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 * Notes on Program-Order guarantees on SMP systems.
1913 * The basic program-order guarantee on SMP systems is that when a task [t]
1914 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1915 * execution on its new cpu [c1].
1917 * For migration (of runnable tasks) this is provided by the following means:
1919 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1920 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1921 * rq(c1)->lock (if not at the same time, then in that order).
1922 * C) LOCK of the rq(c1)->lock scheduling in task
1924 * Transitivity guarantees that B happens after A and C after B.
1925 * Note: we only require RCpc transitivity.
1926 * Note: the cpu doing B need not be c0 or c1
1935 * UNLOCK rq(0)->lock
1937 * LOCK rq(0)->lock // orders against CPU0
1939 * UNLOCK rq(0)->lock
1943 * UNLOCK rq(1)->lock
1945 * LOCK rq(1)->lock // orders against CPU2
1948 * UNLOCK rq(1)->lock
1951 * BLOCKING -- aka. SLEEP + WAKEUP
1953 * For blocking we (obviously) need to provide the same guarantee as for
1954 * migration. However the means are completely different as there is no lock
1955 * chain to provide order. Instead we do:
1957 * 1) smp_store_release(X->on_cpu, 0)
1958 * 2) smp_cond_acquire(!X->on_cpu)
1962 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1964 * LOCK rq(0)->lock LOCK X->pi_lock
1967 * smp_store_release(X->on_cpu, 0);
1969 * smp_cond_acquire(!X->on_cpu);
1975 * X->state = RUNNING
1976 * UNLOCK rq(2)->lock
1978 * LOCK rq(2)->lock // orders against CPU1
1981 * UNLOCK rq(2)->lock
1984 * UNLOCK rq(0)->lock
1987 * However; for wakeups there is a second guarantee we must provide, namely we
1988 * must observe the state that lead to our wakeup. That is, not only must our
1989 * task observe its own prior state, it must also observe the stores prior to
1992 * This means that any means of doing remote wakeups must order the CPU doing
1993 * the wakeup against the CPU the task is going to end up running on. This,
1994 * however, is already required for the regular Program-Order guarantee above,
1995 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
2000 * try_to_wake_up - wake up a thread
2001 * @p: the thread to be awakened
2002 * @state: the mask of task states that can be woken
2003 * @wake_flags: wake modifier flags (WF_*)
2005 * Put it on the run-queue if it's not already there. The "current"
2006 * thread is always on the run-queue (except when the actual
2007 * re-schedule is in progress), and as such you're allowed to do
2008 * the simpler "current->state = TASK_RUNNING" to mark yourself
2009 * runnable without the overhead of this.
2011 * Return: %true if @p was woken up, %false if it was already running.
2012 * or @state didn't match @p's state.
2015 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2017 unsigned long flags;
2018 int cpu, success = 0;
2021 * If we are going to wake up a thread waiting for CONDITION we
2022 * need to ensure that CONDITION=1 done by the caller can not be
2023 * reordered with p->state check below. This pairs with mb() in
2024 * set_current_state() the waiting thread does.
2026 smp_mb__before_spinlock();
2027 raw_spin_lock_irqsave(&p->pi_lock, flags);
2028 if (!(p->state & state))
2031 trace_sched_waking(p);
2033 success = 1; /* we're going to change ->state */
2036 if (p->on_rq && ttwu_remote(p, wake_flags))
2041 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2042 * possible to, falsely, observe p->on_cpu == 0.
2044 * One must be running (->on_cpu == 1) in order to remove oneself
2045 * from the runqueue.
2047 * [S] ->on_cpu = 1; [L] ->on_rq
2051 * [S] ->on_rq = 0; [L] ->on_cpu
2053 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2054 * from the consecutive calls to schedule(); the first switching to our
2055 * task, the second putting it to sleep.
2060 * If the owning (remote) cpu is still in the middle of schedule() with
2061 * this task as prev, wait until its done referencing the task.
2063 * Pairs with the smp_store_release() in finish_lock_switch().
2065 * This ensures that tasks getting woken will be fully ordered against
2066 * their previous state and preserve Program Order.
2068 smp_cond_acquire(!p->on_cpu);
2070 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2071 p->state = TASK_WAKING;
2073 if (p->sched_class->task_waking)
2074 p->sched_class->task_waking(p);
2076 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2077 if (task_cpu(p) != cpu) {
2078 wake_flags |= WF_MIGRATED;
2079 set_task_cpu(p, cpu);
2081 #endif /* CONFIG_SMP */
2085 ttwu_stat(p, cpu, wake_flags);
2087 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2093 * try_to_wake_up_local - try to wake up a local task with rq lock held
2094 * @p: the thread to be awakened
2096 * Put @p on the run-queue if it's not already there. The caller must
2097 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2100 static void try_to_wake_up_local(struct task_struct *p)
2102 struct rq *rq = task_rq(p);
2104 if (WARN_ON_ONCE(rq != this_rq()) ||
2105 WARN_ON_ONCE(p == current))
2108 lockdep_assert_held(&rq->lock);
2110 if (!raw_spin_trylock(&p->pi_lock)) {
2112 * This is OK, because current is on_cpu, which avoids it being
2113 * picked for load-balance and preemption/IRQs are still
2114 * disabled avoiding further scheduler activity on it and we've
2115 * not yet picked a replacement task.
2117 lockdep_unpin_lock(&rq->lock);
2118 raw_spin_unlock(&rq->lock);
2119 raw_spin_lock(&p->pi_lock);
2120 raw_spin_lock(&rq->lock);
2121 lockdep_pin_lock(&rq->lock);
2124 if (!(p->state & TASK_NORMAL))
2127 trace_sched_waking(p);
2129 if (!task_on_rq_queued(p))
2130 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2132 ttwu_do_wakeup(rq, p, 0);
2133 ttwu_stat(p, smp_processor_id(), 0);
2135 raw_spin_unlock(&p->pi_lock);
2139 * wake_up_process - Wake up a specific process
2140 * @p: The process to be woken up.
2142 * Attempt to wake up the nominated process and move it to the set of runnable
2145 * Return: 1 if the process was woken up, 0 if it was already running.
2147 * It may be assumed that this function implies a write memory barrier before
2148 * changing the task state if and only if any tasks are woken up.
2150 int wake_up_process(struct task_struct *p)
2152 return try_to_wake_up(p, TASK_NORMAL, 0);
2154 EXPORT_SYMBOL(wake_up_process);
2156 int wake_up_state(struct task_struct *p, unsigned int state)
2158 return try_to_wake_up(p, state, 0);
2162 * This function clears the sched_dl_entity static params.
2164 void __dl_clear_params(struct task_struct *p)
2166 struct sched_dl_entity *dl_se = &p->dl;
2168 dl_se->dl_runtime = 0;
2169 dl_se->dl_deadline = 0;
2170 dl_se->dl_period = 0;
2174 dl_se->dl_throttled = 0;
2176 dl_se->dl_yielded = 0;
2180 * Perform scheduler related setup for a newly forked process p.
2181 * p is forked by current.
2183 * __sched_fork() is basic setup used by init_idle() too:
2185 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2190 p->se.exec_start = 0;
2191 p->se.sum_exec_runtime = 0;
2192 p->se.prev_sum_exec_runtime = 0;
2193 p->se.nr_migrations = 0;
2195 INIT_LIST_HEAD(&p->se.group_node);
2197 #ifdef CONFIG_SCHEDSTATS
2198 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2201 RB_CLEAR_NODE(&p->dl.rb_node);
2202 init_dl_task_timer(&p->dl);
2203 __dl_clear_params(p);
2205 INIT_LIST_HEAD(&p->rt.run_list);
2207 #ifdef CONFIG_PREEMPT_NOTIFIERS
2208 INIT_HLIST_HEAD(&p->preempt_notifiers);
2211 #ifdef CONFIG_NUMA_BALANCING
2212 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2213 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2214 p->mm->numa_scan_seq = 0;
2217 if (clone_flags & CLONE_VM)
2218 p->numa_preferred_nid = current->numa_preferred_nid;
2220 p->numa_preferred_nid = -1;
2222 p->node_stamp = 0ULL;
2223 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2224 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2225 p->numa_work.next = &p->numa_work;
2226 p->numa_faults = NULL;
2227 p->last_task_numa_placement = 0;
2228 p->last_sum_exec_runtime = 0;
2230 p->numa_group = NULL;
2231 #endif /* CONFIG_NUMA_BALANCING */
2234 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2236 #ifdef CONFIG_NUMA_BALANCING
2238 void set_numabalancing_state(bool enabled)
2241 static_branch_enable(&sched_numa_balancing);
2243 static_branch_disable(&sched_numa_balancing);
2246 #ifdef CONFIG_PROC_SYSCTL
2247 int sysctl_numa_balancing(struct ctl_table *table, int write,
2248 void __user *buffer, size_t *lenp, loff_t *ppos)
2252 int state = static_branch_likely(&sched_numa_balancing);
2254 if (write && !capable(CAP_SYS_ADMIN))
2259 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2263 set_numabalancing_state(state);
2270 * fork()/clone()-time setup:
2272 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2274 unsigned long flags;
2275 int cpu = get_cpu();
2277 __sched_fork(clone_flags, p);
2279 * We mark the process as running here. This guarantees that
2280 * nobody will actually run it, and a signal or other external
2281 * event cannot wake it up and insert it on the runqueue either.
2283 p->state = TASK_RUNNING;
2286 * Make sure we do not leak PI boosting priority to the child.
2288 p->prio = current->normal_prio;
2291 * Revert to default priority/policy on fork if requested.
2293 if (unlikely(p->sched_reset_on_fork)) {
2294 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2295 p->policy = SCHED_NORMAL;
2296 p->static_prio = NICE_TO_PRIO(0);
2298 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2299 p->static_prio = NICE_TO_PRIO(0);
2301 p->prio = p->normal_prio = __normal_prio(p);
2305 * We don't need the reset flag anymore after the fork. It has
2306 * fulfilled its duty:
2308 p->sched_reset_on_fork = 0;
2311 if (dl_prio(p->prio)) {
2314 } else if (rt_prio(p->prio)) {
2315 p->sched_class = &rt_sched_class;
2317 p->sched_class = &fair_sched_class;
2320 if (p->sched_class->task_fork)
2321 p->sched_class->task_fork(p);
2324 * The child is not yet in the pid-hash so no cgroup attach races,
2325 * and the cgroup is pinned to this child due to cgroup_fork()
2326 * is ran before sched_fork().
2328 * Silence PROVE_RCU.
2330 raw_spin_lock_irqsave(&p->pi_lock, flags);
2331 set_task_cpu(p, cpu);
2332 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2334 #ifdef CONFIG_SCHED_INFO
2335 if (likely(sched_info_on()))
2336 memset(&p->sched_info, 0, sizeof(p->sched_info));
2338 #if defined(CONFIG_SMP)
2341 init_task_preempt_count(p);
2343 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2344 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2351 unsigned long to_ratio(u64 period, u64 runtime)
2353 if (runtime == RUNTIME_INF)
2357 * Doing this here saves a lot of checks in all
2358 * the calling paths, and returning zero seems
2359 * safe for them anyway.
2364 return div64_u64(runtime << 20, period);
2368 inline struct dl_bw *dl_bw_of(int i)
2370 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2371 "sched RCU must be held");
2372 return &cpu_rq(i)->rd->dl_bw;
2375 static inline int dl_bw_cpus(int i)
2377 struct root_domain *rd = cpu_rq(i)->rd;
2380 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2381 "sched RCU must be held");
2382 for_each_cpu_and(i, rd->span, cpu_active_mask)
2388 inline struct dl_bw *dl_bw_of(int i)
2390 return &cpu_rq(i)->dl.dl_bw;
2393 static inline int dl_bw_cpus(int i)
2400 * We must be sure that accepting a new task (or allowing changing the
2401 * parameters of an existing one) is consistent with the bandwidth
2402 * constraints. If yes, this function also accordingly updates the currently
2403 * allocated bandwidth to reflect the new situation.
2405 * This function is called while holding p's rq->lock.
2407 * XXX we should delay bw change until the task's 0-lag point, see
2410 static int dl_overflow(struct task_struct *p, int policy,
2411 const struct sched_attr *attr)
2414 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2415 u64 period = attr->sched_period ?: attr->sched_deadline;
2416 u64 runtime = attr->sched_runtime;
2417 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2420 if (new_bw == p->dl.dl_bw)
2424 * Either if a task, enters, leave, or stays -deadline but changes
2425 * its parameters, we may need to update accordingly the total
2426 * allocated bandwidth of the container.
2428 raw_spin_lock(&dl_b->lock);
2429 cpus = dl_bw_cpus(task_cpu(p));
2430 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2431 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2432 __dl_add(dl_b, new_bw);
2434 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2435 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2436 __dl_clear(dl_b, p->dl.dl_bw);
2437 __dl_add(dl_b, new_bw);
2439 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2440 __dl_clear(dl_b, p->dl.dl_bw);
2443 raw_spin_unlock(&dl_b->lock);
2448 extern void init_dl_bw(struct dl_bw *dl_b);
2451 * wake_up_new_task - wake up a newly created task for the first time.
2453 * This function will do some initial scheduler statistics housekeeping
2454 * that must be done for every newly created context, then puts the task
2455 * on the runqueue and wakes it.
