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 static_key_disable(&sched_feat_keys[i]);
170 static void sched_feat_enable(int i)
172 static_key_enable(&sched_feat_keys[i]);
175 static void sched_feat_disable(int i) { };
176 static void sched_feat_enable(int i) { };
177 #endif /* HAVE_JUMP_LABEL */
179 static int sched_feat_set(char *cmp)
184 if (strncmp(cmp, "NO_", 3) == 0) {
189 for (i = 0; i < __SCHED_FEAT_NR; i++) {
190 if (strcmp(cmp, sched_feat_names[i]) == 0) {
192 sysctl_sched_features &= ~(1UL << i);
193 sched_feat_disable(i);
195 sysctl_sched_features |= (1UL << i);
196 sched_feat_enable(i);
206 sched_feat_write(struct file *filp, const char __user *ubuf,
207 size_t cnt, loff_t *ppos)
217 if (copy_from_user(&buf, ubuf, cnt))
223 /* Ensure the static_key remains in a consistent state */
224 inode = file_inode(filp);
225 mutex_lock(&inode->i_mutex);
226 i = sched_feat_set(cmp);
227 mutex_unlock(&inode->i_mutex);
228 if (i == __SCHED_FEAT_NR)
236 static int sched_feat_open(struct inode *inode, struct file *filp)
238 return single_open(filp, sched_feat_show, NULL);
241 static const struct file_operations sched_feat_fops = {
242 .open = sched_feat_open,
243 .write = sched_feat_write,
246 .release = single_release,
249 static __init int sched_init_debug(void)
251 debugfs_create_file("sched_features", 0644, NULL, NULL,
256 late_initcall(sched_init_debug);
257 #endif /* CONFIG_SCHED_DEBUG */
260 * Number of tasks to iterate in a single balance run.
261 * Limited because this is done with IRQs disabled.
263 const_debug unsigned int sysctl_sched_nr_migrate = 32;
266 * period over which we average the RT time consumption, measured
271 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
274 * period over which we measure -rt task cpu usage in us.
277 unsigned int sysctl_sched_rt_period = 1000000;
279 __read_mostly int scheduler_running;
282 * part of the period that we allow rt tasks to run in us.
285 int sysctl_sched_rt_runtime = 950000;
287 /* cpus with isolated domains */
288 cpumask_var_t cpu_isolated_map;
291 * this_rq_lock - lock this runqueue and disable interrupts.
293 static struct rq *this_rq_lock(void)
300 raw_spin_lock(&rq->lock);
305 #ifdef CONFIG_SCHED_HRTICK
307 * Use HR-timers to deliver accurate preemption points.
310 static void hrtick_clear(struct rq *rq)
312 if (hrtimer_active(&rq->hrtick_timer))
313 hrtimer_cancel(&rq->hrtick_timer);
317 * High-resolution timer tick.
318 * Runs from hardirq context with interrupts disabled.
320 static enum hrtimer_restart hrtick(struct hrtimer *timer)
322 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
324 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
326 raw_spin_lock(&rq->lock);
328 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
329 raw_spin_unlock(&rq->lock);
331 return HRTIMER_NORESTART;
336 static void __hrtick_restart(struct rq *rq)
338 struct hrtimer *timer = &rq->hrtick_timer;
340 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
344 * called from hardirq (IPI) context
346 static void __hrtick_start(void *arg)
350 raw_spin_lock(&rq->lock);
351 __hrtick_restart(rq);
352 rq->hrtick_csd_pending = 0;
353 raw_spin_unlock(&rq->lock);
357 * Called to set the hrtick timer state.
359 * called with rq->lock held and irqs disabled
361 void hrtick_start(struct rq *rq, u64 delay)
363 struct hrtimer *timer = &rq->hrtick_timer;
368 * Don't schedule slices shorter than 10000ns, that just
369 * doesn't make sense and can cause timer DoS.
371 delta = max_t(s64, delay, 10000LL);
372 time = ktime_add_ns(timer->base->get_time(), delta);
374 hrtimer_set_expires(timer, time);
376 if (rq == this_rq()) {
377 __hrtick_restart(rq);
378 } else if (!rq->hrtick_csd_pending) {
379 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
380 rq->hrtick_csd_pending = 1;
385 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
387 int cpu = (int)(long)hcpu;
390 case CPU_UP_CANCELED:
391 case CPU_UP_CANCELED_FROZEN:
392 case CPU_DOWN_PREPARE:
393 case CPU_DOWN_PREPARE_FROZEN:
395 case CPU_DEAD_FROZEN:
396 hrtick_clear(cpu_rq(cpu));
403 static __init void init_hrtick(void)
405 hotcpu_notifier(hotplug_hrtick, 0);
409 * Called to set the hrtick timer state.
411 * called with rq->lock held and irqs disabled
413 void hrtick_start(struct rq *rq, u64 delay)
416 * Don't schedule slices shorter than 10000ns, that just
417 * doesn't make sense. Rely on vruntime for fairness.
419 delay = max_t(u64, delay, 10000LL);
420 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
421 HRTIMER_MODE_REL_PINNED);
424 static inline void init_hrtick(void)
427 #endif /* CONFIG_SMP */
429 static void init_rq_hrtick(struct rq *rq)
432 rq->hrtick_csd_pending = 0;
434 rq->hrtick_csd.flags = 0;
435 rq->hrtick_csd.func = __hrtick_start;
436 rq->hrtick_csd.info = rq;
439 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
440 rq->hrtick_timer.function = hrtick;
442 #else /* CONFIG_SCHED_HRTICK */
443 static inline void hrtick_clear(struct rq *rq)
447 static inline void init_rq_hrtick(struct rq *rq)
451 static inline void init_hrtick(void)
454 #endif /* CONFIG_SCHED_HRTICK */
457 * cmpxchg based fetch_or, macro so it works for different integer types
459 #define fetch_or(ptr, val) \
460 ({ typeof(*(ptr)) __old, __val = *(ptr); \
462 __old = cmpxchg((ptr), __val, __val | (val)); \
463 if (__old == __val) \
470 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
472 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
473 * this avoids any races wrt polling state changes and thereby avoids
476 static bool set_nr_and_not_polling(struct task_struct *p)
478 struct thread_info *ti = task_thread_info(p);
479 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
483 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
485 * If this returns true, then the idle task promises to call
486 * sched_ttwu_pending() and reschedule soon.
488 static bool set_nr_if_polling(struct task_struct *p)
490 struct thread_info *ti = task_thread_info(p);
491 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
494 if (!(val & _TIF_POLLING_NRFLAG))
496 if (val & _TIF_NEED_RESCHED)
498 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
507 static bool set_nr_and_not_polling(struct task_struct *p)
509 set_tsk_need_resched(p);
514 static bool set_nr_if_polling(struct task_struct *p)
521 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
523 struct wake_q_node *node = &task->wake_q;
526 * Atomically grab the task, if ->wake_q is !nil already it means
527 * its already queued (either by us or someone else) and will get the
528 * wakeup due to that.
530 * This cmpxchg() implies a full barrier, which pairs with the write
531 * barrier implied by the wakeup in wake_up_list().
533 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
536 get_task_struct(task);
539 * The head is context local, there can be no concurrency.
542 head->lastp = &node->next;
545 void wake_up_q(struct wake_q_head *head)
547 struct wake_q_node *node = head->first;
549 while (node != WAKE_Q_TAIL) {
550 struct task_struct *task;
552 task = container_of(node, struct task_struct, wake_q);
554 /* task can safely be re-inserted now */
556 task->wake_q.next = NULL;
559 * wake_up_process() implies a wmb() to pair with the queueing
560 * in wake_q_add() so as not to miss wakeups.
562 wake_up_process(task);
563 put_task_struct(task);
568 * resched_curr - mark rq's current task 'to be rescheduled now'.
570 * On UP this means the setting of the need_resched flag, on SMP it
571 * might also involve a cross-CPU call to trigger the scheduler on
574 void resched_curr(struct rq *rq)
576 struct task_struct *curr = rq->curr;
579 lockdep_assert_held(&rq->lock);
581 if (test_tsk_need_resched(curr))
586 if (cpu == smp_processor_id()) {
587 set_tsk_need_resched(curr);
588 set_preempt_need_resched();
592 if (set_nr_and_not_polling(curr))
593 smp_send_reschedule(cpu);
595 trace_sched_wake_idle_without_ipi(cpu);
598 void resched_cpu(int cpu)
600 struct rq *rq = cpu_rq(cpu);
603 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
606 raw_spin_unlock_irqrestore(&rq->lock, flags);
610 #ifdef CONFIG_NO_HZ_COMMON
612 * In the semi idle case, use the nearest busy cpu for migrating timers
613 * from an idle cpu. This is good for power-savings.
615 * We don't do similar optimization for completely idle system, as
616 * selecting an idle cpu will add more delays to the timers than intended
617 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
619 int get_nohz_timer_target(void)
621 int i, cpu = smp_processor_id();
622 struct sched_domain *sd;
624 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
628 for_each_domain(cpu, sd) {
629 for_each_cpu(i, sched_domain_span(sd)) {
630 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
637 if (!is_housekeeping_cpu(cpu))
638 cpu = housekeeping_any_cpu();
644 * When add_timer_on() enqueues a timer into the timer wheel of an
645 * idle CPU then this timer might expire before the next timer event
646 * which is scheduled to wake up that CPU. In case of a completely
647 * idle system the next event might even be infinite time into the
648 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
649 * leaves the inner idle loop so the newly added timer is taken into
650 * account when the CPU goes back to idle and evaluates the timer
651 * wheel for the next timer event.
653 static void wake_up_idle_cpu(int cpu)
655 struct rq *rq = cpu_rq(cpu);
657 if (cpu == smp_processor_id())
660 if (set_nr_and_not_polling(rq->idle))
661 smp_send_reschedule(cpu);
663 trace_sched_wake_idle_without_ipi(cpu);
666 static bool wake_up_full_nohz_cpu(int cpu)
669 * We just need the target to call irq_exit() and re-evaluate
670 * the next tick. The nohz full kick at least implies that.
671 * If needed we can still optimize that later with an
674 if (tick_nohz_full_cpu(cpu)) {
675 if (cpu != smp_processor_id() ||
676 tick_nohz_tick_stopped())
677 tick_nohz_full_kick_cpu(cpu);
684 void wake_up_nohz_cpu(int cpu)
686 if (!wake_up_full_nohz_cpu(cpu))
687 wake_up_idle_cpu(cpu);
690 static inline bool got_nohz_idle_kick(void)
692 int cpu = smp_processor_id();
694 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
697 if (idle_cpu(cpu) && !need_resched())
701 * We can't run Idle Load Balance on this CPU for this time so we
702 * cancel it and clear NOHZ_BALANCE_KICK
704 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
708 #else /* CONFIG_NO_HZ_COMMON */
710 static inline bool got_nohz_idle_kick(void)
715 #endif /* CONFIG_NO_HZ_COMMON */
717 #ifdef CONFIG_NO_HZ_FULL
718 bool sched_can_stop_tick(void)
721 * FIFO realtime policy runs the highest priority task. Other runnable
722 * tasks are of a lower priority. The scheduler tick does nothing.
724 if (current->policy == SCHED_FIFO)
728 * Round-robin realtime tasks time slice with other tasks at the same
729 * realtime priority. Is this task the only one at this priority?
731 if (current->policy == SCHED_RR) {
732 struct sched_rt_entity *rt_se = ¤t->rt;
734 return rt_se->run_list.prev == rt_se->run_list.next;
738 * More than one running task need preemption.
739 * nr_running update is assumed to be visible
740 * after IPI is sent from wakers.
742 if (this_rq()->nr_running > 1)
747 #endif /* CONFIG_NO_HZ_FULL */
749 void sched_avg_update(struct rq *rq)
751 s64 period = sched_avg_period();
753 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
755 * Inline assembly required to prevent the compiler
756 * optimising this loop into a divmod call.
757 * See __iter_div_u64_rem() for another example of this.
759 asm("" : "+rm" (rq->age_stamp));
760 rq->age_stamp += period;
765 #endif /* CONFIG_SMP */
767 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
768 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
770 * Iterate task_group tree rooted at *from, calling @down when first entering a
771 * node and @up when leaving it for the final time.
773 * Caller must hold rcu_lock or sufficient equivalent.
775 int walk_tg_tree_from(struct task_group *from,
776 tg_visitor down, tg_visitor up, void *data)
778 struct task_group *parent, *child;
784 ret = (*down)(parent, data);
787 list_for_each_entry_rcu(child, &parent->children, siblings) {
794 ret = (*up)(parent, data);
795 if (ret || parent == from)
799 parent = parent->parent;
806 int tg_nop(struct task_group *tg, void *data)
812 static void set_load_weight(struct task_struct *p)
814 int prio = p->static_prio - MAX_RT_PRIO;
815 struct load_weight *load = &p->se.load;
818 * SCHED_IDLE tasks get minimal weight:
820 if (idle_policy(p->policy)) {
821 load->weight = scale_load(WEIGHT_IDLEPRIO);
822 load->inv_weight = WMULT_IDLEPRIO;
826 load->weight = scale_load(prio_to_weight[prio]);
827 load->inv_weight = prio_to_wmult[prio];
830 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
833 if (!(flags & ENQUEUE_RESTORE))
834 sched_info_queued(rq, p);
835 p->sched_class->enqueue_task(rq, p, flags);
838 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
841 if (!(flags & DEQUEUE_SAVE))
842 sched_info_dequeued(rq, p);
843 p->sched_class->dequeue_task(rq, p, flags);
846 void activate_task(struct rq *rq, struct task_struct *p, int flags)
848 if (task_contributes_to_load(p))
849 rq->nr_uninterruptible--;
851 enqueue_task(rq, p, flags);
854 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
856 if (task_contributes_to_load(p))
857 rq->nr_uninterruptible++;
859 dequeue_task(rq, p, flags);
862 static void update_rq_clock_task(struct rq *rq, s64 delta)
865 * In theory, the compile should just see 0 here, and optimize out the call
866 * to sched_rt_avg_update. But I don't trust it...
868 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
869 s64 steal = 0, irq_delta = 0;
871 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
872 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
875 * Since irq_time is only updated on {soft,}irq_exit, we might run into
876 * this case when a previous update_rq_clock() happened inside a
879 * When this happens, we stop ->clock_task and only update the
880 * prev_irq_time stamp to account for the part that fit, so that a next
881 * update will consume the rest. This ensures ->clock_task is
884 * It does however cause some slight miss-attribution of {soft,}irq
885 * time, a more accurate solution would be to update the irq_time using
886 * the current rq->clock timestamp, except that would require using
889 if (irq_delta > delta)
892 rq->prev_irq_time += irq_delta;
895 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
896 if (static_key_false((¶virt_steal_rq_enabled))) {
897 steal = paravirt_steal_clock(cpu_of(rq));
898 steal -= rq->prev_steal_time_rq;
900 if (unlikely(steal > delta))
903 rq->prev_steal_time_rq += steal;
908 rq->clock_task += delta;
910 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
911 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
912 sched_rt_avg_update(rq, irq_delta + steal);
916 void sched_set_stop_task(int cpu, struct task_struct *stop)
918 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
919 struct task_struct *old_stop = cpu_rq(cpu)->stop;
923 * Make it appear like a SCHED_FIFO task, its something
924 * userspace knows about and won't get confused about.
926 * Also, it will make PI more or less work without too
927 * much confusion -- but then, stop work should not
928 * rely on PI working anyway.
930 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
932 stop->sched_class = &stop_sched_class;
935 cpu_rq(cpu)->stop = stop;
939 * Reset it back to a normal scheduling class so that
940 * it can die in pieces.
942 old_stop->sched_class = &rt_sched_class;
947 * __normal_prio - return the priority that is based on the static prio
949 static inline int __normal_prio(struct task_struct *p)
951 return p->static_prio;
955 * Calculate the expected normal priority: i.e. priority
956 * without taking RT-inheritance into account. Might be
957 * boosted by interactivity modifiers. Changes upon fork,
958 * setprio syscalls, and whenever the interactivity
959 * estimator recalculates.
961 static inline int normal_prio(struct task_struct *p)
965 if (task_has_dl_policy(p))
966 prio = MAX_DL_PRIO-1;
967 else if (task_has_rt_policy(p))
968 prio = MAX_RT_PRIO-1 - p->rt_priority;
970 prio = __normal_prio(p);
975 * Calculate the current priority, i.e. the priority
976 * taken into account by the scheduler. This value might
977 * be boosted by RT tasks, or might be boosted by
978 * interactivity modifiers. Will be RT if the task got
979 * RT-boosted. If not then it returns p->normal_prio.
981 static int effective_prio(struct task_struct *p)
983 p->normal_prio = normal_prio(p);
985 * If we are RT tasks or we were boosted to RT priority,
986 * keep the priority unchanged. Otherwise, update priority
987 * to the normal priority:
989 if (!rt_prio(p->prio))
990 return p->normal_prio;
995 * task_curr - is this task currently executing on a CPU?
996 * @p: the task in question.
998 * Return: 1 if the task is currently executing. 0 otherwise.
1000 inline int task_curr(const struct task_struct *p)
1002 return cpu_curr(task_cpu(p)) == p;
1006 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1007 * use the balance_callback list if you want balancing.
1009 * this means any call to check_class_changed() must be followed by a call to
1010 * balance_callback().
1012 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1013 const struct sched_class *prev_class,
1016 if (prev_class != p->sched_class) {
1017 if (prev_class->switched_from)
1018 prev_class->switched_from(rq, p);
1020 p->sched_class->switched_to(rq, p);
1021 } else if (oldprio != p->prio || dl_task(p))
1022 p->sched_class->prio_changed(rq, p, oldprio);
1025 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1027 const struct sched_class *class;
1029 if (p->sched_class == rq->curr->sched_class) {
1030 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1032 for_each_class(class) {
1033 if (class == rq->curr->sched_class)
1035 if (class == p->sched_class) {
1043 * A queue event has occurred, and we're going to schedule. In
1044 * this case, we can save a useless back to back clock update.
1046 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1047 rq_clock_skip_update(rq, true);
1052 * This is how migration works:
1054 * 1) we invoke migration_cpu_stop() on the target CPU using
1056 * 2) stopper starts to run (implicitly forcing the migrated thread
1058 * 3) it checks whether the migrated task is still in the wrong runqueue.
1059 * 4) if it's in the wrong runqueue then the migration thread removes
1060 * it and puts it into the right queue.
1061 * 5) stopper completes and stop_one_cpu() returns and the migration
1066 * move_queued_task - move a queued task to new rq.
1068 * Returns (locked) new rq. Old rq's lock is released.
1070 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1072 lockdep_assert_held(&rq->lock);
1074 dequeue_task(rq, p, 0);
1075 p->on_rq = TASK_ON_RQ_MIGRATING;
1076 set_task_cpu(p, new_cpu);
1077 raw_spin_unlock(&rq->lock);
1079 rq = cpu_rq(new_cpu);
1081 raw_spin_lock(&rq->lock);
1082 BUG_ON(task_cpu(p) != new_cpu);
1083 p->on_rq = TASK_ON_RQ_QUEUED;
1084 enqueue_task(rq, p, 0);
1085 check_preempt_curr(rq, p, 0);
1090 struct migration_arg {
1091 struct task_struct *task;
1096 * Move (not current) task off this cpu, onto dest cpu. We're doing
1097 * this because either it can't run here any more (set_cpus_allowed()
1098 * away from this CPU, or CPU going down), or because we're
1099 * attempting to rebalance this task on exec (sched_exec).
1101 * So we race with normal scheduler movements, but that's OK, as long
1102 * as the task is no longer on this CPU.
1104 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1106 if (unlikely(!cpu_active(dest_cpu)))
1109 /* Affinity changed (again). */
1110 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1113 rq = move_queued_task(rq, p, dest_cpu);
1119 * migration_cpu_stop - this will be executed by a highprio stopper thread
1120 * and performs thread migration by bumping thread off CPU then
1121 * 'pushing' onto another runqueue.
1123 static int migration_cpu_stop(void *data)
1125 struct migration_arg *arg = data;
1126 struct task_struct *p = arg->task;
1127 struct rq *rq = this_rq();
1130 * The original target cpu might have gone down and we might
1131 * be on another cpu but it doesn't matter.
1133 local_irq_disable();
1135 * We need to explicitly wake pending tasks before running
1136 * __migrate_task() such that we will not miss enforcing cpus_allowed
1137 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1139 sched_ttwu_pending();
1141 raw_spin_lock(&p->pi_lock);
1142 raw_spin_lock(&rq->lock);
1144 * If task_rq(p) != rq, it cannot be migrated here, because we're
1145 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1146 * we're holding p->pi_lock.
1148 if (task_rq(p) == rq && task_on_rq_queued(p))
1149 rq = __migrate_task(rq, p, arg->dest_cpu);
1150 raw_spin_unlock(&rq->lock);
1151 raw_spin_unlock(&p->pi_lock);
1158 * sched_class::set_cpus_allowed must do the below, but is not required to
1159 * actually call this function.
1161 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1163 cpumask_copy(&p->cpus_allowed, new_mask);
1164 p->nr_cpus_allowed = cpumask_weight(new_mask);
1167 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1169 struct rq *rq = task_rq(p);
1170 bool queued, running;
1172 lockdep_assert_held(&p->pi_lock);
1174 queued = task_on_rq_queued(p);
1175 running = task_current(rq, p);
1179 * Because __kthread_bind() calls this on blocked tasks without
1182 lockdep_assert_held(&rq->lock);
1183 dequeue_task(rq, p, DEQUEUE_SAVE);
1186 put_prev_task(rq, p);
1188 p->sched_class->set_cpus_allowed(p, new_mask);
1191 p->sched_class->set_curr_task(rq);
1193 enqueue_task(rq, p, ENQUEUE_RESTORE);
1197 * Change a given task's CPU affinity. Migrate the thread to a
1198 * proper CPU and schedule it away if the CPU it's executing on
1199 * is removed from the allowed bitmask.
1201 * NOTE: the caller must have a valid reference to the task, the
1202 * task must not exit() & deallocate itself prematurely. The
1203 * call is not atomic; no spinlocks may be held.
1205 static int __set_cpus_allowed_ptr(struct task_struct *p,
1206 const struct cpumask *new_mask, bool check)
1208 unsigned long flags;
1210 unsigned int dest_cpu;
1213 rq = task_rq_lock(p, &flags);
1216 * Must re-check here, to close a race against __kthread_bind(),
1217 * sched_setaffinity() is not guaranteed to observe the flag.
1219 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1224 if (cpumask_equal(&p->cpus_allowed, new_mask))
1227 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1232 do_set_cpus_allowed(p, new_mask);
1234 /* Can the task run on the task's current CPU? If so, we're done */
1235 if (cpumask_test_cpu(task_cpu(p), new_mask))
1238 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1239 if (task_running(rq, p) || p->state == TASK_WAKING) {
1240 struct migration_arg arg = { p, dest_cpu };
1241 /* Need help from migration thread: drop lock and wait. */
1242 task_rq_unlock(rq, p, &flags);
1243 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1244 tlb_migrate_finish(p->mm);
1246 } else if (task_on_rq_queued(p)) {
1248 * OK, since we're going to drop the lock immediately
1249 * afterwards anyway.
