use yield() to wait for something, use wait_event().
4551 * If you want to use yield() to be 'nice' for others, use cond_resched().
4552 * If you still want to use yield(), do not!
4554 void __sched yield(void)
4556 set_current_state(TASK_RUNNING);
4559 EXPORT_SYMBOL(yield);
4562 * yield_to - yield the current processor to another thread in
4563 * your thread group, or accelerate that thread toward the
4564 * processor it's on.
4566 * @preempt: whether task preemption is allowed or not
4568 * It's the caller's job to ensure that the target task struct
4569 * can't go away on us before we can do any checks.
4572 * true (>0) if we indeed boosted the target task.
4573 * false (0) if we failed to boost the target.
4574 * -ESRCH if there's no task to yield to.
4576 int __sched yield_to(struct task_struct *p, bool preempt)
4578 struct task_struct *curr = current;
4579 struct rq *rq, *p_rq;
4580 unsigned long flags;
4583 local_irq_save(flags);
4589 * If we're the only runnable task on the rq and target rq also
4590 * has only one task, there's absolutely no point in yielding.
4592 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4597 double_rq_lock(rq, p_rq);
4598 if (task_rq(p) != p_rq) {
4599 double_rq_unlock(rq, p_rq);
4603 if (!curr->sched_class->yield_to_task)
4606 if (curr->sched_class != p->sched_class)
4609 if (task_running(p_rq, p) || p->state)
4612 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4614 schedstat_inc(rq, yld_count);
4616 * Make p's CPU reschedule; pick_next_entity takes care of
4619 if (preempt && rq != p_rq)
4624 double_rq_unlock(rq, p_rq);
4626 local_irq_restore(flags);
4633 EXPORT_SYMBOL_GPL(yield_to);
4636 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4637 * that process accounting knows that this is a task in IO wait state.
4639 long __sched io_schedule_timeout(long timeout)
4641 int old_iowait = current->in_iowait;
4645 current->in_iowait = 1;
4646 blk_schedule_flush_plug(current);
4648 delayacct_blkio_start();
4650 atomic_inc(&rq->nr_iowait);
4651 ret = schedule_timeout(timeout);
4652 current->in_iowait = old_iowait;
4653 atomic_dec(&rq->nr_iowait);
4654 delayacct_blkio_end();
4658 EXPORT_SYMBOL(io_schedule_timeout);
4661 * sys_sched_get_priority_max - return maximum RT priority.
4662 * @policy: scheduling class.
4664 * Return: On success, this syscall returns the maximum
4665 * rt_priority that can be used by a given scheduling class.
4666 * On failure, a negative error code is returned.
4668 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4675 ret = MAX_USER_RT_PRIO-1;
4677 case SCHED_DEADLINE:
4688 * sys_sched_get_priority_min - return minimum RT priority.
4689 * @policy: scheduling class.
4691 * Return: On success, this syscall returns the minimum
4692 * rt_priority that can be used by a given scheduling class.
4693 * On failure, a negative error code is returned.
4695 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4704 case SCHED_DEADLINE:
4714 * sys_sched_rr_get_interval - return the default timeslice of a process.
4715 * @pid: pid of the process.
4716 * @interval: userspace pointer to the timeslice value.
4718 * this syscall writes the default timeslice value of a given process
4719 * into the user-space timespec buffer. A value of '0' means infinity.
4721 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4724 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4725 struct timespec __user *, interval)
4727 struct task_struct *p;
4728 unsigned int time_slice;
4729 unsigned long flags;
4739 p = find_process_by_pid(pid);
4743 retval = security_task_getscheduler(p);
4747 rq = task_rq_lock(p, &flags);
4749 if (p->sched_class->get_rr_interval)
4750 time_slice = p->sched_class->get_rr_interval(rq, p);
4751 task_rq_unlock(rq, p, &flags);
4754 jiffies_to_timespec(time_slice, &t);
4755 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4763 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4765 void sched_show_task(struct task_struct *p)
4767 unsigned long free = 0;
4769 unsigned long state = p->state;
4772 state = __ffs(state) + 1;
4773 printk(KERN_INFO "%-15.15s %c", p->comm,
4774 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4775 #if BITS_PER_LONG == 32
4776 if (state == TASK_RUNNING)
4777 printk(KERN_CONT " running ");
4779 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4781 if (state == TASK_RUNNING)
4782 printk(KERN_CONT " running task ");
4784 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4786 #ifdef CONFIG_DEBUG_STACK_USAGE
4787 free = stack_not_used(p);
4792 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4794 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4795 task_pid_nr(p), ppid,
4796 (unsigned long)task_thread_info(p)->flags);
4798 print_worker_info(KERN_INFO, p);
4799 show_stack(p, NULL);
4802 void show_state_filter(unsigned long state_filter)
4804 struct task_struct *g, *p;
4806 #if BITS_PER_LONG == 32
4808 " task PC stack pid father\n");
4811 " task PC stack pid father\n");
4814 for_each_process_thread(g, p) {
4816 * reset the NMI-timeout, listing all files on a slow
4817 * console might take a lot of time:
4819 touch_nmi_watchdog();
4820 if (!state_filter || (p->state & state_filter))
4824 touch_all_softlockup_watchdogs();
4826 #ifdef CONFIG_SCHED_DEBUG
4827 sysrq_sched_debug_show();
4831 * Only show locks if all tasks are dumped:
4834 debug_show_all_locks();
4837 void init_idle_bootup_task(struct task_struct *idle)
4839 idle->sched_class = &idle_sched_class;
4843 * init_idle - set up an idle thread for a given CPU
4844 * @idle: task in question
4845 * @cpu: cpu the idle task belongs to
4847 * NOTE: this function does not set the idle thread's NEED_RESCHED
4848 * flag, to make booting more robust.
4850 void init_idle(struct task_struct *idle, int cpu)
4852 struct rq *rq = cpu_rq(cpu);
4853 unsigned long flags;
4855 raw_spin_lock_irqsave(&rq->lock, flags);
4857 __sched_fork(0, idle);
4858 idle->state = TASK_RUNNING;
4859 idle->se.exec_start = sched_clock();
4861 do_set_cpus_allowed(idle, cpumask_of(cpu));
4863 * We're having a chicken and egg problem, even though we are
4864 * holding rq->lock, the cpu isn't yet set to this cpu so the
4865 * lockdep check in task_group() will fail.
4867 * Similar case to sched_fork(). / Alternatively we could
4868 * use task_rq_lock() here and obtain the other rq->lock.
4873 __set_task_cpu(idle, cpu);
4876 rq->curr = rq->idle = idle;
4877 idle->on_rq = TASK_ON_RQ_QUEUED;
4878 #if defined(CONFIG_SMP)
4881 raw_spin_unlock_irqrestore(&rq->lock, flags);
4883 /* Set the preempt count _outside_ the spinlocks! */
4884 init_idle_preempt_count(idle, cpu);
4887 * The idle tasks have their own, simple scheduling class:
4889 idle->sched_class = &idle_sched_class;
4890 ftrace_graph_init_idle_task(idle, cpu);
4891 vtime_init_idle(idle, cpu);
4892 #if defined(CONFIG_SMP)
4893 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4897 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4898 const struct cpumask *trial)
4900 int ret = 1, trial_cpus;
4901 struct dl_bw *cur_dl_b;
4902 unsigned long flags;
4904 if (!cpumask_weight(cur))
4907 rcu_read_lock_sched();
4908 cur_dl_b = dl_bw_of(cpumask_any(cur));
4909 trial_cpus = cpumask_weight(trial);
4911 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4912 if (cur_dl_b->bw != -1 &&
4913 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4915 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4916 rcu_read_unlock_sched();
4921 int task_can_attach(struct task_struct *p,
4922 const struct cpumask *cs_cpus_allowed)
4927 * Kthreads which disallow setaffinity shouldn't be moved
4928 * to a new cpuset; we don't want to change their cpu
4929 * affinity and isolating such threads by their set of
4930 * allowed nodes is unnecessary. Thus, cpusets are not
4931 * applicable for such threads. This prevents checking for
4932 * success of set_cpus_allowed_ptr() on all attached tasks
4933 * before cpus_allowed may be changed.
4935 if (p->flags & PF_NO_SETAFFINITY) {
4941 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4943 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4948 unsigned long flags;
4950 rcu_read_lock_sched();
4951 dl_b = dl_bw_of(dest_cpu);
4952 raw_spin_lock_irqsave(&dl_b->lock, flags);
4953 cpus = dl_bw_cpus(dest_cpu);
4954 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4959 * We reserve space for this task in the destination
4960 * root_domain, as we can't fail after this point.
4961 * We will free resources in the source root_domain
4962 * later on (see set_cpus_allowed_dl()).
4964 __dl_add(dl_b, p->dl.dl_bw);
4966 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4967 rcu_read_unlock_sched();
4977 #ifdef CONFIG_NUMA_BALANCING
4978 /* Migrate current task p to target_cpu */
4979 int migrate_task_to(struct task_struct *p, int target_cpu)
4981 struct migration_arg arg = { p, target_cpu };
4982 int curr_cpu = task_cpu(p);
4984 if (curr_cpu == target_cpu)
4987 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4990 /* TODO: This is not properly updating schedstats */
4992 trace_sched_move_numa(p, curr_cpu, target_cpu);
4993 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4997 * Requeue a task on a given node and accurately track the number of NUMA
4998 * tasks on the runqueues
5000 void sched_setnuma(struct task_struct *p, int nid)
5003 unsigned long flags;
5004 bool queued, running;
5006 rq = task_rq_lock(p, &flags);
5007 queued = task_on_rq_queued(p);
5008 running = task_current(rq, p);
5011 dequeue_task(rq, p, 0);
5013 put_prev_task(rq, p);
5015 p->numa_preferred_nid = nid;
5018 p->sched_class->set_curr_task(rq);
5020 enqueue_task(rq, p, 0);
5021 task_rq_unlock(rq, p, &flags);
5023 #endif /* CONFIG_NUMA_BALANCING */
5025 #ifdef CONFIG_HOTPLUG_CPU
5027 * Ensures that the idle task is using init_mm right before its cpu goes
5030 void idle_task_exit(void)
5032 struct mm_struct *mm = current->active_mm;
5034 BUG_ON(cpu_online(smp_processor_id()));
5036 if (mm != &init_mm) {
5037 switch_mm(mm, &init_mm, current);
5038 finish_arch_post_lock_switch();
5044 * Since this CPU is going 'away' for a while, fold any nr_active delta
5045 * we might have. Assumes we're called after migrate_tasks() so that the
5046 * nr_active count is stable.