2457 void wake_up_new_task(struct task_struct *p)
2459 unsigned long flags;
2462 raw_spin_lock_irqsave(&p->pi_lock, flags);
2463 /* Initialize new task's runnable average */
2464 init_entity_runnable_average(&p->se);
2467 * Fork balancing, do it here and not earlier because:
2468 * - cpus_allowed can change in the fork path
2469 * - any previously selected cpu might disappear through hotplug
2471 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2474 rq = __task_rq_lock(p);
2475 activate_task(rq, p, 0);
2476 p->on_rq = TASK_ON_RQ_QUEUED;
2477 trace_sched_wakeup_new(p);
2478 check_preempt_curr(rq, p, WF_FORK);
2480 if (p->sched_class->task_woken) {
2482 * Nothing relies on rq->lock after this, so its fine to
2485 lockdep_unpin_lock(&rq->lock);
2486 p->sched_class->task_woken(rq, p);
2487 lockdep_pin_lock(&rq->lock);
2490 task_rq_unlock(rq, p, &flags);
2493 #ifdef CONFIG_PREEMPT_NOTIFIERS
2495 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2497 void preempt_notifier_inc(void)
2499 static_key_slow_inc(&preempt_notifier_key);
2501 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2503 void preempt_notifier_dec(void)
2505 static_key_slow_dec(&preempt_notifier_key);
2507 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2510 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2511 * @notifier: notifier struct to register
2513 void preempt_notifier_register(struct preempt_notifier *notifier)
2515 if (!static_key_false(&preempt_notifier_key))
2516 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2518 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2520 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2523 * preempt_notifier_unregister - no longer interested in preemption notifications
2524 * @notifier: notifier struct to unregister
2526 * This is *not* safe to call from within a preemption notifier.
2528 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2530 hlist_del(¬ifier->link);
2532 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2534 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2536 struct preempt_notifier *notifier;
2538 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2539 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2542 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2544 if (static_key_false(&preempt_notifier_key))
2545 __fire_sched_in_preempt_notifiers(curr);
2549 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2550 struct task_struct *next)
2552 struct preempt_notifier *notifier;
2554 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2555 notifier->ops->sched_out(notifier, next);
2558 static __always_inline void
2559 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2560 struct task_struct *next)
2562 if (static_key_false(&preempt_notifier_key))
2563 __fire_sched_out_preempt_notifiers(curr, next);
2566 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2568 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2573 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2574 struct task_struct *next)
2578 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2581 * prepare_task_switch - prepare to switch tasks
2582 * @rq: the runqueue preparing to switch
2583 * @prev: the current task that is being switched out
2584 * @next: the task we are going to switch to.
2586 * This is called with the rq lock held and interrupts off. It must
2587 * be paired with a subsequent finish_task_switch after the context
2590 * prepare_task_switch sets up locking and calls architecture specific
2594 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2595 struct task_struct *next)
2597 sched_info_switch(rq, prev, next);
2598 perf_event_task_sched_out(prev, next);
2599 fire_sched_out_preempt_notifiers(prev, next);
2600 prepare_lock_switch(rq, next);
2601 prepare_arch_switch(next);
2605 * finish_task_switch - clean up after a task-switch
2606 * @prev: the thread we just switched away from.
2608 * finish_task_switch must be called after the context switch, paired
2609 * with a prepare_task_switch call before the context switch.
2610 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2611 * and do any other architecture-specific cleanup actions.
2613 * Note that we may have delayed dropping an mm in context_switch(). If
2614 * so, we finish that here outside of the runqueue lock. (Doing it
2615 * with the lock held can cause deadlocks; see schedule() for
2618 * The context switch have flipped the stack from under us and restored the
2619 * local variables which were saved when this task called schedule() in the
2620 * past. prev == current is still correct but we need to recalculate this_rq
2621 * because prev may have moved to another CPU.
2623 static struct rq *finish_task_switch(struct task_struct *prev)
2624 __releases(rq->lock)
2626 struct rq *rq = this_rq();
2627 struct mm_struct *mm = rq->prev_mm;
2631 * The previous task will have left us with a preempt_count of 2
2632 * because it left us after:
2635 * preempt_disable(); // 1
2637 * raw_spin_lock_irq(&rq->lock) // 2
2639 * Also, see FORK_PREEMPT_COUNT.
2641 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2642 "corrupted preempt_count: %s/%d/0x%x\n",
2643 current->comm, current->pid, preempt_count()))
2644 preempt_count_set(FORK_PREEMPT_COUNT);
2649 * A task struct has one reference for the use as "current".
2650 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2651 * schedule one last time. The schedule call will never return, and
2652 * the scheduled task must drop that reference.
2654 * We must observe prev->state before clearing prev->on_cpu (in
2655 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2656 * running on another CPU and we could rave with its RUNNING -> DEAD
2657 * transition, resulting in a double drop.
2659 prev_state = prev->state;
2660 vtime_task_switch(prev);
2661 perf_event_task_sched_in(prev, current);
2662 finish_lock_switch(rq, prev);
2663 finish_arch_post_lock_switch();
2665 fire_sched_in_preempt_notifiers(current);
2668 if (unlikely(prev_state == TASK_DEAD)) {
2669 if (prev->sched_class->task_dead)
2670 prev->sched_class->task_dead(prev);
2673 * Remove function-return probe instances associated with this
2674 * task and put them back on the free list.
2676 kprobe_flush_task(prev);
2677 put_task_struct(prev);
2680 tick_nohz_task_switch();
2686 /* rq->lock is NOT held, but preemption is disabled */
2687 static void __balance_callback(struct rq *rq)
2689 struct callback_head *head, *next;
2690 void (*func)(struct rq *rq);
2691 unsigned long flags;
2693 raw_spin_lock_irqsave(&rq->lock, flags);
2694 head = rq->balance_callback;
2695 rq->balance_callback = NULL;
2697 func = (void (*)(struct rq *))head->func;
2704 raw_spin_unlock_irqrestore(&rq->lock, flags);
2707 static inline void balance_callback(struct rq *rq)
2709 if (unlikely(rq->balance_callback))
2710 __balance_callback(rq);
2715 static inline void balance_callback(struct rq *rq)
2722 * schedule_tail - first thing a freshly forked thread must call.
2723 * @prev: the thread we just switched away from.
2725 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2726 __releases(rq->lock)
2731 * New tasks start with FORK_PREEMPT_COUNT, see there and
2732 * finish_task_switch() for details.
2734 * finish_task_switch() will drop rq->lock() and lower preempt_count
2735 * and the preempt_enable() will end up enabling preemption (on
2736 * PREEMPT_COUNT kernels).
2739 rq = finish_task_switch(prev);
2740 balance_callback(rq);
2743 if (current->set_child_tid)
2744 put_user(task_pid_vnr(current), current->set_child_tid);
2748 * context_switch - switch to the new MM and the new thread's register state.
2750 static inline struct rq *
2751 context_switch(struct rq *rq, struct task_struct *prev,
2752 struct task_struct *next)
2754 struct mm_struct *mm, *oldmm;
2756 prepare_task_switch(rq, prev, next);
2759 oldmm = prev->active_mm;
2761 * For paravirt, this is coupled with an exit in switch_to to
2762 * combine the page table reload and the switch backend into
2765 arch_start_context_switch(prev);
2768 next->active_mm = oldmm;
2769 atomic_inc(&oldmm->mm_count);
2770 enter_lazy_tlb(oldmm, next);
2772 switch_mm(oldmm, mm, next);
2775 prev->active_mm = NULL;
2776 rq->prev_mm = oldmm;
2779 * Since the runqueue lock will be released by the next
2780 * task (which is an invalid locking op but in the case
2781 * of the scheduler it's an obvious special-case), so we
2782 * do an early lockdep release here:
2784 lockdep_unpin_lock(&rq->lock);
2785 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2787 /* Here we just switch the register state and the stack. */
2788 switch_to(prev, next, prev);
2791 return finish_task_switch(prev);
2795 * nr_running and nr_context_switches:
2797 * externally visible scheduler statistics: current number of runnable
2798 * threads, total number of context switches performed since bootup.
2800 unsigned long nr_running(void)
2802 unsigned long i, sum = 0;
2804 for_each_online_cpu(i)
2805 sum += cpu_rq(i)->nr_running;
2811 * Check if only the current task is running on the cpu.
2813 * Caution: this function does not check that the caller has disabled
2814 * preemption, thus the result might have a time-of-check-to-time-of-use
2815 * race. The caller is responsible to use it correctly, for example:
2817 * - from a non-preemptable section (of course)
2819 * - from a thread that is bound to a single CPU
2821 * - in a loop with very short iterations (e.g. a polling loop)
2823 bool single_task_running(void)
2825 return raw_rq()->nr_running == 1;
2827 EXPORT_SYMBOL(single_task_running);
2829 unsigned long long nr_context_switches(void)
2832 unsigned long long sum = 0;
2834 for_each_possible_cpu(i)
2835 sum += cpu_rq(i)->nr_switches;
2840 unsigned long nr_iowait(void)
2842 unsigned long i, sum = 0;
2844 for_each_possible_cpu(i)
2845 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2850 unsigned long nr_iowait_cpu(int cpu)
2852 struct rq *this = cpu_rq(cpu);
2853 return atomic_read(&this->nr_iowait);
2856 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2858 struct rq *rq = this_rq();
2859 *nr_waiters = atomic_read(&rq->nr_iowait);
2860 *load = rq->load.weight;
2866 * sched_exec - execve() is a valuable balancing opportunity, because at
2867 * this point the task has the smallest effective memory and cache footprint.
2869 void sched_exec(void)
2871 struct task_struct *p = current;
2872 unsigned long flags;
2875 raw_spin_lock_irqsave(&p->pi_lock, flags);
2876 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2877 if (dest_cpu == smp_processor_id())
2880 if (likely(cpu_active(dest_cpu))) {
2881 struct migration_arg arg = { p, dest_cpu };
2883 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2884 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2888 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2893 DEFINE_PER_CPU(struct kernel_stat, kstat);
2894 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2896 EXPORT_PER_CPU_SYMBOL(kstat);
2897 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2900 * Return accounted runtime for the task.
2901 * In case the task is currently running, return the runtime plus current's
2902 * pending runtime that have not been accounted yet.
2904 unsigned long long task_sched_runtime(struct task_struct *p)
2906 unsigned long flags;
2910 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2912 * 64-bit doesn't need locks to atomically read a 64bit value.
2913 * So we have a optimization chance when the task's delta_exec is 0.
2914 * Reading ->on_cpu is racy, but this is ok.
2916 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2917 * If we race with it entering cpu, unaccounted time is 0. This is
2918 * indistinguishable from the read occurring a few cycles earlier.
2919 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2920 * been accounted, so we're correct here as well.
2922 if (!p->on_cpu || !task_on_rq_queued(p))
2923 return p->se.sum_exec_runtime;
2926 rq = task_rq_lock(p, &flags);
2928 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2929 * project cycles that may never be accounted to this
2930 * thread, breaking clock_gettime().
2932 if (task_current(rq, p) && task_on_rq_queued(p)) {
2933 update_rq_clock(rq);
2934 p->sched_class->update_curr(rq);
2936 ns = p->se.sum_exec_runtime;
2937 task_rq_unlock(rq, p, &flags);
2943 * This function gets called by the timer code, with HZ frequency.
2944 * We call it with interrupts disabled.
2946 void scheduler_tick(void)
2948 int cpu = smp_processor_id();
2949 struct rq *rq = cpu_rq(cpu);
2950 struct task_struct *curr = rq->curr;
2954 raw_spin_lock(&rq->lock);
2955 update_rq_clock(rq);
2956 curr->sched_class->task_tick(rq, curr, 0);
2957 update_cpu_load_active(rq);
2958 calc_global_load_tick(rq);
2959 raw_spin_unlock(&rq->lock);
2961 perf_event_task_tick();
2964 rq->idle_balance = idle_cpu(cpu);
2965 trigger_load_balance(rq);
2967 rq_last_tick_reset(rq);
2970 #ifdef CONFIG_NO_HZ_FULL
2972 * scheduler_tick_max_deferment
2974 * Keep at least one tick per second when a single
2975 * active task is running because the scheduler doesn't
2976 * yet completely support full dynticks environment.
2978 * This makes sure that uptime, CFS vruntime, load
2979 * balancing, etc... continue to move forward, even
2980 * with a very low granularity.
2982 * Return: Maximum deferment in nanoseconds.
2984 u64 scheduler_tick_max_deferment(void)
2986 struct rq *rq = this_rq();
2987 unsigned long next, now = READ_ONCE(jiffies);
2989 next = rq->last_sched_tick + HZ;
2991 if (time_before_eq(next, now))
2994 return jiffies_to_nsecs(next - now);
2998 notrace unsigned long get_parent_ip(unsigned long addr)
3000 if (in_lock_functions(addr)) {
3001 addr = CALLER_ADDR2;
3002 if (in_lock_functions(addr))
3003 addr = CALLER_ADDR3;
3008 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3009 defined(CONFIG_PREEMPT_TRACER))
3011 void preempt_count_add(int val)
3013 #ifdef CONFIG_DEBUG_PREEMPT
3017 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3020 __preempt_count_add(val);
3021 #ifdef CONFIG_DEBUG_PREEMPT
3023 * Spinlock count overflowing soon?
3025 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3028 if (preempt_count() == val) {
3029 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3030 #ifdef CONFIG_DEBUG_PREEMPT
3031 current->preempt_disable_ip = ip;
3033 trace_preempt_off(CALLER_ADDR0, ip);
3036 EXPORT_SYMBOL(preempt_count_add);
3037 NOKPROBE_SYMBOL(preempt_count_add);
3039 void preempt_count_sub(int val)
3041 #ifdef CONFIG_DEBUG_PREEMPT
3045 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3048 * Is the spinlock portion underflowing?