1251 lockdep_unpin_lock(&rq->lock);
1252 rq = move_queued_task(rq, p, dest_cpu);
1253 lockdep_pin_lock(&rq->lock);
1256 task_rq_unlock(rq, p, &flags);
1261 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1263 return __set_cpus_allowed_ptr(p, new_mask, false);
1265 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1267 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1269 #ifdef CONFIG_SCHED_DEBUG
1271 * We should never call set_task_cpu() on a blocked task,
1272 * ttwu() will sort out the placement.
1274 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1277 #ifdef CONFIG_LOCKDEP
1279 * The caller should hold either p->pi_lock or rq->lock, when changing
1280 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1282 * sched_move_task() holds both and thus holding either pins the cgroup,
1285 * Furthermore, all task_rq users should acquire both locks, see
1288 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1289 lockdep_is_held(&task_rq(p)->lock)));
1293 trace_sched_migrate_task(p, new_cpu);
1295 if (task_cpu(p) != new_cpu) {
1296 if (p->sched_class->migrate_task_rq)
1297 p->sched_class->migrate_task_rq(p);
1298 p->se.nr_migrations++;
1299 perf_event_task_migrate(p);
1302 __set_task_cpu(p, new_cpu);
1305 static void __migrate_swap_task(struct task_struct *p, int cpu)
1307 if (task_on_rq_queued(p)) {
1308 struct rq *src_rq, *dst_rq;
1310 src_rq = task_rq(p);
1311 dst_rq = cpu_rq(cpu);
1313 deactivate_task(src_rq, p, 0);
1314 set_task_cpu(p, cpu);
1315 activate_task(dst_rq, p, 0);
1316 check_preempt_curr(dst_rq, p, 0);
1319 * Task isn't running anymore; make it appear like we migrated
1320 * it before it went to sleep. This means on wakeup we make the
1321 * previous cpu our targer instead of where it really is.
1327 struct migration_swap_arg {
1328 struct task_struct *src_task, *dst_task;
1329 int src_cpu, dst_cpu;
1332 static int migrate_swap_stop(void *data)
1334 struct migration_swap_arg *arg = data;
1335 struct rq *src_rq, *dst_rq;
1338 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1341 src_rq = cpu_rq(arg->src_cpu);
1342 dst_rq = cpu_rq(arg->dst_cpu);
1344 double_raw_lock(&arg->src_task->pi_lock,
1345 &arg->dst_task->pi_lock);
1346 double_rq_lock(src_rq, dst_rq);
1348 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1351 if (task_cpu(arg->src_task) != arg->src_cpu)
1354 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1357 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1360 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1361 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1366 double_rq_unlock(src_rq, dst_rq);
1367 raw_spin_unlock(&arg->dst_task->pi_lock);
1368 raw_spin_unlock(&arg->src_task->pi_lock);
1374 * Cross migrate two tasks
1376 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1378 struct migration_swap_arg arg;
1381 arg = (struct migration_swap_arg){
1383 .src_cpu = task_cpu(cur),
1385 .dst_cpu = task_cpu(p),
1388 if (arg.src_cpu == arg.dst_cpu)
1392 * These three tests are all lockless; this is OK since all of them
1393 * will be re-checked with proper locks held further down the line.
1395 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1398 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1401 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1404 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1405 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1412 * wait_task_inactive - wait for a thread to unschedule.
1414 * If @match_state is nonzero, it's the @p->state value just checked and
1415 * not expected to change. If it changes, i.e. @p might have woken up,
1416 * then return zero. When we succeed in waiting for @p to be off its CPU,
1417 * we return a positive number (its total switch count). If a second call
1418 * a short while later returns the same number, the caller can be sure that
1419 * @p has remained unscheduled the whole time.
1421 * The caller must ensure that the task *will* unschedule sometime soon,
1422 * else this function might spin for a *long* time. This function can't
1423 * be called with interrupts off, or it may introduce deadlock with
1424 * smp_call_function() if an IPI is sent by the same process we are
1425 * waiting to become inactive.
1427 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1429 unsigned long flags;
1430 int running, queued;
1436 * We do the initial early heuristics without holding
1437 * any task-queue locks at all. We'll only try to get
1438 * the runqueue lock when things look like they will
1444 * If the task is actively running on another CPU
1445 * still, just relax and busy-wait without holding
1448 * NOTE! Since we don't hold any locks, it's not
1449 * even sure that "rq" stays as the right runqueue!
1450 * But we don't care, since "task_running()" will
1451 * return false if the runqueue has changed and p
1452 * is actually now running somewhere else!
1454 while (task_running(rq, p)) {
1455 if (match_state && unlikely(p->state != match_state))
1461 * Ok, time to look more closely! We need the rq
1462 * lock now, to be *sure*. If we're wrong, we'll
1463 * just go back and repeat.
1465 rq = task_rq_lock(p, &flags);
1466 trace_sched_wait_task(p);
1467 running = task_running(rq, p);
1468 queued = task_on_rq_queued(p);
1470 if (!match_state || p->state == match_state)
1471 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1472 task_rq_unlock(rq, p, &flags);
1475 * If it changed from the expected state, bail out now.
1477 if (unlikely(!ncsw))
1481 * Was it really running after all now that we
1482 * checked with the proper locks actually held?
1484 * Oops. Go back and try again..
1486 if (unlikely(running)) {
1492 * It's not enough that it's not actively running,
1493 * it must be off the runqueue _entirely_, and not
1496 * So if it was still runnable (but just not actively
1497 * running right now), it's preempted, and we should
1498 * yield - it could be a while.
1500 if (unlikely(queued)) {
1501 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1503 set_current_state(TASK_UNINTERRUPTIBLE);
1504 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1509 * Ahh, all good. It wasn't running, and it wasn't
1510 * runnable, which means that it will never become
1511 * running in the future either. We're all done!
1520 * kick_process - kick a running thread to enter/exit the kernel
1521 * @p: the to-be-kicked thread
1523 * Cause a process which is running on another CPU to enter
1524 * kernel-mode, without any delay. (to get signals handled.)
1526 * NOTE: this function doesn't have to take the runqueue lock,
1527 * because all it wants to ensure is that the remote task enters
1528 * the kernel. If the IPI races and the task has been migrated
1529 * to another CPU then no harm is done and the purpose has been
1532 void kick_process(struct task_struct *p)
1538 if ((cpu != smp_processor_id()) && task_curr(p))
1539 smp_send_reschedule(cpu);
1542 EXPORT_SYMBOL_GPL(kick_process);
1545 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1547 static int select_fallback_rq(int cpu, struct task_struct *p)
1549 int nid = cpu_to_node(cpu);
1550 const struct cpumask *nodemask = NULL;
1551 enum { cpuset, possible, fail } state = cpuset;
1555 * If the node that the cpu is on has been offlined, cpu_to_node()
1556 * will return -1. There is no cpu on the node, and we should
1557 * select the cpu on the other node.
1560 nodemask = cpumask_of_node(nid);
1562 /* Look for allowed, online CPU in same node. */
1563 for_each_cpu(dest_cpu, nodemask) {
1564 if (!cpu_online(dest_cpu))
1566 if (!cpu_active(dest_cpu))
1568 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1574 /* Any allowed, online CPU? */
1575 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1576 if (!cpu_online(dest_cpu))
1578 if (!cpu_active(dest_cpu))
1583 /* No more Mr. Nice Guy. */
1586 if (IS_ENABLED(CONFIG_CPUSETS)) {
1587 cpuset_cpus_allowed_fallback(p);
1593 do_set_cpus_allowed(p, cpu_possible_mask);
1604 if (state != cpuset) {
1606 * Don't tell them about moving exiting tasks or
1607 * kernel threads (both mm NULL), since they never
1610 if (p->mm && printk_ratelimit()) {
1611 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1612 task_pid_nr(p), p->comm, cpu);
1620 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1623 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1625 lockdep_assert_held(&p->pi_lock);
1627 if (p->nr_cpus_allowed > 1)
1628 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1631 * In order not to call set_task_cpu() on a blocking task we need
1632 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1635 * Since this is common to all placement strategies, this lives here.
1637 * [ this allows ->select_task() to simply return task_cpu(p) and
1638 * not worry about this generic constraint ]
1640 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1642 cpu = select_fallback_rq(task_cpu(p), p);
1647 static void update_avg(u64 *avg, u64 sample)
1649 s64 diff = sample - *avg;
1655 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1656 const struct cpumask *new_mask, bool check)
1658 return set_cpus_allowed_ptr(p, new_mask);
1661 #endif /* CONFIG_SMP */
1664 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1666 #ifdef CONFIG_SCHEDSTATS
1667 struct rq *rq = this_rq();
1670 int this_cpu = smp_processor_id();
1672 if (cpu == this_cpu) {
1673 schedstat_inc(rq, ttwu_local);
1674 schedstat_inc(p, se.statistics.nr_wakeups_local);
1676 struct sched_domain *sd;
1678 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1680 for_each_domain(this_cpu, sd) {
1681 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1682 schedstat_inc(sd, ttwu_wake_remote);
1689 if (wake_flags & WF_MIGRATED)
1690 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1692 #endif /* CONFIG_SMP */
1694 schedstat_inc(rq, ttwu_count);
1695 schedstat_inc(p, se.statistics.nr_wakeups);
1697 if (wake_flags & WF_SYNC)
1698 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1700 #endif /* CONFIG_SCHEDSTATS */
1703 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1705 activate_task(rq, p, en_flags);
1706 p->on_rq = TASK_ON_RQ_QUEUED;
1708 /* if a worker is waking up, notify workqueue */
1709 if (p->flags & PF_WQ_WORKER)
1710 wq_worker_waking_up(p, cpu_of(rq));
1714 * Mark the task runnable and perform wakeup-preemption.
1717 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1719 check_preempt_curr(rq, p, wake_flags);
1720 p->state = TASK_RUNNING;
1721 trace_sched_wakeup(p);
1724 if (p->sched_class->task_woken) {
1726 * Our task @p is fully woken up and running; so its safe to
1727 * drop the rq->lock, hereafter rq is only used for statistics.
1729 lockdep_unpin_lock(&rq->lock);
1730 p->sched_class->task_woken(rq, p);
1731 lockdep_pin_lock(&rq->lock);
1734 if (rq->idle_stamp) {
1735 u64 delta = rq_clock(rq) - rq->idle_stamp;
1736 u64 max = 2*rq->max_idle_balance_cost;
1738 update_avg(&rq->avg_idle, delta);
1740 if (rq->avg_idle > max)
1749 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1751 lockdep_assert_held(&rq->lock);
1754 if (p->sched_contributes_to_load)
1755 rq->nr_uninterruptible--;
1758 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1759 ttwu_do_wakeup(rq, p, wake_flags);
1763 * Called in case the task @p isn't fully descheduled from its runqueue,
1764 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1765 * since all we need to do is flip p->state to TASK_RUNNING, since
1766 * the task is still ->on_rq.
1768 static int ttwu_remote(struct task_struct *p, int wake_flags)
1773 rq = __task_rq_lock(p);
1774 if (task_on_rq_queued(p)) {
1775 /* check_preempt_curr() may use rq clock */
1776 update_rq_clock(rq);
1777 ttwu_do_wakeup(rq, p, wake_flags);
1780 __task_rq_unlock(rq);
1786 void sched_ttwu_pending(void)
1788 struct rq *rq = this_rq();
1789 struct llist_node *llist = llist_del_all(&rq->wake_list);
1790 struct task_struct *p;
1791 unsigned long flags;
1796 raw_spin_lock_irqsave(&rq->lock, flags);
1797 lockdep_pin_lock(&rq->lock);
1800 p = llist_entry(llist, struct task_struct, wake_entry);
1801 llist = llist_next(llist);
1802 ttwu_do_activate(rq, p, 0);
1805 lockdep_unpin_lock(&rq->lock);
1806 raw_spin_unlock_irqrestore(&rq->lock, flags);
1809 void scheduler_ipi(void)
1812 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1813 * TIF_NEED_RESCHED remotely (for the first time) will also send
1816 preempt_fold_need_resched();
1818 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1822 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1823 * traditionally all their work was done from the interrupt return
1824 * path. Now that we actually do some work, we need to make sure
1827 * Some archs already do call them, luckily irq_enter/exit nest
1830 * Arguably we should visit all archs and update all handlers,
1831 * however a fair share of IPIs are still resched only so this would
1832 * somewhat pessimize the simple resched case.
1835 sched_ttwu_pending();
1838 * Check if someone kicked us for doing the nohz idle load balance.
1840 if (unlikely(got_nohz_idle_kick())) {
1841 this_rq()->idle_balance = 1;
1842 raise_softirq_irqoff(SCHED_SOFTIRQ);
1847 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1849 struct rq *rq = cpu_rq(cpu);
1851 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1852 if (!set_nr_if_polling(rq->idle))
1853 smp_send_reschedule(cpu);
1855 trace_sched_wake_idle_without_ipi(cpu);
1859 void wake_up_if_idle(int cpu)
1861 struct rq *rq = cpu_rq(cpu);
1862 unsigned long flags;
1866 if (!is_idle_task(rcu_dereference(rq->curr)))
1869 if (set_nr_if_polling(rq->idle)) {
1870 trace_sched_wake_idle_without_ipi(cpu);
1872 raw_spin_lock_irqsave(&rq->lock, flags);
1873 if (is_idle_task(rq->curr))
1874 smp_send_reschedule(cpu);
1875 /* Else cpu is not in idle, do nothing here */
1876 raw_spin_unlock_irqrestore(&rq->lock, flags);
1883 bool cpus_share_cache(int this_cpu, int that_cpu)
1885 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1887 #endif /* CONFIG_SMP */
1889 static void ttwu_queue(struct task_struct *p, int cpu)
1891 struct rq *rq = cpu_rq(cpu);
1893 #if defined(CONFIG_SMP)
1894 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1895 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1896 ttwu_queue_remote(p, cpu);
1901 raw_spin_lock(&rq->lock);
1902 lockdep_pin_lock(&rq->lock);
1903 ttwu_do_activate(rq, p, 0);
1904 lockdep_unpin_lock(&rq->lock);
1905 raw_spin_unlock(&rq->lock);
1909 * try_to_wake_up - wake up a thread
1910 * @p: the thread to be awakened
1911 * @state: the mask of task states that can be woken
1912 * @wake_flags: wake modifier flags (WF_*)
1914 * Put it on the run-queue if it's not already there. The "current"
1915 * thread is always on the run-queue (except when the actual
1916 * re-schedule is in progress), and as such you're allowed to do
1917 * the simpler "current->state = TASK_RUNNING" to mark yourself
1918 * runnable without the overhead of this.
1920 * Return: %true if @p was woken up, %false if it was already running.
1921 * or @state didn't match @p's state.
1924 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1926 unsigned long flags;
1927 int cpu, success = 0;
1930 * If we are going to wake up a thread waiting for CONDITION we
1931 * need to ensure that CONDITION=1 done by the caller can not be
1932 * reordered with p->state check below. This pairs with mb() in
1933 * set_current_state() the waiting thread does.
1935 smp_mb__before_spinlock();
1936 raw_spin_lock_irqsave(&p->pi_lock, flags);
1937 if (!(p->state & state))
1940 trace_sched_waking(p);
1942 success = 1; /* we're going to change ->state */
1945 if (p->on_rq && ttwu_remote(p, wake_flags))
1950 * If the owning (remote) cpu is still in the middle of schedule() with
1951 * this task as prev, wait until its done referencing the task.
1956 * Combined with the control dependency above, we have an effective
1957 * smp_load_acquire() without the need for full barriers.
1959 * Pairs with the smp_store_release() in finish_lock_switch().
1961 * This ensures that tasks getting woken will be fully ordered against
1962 * their previous state and preserve Program Order.
1966 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1967 p->state = TASK_WAKING;
1969 if (p->sched_class->task_waking)
1970 p->sched_class->task_waking(p);
1972 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1973 if (task_cpu(p) != cpu) {
1974 wake_flags |= WF_MIGRATED;
1975 set_task_cpu(p, cpu);
1977 #endif /* CONFIG_SMP */
1981 ttwu_stat(p, cpu, wake_flags);
1983 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1989 * try_to_wake_up_local - try to wake up a local task with rq lock held
1990 * @p: the thread to be awakened
1992 * Put @p on the run-queue if it's not already there. The caller must
1993 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1996 static void try_to_wake_up_local(struct task_struct *p)
1998 struct rq *rq = task_rq(p);
2000 if (WARN_ON_ONCE(rq != this_rq()) ||
2001 WARN_ON_ONCE(p == current))
2004 lockdep_assert_held(&rq->lock);
2006 if (!raw_spin_trylock(&p->pi_lock)) {
2008 * This is OK, because current is on_cpu, which avoids it being
2009 * picked for load-balance and preemption/IRQs are still
2010 * disabled avoiding further scheduler activity on it and we've
2011 * not yet picked a replacement task.
2013 lockdep_unpin_lock(&rq->lock);
2014 raw_spin_unlock(&rq->lock);
2015 raw_spin_lock(&p->pi_lock);
2016 raw_spin_lock(&rq->lock);
2017 lockdep_pin_lock(&rq->lock);
2020 if (!(p->state & TASK_NORMAL))
2023 trace_sched_waking(p);
2025 if (!task_on_rq_queued(p))
2026 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2028 ttwu_do_wakeup(rq, p, 0);
2029 ttwu_stat(p, smp_processor_id(), 0);
2031 raw_spin_unlock(&p->pi_lock);
2035 * wake_up_process - Wake up a specific process
2036 * @p: The process to be woken up.
2038 * Attempt to wake up the nominated process and move it to the set of runnable
2041 * Return: 1 if the process was woken up, 0 if it was already running.
2043 * It may be assumed that this function implies a write memory barrier before
2044 * changing the task state if and only if any tasks are woken up.
2046 int wake_up_process(struct task_struct *p)
2048 return try_to_wake_up(p, TASK_NORMAL, 0);
2050 EXPORT_SYMBOL(wake_up_process);
2052 int wake_up_state(struct task_struct *p, unsigned int state)
2054 return try_to_wake_up(p, state, 0);
2058 * This function clears the sched_dl_entity static params.
2060 void __dl_clear_params(struct task_struct *p)
2062 struct sched_dl_entity *dl_se = &p->dl;
2064 dl_se->dl_runtime = 0;
2065 dl_se->dl_deadline = 0;
2066 dl_se->dl_period = 0;
2070 dl_se->dl_throttled = 0;
2072 dl_se->dl_yielded = 0;
2076 * Perform scheduler related setup for a newly forked process p.
2077 * p is forked by current.
2079 * __sched_fork() is basic setup used by init_idle() too:
2081 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2086 p->se.exec_start = 0;
2087 p->se.sum_exec_runtime = 0;
2088 p->se.prev_sum_exec_runtime = 0;
2089 p->se.nr_migrations = 0;
2091 INIT_LIST_HEAD(&p->se.group_node);
2093 #ifdef CONFIG_SCHEDSTATS
2094 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2097 RB_CLEAR_NODE(&p->dl.rb_node);
2098 init_dl_task_timer(&p->dl);
2099 __dl_clear_params(p);
2101 INIT_LIST_HEAD(&p->rt.run_list);
2103 #ifdef CONFIG_PREEMPT_NOTIFIERS
2104 INIT_HLIST_HEAD(&p->preempt_notifiers);
2107 #ifdef CONFIG_NUMA_BALANCING
2108 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2109 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2110 p->mm->numa_scan_seq = 0;
2113 if (clone_flags & CLONE_VM)
2114 p->numa_preferred_nid = current->numa_preferred_nid;
2116 p->numa_preferred_nid = -1;
2118 p->node_stamp = 0ULL;
2119 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2120 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2121 p->numa_work.next = &p->numa_work;
2122 p->numa_faults = NULL;
2123 p->last_task_numa_placement = 0;
2124 p->last_sum_exec_runtime = 0;
2126 p->numa_group = NULL;
2127 #endif /* CONFIG_NUMA_BALANCING */
2130 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2132 #ifdef CONFIG_NUMA_BALANCING
2134 void set_numabalancing_state(bool enabled)
2137 static_branch_enable(&sched_numa_balancing);
2139 static_branch_disable(&sched_numa_balancing);
2142 #ifdef CONFIG_PROC_SYSCTL
2143 int sysctl_numa_balancing(struct ctl_table *table, int write,
2144 void __user *buffer, size_t *lenp, loff_t *ppos)
2148 int state = static_branch_likely(&sched_numa_balancing);
2150 if (write && !capable(CAP_SYS_ADMIN))
2155 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2159 set_numabalancing_state(state);
2166 * fork()/clone()-time setup:
2168 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2170 unsigned long flags;
2171 int cpu = get_cpu();
2173 __sched_fork(clone_flags, p);
2175 * We mark the process as running here. This guarantees that
2176 * nobody will actually run it, and a signal or other external
2177 * event cannot wake it up and insert it on the runqueue either.
2179 p->state = TASK_RUNNING;
2182 * Make sure we do not leak PI boosting priority to the child.
2184 p->prio = current->normal_prio;
2187 * Revert to default priority/policy on fork if requested.
2189 if (unlikely(p->sched_reset_on_fork)) {
2190 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2191 p->policy = SCHED_NORMAL;
2192 p->static_prio = NICE_TO_PRIO(0);
2194 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2195 p->static_prio = NICE_TO_PRIO(0);
2197 p->prio = p->normal_prio = __normal_prio(p);
2201 * We don't need the reset flag anymore after the fork. It has
2202 * fulfilled its duty:
2204 p->sched_reset_on_fork = 0;
2207 if (dl_prio(p->prio)) {
2210 } else if (rt_prio(p->prio)) {
2211 p->sched_class = &rt_sched_class;
2213 p->sched_class = &fair_sched_class;
2216 if (p->sched_class->task_fork)
2217 p->sched_class->task_fork(p);
2220 * The child is not yet in the pid-hash so no cgroup attach races,
2221 * and the cgroup is pinned to this child due to cgroup_fork()
2222 * is ran before sched_fork().
2224 * Silence PROVE_RCU.
2226 raw_spin_lock_irqsave(&p->pi_lock, flags);
2227 set_task_cpu(p, cpu);
2228 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2230 #ifdef CONFIG_SCHED_INFO
2231 if (likely(sched_info_on()))
2232 memset(&p->sched_info, 0, sizeof(p->sched_info));
2234 #if defined(CONFIG_SMP)
2237 init_task_preempt_count(p);
2239 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2240 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2247 unsigned long to_ratio(u64 period, u64 runtime)
2249 if (runtime == RUNTIME_INF)
2253 * Doing this here saves a lot of checks in all
2254 * the calling paths, and returning zero seems
2255 * safe for them anyway.