5048 * Also see the comment "Global load-average calculations".
5050 static void calc_load_migrate(struct rq *rq)
5052 long delta = calc_load_fold_active(rq);
5054 atomic_long_add(delta, &calc_load_tasks);
5057 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5061 static const struct sched_class fake_sched_class = {
5062 .put_prev_task = put_prev_task_fake,
5065 static struct task_struct fake_task = {
5067 * Avoid pull_{rt,dl}_task()
5069 .prio = MAX_PRIO + 1,
5070 .sched_class = &fake_sched_class,
5074 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5075 * try_to_wake_up()->select_task_rq().
5077 * Called with rq->lock held even though we'er in stop_machine() and
5078 * there's no concurrency possible, we hold the required locks anyway
5079 * because of lock validation efforts.
5081 static void migrate_tasks(struct rq *dead_rq)
5083 struct rq *rq = dead_rq;
5084 struct task_struct *next, *stop = rq->stop;
5088 * Fudge the rq selection such that the below task selection loop
5089 * doesn't get stuck on the currently eligible stop task.
5091 * We're currently inside stop_machine() and the rq is either stuck
5092 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5093 * either way we should never end up calling schedule() until we're
5099 * put_prev_task() and pick_next_task() sched
5100 * class method both need to have an up-to-date
5101 * value of rq->clock[_task]
5103 update_rq_clock(rq);
5107 * There's this thread running, bail when that's the only
5110 if (rq->nr_running == 1)
5114 * Ensure rq->lock covers the entire task selection
5115 * until the migration.
5117 lockdep_pin_lock(&rq->lock);
5118 next = pick_next_task(rq, &fake_task);
5120 next->sched_class->put_prev_task(rq, next);
5122 /* Find suitable destination for @next, with force if needed. */
5123 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5125 lockdep_unpin_lock(&rq->lock);
5126 rq = __migrate_task(rq, next, dest_cpu);
5127 if (rq != dead_rq) {
5128 raw_spin_unlock(&rq->lock);
5130 raw_spin_lock(&rq->lock);
5136 #endif /* CONFIG_HOTPLUG_CPU */
5138 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5140 static struct ctl_table sd_ctl_dir[] = {
5142 .procname = "sched_domain",
5148 static struct ctl_table sd_ctl_root[] = {
5150 .procname = "kernel",
5152 .child = sd_ctl_dir,
5157 static struct ctl_table *sd_alloc_ctl_entry(int n)
5159 struct ctl_table *entry =
5160 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5165 static void sd_free_ctl_entry(struct ctl_table **tablep)
5167 struct ctl_table *entry;
5170 * In the intermediate directories, both the child directory and
5171 * procname are dynamically allocated and could fail but the mode
5172 * will always be set. In the lowest directory the names are
5173 * static strings and all have proc handlers.
5175 for (entry = *tablep; entry->mode; entry++) {
5177 sd_free_ctl_entry(&entry->child);
5178 if (entry->proc_handler == NULL)
5179 kfree(entry->procname);
5186 static int min_load_idx = 0;
5187 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5190 set_table_entry(struct ctl_table *entry,
5191 const char *procname, void *data, int maxlen,
5192 umode_t mode, proc_handler *proc_handler,
5195 entry->procname = procname;
5197 entry->maxlen = maxlen;
5199 entry->proc_handler = proc_handler;
5202 entry->extra1 = &min_load_idx;
5203 entry->extra2 = &max_load_idx;
5207 static struct ctl_table *
5208 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5210 struct ctl_table *table = sd_alloc_ctl_entry(14);
5215 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5216 sizeof(long), 0644, proc_doulongvec_minmax, false);
5217 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5218 sizeof(long), 0644, proc_doulongvec_minmax, false);
5219 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5220 sizeof(int), 0644, proc_dointvec_minmax, true);
5221 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5222 sizeof(int), 0644, proc_dointvec_minmax, true);
5223 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5224 sizeof(int), 0644, proc_dointvec_minmax, true);
5225 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5226 sizeof(int), 0644, proc_dointvec_minmax, true);
5227 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5228 sizeof(int), 0644, proc_dointvec_minmax, true);
5229 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5230 sizeof(int), 0644, proc_dointvec_minmax, false);
5231 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5232 sizeof(int), 0644, proc_dointvec_minmax, false);
5233 set_table_entry(&table[9], "cache_nice_tries",
5234 &sd->cache_nice_tries,
5235 sizeof(int), 0644, proc_dointvec_minmax, false);
5236 set_table_entry(&table[10], "flags", &sd->flags,
5237 sizeof(int), 0644, proc_dointvec_minmax, false);
5238 set_table_entry(&table[11], "max_newidle_lb_cost",
5239 &sd->max_newidle_lb_cost,
5240 sizeof(long), 0644, proc_doulongvec_minmax, false);
5241 set_table_entry(&table[12], "name", sd->name,
5242 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5243 /* &table[13] is terminator */
5248 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5250 struct ctl_table *entry, *table;
5251 struct sched_domain *sd;
5252 int domain_num = 0, i;
5255 for_each_domain(cpu, sd)
5257 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5262 for_each_domain(cpu, sd) {
5263 snprintf(buf, 32, "domain%d", i);
5264 entry->procname = kstrdup(buf, GFP_KERNEL);
5266 entry->child = sd_alloc_ctl_domain_table(sd);
5273 static struct ctl_table_header *sd_sysctl_header;
5274 static void register_sched_domain_sysctl(void)
5276 int i, cpu_num = num_possible_cpus();
5277 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5280 WARN_ON(sd_ctl_dir[0].child);
5281 sd_ctl_dir[0].child = entry;
5286 for_each_possible_cpu(i) {
5287 snprintf(buf, 32, "cpu%d", i);
5288 entry->procname = kstrdup(buf, GFP_KERNEL);
5290 entry->child = sd_alloc_ctl_cpu_table(i);
5294 WARN_ON(sd_sysctl_header);
5295 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5298 /* may be called multiple times per register */
5299 static void unregister_sched_domain_sysctl(void)
5301 if (sd_sysctl_header)
5302 unregister_sysctl_table(sd_sysctl_header);
5303 sd_sysctl_header = NULL;
5304 if (sd_ctl_dir[0].child)
5305 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5308 static void register_sched_domain_sysctl(void)
5311 static void unregister_sched_domain_sysctl(void)
5314 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5316 static void set_rq_online(struct rq *rq)
5319 const struct sched_class *class;
5321 cpumask_set_cpu(rq->cpu, rq->rd->online);
5324 for_each_class(class) {
5325 if (class->rq_online)
5326 class->rq_online(rq);
5331 static void set_rq_offline(struct rq *rq)
5334 const struct sched_class *class;
5336 for_each_class(class) {
5337 if (class->rq_offline)
5338 class->rq_offline(rq);
5341 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5347 * migration_call - callback that gets triggered when a CPU is added.
5348 * Here we can start up the necessary migration thread for the new CPU.
5351 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5353 int cpu = (long)hcpu;
5354 unsigned long flags;
5355 struct rq *rq = cpu_rq(cpu);
5357 switch (action & ~CPU_TASKS_FROZEN) {
5359 case CPU_UP_PREPARE:
5360 rq->calc_load_update = calc_load_update;
5364 /* Update our root-domain */
5365 raw_spin_lock_irqsave(&rq->lock, flags);
5367 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5371 raw_spin_unlock_irqrestore(&rq->lock, flags);
5374 #ifdef CONFIG_HOTPLUG_CPU
5376 sched_ttwu_pending();
5377 /* Update our root-domain */
5378 raw_spin_lock_irqsave(&rq->lock, flags);
5380 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5384 BUG_ON(rq->nr_running != 1); /* the migration thread */
5385 raw_spin_unlock_irqrestore(&rq->lock, flags);
5389 calc_load_migrate(rq);
5394 update_max_interval();
5400 * Register at high priority so that task migration (migrate_all_tasks)
5401 * happens before everything else. This has to be lower priority than
5402 * the notifier in the perf_event subsystem, though.