3050 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3051 !(preempt_count() & PREEMPT_MASK)))
3055 if (preempt_count() == val)
3056 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3057 __preempt_count_sub(val);
3059 EXPORT_SYMBOL(preempt_count_sub);
3060 NOKPROBE_SYMBOL(preempt_count_sub);
3065 * Print scheduling while atomic bug:
3067 static noinline void __schedule_bug(struct task_struct *prev)
3069 if (oops_in_progress)
3072 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3073 prev->comm, prev->pid, preempt_count());
3075 debug_show_held_locks(prev);
3077 if (irqs_disabled())
3078 print_irqtrace_events(prev);
3079 #ifdef CONFIG_DEBUG_PREEMPT
3080 if (in_atomic_preempt_off()) {
3081 pr_err("Preemption disabled at:");
3082 print_ip_sym(current->preempt_disable_ip);
3087 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3091 * Various schedule()-time debugging checks and statistics:
3093 static inline void schedule_debug(struct task_struct *prev)
3095 #ifdef CONFIG_SCHED_STACK_END_CHECK
3096 BUG_ON(task_stack_end_corrupted(prev));
3099 if (unlikely(in_atomic_preempt_off())) {
3100 __schedule_bug(prev);
3101 preempt_count_set(PREEMPT_DISABLED);
3105 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3107 schedstat_inc(this_rq(), sched_count);
3111 * Pick up the highest-prio task:
3113 static inline struct task_struct *
3114 pick_next_task(struct rq *rq, struct task_struct *prev)
3116 const struct sched_class *class = &fair_sched_class;
3117 struct task_struct *p;
3120 * Optimization: we know that if all tasks are in
3121 * the fair class we can call that function directly:
3123 if (likely(prev->sched_class == class &&
3124 rq->nr_running == rq->cfs.h_nr_running)) {
3125 p = fair_sched_class.pick_next_task(rq, prev);
3126 if (unlikely(p == RETRY_TASK))
3129 /* assumes fair_sched_class->next == idle_sched_class */
3131 p = idle_sched_class.pick_next_task(rq, prev);
3137 for_each_class(class) {
3138 p = class->pick_next_task(rq, prev);
3140 if (unlikely(p == RETRY_TASK))
3146 BUG(); /* the idle class will always have a runnable task */
3150 * __schedule() is the main scheduler function.
3152 * The main means of driving the scheduler and thus entering this function are:
3154 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3156 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3157 * paths. For example, see arch/x86/entry_64.S.
3159 * To drive preemption between tasks, the scheduler sets the flag in timer
3160 * interrupt handler scheduler_tick().
3162 * 3. Wakeups don't really cause entry into schedule(). They add a
3163 * task to the run-queue and that's it.
3165 * Now, if the new task added to the run-queue preempts the current
3166 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3167 * called on the nearest possible occasion:
3169 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3171 * - in syscall or exception context, at the next outmost
3172 * preempt_enable(). (this might be as soon as the wake_up()'s
3175 * - in IRQ context, return from interrupt-handler to
3176 * preemptible context
3178 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3181 * - cond_resched() call
3182 * - explicit schedule() call
3183 * - return from syscall or exception to user-space
3184 * - return from interrupt-handler to user-space
3186 * WARNING: must be called with preemption disabled!
3188 static void __sched notrace __schedule(bool preempt)
3190 struct task_struct *prev, *next;
3191 unsigned long *switch_count;
3195 cpu = smp_processor_id();
3200 * do_exit() calls schedule() with preemption disabled as an exception;
3201 * however we must fix that up, otherwise the next task will see an
3202 * inconsistent (higher) preempt count.
3204 * It also avoids the below schedule_debug() test from complaining
3207 if (unlikely(prev->state == TASK_DEAD))
3208 preempt_enable_no_resched_notrace();
3210 schedule_debug(prev);
3212 if (sched_feat(HRTICK))
3215 local_irq_disable();
3216 rcu_note_context_switch();
3219 * Make sure that signal_pending_state()->signal_pending() below
3220 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3221 * done by the caller to avoid the race with signal_wake_up().
3223 smp_mb__before_spinlock();
3224 raw_spin_lock(&rq->lock);
3225 lockdep_pin_lock(&rq->lock);
3227 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3229 switch_count = &prev->nivcsw;
3230 if (!preempt && prev->state) {
3231 if (unlikely(signal_pending_state(prev->state, prev))) {
3232 prev->state = TASK_RUNNING;
3234 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3238 * If a worker went to sleep, notify and ask workqueue
3239 * whether it wants to wake up a task to maintain
3242 if (prev->flags & PF_WQ_WORKER) {
3243 struct task_struct *to_wakeup;
3245 to_wakeup = wq_worker_sleeping(prev, cpu);
3247 try_to_wake_up_local(to_wakeup);
3250 switch_count = &prev->nvcsw;
3253 if (task_on_rq_queued(prev))
3254 update_rq_clock(rq);
3256 next = pick_next_task(rq, prev);
3257 clear_tsk_need_resched(prev);
3258 clear_preempt_need_resched();
3259 rq->clock_skip_update = 0;
3261 if (likely(prev != next)) {
3266 trace_sched_switch(preempt, prev, next);
3267 rq = context_switch(rq, prev, next); /* unlocks the rq */
3270 lockdep_unpin_lock(&rq->lock);
3271 raw_spin_unlock_irq(&rq->lock);
3274 balance_callback(rq);
3277 static inline void sched_submit_work(struct task_struct *tsk)
3279 if (!tsk->state || tsk_is_pi_blocked(tsk))
3282 * If we are going to sleep and we have plugged IO queued,
3283 * make sure to submit it to avoid deadlocks.
3285 if (blk_needs_flush_plug(tsk))
3286 blk_schedule_flush_plug(tsk);
3289 asmlinkage __visible void __sched schedule(void)
3291 struct task_struct *tsk = current;
3293 sched_submit_work(tsk);
3297 sched_preempt_enable_no_resched();
3298 } while (need_resched());
3300 EXPORT_SYMBOL(schedule);
3302 #ifdef CONFIG_CONTEXT_TRACKING
3303 asmlinkage __visible void __sched schedule_user(void)
3306 * If we come here after a random call to set_need_resched(),
3307 * or we have been woken up remotely but the IPI has not yet arrived,
3308 * we haven't yet exited the RCU idle mode. Do it here manually until
3309 * we find a better solution.
3311 * NB: There are buggy callers of this function. Ideally we
3312 * should warn if prev_state != CONTEXT_USER, but that will trigger
3313 * too frequently to make sense yet.
3315 enum ctx_state prev_state = exception_enter();
3317 exception_exit(prev_state);
3322 * schedule_preempt_disabled - called with preemption disabled
3324 * Returns with preemption disabled. Note: preempt_count must be 1
3326 void __sched schedule_preempt_disabled(void)
3328 sched_preempt_enable_no_resched();
3333 static void __sched notrace preempt_schedule_common(void)
3336 preempt_disable_notrace();
3338 preempt_enable_no_resched_notrace();
3341 * Check again in case we missed a preemption opportunity
3342 * between schedule and now.
3344 } while (need_resched());
3347 #ifdef CONFIG_PREEMPT
3349 * this is the entry point to schedule() from in-kernel preemption
3350 * off of preempt_enable. Kernel preemptions off return from interrupt
3351 * occur there and call schedule directly.
3353 asmlinkage __visible void __sched notrace preempt_schedule(void)
3356 * If there is a non-zero preempt_count or interrupts are disabled,
3357 * we do not want to preempt the current task. Just return..
3359 if (likely(!preemptible()))
3362 preempt_schedule_common();
3364 NOKPROBE_SYMBOL(preempt_schedule);
3365 EXPORT_SYMBOL(preempt_schedule);
3368 * preempt_schedule_notrace - preempt_schedule called by tracing
3370 * The tracing infrastructure uses preempt_enable_notrace to prevent
3371 * recursion and tracing preempt enabling caused by the tracing
3372 * infrastructure itself. But as tracing can happen in areas coming
3373 * from userspace or just about to enter userspace, a preempt enable
3374 * can occur before user_exit() is called. This will cause the scheduler
3375 * to be called when the system is still in usermode.
3377 * To prevent this, the preempt_enable_notrace will use this function
3378 * instead of preempt_schedule() to exit user context if needed before
3379 * calling the scheduler.
3381 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3383 enum ctx_state prev_ctx;
3385 if (likely(!preemptible()))
3389 preempt_disable_notrace();
3391 * Needs preempt disabled in case user_exit() is traced
3392 * and the tracer calls preempt_enable_notrace() causing
3393 * an infinite recursion.
3395 prev_ctx = exception_enter();
3397 exception_exit(prev_ctx);
3399 preempt_enable_no_resched_notrace();
3400 } while (need_resched());
3402 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3404 #endif /* CONFIG_PREEMPT */
3407 * this is the entry point to schedule() from kernel preemption
3408 * off of irq context.
3409 * Note, that this is called and return with irqs disabled. This will
3410 * protect us against recursive calling from irq.
3412 asmlinkage __visible void __sched preempt_schedule_irq(void)
3414 enum ctx_state prev_state;
3416 /* Catch callers which need to be fixed */
3417 BUG_ON(preempt_count() || !irqs_disabled());
3419 prev_state = exception_enter();
3425 local_irq_disable();
3426 sched_preempt_enable_no_resched();
3427 } while (need_resched());
3429 exception_exit(prev_state);
3432 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3435 return try_to_wake_up(curr->private, mode, wake_flags);
3437 EXPORT_SYMBOL(default_wake_function);
3439 #ifdef CONFIG_RT_MUTEXES
3442 * rt_mutex_setprio - set the current priority of a task
3444 * @prio: prio value (kernel-internal form)
3446 * This function changes the 'effective' priority of a task. It does
3447 * not touch ->normal_prio like __setscheduler().
3449 * Used by the rt_mutex code to implement priority inheritance
3450 * logic. Call site only calls if the priority of the task changed.
3452 void rt_mutex_setprio(struct task_struct *p, int prio)
3454 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3456 const struct sched_class *prev_class;
3458 BUG_ON(prio > MAX_PRIO);
3460 rq = __task_rq_lock(p);
3463 * Idle task boosting is a nono in general. There is one
3464 * exception, when PREEMPT_RT and NOHZ is active:
3466 * The idle task calls get_next_timer_interrupt() and holds
3467 * the timer wheel base->lock on the CPU and another CPU wants
3468 * to access the timer (probably to cancel it). We can safely
3469 * ignore the boosting request, as the idle CPU runs this code
3470 * with interrupts disabled and will complete the lock
3471 * protected section without being interrupted. So there is no
3472 * real need to boost.
3474 if (unlikely(p == rq->idle)) {
3475 WARN_ON(p != rq->curr);
3476 WARN_ON(p->pi_blocked_on);
3480 trace_sched_pi_setprio(p, prio);
3482 prev_class = p->sched_class;
3483 queued = task_on_rq_queued(p);
3484 running = task_current(rq, p);
3486 dequeue_task(rq, p, DEQUEUE_SAVE);
3488 put_prev_task(rq, p);
3491 * Boosting condition are:
3492 * 1. -rt task is running and holds mutex A
3493 * --> -dl task blocks on mutex A
3495 * 2. -dl task is running and holds mutex A
3496 * --> -dl task blocks on mutex A and could preempt the
3499 if (dl_prio(prio)) {
3500 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3501 if (!dl_prio(p->normal_prio) ||
3502 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3503 p->dl.dl_boosted = 1;
3504 enqueue_flag |= ENQUEUE_REPLENISH;
3506 p->dl.dl_boosted = 0;
3507 p->sched_class = &dl_sched_class;
3508 } else if (rt_prio(prio)) {
3509 if (dl_prio(oldprio))
3510 p->dl.dl_boosted = 0;
3512 enqueue_flag |= ENQUEUE_HEAD;
3513 p->sched_class = &rt_sched_class;
3515 if (dl_prio(oldprio))
3516 p->dl.dl_boosted = 0;
3517 if (rt_prio(oldprio))
3519 p->sched_class = &fair_sched_class;
3525 p->sched_class->set_curr_task(rq);
3527 enqueue_task(rq, p, enqueue_flag);
3529 check_class_changed(rq, p, prev_class, oldprio);
3531 preempt_disable(); /* avoid rq from going away on us */
3532 __task_rq_unlock(rq);
3534 balance_callback(rq);
3539 void set_user_nice(struct task_struct *p, long nice)
3541 int old_prio, delta, queued;
3542 unsigned long flags;
3545 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3548 * We have to be careful, if called from sys_setpriority(),
3549 * the task might be in the middle of scheduling on another CPU.
3551 rq = task_rq_lock(p, &flags);
3553 * The RT priorities are set via sched_setscheduler(), but we still
3554 * allow the 'normal' nice value to be set - but as expected
3555 * it wont have any effect on scheduling until the task is
3556 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3558 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3559 p->static_prio = NICE_TO_PRIO(nice);
3562 queued = task_on_rq_queued(p);
3564 dequeue_task(rq, p, DEQUEUE_SAVE);
3566 p->static_prio = NICE_TO_PRIO(nice);
3569 p->prio = effective_prio(p);
3570 delta = p->prio - old_prio;
3573 enqueue_task(rq, p, ENQUEUE_RESTORE);
3575 * If the task increased its priority or is running and
3576 * lowered its priority, then reschedule its CPU:
3578 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3582 task_rq_unlock(rq, p, &flags);
3584 EXPORT_SYMBOL(set_user_nice);
3587 * can_nice - check if a task can reduce its nice value
3591 int can_nice(const struct task_struct *p, const int nice)
3593 /* convert nice value [19,-20] to rlimit style value [1,40] */
3594 int nice_rlim = nice_to_rlimit(nice);
3596 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3597 capable(CAP_SYS_NICE));
3600 #ifdef __ARCH_WANT_SYS_NICE
3603 * sys_nice - change the priority of the current process.