2260 return div64_u64(runtime << 20, period);
2264 inline struct dl_bw *dl_bw_of(int i)
2266 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2267 "sched RCU must be held");
2268 return &cpu_rq(i)->rd->dl_bw;
2271 static inline int dl_bw_cpus(int i)
2273 struct root_domain *rd = cpu_rq(i)->rd;
2276 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2277 "sched RCU must be held");
2278 for_each_cpu_and(i, rd->span, cpu_active_mask)
2284 inline struct dl_bw *dl_bw_of(int i)
2286 return &cpu_rq(i)->dl.dl_bw;
2289 static inline int dl_bw_cpus(int i)
2296 * We must be sure that accepting a new task (or allowing changing the
2297 * parameters of an existing one) is consistent with the bandwidth
2298 * constraints. If yes, this function also accordingly updates the currently
2299 * allocated bandwidth to reflect the new situation.
2301 * This function is called while holding p's rq->lock.
2303 * XXX we should delay bw change until the task's 0-lag point, see
2306 static int dl_overflow(struct task_struct *p, int policy,
2307 const struct sched_attr *attr)
2310 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2311 u64 period = attr->sched_period ?: attr->sched_deadline;
2312 u64 runtime = attr->sched_runtime;
2313 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2316 if (new_bw == p->dl.dl_bw)
2320 * Either if a task, enters, leave, or stays -deadline but changes
2321 * its parameters, we may need to update accordingly the total
2322 * allocated bandwidth of the container.
2324 raw_spin_lock(&dl_b->lock);
2325 cpus = dl_bw_cpus(task_cpu(p));
2326 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2327 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2328 __dl_add(dl_b, new_bw);
2330 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2331 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2332 __dl_clear(dl_b, p->dl.dl_bw);
2333 __dl_add(dl_b, new_bw);
2335 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2336 __dl_clear(dl_b, p->dl.dl_bw);
2339 raw_spin_unlock(&dl_b->lock);
2344 extern void init_dl_bw(struct dl_bw *dl_b);
2347 * wake_up_new_task - wake up a newly created task for the first time.
2349 * This function will do some initial scheduler statistics housekeeping
2350 * that must be done for every newly created context, then puts the task
2351 * on the runqueue and wakes it.
2353 void wake_up_new_task(struct task_struct *p)
2355 unsigned long flags;
2358 raw_spin_lock_irqsave(&p->pi_lock, flags);
2359 /* Initialize new task's runnable average */
2360 init_entity_runnable_average(&p->se);
2363 * Fork balancing, do it here and not earlier because:
2364 * - cpus_allowed can change in the fork path
2365 * - any previously selected cpu might disappear through hotplug
2367 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2370 rq = __task_rq_lock(p);
2371 activate_task(rq, p, 0);
2372 p->on_rq = TASK_ON_RQ_QUEUED;
2373 trace_sched_wakeup_new(p);
2374 check_preempt_curr(rq, p, WF_FORK);
2376 if (p->sched_class->task_woken) {
2378 * Nothing relies on rq->lock after this, so its fine to
2381 lockdep_unpin_lock(&rq->lock);
2382 p->sched_class->task_woken(rq, p);
2383 lockdep_pin_lock(&rq->lock);
2386 task_rq_unlock(rq, p, &flags);
2389 #ifdef CONFIG_PREEMPT_NOTIFIERS
2391 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2393 void preempt_notifier_inc(void)
2395 static_key_slow_inc(&preempt_notifier_key);
2397 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2399 void preempt_notifier_dec(void)
2401 static_key_slow_dec(&preempt_notifier_key);
2403 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2406 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2407 * @notifier: notifier struct to register
2409 void preempt_notifier_register(struct preempt_notifier *notifier)
2411 if (!static_key_false(&preempt_notifier_key))
2412 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2414 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2416 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2419 * preempt_notifier_unregister - no longer interested in preemption notifications
2420 * @notifier: notifier struct to unregister
2422 * This is *not* safe to call from within a preemption notifier.
2424 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2426 hlist_del(¬ifier->link);
2428 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2430 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2432 struct preempt_notifier *notifier;
2434 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2435 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2438 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2440 if (static_key_false(&preempt_notifier_key))
2441 __fire_sched_in_preempt_notifiers(curr);
2445 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2446 struct task_struct *next)
2448 struct preempt_notifier *notifier;
2450 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2451 notifier->ops->sched_out(notifier, next);
2454 static __always_inline void
2455 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2456 struct task_struct *next)
2458 if (static_key_false(&preempt_notifier_key))
2459 __fire_sched_out_preempt_notifiers(curr, next);
2462 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2464 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2469 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2470 struct task_struct *next)
2474 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2477 * prepare_task_switch - prepare to switch tasks
2478 * @rq: the runqueue preparing to switch
2479 * @prev: the current task that is being switched out
2480 * @next: the task we are going to switch to.
2482 * This is called with the rq lock held and interrupts off. It must
2483 * be paired with a subsequent finish_task_switch after the context
2486 * prepare_task_switch sets up locking and calls architecture specific
2490 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2491 struct task_struct *next)
2493 sched_info_switch(rq, prev, next);
2494 perf_event_task_sched_out(prev, next);
2495 fire_sched_out_preempt_notifiers(prev, next);
2496 prepare_lock_switch(rq, next);
2497 prepare_arch_switch(next);
2501 * finish_task_switch - clean up after a task-switch
2502 * @prev: the thread we just switched away from.
2504 * finish_task_switch must be called after the context switch, paired
2505 * with a prepare_task_switch call before the context switch.
2506 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2507 * and do any other architecture-specific cleanup actions.
2509 * Note that we may have delayed dropping an mm in context_switch(). If
2510 * so, we finish that here outside of the runqueue lock. (Doing it
2511 * with the lock held can cause deadlocks; see schedule() for
2514 * The context switch have flipped the stack from under us and restored the
2515 * local variables which were saved when this task called schedule() in the
2516 * past. prev == current is still correct but we need to recalculate this_rq
2517 * because prev may have moved to another CPU.
2519 static struct rq *finish_task_switch(struct task_struct *prev)
2520 __releases(rq->lock)
2522 struct rq *rq = this_rq();
2523 struct mm_struct *mm = rq->prev_mm;
2527 * The previous task will have left us with a preempt_count of 2
2528 * because it left us after:
2531 * preempt_disable(); // 1
2533 * raw_spin_lock_irq(&rq->lock) // 2
2535 * Also, see FORK_PREEMPT_COUNT.
2537 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2538 "corrupted preempt_count: %s/%d/0x%x\n",
2539 current->comm, current->pid, preempt_count()))
2540 preempt_count_set(FORK_PREEMPT_COUNT);
2545 * A task struct has one reference for the use as "current".
2546 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2547 * schedule one last time. The schedule call will never return, and
2548 * the scheduled task must drop that reference.
2550 * We must observe prev->state before clearing prev->on_cpu (in
2551 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2552 * running on another CPU and we could rave with its RUNNING -> DEAD
2553 * transition, resulting in a double drop.
2555 prev_state = prev->state;
2556 vtime_task_switch(prev);
2557 perf_event_task_sched_in(prev, current);
2558 finish_lock_switch(rq, prev);
2559 finish_arch_post_lock_switch();
2561 fire_sched_in_preempt_notifiers(current);
2564 if (unlikely(prev_state == TASK_DEAD)) {
2565 if (prev->sched_class->task_dead)
2566 prev->sched_class->task_dead(prev);
2569 * Remove function-return probe instances associated with this
2570 * task and put them back on the free list.
2572 kprobe_flush_task(prev);
2573 put_task_struct(prev);
2576 tick_nohz_task_switch();
2582 /* rq->lock is NOT held, but preemption is disabled */
2583 static void __balance_callback(struct rq *rq)
2585 struct callback_head *head, *next;
2586 void (*func)(struct rq *rq);
2587 unsigned long flags;
2589 raw_spin_lock_irqsave(&rq->lock, flags);
2590 head = rq->balance_callback;
2591 rq->balance_callback = NULL;
2593 func = (void (*)(struct rq *))head->func;
2600 raw_spin_unlock_irqrestore(&rq->lock, flags);
2603 static inline void balance_callback(struct rq *rq)
2605 if (unlikely(rq->balance_callback))
2606 __balance_callback(rq);
2611 static inline void balance_callback(struct rq *rq)
2618 * schedule_tail - first thing a freshly forked thread must call.
2619 * @prev: the thread we just switched away from.
2621 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2622 __releases(rq->lock)
2627 * New tasks start with FORK_PREEMPT_COUNT, see there and
2628 * finish_task_switch() for details.
2630 * finish_task_switch() will drop rq->lock() and lower preempt_count
2631 * and the preempt_enable() will end up enabling preemption (on
2632 * PREEMPT_COUNT kernels).
2635 rq = finish_task_switch(prev);
2636 balance_callback(rq);
2639 if (current->set_child_tid)
2640 put_user(task_pid_vnr(current), current->set_child_tid);
2644 * context_switch - switch to the new MM and the new thread's register state.
2646 static inline struct rq *
2647 context_switch(struct rq *rq, struct task_struct *prev,
2648 struct task_struct *next)
2650 struct mm_struct *mm, *oldmm;
2652 prepare_task_switch(rq, prev, next);
2655 oldmm = prev->active_mm;
2657 * For paravirt, this is coupled with an exit in switch_to to
2658 * combine the page table reload and the switch backend into
2661 arch_start_context_switch(prev);
2664 next->active_mm = oldmm;
2665 atomic_inc(&oldmm->mm_count);
2666 enter_lazy_tlb(oldmm, next);
2668 switch_mm(oldmm, mm, next);
2671 prev->active_mm = NULL;
2672 rq->prev_mm = oldmm;
2675 * Since the runqueue lock will be released by the next
2676 * task (which is an invalid locking op but in the case
2677 * of the scheduler it's an obvious special-case), so we
2678 * do an early lockdep release here:
2680 lockdep_unpin_lock(&rq->lock);
2681 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2683 /* Here we just switch the register state and the stack. */
2684 switch_to(prev, next, prev);
2687 return finish_task_switch(prev);
2691 * nr_running and nr_context_switches:
2693 * externally visible scheduler statistics: current number of runnable
2694 * threads, total number of context switches performed since bootup.
2696 unsigned long nr_running(void)
2698 unsigned long i, sum = 0;
2700 for_each_online_cpu(i)
2701 sum += cpu_rq(i)->nr_running;
2707 * Check if only the current task is running on the cpu.
2709 * Caution: this function does not check that the caller has disabled
2710 * preemption, thus the result might have a time-of-check-to-time-of-use
2711 * race. The caller is responsible to use it correctly, for example:
2713 * - from a non-preemptable section (of course)
2715 * - from a thread that is bound to a single CPU
2717 * - in a loop with very short iterations (e.g. a polling loop)
2719 bool single_task_running(void)
2721 return raw_rq()->nr_running == 1;
2723 EXPORT_SYMBOL(single_task_running);
2725 unsigned long long nr_context_switches(void)
2728 unsigned long long sum = 0;
2730 for_each_possible_cpu(i)
2731 sum += cpu_rq(i)->nr_switches;
2736 unsigned long nr_iowait(void)
2738 unsigned long i, sum = 0;
2740 for_each_possible_cpu(i)
2741 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2746 unsigned long nr_iowait_cpu(int cpu)
2748 struct rq *this = cpu_rq(cpu);
2749 return atomic_read(&this->nr_iowait);
2752 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2754 struct rq *rq = this_rq();
2755 *nr_waiters = atomic_read(&rq->nr_iowait);
2756 *load = rq->load.weight;
2762 * sched_exec - execve() is a valuable balancing opportunity, because at
2763 * this point the task has the smallest effective memory and cache footprint.
2765 void sched_exec(void)
2767 struct task_struct *p = current;
2768 unsigned long flags;
2771 raw_spin_lock_irqsave(&p->pi_lock, flags);
2772 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2773 if (dest_cpu == smp_processor_id())
2776 if (likely(cpu_active(dest_cpu))) {
2777 struct migration_arg arg = { p, dest_cpu };
2779 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2780 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2784 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2789 DEFINE_PER_CPU(struct kernel_stat, kstat);
2790 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2792 EXPORT_PER_CPU_SYMBOL(kstat);
2793 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2796 * Return accounted runtime for the task.
2797 * In case the task is currently running, return the runtime plus current's
2798 * pending runtime that have not been accounted yet.
2800 unsigned long long task_sched_runtime(struct task_struct *p)
2802 unsigned long flags;
2806 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2808 * 64-bit doesn't need locks to atomically read a 64bit value.
2809 * So we have a optimization chance when the task's delta_exec is 0.
2810 * Reading ->on_cpu is racy, but this is ok.
2812 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2813 * If we race with it entering cpu, unaccounted time is 0. This is
2814 * indistinguishable from the read occurring a few cycles earlier.
2815 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2816 * been accounted, so we're correct here as well.
2818 if (!p->on_cpu || !task_on_rq_queued(p))
2819 return p->se.sum_exec_runtime;
2822 rq = task_rq_lock(p, &flags);
2824 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2825 * project cycles that may never be accounted to this
2826 * thread, breaking clock_gettime().
2828 if (task_current(rq, p) && task_on_rq_queued(p)) {
2829 update_rq_clock(rq);
2830 p->sched_class->update_curr(rq);
2832 ns = p->se.sum_exec_runtime;
2833 task_rq_unlock(rq, p, &flags);
2839 * This function gets called by the timer code, with HZ frequency.
2840 * We call it with interrupts disabled.
2842 void scheduler_tick(void)
2844 int cpu = smp_processor_id();
2845 struct rq *rq = cpu_rq(cpu);
2846 struct task_struct *curr = rq->curr;
2850 raw_spin_lock(&rq->lock);
2851 update_rq_clock(rq);
2852 curr->sched_class->task_tick(rq, curr, 0);
2853 update_cpu_load_active(rq);
2854 calc_global_load_tick(rq);
2855 raw_spin_unlock(&rq->lock);
2857 perf_event_task_tick();
2860 rq->idle_balance = idle_cpu(cpu);
2861 trigger_load_balance(rq);
2863 rq_last_tick_reset(rq);
2866 #ifdef CONFIG_NO_HZ_FULL
2868 * scheduler_tick_max_deferment
2870 * Keep at least one tick per second when a single
2871 * active task is running because the scheduler doesn't
2872 * yet completely support full dynticks environment.
2874 * This makes sure that uptime, CFS vruntime, load
2875 * balancing, etc... continue to move forward, even
2876 * with a very low granularity.
2878 * Return: Maximum deferment in nanoseconds.
2880 u64 scheduler_tick_max_deferment(void)
2882 struct rq *rq = this_rq();
2883 unsigned long next, now = READ_ONCE(jiffies);
2885 next = rq->last_sched_tick + HZ;
2887 if (time_before_eq(next, now))
2890 return jiffies_to_nsecs(next - now);
2894 notrace unsigned long get_parent_ip(unsigned long addr)
2896 if (in_lock_functions(addr)) {
2897 addr = CALLER_ADDR2;
2898 if (in_lock_functions(addr))
2899 addr = CALLER_ADDR3;
2904 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2905 defined(CONFIG_PREEMPT_TRACER))
2907 void preempt_count_add(int val)
2909 #ifdef CONFIG_DEBUG_PREEMPT
2913 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2916 __preempt_count_add(val);
2917 #ifdef CONFIG_DEBUG_PREEMPT
2919 * Spinlock count overflowing soon?
2921 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2924 if (preempt_count() == val) {
2925 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2926 #ifdef CONFIG_DEBUG_PREEMPT
2927 current->preempt_disable_ip = ip;
2929 trace_preempt_off(CALLER_ADDR0, ip);
2932 EXPORT_SYMBOL(preempt_count_add);
2933 NOKPROBE_SYMBOL(preempt_count_add);
2935 void preempt_count_sub(int val)
2937 #ifdef CONFIG_DEBUG_PREEMPT
2941 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2944 * Is the spinlock portion underflowing?
2946 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2947 !(preempt_count() & PREEMPT_MASK)))
2951 if (preempt_count() == val)
2952 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2953 __preempt_count_sub(val);
2955 EXPORT_SYMBOL(preempt_count_sub);
2956 NOKPROBE_SYMBOL(preempt_count_sub);
2961 * Print scheduling while atomic bug:
2963 static noinline void __schedule_bug(struct task_struct *prev)
2965 if (oops_in_progress)
2968 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2969 prev->comm, prev->pid, preempt_count());
2971 debug_show_held_locks(prev);
2973 if (irqs_disabled())
2974 print_irqtrace_events(prev);
2975 #ifdef CONFIG_DEBUG_PREEMPT
2976 if (in_atomic_preempt_off()) {
2977 pr_err("Preemption disabled at:");
2978 print_ip_sym(current->preempt_disable_ip);
2983 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2987 * Various schedule()-time debugging checks and statistics:
2989 static inline void schedule_debug(struct task_struct *prev)
2991 #ifdef CONFIG_SCHED_STACK_END_CHECK
2992 BUG_ON(task_stack_end_corrupted(prev));
2995 if (unlikely(in_atomic_preempt_off())) {
2996 __schedule_bug(prev);
2997 preempt_count_set(PREEMPT_DISABLED);
3001 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3003 schedstat_inc(this_rq(), sched_count);
3007 * Pick up the highest-prio task:
3009 static inline struct task_struct *
3010 pick_next_task(struct rq *rq, struct task_struct *prev)
3012 const struct sched_class *class = &fair_sched_class;
3013 struct task_struct *p;
3016 * Optimization: we know that if all tasks are in
3017 * the fair class we can call that function directly:
3019 if (likely(prev->sched_class == class &&
3020 rq->nr_running == rq->cfs.h_nr_running)) {
3021 p = fair_sched_class.pick_next_task(rq, prev);
3022 if (unlikely(p == RETRY_TASK))
3025 /* assumes fair_sched_class->next == idle_sched_class */
3027 p = idle_sched_class.pick_next_task(rq, prev);
3033 for_each_class(class) {
3034 p = class->pick_next_task(rq, prev);
3036 if (unlikely(p == RETRY_TASK))
3042 BUG(); /* the idle class will always have a runnable task */
3046 * __schedule() is the main scheduler function.
3048 * The main means of driving the scheduler and thus entering this function are:
3050 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3052 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3053 * paths. For example, see arch/x86/entry_64.S.
3055 * To drive preemption between tasks, the scheduler sets the flag in timer
3056 * interrupt handler scheduler_tick().
3058 * 3. Wakeups don't really cause entry into schedule(). They add a
3059 * task to the run-queue and that's it.
3061 * Now, if the new task added to the run-queue preempts the current
3062 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3063 * called on the nearest possible occasion:
3065 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3067 * - in syscall or exception context, at the next outmost
3068 * preempt_enable(). (this might be as soon as the wake_up()'s
3071 * - in IRQ context, return from interrupt-handler to
3072 * preemptible context
3074 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3077 * - cond_resched() call
3078 * - explicit schedule() call
3079 * - return from syscall or exception to user-space
3080 * - return from interrupt-handler to user-space
3082 * WARNING: must be called with preemption disabled!
3084 static void __sched notrace __schedule(bool preempt)
3086 struct task_struct *prev, *next;
3087 unsigned long *switch_count;
3091 cpu = smp_processor_id();
3093 rcu_note_context_switch();
3097 * do_exit() calls schedule() with preemption disabled as an exception;
3098 * however we must fix that up, otherwise the next task will see an
3099 * inconsistent (higher) preempt count.
3101 * It also avoids the below schedule_debug() test from complaining
3104 if (unlikely(prev->state == TASK_DEAD))
3105 preempt_enable_no_resched_notrace();
3107 schedule_debug(prev);
3109 if (sched_feat(HRTICK))
3113 * Make sure that signal_pending_state()->signal_pending() below
3114 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3115 * done by the caller to avoid the race with signal_wake_up().
3117 smp_mb__before_spinlock();
3118 raw_spin_lock_irq(&rq->lock);
3119 lockdep_pin_lock(&rq->lock);
3121 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3123 switch_count = &prev->nivcsw;
3124 if (!preempt && prev->state) {
3125 if (unlikely(signal_pending_state(prev->state, prev))) {
3126 prev->state = TASK_RUNNING;
3128 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3132 * If a worker went to sleep, notify and ask workqueue
3133 * whether it wants to wake up a task to maintain
3136 if (prev->flags & PF_WQ_WORKER) {
3137 struct task_struct *to_wakeup;
3139 to_wakeup = wq_worker_sleeping(prev, cpu);
3141 try_to_wake_up_local(to_wakeup);
3144 switch_count = &prev->nvcsw;
3147 if (task_on_rq_queued(prev))
3148 update_rq_clock(rq);
3150 next = pick_next_task(rq, prev);
3151 clear_tsk_need_resched(prev);
3152 clear_preempt_need_resched();
3153 rq->clock_skip_update = 0;
3155 if (likely(prev != next)) {
3160 trace_sched_switch(preempt, prev, next);
3161 rq = context_switch(rq, prev, next); /* unlocks the rq */
3164 lockdep_unpin_lock(&rq->lock);
3165 raw_spin_unlock_irq(&rq->lock);
3168 balance_callback(rq);
3171 static inline void sched_submit_work(struct task_struct *tsk)
3173 if (!tsk->state || tsk_is_pi_blocked(tsk))
3176 * If we are going to sleep and we have plugged IO queued,
3177 * make sure to submit it to avoid deadlocks.
3179 if (blk_needs_flush_plug(tsk))
3180 blk_schedule_flush_plug(tsk);
3183 asmlinkage __visible void __sched schedule(void)
3185 struct task_struct *tsk = current;
3187 sched_submit_work(tsk);
3191 sched_preempt_enable_no_resched();
3192 } while (need_resched());
3194 EXPORT_SYMBOL(schedule);
3196 #ifdef CONFIG_CONTEXT_TRACKING
3197 asmlinkage __visible void __sched schedule_user(void)
3200 * If we come here after a random call to set_need_resched(),
3201 * or we have been woken up remotely but the IPI has not yet arrived,
3202 * we haven't yet exited the RCU idle mode. Do it here manually until
3203 * we find a better solution.
3205 * NB: There are buggy callers of this function. Ideally we
3206 * should warn if prev_state != CONTEXT_USER, but that will trigger
3207 * too frequently to make sense yet.
3209 enum ctx_state prev_state = exception_enter();
3211 exception_exit(prev_state);
3216 * schedule_preempt_disabled - called with preemption disabled
3218 * Returns with preemption disabled. Note: preempt_count must be 1
3220 void __sched schedule_preempt_disabled(void)
3222 sched_preempt_enable_no_resched();
3227 static void __sched notrace preempt_schedule_common(void)
3230 preempt_disable_notrace();
3232 preempt_enable_no_resched_notrace();
3235 * Check again in case we missed a preemption opportunity
3236 * between schedule and now.