5404 static struct notifier_block migration_notifier = {
5405 .notifier_call = migration_call,
5406 .priority = CPU_PRI_MIGRATION,
5409 static void set_cpu_rq_start_time(void)
5411 int cpu = smp_processor_id();
5412 struct rq *rq = cpu_rq(cpu);
5413 rq->age_stamp = sched_clock_cpu(cpu);
5416 static int sched_cpu_active(struct notifier_block *nfb,
5417 unsigned long action, void *hcpu)
5419 switch (action & ~CPU_TASKS_FROZEN) {
5421 set_cpu_rq_start_time();
5423 case CPU_DOWN_FAILED:
5424 set_cpu_active((long)hcpu, true);
5431 static int sched_cpu_inactive(struct notifier_block *nfb,
5432 unsigned long action, void *hcpu)
5434 switch (action & ~CPU_TASKS_FROZEN) {
5435 case CPU_DOWN_PREPARE:
5436 set_cpu_active((long)hcpu, false);
5443 static int __init migration_init(void)
5445 void *cpu = (void *)(long)smp_processor_id();
5448 /* Initialize migration for the boot CPU */
5449 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5450 BUG_ON(err == NOTIFY_BAD);
5451 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5452 register_cpu_notifier(&migration_notifier);
5454 /* Register cpu active notifiers */
5455 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5456 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5460 early_initcall(migration_init);
5462 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5464 #ifdef CONFIG_SCHED_DEBUG
5466 static __read_mostly int sched_debug_enabled;
5468 static int __init sched_debug_setup(char *str)
5470 sched_debug_enabled = 1;
5474 early_param("sched_debug", sched_debug_setup);
5476 static inline bool sched_debug(void)
5478 return sched_debug_enabled;
5481 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5482 struct cpumask *groupmask)
5484 struct sched_group *group = sd->groups;
5486 cpumask_clear(groupmask);
5488 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5490 if (!(sd->flags & SD_LOAD_BALANCE)) {
5491 printk("does not load-balance\n");
5493 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5498 printk(KERN_CONT "span %*pbl level %s\n",
5499 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5501 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5502 printk(KERN_ERR "ERROR: domain->span does not contain "
5505 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5506 printk(KERN_ERR "ERROR: domain->groups does not contain"
5510 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5514 printk(KERN_ERR "ERROR: group is NULL\n");
5518 if (!cpumask_weight(sched_group_cpus(group))) {
5519 printk(KERN_CONT "\n");
5520 printk(KERN_ERR "ERROR: empty group\n");
5524 if (!(sd->flags & SD_OVERLAP) &&
5525 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5526 printk(KERN_CONT "\n");
5527 printk(KERN_ERR "ERROR: repeated CPUs\n");
5531 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5533 printk(KERN_CONT " %*pbl",
5534 cpumask_pr_args(sched_group_cpus(group)));
5535 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5536 printk(KERN_CONT " (cpu_capacity = %d)",
5537 group->sgc->capacity);
5540 group = group->next;
5541 } while (group != sd->groups);
5542 printk(KERN_CONT "\n");
5544 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5545 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5548 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5549 printk(KERN_ERR "ERROR: parent span is not a superset "
5550 "of domain->span\n");
5554 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5558 if (!sched_debug_enabled)
5562 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5566 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5569 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5577 #else /* !CONFIG_SCHED_DEBUG */
5578 # define sched_domain_debug(sd, cpu) do { } while (0)
5579 static inline bool sched_debug(void)
5583 #endif /* CONFIG_SCHED_DEBUG */
5585 static int sd_degenerate(struct sched_domain *sd)
5587 if (cpumask_weight(sched_domain_span(sd)) == 1)
5590 /* Following flags need at least 2 groups */
5591 if (sd->flags & (SD_LOAD_BALANCE |
5592 SD_BALANCE_NEWIDLE |
5595 SD_SHARE_CPUCAPACITY |
5596 SD_SHARE_PKG_RESOURCES |
5597 SD_SHARE_POWERDOMAIN)) {
5598 if (sd->groups != sd->groups->next)
5602 /* Following flags don't use groups */
5603 if (sd->flags & (SD_WAKE_AFFINE))
5610 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5612 unsigned long cflags = sd->flags, pflags = parent->flags;
5614 if (sd_degenerate(parent))
5617 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5620 /* Flags needing groups don't count if only 1 group in parent */
5621 if (parent->groups == parent->groups->next) {
5622 pflags &= ~(SD_LOAD_BALANCE |
5623 SD_BALANCE_NEWIDLE |
5626 SD_SHARE_CPUCAPACITY |
5627 SD_SHARE_PKG_RESOURCES |
5629 SD_SHARE_POWERDOMAIN);
5630 if (nr_node_ids == 1)
5631 pflags &= ~SD_SERIALIZE;
5633 if (~cflags & pflags)
5639 static void free_rootdomain(struct rcu_head *rcu)
5641 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5643 cpupri_cleanup(&rd->cpupri);
5644 cpudl_cleanup(&rd->cpudl);
5645 free_cpumask_var(rd->dlo_mask);
5646 free_cpumask_var(rd->rto_mask);
5647 free_cpumask_var(rd->online);
5648 free_cpumask_var(rd->span);
5652 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5654 struct root_domain *old_rd = NULL;
5655 unsigned long flags;
5657 raw_spin_lock_irqsave(&rq->lock, flags);
5662 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5665 cpumask_clear_cpu(rq->cpu, old_rd->span);
5668 * If we dont want to free the old_rd yet then
5669 * set old_rd to NULL to skip the freeing later
5672 if (!atomic_dec_and_test(&old_rd->refcount))
5676 atomic_inc(&rd->refcount);
5679 cpumask_set_cpu(rq->cpu, rd->span);
5680 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5683 raw_spin_unlock_irqrestore(&rq->lock, flags);
5686 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5689 static int init_rootdomain(struct root_domain *rd)
5691 memset(rd, 0, sizeof(*rd));
5693 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5695 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5697 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5699 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5702 init_dl_bw(&rd->dl_bw);
5703 if (cpudl_init(&rd->cpudl) != 0)
5706 if (cpupri_init(&rd->cpupri) != 0)
5711 free_cpumask_var(rd->rto_mask);
5713 free_cpumask_var(rd->dlo_mask);
5715 free_cpumask_var(rd->online);
5717 free_cpumask_var(rd->span);
5723 * By default the system creates a single root-domain with all cpus as
5724 * members (mimicking the global state we have today).
5726 struct root_domain def_root_domain;
5728 static void init_defrootdomain(void)
5730 init_rootdomain(&def_root_domain);
5732 atomic_set(&def_root_domain.refcount, 1);
5735 static struct root_domain *alloc_rootdomain(void)
5737 struct root_domain *rd;
5739 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5743 if (init_rootdomain(rd) != 0) {
5751 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5753 struct sched_group *tmp, *first;
5762 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5767 } while (sg != first);
5770 static void free_sched_domain(struct rcu_head *rcu)
5772 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5775 * If its an overlapping domain it has private groups, iterate and
5778 if (sd->flags & SD_OVERLAP) {
5779 free_sched_groups(sd->groups, 1);
5780 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5781 kfree(sd->groups->sgc);
5787 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5789 call_rcu(&sd->rcu, free_sched_domain);
5792 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5794 for (; sd; sd = sd->parent)
5795 destroy_sched_domain(sd, cpu);
5799 * Keep a special pointer to the highest sched_domain that has
5800 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5801 * allows us to avoid some pointer chasing select_idle_sibling().
5803 * Also keep a unique ID per domain (we use the first cpu number in
5804 * the cpumask of the domain), this allows us to quickly tell if
5805 * two cpus are in the same cache domain, see cpus_share_cache().
5807 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5808 DEFINE_PER_CPU(int, sd_llc_size);
5809 DEFINE_PER_CPU(int, sd_llc_id);
5810 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5811 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5812 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5814 static void update_top_cache_domain(int cpu)
5816 struct sched_domain *sd;
5817 struct sched_domain *busy_sd = NULL;
5821 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5823 id = cpumask_first(sched_domain_span(sd));
5824 size = cpumask_weight(sched_domain_span(sd));
5825 busy_sd = sd->parent; /* sd_busy */
5827 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5829 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5830 per_cpu(sd_llc_size, cpu) = size;
5831 per_cpu(sd_llc_id, cpu) = id;
5833 sd = lowest_flag_domain(cpu, SD_NUMA);
5834 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5836 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5837 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5841 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5842 * hold the hotplug lock.
5845 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5847 struct rq *rq = cpu_rq(cpu);
5848 struct sched_domain *tmp;
5850 /* Remove the sched domains which do not contribute to scheduling. */
5851 for (tmp = sd; tmp; ) {
5852 struct sched_domain *parent = tmp->parent;
5856 if (sd_parent_degenerate(tmp, parent)) {
5857 tmp->parent = parent->parent;
5859 parent->parent->child = tmp;
5861 * Transfer SD_PREFER_SIBLING down in case of a
5862 * degenerate parent; the spans match for this
5863 * so the property transfers.
5865 if (parent->flags & SD_PREFER_SIBLING)
5866 tmp->flags |= SD_PREFER_SIBLING;
5867 destroy_sched_domain(parent, cpu);
5872 if (sd && sd_degenerate(sd)) {
5875 destroy_sched_domain(tmp, cpu);
5880 sched_domain_debug(sd, cpu);
5882 rq_attach_root(rq, rd);
5884 rcu_assign_pointer(rq->sd, sd);
5885 destroy_sched_domains(tmp, cpu);
5887 update_top_cache_domain(cpu);
5890 /* Setup the mask of cpus configured for isolated domains */
5891 static int __init isolated_cpu_setup(char *str)
5893 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5894 cpulist_parse(str, cpu_isolated_map);
5898 __setup("isolcpus=", isolated_cpu_setup);
5901 struct sched_domain ** __percpu sd;
5902 struct root_domain *rd;
5913 * Build an iteration mask that can exclude certain CPUs from the upwards
5916 * Asymmetric node setups can result in situations where the domain tree is of
5917 * unequal depth, make sure to skip domains that already cover the entire
5920 * In that case build_sched_domains() will have terminated the iteration early
5921 * and our sibling sd spans will be empty. Domains should always include the
5922 * cpu they're built on, so check that.
5925 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5927 const struct cpumask *span = sched_domain_span(sd);
5928 struct sd_data *sdd = sd->private;
5929 struct sched_domain *sibling;
5932 for_each_cpu(i, span) {
5933 sibling = *per_cpu_ptr(sdd->sd, i);
5934 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5937 cpumask_set_cpu(i, sched_group_mask(sg));
5942 * Return the canonical balance cpu for this group, this is the first cpu
5943 * of this group that's also in the iteration mask.
5945 int group_balance_cpu(struct sched_group *sg)
5947 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5951 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5953 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5954 const struct cpumask *span = sched_domain_span(sd);
5955 struct cpumask *covered = sched_domains_tmpmask;
5956 struct sd_data *sdd = sd->private;
5957 struct sched_domain *sibling;
5960 cpumask_clear(covered);
5962 for_each_cpu(i, span) {
5963 struct cpumask *sg_span;
5965 if (cpumask_test_cpu(i, covered))
5968 sibling = *per_cpu_ptr(sdd->sd, i);
5970 /* See the comment near build_group_mask(). */
5971 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5974 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5975 GFP_KERNEL, cpu_to_node(cpu));
5980 sg_span = sched_group_cpus(sg);
5982 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5984 cpumask_set_cpu(i, sg_span);
5986 cpumask_or(covered, covered, sg_span);
5988 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5989 if (atomic_inc_return(&sg->sgc->ref) == 1)
5990 build_group_mask(sd, sg);
5993 * Initialize sgc->capacity such that even if we mess up the
5994 * domains and no possible iteration will get us here, we won't
5997 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6000 * Make sure the first group of this domain contains the
6001 * canonical balance cpu. Otherwise the sched_domain iteration
6002 * breaks. See update_sg_lb_stats().
6004 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6005 group_balance_cpu(sg) == cpu)
6015 sd->groups = groups;
6020 free_sched_groups(first, 0);
6025 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6027 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6028 struct sched_domain *child = sd->child;
6031 cpu = cpumask_first(sched_domain_span(child));
6034 *sg = *per_cpu_ptr(sdd->sg, cpu);
6035 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6036 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6043 * build_sched_groups will build a circular linked list of the groups
6044 * covered by the given span, and will set each group's ->cpumask correctly,
6045 * and ->cpu_capacity to 0.
6047 * Assumes the sched_domain tree is fully constructed
6050 build_sched_groups(struct sched_domain *sd, int cpu)
6052 struct sched_group *first = NULL, *last = NULL;
6053 struct sd_data *sdd = sd->private;
6054 const struct cpumask *span = sched_domain_span(sd);
6055 struct cpumask *covered;
6058 get_group(cpu, sdd, &sd->groups);
6059 atomic_inc(&sd->groups->ref);
6061 if (cpu != cpumask_first(span))
6064 lockdep_assert_held(&sched_domains_mutex);
6065 covered = sched_domains_tmpmask;
6067 cpumask_clear(covered);
6069 for_each_cpu(i, span) {
6070 struct sched_group *sg;
6073 if (cpumask_test_cpu(i, covered))
6076 group = get_group(i, sdd, &sg);
6077 cpumask_setall(sched_group_mask(sg));
6079 for_each_cpu(j, span) {
6080 if (get_group(j, sdd, NULL) != group)
6083 cpumask_set_cpu(j, covered);
6084 cpumask_set_cpu(j, sched_group_cpus(sg));
6099 * Initialize sched groups cpu_capacity.