3604 * @increment: priority increment
3606 * sys_setpriority is a more generic, but much slower function that
3607 * does similar things.
3609 SYSCALL_DEFINE1(nice, int, increment)
3614 * Setpriority might change our priority at the same moment.
3615 * We don't have to worry. Conceptually one call occurs first
3616 * and we have a single winner.
3618 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3619 nice = task_nice(current) + increment;
3621 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3622 if (increment < 0 && !can_nice(current, nice))
3625 retval = security_task_setnice(current, nice);
3629 set_user_nice(current, nice);
3636 * task_prio - return the priority value of a given task.
3637 * @p: the task in question.
3639 * Return: The priority value as seen by users in /proc.
3640 * RT tasks are offset by -200. Normal tasks are centered
3641 * around 0, value goes from -16 to +15.
3643 int task_prio(const struct task_struct *p)
3645 return p->prio - MAX_RT_PRIO;
3649 * idle_cpu - is a given cpu idle currently?
3650 * @cpu: the processor in question.
3652 * Return: 1 if the CPU is currently idle. 0 otherwise.
3654 int idle_cpu(int cpu)
3656 struct rq *rq = cpu_rq(cpu);
3658 if (rq->curr != rq->idle)
3665 if (!llist_empty(&rq->wake_list))
3673 * idle_task - return the idle task for a given cpu.
3674 * @cpu: the processor in question.
3676 * Return: The idle task for the cpu @cpu.
3678 struct task_struct *idle_task(int cpu)
3680 return cpu_rq(cpu)->idle;
3684 * find_process_by_pid - find a process with a matching PID value.
3685 * @pid: the pid in question.
3687 * The task of @pid, if found. %NULL otherwise.
3689 static struct task_struct *find_process_by_pid(pid_t pid)
3691 return pid ? find_task_by_vpid(pid) : current;
3695 * This function initializes the sched_dl_entity of a newly becoming
3696 * SCHED_DEADLINE task.
3698 * Only the static values are considered here, the actual runtime and the
3699 * absolute deadline will be properly calculated when the task is enqueued
3700 * for the first time with its new policy.
3703 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3705 struct sched_dl_entity *dl_se = &p->dl;
3707 dl_se->dl_runtime = attr->sched_runtime;
3708 dl_se->dl_deadline = attr->sched_deadline;
3709 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3710 dl_se->flags = attr->sched_flags;
3711 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3714 * Changing the parameters of a task is 'tricky' and we're not doing
3715 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3717 * What we SHOULD do is delay the bandwidth release until the 0-lag
3718 * point. This would include retaining the task_struct until that time
3719 * and change dl_overflow() to not immediately decrement the current
3722 * Instead we retain the current runtime/deadline and let the new
3723 * parameters take effect after the current reservation period lapses.
3724 * This is safe (albeit pessimistic) because the 0-lag point is always
3725 * before the current scheduling deadline.
3727 * We can still have temporary overloads because we do not delay the
3728 * change in bandwidth until that time; so admission control is
3729 * not on the safe side. It does however guarantee tasks will never
3730 * consume more than promised.
3735 * sched_setparam() passes in -1 for its policy, to let the functions
3736 * it calls know not to change it.
3738 #define SETPARAM_POLICY -1
3740 static void __setscheduler_params(struct task_struct *p,
3741 const struct sched_attr *attr)
3743 int policy = attr->sched_policy;
3745 if (policy == SETPARAM_POLICY)
3750 if (dl_policy(policy))
3751 __setparam_dl(p, attr);
3752 else if (fair_policy(policy))
3753 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3756 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3757 * !rt_policy. Always setting this ensures that things like
3758 * getparam()/getattr() don't report silly values for !rt tasks.
3760 p->rt_priority = attr->sched_priority;
3761 p->normal_prio = normal_prio(p);
3765 /* Actually do priority change: must hold pi & rq lock. */
3766 static void __setscheduler(struct rq *rq, struct task_struct *p,
3767 const struct sched_attr *attr, bool keep_boost)
3769 __setscheduler_params(p, attr);
3772 * Keep a potential priority boosting if called from
3773 * sched_setscheduler().
3776 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3778 p->prio = normal_prio(p);
3780 if (dl_prio(p->prio))
3781 p->sched_class = &dl_sched_class;
3782 else if (rt_prio(p->prio))
3783 p->sched_class = &rt_sched_class;
3785 p->sched_class = &fair_sched_class;
3789 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3791 struct sched_dl_entity *dl_se = &p->dl;
3793 attr->sched_priority = p->rt_priority;
3794 attr->sched_runtime = dl_se->dl_runtime;
3795 attr->sched_deadline = dl_se->dl_deadline;
3796 attr->sched_period = dl_se->dl_period;
3797 attr->sched_flags = dl_se->flags;
3801 * This function validates the new parameters of a -deadline task.
3802 * We ask for the deadline not being zero, and greater or equal
3803 * than the runtime, as well as the period of being zero or
3804 * greater than deadline. Furthermore, we have to be sure that
3805 * user parameters are above the internal resolution of 1us (we
3806 * check sched_runtime only since it is always the smaller one) and
3807 * below 2^63 ns (we have to check both sched_deadline and
3808 * sched_period, as the latter can be zero).
3811 __checkparam_dl(const struct sched_attr *attr)
3814 if (attr->sched_deadline == 0)
3818 * Since we truncate DL_SCALE bits, make sure we're at least
3821 if (attr->sched_runtime < (1ULL << DL_SCALE))
3825 * Since we use the MSB for wrap-around and sign issues, make
3826 * sure it's not set (mind that period can be equal to zero).
3828 if (attr->sched_deadline & (1ULL << 63) ||
3829 attr->sched_period & (1ULL << 63))
3832 /* runtime <= deadline <= period (if period != 0) */
3833 if ((attr->sched_period != 0 &&
3834 attr->sched_period < attr->sched_deadline) ||
3835 attr->sched_deadline < attr->sched_runtime)
3842 * check the target process has a UID that matches the current process's
3844 static bool check_same_owner(struct task_struct *p)
3846 const struct cred *cred = current_cred(), *pcred;
3850 pcred = __task_cred(p);
3851 match = (uid_eq(cred->euid, pcred->euid) ||
3852 uid_eq(cred->euid, pcred->uid));
3857 static bool dl_param_changed(struct task_struct *p,
3858 const struct sched_attr *attr)
3860 struct sched_dl_entity *dl_se = &p->dl;
3862 if (dl_se->dl_runtime != attr->sched_runtime ||
3863 dl_se->dl_deadline != attr->sched_deadline ||
3864 dl_se->dl_period != attr->sched_period ||
3865 dl_se->flags != attr->sched_flags)
3871 static int __sched_setscheduler(struct task_struct *p,
3872 const struct sched_attr *attr,
3875 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3876 MAX_RT_PRIO - 1 - attr->sched_priority;
3877 int retval, oldprio, oldpolicy = -1, queued, running;
3878 int new_effective_prio, policy = attr->sched_policy;
3879 unsigned long flags;
3880 const struct sched_class *prev_class;
3884 /* may grab non-irq protected spin_locks */
3885 BUG_ON(in_interrupt());
3887 /* double check policy once rq lock held */
3889 reset_on_fork = p->sched_reset_on_fork;
3890 policy = oldpolicy = p->policy;
3892 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3894 if (!valid_policy(policy))
3898 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3902 * Valid priorities for SCHED_FIFO and SCHED_RR are
3903 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3904 * SCHED_BATCH and SCHED_IDLE is 0.
3906 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3907 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3909 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3910 (rt_policy(policy) != (attr->sched_priority != 0)))
3914 * Allow unprivileged RT tasks to decrease priority:
3916 if (user && !capable(CAP_SYS_NICE)) {
3917 if (fair_policy(policy)) {
3918 if (attr->sched_nice < task_nice(p) &&
3919 !can_nice(p, attr->sched_nice))
3923 if (rt_policy(policy)) {
3924 unsigned long rlim_rtprio =
3925 task_rlimit(p, RLIMIT_RTPRIO);
3927 /* can't set/change the rt policy */
3928 if (policy != p->policy && !rlim_rtprio)
3931 /* can't increase priority */
3932 if (attr->sched_priority > p->rt_priority &&
3933 attr->sched_priority > rlim_rtprio)
3938 * Can't set/change SCHED_DEADLINE policy at all for now
3939 * (safest behavior); in the future we would like to allow
3940 * unprivileged DL tasks to increase their relative deadline
3941 * or reduce their runtime (both ways reducing utilization)
3943 if (dl_policy(policy))
3947 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3948 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3950 if (idle_policy(p->policy) && !idle_policy(policy)) {
3951 if (!can_nice(p, task_nice(p)))
3955 /* can't change other user's priorities */
3956 if (!check_same_owner(p))
3959 /* Normal users shall not reset the sched_reset_on_fork flag */
3960 if (p->sched_reset_on_fork && !reset_on_fork)
3965 retval = security_task_setscheduler(p);
3971 * make sure no PI-waiters arrive (or leave) while we are
3972 * changing the priority of the task:
3974 * To be able to change p->policy safely, the appropriate
3975 * runqueue lock must be held.
3977 rq = task_rq_lock(p, &flags);
3980 * Changing the policy of the stop threads its a very bad idea
3982 if (p == rq->stop) {
3983 task_rq_unlock(rq, p, &flags);
3988 * If not changing anything there's no need to proceed further,
3989 * but store a possible modification of reset_on_fork.
3991 if (unlikely(policy == p->policy)) {
3992 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3994 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3996 if (dl_policy(policy) && dl_param_changed(p, attr))
3999 p->sched_reset_on_fork = reset_on_fork;
4000 task_rq_unlock(rq, p, &flags);
4006 #ifdef CONFIG_RT_GROUP_SCHED
4008 * Do not allow realtime tasks into groups that have no runtime
4011 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4012 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4013 !task_group_is_autogroup(task_group(p))) {
4014 task_rq_unlock(rq, p, &flags);
4019 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4020 cpumask_t *span = rq->rd->span;
4023 * Don't allow tasks with an affinity mask smaller than
4024 * the entire root_domain to become SCHED_DEADLINE. We
4025 * will also fail if there's no bandwidth available.
4027 if (!cpumask_subset(span, &p->cpus_allowed) ||
4028 rq->rd->dl_bw.bw == 0) {
4029 task_rq_unlock(rq, p, &flags);
4036 /* recheck policy now with rq lock held */
4037 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4038 policy = oldpolicy = -1;
4039 task_rq_unlock(rq, p, &flags);
4044 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4045 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4048 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4049 task_rq_unlock(rq, p, &flags);
4053 p->sched_reset_on_fork = reset_on_fork;
4058 * Take priority boosted tasks into account. If the new
4059 * effective priority is unchanged, we just store the new
4060 * normal parameters and do not touch the scheduler class and
4061 * the runqueue. This will be done when the task deboost
4064 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4065 if (new_effective_prio == oldprio) {
4066 __setscheduler_params(p, attr);
4067 task_rq_unlock(rq, p, &flags);
4072 queued = task_on_rq_queued(p);
4073 running = task_current(rq, p);
4075 dequeue_task(rq, p, DEQUEUE_SAVE);
4077 put_prev_task(rq, p);
4079 prev_class = p->sched_class;
4080 __setscheduler(rq, p, attr, pi);
4083 p->sched_class->set_curr_task(rq);
4085 int enqueue_flags = ENQUEUE_RESTORE;
4087 * We enqueue to tail when the priority of a task is
4088 * increased (user space view).
4090 if (oldprio <= p->prio)
4091 enqueue_flags |= ENQUEUE_HEAD;
4093 enqueue_task(rq, p, enqueue_flags);
4096 check_class_changed(rq, p, prev_class, oldprio);
4097 preempt_disable(); /* avoid rq from going away on us */
4098 task_rq_unlock(rq, p, &flags);
4101 rt_mutex_adjust_pi(p);
4104 * Run balance callbacks after we've adjusted the PI chain.
4106 balance_callback(rq);
4112 static int _sched_setscheduler(struct task_struct *p, int policy,
4113 const struct sched_param *param, bool check)
4115 struct sched_attr attr = {
4116 .sched_policy = policy,
4117 .sched_priority = param->sched_priority,
4118 .sched_nice = PRIO_TO_NICE(p->static_prio),
4121 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4122 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4123 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4124 policy &= ~SCHED_RESET_ON_FORK;
4125 attr.sched_policy = policy;
4128 return __sched_setscheduler(p, &attr, check, true);
4131 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4132 * @p: the task in question.
4133 * @policy: new policy.
4134 * @param: structure containing the new RT priority.
4136 * Return: 0 on success. An error code otherwise.
4138 * NOTE that the task may be already dead.
4140 int sched_setscheduler(struct task_struct *p, int policy,
4141 const struct sched_param *param)
4143 return _sched_setscheduler(p, policy, param, true);
4145 EXPORT_SYMBOL_GPL(sched_setscheduler);
4147 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4149 return __sched_setscheduler(p, attr, true, true);
4151 EXPORT_SYMBOL_GPL(sched_setattr);
4154 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4155 * @p: the task in question.
4156 * @policy: new policy.
4157 * @param: structure containing the new RT priority.
4159 * Just like sched_setscheduler, only don't bother checking if the
4160 * current context has permission. For example, this is needed in
4161 * stop_machine(): we create temporary high priority worker threads,
4162 * but our caller might not have that capability.