3238 } while (need_resched());
3241 #ifdef CONFIG_PREEMPT
3243 * this is the entry point to schedule() from in-kernel preemption
3244 * off of preempt_enable. Kernel preemptions off return from interrupt
3245 * occur there and call schedule directly.
3247 asmlinkage __visible void __sched notrace preempt_schedule(void)
3250 * If there is a non-zero preempt_count or interrupts are disabled,
3251 * we do not want to preempt the current task. Just return..
3253 if (likely(!preemptible()))
3256 preempt_schedule_common();
3258 NOKPROBE_SYMBOL(preempt_schedule);
3259 EXPORT_SYMBOL(preempt_schedule);
3262 * preempt_schedule_notrace - preempt_schedule called by tracing
3264 * The tracing infrastructure uses preempt_enable_notrace to prevent
3265 * recursion and tracing preempt enabling caused by the tracing
3266 * infrastructure itself. But as tracing can happen in areas coming
3267 * from userspace or just about to enter userspace, a preempt enable
3268 * can occur before user_exit() is called. This will cause the scheduler
3269 * to be called when the system is still in usermode.
3271 * To prevent this, the preempt_enable_notrace will use this function
3272 * instead of preempt_schedule() to exit user context if needed before
3273 * calling the scheduler.
3275 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3277 enum ctx_state prev_ctx;
3279 if (likely(!preemptible()))
3283 preempt_disable_notrace();
3285 * Needs preempt disabled in case user_exit() is traced
3286 * and the tracer calls preempt_enable_notrace() causing
3287 * an infinite recursion.
3289 prev_ctx = exception_enter();
3291 exception_exit(prev_ctx);
3293 preempt_enable_no_resched_notrace();
3294 } while (need_resched());
3296 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3298 #endif /* CONFIG_PREEMPT */
3301 * this is the entry point to schedule() from kernel preemption
3302 * off of irq context.
3303 * Note, that this is called and return with irqs disabled. This will
3304 * protect us against recursive calling from irq.
3306 asmlinkage __visible void __sched preempt_schedule_irq(void)
3308 enum ctx_state prev_state;
3310 /* Catch callers which need to be fixed */
3311 BUG_ON(preempt_count() || !irqs_disabled());
3313 prev_state = exception_enter();
3319 local_irq_disable();
3320 sched_preempt_enable_no_resched();
3321 } while (need_resched());
3323 exception_exit(prev_state);
3326 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3329 return try_to_wake_up(curr->private, mode, wake_flags);
3331 EXPORT_SYMBOL(default_wake_function);
3333 #ifdef CONFIG_RT_MUTEXES
3336 * rt_mutex_setprio - set the current priority of a task
3338 * @prio: prio value (kernel-internal form)
3340 * This function changes the 'effective' priority of a task. It does
3341 * not touch ->normal_prio like __setscheduler().
3343 * Used by the rt_mutex code to implement priority inheritance
3344 * logic. Call site only calls if the priority of the task changed.
3346 void rt_mutex_setprio(struct task_struct *p, int prio)
3348 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3350 const struct sched_class *prev_class;
3352 BUG_ON(prio > MAX_PRIO);
3354 rq = __task_rq_lock(p);
3357 * Idle task boosting is a nono in general. There is one
3358 * exception, when PREEMPT_RT and NOHZ is active:
3360 * The idle task calls get_next_timer_interrupt() and holds
3361 * the timer wheel base->lock on the CPU and another CPU wants
3362 * to access the timer (probably to cancel it). We can safely
3363 * ignore the boosting request, as the idle CPU runs this code
3364 * with interrupts disabled and will complete the lock
3365 * protected section without being interrupted. So there is no
3366 * real need to boost.
3368 if (unlikely(p == rq->idle)) {
3369 WARN_ON(p != rq->curr);
3370 WARN_ON(p->pi_blocked_on);
3374 trace_sched_pi_setprio(p, prio);
3376 prev_class = p->sched_class;
3377 queued = task_on_rq_queued(p);
3378 running = task_current(rq, p);
3380 dequeue_task(rq, p, DEQUEUE_SAVE);
3382 put_prev_task(rq, p);
3385 * Boosting condition are:
3386 * 1. -rt task is running and holds mutex A
3387 * --> -dl task blocks on mutex A
3389 * 2. -dl task is running and holds mutex A
3390 * --> -dl task blocks on mutex A and could preempt the
3393 if (dl_prio(prio)) {
3394 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3395 if (!dl_prio(p->normal_prio) ||
3396 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3397 p->dl.dl_boosted = 1;
3398 enqueue_flag |= ENQUEUE_REPLENISH;
3400 p->dl.dl_boosted = 0;
3401 p->sched_class = &dl_sched_class;
3402 } else if (rt_prio(prio)) {
3403 if (dl_prio(oldprio))
3404 p->dl.dl_boosted = 0;
3406 enqueue_flag |= ENQUEUE_HEAD;
3407 p->sched_class = &rt_sched_class;
3409 if (dl_prio(oldprio))
3410 p->dl.dl_boosted = 0;
3411 if (rt_prio(oldprio))
3413 p->sched_class = &fair_sched_class;
3419 p->sched_class->set_curr_task(rq);
3421 enqueue_task(rq, p, enqueue_flag);
3423 check_class_changed(rq, p, prev_class, oldprio);
3425 preempt_disable(); /* avoid rq from going away on us */
3426 __task_rq_unlock(rq);
3428 balance_callback(rq);
3433 void set_user_nice(struct task_struct *p, long nice)
3435 int old_prio, delta, queued;
3436 unsigned long flags;
3439 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3442 * We have to be careful, if called from sys_setpriority(),
3443 * the task might be in the middle of scheduling on another CPU.
3445 rq = task_rq_lock(p, &flags);
3447 * The RT priorities are set via sched_setscheduler(), but we still
3448 * allow the 'normal' nice value to be set - but as expected
3449 * it wont have any effect on scheduling until the task is
3450 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3452 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3453 p->static_prio = NICE_TO_PRIO(nice);
3456 queued = task_on_rq_queued(p);
3458 dequeue_task(rq, p, DEQUEUE_SAVE);
3460 p->static_prio = NICE_TO_PRIO(nice);
3463 p->prio = effective_prio(p);
3464 delta = p->prio - old_prio;
3467 enqueue_task(rq, p, ENQUEUE_RESTORE);
3469 * If the task increased its priority or is running and
3470 * lowered its priority, then reschedule its CPU:
3472 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3476 task_rq_unlock(rq, p, &flags);
3478 EXPORT_SYMBOL(set_user_nice);
3481 * can_nice - check if a task can reduce its nice value
3485 int can_nice(const struct task_struct *p, const int nice)
3487 /* convert nice value [19,-20] to rlimit style value [1,40] */
3488 int nice_rlim = nice_to_rlimit(nice);
3490 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3491 capable(CAP_SYS_NICE));
3494 #ifdef __ARCH_WANT_SYS_NICE
3497 * sys_nice - change the priority of the current process.
3498 * @increment: priority increment
3500 * sys_setpriority is a more generic, but much slower function that
3501 * does similar things.
3503 SYSCALL_DEFINE1(nice, int, increment)
3508 * Setpriority might change our priority at the same moment.
3509 * We don't have to worry. Conceptually one call occurs first
3510 * and we have a single winner.
3512 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3513 nice = task_nice(current) + increment;
3515 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3516 if (increment < 0 && !can_nice(current, nice))
3519 retval = security_task_setnice(current, nice);
3523 set_user_nice(current, nice);
3530 * task_prio - return the priority value of a given task.
3531 * @p: the task in question.
3533 * Return: The priority value as seen by users in /proc.
3534 * RT tasks are offset by -200. Normal tasks are centered
3535 * around 0, value goes from -16 to +15.
3537 int task_prio(const struct task_struct *p)
3539 return p->prio - MAX_RT_PRIO;
3543 * idle_cpu - is a given cpu idle currently?
3544 * @cpu: the processor in question.
3546 * Return: 1 if the CPU is currently idle. 0 otherwise.
3548 int idle_cpu(int cpu)
3550 struct rq *rq = cpu_rq(cpu);
3552 if (rq->curr != rq->idle)
3559 if (!llist_empty(&rq->wake_list))
3567 * idle_task - return the idle task for a given cpu.
3568 * @cpu: the processor in question.
3570 * Return: The idle task for the cpu @cpu.
3572 struct task_struct *idle_task(int cpu)
3574 return cpu_rq(cpu)->idle;
3578 * find_process_by_pid - find a process with a matching PID value.
3579 * @pid: the pid in question.
3581 * The task of @pid, if found. %NULL otherwise.
3583 static struct task_struct *find_process_by_pid(pid_t pid)
3585 return pid ? find_task_by_vpid(pid) : current;
3589 * This function initializes the sched_dl_entity of a newly becoming
3590 * SCHED_DEADLINE task.
3592 * Only the static values are considered here, the actual runtime and the
3593 * absolute deadline will be properly calculated when the task is enqueued
3594 * for the first time with its new policy.
3597 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3599 struct sched_dl_entity *dl_se = &p->dl;
3601 dl_se->dl_runtime = attr->sched_runtime;
3602 dl_se->dl_deadline = attr->sched_deadline;
3603 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3604 dl_se->flags = attr->sched_flags;
3605 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3608 * Changing the parameters of a task is 'tricky' and we're not doing
3609 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3611 * What we SHOULD do is delay the bandwidth release until the 0-lag
3612 * point. This would include retaining the task_struct until that time
3613 * and change dl_overflow() to not immediately decrement the current
3616 * Instead we retain the current runtime/deadline and let the new
3617 * parameters take effect after the current reservation period lapses.
3618 * This is safe (albeit pessimistic) because the 0-lag point is always
3619 * before the current scheduling deadline.
3621 * We can still have temporary overloads because we do not delay the
3622 * change in bandwidth until that time; so admission control is
3623 * not on the safe side. It does however guarantee tasks will never
3624 * consume more than promised.
3629 * sched_setparam() passes in -1 for its policy, to let the functions
3630 * it calls know not to change it.
3632 #define SETPARAM_POLICY -1
3634 static void __setscheduler_params(struct task_struct *p,
3635 const struct sched_attr *attr)
3637 int policy = attr->sched_policy;
3639 if (policy == SETPARAM_POLICY)
3644 if (dl_policy(policy))
3645 __setparam_dl(p, attr);
3646 else if (fair_policy(policy))
3647 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3650 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3651 * !rt_policy. Always setting this ensures that things like
3652 * getparam()/getattr() don't report silly values for !rt tasks.
3654 p->rt_priority = attr->sched_priority;
3655 p->normal_prio = normal_prio(p);
3659 /* Actually do priority change: must hold pi & rq lock. */
3660 static void __setscheduler(struct rq *rq, struct task_struct *p,
3661 const struct sched_attr *attr, bool keep_boost)
3663 __setscheduler_params(p, attr);
3666 * Keep a potential priority boosting if called from
3667 * sched_setscheduler().
3670 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3672 p->prio = normal_prio(p);
3674 if (dl_prio(p->prio))
3675 p->sched_class = &dl_sched_class;
3676 else if (rt_prio(p->prio))
3677 p->sched_class = &rt_sched_class;
3679 p->sched_class = &fair_sched_class;
3683 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3685 struct sched_dl_entity *dl_se = &p->dl;
3687 attr->sched_priority = p->rt_priority;
3688 attr->sched_runtime = dl_se->dl_runtime;
3689 attr->sched_deadline = dl_se->dl_deadline;
3690 attr->sched_period = dl_se->dl_period;
3691 attr->sched_flags = dl_se->flags;
3695 * This function validates the new parameters of a -deadline task.
3696 * We ask for the deadline not being zero, and greater or equal
3697 * than the runtime, as well as the period of being zero or
3698 * greater than deadline. Furthermore, we have to be sure that
3699 * user parameters are above the internal resolution of 1us (we
3700 * check sched_runtime only since it is always the smaller one) and
3701 * below 2^63 ns (we have to check both sched_deadline and
3702 * sched_period, as the latter can be zero).
3705 __checkparam_dl(const struct sched_attr *attr)
3708 if (attr->sched_deadline == 0)
3712 * Since we truncate DL_SCALE bits, make sure we're at least
3715 if (attr->sched_runtime < (1ULL << DL_SCALE))
3719 * Since we use the MSB for wrap-around and sign issues, make
3720 * sure it's not set (mind that period can be equal to zero).
3722 if (attr->sched_deadline & (1ULL << 63) ||
3723 attr->sched_period & (1ULL << 63))
3726 /* runtime <= deadline <= period (if period != 0) */
3727 if ((attr->sched_period != 0 &&
3728 attr->sched_period < attr->sched_deadline) ||
3729 attr->sched_deadline < attr->sched_runtime)
3736 * check the target process has a UID that matches the current process's
3738 static bool check_same_owner(struct task_struct *p)
3740 const struct cred *cred = current_cred(), *pcred;
3744 pcred = __task_cred(p);
3745 match = (uid_eq(cred->euid, pcred->euid) ||
3746 uid_eq(cred->euid, pcred->uid));
3751 static bool dl_param_changed(struct task_struct *p,
3752 const struct sched_attr *attr)
3754 struct sched_dl_entity *dl_se = &p->dl;
3756 if (dl_se->dl_runtime != attr->sched_runtime ||
3757 dl_se->dl_deadline != attr->sched_deadline ||
3758 dl_se->dl_period != attr->sched_period ||
3759 dl_se->flags != attr->sched_flags)
3765 static int __sched_setscheduler(struct task_struct *p,
3766 const struct sched_attr *attr,
3769 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3770 MAX_RT_PRIO - 1 - attr->sched_priority;
3771 int retval, oldprio, oldpolicy = -1, queued, running;
3772 int new_effective_prio, policy = attr->sched_policy;
3773 unsigned long flags;
3774 const struct sched_class *prev_class;
3778 /* may grab non-irq protected spin_locks */
3779 BUG_ON(in_interrupt());
3781 /* double check policy once rq lock held */
3783 reset_on_fork = p->sched_reset_on_fork;
3784 policy = oldpolicy = p->policy;
3786 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3788 if (!valid_policy(policy))
3792 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3796 * Valid priorities for SCHED_FIFO and SCHED_RR are
3797 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3798 * SCHED_BATCH and SCHED_IDLE is 0.
3800 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3801 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3803 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3804 (rt_policy(policy) != (attr->sched_priority != 0)))
3808 * Allow unprivileged RT tasks to decrease priority:
3810 if (user && !capable(CAP_SYS_NICE)) {
3811 if (fair_policy(policy)) {
3812 if (attr->sched_nice < task_nice(p) &&
3813 !can_nice(p, attr->sched_nice))
3817 if (rt_policy(policy)) {
3818 unsigned long rlim_rtprio =
3819 task_rlimit(p, RLIMIT_RTPRIO);
3821 /* can't set/change the rt policy */
3822 if (policy != p->policy && !rlim_rtprio)
3825 /* can't increase priority */
3826 if (attr->sched_priority > p->rt_priority &&
3827 attr->sched_priority > rlim_rtprio)
3832 * Can't set/change SCHED_DEADLINE policy at all for now
3833 * (safest behavior); in the future we would like to allow
3834 * unprivileged DL tasks to increase their relative deadline
3835 * or reduce their runtime (both ways reducing utilization)
3837 if (dl_policy(policy))
3841 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3842 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3844 if (idle_policy(p->policy) && !idle_policy(policy)) {
3845 if (!can_nice(p, task_nice(p)))
3849 /* can't change other user's priorities */
3850 if (!check_same_owner(p))
3853 /* Normal users shall not reset the sched_reset_on_fork flag */
3854 if (p->sched_reset_on_fork && !reset_on_fork)
3859 retval = security_task_setscheduler(p);
3865 * make sure no PI-waiters arrive (or leave) while we are
3866 * changing the priority of the task:
3868 * To be able to change p->policy safely, the appropriate
3869 * runqueue lock must be held.
3871 rq = task_rq_lock(p, &flags);
3874 * Changing the policy of the stop threads its a very bad idea
3876 if (p == rq->stop) {
3877 task_rq_unlock(rq, p, &flags);
3882 * If not changing anything there's no need to proceed further,
3883 * but store a possible modification of reset_on_fork.
3885 if (unlikely(policy == p->policy)) {
3886 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3888 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3890 if (dl_policy(policy) && dl_param_changed(p, attr))
3893 p->sched_reset_on_fork = reset_on_fork;
3894 task_rq_unlock(rq, p, &flags);
3900 #ifdef CONFIG_RT_GROUP_SCHED
3902 * Do not allow realtime tasks into groups that have no runtime
3905 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3906 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3907 !task_group_is_autogroup(task_group(p))) {
3908 task_rq_unlock(rq, p, &flags);
3913 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3914 cpumask_t *span = rq->rd->span;
3917 * Don't allow tasks with an affinity mask smaller than
3918 * the entire root_domain to become SCHED_DEADLINE. We
3919 * will also fail if there's no bandwidth available.
3921 if (!cpumask_subset(span, &p->cpus_allowed) ||
3922 rq->rd->dl_bw.bw == 0) {
3923 task_rq_unlock(rq, p, &flags);
3930 /* recheck policy now with rq lock held */
3931 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3932 policy = oldpolicy = -1;
3933 task_rq_unlock(rq, p, &flags);
3938 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3939 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3942 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3943 task_rq_unlock(rq, p, &flags);
3947 p->sched_reset_on_fork = reset_on_fork;
3952 * Take priority boosted tasks into account. If the new
3953 * effective priority is unchanged, we just store the new
3954 * normal parameters and do not touch the scheduler class and
3955 * the runqueue. This will be done when the task deboost
3958 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3959 if (new_effective_prio == oldprio) {
3960 __setscheduler_params(p, attr);
3961 task_rq_unlock(rq, p, &flags);
3966 queued = task_on_rq_queued(p);
3967 running = task_current(rq, p);
3969 dequeue_task(rq, p, DEQUEUE_SAVE);
3971 put_prev_task(rq, p);
3973 prev_class = p->sched_class;
3974 __setscheduler(rq, p, attr, pi);
3977 p->sched_class->set_curr_task(rq);
3979 int enqueue_flags = ENQUEUE_RESTORE;
3981 * We enqueue to tail when the priority of a task is
3982 * increased (user space view).
3984 if (oldprio <= p->prio)
3985 enqueue_flags |= ENQUEUE_HEAD;
3987 enqueue_task(rq, p, enqueue_flags);
3990 check_class_changed(rq, p, prev_class, oldprio);
3991 preempt_disable(); /* avoid rq from going away on us */
3992 task_rq_unlock(rq, p, &flags);
3995 rt_mutex_adjust_pi(p);
3998 * Run balance callbacks after we've adjusted the PI chain.
4000 balance_callback(rq);
4006 static int _sched_setscheduler(struct task_struct *p, int policy,
4007 const struct sched_param *param, bool check)
4009 struct sched_attr attr = {
4010 .sched_policy = policy,
4011 .sched_priority = param->sched_priority,
4012 .sched_nice = PRIO_TO_NICE(p->static_prio),
4015 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4016 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4017 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4018 policy &= ~SCHED_RESET_ON_FORK;
4019 attr.sched_policy = policy;
4022 return __sched_setscheduler(p, &attr, check, true);
4025 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4026 * @p: the task in question.
4027 * @policy: new policy.
4028 * @param: structure containing the new RT priority.
4030 * Return: 0 on success. An error code otherwise.
4032 * NOTE that the task may be already dead.
4034 int sched_setscheduler(struct task_struct *p, int policy,
4035 const struct sched_param *param)
4037 return _sched_setscheduler(p, policy, param, true);
4039 EXPORT_SYMBOL_GPL(sched_setscheduler);
4041 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4043 return __sched_setscheduler(p, attr, true, true);
4045 EXPORT_SYMBOL_GPL(sched_setattr);
4048 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4049 * @p: the task in question.
4050 * @policy: new policy.
4051 * @param: structure containing the new RT priority.
4053 * Just like sched_setscheduler, only don't bother checking if the
4054 * current context has permission. For example, this is needed in
4055 * stop_machine(): we create temporary high priority worker threads,
4056 * but our caller might not have that capability.
4058 * Return: 0 on success. An error code otherwise.
4060 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4061 const struct sched_param *param)
4063 return _sched_setscheduler(p, policy, param, false);
4065 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4068 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4070 struct sched_param lparam;
4071 struct task_struct *p;
4074 if (!param || pid < 0)
4076 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4081 p = find_process_by_pid(pid);
4083 retval = sched_setscheduler(p, policy, &lparam);
4090 * Mimics kernel/events/core.c perf_copy_attr().
4092 static int sched_copy_attr(struct sched_attr __user *uattr,
4093 struct sched_attr *attr)
4098 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4102 * zero the full structure, so that a short copy will be nice.
4104 memset(attr, 0, sizeof(*attr));
4106 ret = get_user(size, &uattr->size);
4110 if (size > PAGE_SIZE) /* silly large */
4113 if (!size) /* abi compat */
4114 size = SCHED_ATTR_SIZE_VER0;
4116 if (size < SCHED_ATTR_SIZE_VER0)
4120 * If we're handed a bigger struct than we know of,
4121 * ensure all the unknown bits are 0 - i.e. new
4122 * user-space does not rely on any kernel feature
4123 * extensions we dont know about yet.
4125 if (size > sizeof(*attr)) {
4126 unsigned char __user *addr;
4127 unsigned char __user *end;
4130 addr = (void __user *)uattr + sizeof(*attr);
4131 end = (void __user *)uattr + size;
4133 for (; addr < end; addr++) {
4134 ret = get_user(val, addr);
4140 size = sizeof(*attr);
4143 ret = copy_from_user(attr, uattr, size);
4148 * XXX: do we want to be lenient like existing syscalls; or do we want
4149 * to be strict and return an error on out-of-bounds values?
4151 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4156 put_user(sizeof(*attr), &uattr->size);
4161 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4162 * @pid: the pid in question.
4163 * @policy: new policy.
4164 * @param: structure containing the new RT priority.
4166 * Return: 0 on success. An error code otherwise.
4168 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4169 struct sched_param __user *, param)
4171 /* negative values for policy are not valid */
4175 return do_sched_setscheduler(pid, policy, param);
4179 * sys_sched_setparam - set/change the RT priority of a thread
4180 * @pid: the pid in question.
4181 * @param: structure containing the new RT priority.
4183 * Return: 0 on success. An error code otherwise.
4185 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4187 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4191 * sys_sched_setattr - same as above, but with extended sched_attr
4192 * @pid: the pid in question.
4193 * @uattr: structure containing the extended parameters.
4194 * @flags: for future extension.
4196 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4197 unsigned int, flags)
4199 struct sched_attr attr;
4200 struct task_struct *p;
4203 if (!uattr || pid < 0 || flags)
4206 retval = sched_copy_attr(uattr, &attr);
4210 if ((int)attr.sched_policy < 0)
4215 p = find_process_by_pid(pid);
4217 retval = sched_setattr(p, &attr);
4224 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4225 * @pid: the pid in question.