6101 * cpu_capacity indicates the capacity of sched group, which is used while
6102 * distributing the load between different sched groups in a sched domain.
6103 * Typically cpu_capacity for all the groups in a sched domain will be same
6104 * unless there are asymmetries in the topology. If there are asymmetries,
6105 * group having more cpu_capacity will pickup more load compared to the
6106 * group having less cpu_capacity.
6108 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6110 struct sched_group *sg = sd->groups;
6115 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6117 } while (sg != sd->groups);
6119 if (cpu != group_balance_cpu(sg))
6122 update_group_capacity(sd, cpu);
6123 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6127 * Initializers for schedule domains
6128 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6131 static int default_relax_domain_level = -1;
6132 int sched_domain_level_max;
6134 static int __init setup_relax_domain_level(char *str)
6136 if (kstrtoint(str, 0, &default_relax_domain_level))
6137 pr_warn("Unable to set relax_domain_level\n");
6141 __setup("relax_domain_level=", setup_relax_domain_level);
6143 static void set_domain_attribute(struct sched_domain *sd,
6144 struct sched_domain_attr *attr)
6148 if (!attr || attr->relax_domain_level < 0) {
6149 if (default_relax_domain_level < 0)
6152 request = default_relax_domain_level;
6154 request = attr->relax_domain_level;
6155 if (request < sd->level) {
6156 /* turn off idle balance on this domain */
6157 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6159 /* turn on idle balance on this domain */
6160 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6164 static void __sdt_free(const struct cpumask *cpu_map);
6165 static int __sdt_alloc(const struct cpumask *cpu_map);
6167 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6168 const struct cpumask *cpu_map)
6172 if (!atomic_read(&d->rd->refcount))
6173 free_rootdomain(&d->rd->rcu); /* fall through */
6175 free_percpu(d->sd); /* fall through */
6177 __sdt_free(cpu_map); /* fall through */
6183 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6184 const struct cpumask *cpu_map)
6186 memset(d, 0, sizeof(*d));
6188 if (__sdt_alloc(cpu_map))
6189 return sa_sd_storage;
6190 d->sd = alloc_percpu(struct sched_domain *);
6192 return sa_sd_storage;
6193 d->rd = alloc_rootdomain();
6196 return sa_rootdomain;
6200 * NULL the sd_data elements we've used to build the sched_domain and
6201 * sched_group structure so that the subsequent __free_domain_allocs()
6202 * will not free the data we're using.
6204 static void claim_allocations(int cpu, struct sched_domain *sd)
6206 struct sd_data *sdd = sd->private;
6208 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6209 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6211 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6212 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6214 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6215 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6219 static int sched_domains_numa_levels;
6220 enum numa_topology_type sched_numa_topology_type;
6221 static int *sched_domains_numa_distance;
6222 int sched_max_numa_distance;
6223 static struct cpumask ***sched_domains_numa_masks;
6224 static int sched_domains_curr_level;
6228 * SD_flags allowed in topology descriptions.
6230 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6231 * SD_SHARE_PKG_RESOURCES - describes shared caches
6232 * SD_NUMA - describes NUMA topologies
6233 * SD_SHARE_POWERDOMAIN - describes shared power domain
6236 * SD_ASYM_PACKING - describes SMT quirks
6238 #define TOPOLOGY_SD_FLAGS \
6239 (SD_SHARE_CPUCAPACITY | \
6240 SD_SHARE_PKG_RESOURCES | \
6243 SD_SHARE_POWERDOMAIN)
6245 static struct sched_domain *
6246 sd_init(struct sched_domain_topology_level *tl, int cpu)
6248 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6249 int sd_weight, sd_flags = 0;
6253 * Ugly hack to pass state to sd_numa_mask()...
6255 sched_domains_curr_level = tl->numa_level;
6258 sd_weight = cpumask_weight(tl->mask(cpu));
6261 sd_flags = (*tl->sd_flags)();
6262 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6263 "wrong sd_flags in topology description\n"))
6264 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6266 *sd = (struct sched_domain){
6267 .min_interval = sd_weight,
6268 .max_interval = 2*sd_weight,
6270 .imbalance_pct = 125,
6272 .cache_nice_tries = 0,
6279 .flags = 1*SD_LOAD_BALANCE
6280 | 1*SD_BALANCE_NEWIDLE
6285 | 0*SD_SHARE_CPUCAPACITY
6286 | 0*SD_SHARE_PKG_RESOURCES
6288 | 0*SD_PREFER_SIBLING
6293 .last_balance = jiffies,
6294 .balance_interval = sd_weight,
6296 .max_newidle_lb_cost = 0,
6297 .next_decay_max_lb_cost = jiffies,
6298 #ifdef CONFIG_SCHED_DEBUG
6304 * Convert topological properties into behaviour.
6307 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6308 sd->flags |= SD_PREFER_SIBLING;
6309 sd->imbalance_pct = 110;
6310 sd->smt_gain = 1178; /* ~15% */
6312 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6313 sd->imbalance_pct = 117;
6314 sd->cache_nice_tries = 1;
6318 } else if (sd->flags & SD_NUMA) {
6319 sd->cache_nice_tries = 2;
6323 sd->flags |= SD_SERIALIZE;
6324 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6325 sd->flags &= ~(SD_BALANCE_EXEC |
6332 sd->flags |= SD_PREFER_SIBLING;
6333 sd->cache_nice_tries = 1;
6338 sd->private = &tl->data;
6344 * Topology list, bottom-up.
6346 static struct sched_domain_topology_level default_topology[] = {
6347 #ifdef CONFIG_SCHED_SMT
6348 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6350 #ifdef CONFIG_SCHED_MC
6351 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6353 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6357 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6359 #define for_each_sd_topology(tl) \
6360 for (tl = sched_domain_topology; tl->mask; tl++)
6362 void set_sched_topology(struct sched_domain_topology_level *tl)
6364 sched_domain_topology = tl;
6369 static const struct cpumask *sd_numa_mask(int cpu)
6371 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6374 static void sched_numa_warn(const char *str)
6376 static int done = false;
6384 printk(KERN_WARNING "ERROR: %s\n\n", str);
6386 for (i = 0; i < nr_node_ids; i++) {
6387 printk(KERN_WARNING " ");
6388 for (j = 0; j < nr_node_ids; j++)
6389 printk(KERN_CONT "%02d ", node_distance(i,j));
6390 printk(KERN_CONT "\n");
6392 printk(KERN_WARNING "\n");
6395 bool find_numa_distance(int distance)
6399 if (distance == node_distance(0, 0))
6402 for (i = 0; i < sched_domains_numa_levels; i++) {
6403 if (sched_domains_numa_distance[i] == distance)
6411 * A system can have three types of NUMA topology:
6412 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6413 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6414 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6416 * The difference between a glueless mesh topology and a backplane
6417 * topology lies in whether communication between not directly
6418 * connected nodes goes through intermediary nodes (where programs
6419 * could run), or through backplane controllers. This affects
6420 * placement of programs.
6422 * The type of topology can be discerned with the following tests:
6423 * - If the maximum distance between any nodes is 1 hop, the system
6424 * is directly connected.
6425 * - If for two nodes A and B, located N > 1 hops away from each other,
6426 * there is an intermediary node C, which is < N hops away from both
6427 * nodes A and B, the system is a glueless mesh.
6429 static void init_numa_topology_type(void)
6433 n = sched_max_numa_distance;
6436 sched_numa_topology_type = NUMA_DIRECT;
6438 for_each_online_node(a) {
6439 for_each_online_node(b) {
6440 /* Find two nodes furthest removed from each other. */
6441 if (node_distance(a, b) < n)
6444 /* Is there an intermediary node between a and b? */
6445 for_each_online_node(c) {
6446 if (node_distance(a, c) < n &&
6447 node_distance(b, c) < n) {
6448 sched_numa_topology_type =
6454 sched_numa_topology_type = NUMA_BACKPLANE;
6460 static void sched_init_numa(void)
6462 int next_distance, curr_distance = node_distance(0, 0);
6463 struct sched_domain_topology_level *tl;
6467 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6468 if (!sched_domains_numa_distance)
6472 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6473 * unique distances in the node_distance() table.
6475 * Assumes node_distance(0,j) includes all distances in
6476 * node_distance(i,j) in order to avoid cubic time.
6478 next_distance = curr_distance;
6479 for (i = 0; i < nr_node_ids; i++) {
6480 for (j = 0; j < nr_node_ids; j++) {
6481 for (k = 0; k < nr_node_ids; k++) {
6482 int distance = node_distance(i, k);
6484 if (distance > curr_distance &&
6485 (distance < next_distance ||
6486 next_distance == curr_distance))
6487 next_distance = distance;
6490 * While not a strong assumption it would be nice to know
6491 * about cases where if node A is connected to B, B is not
6492 * equally connected to A.
6494 if (sched_debug() && node_distance(k, i) != distance)
6495 sched_numa_warn("Node-distance not symmetric");
6497 if (sched_debug() && i && !find_numa_distance(distance))
6498 sched_numa_warn("Node-0 not representative");
6500 if (next_distance != curr_distance) {
6501 sched_domains_numa_distance[level++] = next_distance;
6502 sched_domains_numa_levels = level;
6503 curr_distance = next_distance;
6508 * In case of sched_debug() we verify the above assumption.
6518 * 'level' contains the number of unique distances, excluding the
6519 * identity distance node_distance(i,i).
6521 * The sched_domains_numa_distance[] array includes the actual distance
6526 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6527 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6528 * the array will contain less then 'level' members. This could be
6529 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6530 * in other functions.
6532 * We reset it to 'level' at the end of this function.
6534 sched_domains_numa_levels = 0;
6536 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6537 if (!sched_domains_numa_masks)
6541 * Now for each level, construct a mask per node which contains all
6542 * cpus of nodes that are that many hops away from us.
6544 for (i = 0; i < level; i++) {
6545 sched_domains_numa_masks[i] =
6546 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6547 if (!sched_domains_numa_masks[i])
6550 for (j = 0; j < nr_node_ids; j++) {
6551 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6555 sched_domains_numa_masks[i][j] = mask;
6557 for (k = 0; k < nr_node_ids; k++) {
6558 if (node_distance(j, k) > sched_domains_numa_distance[i])
6561 cpumask_or(mask, mask, cpumask_of_node(k));
6566 /* Compute default topology size */
6567 for (i = 0; sched_domain_topology[i].mask; i++);
6569 tl = kzalloc((i + level + 1) *
6570 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6575 * Copy the default topology bits..