4164 * Return: 0 on success. An error code otherwise.
4166 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4167 const struct sched_param *param)
4169 return _sched_setscheduler(p, policy, param, false);
4171 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4174 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4176 struct sched_param lparam;
4177 struct task_struct *p;
4180 if (!param || pid < 0)
4182 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4187 p = find_process_by_pid(pid);
4189 retval = sched_setscheduler(p, policy, &lparam);
4196 * Mimics kernel/events/core.c perf_copy_attr().
4198 static int sched_copy_attr(struct sched_attr __user *uattr,
4199 struct sched_attr *attr)
4204 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4208 * zero the full structure, so that a short copy will be nice.
4210 memset(attr, 0, sizeof(*attr));
4212 ret = get_user(size, &uattr->size);
4216 if (size > PAGE_SIZE) /* silly large */
4219 if (!size) /* abi compat */
4220 size = SCHED_ATTR_SIZE_VER0;
4222 if (size < SCHED_ATTR_SIZE_VER0)
4226 * If we're handed a bigger struct than we know of,
4227 * ensure all the unknown bits are 0 - i.e. new
4228 * user-space does not rely on any kernel feature
4229 * extensions we dont know about yet.
4231 if (size > sizeof(*attr)) {
4232 unsigned char __user *addr;
4233 unsigned char __user *end;
4236 addr = (void __user *)uattr + sizeof(*attr);
4237 end = (void __user *)uattr + size;
4239 for (; addr < end; addr++) {
4240 ret = get_user(val, addr);
4246 size = sizeof(*attr);
4249 ret = copy_from_user(attr, uattr, size);
4254 * XXX: do we want to be lenient like existing syscalls; or do we want
4255 * to be strict and return an error on out-of-bounds values?
4257 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4262 put_user(sizeof(*attr), &uattr->size);
4267 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4268 * @pid: the pid in question.
4269 * @policy: new policy.
4270 * @param: structure containing the new RT priority.
4272 * Return: 0 on success. An error code otherwise.
4274 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4275 struct sched_param __user *, param)
4277 /* negative values for policy are not valid */
4281 return do_sched_setscheduler(pid, policy, param);
4285 * sys_sched_setparam - set/change the RT priority of a thread
4286 * @pid: the pid in question.
4287 * @param: structure containing the new RT priority.
4289 * Return: 0 on success. An error code otherwise.
4291 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4293 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4297 * sys_sched_setattr - same as above, but with extended sched_attr
4298 * @pid: the pid in question.
4299 * @uattr: structure containing the extended parameters.
4300 * @flags: for future extension.
4302 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4303 unsigned int, flags)
4305 struct sched_attr attr;
4306 struct task_struct *p;
4309 if (!uattr || pid < 0 || flags)
4312 retval = sched_copy_attr(uattr, &attr);
4316 if ((int)attr.sched_policy < 0)
4321 p = find_process_by_pid(pid);
4323 retval = sched_setattr(p, &attr);
4330 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4331 * @pid: the pid in question.
4333 * Return: On success, the policy of the thread. Otherwise, a negative error
4336 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4338 struct task_struct *p;
4346 p = find_process_by_pid(pid);
4348 retval = security_task_getscheduler(p);
4351 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4358 * sys_sched_getparam - get the RT priority of a thread
4359 * @pid: the pid in question.
4360 * @param: structure containing the RT priority.
4362 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4365 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4367 struct sched_param lp = { .sched_priority = 0 };
4368 struct task_struct *p;
4371 if (!param || pid < 0)
4375 p = find_process_by_pid(pid);
4380 retval = security_task_getscheduler(p);
4384 if (task_has_rt_policy(p))
4385 lp.sched_priority = p->rt_priority;
4389 * This one might sleep, we cannot do it with a spinlock held ...
4391 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4400 static int sched_read_attr(struct sched_attr __user *uattr,
4401 struct sched_attr *attr,
4406 if (!access_ok(VERIFY_WRITE, uattr, usize))
4410 * If we're handed a smaller struct than we know of,
4411 * ensure all the unknown bits are 0 - i.e. old
4412 * user-space does not get uncomplete information.
4414 if (usize < sizeof(*attr)) {
4415 unsigned char *addr;
4418 addr = (void *)attr + usize;
4419 end = (void *)attr + sizeof(*attr);
4421 for (; addr < end; addr++) {
4429 ret = copy_to_user(uattr, attr, attr->size);
4437 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4438 * @pid: the pid in question.
4439 * @uattr: structure containing the extended parameters.
4440 * @size: sizeof(attr) for fwd/bwd comp.
4441 * @flags: for future extension.
4443 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4444 unsigned int, size, unsigned int, flags)
4446 struct sched_attr attr = {
4447 .size = sizeof(struct sched_attr),
4449 struct task_struct *p;
4452 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4453 size < SCHED_ATTR_SIZE_VER0 || flags)
4457 p = find_process_by_pid(pid);
4462 retval = security_task_getscheduler(p);
4466 attr.sched_policy = p->policy;
4467 if (p->sched_reset_on_fork)
4468 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4469 if (task_has_dl_policy(p))
4470 __getparam_dl(p, &attr);
4471 else if (task_has_rt_policy(p))
4472 attr.sched_priority = p->rt_priority;
4474 attr.sched_nice = task_nice(p);
4478 retval = sched_read_attr(uattr, &attr, size);
4486 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4488 cpumask_var_t cpus_allowed, new_mask;
4489 struct task_struct *p;
4494 p = find_process_by_pid(pid);
4500 /* Prevent p going away */
4504 if (p->flags & PF_NO_SETAFFINITY) {
4508 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4512 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4514 goto out_free_cpus_allowed;
4517 if (!check_same_owner(p)) {
4519 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4521 goto out_free_new_mask;
4526 retval = security_task_setscheduler(p);
4528 goto out_free_new_mask;
4531 cpuset_cpus_allowed(p, cpus_allowed);
4532 cpumask_and(new_mask, in_mask, cpus_allowed);
4535 * Since bandwidth control happens on root_domain basis,
4536 * if admission test is enabled, we only admit -deadline
4537 * tasks allowed to run on all the CPUs in the task's
4541 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4543 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4546 goto out_free_new_mask;
4552 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4555 cpuset_cpus_allowed(p, cpus_allowed);
4556 if (!cpumask_subset(new_mask, cpus_allowed)) {
4558 * We must have raced with a concurrent cpuset
4559 * update. Just reset the cpus_allowed to the
4560 * cpuset's cpus_allowed
4562 cpumask_copy(new_mask, cpus_allowed);
4567 free_cpumask_var(new_mask);
4568 out_free_cpus_allowed:
4569 free_cpumask_var(cpus_allowed);
4575 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4576 struct cpumask *new_mask)
4578 if (len < cpumask_size())
4579 cpumask_clear(new_mask);
4580 else if (len > cpumask_size())
4581 len = cpumask_size();
4583 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4587 * sys_sched_setaffinity - set the cpu affinity of a process
4588 * @pid: pid of the process
4589 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4590 * @user_mask_ptr: user-space pointer to the new cpu mask
4592 * Return: 0 on success. An error code otherwise.
4594 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4595 unsigned long __user *, user_mask_ptr)
4597 cpumask_var_t new_mask;
4600 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4603 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4605 retval = sched_setaffinity(pid, new_mask);
4606 free_cpumask_var(new_mask);
4610 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4612 struct task_struct *p;
4613 unsigned long flags;
4619 p = find_process_by_pid(pid);
4623 retval = security_task_getscheduler(p);
4627 raw_spin_lock_irqsave(&p->pi_lock, flags);
4628 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4629 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4638 * sys_sched_getaffinity - get the cpu affinity of a process
4639 * @pid: pid of the process
4640 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4641 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4643 * Return: 0 on success. An error code otherwise.
4645 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4646 unsigned long __user *, user_mask_ptr)
4651 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4653 if (len & (sizeof(unsigned long)-1))
4656 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4659 ret = sched_getaffinity(pid, mask);
4661 size_t retlen = min_t(size_t, len, cpumask_size());
4663 if (copy_to_user(user_mask_ptr, mask, retlen))
4668 free_cpumask_var(mask);
4674 * sys_sched_yield - yield the current processor to other threads.
4676 * This function yields the current CPU to other tasks. If there are no
4677 * other threads running on this CPU then this function will return.
4681 SYSCALL_DEFINE0(sched_yield)
4683 struct rq *rq = this_rq_lock();
4685 schedstat_inc(rq, yld_count);
4686 current->sched_class->yield_task(rq);
4689 * Since we are going to call schedule() anyway, there's
4690 * no need to preempt or enable interrupts:
4692 __release(rq->lock);
4693 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4694 do_raw_spin_unlock(&rq->lock);
4695 sched_preempt_enable_no_resched();
4702 int __sched _cond_resched(void)
4704 if (should_resched(0)) {
4705 preempt_schedule_common();
4710 EXPORT_SYMBOL(_cond_resched);
4713 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4714 * call schedule, and on return reacquire the lock.
4716 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4717 * operations here to prevent schedule() from being called twice (once via
4718 * spin_unlock(), once by hand).
4720 int __cond_resched_lock(spinlock_t *lock)
4722 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4725 lockdep_assert_held(lock);
4727 if (spin_needbreak(lock) || resched) {
4730 preempt_schedule_common();
4738 EXPORT_SYMBOL(__cond_resched_lock);
4740 int __sched __cond_resched_softirq(void)
4742 BUG_ON(!in_softirq());
4744 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4746 preempt_schedule_common();
4752 EXPORT_SYMBOL(__cond_resched_softirq);
4755 * yield - yield the current processor to other threads.
4757 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4759 * The scheduler is at all times free to pick the calling task as the most
4760 * eligible task to run, if removing the yield() call from your code breaks
4761 * it, its already broken.
4763 * Typical broken usage is:
4768 * where one assumes that yield() will let 'the other' process run that will
4769 * make event true. If the current task is a SCHED_FIFO task that will never
4770 * happen. Never use yield() as a progress guarantee!!
4772 * If you want to use yield() to wait for something, use wait_event().
4773 * If you want to use yield() to be 'nice' for others, use cond_resched().
4774 * If you still want to use yield(), do not!
4776 void __sched yield(void)
4778 set_current_state(TASK_RUNNING);
4781 EXPORT_SYMBOL(yield);
4784 * yield_to - yield the current processor to another thread in
4785 * your thread group, or accelerate that thread toward the
4786 * processor it's on.
4788 * @preempt: whether task preemption is allowed or not
4790 * It's the caller's job to ensure that the target task struct
4791 * can't go away on us before we can do any checks.
4794 * true (>0) if we indeed boosted the target task.
4795 * false (0) if we failed to boost the target.
4796 * -ESRCH if there's no task to yield to.
4798 int __sched yield_to(struct task_struct *p, bool preempt)
4800 struct task_struct *curr = current;
4801 struct rq *rq, *p_rq;
4802 unsigned long flags;
4805 local_irq_save(flags);
4811 * If we're the only runnable task on the rq and target rq also
4812 * has only one task, there's absolutely no point in yielding.
4814 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4819 double_rq_lock(rq, p_rq);
4820 if (task_rq(p) != p_rq) {
4821 double_rq_unlock(rq, p_rq);
4825 if (!curr->sched_class->yield_to_task)
4828 if (curr->sched_class != p->sched_class)
4831 if (task_running(p_rq, p) || p->state)
4834 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4836 schedstat_inc(rq, yld_count);
4838 * Make p's CPU reschedule; pick_next_entity takes care of
4841 if (preempt && rq != p_rq)
4846 double_rq_unlock(rq, p_rq);
4848 local_irq_restore(flags);
4855 EXPORT_SYMBOL_GPL(yield_to);
4858 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4859 * that process accounting knows that this is a task in IO wait state.
4861 long __sched io_schedule_timeout(long timeout)
4863 int old_iowait = current->in_iowait;
4867 current->in_iowait = 1;
4868 blk_schedule_flush_plug(current);
4870 delayacct_blkio_start();
4872 atomic_inc(&rq->nr_iowait);
4873 ret = schedule_timeout(timeout);
4874 current->in_iowait = old_iowait;
4875 atomic_dec(&rq->nr_iowait);
4876 delayacct_blkio_end();
4880 EXPORT_SYMBOL(io_schedule_timeout);
4883 * sys_sched_get_priority_max - return maximum RT priority.
4884 * @policy: scheduling class.
4886 * Return: On success, this syscall returns the maximum
4887 * rt_priority that can be used by a given scheduling class.
4888 * On failure, a negative error code is returned.
4890 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4897 ret = MAX_USER_RT_PRIO-1;
4899 case SCHED_DEADLINE:
4910 * sys_sched_get_priority_min - return minimum RT priority.
4911 * @policy: scheduling class.
4913 * Return: On success, this syscall returns the minimum
4914 * rt_priority that can be used by a given scheduling class.
4915 * On failure, a negative error code is returned.
4917 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4926 case SCHED_DEADLINE:
4936 * sys_sched_rr_get_interval - return the default timeslice of a process.
4937 * @pid: pid of the process.
4938 * @interval: userspace pointer to the timeslice value.
4940 * this syscall writes the default timeslice value of a given process
4941 * into the user-space timespec buffer. A value of '0' means infinity.