4227 * Return: On success, the policy of the thread. Otherwise, a negative error
4230 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4232 struct task_struct *p;
4240 p = find_process_by_pid(pid);
4242 retval = security_task_getscheduler(p);
4245 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4252 * sys_sched_getparam - get the RT priority of a thread
4253 * @pid: the pid in question.
4254 * @param: structure containing the RT priority.
4256 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4259 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4261 struct sched_param lp = { .sched_priority = 0 };
4262 struct task_struct *p;
4265 if (!param || pid < 0)
4269 p = find_process_by_pid(pid);
4274 retval = security_task_getscheduler(p);
4278 if (task_has_rt_policy(p))
4279 lp.sched_priority = p->rt_priority;
4283 * This one might sleep, we cannot do it with a spinlock held ...
4285 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4294 static int sched_read_attr(struct sched_attr __user *uattr,
4295 struct sched_attr *attr,
4300 if (!access_ok(VERIFY_WRITE, uattr, usize))
4304 * If we're handed a smaller struct than we know of,
4305 * ensure all the unknown bits are 0 - i.e. old
4306 * user-space does not get uncomplete information.
4308 if (usize < sizeof(*attr)) {
4309 unsigned char *addr;
4312 addr = (void *)attr + usize;
4313 end = (void *)attr + sizeof(*attr);
4315 for (; addr < end; addr++) {
4323 ret = copy_to_user(uattr, attr, attr->size);
4331 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4332 * @pid: the pid in question.
4333 * @uattr: structure containing the extended parameters.
4334 * @size: sizeof(attr) for fwd/bwd comp.
4335 * @flags: for future extension.
4337 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4338 unsigned int, size, unsigned int, flags)
4340 struct sched_attr attr = {
4341 .size = sizeof(struct sched_attr),
4343 struct task_struct *p;
4346 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4347 size < SCHED_ATTR_SIZE_VER0 || flags)
4351 p = find_process_by_pid(pid);
4356 retval = security_task_getscheduler(p);
4360 attr.sched_policy = p->policy;
4361 if (p->sched_reset_on_fork)
4362 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4363 if (task_has_dl_policy(p))
4364 __getparam_dl(p, &attr);
4365 else if (task_has_rt_policy(p))
4366 attr.sched_priority = p->rt_priority;
4368 attr.sched_nice = task_nice(p);
4372 retval = sched_read_attr(uattr, &attr, size);
4380 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4382 cpumask_var_t cpus_allowed, new_mask;
4383 struct task_struct *p;
4388 p = find_process_by_pid(pid);
4394 /* Prevent p going away */
4398 if (p->flags & PF_NO_SETAFFINITY) {
4402 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4406 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4408 goto out_free_cpus_allowed;
4411 if (!check_same_owner(p)) {
4413 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4415 goto out_free_new_mask;
4420 retval = security_task_setscheduler(p);
4422 goto out_free_new_mask;
4425 cpuset_cpus_allowed(p, cpus_allowed);
4426 cpumask_and(new_mask, in_mask, cpus_allowed);
4429 * Since bandwidth control happens on root_domain basis,
4430 * if admission test is enabled, we only admit -deadline
4431 * tasks allowed to run on all the CPUs in the task's
4435 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4437 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4440 goto out_free_new_mask;
4446 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4449 cpuset_cpus_allowed(p, cpus_allowed);
4450 if (!cpumask_subset(new_mask, cpus_allowed)) {
4452 * We must have raced with a concurrent cpuset
4453 * update. Just reset the cpus_allowed to the
4454 * cpuset's cpus_allowed
4456 cpumask_copy(new_mask, cpus_allowed);
4461 free_cpumask_var(new_mask);
4462 out_free_cpus_allowed:
4463 free_cpumask_var(cpus_allowed);
4469 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4470 struct cpumask *new_mask)
4472 if (len < cpumask_size())
4473 cpumask_clear(new_mask);
4474 else if (len > cpumask_size())
4475 len = cpumask_size();
4477 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4481 * sys_sched_setaffinity - set the cpu affinity of a process
4482 * @pid: pid of the process
4483 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4484 * @user_mask_ptr: user-space pointer to the new cpu mask
4486 * Return: 0 on success. An error code otherwise.
4488 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4489 unsigned long __user *, user_mask_ptr)
4491 cpumask_var_t new_mask;
4494 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4497 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4499 retval = sched_setaffinity(pid, new_mask);
4500 free_cpumask_var(new_mask);
4504 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4506 struct task_struct *p;
4507 unsigned long flags;
4513 p = find_process_by_pid(pid);
4517 retval = security_task_getscheduler(p);
4521 raw_spin_lock_irqsave(&p->pi_lock, flags);
4522 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4523 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4532 * sys_sched_getaffinity - get the cpu affinity of a process
4533 * @pid: pid of the process
4534 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4535 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4537 * Return: 0 on success. An error code otherwise.
4539 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4540 unsigned long __user *, user_mask_ptr)
4545 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4547 if (len & (sizeof(unsigned long)-1))
4550 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4553 ret = sched_getaffinity(pid, mask);
4555 size_t retlen = min_t(size_t, len, cpumask_size());
4557 if (copy_to_user(user_mask_ptr, mask, retlen))
4562 free_cpumask_var(mask);
4568 * sys_sched_yield - yield the current processor to other threads.
4570 * This function yields the current CPU to other tasks. If there are no
4571 * other threads running on this CPU then this function will return.
4575 SYSCALL_DEFINE0(sched_yield)
4577 struct rq *rq = this_rq_lock();
4579 schedstat_inc(rq, yld_count);
4580 current->sched_class->yield_task(rq);
4583 * Since we are going to call schedule() anyway, there's
4584 * no need to preempt or enable interrupts:
4586 __release(rq->lock);
4587 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4588 do_raw_spin_unlock(&rq->lock);
4589 sched_preempt_enable_no_resched();
4596 int __sched _cond_resched(void)
4598 if (should_resched(0)) {
4599 preempt_schedule_common();
4604 EXPORT_SYMBOL(_cond_resched);
4607 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4608 * call schedule, and on return reacquire the lock.
4610 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4611 * operations here to prevent schedule() from being called twice (once via
4612 * spin_unlock(), once by hand).
4614 int __cond_resched_lock(spinlock_t *lock)
4616 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4619 lockdep_assert_held(lock);
4621 if (spin_needbreak(lock) || resched) {
4624 preempt_schedule_common();
4632 EXPORT_SYMBOL(__cond_resched_lock);
4634 int __sched __cond_resched_softirq(void)
4636 BUG_ON(!in_softirq());
4638 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4640 preempt_schedule_common();
4646 EXPORT_SYMBOL(__cond_resched_softirq);
4649 * yield - yield the current processor to other threads.
4651 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4653 * The scheduler is at all times free to pick the calling task as the most
4654 * eligible task to run, if removing the yield() call from your code breaks
4655 * it, its already broken.
4657 * Typical broken usage is:
4662 * where one assumes that yield() will let 'the other' process run that will
4663 * make event true. If the current task is a SCHED_FIFO task that will never
4664 * happen. Never use yield() as a progress guarantee!!
4666 * If you want to use yield() to wait for something, use wait_event().
4667 * If you want to use yield() to be 'nice' for others, use cond_resched().
4668 * If you still want to use yield(), do not!
4670 void __sched yield(void)
4672 set_current_state(TASK_RUNNING);
4675 EXPORT_SYMBOL(yield);
4678 * yield_to - yield the current processor to another thread in
4679 * your thread group, or accelerate that thread toward the
4680 * processor it's on.
4682 * @preempt: whether task preemption is allowed or not
4684 * It's the caller's job to ensure that the target task struct
4685 * can't go away on us before we can do any checks.
4688 * true (>0) if we indeed boosted the target task.
4689 * false (0) if we failed to boost the target.
4690 * -ESRCH if there's no task to yield to.
4692 int __sched yield_to(struct task_struct *p, bool preempt)
4694 struct task_struct *curr = current;
4695 struct rq *rq, *p_rq;
4696 unsigned long flags;
4699 local_irq_save(flags);
4705 * If we're the only runnable task on the rq and target rq also
4706 * has only one task, there's absolutely no point in yielding.
4708 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4713 double_rq_lock(rq, p_rq);
4714 if (task_rq(p) != p_rq) {
4715 double_rq_unlock(rq, p_rq);
4719 if (!curr->sched_class->yield_to_task)
4722 if (curr->sched_class != p->sched_class)
4725 if (task_running(p_rq, p) || p->state)
4728 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4730 schedstat_inc(rq, yld_count);
4732 * Make p's CPU reschedule; pick_next_entity takes care of
4735 if (preempt && rq != p_rq)
4740 double_rq_unlock(rq, p_rq);
4742 local_irq_restore(flags);
4749 EXPORT_SYMBOL_GPL(yield_to);
4752 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4753 * that process accounting knows that this is a task in IO wait state.
4755 long __sched io_schedule_timeout(long timeout)
4757 int old_iowait = current->in_iowait;
4761 current->in_iowait = 1;
4762 blk_schedule_flush_plug(current);
4764 delayacct_blkio_start();
4766 atomic_inc(&rq->nr_iowait);
4767 ret = schedule_timeout(timeout);
4768 current->in_iowait = old_iowait;
4769 atomic_dec(&rq->nr_iowait);
4770 delayacct_blkio_end();
4774 EXPORT_SYMBOL(io_schedule_timeout);
4777 * sys_sched_get_priority_max - return maximum RT priority.
4778 * @policy: scheduling class.
4780 * Return: On success, this syscall returns the maximum
4781 * rt_priority that can be used by a given scheduling class.
4782 * On failure, a negative error code is returned.
4784 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4791 ret = MAX_USER_RT_PRIO-1;
4793 case SCHED_DEADLINE:
4804 * sys_sched_get_priority_min - return minimum RT priority.
4805 * @policy: scheduling class.
4807 * Return: On success, this syscall returns the minimum
4808 * rt_priority that can be used by a given scheduling class.
4809 * On failure, a negative error code is returned.
4811 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4820 case SCHED_DEADLINE:
4830 * sys_sched_rr_get_interval - return the default timeslice of a process.
4831 * @pid: pid of the process.
4832 * @interval: userspace pointer to the timeslice value.
4834 * this syscall writes the default timeslice value of a given process
4835 * into the user-space timespec buffer. A value of '0' means infinity.
4837 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4840 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4841 struct timespec __user *, interval)
4843 struct task_struct *p;
4844 unsigned int time_slice;
4845 unsigned long flags;
4855 p = find_process_by_pid(pid);
4859 retval = security_task_getscheduler(p);
4863 rq = task_rq_lock(p, &flags);
4865 if (p->sched_class->get_rr_interval)
4866 time_slice = p->sched_class->get_rr_interval(rq, p);
4867 task_rq_unlock(rq, p, &flags);
4870 jiffies_to_timespec(time_slice, &t);
4871 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4879 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4881 void sched_show_task(struct task_struct *p)
4883 unsigned long free = 0;
4885 unsigned long state = p->state;
4888 state = __ffs(state) + 1;
4889 printk(KERN_INFO "%-15.15s %c", p->comm,
4890 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4891 #if BITS_PER_LONG == 32
4892 if (state == TASK_RUNNING)
4893 printk(KERN_CONT " running ");
4895 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4897 if (state == TASK_RUNNING)
4898 printk(KERN_CONT " running task ");
4900 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4902 #ifdef CONFIG_DEBUG_STACK_USAGE
4903 free = stack_not_used(p);
4908 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4910 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4911 task_pid_nr(p), ppid,
4912 (unsigned long)task_thread_info(p)->flags);
4914 print_worker_info(KERN_INFO, p);
4915 show_stack(p, NULL);
4918 void show_state_filter(unsigned long state_filter)
4920 struct task_struct *g, *p;
4922 #if BITS_PER_LONG == 32
4924 " task PC stack pid father\n");
4927 " task PC stack pid father\n");
4930 for_each_process_thread(g, p) {
4932 * reset the NMI-timeout, listing all files on a slow
4933 * console might take a lot of time:
4935 touch_nmi_watchdog();
4936 if (!state_filter || (p->state & state_filter))
4940 touch_all_softlockup_watchdogs();
4942 #ifdef CONFIG_SCHED_DEBUG
4943 sysrq_sched_debug_show();
4947 * Only show locks if all tasks are dumped:
4950 debug_show_all_locks();
4953 void init_idle_bootup_task(struct task_struct *idle)
4955 idle->sched_class = &idle_sched_class;
4959 * init_idle - set up an idle thread for a given CPU
4960 * @idle: task in question
4961 * @cpu: cpu the idle task belongs to
4963 * NOTE: this function does not set the idle thread's NEED_RESCHED
4964 * flag, to make booting more robust.
4966 void init_idle(struct task_struct *idle, int cpu)
4968 struct rq *rq = cpu_rq(cpu);
4969 unsigned long flags;
4971 raw_spin_lock_irqsave(&idle->pi_lock, flags);
4972 raw_spin_lock(&rq->lock);
4974 __sched_fork(0, idle);
4975 idle->state = TASK_RUNNING;
4976 idle->se.exec_start = sched_clock();
4980 * Its possible that init_idle() gets called multiple times on a task,
4981 * in that case do_set_cpus_allowed() will not do the right thing.
4983 * And since this is boot we can forgo the serialization.
4985 set_cpus_allowed_common(idle, cpumask_of(cpu));
4988 * We're having a chicken and egg problem, even though we are
4989 * holding rq->lock, the cpu isn't yet set to this cpu so the
4990 * lockdep check in task_group() will fail.
4992 * Similar case to sched_fork(). / Alternatively we could
4993 * use task_rq_lock() here and obtain the other rq->lock.
4998 __set_task_cpu(idle, cpu);
5001 rq->curr = rq->idle = idle;
5002 idle->on_rq = TASK_ON_RQ_QUEUED;
5006 raw_spin_unlock(&rq->lock);
5007 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5009 /* Set the preempt count _outside_ the spinlocks! */
5010 init_idle_preempt_count(idle, cpu);
5013 * The idle tasks have their own, simple scheduling class:
5015 idle->sched_class = &idle_sched_class;
5016 ftrace_graph_init_idle_task(idle, cpu);
5017 vtime_init_idle(idle, cpu);
5019 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5023 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5024 const struct cpumask *trial)
5026 int ret = 1, trial_cpus;
5027 struct dl_bw *cur_dl_b;
5028 unsigned long flags;
5030 if (!cpumask_weight(cur))
5033 rcu_read_lock_sched();
5034 cur_dl_b = dl_bw_of(cpumask_any(cur));
5035 trial_cpus = cpumask_weight(trial);
5037 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5038 if (cur_dl_b->bw != -1 &&
5039 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5041 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5042 rcu_read_unlock_sched();
5047 int task_can_attach(struct task_struct *p,
5048 const struct cpumask *cs_cpus_allowed)
5053 * Kthreads which disallow setaffinity shouldn't be moved
5054 * to a new cpuset; we don't want to change their cpu
5055 * affinity and isolating such threads by their set of
5056 * allowed nodes is unnecessary. Thus, cpusets are not
5057 * applicable for such threads. This prevents checking for
5058 * success of set_cpus_allowed_ptr() on all attached tasks
5059 * before cpus_allowed may be changed.
5061 if (p->flags & PF_NO_SETAFFINITY) {
5067 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5069 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5074 unsigned long flags;
5076 rcu_read_lock_sched();
5077 dl_b = dl_bw_of(dest_cpu);
5078 raw_spin_lock_irqsave(&dl_b->lock, flags);
5079 cpus = dl_bw_cpus(dest_cpu);
5080 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5085 * We reserve space for this task in the destination
5086 * root_domain, as we can't fail after this point.
5087 * We will free resources in the source root_domain
5088 * later on (see set_cpus_allowed_dl()).
5090 __dl_add(dl_b, p->dl.dl_bw);
5092 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5093 rcu_read_unlock_sched();
5103 #ifdef CONFIG_NUMA_BALANCING
5104 /* Migrate current task p to target_cpu */
5105 int migrate_task_to(struct task_struct *p, int target_cpu)
5107 struct migration_arg arg = { p, target_cpu };
5108 int curr_cpu = task_cpu(p);
5110 if (curr_cpu == target_cpu)
5113 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5116 /* TODO: This is not properly updating schedstats */
5118 trace_sched_move_numa(p, curr_cpu, target_cpu);
5119 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5123 * Requeue a task on a given node and accurately track the number of NUMA
5124 * tasks on the runqueues
5126 void sched_setnuma(struct task_struct *p, int nid)
5129 unsigned long flags;
5130 bool queued, running;
5132 rq = task_rq_lock(p, &flags);
5133 queued = task_on_rq_queued(p);
5134 running = task_current(rq, p);
5137 dequeue_task(rq, p, DEQUEUE_SAVE);
5139 put_prev_task(rq, p);
5141 p->numa_preferred_nid = nid;
5144 p->sched_class->set_curr_task(rq);
5146 enqueue_task(rq, p, ENQUEUE_RESTORE);
5147 task_rq_unlock(rq, p, &flags);
5149 #endif /* CONFIG_NUMA_BALANCING */
5151 #ifdef CONFIG_HOTPLUG_CPU
5153 * Ensures that the idle task is using init_mm right before its cpu goes
5156 void idle_task_exit(void)
5158 struct mm_struct *mm = current->active_mm;
5160 BUG_ON(cpu_online(smp_processor_id()));
5162 if (mm != &init_mm) {
5163 switch_mm(mm, &init_mm, current);
5164 finish_arch_post_lock_switch();
5170 * Since this CPU is going 'away' for a while, fold any nr_active delta
5171 * we might have. Assumes we're called after migrate_tasks() so that the
5172 * nr_active count is stable.
5174 * Also see the comment "Global load-average calculations".
5176 static void calc_load_migrate(struct rq *rq)
5178 long delta = calc_load_fold_active(rq);
5180 atomic_long_add(delta, &calc_load_tasks);
5183 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5187 static const struct sched_class fake_sched_class = {
5188 .put_prev_task = put_prev_task_fake,
5191 static struct task_struct fake_task = {
5193 * Avoid pull_{rt,dl}_task()
5195 .prio = MAX_PRIO + 1,
5196 .sched_class = &fake_sched_class,
5200 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5201 * try_to_wake_up()->select_task_rq().
5203 * Called with rq->lock held even though we'er in stop_machine() and
5204 * there's no concurrency possible, we hold the required locks anyway
5205 * because of lock validation efforts.
5207 static void migrate_tasks(struct rq *dead_rq)
5209 struct rq *rq = dead_rq;
5210 struct task_struct *next, *stop = rq->stop;
5214 * Fudge the rq selection such that the below task selection loop
5215 * doesn't get stuck on the currently eligible stop task.
5217 * We're currently inside stop_machine() and the rq is either stuck
5218 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5219 * either way we should never end up calling schedule() until we're
5225 * put_prev_task() and pick_next_task() sched
5226 * class method both need to have an up-to-date
5227 * value of rq->clock[_task]
5229 update_rq_clock(rq);
5233 * There's this thread running, bail when that's the only
5236 if (rq->nr_running == 1)
5240 * pick_next_task assumes pinned rq->lock.
5242 lockdep_pin_lock(&rq->lock);
5243 next = pick_next_task(rq, &fake_task);
5245 next->sched_class->put_prev_task(rq, next);
5248 * Rules for changing task_struct::cpus_allowed are holding
5249 * both pi_lock and rq->lock, such that holding either
5250 * stabilizes the mask.
5252 * Drop rq->lock is not quite as disastrous as it usually is
5253 * because !cpu_active at this point, which means load-balance
5254 * will not interfere. Also, stop-machine.
5256 lockdep_unpin_lock(&rq->lock);
5257 raw_spin_unlock(&rq->lock);
5258 raw_spin_lock(&next->pi_lock);
5259 raw_spin_lock(&rq->lock);
5262 * Since we're inside stop-machine, _nothing_ should have
5263 * changed the task, WARN if weird stuff happened, because in
5264 * that case the above rq->lock drop is a fail too.
5266 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5267 raw_spin_unlock(&next->pi_lock);
5271 /* Find suitable destination for @next, with force if needed. */
5272 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5274 rq = __migrate_task(rq, next, dest_cpu);
5275 if (rq != dead_rq) {
5276 raw_spin_unlock(&rq->lock);
5278 raw_spin_lock(&rq->lock);
5280 raw_spin_unlock(&next->pi_lock);
5285 #endif /* CONFIG_HOTPLUG_CPU */
5287 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5289 static struct ctl_table sd_ctl_dir[] = {
5291 .procname = "sched_domain",
5297 static struct ctl_table sd_ctl_root[] = {
5299 .procname = "kernel",
5301 .child = sd_ctl_dir,
5306 static struct ctl_table *sd_alloc_ctl_entry(int n)
5308 struct ctl_table *entry =
5309 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5314 static void sd_free_ctl_entry(struct ctl_table **tablep)
5316 struct ctl_table *entry;
5319 * In the intermediate directories, both the child directory and
5320 * procname are dynamically allocated and could fail but the mode
5321 * will always be set. In the lowest directory the names are
5322 * static strings and all have proc handlers.