6577 for (i = 0; sched_domain_topology[i].mask; i++)
6578 tl[i] = sched_domain_topology[i];
6581 * .. and append 'j' levels of NUMA goodness.
6583 for (j = 0; j < level; i++, j++) {
6584 tl[i] = (struct sched_domain_topology_level){
6585 .mask = sd_numa_mask,
6586 .sd_flags = cpu_numa_flags,
6587 .flags = SDTL_OVERLAP,
6593 sched_domain_topology = tl;
6595 sched_domains_numa_levels = level;
6596 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6598 init_numa_topology_type();
6601 static void sched_domains_numa_masks_set(int cpu)
6604 int node = cpu_to_node(cpu);
6606 for (i = 0; i < sched_domains_numa_levels; i++) {
6607 for (j = 0; j < nr_node_ids; j++) {
6608 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6609 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6614 static void sched_domains_numa_masks_clear(int cpu)
6617 for (i = 0; i < sched_domains_numa_levels; i++) {
6618 for (j = 0; j < nr_node_ids; j++)
6619 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6624 * Update sched_domains_numa_masks[level][node] array when new cpus
6627 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6628 unsigned long action,
6631 int cpu = (long)hcpu;
6633 switch (action & ~CPU_TASKS_FROZEN) {
6635 sched_domains_numa_masks_set(cpu);
6639 sched_domains_numa_masks_clear(cpu);
6649 static inline void sched_init_numa(void)
6653 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6654 unsigned long action,
6659 #endif /* CONFIG_NUMA */
6661 static int __sdt_alloc(const struct cpumask *cpu_map)
6663 struct sched_domain_topology_level *tl;
6666 for_each_sd_topology(tl) {
6667 struct sd_data *sdd = &tl->data;
6669 sdd->sd = alloc_percpu(struct sched_domain *);
6673 sdd->sg = alloc_percpu(struct sched_group *);
6677 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6681 for_each_cpu(j, cpu_map) {
6682 struct sched_domain *sd;
6683 struct sched_group *sg;
6684 struct sched_group_capacity *sgc;
6686 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6687 GFP_KERNEL, cpu_to_node(j));
6691 *per_cpu_ptr(sdd->sd, j) = sd;
6693 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6694 GFP_KERNEL, cpu_to_node(j));
6700 *per_cpu_ptr(sdd->sg, j) = sg;
6702 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6703 GFP_KERNEL, cpu_to_node(j));
6707 *per_cpu_ptr(sdd->sgc, j) = sgc;
6714 static void __sdt_free(const struct cpumask *cpu_map)
6716 struct sched_domain_topology_level *tl;
6719 for_each_sd_topology(tl) {
6720 struct sd_data *sdd = &tl->data;
6722 for_each_cpu(j, cpu_map) {
6723 struct sched_domain *sd;
6726 sd = *per_cpu_ptr(sdd->sd, j);
6727 if (sd && (sd->flags & SD_OVERLAP))
6728 free_sched_groups(sd->groups, 0);
6729 kfree(*per_cpu_ptr(sdd->sd, j));
6733 kfree(*per_cpu_ptr(sdd->sg, j));
6735 kfree(*per_cpu_ptr(sdd->sgc, j));
6737 free_percpu(sdd->sd);
6739 free_percpu(sdd->sg);
6741 free_percpu(sdd->sgc);
6746 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6747 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6748 struct sched_domain *child, int cpu)
6750 struct sched_domain *sd = sd_init(tl, cpu);
6754 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6756 sd->level = child->level + 1;
6757 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6761 if (!cpumask_subset(sched_domain_span(child),
6762 sched_domain_span(sd))) {
6763 pr_err("BUG: arch topology borken\n");
6764 #ifdef CONFIG_SCHED_DEBUG
6765 pr_err(" the %s domain not a subset of the %s domain\n",
6766 child->name, sd->name);
6768 /* Fixup, ensure @sd has at least @child cpus. */
6769 cpumask_or(sched_domain_span(sd),
6770 sched_domain_span(sd),
6771 sched_domain_span(child));
6775 set_domain_attribute(sd, attr);
6781 * Build sched domains for a given set of cpus and attach the sched domains
6782 * to the individual cpus
6784 static int build_sched_domains(const struct cpumask *cpu_map,
6785 struct sched_domain_attr *attr)
6787 enum s_alloc alloc_state;
6788 struct sched_domain *sd;
6790 int i, ret = -ENOMEM;
6792 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6793 if (alloc_state != sa_rootdomain)
6796 /* Set up domains for cpus specified by the cpu_map. */
6797 for_each_cpu(i, cpu_map) {
6798 struct sched_domain_topology_level *tl;
6801 for_each_sd_topology(tl) {
6802 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6803 if (tl == sched_domain_topology)
6804 *per_cpu_ptr(d.sd, i) = sd;
6805 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6806 sd->flags |= SD_OVERLAP;
6807 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6812 /* Build the groups for the domains */
6813 for_each_cpu(i, cpu_map) {
6814 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6815 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6816 if (sd->flags & SD_OVERLAP) {
6817 if (build_overlap_sched_groups(sd, i))
6820 if (build_sched_groups(sd, i))
6826 /* Calculate CPU capacity for physical packages and nodes */
6827 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6828 if (!cpumask_test_cpu(i, cpu_map))
6831 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6832 claim_allocations(i, sd);
6833 init_sched_groups_capacity(i, sd);
6837 /* Attach the domains */
6839 for_each_cpu(i, cpu_map) {
6840 sd = *per_cpu_ptr(d.sd, i);
6841 cpu_attach_domain(sd, d.rd, i);
6847 __free_domain_allocs(&d, alloc_state, cpu_map);
6851 static cpumask_var_t *doms_cur; /* current sched domains */
6852 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6853 static struct sched_domain_attr *dattr_cur;
6854 /* attribues of custom domains in 'doms_cur' */
6857 * Special case: If a kmalloc of a doms_cur partition (array of
6858 * cpumask) fails, then fallback to a single sched domain,
6859 * as determined by the single cpumask fallback_doms.
6861 static cpumask_var_t fallback_doms;
6864 * arch_update_cpu_topology lets virtualized architectures update the
6865 * cpu core maps. It is supposed to return 1 if the topology changed
6866 * or 0 if it stayed the same.
6868 int __weak arch_update_cpu_topology(void)
6873 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6876 cpumask_var_t *doms;
6878 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6881 for (i = 0; i < ndoms; i++) {
6882 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6883 free_sched_domains(doms, i);
6890 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6893 for (i = 0; i < ndoms; i++)
6894 free_cpumask_var(doms[i]);
6899 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6900 * For now this just excludes isolated cpus, but could be used to
6901 * exclude other special cases in the future.
6903 static int init_sched_domains(const struct cpumask *cpu_map)
6907 arch_update_cpu_topology();
6909 doms_cur = alloc_sched_domains(ndoms_cur);
6911 doms_cur = &fallback_doms;
6912 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6913 err = build_sched_domains(doms_cur[0], NULL);
6914 register_sched_domain_sysctl();
6920 * Detach sched domains from a group of cpus specified in cpu_map
6921 * These cpus will now be attached to the NULL domain
6923 static void detach_destroy_domains(const struct cpumask *cpu_map)
6928 for_each_cpu(i, cpu_map)
6929 cpu_attach_domain(NULL, &def_root_domain, i);
6933 /* handle null as "default" */
6934 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6935 struct sched_domain_attr *new, int idx_new)
6937 struct sched_domain_attr tmp;
6944 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6945 new ? (new + idx_new) : &tmp,
6946 sizeof(struct sched_domain_attr));
6950 * Partition sched domains as specified by the 'ndoms_new'
6951 * cpumasks in the array doms_new[] of cpumasks. This compares
6952 * doms_new[] to the current sched domain partitioning, doms_cur[].
6953 * It destroys each deleted domain and builds each new domain.
6955 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6956 * The masks don't intersect (don't overlap.) We should setup one
6957 * sched domain for each mask. CPUs not in any of the cpumasks will
6958 * not be load balanced. If the same cpumask appears both in the
6959 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6962 * The passed in 'doms_new' should be allocated using
6963 * alloc_sched_domains. This routine takes ownership of it and will
6964 * free_sched_domains it when done with it. If the caller failed the
6965 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6966 * and partition_sched_domains() will fallback to the single partition
6967 * 'fallback_doms', it also forces the domains to be rebuilt.
6969 * If doms_new == NULL it will be replaced with cpu_online_mask.
6970 * ndoms_new == 0 is a special case for destroying existing domains,
6971 * and it will not create the default domain.
6973 * Call with hotplug lock held
6975 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6976 struct sched_domain_attr *dattr_new)
6981 mutex_lock(&sched_domains_mutex);
6983 /* always unregister in case we don't destroy any domains */
6984 unregister_sched_domain_sysctl();
6986 /* Let architecture update cpu core mappings. */
6987 new_topology = arch_update_cpu_topology();
6989 n = doms_new ? ndoms_new : 0;
6991 /* Destroy deleted domains */
6992 for (i = 0; i < ndoms_cur; i++) {
6993 for (j = 0; j < n && !new_topology; j++) {
6994 if (cpumask_equal(doms_cur[i], doms_new[j])
6995 && dattrs_equal(dattr_cur, i, dattr_new, j))
6998 /* no match - a current sched domain not in new doms_new[] */
6999 detach_destroy_domains(doms_cur[i]);
7005 if (doms_new == NULL) {
7007 doms_new = &fallback_doms;
7008 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7009 WARN_ON_ONCE(dattr_new);
7012 /* Build new domains */
7013 for (i = 0; i < ndoms_new; i++) {
7014 for (j = 0; j < n && !new_topology; j++) {
7015 if (cpumask_equal(doms_new[i], doms_cur[j])
7016 && dattrs_equal(dattr_new, i, dattr_cur, j))
7019 /* no match - add a new doms_new */
7020 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7025 /* Remember the new sched domains */
7026 if (doms_cur != &fallback_doms)
7027 free_sched_domains(doms_cur, ndoms_cur);
7028 kfree(dattr_cur); /* kfree(NULL) is safe */
7029 doms_cur = doms_new;
7030 dattr_cur = dattr_new;
7031 ndoms_cur = ndoms_new;
7033 register_sched_domain_sysctl();
7035 mutex_unlock(&sched_domains_mutex);
7038 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7041 * Update cpusets according to cpu_active mask. If cpusets are
7042 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7043 * around partition_sched_domains().