4943 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4946 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4947 struct timespec __user *, interval)
4949 struct task_struct *p;
4950 unsigned int time_slice;
4951 unsigned long flags;
4961 p = find_process_by_pid(pid);
4965 retval = security_task_getscheduler(p);
4969 rq = task_rq_lock(p, &flags);
4971 if (p->sched_class->get_rr_interval)
4972 time_slice = p->sched_class->get_rr_interval(rq, p);
4973 task_rq_unlock(rq, p, &flags);
4976 jiffies_to_timespec(time_slice, &t);
4977 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4985 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4987 void sched_show_task(struct task_struct *p)
4989 unsigned long free = 0;
4991 unsigned long state = p->state;
4994 state = __ffs(state) + 1;
4995 printk(KERN_INFO "%-15.15s %c", p->comm,
4996 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4997 #if BITS_PER_LONG == 32
4998 if (state == TASK_RUNNING)
4999 printk(KERN_CONT " running ");
5001 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5003 if (state == TASK_RUNNING)
5004 printk(KERN_CONT " running task ");
5006 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5008 #ifdef CONFIG_DEBUG_STACK_USAGE
5009 free = stack_not_used(p);
5014 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5016 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5017 task_pid_nr(p), ppid,
5018 (unsigned long)task_thread_info(p)->flags);
5020 print_worker_info(KERN_INFO, p);
5021 show_stack(p, NULL);
5024 void show_state_filter(unsigned long state_filter)
5026 struct task_struct *g, *p;
5028 #if BITS_PER_LONG == 32
5030 " task PC stack pid father\n");
5033 " task PC stack pid father\n");
5036 for_each_process_thread(g, p) {
5038 * reset the NMI-timeout, listing all files on a slow
5039 * console might take a lot of time:
5041 touch_nmi_watchdog();
5042 if (!state_filter || (p->state & state_filter))
5046 touch_all_softlockup_watchdogs();
5048 #ifdef CONFIG_SCHED_DEBUG
5049 sysrq_sched_debug_show();
5053 * Only show locks if all tasks are dumped:
5056 debug_show_all_locks();
5059 void init_idle_bootup_task(struct task_struct *idle)
5061 idle->sched_class = &idle_sched_class;
5065 * init_idle - set up an idle thread for a given CPU
5066 * @idle: task in question
5067 * @cpu: cpu the idle task belongs to
5069 * NOTE: this function does not set the idle thread's NEED_RESCHED
5070 * flag, to make booting more robust.
5072 void init_idle(struct task_struct *idle, int cpu)
5074 struct rq *rq = cpu_rq(cpu);
5075 unsigned long flags;
5077 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5078 raw_spin_lock(&rq->lock);
5080 __sched_fork(0, idle);
5081 idle->state = TASK_RUNNING;
5082 idle->se.exec_start = sched_clock();
5086 * Its possible that init_idle() gets called multiple times on a task,
5087 * in that case do_set_cpus_allowed() will not do the right thing.
5089 * And since this is boot we can forgo the serialization.
5091 set_cpus_allowed_common(idle, cpumask_of(cpu));
5094 * We're having a chicken and egg problem, even though we are
5095 * holding rq->lock, the cpu isn't yet set to this cpu so the
5096 * lockdep check in task_group() will fail.
5098 * Similar case to sched_fork(). / Alternatively we could
5099 * use task_rq_lock() here and obtain the other rq->lock.
5104 __set_task_cpu(idle, cpu);
5107 rq->curr = rq->idle = idle;
5108 idle->on_rq = TASK_ON_RQ_QUEUED;
5112 raw_spin_unlock(&rq->lock);
5113 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5115 /* Set the preempt count _outside_ the spinlocks! */
5116 init_idle_preempt_count(idle, cpu);
5119 * The idle tasks have their own, simple scheduling class:
5121 idle->sched_class = &idle_sched_class;
5122 ftrace_graph_init_idle_task(idle, cpu);
5123 vtime_init_idle(idle, cpu);
5125 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5129 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5130 const struct cpumask *trial)
5132 int ret = 1, trial_cpus;
5133 struct dl_bw *cur_dl_b;
5134 unsigned long flags;
5136 if (!cpumask_weight(cur))
5139 rcu_read_lock_sched();
5140 cur_dl_b = dl_bw_of(cpumask_any(cur));
5141 trial_cpus = cpumask_weight(trial);
5143 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5144 if (cur_dl_b->bw != -1 &&
5145 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5147 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5148 rcu_read_unlock_sched();
5153 int task_can_attach(struct task_struct *p,
5154 const struct cpumask *cs_cpus_allowed)
5159 * Kthreads which disallow setaffinity shouldn't be moved
5160 * to a new cpuset; we don't want to change their cpu
5161 * affinity and isolating such threads by their set of
5162 * allowed nodes is unnecessary. Thus, cpusets are not
5163 * applicable for such threads. This prevents checking for
5164 * success of set_cpus_allowed_ptr() on all attached tasks
5165 * before cpus_allowed may be changed.
5167 if (p->flags & PF_NO_SETAFFINITY) {
5173 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5175 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5180 unsigned long flags;
5182 rcu_read_lock_sched();
5183 dl_b = dl_bw_of(dest_cpu);
5184 raw_spin_lock_irqsave(&dl_b->lock, flags);
5185 cpus = dl_bw_cpus(dest_cpu);
5186 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5191 * We reserve space for this task in the destination
5192 * root_domain, as we can't fail after this point.
5193 * We will free resources in the source root_domain
5194 * later on (see set_cpus_allowed_dl()).
5196 __dl_add(dl_b, p->dl.dl_bw);
5198 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5199 rcu_read_unlock_sched();
5209 #ifdef CONFIG_NUMA_BALANCING
5210 /* Migrate current task p to target_cpu */
5211 int migrate_task_to(struct task_struct *p, int target_cpu)
5213 struct migration_arg arg = { p, target_cpu };
5214 int curr_cpu = task_cpu(p);
5216 if (curr_cpu == target_cpu)
5219 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5222 /* TODO: This is not properly updating schedstats */
5224 trace_sched_move_numa(p, curr_cpu, target_cpu);
5225 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5229 * Requeue a task on a given node and accurately track the number of NUMA
5230 * tasks on the runqueues
5232 void sched_setnuma(struct task_struct *p, int nid)
5235 unsigned long flags;
5236 bool queued, running;
5238 rq = task_rq_lock(p, &flags);
5239 queued = task_on_rq_queued(p);
5240 running = task_current(rq, p);
5243 dequeue_task(rq, p, DEQUEUE_SAVE);
5245 put_prev_task(rq, p);
5247 p->numa_preferred_nid = nid;
5250 p->sched_class->set_curr_task(rq);
5252 enqueue_task(rq, p, ENQUEUE_RESTORE);
5253 task_rq_unlock(rq, p, &flags);
5255 #endif /* CONFIG_NUMA_BALANCING */
5257 #ifdef CONFIG_HOTPLUG_CPU
5259 * Ensures that the idle task is using init_mm right before its cpu goes
5262 void idle_task_exit(void)
5264 struct mm_struct *mm = current->active_mm;
5266 BUG_ON(cpu_online(smp_processor_id()));
5268 if (mm != &init_mm) {
5269 switch_mm(mm, &init_mm, current);
5270 finish_arch_post_lock_switch();
5276 * Since this CPU is going 'away' for a while, fold any nr_active delta
5277 * we might have. Assumes we're called after migrate_tasks() so that the
5278 * nr_active count is stable.
5280 * Also see the comment "Global load-average calculations".
5282 static void calc_load_migrate(struct rq *rq)
5284 long delta = calc_load_fold_active(rq);
5286 atomic_long_add(delta, &calc_load_tasks);
5289 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5293 static const struct sched_class fake_sched_class = {
5294 .put_prev_task = put_prev_task_fake,
5297 static struct task_struct fake_task = {
5299 * Avoid pull_{rt,dl}_task()
5301 .prio = MAX_PRIO + 1,
5302 .sched_class = &fake_sched_class,
5306 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5307 * try_to_wake_up()->select_task_rq().
5309 * Called with rq->lock held even though we'er in stop_machine() and
5310 * there's no concurrency possible, we hold the required locks anyway
5311 * because of lock validation efforts.
5313 static void migrate_tasks(struct rq *dead_rq)
5315 struct rq *rq = dead_rq;
5316 struct task_struct *next, *stop = rq->stop;
5320 * Fudge the rq selection such that the below task selection loop
5321 * doesn't get stuck on the currently eligible stop task.
5323 * We're currently inside stop_machine() and the rq is either stuck
5324 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5325 * either way we should never end up calling schedule() until we're
5331 * put_prev_task() and pick_next_task() sched
5332 * class method both need to have an up-to-date
5333 * value of rq->clock[_task]
5335 update_rq_clock(rq);
5339 * There's this thread running, bail when that's the only
5342 if (rq->nr_running == 1)
5346 * pick_next_task assumes pinned rq->lock.
5348 lockdep_pin_lock(&rq->lock);
5349 next = pick_next_task(rq, &fake_task);
5351 next->sched_class->put_prev_task(rq, next);
5354 * Rules for changing task_struct::cpus_allowed are holding
5355 * both pi_lock and rq->lock, such that holding either
5356 * stabilizes the mask.
5358 * Drop rq->lock is not quite as disastrous as it usually is
5359 * because !cpu_active at this point, which means load-balance
5360 * will not interfere. Also, stop-machine.
5362 lockdep_unpin_lock(&rq->lock);
5363 raw_spin_unlock(&rq->lock);
5364 raw_spin_lock(&next->pi_lock);
5365 raw_spin_lock(&rq->lock);
5368 * Since we're inside stop-machine, _nothing_ should have
5369 * changed the task, WARN if weird stuff happened, because in
5370 * that case the above rq->lock drop is a fail too.
5372 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5373 raw_spin_unlock(&next->pi_lock);
5377 /* Find suitable destination for @next, with force if needed. */
5378 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5380 rq = __migrate_task(rq, next, dest_cpu);
5381 if (rq != dead_rq) {
5382 raw_spin_unlock(&rq->lock);
5384 raw_spin_lock(&rq->lock);
5386 raw_spin_unlock(&next->pi_lock);
5391 #endif /* CONFIG_HOTPLUG_CPU */
5393 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5395 static struct ctl_table sd_ctl_dir[] = {
5397 .procname = "sched_domain",
5403 static struct ctl_table sd_ctl_root[] = {
5405 .procname = "kernel",
5407 .child = sd_ctl_dir,
5412 static struct ctl_table *sd_alloc_ctl_entry(int n)
5414 struct ctl_table *entry =
5415 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5420 static void sd_free_ctl_entry(struct ctl_table **tablep)
5422 struct ctl_table *entry;
5425 * In the intermediate directories, both the child directory and
5426 * procname are dynamically allocated and could fail but the mode
5427 * will always be set. In the lowest directory the names are
5428 * static strings and all have proc handlers.
5430 for (entry = *tablep; entry->mode; entry++) {
5432 sd_free_ctl_entry(&entry->child);
5433 if (entry->proc_handler == NULL)
5434 kfree(entry->procname);
5441 static int min_load_idx = 0;
5442 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5445 set_table_entry(struct ctl_table *entry,
5446 const char *procname, void *data, int maxlen,
5447 umode_t mode, proc_handler *proc_handler,
5450 entry->procname = procname;
5452 entry->maxlen = maxlen;
5454 entry->proc_handler = proc_handler;
5457 entry->extra1 = &min_load_idx;
5458 entry->extra2 = &max_load_idx;
5462 static struct ctl_table *
5463 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5465 struct ctl_table *table = sd_alloc_ctl_entry(14);
5470 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5471 sizeof(long), 0644, proc_doulongvec_minmax, false);
5472 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5473 sizeof(long), 0644, proc_doulongvec_minmax, false);
5474 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5475 sizeof(int), 0644, proc_dointvec_minmax, true);
5476 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5477 sizeof(int), 0644, proc_dointvec_minmax, true);
5478 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5479 sizeof(int), 0644, proc_dointvec_minmax, true);
5480 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5481 sizeof(int), 0644, proc_dointvec_minmax, true);
5482 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5483 sizeof(int), 0644, proc_dointvec_minmax, true);
5484 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5485 sizeof(int), 0644, proc_dointvec_minmax, false);
5486 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5487 sizeof(int), 0644, proc_dointvec_minmax, false);
5488 set_table_entry(&table[9], "cache_nice_tries",
5489 &sd->cache_nice_tries,
5490 sizeof(int), 0644, proc_dointvec_minmax, false);
5491 set_table_entry(&table[10], "flags", &sd->flags,
5492 sizeof(int), 0644, proc_dointvec_minmax, false);
5493 set_table_entry(&table[11], "max_newidle_lb_cost",
5494 &sd->max_newidle_lb_cost,
5495 sizeof(long), 0644, proc_doulongvec_minmax, false);
5496 set_table_entry(&table[12], "name", sd->name,
5497 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5498 /* &table[13] is terminator */
5503 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5505 struct ctl_table *entry, *table;
5506 struct sched_domain *sd;
5507 int domain_num = 0, i;
5510 for_each_domain(cpu, sd)
5512 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5517 for_each_domain(cpu, sd) {
5518 snprintf(buf, 32, "domain%d", i);
5519 entry->procname = kstrdup(buf, GFP_KERNEL);
5521 entry->child = sd_alloc_ctl_domain_table(sd);
5528 static struct ctl_table_header *sd_sysctl_header;
5529 static void register_sched_domain_sysctl(void)
5531 int i, cpu_num = num_possible_cpus();
5532 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5535 WARN_ON(sd_ctl_dir[0].child);
5536 sd_ctl_dir[0].child = entry;
5541 for_each_possible_cpu(i) {
5542 snprintf(buf, 32, "cpu%d", i);
5543 entry->procname = kstrdup(buf, GFP_KERNEL);
5545 entry->child = sd_alloc_ctl_cpu_table(i);
5549 WARN_ON(sd_sysctl_header);
5550 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5553 /* may be called multiple times per register */
5554 static void unregister_sched_domain_sysctl(void)
5556 unregister_sysctl_table(sd_sysctl_header);
5557 sd_sysctl_header = NULL;
5558 if (sd_ctl_dir[0].child)
5559 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5562 static void register_sched_domain_sysctl(void)
5565 static void unregister_sched_domain_sysctl(void)
5568 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5570 static void set_rq_online(struct rq *rq)
5573 const struct sched_class *class;
5575 cpumask_set_cpu(rq->cpu, rq->rd->online);
5578 for_each_class(class) {
5579 if (class->rq_online)
5580 class->rq_online(rq);
5585 static void set_rq_offline(struct rq *rq)
5588 const struct sched_class *class;
5590 for_each_class(class) {
5591 if (class->rq_offline)
5592 class->rq_offline(rq);
5595 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5601 * migration_call - callback that gets triggered when a CPU is added.