5324 for (entry = *tablep; entry->mode; entry++) {
5326 sd_free_ctl_entry(&entry->child);
5327 if (entry->proc_handler == NULL)
5328 kfree(entry->procname);
5335 static int min_load_idx = 0;
5336 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5339 set_table_entry(struct ctl_table *entry,
5340 const char *procname, void *data, int maxlen,
5341 umode_t mode, proc_handler *proc_handler,
5344 entry->procname = procname;
5346 entry->maxlen = maxlen;
5348 entry->proc_handler = proc_handler;
5351 entry->extra1 = &min_load_idx;
5352 entry->extra2 = &max_load_idx;
5356 static struct ctl_table *
5357 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5359 struct ctl_table *table = sd_alloc_ctl_entry(14);
5364 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5365 sizeof(long), 0644, proc_doulongvec_minmax, false);
5366 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5367 sizeof(long), 0644, proc_doulongvec_minmax, false);
5368 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5369 sizeof(int), 0644, proc_dointvec_minmax, true);
5370 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5371 sizeof(int), 0644, proc_dointvec_minmax, true);
5372 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5373 sizeof(int), 0644, proc_dointvec_minmax, true);
5374 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5375 sizeof(int), 0644, proc_dointvec_minmax, true);
5376 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5377 sizeof(int), 0644, proc_dointvec_minmax, true);
5378 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5379 sizeof(int), 0644, proc_dointvec_minmax, false);
5380 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5381 sizeof(int), 0644, proc_dointvec_minmax, false);
5382 set_table_entry(&table[9], "cache_nice_tries",
5383 &sd->cache_nice_tries,
5384 sizeof(int), 0644, proc_dointvec_minmax, false);
5385 set_table_entry(&table[10], "flags", &sd->flags,
5386 sizeof(int), 0644, proc_dointvec_minmax, false);
5387 set_table_entry(&table[11], "max_newidle_lb_cost",
5388 &sd->max_newidle_lb_cost,
5389 sizeof(long), 0644, proc_doulongvec_minmax, false);
5390 set_table_entry(&table[12], "name", sd->name,
5391 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5392 /* &table[13] is terminator */
5397 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5399 struct ctl_table *entry, *table;
5400 struct sched_domain *sd;
5401 int domain_num = 0, i;
5404 for_each_domain(cpu, sd)
5406 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5411 for_each_domain(cpu, sd) {
5412 snprintf(buf, 32, "domain%d", i);
5413 entry->procname = kstrdup(buf, GFP_KERNEL);
5415 entry->child = sd_alloc_ctl_domain_table(sd);
5422 static struct ctl_table_header *sd_sysctl_header;
5423 static void register_sched_domain_sysctl(void)
5425 int i, cpu_num = num_possible_cpus();
5426 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5429 WARN_ON(sd_ctl_dir[0].child);
5430 sd_ctl_dir[0].child = entry;
5435 for_each_possible_cpu(i) {
5436 snprintf(buf, 32, "cpu%d", i);
5437 entry->procname = kstrdup(buf, GFP_KERNEL);
5439 entry->child = sd_alloc_ctl_cpu_table(i);
5443 WARN_ON(sd_sysctl_header);
5444 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5447 /* may be called multiple times per register */
5448 static void unregister_sched_domain_sysctl(void)
5450 unregister_sysctl_table(sd_sysctl_header);
5451 sd_sysctl_header = NULL;
5452 if (sd_ctl_dir[0].child)
5453 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5456 static void register_sched_domain_sysctl(void)
5459 static void unregister_sched_domain_sysctl(void)
5462 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5464 static void set_rq_online(struct rq *rq)
5467 const struct sched_class *class;
5469 cpumask_set_cpu(rq->cpu, rq->rd->online);
5472 for_each_class(class) {
5473 if (class->rq_online)
5474 class->rq_online(rq);
5479 static void set_rq_offline(struct rq *rq)
5482 const struct sched_class *class;
5484 for_each_class(class) {
5485 if (class->rq_offline)
5486 class->rq_offline(rq);
5489 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5495 * migration_call - callback that gets triggered when a CPU is added.
5496 * Here we can start up the necessary migration thread for the new CPU.
5499 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5501 int cpu = (long)hcpu;
5502 unsigned long flags;
5503 struct rq *rq = cpu_rq(cpu);
5505 switch (action & ~CPU_TASKS_FROZEN) {
5507 case CPU_UP_PREPARE:
5508 rq->calc_load_update = calc_load_update;
5512 /* Update our root-domain */
5513 raw_spin_lock_irqsave(&rq->lock, flags);
5515 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5519 raw_spin_unlock_irqrestore(&rq->lock, flags);
5522 #ifdef CONFIG_HOTPLUG_CPU
5524 sched_ttwu_pending();
5525 /* Update our root-domain */
5526 raw_spin_lock_irqsave(&rq->lock, flags);
5528 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5532 BUG_ON(rq->nr_running != 1); /* the migration thread */
5533 raw_spin_unlock_irqrestore(&rq->lock, flags);
5537 calc_load_migrate(rq);
5542 update_max_interval();
5548 * Register at high priority so that task migration (migrate_all_tasks)
5549 * happens before everything else. This has to be lower priority than
5550 * the notifier in the perf_event subsystem, though.
5552 static struct notifier_block migration_notifier = {
5553 .notifier_call = migration_call,
5554 .priority = CPU_PRI_MIGRATION,
5557 static void set_cpu_rq_start_time(void)
5559 int cpu = smp_processor_id();
5560 struct rq *rq = cpu_rq(cpu);
5561 rq->age_stamp = sched_clock_cpu(cpu);
5564 static int sched_cpu_active(struct notifier_block *nfb,
5565 unsigned long action, void *hcpu)
5567 int cpu = (long)hcpu;
5569 switch (action & ~CPU_TASKS_FROZEN) {
5571 set_cpu_rq_start_time();
5576 * At this point a starting CPU has marked itself as online via
5577 * set_cpu_online(). But it might not yet have marked itself
5578 * as active, which is essential from here on.
5580 set_cpu_active(cpu, true);
5581 stop_machine_unpark(cpu);
5584 case CPU_DOWN_FAILED:
5585 set_cpu_active(cpu, true);
5593 static int sched_cpu_inactive(struct notifier_block *nfb,
5594 unsigned long action, void *hcpu)
5596 switch (action & ~CPU_TASKS_FROZEN) {
5597 case CPU_DOWN_PREPARE:
5598 set_cpu_active((long)hcpu, false);
5605 static int __init migration_init(void)
5607 void *cpu = (void *)(long)smp_processor_id();
5610 /* Initialize migration for the boot CPU */
5611 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5612 BUG_ON(err == NOTIFY_BAD);
5613 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5614 register_cpu_notifier(&migration_notifier);
5616 /* Register cpu active notifiers */
5617 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5618 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5622 early_initcall(migration_init);
5624 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5626 #ifdef CONFIG_SCHED_DEBUG
5628 static __read_mostly int sched_debug_enabled;
5630 static int __init sched_debug_setup(char *str)
5632 sched_debug_enabled = 1;
5636 early_param("sched_debug", sched_debug_setup);
5638 static inline bool sched_debug(void)
5640 return sched_debug_enabled;
5643 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5644 struct cpumask *groupmask)
5646 struct sched_group *group = sd->groups;
5648 cpumask_clear(groupmask);
5650 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5652 if (!(sd->flags & SD_LOAD_BALANCE)) {
5653 printk("does not load-balance\n");
5655 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5660 printk(KERN_CONT "span %*pbl level %s\n",
5661 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5663 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5664 printk(KERN_ERR "ERROR: domain->span does not contain "
5667 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5668 printk(KERN_ERR "ERROR: domain->groups does not contain"
5672 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5676 printk(KERN_ERR "ERROR: group is NULL\n");
5680 if (!cpumask_weight(sched_group_cpus(group))) {
5681 printk(KERN_CONT "\n");
5682 printk(KERN_ERR "ERROR: empty group\n");
5686 if (!(sd->flags & SD_OVERLAP) &&
5687 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5688 printk(KERN_CONT "\n");
5689 printk(KERN_ERR "ERROR: repeated CPUs\n");
5693 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5695 printk(KERN_CONT " %*pbl",
5696 cpumask_pr_args(sched_group_cpus(group)));
5697 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5698 printk(KERN_CONT " (cpu_capacity = %d)",
5699 group->sgc->capacity);
5702 group = group->next;
5703 } while (group != sd->groups);
5704 printk(KERN_CONT "\n");
5706 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5707 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5710 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5711 printk(KERN_ERR "ERROR: parent span is not a superset "
5712 "of domain->span\n");
5716 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5720 if (!sched_debug_enabled)
5724 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5728 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5731 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5739 #else /* !CONFIG_SCHED_DEBUG */
5740 # define sched_domain_debug(sd, cpu) do { } while (0)
5741 static inline bool sched_debug(void)
5745 #endif /* CONFIG_SCHED_DEBUG */
5747 static int sd_degenerate(struct sched_domain *sd)
5749 if (cpumask_weight(sched_domain_span(sd)) == 1)
5752 /* Following flags need at least 2 groups */
5753 if (sd->flags & (SD_LOAD_BALANCE |
5754 SD_BALANCE_NEWIDLE |
5757 SD_SHARE_CPUCAPACITY |
5758 SD_SHARE_PKG_RESOURCES |
5759 SD_SHARE_POWERDOMAIN)) {
5760 if (sd->groups != sd->groups->next)
5764 /* Following flags don't use groups */
5765 if (sd->flags & (SD_WAKE_AFFINE))
5772 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5774 unsigned long cflags = sd->flags, pflags = parent->flags;
5776 if (sd_degenerate(parent))
5779 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5782 /* Flags needing groups don't count if only 1 group in parent */
5783 if (parent->groups == parent->groups->next) {
5784 pflags &= ~(SD_LOAD_BALANCE |
5785 SD_BALANCE_NEWIDLE |
5788 SD_SHARE_CPUCAPACITY |
5789 SD_SHARE_PKG_RESOURCES |
5791 SD_SHARE_POWERDOMAIN);
5792 if (nr_node_ids == 1)
5793 pflags &= ~SD_SERIALIZE;
5795 if (~cflags & pflags)
5801 static void free_rootdomain(struct rcu_head *rcu)
5803 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5805 cpupri_cleanup(&rd->cpupri);
5806 cpudl_cleanup(&rd->cpudl);
5807 free_cpumask_var(rd->dlo_mask);
5808 free_cpumask_var(rd->rto_mask);
5809 free_cpumask_var(rd->online);
5810 free_cpumask_var(rd->span);
5814 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5816 struct root_domain *old_rd = NULL;
5817 unsigned long flags;
5819 raw_spin_lock_irqsave(&rq->lock, flags);
5824 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5827 cpumask_clear_cpu(rq->cpu, old_rd->span);
5830 * If we dont want to free the old_rd yet then
5831 * set old_rd to NULL to skip the freeing later
5834 if (!atomic_dec_and_test(&old_rd->refcount))
5838 atomic_inc(&rd->refcount);
5841 cpumask_set_cpu(rq->cpu, rd->span);
5842 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5845 raw_spin_unlock_irqrestore(&rq->lock, flags);
5848 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5851 static int init_rootdomain(struct root_domain *rd)
5853 memset(rd, 0, sizeof(*rd));
5855 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5857 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5859 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5861 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5864 init_dl_bw(&rd->dl_bw);
5865 if (cpudl_init(&rd->cpudl) != 0)
5868 if (cpupri_init(&rd->cpupri) != 0)
5873 free_cpumask_var(rd->rto_mask);
5875 free_cpumask_var(rd->dlo_mask);
5877 free_cpumask_var(rd->online);
5879 free_cpumask_var(rd->span);
5885 * By default the system creates a single root-domain with all cpus as
5886 * members (mimicking the global state we have today).
5888 struct root_domain def_root_domain;
5890 static void init_defrootdomain(void)
5892 init_rootdomain(&def_root_domain);
5894 atomic_set(&def_root_domain.refcount, 1);
5897 static struct root_domain *alloc_rootdomain(void)
5899 struct root_domain *rd;
5901 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5905 if (init_rootdomain(rd) != 0) {
5913 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5915 struct sched_group *tmp, *first;
5924 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5929 } while (sg != first);
5932 static void free_sched_domain(struct rcu_head *rcu)
5934 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5937 * If its an overlapping domain it has private groups, iterate and
5940 if (sd->flags & SD_OVERLAP) {
5941 free_sched_groups(sd->groups, 1);
5942 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5943 kfree(sd->groups->sgc);
5949 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5951 call_rcu(&sd->rcu, free_sched_domain);
5954 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5956 for (; sd; sd = sd->parent)
5957 destroy_sched_domain(sd, cpu);
5961 * Keep a special pointer to the highest sched_domain that has
5962 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5963 * allows us to avoid some pointer chasing select_idle_sibling().
5965 * Also keep a unique ID per domain (we use the first cpu number in
5966 * the cpumask of the domain), this allows us to quickly tell if
5967 * two cpus are in the same cache domain, see cpus_share_cache().
5969 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5970 DEFINE_PER_CPU(int, sd_llc_size);
5971 DEFINE_PER_CPU(int, sd_llc_id);
5972 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5973 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5974 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5976 static void update_top_cache_domain(int cpu)
5978 struct sched_domain *sd;
5979 struct sched_domain *busy_sd = NULL;
5983 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5985 id = cpumask_first(sched_domain_span(sd));
5986 size = cpumask_weight(sched_domain_span(sd));
5987 busy_sd = sd->parent; /* sd_busy */
5989 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5991 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5992 per_cpu(sd_llc_size, cpu) = size;
5993 per_cpu(sd_llc_id, cpu) = id;
5995 sd = lowest_flag_domain(cpu, SD_NUMA);
5996 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5998 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5999 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6003 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6004 * hold the hotplug lock.
6007 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6009 struct rq *rq = cpu_rq(cpu);
6010 struct sched_domain *tmp;
6012 /* Remove the sched domains which do not contribute to scheduling. */
6013 for (tmp = sd; tmp; ) {
6014 struct sched_domain *parent = tmp->parent;
6018 if (sd_parent_degenerate(tmp, parent)) {
6019 tmp->parent = parent->parent;
6021 parent->parent->child = tmp;
6023 * Transfer SD_PREFER_SIBLING down in case of a
6024 * degenerate parent; the spans match for this
6025 * so the property transfers.
6027 if (parent->flags & SD_PREFER_SIBLING)
6028 tmp->flags |= SD_PREFER_SIBLING;
6029 destroy_sched_domain(parent, cpu);
6034 if (sd && sd_degenerate(sd)) {
6037 destroy_sched_domain(tmp, cpu);
6042 sched_domain_debug(sd, cpu);
6044 rq_attach_root(rq, rd);
6046 rcu_assign_pointer(rq->sd, sd);
6047 destroy_sched_domains(tmp, cpu);
6049 update_top_cache_domain(cpu);
6052 /* Setup the mask of cpus configured for isolated domains */
6053 static int __init isolated_cpu_setup(char *str)
6055 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6056 cpulist_parse(str, cpu_isolated_map);
6060 __setup("isolcpus=", isolated_cpu_setup);
6063 struct sched_domain ** __percpu sd;
6064 struct root_domain *rd;
6075 * Build an iteration mask that can exclude certain CPUs from the upwards
6078 * Asymmetric node setups can result in situations where the domain tree is of
6079 * unequal depth, make sure to skip domains that already cover the entire
6082 * In that case build_sched_domains() will have terminated the iteration early
6083 * and our sibling sd spans will be empty. Domains should always include the
6084 * cpu they're built on, so check that.
6087 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6089 const struct cpumask *span = sched_domain_span(sd);
6090 struct sd_data *sdd = sd->private;
6091 struct sched_domain *sibling;
6094 for_each_cpu(i, span) {
6095 sibling = *per_cpu_ptr(sdd->sd, i);
6096 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6099 cpumask_set_cpu(i, sched_group_mask(sg));
6104 * Return the canonical balance cpu for this group, this is the first cpu
6105 * of this group that's also in the iteration mask.
6107 int group_balance_cpu(struct sched_group *sg)
6109 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6113 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6115 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6116 const struct cpumask *span = sched_domain_span(sd);
6117 struct cpumask *covered = sched_domains_tmpmask;
6118 struct sd_data *sdd = sd->private;
6119 struct sched_domain *sibling;
6122 cpumask_clear(covered);
6124 for_each_cpu(i, span) {
6125 struct cpumask *sg_span;
6127 if (cpumask_test_cpu(i, covered))
6130 sibling = *per_cpu_ptr(sdd->sd, i);
6132 /* See the comment near build_group_mask(). */
6133 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6136 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6137 GFP_KERNEL, cpu_to_node(cpu));
6142 sg_span = sched_group_cpus(sg);
6144 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6146 cpumask_set_cpu(i, sg_span);
6148 cpumask_or(covered, covered, sg_span);
6150 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6151 if (atomic_inc_return(&sg->sgc->ref) == 1)
6152 build_group_mask(sd, sg);
6155 * Initialize sgc->capacity such that even if we mess up the
6156 * domains and no possible iteration will get us here, we won't
6159 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6162 * Make sure the first group of this domain contains the
6163 * canonical balance cpu. Otherwise the sched_domain iteration
6164 * breaks. See update_sg_lb_stats().
6166 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6167 group_balance_cpu(sg) == cpu)
6177 sd->groups = groups;
6182 free_sched_groups(first, 0);
6187 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6189 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6190 struct sched_domain *child = sd->child;
6193 cpu = cpumask_first(sched_domain_span(child));
6196 *sg = *per_cpu_ptr(sdd->sg, cpu);
6197 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6198 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6205 * build_sched_groups will build a circular linked list of the groups
6206 * covered by the given span, and will set each group's ->cpumask correctly,
6207 * and ->cpu_capacity to 0.
6209 * Assumes the sched_domain tree is fully constructed
6212 build_sched_groups(struct sched_domain *sd, int cpu)
6214 struct sched_group *first = NULL, *last = NULL;
6215 struct sd_data *sdd = sd->private;
6216 const struct cpumask *span = sched_domain_span(sd);
6217 struct cpumask *covered;
6220 get_group(cpu, sdd, &sd->groups);
6221 atomic_inc(&sd->groups->ref);
6223 if (cpu != cpumask_first(span))
6226 lockdep_assert_held(&sched_domains_mutex);
6227 covered = sched_domains_tmpmask;
6229 cpumask_clear(covered);
6231 for_each_cpu(i, span) {
6232 struct sched_group *sg;
6235 if (cpumask_test_cpu(i, covered))
6238 group = get_group(i, sdd, &sg);
6239 cpumask_setall(sched_group_mask(sg));
6241 for_each_cpu(j, span) {
6242 if (get_group(j, sdd, NULL) != group)
6245 cpumask_set_cpu(j, covered);
6246 cpumask_set_cpu(j, sched_group_cpus(sg));
6261 * Initialize sched groups cpu_capacity.
6263 * cpu_capacity indicates the capacity of sched group, which is used while
6264 * distributing the load between different sched groups in a sched domain.
6265 * Typically cpu_capacity for all the groups in a sched domain will be same
6266 * unless there are asymmetries in the topology. If there are asymmetries,
6267 * group having more cpu_capacity will pickup more load compared to the
6268 * group having less cpu_capacity.
6270 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6272 struct sched_group *sg = sd->groups;
6277 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6279 } while (sg != sd->groups);
6281 if (cpu != group_balance_cpu(sg))
6284 update_group_capacity(sd, cpu);
6285 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6289 * Initializers for schedule domains
6290 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6293 static int default_relax_domain_level = -1;
6294 int sched_domain_level_max;
6296 static int __init setup_relax_domain_level(char *str)
6298 if (kstrtoint(str, 0, &default_relax_domain_level))
6299 pr_warn("Unable to set relax_domain_level\n");
6303 __setup("relax_domain_level=", setup_relax_domain_level);
6305 static void set_domain_attribute(struct sched_domain *sd,
6306 struct sched_domain_attr *attr)
6310 if (!attr || attr->relax_domain_level < 0) {
6311 if (default_relax_domain_level < 0)
6314 request = default_relax_domain_level;
6316 request = attr->relax_domain_level;
6317 if (request < sd->level) {
6318 /* turn off idle balance on this domain */
6319 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6321 /* turn on idle balance on this domain */
6322 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6326 static void __sdt_free(const struct cpumask *cpu_map);
6327 static int __sdt_alloc(const struct cpumask *cpu_map);
6329 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6330 const struct cpumask *cpu_map)
6334 if (!atomic_read(&d->rd->refcount))
6335 free_rootdomain(&d->rd->rcu); /* fall through */
6337 free_percpu(d->sd); /* fall through */
6339 __sdt_free(cpu_map); /* fall through */
6345 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6346 const struct cpumask *cpu_map)
6348 memset(d, 0, sizeof(*d));
6350 if (__sdt_alloc(cpu_map))
6351 return sa_sd_storage;
6352 d->sd = alloc_percpu(struct sched_domain *);
6354 return sa_sd_storage;
6355 d->rd = alloc_rootdomain();
6358 return sa_rootdomain;
6362 * NULL the sd_data elements we've used to build the sched_domain and
6363 * sched_group structure so that the subsequent __free_domain_allocs()
6364 * will not free the data we're using.
6366 static void claim_allocations(int cpu, struct sched_domain *sd)
6368 struct sd_data *sdd = sd->private;
6370 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6371 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6373 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6374 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6376 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6377 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6381 static int sched_domains_numa_levels;
6382 enum numa_topology_type sched_numa_topology_type;
6383 static int *sched_domains_numa_distance;
6384 int sched_max_numa_distance;
6385 static struct cpumask ***sched_domains_numa_masks;
6386 static int sched_domains_curr_level;
6390 * SD_flags allowed in topology descriptions.
6392 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6393 * SD_SHARE_PKG_RESOURCES - describes shared caches
6394 * SD_NUMA - describes NUMA topologies
6395 * SD_SHARE_POWERDOMAIN - describes shared power domain
6398 * SD_ASYM_PACKING - describes SMT quirks
6400 #define TOPOLOGY_SD_FLAGS \
6401 (SD_SHARE_CPUCAPACITY | \
6402 SD_SHARE_PKG_RESOURCES | \
6405 SD_SHARE_POWERDOMAIN)
6407 static struct sched_domain *
6408 sd_init(struct sched_domain_topology_level *tl, int cpu)
6410 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6411 int sd_weight, sd_flags = 0;
6415 * Ugly hack to pass state to sd_numa_mask()...
6417 sched_domains_curr_level = tl->numa_level;
6420 sd_weight = cpumask_weight(tl->mask(cpu));
6423 sd_flags = (*tl->sd_flags)();
6424 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6425 "wrong sd_flags in topology description\n"))
6426 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6428 *sd = (struct sched_domain){
6429 .min_interval = sd_weight,
6430 .max_interval = 2*sd_weight,
6432 .imbalance_pct = 125,
6434 .cache_nice_tries = 0,
6441 .flags = 1*SD_LOAD_BALANCE
6442 | 1*SD_BALANCE_NEWIDLE
6447 | 0*SD_SHARE_CPUCAPACITY
6448 | 0*SD_SHARE_PKG_RESOURCES
6450 | 0*SD_PREFER_SIBLING
6455 .last_balance = jiffies,
6456 .balance_interval = sd_weight,
6458 .max_newidle_lb_cost = 0,
6459 .next_decay_max_lb_cost = jiffies,
6460 #ifdef CONFIG_SCHED_DEBUG
6466 * Convert topological properties into behaviour.
6469 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6470 sd->flags |= SD_PREFER_SIBLING;
6471 sd->imbalance_pct = 110;
6472 sd->smt_gain = 1178; /* ~15% */
6474 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6475 sd->imbalance_pct = 117;
6476 sd->cache_nice_tries = 1;
6480 } else if (sd->flags & SD_NUMA) {
6481 sd->cache_nice_tries = 2;
6485 sd->flags |= SD_SERIALIZE;
6486 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6487 sd->flags &= ~(SD_BALANCE_EXEC |
6494 sd->flags |= SD_PREFER_SIBLING;
6495 sd->cache_nice_tries = 1;
6500 sd->private = &tl->data;
6506 * Topology list, bottom-up.