7045 * If we come here as part of a suspend/resume, don't touch cpusets because we
7046 * want to restore it back to its original state upon resume anyway.
7048 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7052 case CPU_ONLINE_FROZEN:
7053 case CPU_DOWN_FAILED_FROZEN:
7056 * num_cpus_frozen tracks how many CPUs are involved in suspend
7057 * resume sequence. As long as this is not the last online
7058 * operation in the resume sequence, just build a single sched
7059 * domain, ignoring cpusets.
7062 if (likely(num_cpus_frozen)) {
7063 partition_sched_domains(1, NULL, NULL);
7068 * This is the last CPU online operation. So fall through and
7069 * restore the original sched domains by considering the
7070 * cpuset configurations.
7074 cpuset_update_active_cpus(true);
7082 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7085 unsigned long flags;
7086 long cpu = (long)hcpu;
7092 case CPU_DOWN_PREPARE:
7093 rcu_read_lock_sched();
7094 dl_b = dl_bw_of(cpu);
7096 raw_spin_lock_irqsave(&dl_b->lock, flags);
7097 cpus = dl_bw_cpus(cpu);
7098 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7099 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7101 rcu_read_unlock_sched();
7104 return notifier_from_errno(-EBUSY);
7105 cpuset_update_active_cpus(false);
7107 case CPU_DOWN_PREPARE_FROZEN:
7109 partition_sched_domains(1, NULL, NULL);
7117 void __init sched_init_smp(void)
7119 cpumask_var_t non_isolated_cpus;
7121 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7122 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7124 /* nohz_full won't take effect without isolating the cpus. */
7125 tick_nohz_full_add_cpus_to(cpu_isolated_map);
7130 * There's no userspace yet to cause hotplug operations; hence all the
7131 * cpu masks are stable and all blatant races in the below code cannot
7134 mutex_lock(&sched_domains_mutex);
7135 init_sched_domains(cpu_active_mask);
7136 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7137 if (cpumask_empty(non_isolated_cpus))
7138 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7139 mutex_unlock(&sched_domains_mutex);
7141 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7142 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7143 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7147 /* Move init over to a non-isolated CPU */
7148 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7150 sched_init_granularity();
7151 free_cpumask_var(non_isolated_cpus);
7153 init_sched_rt_class();
7154 init_sched_dl_class();
7157 void __init sched_init_smp(void)
7159 sched_init_granularity();
7161 #endif /* CONFIG_SMP */
7163 int in_sched_functions(unsigned long addr)
7165 return in_lock_functions(addr) ||
7166 (addr >= (unsigned long)__sched_text_start
7167 && addr < (unsigned long)__sched_text_end);
7170 #ifdef CONFIG_CGROUP_SCHED
7172 * Default task group.
7173 * Every task in system belongs to this group at bootup.
7175 struct task_group root_task_group;
7176 LIST_HEAD(task_groups);
7179 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7181 void __init sched_init(void)
7184 unsigned long alloc_size = 0, ptr;
7186 #ifdef CONFIG_FAIR_GROUP_SCHED
7187 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7189 #ifdef CONFIG_RT_GROUP_SCHED
7190 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7193 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7195 #ifdef CONFIG_FAIR_GROUP_SCHED
7196 root_task_group.se = (struct sched_entity **)ptr;
7197 ptr += nr_cpu_ids * sizeof(void **);
7199 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7200 ptr += nr_cpu_ids * sizeof(void **);
7202 #endif /* CONFIG_FAIR_GROUP_SCHED */
7203 #ifdef CONFIG_RT_GROUP_SCHED
7204 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7205 ptr += nr_cpu_ids * sizeof(void **);
7207 root_task_group.rt_rq = (struct rt_rq **)ptr;
7208 ptr += nr_cpu_ids * sizeof(void **);
7210 #endif /* CONFIG_RT_GROUP_SCHED */
7212 #ifdef CONFIG_CPUMASK_OFFSTACK
7213 for_each_possible_cpu(i) {
7214 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7215 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7217 #endif /* CONFIG_CPUMASK_OFFSTACK */
7219 init_rt_bandwidth(&def_rt_bandwidth,
7220 global_rt_period(), global_rt_runtime());
7221 init_dl_bandwidth(&def_dl_bandwidth,
7222 global_rt_period(), global_rt_runtime());
7225 init_defrootdomain();
7228 #ifdef CONFIG_RT_GROUP_SCHED
7229 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7230 global_rt_period(), global_rt_runtime());
7231 #endif /* CONFIG_RT_GROUP_SCHED */
7233 #ifdef CONFIG_CGROUP_SCHED
7234 list_add(&root_task_group.list, &task_groups);
7235 INIT_LIST_HEAD(&root_task_group.children);
7236 INIT_LIST_HEAD(&root_task_group.siblings);
7237 autogroup_init(&init_task);
7239 #endif /* CONFIG_CGROUP_SCHED */
7241 for_each_possible_cpu(i) {
7245 raw_spin_lock_init(&rq->lock);
7247 rq->calc_load_active = 0;
7248 rq->calc_load_update = jiffies + LOAD_FREQ;
7249 init_cfs_rq(&rq->cfs);
7250 init_rt_rq(&rq->rt);
7251 init_dl_rq(&rq->dl);
7252 #ifdef CONFIG_FAIR_GROUP_SCHED
7253 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7254 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7256 * How much cpu bandwidth does root_task_group get?
7258 * In case of task-groups formed thr' the cgroup filesystem, it
7259 * gets 100% of the cpu resources in the system. This overall
7260 * system cpu resource is divided among the tasks of
7261 * root_task_group and its child task-groups in a fair manner,
7262 * based on each entity's (task or task-group's) weight
7263 * (se->load.weight).
7265 * In other words, if root_task_group has 10 tasks of weight
7266 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7267 * then A0's share of the cpu resource is:
7269 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7271 * We achieve this by letting root_task_group's tasks sit
7272 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7274 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7275 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7276 #endif /* CONFIG_FAIR_GROUP_SCHED */
7278 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7279 #ifdef CONFIG_RT_GROUP_SCHED
7280 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7283 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7284 rq->cpu_load[j] = 0;
7286 rq->last_load_update_tick = jiffies;
7291 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7292 rq->balance_callback = NULL;
7293 rq->active_balance = 0;
7294 rq->next_balance = jiffies;
7299 rq->avg_idle = 2*sysctl_sched_migration_cost;
7300 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7302 INIT_LIST_HEAD(&rq->cfs_tasks);
7304 rq_attach_root(rq, &def_root_domain);
7305 #ifdef CONFIG_NO_HZ_COMMON
7308 #ifdef CONFIG_NO_HZ_FULL
7309 rq->last_sched_tick = 0;
7313 atomic_set(&rq->nr_iowait, 0);
7316 set_load_weight(&init_task);
7318 #ifdef CONFIG_PREEMPT_NOTIFIERS
7319 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7323 * The boot idle thread does lazy MMU switching as well:
7325 atomic_inc(&init_mm.mm_count);
7326 enter_lazy_tlb(&init_mm, current);
7329 * During early bootup we pretend to be a normal task:
7331 current->sched_class = &fair_sched_class;
7334 * Make us the idle thread. Technically, schedule() should not be
7335 * called from this thread, however somewhere below it might be,
7336 * but because we are the idle thread, we just pick up running again
7337 * when this runqueue becomes "idle".
7339 init_idle(current, smp_processor_id());
7341 calc_load_update = jiffies + LOAD_FREQ;
7344 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7345 /* May be allocated at isolcpus cmdline parse time */
7346 if (cpu_isolated_map == NULL)
7347 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7348 idle_thread_set_boot_cpu();
7349 set_cpu_rq_start_time();
7351 init_sched_fair_class();
7353 scheduler_running = 1;
7356 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7357 static inline int preempt_count_equals(int preempt_offset)
7359 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7361 return (nested == preempt_offset);
7364 void __might_sleep(const char *file, int line, int preempt_offset)
7367 * Blocking primitives will set (and therefore destroy) current->state,
7368 * since we will exit with TASK_RUNNING make sure we enter with it,
7369 * otherwise we will destroy state.
7371 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7372 "do not call blocking ops when !TASK_RUNNING; "
7373 "state=%lx set at [<%p>] %pS\n",
7375 (void *)current->task_state_change,
7376 (void *)current->task_state_change);
7378 ___might_sleep(file, line, preempt_offset);
7380 EXPORT_SYMBOL(__might_sleep);
7382 void ___might_sleep(const char *file, int line, int preempt_offset)
7384 static unsigned long prev_jiffy; /* ratelimiting */
7386 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7387 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7388 !is_idle_task(current)) ||
7389 system_state != SYSTEM_RUNNING || oops_in_progress)
7391 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7393 prev_jiffy = jiffies;
7396 "BUG: sleeping function called from invalid context at %s:%d\n",
7399 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7400 in_atomic(), irqs_disabled(),
7401 current->pid, current->comm);
7403 if (task_stack_end_corrupted(current))
7404 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7406 debug_show_held_locks(current);
7407 if (irqs_disabled())
7408 print_irqtrace_events(current);
7409 #ifdef CONFIG_DEBUG_PREEMPT
7410 if (!preempt_count_equals(preempt_offset)) {
7411 pr_err("Preemption disabled at:");
7412 print_ip_sym(current->preempt_disable_ip);
7418 EXPORT_SYMBOL(___might_sleep);
7421 #ifdef CONFIG_MAGIC_SYSRQ
7422 void normalize_rt_tasks(void)
7424 struct task_struct *g, *p;
7425 struct sched_attr attr = {
7426 .sched_policy = SCHED_NORMAL,
7429 read_lock(&tasklist_lock);
7430 for_each_process_thread(g, p) {
7432 * Only normalize user tasks:
7434 if (p->flags & PF_KTHREAD)
7437 p->se.exec_start = 0;
7438 #ifdef CONFIG_SCHEDSTATS
7439 p->se.statistics.wait_start = 0;
7440 p->se.statistics.sleep_start = 0;
7441 p->se.statistics.block_start = 0;
7444 if (!dl_task(p) && !rt_task(p)) {
7446 * Renice negative nice level userspace
7449 if (task_nice(p) < 0)
7450 set_user_nice(p, 0);
7454 __sched_setscheduler(p, &attr, false, false);
7456 read_unlock(&tasklist_lock);
7459 #endif /* CONFIG_MAGIC_SYSRQ */
7461 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7463 * These functions are only useful for the IA64 MCA handling, or kdb.