5602 * Here we can start up the necessary migration thread for the new CPU.
5605 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5607 int cpu = (long)hcpu;
5608 unsigned long flags;
5609 struct rq *rq = cpu_rq(cpu);
5611 switch (action & ~CPU_TASKS_FROZEN) {
5613 case CPU_UP_PREPARE:
5614 rq->calc_load_update = calc_load_update;
5618 /* Update our root-domain */
5619 raw_spin_lock_irqsave(&rq->lock, flags);
5621 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5625 raw_spin_unlock_irqrestore(&rq->lock, flags);
5628 #ifdef CONFIG_HOTPLUG_CPU
5630 sched_ttwu_pending();
5631 /* Update our root-domain */
5632 raw_spin_lock_irqsave(&rq->lock, flags);
5634 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5638 BUG_ON(rq->nr_running != 1); /* the migration thread */
5639 raw_spin_unlock_irqrestore(&rq->lock, flags);
5643 calc_load_migrate(rq);
5648 update_max_interval();
5654 * Register at high priority so that task migration (migrate_all_tasks)
5655 * happens before everything else. This has to be lower priority than
5656 * the notifier in the perf_event subsystem, though.
5658 static struct notifier_block migration_notifier = {
5659 .notifier_call = migration_call,
5660 .priority = CPU_PRI_MIGRATION,
5663 static void set_cpu_rq_start_time(void)
5665 int cpu = smp_processor_id();
5666 struct rq *rq = cpu_rq(cpu);
5667 rq->age_stamp = sched_clock_cpu(cpu);
5670 static int sched_cpu_active(struct notifier_block *nfb,
5671 unsigned long action, void *hcpu)
5673 int cpu = (long)hcpu;
5675 switch (action & ~CPU_TASKS_FROZEN) {
5677 set_cpu_rq_start_time();
5682 * At this point a starting CPU has marked itself as online via
5683 * set_cpu_online(). But it might not yet have marked itself
5684 * as active, which is essential from here on.
5686 set_cpu_active(cpu, true);
5687 stop_machine_unpark(cpu);
5690 case CPU_DOWN_FAILED:
5691 set_cpu_active(cpu, true);
5699 static int sched_cpu_inactive(struct notifier_block *nfb,
5700 unsigned long action, void *hcpu)
5702 switch (action & ~CPU_TASKS_FROZEN) {
5703 case CPU_DOWN_PREPARE:
5704 set_cpu_active((long)hcpu, false);
5711 static int __init migration_init(void)
5713 void *cpu = (void *)(long)smp_processor_id();
5716 /* Initialize migration for the boot CPU */
5717 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5718 BUG_ON(err == NOTIFY_BAD);
5719 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5720 register_cpu_notifier(&migration_notifier);
5722 /* Register cpu active notifiers */
5723 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5724 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5728 early_initcall(migration_init);
5730 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5732 #ifdef CONFIG_SCHED_DEBUG
5734 static __read_mostly int sched_debug_enabled;
5736 static int __init sched_debug_setup(char *str)
5738 sched_debug_enabled = 1;
5742 early_param("sched_debug", sched_debug_setup);
5744 static inline bool sched_debug(void)
5746 return sched_debug_enabled;
5749 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5750 struct cpumask *groupmask)
5752 struct sched_group *group = sd->groups;
5754 cpumask_clear(groupmask);
5756 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5758 if (!(sd->flags & SD_LOAD_BALANCE)) {
5759 printk("does not load-balance\n");
5761 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5766 printk(KERN_CONT "span %*pbl level %s\n",
5767 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5769 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5770 printk(KERN_ERR "ERROR: domain->span does not contain "
5773 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5774 printk(KERN_ERR "ERROR: domain->groups does not contain"
5778 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5782 printk(KERN_ERR "ERROR: group is NULL\n");
5786 if (!cpumask_weight(sched_group_cpus(group))) {
5787 printk(KERN_CONT "\n");
5788 printk(KERN_ERR "ERROR: empty group\n");
5792 if (!(sd->flags & SD_OVERLAP) &&
5793 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5794 printk(KERN_CONT "\n");
5795 printk(KERN_ERR "ERROR: repeated CPUs\n");
5799 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5801 printk(KERN_CONT " %*pbl",
5802 cpumask_pr_args(sched_group_cpus(group)));
5803 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5804 printk(KERN_CONT " (cpu_capacity = %d)",
5805 group->sgc->capacity);
5808 group = group->next;
5809 } while (group != sd->groups);
5810 printk(KERN_CONT "\n");
5812 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5813 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5816 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5817 printk(KERN_ERR "ERROR: parent span is not a superset "
5818 "of domain->span\n");
5822 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5826 if (!sched_debug_enabled)
5830 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5834 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5837 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5845 #else /* !CONFIG_SCHED_DEBUG */
5846 # define sched_domain_debug(sd, cpu) do { } while (0)
5847 static inline bool sched_debug(void)
5851 #endif /* CONFIG_SCHED_DEBUG */
5853 static int sd_degenerate(struct sched_domain *sd)
5855 if (cpumask_weight(sched_domain_span(sd)) == 1)
5858 /* Following flags need at least 2 groups */
5859 if (sd->flags & (SD_LOAD_BALANCE |
5860 SD_BALANCE_NEWIDLE |
5863 SD_SHARE_CPUCAPACITY |
5864 SD_SHARE_PKG_RESOURCES |
5865 SD_SHARE_POWERDOMAIN)) {
5866 if (sd->groups != sd->groups->next)
5870 /* Following flags don't use groups */
5871 if (sd->flags & (SD_WAKE_AFFINE))
5878 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5880 unsigned long cflags = sd->flags, pflags = parent->flags;
5882 if (sd_degenerate(parent))
5885 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5888 /* Flags needing groups don't count if only 1 group in parent */
5889 if (parent->groups == parent->groups->next) {
5890 pflags &= ~(SD_LOAD_BALANCE |
5891 SD_BALANCE_NEWIDLE |
5894 SD_SHARE_CPUCAPACITY |
5895 SD_SHARE_PKG_RESOURCES |
5897 SD_SHARE_POWERDOMAIN);
5898 if (nr_node_ids == 1)
5899 pflags &= ~SD_SERIALIZE;
5901 if (~cflags & pflags)
5907 static void free_rootdomain(struct rcu_head *rcu)
5909 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5911 cpupri_cleanup(&rd->cpupri);
5912 cpudl_cleanup(&rd->cpudl);
5913 free_cpumask_var(rd->dlo_mask);
5914 free_cpumask_var(rd->rto_mask);
5915 free_cpumask_var(rd->online);
5916 free_cpumask_var(rd->span);
5920 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5922 struct root_domain *old_rd = NULL;
5923 unsigned long flags;
5925 raw_spin_lock_irqsave(&rq->lock, flags);
5930 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5933 cpumask_clear_cpu(rq->cpu, old_rd->span);
5936 * If we dont want to free the old_rd yet then
5937 * set old_rd to NULL to skip the freeing later
5940 if (!atomic_dec_and_test(&old_rd->refcount))
5944 atomic_inc(&rd->refcount);
5947 cpumask_set_cpu(rq->cpu, rd->span);
5948 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5951 raw_spin_unlock_irqrestore(&rq->lock, flags);
5954 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5957 static int init_rootdomain(struct root_domain *rd)
5959 memset(rd, 0, sizeof(*rd));
5961 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5963 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5965 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5967 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5970 init_dl_bw(&rd->dl_bw);
5971 if (cpudl_init(&rd->cpudl) != 0)
5974 if (cpupri_init(&rd->cpupri) != 0)
5979 free_cpumask_var(rd->rto_mask);
5981 free_cpumask_var(rd->dlo_mask);
5983 free_cpumask_var(rd->online);
5985 free_cpumask_var(rd->span);
5991 * By default the system creates a single root-domain with all cpus as
5992 * members (mimicking the global state we have today).
5994 struct root_domain def_root_domain;
5996 static void init_defrootdomain(void)
5998 init_rootdomain(&def_root_domain);
6000 atomic_set(&def_root_domain.refcount, 1);
6003 static struct root_domain *alloc_rootdomain(void)
6005 struct root_domain *rd;
6007 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6011 if (init_rootdomain(rd) != 0) {
6019 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6021 struct sched_group *tmp, *first;
6030 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6035 } while (sg != first);
6038 static void free_sched_domain(struct rcu_head *rcu)
6040 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6043 * If its an overlapping domain it has private groups, iterate and
6046 if (sd->flags & SD_OVERLAP) {
6047 free_sched_groups(sd->groups, 1);
6048 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6049 kfree(sd->groups->sgc);
6055 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6057 call_rcu(&sd->rcu, free_sched_domain);
6060 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6062 for (; sd; sd = sd->parent)
6063 destroy_sched_domain(sd, cpu);
6067 * Keep a special pointer to the highest sched_domain that has
6068 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6069 * allows us to avoid some pointer chasing select_idle_sibling().
6071 * Also keep a unique ID per domain (we use the first cpu number in
6072 * the cpumask of the domain), this allows us to quickly tell if
6073 * two cpus are in the same cache domain, see cpus_share_cache().
6075 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6076 DEFINE_PER_CPU(int, sd_llc_size);
6077 DEFINE_PER_CPU(int, sd_llc_id);
6078 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6079 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6080 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6082 static void update_top_cache_domain(int cpu)
6084 struct sched_domain *sd;
6085 struct sched_domain *busy_sd = NULL;
6089 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6091 id = cpumask_first(sched_domain_span(sd));
6092 size = cpumask_weight(sched_domain_span(sd));
6093 busy_sd = sd->parent; /* sd_busy */
6095 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6097 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6098 per_cpu(sd_llc_size, cpu) = size;
6099 per_cpu(sd_llc_id, cpu) = id;
6101 sd = lowest_flag_domain(cpu, SD_NUMA);
6102 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6104 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6105 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6109 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6110 * hold the hotplug lock.
6113 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6115 struct rq *rq = cpu_rq(cpu);
6116 struct sched_domain *tmp;
6118 /* Remove the sched domains which do not contribute to scheduling. */
6119 for (tmp = sd; tmp; ) {
6120 struct sched_domain *parent = tmp->parent;
6124 if (sd_parent_degenerate(tmp, parent)) {
6125 tmp->parent = parent->parent;
6127 parent->parent->child = tmp;
6129 * Transfer SD_PREFER_SIBLING down in case of a
6130 * degenerate parent; the spans match for this
6131 * so the property transfers.
6133 if (parent->flags & SD_PREFER_SIBLING)
6134 tmp->flags |= SD_PREFER_SIBLING;
6135 destroy_sched_domain(parent, cpu);
6140 if (sd && sd_degenerate(sd)) {
6143 destroy_sched_domain(tmp, cpu);
6148 sched_domain_debug(sd, cpu);
6150 rq_attach_root(rq, rd);
6152 rcu_assign_pointer(rq->sd, sd);
6153 destroy_sched_domains(tmp, cpu);
6155 update_top_cache_domain(cpu);
6158 /* Setup the mask of cpus configured for isolated domains */
6159 static int __init isolated_cpu_setup(char *str)
6161 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6162 cpulist_parse(str, cpu_isolated_map);
6166 __setup("isolcpus=", isolated_cpu_setup);
6169 struct sched_domain ** __percpu sd;
6170 struct root_domain *rd;
6181 * Build an iteration mask that can exclude certain CPUs from the upwards
6184 * Asymmetric node setups can result in situations where the domain tree is of
6185 * unequal depth, make sure to skip domains that already cover the entire
6188 * In that case build_sched_domains() will have terminated the iteration early
6189 * and our sibling sd spans will be empty. Domains should always include the
6190 * cpu they're built on, so check that.
6193 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6195 const struct cpumask *span = sched_domain_span(sd);
6196 struct sd_data *sdd = sd->private;
6197 struct sched_domain *sibling;
6200 for_each_cpu(i, span) {
6201 sibling = *per_cpu_ptr(sdd->sd, i);
6202 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6205 cpumask_set_cpu(i, sched_group_mask(sg));
6210 * Return the canonical balance cpu for this group, this is the first cpu
6211 * of this group that's also in the iteration mask.