6508 static struct sched_domain_topology_level default_topology[] = {
6509 #ifdef CONFIG_SCHED_SMT
6510 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6512 #ifdef CONFIG_SCHED_MC
6513 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6515 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6519 static struct sched_domain_topology_level *sched_domain_topology =
6522 #define for_each_sd_topology(tl) \
6523 for (tl = sched_domain_topology; tl->mask; tl++)
6525 void set_sched_topology(struct sched_domain_topology_level *tl)
6527 sched_domain_topology = tl;
6532 static const struct cpumask *sd_numa_mask(int cpu)
6534 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6537 static void sched_numa_warn(const char *str)
6539 static int done = false;
6547 printk(KERN_WARNING "ERROR: %s\n\n", str);
6549 for (i = 0; i < nr_node_ids; i++) {
6550 printk(KERN_WARNING " ");
6551 for (j = 0; j < nr_node_ids; j++)
6552 printk(KERN_CONT "%02d ", node_distance(i,j));
6553 printk(KERN_CONT "\n");
6555 printk(KERN_WARNING "\n");
6558 bool find_numa_distance(int distance)
6562 if (distance == node_distance(0, 0))
6565 for (i = 0; i < sched_domains_numa_levels; i++) {
6566 if (sched_domains_numa_distance[i] == distance)
6574 * A system can have three types of NUMA topology:
6575 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6576 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6577 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6579 * The difference between a glueless mesh topology and a backplane
6580 * topology lies in whether communication between not directly
6581 * connected nodes goes through intermediary nodes (where programs
6582 * could run), or through backplane controllers. This affects
6583 * placement of programs.
6585 * The type of topology can be discerned with the following tests:
6586 * - If the maximum distance between any nodes is 1 hop, the system
6587 * is directly connected.
6588 * - If for two nodes A and B, located N > 1 hops away from each other,
6589 * there is an intermediary node C, which is < N hops away from both
6590 * nodes A and B, the system is a glueless mesh.
6592 static void init_numa_topology_type(void)
6596 n = sched_max_numa_distance;
6598 if (sched_domains_numa_levels <= 1) {
6599 sched_numa_topology_type = NUMA_DIRECT;
6603 for_each_online_node(a) {
6604 for_each_online_node(b) {
6605 /* Find two nodes furthest removed from each other. */
6606 if (node_distance(a, b) < n)
6609 /* Is there an intermediary node between a and b? */
6610 for_each_online_node(c) {
6611 if (node_distance(a, c) < n &&
6612 node_distance(b, c) < n) {
6613 sched_numa_topology_type =
6619 sched_numa_topology_type = NUMA_BACKPLANE;
6625 static void sched_init_numa(void)
6627 int next_distance, curr_distance = node_distance(0, 0);
6628 struct sched_domain_topology_level *tl;
6632 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6633 if (!sched_domains_numa_distance)
6637 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6638 * unique distances in the node_distance() table.
6640 * Assumes node_distance(0,j) includes all distances in
6641 * node_distance(i,j) in order to avoid cubic time.
6643 next_distance = curr_distance;
6644 for (i = 0; i < nr_node_ids; i++) {
6645 for (j = 0; j < nr_node_ids; j++) {
6646 for (k = 0; k < nr_node_ids; k++) {
6647 int distance = node_distance(i, k);
6649 if (distance > curr_distance &&
6650 (distance < next_distance ||
6651 next_distance == curr_distance))
6652 next_distance = distance;
6655 * While not a strong assumption it would be nice to know
6656 * about cases where if node A is connected to B, B is not
6657 * equally connected to A.
6659 if (sched_debug() && node_distance(k, i) != distance)
6660 sched_numa_warn("Node-distance not symmetric");
6662 if (sched_debug() && i && !find_numa_distance(distance))
6663 sched_numa_warn("Node-0 not representative");
6665 if (next_distance != curr_distance) {
6666 sched_domains_numa_distance[level++] = next_distance;
6667 sched_domains_numa_levels = level;
6668 curr_distance = next_distance;
6673 * In case of sched_debug() we verify the above assumption.
6683 * 'level' contains the number of unique distances, excluding the
6684 * identity distance node_distance(i,i).
6686 * The sched_domains_numa_distance[] array includes the actual distance
6691 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6692 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6693 * the array will contain less then 'level' members. This could be
6694 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6695 * in other functions.
6697 * We reset it to 'level' at the end of this function.
6699 sched_domains_numa_levels = 0;
6701 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6702 if (!sched_domains_numa_masks)
6706 * Now for each level, construct a mask per node which contains all
6707 * cpus of nodes that are that many hops away from us.
6709 for (i = 0; i < level; i++) {
6710 sched_domains_numa_masks[i] =
6711 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6712 if (!sched_domains_numa_masks[i])
6715 for (j = 0; j < nr_node_ids; j++) {
6716 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6720 sched_domains_numa_masks[i][j] = mask;
6722 for (k = 0; k < nr_node_ids; k++) {
6723 if (node_distance(j, k) > sched_domains_numa_distance[i])
6726 cpumask_or(mask, mask, cpumask_of_node(k));
6731 /* Compute default topology size */
6732 for (i = 0; sched_domain_topology[i].mask; i++);
6734 tl = kzalloc((i + level + 1) *
6735 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6740 * Copy the default topology bits..
6742 for (i = 0; sched_domain_topology[i].mask; i++)
6743 tl[i] = sched_domain_topology[i];
6746 * .. and append 'j' levels of NUMA goodness.
6748 for (j = 0; j < level; i++, j++) {
6749 tl[i] = (struct sched_domain_topology_level){
6750 .mask = sd_numa_mask,
6751 .sd_flags = cpu_numa_flags,
6752 .flags = SDTL_OVERLAP,
6758 sched_domain_topology = tl;
6760 sched_domains_numa_levels = level;
6761 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6763 init_numa_topology_type();
6766 static void sched_domains_numa_masks_set(int cpu)
6769 int node = cpu_to_node(cpu);
6771 for (i = 0; i < sched_domains_numa_levels; i++) {
6772 for (j = 0; j < nr_node_ids; j++) {
6773 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6774 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6779 static void sched_domains_numa_masks_clear(int cpu)
6782 for (i = 0; i < sched_domains_numa_levels; i++) {
6783 for (j = 0; j < nr_node_ids; j++)
6784 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6789 * Update sched_domains_numa_masks[level][node] array when new cpus
6792 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6793 unsigned long action,
6796 int cpu = (long)hcpu;
6798 switch (action & ~CPU_TASKS_FROZEN) {
6800 sched_domains_numa_masks_set(cpu);
6804 sched_domains_numa_masks_clear(cpu);
6814 static inline void sched_init_numa(void)
6818 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6819 unsigned long action,
6824 #endif /* CONFIG_NUMA */
6826 static int __sdt_alloc(const struct cpumask *cpu_map)
6828 struct sched_domain_topology_level *tl;
6831 for_each_sd_topology(tl) {
6832 struct sd_data *sdd = &tl->data;
6834 sdd->sd = alloc_percpu(struct sched_domain *);
6838 sdd->sg = alloc_percpu(struct sched_group *);
6842 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6846 for_each_cpu(j, cpu_map) {
6847 struct sched_domain *sd;
6848 struct sched_group *sg;
6849 struct sched_group_capacity *sgc;
6851 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6852 GFP_KERNEL, cpu_to_node(j));
6856 *per_cpu_ptr(sdd->sd, j) = sd;
6858 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6859 GFP_KERNEL, cpu_to_node(j));
6865 *per_cpu_ptr(sdd->sg, j) = sg;
6867 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6868 GFP_KERNEL, cpu_to_node(j));
6872 *per_cpu_ptr(sdd->sgc, j) = sgc;
6879 static void __sdt_free(const struct cpumask *cpu_map)
6881 struct sched_domain_topology_level *tl;
6884 for_each_sd_topology(tl) {
6885 struct sd_data *sdd = &tl->data;
6887 for_each_cpu(j, cpu_map) {
6888 struct sched_domain *sd;
6891 sd = *per_cpu_ptr(sdd->sd, j);
6892 if (sd && (sd->flags & SD_OVERLAP))
6893 free_sched_groups(sd->groups, 0);
6894 kfree(*per_cpu_ptr(sdd->sd, j));
6898 kfree(*per_cpu_ptr(sdd->sg, j));
6900 kfree(*per_cpu_ptr(sdd->sgc, j));
6902 free_percpu(sdd->sd);
6904 free_percpu(sdd->sg);
6906 free_percpu(sdd->sgc);
6911 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6912 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6913 struct sched_domain *child, int cpu)
6915 struct sched_domain *sd = sd_init(tl, cpu);
6919 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6921 sd->level = child->level + 1;
6922 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6926 if (!cpumask_subset(sched_domain_span(child),
6927 sched_domain_span(sd))) {
6928 pr_err("BUG: arch topology borken\n");
6929 #ifdef CONFIG_SCHED_DEBUG
6930 pr_err(" the %s domain not a subset of the %s domain\n",
6931 child->name, sd->name);
6933 /* Fixup, ensure @sd has at least @child cpus. */
6934 cpumask_or(sched_domain_span(sd),
6935 sched_domain_span(sd),
6936 sched_domain_span(child));
6940 set_domain_attribute(sd, attr);
6946 * Build sched domains for a given set of cpus and attach the sched domains
6947 * to the individual cpus
6949 static int build_sched_domains(const struct cpumask *cpu_map,
6950 struct sched_domain_attr *attr)
6952 enum s_alloc alloc_state;
6953 struct sched_domain *sd;
6955 int i, ret = -ENOMEM;
6957 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6958 if (alloc_state != sa_rootdomain)
6961 /* Set up domains for cpus specified by the cpu_map. */
6962 for_each_cpu(i, cpu_map) {
6963 struct sched_domain_topology_level *tl;
6966 for_each_sd_topology(tl) {
6967 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6968 if (tl == sched_domain_topology)
6969 *per_cpu_ptr(d.sd, i) = sd;
6970 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6971 sd->flags |= SD_OVERLAP;
6972 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6977 /* Build the groups for the domains */
6978 for_each_cpu(i, cpu_map) {
6979 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6980 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6981 if (sd->flags & SD_OVERLAP) {
6982 if (build_overlap_sched_groups(sd, i))
6985 if (build_sched_groups(sd, i))
6991 /* Calculate CPU capacity for physical packages and nodes */
6992 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6993 if (!cpumask_test_cpu(i, cpu_map))
6996 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6997 claim_allocations(i, sd);
6998 init_sched_groups_capacity(i, sd);
7002 /* Attach the domains */
7004 for_each_cpu(i, cpu_map) {
7005 sd = *per_cpu_ptr(d.sd, i);
7006 cpu_attach_domain(sd, d.rd, i);
7012 __free_domain_allocs(&d, alloc_state, cpu_map);
7016 static cpumask_var_t *doms_cur; /* current sched domains */
7017 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7018 static struct sched_domain_attr *dattr_cur;
7019 /* attribues of custom domains in 'doms_cur' */
7022 * Special case: If a kmalloc of a doms_cur partition (array of
7023 * cpumask) fails, then fallback to a single sched domain,
7024 * as determined by the single cpumask fallback_doms.
7026 static cpumask_var_t fallback_doms;
7029 * arch_update_cpu_topology lets virtualized architectures update the
7030 * cpu core maps. It is supposed to return 1 if the topology changed
7031 * or 0 if it stayed the same.
7033 int __weak arch_update_cpu_topology(void)
7038 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7041 cpumask_var_t *doms;
7043 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7046 for (i = 0; i < ndoms; i++) {
7047 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7048 free_sched_domains(doms, i);
7055 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7058 for (i = 0; i < ndoms; i++)
7059 free_cpumask_var(doms[i]);
7064 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7065 * For now this just excludes isolated cpus, but could be used to
7066 * exclude other special cases in the future.
7068 static int init_sched_domains(const struct cpumask *cpu_map)
7072 arch_update_cpu_topology();
7074 doms_cur = alloc_sched_domains(ndoms_cur);
7076 doms_cur = &fallback_doms;
7077 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7078 err = build_sched_domains(doms_cur[0], NULL);
7079 register_sched_domain_sysctl();
7085 * Detach sched domains from a group of cpus specified in cpu_map
7086 * These cpus will now be attached to the NULL domain
7088 static void detach_destroy_domains(const struct cpumask *cpu_map)
7093 for_each_cpu(i, cpu_map)
7094 cpu_attach_domain(NULL, &def_root_domain, i);
7098 /* handle null as "default" */
7099 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7100 struct sched_domain_attr *new, int idx_new)
7102 struct sched_domain_attr tmp;
7109 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7110 new ? (new + idx_new) : &tmp,
7111 sizeof(struct sched_domain_attr));
7115 * Partition sched domains as specified by the 'ndoms_new'
7116 * cpumasks in the array doms_new[] of cpumasks. This compares
7117 * doms_new[] to the current sched domain partitioning, doms_cur[].
7118 * It destroys each deleted domain and builds each new domain.
7120 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7121 * The masks don't intersect (don't overlap.) We should setup one
7122 * sched domain for each mask. CPUs not in any of the cpumasks will
7123 * not be load balanced. If the same cpumask appears both in the
7124 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7127 * The passed in 'doms_new' should be allocated using
7128 * alloc_sched_domains. This routine takes ownership of it and will
7129 * free_sched_domains it when done with it. If the caller failed the
7130 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7131 * and partition_sched_domains() will fallback to the single partition
7132 * 'fallback_doms', it also forces the domains to be rebuilt.
7134 * If doms_new == NULL it will be replaced with cpu_online_mask.
7135 * ndoms_new == 0 is a special case for destroying existing domains,
7136 * and it will not create the default domain.
7138 * Call with hotplug lock held
7140 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7141 struct sched_domain_attr *dattr_new)
7146 mutex_lock(&sched_domains_mutex);
7148 /* always unregister in case we don't destroy any domains */
7149 unregister_sched_domain_sysctl();
7151 /* Let architecture update cpu core mappings. */
7152 new_topology = arch_update_cpu_topology();
7154 n = doms_new ? ndoms_new : 0;
7156 /* Destroy deleted domains */
7157 for (i = 0; i < ndoms_cur; i++) {
7158 for (j = 0; j < n && !new_topology; j++) {
7159 if (cpumask_equal(doms_cur[i], doms_new[j])
7160 && dattrs_equal(dattr_cur, i, dattr_new, j))
7163 /* no match - a current sched domain not in new doms_new[] */
7164 detach_destroy_domains(doms_cur[i]);
7170 if (doms_new == NULL) {
7172 doms_new = &fallback_doms;
7173 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7174 WARN_ON_ONCE(dattr_new);
7177 /* Build new domains */
7178 for (i = 0; i < ndoms_new; i++) {
7179 for (j = 0; j < n && !new_topology; j++) {
7180 if (cpumask_equal(doms_new[i], doms_cur[j])
7181 && dattrs_equal(dattr_new, i, dattr_cur, j))
7184 /* no match - add a new doms_new */
7185 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7190 /* Remember the new sched domains */
7191 if (doms_cur != &fallback_doms)
7192 free_sched_domains(doms_cur, ndoms_cur);
7193 kfree(dattr_cur); /* kfree(NULL) is safe */
7194 doms_cur = doms_new;
7195 dattr_cur = dattr_new;
7196 ndoms_cur = ndoms_new;
7198 register_sched_domain_sysctl();
7200 mutex_unlock(&sched_domains_mutex);
7203 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7206 * Update cpusets according to cpu_active mask. If cpusets are
7207 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7208 * around partition_sched_domains().
7210 * If we come here as part of a suspend/resume, don't touch cpusets because we
7211 * want to restore it back to its original state upon resume anyway.
7213 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7217 case CPU_ONLINE_FROZEN:
7218 case CPU_DOWN_FAILED_FROZEN:
7221 * num_cpus_frozen tracks how many CPUs are involved in suspend
7222 * resume sequence. As long as this is not the last online
7223 * operation in the resume sequence, just build a single sched
7224 * domain, ignoring cpusets.
7227 if (likely(num_cpus_frozen)) {
7228 partition_sched_domains(1, NULL, NULL);
7233 * This is the last CPU online operation. So fall through and
7234 * restore the original sched domains by considering the
7235 * cpuset configurations.
7239 cpuset_update_active_cpus(true);
7247 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7250 unsigned long flags;
7251 long cpu = (long)hcpu;
7257 case CPU_DOWN_PREPARE:
7258 rcu_read_lock_sched();
7259 dl_b = dl_bw_of(cpu);
7261 raw_spin_lock_irqsave(&dl_b->lock, flags);
7262 cpus = dl_bw_cpus(cpu);
7263 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7264 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7266 rcu_read_unlock_sched();
7269 return notifier_from_errno(-EBUSY);
7270 cpuset_update_active_cpus(false);
7272 case CPU_DOWN_PREPARE_FROZEN:
7274 partition_sched_domains(1, NULL, NULL);
7282 void __init sched_init_smp(void)
7284 cpumask_var_t non_isolated_cpus;
7286 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7287 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7292 * There's no userspace yet to cause hotplug operations; hence all the
7293 * cpu masks are stable and all blatant races in the below code cannot
7296 mutex_lock(&sched_domains_mutex);
7297 init_sched_domains(cpu_active_mask);
7298 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7299 if (cpumask_empty(non_isolated_cpus))
7300 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7301 mutex_unlock(&sched_domains_mutex);
7303 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7304 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7305 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7309 /* Move init over to a non-isolated CPU */
7310 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7312 sched_init_granularity();
7313 free_cpumask_var(non_isolated_cpus);
7315 init_sched_rt_class();
7316 init_sched_dl_class();
7319 void __init sched_init_smp(void)
7321 sched_init_granularity();
7323 #endif /* CONFIG_SMP */
7325 int in_sched_functions(unsigned long addr)
7327 return in_lock_functions(addr) ||
7328 (addr >= (unsigned long)__sched_text_start
7329 && addr < (unsigned long)__sched_text_end);
7332 #ifdef CONFIG_CGROUP_SCHED
7334 * Default task group.
7335 * Every task in system belongs to this group at bootup.
7337 struct task_group root_task_group;
7338 LIST_HEAD(task_groups);
7341 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7343 void __init sched_init(void)
7346 unsigned long alloc_size = 0, ptr;
7348 #ifdef CONFIG_FAIR_GROUP_SCHED
7349 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7351 #ifdef CONFIG_RT_GROUP_SCHED
7352 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7355 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7357 #ifdef CONFIG_FAIR_GROUP_SCHED
7358 root_task_group.se = (struct sched_entity **)ptr;
7359 ptr += nr_cpu_ids * sizeof(void **);
7361 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7362 ptr += nr_cpu_ids * sizeof(void **);
7364 #endif /* CONFIG_FAIR_GROUP_SCHED */
7365 #ifdef CONFIG_RT_GROUP_SCHED
7366 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7367 ptr += nr_cpu_ids * sizeof(void **);
7369 root_task_group.rt_rq = (struct rt_rq **)ptr;
7370 ptr += nr_cpu_ids * sizeof(void **);
7372 #endif /* CONFIG_RT_GROUP_SCHED */
7374 #ifdef CONFIG_CPUMASK_OFFSTACK
7375 for_each_possible_cpu(i) {
7376 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7377 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7379 #endif /* CONFIG_CPUMASK_OFFSTACK */
7381 init_rt_bandwidth(&def_rt_bandwidth,
7382 global_rt_period(), global_rt_runtime());
7383 init_dl_bandwidth(&def_dl_bandwidth,
7384 global_rt_period(), global_rt_runtime());
7387 init_defrootdomain();
7390 #ifdef CONFIG_RT_GROUP_SCHED
7391 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7392 global_rt_period(), global_rt_runtime());
7393 #endif /* CONFIG_RT_GROUP_SCHED */
7395 #ifdef CONFIG_CGROUP_SCHED
7396 list_add(&root_task_group.list, &task_groups);
7397 INIT_LIST_HEAD(&root_task_group.children);
7398 INIT_LIST_HEAD(&root_task_group.siblings);
7399 autogroup_init(&init_task);
7401 #endif /* CONFIG_CGROUP_SCHED */
7403 for_each_possible_cpu(i) {
7407 raw_spin_lock_init(&rq->lock);
7409 rq->calc_load_active = 0;
7410 rq->calc_load_update = jiffies + LOAD_FREQ;
7411 init_cfs_rq(&rq->cfs);
7412 init_rt_rq(&rq->rt);
7413 init_dl_rq(&rq->dl);
7414 #ifdef CONFIG_FAIR_GROUP_SCHED
7415 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7416 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7418 * How much cpu bandwidth does root_task_group get?
7420 * In case of task-groups formed thr' the cgroup filesystem, it
7421 * gets 100% of the cpu resources in the system. This overall
7422 * system cpu resource is divided among the tasks of
7423 * root_task_group and its child task-groups in a fair manner,
7424 * based on each entity's (task or task-group's) weight
7425 * (se->load.weight).
7427 * In other words, if root_task_group has 10 tasks of weight
7428 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7429 * then A0's share of the cpu resource is:
7431 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7433 * We achieve this by letting root_task_group's tasks sit
7434 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7436 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7437 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7438 #endif /* CONFIG_FAIR_GROUP_SCHED */
7440 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7441 #ifdef CONFIG_RT_GROUP_SCHED
7442 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7445 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7446 rq->cpu_load[j] = 0;
7448 rq->last_load_update_tick = jiffies;
7453 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7454 rq->balance_callback = NULL;
7455 rq->active_balance = 0;
7456 rq->next_balance = jiffies;
7461 rq->avg_idle = 2*sysctl_sched_migration_cost;
7462 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7464 INIT_LIST_HEAD(&rq->cfs_tasks);
7466 rq_attach_root(rq, &def_root_domain);
7467 #ifdef CONFIG_NO_HZ_COMMON
7470 #ifdef CONFIG_NO_HZ_FULL
7471 rq->last_sched_tick = 0;
7475 atomic_set(&rq->nr_iowait, 0);
7478 set_load_weight(&init_task);
7480 #ifdef CONFIG_PREEMPT_NOTIFIERS
7481 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7485 * The boot idle thread does lazy MMU switching as well:
7487 atomic_inc(&init_mm.mm_count);
7488 enter_lazy_tlb(&init_mm, current);
7491 * During early bootup we pretend to be a normal task:
7493 current->sched_class = &fair_sched_class;
7496 * Make us the idle thread. Technically, schedule() should not be
7497 * called from this thread, however somewhere below it might be,
7498 * but because we are the idle thread, we just pick up running again
7499 * when this runqueue becomes "idle".
7501 init_idle(current, smp_processor_id());
7503 calc_load_update = jiffies + LOAD_FREQ;
7506 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7507 /* May be allocated at isolcpus cmdline parse time */
7508 if (cpu_isolated_map == NULL)
7509 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7510 idle_thread_set_boot_cpu();
7511 set_cpu_rq_start_time();
7513 init_sched_fair_class();
7515 scheduler_running = 1;
7518 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7519 static inline int preempt_count_equals(int preempt_offset)
7521 int nested = preempt_count() + rcu_preempt_depth();
7523 return (nested == preempt_offset);
7526 void __might_sleep(const char *file, int line, int preempt_offset)
7529 * Blocking primitives will set (and therefore destroy) current->state,
7530 * since we will exit with TASK_RUNNING make sure we enter with it,
7531 * otherwise we will destroy state.