7465 * They can only be called when the whole system has been
7466 * stopped - every CPU needs to be quiescent, and no scheduling
7467 * activity can take place. Using them for anything else would
7468 * be a serious bug, and as a result, they aren't even visible
7469 * under any other configuration.
7473 * curr_task - return the current task for a given cpu.
7474 * @cpu: the processor in question.
7476 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7478 * Return: The current task for @cpu.
7480 struct task_struct *curr_task(int cpu)
7482 return cpu_curr(cpu);
7485 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7489 * set_curr_task - set the current task for a given cpu.
7490 * @cpu: the processor in question.
7491 * @p: the task pointer to set.
7493 * Description: This function must only be used when non-maskable interrupts
7494 * are serviced on a separate stack. It allows the architecture to switch the
7495 * notion of the current task on a cpu in a non-blocking manner. This function
7496 * must be called with all CPU's synchronized, and interrupts disabled, the
7497 * and caller must save the original value of the current task (see
7498 * curr_task() above) and restore that value before reenabling interrupts and
7499 * re-starting the system.
7501 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7503 void set_curr_task(int cpu, struct task_struct *p)
7510 #ifdef CONFIG_CGROUP_SCHED
7511 /* task_group_lock serializes the addition/removal of task groups */
7512 static DEFINE_SPINLOCK(task_group_lock);
7514 static void free_sched_group(struct task_group *tg)
7516 free_fair_sched_group(tg);
7517 free_rt_sched_group(tg);
7522 /* allocate runqueue etc for a new task group */
7523 struct task_group *sched_create_group(struct task_group *parent)
7525 struct task_group *tg;
7527 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7529 return ERR_PTR(-ENOMEM);
7531 if (!alloc_fair_sched_group(tg, parent))
7534 if (!alloc_rt_sched_group(tg, parent))
7540 free_sched_group(tg);
7541 return ERR_PTR(-ENOMEM);
7544 void sched_online_group(struct task_group *tg, struct task_group *parent)
7546 unsigned long flags;
7548 spin_lock_irqsave(&task_group_lock, flags);
7549 list_add_rcu(&tg->list, &task_groups);
7551 WARN_ON(!parent); /* root should already exist */
7553 tg->parent = parent;
7554 INIT_LIST_HEAD(&tg->children);
7555 list_add_rcu(&tg->siblings, &parent->children);
7556 spin_unlock_irqrestore(&task_group_lock, flags);
7559 /* rcu callback to free various structures associated with a task group */
7560 static void free_sched_group_rcu(struct rcu_head *rhp)
7562 /* now it should be safe to free those cfs_rqs */
7563 free_sched_group(container_of(rhp, struct task_group, rcu));
7566 /* Destroy runqueue etc associated with a task group */
7567 void sched_destroy_group(struct task_group *tg)
7569 /* wait for possible concurrent references to cfs_rqs complete */
7570 call_rcu(&tg->rcu, free_sched_group_rcu);
7573 void sched_offline_group(struct task_group *tg)
7575 unsigned long flags;
7578 /* end participation in shares distribution */
7579 for_each_possible_cpu(i)
7580 unregister_fair_sched_group(tg, i);
7582 spin_lock_irqsave(&task_group_lock, flags);
7583 list_del_rcu(&tg->list);
7584 list_del_rcu(&tg->siblings);
7585 spin_unlock_irqrestore(&task_group_lock, flags);
7588 /* change task's runqueue when it moves between groups.
7589 * The caller of this function should have put the task in its new group
7590 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7591 * reflect its new group.
7593 void sched_move_task(struct task_struct *tsk)
7595 struct task_group *tg;
7596 int queued, running;
7597 unsigned long flags;
7600 rq = task_rq_lock(tsk, &flags);
7602 running = task_current(rq, tsk);
7603 queued = task_on_rq_queued(tsk);
7606 dequeue_task(rq, tsk, 0);
7607 if (unlikely(running))
7608 put_prev_task(rq, tsk);
7611 * All callers are synchronized by task_rq_lock(); we do not use RCU
7612 * which is pointless here. Thus, we pass "true" to task_css_check()
7613 * to prevent lockdep warnings.
7615 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7616 struct task_group, css);
7617 tg = autogroup_task_group(tsk, tg);
7618 tsk->sched_task_group = tg;
7620 #ifdef CONFIG_FAIR_GROUP_SCHED
7621 if (tsk->sched_class->task_move_group)
7622 tsk->sched_class->task_move_group(tsk, queued);
7625 set_task_rq(tsk, task_cpu(tsk));
7627 if (unlikely(running))
7628 tsk->sched_class->set_curr_task(rq);
7630 enqueue_task(rq, tsk, 0);
7632 task_rq_unlock(rq, tsk, &flags);
7634 #endif /* CONFIG_CGROUP_SCHED */
7636 #ifdef CONFIG_RT_GROUP_SCHED
7638 * Ensure that the real time constraints are schedulable.
7640 static DEFINE_MUTEX(rt_constraints_mutex);
7642 /* Must be called with tasklist_lock held */
7643 static inline int tg_has_rt_tasks(struct task_group *tg)
7645 struct task_struct *g, *p;
7648 * Autogroups do not have RT tasks; see autogroup_create().
7650 if (task_group_is_autogroup(tg))
7653 for_each_process_thread(g, p) {
7654 if (rt_task(p) && task_group(p) == tg)
7661 struct rt_schedulable_data {
7662 struct task_group *tg;
7667 static int tg_rt_schedulable(struct task_group *tg, void *data)
7669 struct rt_schedulable_data *d = data;
7670 struct task_group *child;
7671 unsigned long total, sum = 0;
7672 u64 period, runtime;
7674 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7675 runtime = tg->rt_bandwidth.rt_runtime;
7678 period = d->rt_period;
7679 runtime = d->rt_runtime;
7683 * Cannot have more runtime than the period.
7685 if (runtime > period && runtime != RUNTIME_INF)
7689 * Ensure we don't starve existing RT tasks.
7691 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7694 total = to_ratio(period, runtime);
7697 * Nobody can have more than the global setting allows.
7699 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7703 * The sum of our children's runtime should not exceed our own.
7705 list_for_each_entry_rcu(child, &tg->children, siblings) {
7706 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7707 runtime = child->rt_bandwidth.rt_runtime;
7709 if (child == d->tg) {
7710 period = d->rt_period;
7711 runtime = d->rt_runtime;
7714 sum += to_ratio(period, runtime);
7723 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7727 struct rt_schedulable_data data = {
7729 .rt_period = period,
7730 .rt_runtime = runtime,
7734 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7740 static int tg_set_rt_bandwidth(struct task_group *tg,
7741 u64 rt_period, u64 rt_runtime)
7746 * Disallowing the root group RT runtime is BAD, it would disallow the
7747 * kernel creating (and or operating) RT threads.
7749 if (tg == &root_task_group && rt_runtime == 0)
7752 /* No period doesn't make any sense. */
7756 mutex_lock(&rt_constraints_mutex);
7757 read_lock(&tasklist_lock);
7758 err = __rt_schedulable(tg, rt_period, rt_runtime);
7762 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7763 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7764 tg->rt_bandwidth.rt_runtime = rt_runtime;
7766 for_each_possible_cpu(i) {
7767 struct rt_rq *rt_rq = tg->rt_rq[i];
7769 raw_spin_lock(&rt_rq->rt_runtime_lock);
7770 rt_rq->rt_runtime = rt_runtime;
7771 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7773 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7775 read_unlock(&tasklist_lock);
7776 mutex_unlock(&rt_constraints_mutex);
7781 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7783 u64 rt_runtime, rt_period;
7785 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7786 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7787 if (rt_runtime_us < 0)
7788 rt_runtime = RUNTIME_INF;
7790 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7793 static long sched_group_rt_runtime(struct task_group *tg)
7797 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7800 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7801 do_div(rt_runtime_us, NSEC_PER_USEC);
7802 return rt_runtime_us;
7805 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7807 u64 rt_runtime, rt_period;
7809 rt_period = rt_period_us * NSEC_PER_USEC;
7810 rt_runtime = tg->rt_bandwidth.rt_runtime;
7812 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7815 static long sched_group_rt_period(struct task_group *tg)
7819 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7820 do_div(rt_period_us, NSEC_PER_USEC);
7821 return rt_period_us;
7823 #endif /* CONFIG_RT_GROUP_SCHED */
7825 #ifdef CONFIG_RT_GROUP_SCHED
7826 static int sched_rt_global_constraints(void)
7830 mutex_lock(&rt_constraints_mutex);
7831 read_lock(&tasklist_lock);
7832 ret = __rt_schedulable(NULL, 0, 0);
7833 read_unlock(&tasklist_lock);
7834 mutex_unlock(&rt_constraints_mutex);
7839 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7841 /* Don't accept realtime tasks when there is no way for them to run */
7842 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7848 #else /* !CONFIG_RT_GROUP_SCHED */
7849 static int sched_rt_global_constraints(void)
7851 unsigned long flags;
7854 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7855 for_each_possible_cpu(i) {
7856 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7858 raw_spin_lock(&rt_rq->rt_runtime_lock);
7859 rt_rq->rt_runtime = global_rt_runtime();
7860 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7862 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7866 #endif /* CONFIG_RT_GROUP_SCHED */
7868 static int sched_dl_global_validate(void)
7870 u64 runtime = global_rt_runtime();
7871 u64 period = global_rt_period();
7872 u64 new_bw = to_ratio(period, runtime);
7875 unsigned long flags;
7878 * Here we want to check the bandwidth not being set to some
7879 * value smaller than the currently allocated bandwidth in
7880 * any of the root_domains.
7882 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7883 * cycling on root_domains... Discussion on different/better
7884 * solutions is welcome!
7886 for_each_possible_cpu(cpu) {
7887 rcu_read_lock_sched();
7888 dl_b = dl_bw_of(cpu);
7890 raw_spin_lock_irqsave(&dl_b->lock, flags);
7891 if (new_bw < dl_b->total_bw)
7893 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7895 rcu_read_unlock_sched();
7904 static void sched_dl_do_global(void)
7909 unsigned long flags;
7911 def_dl_bandwidth.dl_period = global_rt_period();
7912 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7914 if (global_rt_runtime() != RUNTIME_INF)
7915 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7918 * FIXME: As above...