7533 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7534 "do not call blocking ops when !TASK_RUNNING; "
7535 "state=%lx set at [<%p>] %pS\n",
7537 (void *)current->task_state_change,
7538 (void *)current->task_state_change);
7540 ___might_sleep(file, line, preempt_offset);
7542 EXPORT_SYMBOL(__might_sleep);
7544 void ___might_sleep(const char *file, int line, int preempt_offset)
7546 static unsigned long prev_jiffy; /* ratelimiting */
7548 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7549 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7550 !is_idle_task(current)) ||
7551 system_state != SYSTEM_RUNNING || oops_in_progress)
7553 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7555 prev_jiffy = jiffies;
7558 "BUG: sleeping function called from invalid context at %s:%d\n",
7561 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7562 in_atomic(), irqs_disabled(),
7563 current->pid, current->comm);
7565 if (task_stack_end_corrupted(current))
7566 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7568 debug_show_held_locks(current);
7569 if (irqs_disabled())
7570 print_irqtrace_events(current);
7571 #ifdef CONFIG_DEBUG_PREEMPT
7572 if (!preempt_count_equals(preempt_offset)) {
7573 pr_err("Preemption disabled at:");
7574 print_ip_sym(current->preempt_disable_ip);
7580 EXPORT_SYMBOL(___might_sleep);
7583 #ifdef CONFIG_MAGIC_SYSRQ
7584 void normalize_rt_tasks(void)
7586 struct task_struct *g, *p;
7587 struct sched_attr attr = {
7588 .sched_policy = SCHED_NORMAL,
7591 read_lock(&tasklist_lock);
7592 for_each_process_thread(g, p) {
7594 * Only normalize user tasks:
7596 if (p->flags & PF_KTHREAD)
7599 p->se.exec_start = 0;
7600 #ifdef CONFIG_SCHEDSTATS
7601 p->se.statistics.wait_start = 0;
7602 p->se.statistics.sleep_start = 0;
7603 p->se.statistics.block_start = 0;
7606 if (!dl_task(p) && !rt_task(p)) {
7608 * Renice negative nice level userspace
7611 if (task_nice(p) < 0)
7612 set_user_nice(p, 0);
7616 __sched_setscheduler(p, &attr, false, false);
7618 read_unlock(&tasklist_lock);
7621 #endif /* CONFIG_MAGIC_SYSRQ */
7623 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7625 * These functions are only useful for the IA64 MCA handling, or kdb.
7627 * They can only be called when the whole system has been
7628 * stopped - every CPU needs to be quiescent, and no scheduling
7629 * activity can take place. Using them for anything else would
7630 * be a serious bug, and as a result, they aren't even visible
7631 * under any other configuration.
7635 * curr_task - return the current task for a given cpu.
7636 * @cpu: the processor in question.
7638 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7640 * Return: The current task for @cpu.
7642 struct task_struct *curr_task(int cpu)
7644 return cpu_curr(cpu);
7647 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7651 * set_curr_task - set the current task for a given cpu.
7652 * @cpu: the processor in question.
7653 * @p: the task pointer to set.
7655 * Description: This function must only be used when non-maskable interrupts
7656 * are serviced on a separate stack. It allows the architecture to switch the
7657 * notion of the current task on a cpu in a non-blocking manner. This function
7658 * must be called with all CPU's synchronized, and interrupts disabled, the
7659 * and caller must save the original value of the current task (see
7660 * curr_task() above) and restore that value before reenabling interrupts and
7661 * re-starting the system.
7663 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7665 void set_curr_task(int cpu, struct task_struct *p)
7672 #ifdef CONFIG_CGROUP_SCHED
7673 /* task_group_lock serializes the addition/removal of task groups */
7674 static DEFINE_SPINLOCK(task_group_lock);
7676 static void free_sched_group(struct task_group *tg)
7678 free_fair_sched_group(tg);
7679 free_rt_sched_group(tg);
7684 /* allocate runqueue etc for a new task group */
7685 struct task_group *sched_create_group(struct task_group *parent)
7687 struct task_group *tg;
7689 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7691 return ERR_PTR(-ENOMEM);
7693 if (!alloc_fair_sched_group(tg, parent))
7696 if (!alloc_rt_sched_group(tg, parent))
7702 free_sched_group(tg);
7703 return ERR_PTR(-ENOMEM);
7706 void sched_online_group(struct task_group *tg, struct task_group *parent)
7708 unsigned long flags;
7710 spin_lock_irqsave(&task_group_lock, flags);
7711 list_add_rcu(&tg->list, &task_groups);
7713 WARN_ON(!parent); /* root should already exist */
7715 tg->parent = parent;
7716 INIT_LIST_HEAD(&tg->children);
7717 list_add_rcu(&tg->siblings, &parent->children);
7718 spin_unlock_irqrestore(&task_group_lock, flags);
7721 /* rcu callback to free various structures associated with a task group */
7722 static void free_sched_group_rcu(struct rcu_head *rhp)
7724 /* now it should be safe to free those cfs_rqs */
7725 free_sched_group(container_of(rhp, struct task_group, rcu));
7728 /* Destroy runqueue etc associated with a task group */
7729 void sched_destroy_group(struct task_group *tg)
7731 /* wait for possible concurrent references to cfs_rqs complete */
7732 call_rcu(&tg->rcu, free_sched_group_rcu);
7735 void sched_offline_group(struct task_group *tg)
7737 unsigned long flags;
7740 /* end participation in shares distribution */
7741 for_each_possible_cpu(i)
7742 unregister_fair_sched_group(tg, i);
7744 spin_lock_irqsave(&task_group_lock, flags);
7745 list_del_rcu(&tg->list);
7746 list_del_rcu(&tg->siblings);
7747 spin_unlock_irqrestore(&task_group_lock, flags);
7750 /* change task's runqueue when it moves between groups.
7751 * The caller of this function should have put the task in its new group
7752 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7753 * reflect its new group.
7755 void sched_move_task(struct task_struct *tsk)
7757 struct task_group *tg;
7758 int queued, running;
7759 unsigned long flags;
7762 rq = task_rq_lock(tsk, &flags);
7764 running = task_current(rq, tsk);
7765 queued = task_on_rq_queued(tsk);
7768 dequeue_task(rq, tsk, DEQUEUE_SAVE);
7769 if (unlikely(running))
7770 put_prev_task(rq, tsk);
7773 * All callers are synchronized by task_rq_lock(); we do not use RCU
7774 * which is pointless here. Thus, we pass "true" to task_css_check()
7775 * to prevent lockdep warnings.
7777 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7778 struct task_group, css);
7779 tg = autogroup_task_group(tsk, tg);
7780 tsk->sched_task_group = tg;
7782 #ifdef CONFIG_FAIR_GROUP_SCHED
7783 if (tsk->sched_class->task_move_group)
7784 tsk->sched_class->task_move_group(tsk);
7787 set_task_rq(tsk, task_cpu(tsk));
7789 if (unlikely(running))
7790 tsk->sched_class->set_curr_task(rq);
7792 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
7794 task_rq_unlock(rq, tsk, &flags);
7796 #endif /* CONFIG_CGROUP_SCHED */
7798 #ifdef CONFIG_RT_GROUP_SCHED
7800 * Ensure that the real time constraints are schedulable.
7802 static DEFINE_MUTEX(rt_constraints_mutex);
7804 /* Must be called with tasklist_lock held */
7805 static inline int tg_has_rt_tasks(struct task_group *tg)
7807 struct task_struct *g, *p;
7810 * Autogroups do not have RT tasks; see autogroup_create().
7812 if (task_group_is_autogroup(tg))
7815 for_each_process_thread(g, p) {
7816 if (rt_task(p) && task_group(p) == tg)
7823 struct rt_schedulable_data {
7824 struct task_group *tg;
7829 static int tg_rt_schedulable(struct task_group *tg, void *data)
7831 struct rt_schedulable_data *d = data;
7832 struct task_group *child;
7833 unsigned long total, sum = 0;
7834 u64 period, runtime;
7836 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7837 runtime = tg->rt_bandwidth.rt_runtime;
7840 period = d->rt_period;
7841 runtime = d->rt_runtime;
7845 * Cannot have more runtime than the period.
7847 if (runtime > period && runtime != RUNTIME_INF)
7851 * Ensure we don't starve existing RT tasks.
7853 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7856 total = to_ratio(period, runtime);
7859 * Nobody can have more than the global setting allows.
7861 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7865 * The sum of our children's runtime should not exceed our own.
7867 list_for_each_entry_rcu(child, &tg->children, siblings) {
7868 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7869 runtime = child->rt_bandwidth.rt_runtime;
7871 if (child == d->tg) {
7872 period = d->rt_period;
7873 runtime = d->rt_runtime;
7876 sum += to_ratio(period, runtime);
7885 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7889 struct rt_schedulable_data data = {
7891 .rt_period = period,
7892 .rt_runtime = runtime,
7896 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7902 static int tg_set_rt_bandwidth(struct task_group *tg,
7903 u64 rt_period, u64 rt_runtime)
7908 * Disallowing the root group RT runtime is BAD, it would disallow the
7909 * kernel creating (and or operating) RT threads.
7911 if (tg == &root_task_group && rt_runtime == 0)
7914 /* No period doesn't make any sense. */
7918 mutex_lock(&rt_constraints_mutex);
7919 read_lock(&tasklist_lock);
7920 err = __rt_schedulable(tg, rt_period, rt_runtime);
7924 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7925 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7926 tg->rt_bandwidth.rt_runtime = rt_runtime;
7928 for_each_possible_cpu(i) {
7929 struct rt_rq *rt_rq = tg->rt_rq[i];
7931 raw_spin_lock(&rt_rq->rt_runtime_lock);
7932 rt_rq->rt_runtime = rt_runtime;
7933 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7935 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7937 read_unlock(&tasklist_lock);
7938 mutex_unlock(&rt_constraints_mutex);
7943 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7945 u64 rt_runtime, rt_period;
7947 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7948 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7949 if (rt_runtime_us < 0)
7950 rt_runtime = RUNTIME_INF;
7952 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7955 static long sched_group_rt_runtime(struct task_group *tg)
7959 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7962 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7963 do_div(rt_runtime_us, NSEC_PER_USEC);
7964 return rt_runtime_us;
7967 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7969 u64 rt_runtime, rt_period;
7971 rt_period = rt_period_us * NSEC_PER_USEC;
7972 rt_runtime = tg->rt_bandwidth.rt_runtime;
7974 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7977 static long sched_group_rt_period(struct task_group *tg)
7981 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7982 do_div(rt_period_us, NSEC_PER_USEC);
7983 return rt_period_us;
7985 #endif /* CONFIG_RT_GROUP_SCHED */
7987 #ifdef CONFIG_RT_GROUP_SCHED
7988 static int sched_rt_global_constraints(void)
7992 mutex_lock(&rt_constraints_mutex);
7993 read_lock(&tasklist_lock);
7994 ret = __rt_schedulable(NULL, 0, 0);
7995 read_unlock(&tasklist_lock);
7996 mutex_unlock(&rt_constraints_mutex);
8001 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8003 /* Don't accept realtime tasks when there is no way for them to run */
8004 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8010 #else /* !CONFIG_RT_GROUP_SCHED */
8011 static int sched_rt_global_constraints(void)
8013 unsigned long flags;
8016 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8017 for_each_possible_cpu(i) {
8018 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8020 raw_spin_lock(&rt_rq->rt_runtime_lock);
8021 rt_rq->rt_runtime = global_rt_runtime();
8022 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8024 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8028 #endif /* CONFIG_RT_GROUP_SCHED */
8030 static int sched_dl_global_validate(void)
8032 u64 runtime = global_rt_runtime();
8033 u64 period = global_rt_period();
8034 u64 new_bw = to_ratio(period, runtime);
8037 unsigned long flags;
8040 * Here we want to check the bandwidth not being set to some
8041 * value smaller than the currently allocated bandwidth in
8042 * any of the root_domains.
8044 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8045 * cycling on root_domains... Discussion on different/better
8046 * solutions is welcome!
8048 for_each_possible_cpu(cpu) {
8049 rcu_read_lock_sched();
8050 dl_b = dl_bw_of(cpu);
8052 raw_spin_lock_irqsave(&dl_b->lock, flags);
8053 if (new_bw < dl_b->total_bw)
8055 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8057 rcu_read_unlock_sched();
8066 static void sched_dl_do_global(void)
8071 unsigned long flags;
8073 def_dl_bandwidth.dl_period = global_rt_period();
8074 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8076 if (global_rt_runtime() != RUNTIME_INF)
8077 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8080 * FIXME: As above...
8082 for_each_possible_cpu(cpu) {
8083 rcu_read_lock_sched();
8084 dl_b = dl_bw_of(cpu);
8086 raw_spin_lock_irqsave(&dl_b->lock, flags);
8088 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8090 rcu_read_unlock_sched();
8094 static int sched_rt_global_validate(void)
8096 if (sysctl_sched_rt_period <= 0)
8099 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8100 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8106 static void sched_rt_do_global(void)
8108 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8109 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8112 int sched_rt_handler(struct ctl_table *table, int write,
8113 void __user *buffer, size_t *lenp,
8116 int old_period, old_runtime;
8117 static DEFINE_MUTEX(mutex);
8121 old_period = sysctl_sched_rt_period;
8122 old_runtime = sysctl_sched_rt_runtime;
8124 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8126 if (!ret && write) {
8127 ret = sched_rt_global_validate();
8131 ret = sched_dl_global_validate();
8135 ret = sched_rt_global_constraints();
8139 sched_rt_do_global();
8140 sched_dl_do_global();
8144 sysctl_sched_rt_period = old_period;
8145 sysctl_sched_rt_runtime = old_runtime;
8147 mutex_unlock(&mutex);
8152 int sched_rr_handler(struct ctl_table *table, int write,
8153 void __user *buffer, size_t *lenp,
8157 static DEFINE_MUTEX(mutex);
8160 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8161 /* make sure that internally we keep jiffies */
8162 /* also, writing zero resets timeslice to default */
8163 if (!ret && write) {
8164 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8165 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8167 mutex_unlock(&mutex);
8171 #ifdef CONFIG_CGROUP_SCHED
8173 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8175 return css ? container_of(css, struct task_group, css) : NULL;
8178 static struct cgroup_subsys_state *
8179 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8181 struct task_group *parent = css_tg(parent_css);
8182 struct task_group *tg;
8185 /* This is early initialization for the top cgroup */
8186 return &root_task_group.css;
8189 tg = sched_create_group(parent);
8191 return ERR_PTR(-ENOMEM);
8196 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8198 struct task_group *tg = css_tg(css);
8199 struct task_group *parent = css_tg(css->parent);
8202 sched_online_group(tg, parent);
8206 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8208 struct task_group *tg = css_tg(css);
8210 sched_destroy_group(tg);
8213 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8215 struct task_group *tg = css_tg(css);
8217 sched_offline_group(tg);
8220 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8222 sched_move_task(task);
8225 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8226 struct cgroup_taskset *tset)
8228 struct task_struct *task;
8230 cgroup_taskset_for_each(task, tset) {
8231 #ifdef CONFIG_RT_GROUP_SCHED
8232 if (!sched_rt_can_attach(css_tg(css), task))
8235 /* We don't support RT-tasks being in separate groups */
8236 if (task->sched_class != &fair_sched_class)
8243 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8244 struct cgroup_taskset *tset)
8246 struct task_struct *task;
8248 cgroup_taskset_for_each(task, tset)
8249 sched_move_task(task);
8252 #ifdef CONFIG_FAIR_GROUP_SCHED
8253 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8254 struct cftype *cftype, u64 shareval)
8256 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8259 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8262 struct task_group *tg = css_tg(css);
8264 return (u64) scale_load_down(tg->shares);
8267 #ifdef CONFIG_CFS_BANDWIDTH
8268 static DEFINE_MUTEX(cfs_constraints_mutex);
8270 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8271 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8273 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8275 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8277 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8278 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8280 if (tg == &root_task_group)
8284 * Ensure we have at some amount of bandwidth every period. This is
8285 * to prevent reaching a state of large arrears when throttled via
8286 * entity_tick() resulting in prolonged exit starvation.
8288 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8292 * Likewise, bound things on the otherside by preventing insane quota
8293 * periods. This also allows us to normalize in computing quota
8296 if (period > max_cfs_quota_period)
8300 * Prevent race between setting of cfs_rq->runtime_enabled and
8301 * unthrottle_offline_cfs_rqs().
8304 mutex_lock(&cfs_constraints_mutex);
8305 ret = __cfs_schedulable(tg, period, quota);
8309 runtime_enabled = quota != RUNTIME_INF;
8310 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8312 * If we need to toggle cfs_bandwidth_used, off->on must occur
8313 * before making related changes, and on->off must occur afterwards
8315 if (runtime_enabled && !runtime_was_enabled)
8316 cfs_bandwidth_usage_inc();
8317 raw_spin_lock_irq(&cfs_b->lock);
8318 cfs_b->period = ns_to_ktime(period);
8319 cfs_b->quota = quota;
8321 __refill_cfs_bandwidth_runtime(cfs_b);
8322 /* restart the period timer (if active) to handle new period expiry */
8323 if (runtime_enabled)
8324 start_cfs_bandwidth(cfs_b);
8325 raw_spin_unlock_irq(&cfs_b->lock);
8327 for_each_online_cpu(i) {
8328 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8329 struct rq *rq = cfs_rq->rq;
8331 raw_spin_lock_irq(&rq->lock);
8332 cfs_rq->runtime_enabled = runtime_enabled;
8333 cfs_rq->runtime_remaining = 0;
8335 if (cfs_rq->throttled)
8336 unthrottle_cfs_rq(cfs_rq);
8337 raw_spin_unlock_irq(&rq->lock);
8339 if (runtime_was_enabled && !runtime_enabled)
8340 cfs_bandwidth_usage_dec();
8342 mutex_unlock(&cfs_constraints_mutex);
8348 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8352 period = ktime_to_ns(tg->cfs_bandwidth.period);
8353 if (cfs_quota_us < 0)
8354 quota = RUNTIME_INF;
8356 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8358 return tg_set_cfs_bandwidth(tg, period, quota);
8361 long tg_get_cfs_quota(struct task_group *tg)
8365 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8368 quota_us = tg->cfs_bandwidth.quota;
8369 do_div(quota_us, NSEC_PER_USEC);
8374 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8378 period = (u64)cfs_period_us * NSEC_PER_USEC;
8379 quota = tg->cfs_bandwidth.quota;
8381 return tg_set_cfs_bandwidth(tg, period, quota);
8384 long tg_get_cfs_period(struct task_group *tg)
8388 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8389 do_div(cfs_period_us, NSEC_PER_USEC);
8391 return cfs_period_us;
8394 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8397 return tg_get_cfs_quota(css_tg(css));
8400 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8401 struct cftype *cftype, s64 cfs_quota_us)
8403 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8406 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8409 return tg_get_cfs_period(css_tg(css));
8412 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8413 struct cftype *cftype, u64 cfs_period_us)
8415 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8418 struct cfs_schedulable_data {
8419 struct task_group *tg;
8424 * normalize group quota/period to be quota/max_period
8425 * note: units are usecs
8427 static u64 normalize_cfs_quota(struct task_group *tg,
8428 struct cfs_schedulable_data *d)
8436 period = tg_get_cfs_period(tg);
8437 quota = tg_get_cfs_quota(tg);
8440 /* note: these should typically be equivalent */
8441 if (quota == RUNTIME_INF || quota == -1)
8444 return to_ratio(period, quota);
8447 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8449 struct cfs_schedulable_data *d = data;
8450 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8451 s64 quota = 0, parent_quota = -1;
8454 quota = RUNTIME_INF;
8456 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8458 quota = normalize_cfs_quota(tg, d);
8459 parent_quota = parent_b->hierarchical_quota;
8462 * ensure max(child_quota) <= parent_quota, inherit when no
8465 if (quota == RUNTIME_INF)
8466 quota = parent_quota;
8467 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8470 cfs_b->hierarchical_quota = quota;
8475 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8478 struct cfs_schedulable_data data = {
8484 if (quota != RUNTIME_INF) {
8485 do_div(data.period, NSEC_PER_USEC);
8486 do_div(data.quota, NSEC_PER_USEC);
8490 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8496 static int cpu_stats_show(struct seq_file *sf, void *v)
8498 struct task_group *tg = css_tg(seq_css(sf));
8499 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8501 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8502 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8503 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8507 #endif /* CONFIG_CFS_BANDWIDTH */
8508 #endif /* CONFIG_FAIR_GROUP_SCHED */
8510 #ifdef CONFIG_RT_GROUP_SCHED
8511 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8512 struct cftype *cft, s64 val)
8514 return sched_group_set_rt_runtime(css_tg(css), val);
8517 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8520 return sched_group_rt_runtime(css_tg(css));
8523 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8524 struct cftype *cftype, u64 rt_period_us)
8526 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8529 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8532 return sched_group_rt_period(css_tg(css));
8534 #endif /* CONFIG_RT_GROUP_SCHED */
8536 static struct cftype cpu_files[] = {
8537 #ifdef CONFIG_FAIR_GROUP_SCHED
8540 .read_u64 = cpu_shares_read_u64,
8541 .write_u64 = cpu_shares_write_u64,
8544 #ifdef CONFIG_CFS_BANDWIDTH
8546 .name = "cfs_quota_us",
8547 .read_s64 = cpu_cfs_quota_read_s64,
8548 .write_s64 = cpu_cfs_quota_write_s64,
8551 .name = "cfs_period_us",
8552 .read_u64 = cpu_cfs_period_read_u64,
8553 .write_u64 = cpu_cfs_period_write_u64,
8557 .seq_show = cpu_stats_show,
8560 #ifdef CONFIG_RT_GROUP_SCHED
8562 .name = "rt_runtime_us",
8563 .read_s64 = cpu_rt_runtime_read,
8564 .write_s64 = cpu_rt_runtime_write,
8567 .name = "rt_period_us",
8568 .read_u64 = cpu_rt_period_read_uint,
8569 .write_u64 = cpu_rt_period_write_uint,
8575 struct cgroup_subsys cpu_cgrp_subsys = {
8576 .css_alloc = cpu_cgroup_css_alloc,
8577 .css_free = cpu_cgroup_css_free,
8578 .css_online = cpu_cgroup_css_online,
8579 .css_offline = cpu_cgroup_css_offline,
8580 .fork = cpu_cgroup_fork,
8581 .can_attach = cpu_cgroup_can_attach,
8582 .attach = cpu_cgroup_attach,
8583 .legacy_cftypes = cpu_files,
8587 #endif /* CONFIG_CGROUP_SCHED */
8589 void dump_cpu_task(int cpu)
8591 pr_info("Task dump for CPU %d:\n", cpu);
8592 sched_show_task(cpu_curr(cpu));