7920 for_each_possible_cpu(cpu) {
7921 rcu_read_lock_sched();
7922 dl_b = dl_bw_of(cpu);
7924 raw_spin_lock_irqsave(&dl_b->lock, flags);
7926 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7928 rcu_read_unlock_sched();
7932 static int sched_rt_global_validate(void)
7934 if (sysctl_sched_rt_period <= 0)
7937 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7938 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7944 static void sched_rt_do_global(void)
7946 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7947 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7950 int sched_rt_handler(struct ctl_table *table, int write,
7951 void __user *buffer, size_t *lenp,
7954 int old_period, old_runtime;
7955 static DEFINE_MUTEX(mutex);
7959 old_period = sysctl_sched_rt_period;
7960 old_runtime = sysctl_sched_rt_runtime;
7962 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7964 if (!ret && write) {
7965 ret = sched_rt_global_validate();
7969 ret = sched_dl_global_validate();
7973 ret = sched_rt_global_constraints();
7977 sched_rt_do_global();
7978 sched_dl_do_global();
7982 sysctl_sched_rt_period = old_period;
7983 sysctl_sched_rt_runtime = old_runtime;
7985 mutex_unlock(&mutex);
7990 int sched_rr_handler(struct ctl_table *table, int write,
7991 void __user *buffer, size_t *lenp,
7995 static DEFINE_MUTEX(mutex);
7998 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7999 /* make sure that internally we keep jiffies */
8000 /* also, writing zero resets timeslice to default */
8001 if (!ret && write) {
8002 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8003 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8005 mutex_unlock(&mutex);
8009 #ifdef CONFIG_CGROUP_SCHED
8011 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8013 return css ? container_of(css, struct task_group, css) : NULL;
8016 static struct cgroup_subsys_state *
8017 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8019 struct task_group *parent = css_tg(parent_css);
8020 struct task_group *tg;
8023 /* This is early initialization for the top cgroup */
8024 return &root_task_group.css;
8027 tg = sched_create_group(parent);
8029 return ERR_PTR(-ENOMEM);
8034 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8036 struct task_group *tg = css_tg(css);
8037 struct task_group *parent = css_tg(css->parent);
8040 sched_online_group(tg, parent);
8044 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8046 struct task_group *tg = css_tg(css);
8048 sched_destroy_group(tg);
8051 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8053 struct task_group *tg = css_tg(css);
8055 sched_offline_group(tg);
8058 static void cpu_cgroup_fork(struct task_struct *task)
8060 sched_move_task(task);
8063 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8064 struct cgroup_taskset *tset)
8066 struct task_struct *task;
8068 cgroup_taskset_for_each(task, tset) {
8069 #ifdef CONFIG_RT_GROUP_SCHED
8070 if (!sched_rt_can_attach(css_tg(css), task))
8073 /* We don't support RT-tasks being in separate groups */
8074 if (task->sched_class != &fair_sched_class)
8081 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8082 struct cgroup_taskset *tset)
8084 struct task_struct *task;
8086 cgroup_taskset_for_each(task, tset)
8087 sched_move_task(task);
8090 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8091 struct cgroup_subsys_state *old_css,
8092 struct task_struct *task)
8095 * cgroup_exit() is called in the copy_process() failure path.
8096 * Ignore this case since the task hasn't ran yet, this avoids
8097 * trying to poke a half freed task state from generic code.
8099 if (!(task->flags & PF_EXITING))
8102 sched_move_task(task);
8105 #ifdef CONFIG_FAIR_GROUP_SCHED
8106 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8107 struct cftype *cftype, u64 shareval)
8109 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8112 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8115 struct task_group *tg = css_tg(css);
8117 return (u64) scale_load_down(tg->shares);
8120 #ifdef CONFIG_CFS_BANDWIDTH
8121 static DEFINE_MUTEX(cfs_constraints_mutex);
8123 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8124 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8126 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8128 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8130 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8131 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8133 if (tg == &root_task_group)
8137 * Ensure we have at some amount of bandwidth every period. This is
8138 * to prevent reaching a state of large arrears when throttled via
8139 * entity_tick() resulting in prolonged exit starvation.
8141 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8145 * Likewise, bound things on the otherside by preventing insane quota
8146 * periods. This also allows us to normalize in computing quota
8149 if (period > max_cfs_quota_period)
8153 * Prevent race between setting of cfs_rq->runtime_enabled and
8154 * unthrottle_offline_cfs_rqs().
8157 mutex_lock(&cfs_constraints_mutex);
8158 ret = __cfs_schedulable(tg, period, quota);
8162 runtime_enabled = quota != RUNTIME_INF;
8163 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8165 * If we need to toggle cfs_bandwidth_used, off->on must occur
8166 * before making related changes, and on->off must occur afterwards
8168 if (runtime_enabled && !runtime_was_enabled)
8169 cfs_bandwidth_usage_inc();
8170 raw_spin_lock_irq(&cfs_b->lock);
8171 cfs_b->period = ns_to_ktime(period);
8172 cfs_b->quota = quota;
8174 __refill_cfs_bandwidth_runtime(cfs_b);
8175 /* restart the period timer (if active) to handle new period expiry */
8176 if (runtime_enabled)
8177 start_cfs_bandwidth(cfs_b);
8178 raw_spin_unlock_irq(&cfs_b->lock);
8180 for_each_online_cpu(i) {
8181 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8182 struct rq *rq = cfs_rq->rq;
8184 raw_spin_lock_irq(&rq->lock);
8185 cfs_rq->runtime_enabled = runtime_enabled;
8186 cfs_rq->runtime_remaining = 0;
8188 if (cfs_rq->throttled)
8189 unthrottle_cfs_rq(cfs_rq);
8190 raw_spin_unlock_irq(&rq->lock);
8192 if (runtime_was_enabled && !runtime_enabled)
8193 cfs_bandwidth_usage_dec();
8195 mutex_unlock(&cfs_constraints_mutex);
8201 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8205 period = ktime_to_ns(tg->cfs_bandwidth.period);
8206 if (cfs_quota_us < 0)
8207 quota = RUNTIME_INF;
8209 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8211 return tg_set_cfs_bandwidth(tg, period, quota);
8214 long tg_get_cfs_quota(struct task_group *tg)
8218 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8221 quota_us = tg->cfs_bandwidth.quota;
8222 do_div(quota_us, NSEC_PER_USEC);
8227 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8231 period = (u64)cfs_period_us * NSEC_PER_USEC;
8232 quota = tg->cfs_bandwidth.quota;
8234 return tg_set_cfs_bandwidth(tg, period, quota);
8237 long tg_get_cfs_period(struct task_group *tg)
8241 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8242 do_div(cfs_period_us, NSEC_PER_USEC);
8244 return cfs_period_us;
8247 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8250 return tg_get_cfs_quota(css_tg(css));
8253 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8254 struct cftype *cftype, s64 cfs_quota_us)
8256 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8259 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8262 return tg_get_cfs_period(css_tg(css));
8265 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8266 struct cftype *cftype, u64 cfs_period_us)
8268 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8271 struct cfs_schedulable_data {
8272 struct task_group *tg;
8277 * normalize group quota/period to be quota/max_period
8278 * note: units are usecs
8280 static u64 normalize_cfs_quota(struct task_group *tg,
8281 struct cfs_schedulable_data *d)
8289 period = tg_get_cfs_period(tg);
8290 quota = tg_get_cfs_quota(tg);
8293 /* note: these should typically be equivalent */
8294 if (quota == RUNTIME_INF || quota == -1)
8297 return to_ratio(period, quota);
8300 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8302 struct cfs_schedulable_data *d = data;
8303 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8304 s64 quota = 0, parent_quota = -1;
8307 quota = RUNTIME_INF;
8309 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8311 quota = normalize_cfs_quota(tg, d);
8312 parent_quota = parent_b->hierarchical_quota;
8315 * ensure max(child_quota) <= parent_quota, inherit when no
8318 if (quota == RUNTIME_INF)
8319 quota = parent_quota;
8320 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8323 cfs_b->hierarchical_quota = quota;
8328 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8331 struct cfs_schedulable_data data = {
8337 if (quota != RUNTIME_INF) {
8338 do_div(data.period, NSEC_PER_USEC);
8339 do_div(data.quota, NSEC_PER_USEC);
8343 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8349 static int cpu_stats_show(struct seq_file *sf, void *v)
8351 struct task_group *tg = css_tg(seq_css(sf));
8352 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8354 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8355 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8356 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8360 #endif /* CONFIG_CFS_BANDWIDTH */
8361 #endif /* CONFIG_FAIR_GROUP_SCHED */
8363 #ifdef CONFIG_RT_GROUP_SCHED
8364 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8365 struct cftype *cft, s64 val)
8367 return sched_group_set_rt_runtime(css_tg(css), val);
8370 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8373 return sched_group_rt_runtime(css_tg(css));
8376 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8377 struct cftype *cftype, u64 rt_period_us)
8379 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8382 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8385 return sched_group_rt_period(css_tg(css));
8387 #endif /* CONFIG_RT_GROUP_SCHED */
8389 static struct cftype cpu_files[] = {
8390 #ifdef CONFIG_FAIR_GROUP_SCHED
8393 .read_u64 = cpu_shares_read_u64,
8394 .write_u64 = cpu_shares_write_u64,
8397 #ifdef CONFIG_CFS_BANDWIDTH
8399 .name = "cfs_quota_us",
8400 .read_s64 = cpu_cfs_quota_read_s64,
8401 .write_s64 = cpu_cfs_quota_write_s64,
8404 .name = "cfs_period_us",
8405 .read_u64 = cpu_cfs_period_read_u64,
8406 .write_u64 = cpu_cfs_period_write_u64,
8410 .seq_show = cpu_stats_show,
8413 #ifdef CONFIG_RT_GROUP_SCHED
8415 .name = "rt_runtime_us",
8416 .read_s64 = cpu_rt_runtime_read,
8417 .write_s64 = cpu_rt_runtime_write,
8420 .name = "rt_period_us",
8421 .read_u64 = cpu_rt_period_read_uint,
8422 .write_u64 = cpu_rt_period_write_uint,
8428 struct cgroup_subsys cpu_cgrp_subsys = {
8429 .css_alloc = cpu_cgroup_css_alloc,
8430 .css_free = cpu_cgroup_css_free,
8431 .css_online = cpu_cgroup_css_online,
8432 .css_offline = cpu_cgroup_css_offline,
8433 .fork = cpu_cgroup_fork,
8434 .can_attach = cpu_cgroup_can_attach,
8435 .attach = cpu_cgroup_attach,
8436 .exit = cpu_cgroup_exit,
8437 .legacy_cftypes = cpu_files,
8441 #endif /* CONFIG_CGROUP_SCHED */
8443 void dump_cpu_task(int cpu)
8445 pr_info("Task dump for CPU %d:\n", cpu);
8446 sched_show_task(cpu_curr(cpu));