6213 int group_balance_cpu(struct sched_group *sg)
6215 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6219 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6221 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6222 const struct cpumask *span = sched_domain_span(sd);
6223 struct cpumask *covered = sched_domains_tmpmask;
6224 struct sd_data *sdd = sd->private;
6225 struct sched_domain *sibling;
6228 cpumask_clear(covered);
6230 for_each_cpu(i, span) {
6231 struct cpumask *sg_span;
6233 if (cpumask_test_cpu(i, covered))
6236 sibling = *per_cpu_ptr(sdd->sd, i);
6238 /* See the comment near build_group_mask(). */
6239 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6242 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6243 GFP_KERNEL, cpu_to_node(cpu));
6248 sg_span = sched_group_cpus(sg);
6250 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6252 cpumask_set_cpu(i, sg_span);
6254 cpumask_or(covered, covered, sg_span);
6256 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6257 if (atomic_inc_return(&sg->sgc->ref) == 1)
6258 build_group_mask(sd, sg);
6261 * Initialize sgc->capacity such that even if we mess up the
6262 * domains and no possible iteration will get us here, we won't
6265 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6268 * Make sure the first group of this domain contains the
6269 * canonical balance cpu. Otherwise the sched_domain iteration
6270 * breaks. See update_sg_lb_stats().
6272 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6273 group_balance_cpu(sg) == cpu)
6283 sd->groups = groups;
6288 free_sched_groups(first, 0);
6293 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6295 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6296 struct sched_domain *child = sd->child;
6299 cpu = cpumask_first(sched_domain_span(child));
6302 *sg = *per_cpu_ptr(sdd->sg, cpu);
6303 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6304 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6311 * build_sched_groups will build a circular linked list of the groups
6312 * covered by the given span, and will set each group's ->cpumask correctly,
6313 * and ->cpu_capacity to 0.
6315 * Assumes the sched_domain tree is fully constructed
6318 build_sched_groups(struct sched_domain *sd, int cpu)
6320 struct sched_group *first = NULL, *last = NULL;
6321 struct sd_data *sdd = sd->private;
6322 const struct cpumask *span = sched_domain_span(sd);
6323 struct cpumask *covered;
6326 get_group(cpu, sdd, &sd->groups);
6327 atomic_inc(&sd->groups->ref);
6329 if (cpu != cpumask_first(span))
6332 lockdep_assert_held(&sched_domains_mutex);
6333 covered = sched_domains_tmpmask;
6335 cpumask_clear(covered);
6337 for_each_cpu(i, span) {
6338 struct sched_group *sg;
6341 if (cpumask_test_cpu(i, covered))
6344 group = get_group(i, sdd, &sg);
6345 cpumask_setall(sched_group_mask(sg));
6347 for_each_cpu(j, span) {
6348 if (get_group(j, sdd, NULL) != group)
6351 cpumask_set_cpu(j, covered);
6352 cpumask_set_cpu(j, sched_group_cpus(sg));
6367 * Initialize sched groups cpu_capacity.
6369 * cpu_capacity indicates the capacity of sched group, which is used while
6370 * distributing the load between different sched groups in a sched domain.
6371 * Typically cpu_capacity for all the groups in a sched domain will be same
6372 * unless there are asymmetries in the topology. If there are asymmetries,
6373 * group having more cpu_capacity will pickup more load compared to the
6374 * group having less cpu_capacity.
6376 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6378 struct sched_group *sg = sd->groups;
6383 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6385 } while (sg != sd->groups);
6387 if (cpu != group_balance_cpu(sg))
6390 update_group_capacity(sd, cpu);
6391 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6395 * Initializers for schedule domains
6396 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6399 static int default_relax_domain_level = -1;
6400 int sched_domain_level_max;
6402 static int __init setup_relax_domain_level(char *str)
6404 if (kstrtoint(str, 0, &default_relax_domain_level))
6405 pr_warn("Unable to set relax_domain_level\n");
6409 __setup("relax_domain_level=", setup_relax_domain_level);
6411 static void set_domain_attribute(struct sched_domain *sd,
6412 struct sched_domain_attr *attr)
6416 if (!attr || attr->relax_domain_level < 0) {
6417 if (default_relax_domain_level < 0)
6420 request = default_relax_domain_level;
6422 request = attr->relax_domain_level;
6423 if (request < sd->level) {
6424 /* turn off idle balance on this domain */
6425 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6427 /* turn on idle balance on this domain */
6428 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6432 static void __sdt_free(const struct cpumask *cpu_map);
6433 static int __sdt_alloc(const struct cpumask *cpu_map);
6435 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6436 const struct cpumask *cpu_map)
6440 if (!atomic_read(&d->rd->refcount))
6441 free_rootdomain(&d->rd->rcu); /* fall through */
6443 free_percpu(d->sd); /* fall through */
6445 __sdt_free(cpu_map); /* fall through */
6451 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6452 const struct cpumask *cpu_map)
6454 memset(d, 0, sizeof(*d));
6456 if (__sdt_alloc(cpu_map))
6457 return sa_sd_storage;
6458 d->sd = alloc_percpu(struct sched_domain *);
6460 return sa_sd_storage;
6461 d->rd = alloc_rootdomain();
6464 return sa_rootdomain;
6468 * NULL the sd_data elements we've used to build the sched_domain and
6469 * sched_group structure so that the subsequent __free_domain_allocs()
6470 * will not free the data we're using.
6472 static void claim_allocations(int cpu, struct sched_domain *sd)
6474 struct sd_data *sdd = sd->private;
6476 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6477 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6479 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6480 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6482 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6483 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6487 static int sched_domains_numa_levels;
6488 enum numa_topology_type sched_numa_topology_type;
6489 static int *sched_domains_numa_distance;
6490 int sched_max_numa_distance;
6491 static struct cpumask ***sched_domains_numa_masks;
6492 static int sched_domains_curr_level;
6496 * SD_flags allowed in topology descriptions.
6498 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6499 * SD_SHARE_PKG_RESOURCES - describes shared caches
6500 * SD_NUMA - describes NUMA topologies
6501 * SD_SHARE_POWERDOMAIN - describes shared power domain
6504 * SD_ASYM_PACKING - describes SMT quirks
6506 #define TOPOLOGY_SD_FLAGS \
6507 (SD_SHARE_CPUCAPACITY | \
6508 SD_SHARE_PKG_RESOURCES | \
6511 SD_SHARE_POWERDOMAIN)
6513 static struct sched_domain *
6514 sd_init(struct sched_domain_topology_level *tl, int cpu)
6516 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6517 int sd_weight, sd_flags = 0;
6521 * Ugly hack to pass state to sd_numa_mask()...
6523 sched_domains_curr_level = tl->numa_level;
6526 sd_weight = cpumask_weight(tl->mask(cpu));
6529 sd_flags = (*tl->sd_flags)();
6530 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6531 "wrong sd_flags in topology description\n"))
6532 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6534 *sd = (struct sched_domain){
6535 .min_interval = sd_weight,
6536 .max_interval = 2*sd_weight,
6538 .imbalance_pct = 125,
6540 .cache_nice_tries = 0,
6547 .flags = 1*SD_LOAD_BALANCE
6548 | 1*SD_BALANCE_NEWIDLE
6553 | 0*SD_SHARE_CPUCAPACITY
6554 | 0*SD_SHARE_PKG_RESOURCES
6556 | 0*SD_PREFER_SIBLING
6561 .last_balance = jiffies,
6562 .balance_interval = sd_weight,
6564 .max_newidle_lb_cost = 0,
6565 .next_decay_max_lb_cost = jiffies,
6566 #ifdef CONFIG_SCHED_DEBUG
6572 * Convert topological properties into behaviour.
6575 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6576 sd->flags |= SD_PREFER_SIBLING;
6577 sd->imbalance_pct = 110;
6578 sd->smt_gain = 1178; /* ~15% */
6580 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6581 sd->imbalance_pct = 117;
6582 sd->cache_nice_tries = 1;
6586 } else if (sd->flags & SD_NUMA) {
6587 sd->cache_nice_tries = 2;
6591 sd->flags |= SD_SERIALIZE;
6592 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6593 sd->flags &= ~(SD_BALANCE_EXEC |
6600 sd->flags |= SD_PREFER_SIBLING;
6601 sd->cache_nice_tries = 1;
6606 sd->private = &tl->data;
6612 * Topology list, bottom-up.
6614 static struct sched_domain_topology_level default_topology[] = {
6615 #ifdef CONFIG_SCHED_SMT
6616 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6618 #ifdef CONFIG_SCHED_MC
6619 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6621 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6625 static struct sched_domain_topology_level *sched_domain_topology =
6628 #define for_each_sd_topology(tl) \
6629 for (tl = sched_domain_topology; tl->mask; tl++)
6631 void set_sched_topology(struct sched_domain_topology_level *tl)
6633 sched_domain_topology = tl;
6638 static const struct cpumask *sd_numa_mask(int cpu)
6640 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6643 static void sched_numa_warn(const char *str)
6645 static int done = false;
6653 printk(KERN_WARNING "ERROR: %s\n\n", str);
6655 for (i = 0; i < nr_node_ids; i++) {
6656 printk(KERN_WARNING " ");
6657 for (j = 0; j < nr_node_ids; j++)
6658 printk(KERN_CONT "%02d ", node_distance(i,j));
6659 printk(KERN_CONT "\n");
6661 printk(KERN_WARNING "\n");
6664 bool find_numa_distance(int distance)
6668 if (distance == node_distance(0, 0))
6671 for (i = 0; i < sched_domains_numa_levels; i++) {
6672 if (sched_domains_numa_distance[i] == distance)
6680 * A system can have three types of NUMA topology:
6681 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6682 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6683 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6685 * The difference between a glueless mesh topology and a backplane
6686 * topology lies in whether communication between not directly
6687 * connected nodes goes through intermediary nodes (where programs
6688 * could run), or through backplane controllers. This affects
6689 * placement of programs.
6691 * The type of topology can be discerned with the following tests:
6692 * - If the maximum distance between any nodes is 1 hop, the system
6693 * is directly connected.
6694 * - If for two nodes A and B, located N > 1 hops away from each other,
6695 * there is an intermediary node C, which is < N hops away from both
6696 * nodes A and B, the system is a glueless mesh.
6698 static void init_numa_topology_type(void)
6702 n = sched_max_numa_distance;
6704 if (sched_domains_numa_levels <= 1) {
6705 sched_numa_topology_type = NUMA_DIRECT;
6709 for_each_online_node(a) {
6710 for_each_online_node(b) {
6711 /* Find two nodes furthest removed from each other. */
6712 if (node_distance(a, b) < n)
6715 /* Is there an intermediary node between a and b? */
6716 for_each_online_node(c) {
6717 if (node_distance(a, c) < n &&
6718 node_distance(b, c) < n) {
6719 sched_numa_topology_type =
6725 sched_numa_topology_type = NUMA_BACKPLANE;
6731 static void sched_init_numa(void)
6733 int next_distance, curr_distance = node_distance(0, 0);
6734 struct sched_domain_topology_level *tl;
6738 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6739 if (!sched_domains_numa_distance)
6743 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6744 * unique distances in the node_distance() table.
6746 * Assumes node_distance(0,j) includes all distances in
6747 * node_distance(i,j) in order to avoid cubic time.
6749 next_distance = curr_distance;
6750 for (i = 0; i < nr_node_ids; i++) {
6751 for (j = 0; j < nr_node_ids; j++) {
6752 for (k = 0; k < nr_node_ids; k++) {
6753 int distance = node_distance(i, k);
6755 if (distance > curr_distance &&
6756 (distance < next_distance ||
6757 next_distance == curr_distance))
6758 next_distance = distance;
6761 * While not a strong assumption it would be nice to know
6762 * about cases where if node A is connected to B, B is not
6763 * equally connected to A.
6765 if (sched_debug() && node_distance(k, i) != distance)
6766 sched_numa_warn("Node-distance not symmetric");
6768 if (sched_debug() && i && !find_numa_distance(distance))
6769 sched_numa_warn("Node-0 not representative");
6771 if (next_distance != curr_distance) {
6772 sched_domains_numa_distance[level++] = next_distance;
6773 sched_domains_numa_levels = level;
6774 curr_distance = next_distance;
6779 * In case of sched_debug() we verify the above assumption.
6789 * 'level' contains the number of unique distances, excluding the
6790 * identity distance node_distance(i,i).
6792 * The sched_domains_numa_distance[] array includes the actual distance
6797 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6798 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6799 * the array will contain less then 'level' members. This could be
6800 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6801 * in other functions.
6803 * We reset it to 'level' at the end of this function.
6805 sched_domains_numa_levels = 0;
6807 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6808 if (!sched_domains_numa_masks)
6812 * Now for each level, construct a mask per node which contains all
6813 * cpus of nodes that are that many hops away from us.
6815 for (i = 0; i < level; i++) {
6816 sched_domains_numa_masks[i] =
6817 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6818 if (!sched_domains_numa_masks[i])
6821 for (j = 0; j < nr_node_ids; j++) {
6822 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6826 sched_domains_numa_masks[i][j] = mask;
6828 for (k = 0; k < nr_node_ids; k++) {
6829 if (node_distance(j, k) > sched_domains_numa_distance[i])
6832 cpumask_or(mask, mask, cpumask_of_node(k));
6837 /* Compute default topology size */
6838 for (i = 0; sched_domain_topology[i].mask; i++);
6840 tl = kzalloc((i + level + 1) *
6841 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6846 * Copy the default topology bits..
6848 for (i = 0; sched_domain_topology[i].mask; i++)
6849 tl[i] = sched_domain_topology[i];
6852 * .. and append 'j' levels of NUMA goodness.
6854 for (j = 0; j < level; i++, j++) {
6855 tl[i] = (struct sched_domain_topology_level){
6856 .mask = sd_numa_mask,
6857 .sd_flags = cpu_numa_flags,
6858 .flags = SDTL_OVERLAP,
6864 sched_domain_topology = tl;
6866 sched_domains_numa_levels = level;
6867 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6869 init_numa_topology_type();
6872 static void sched_domains_numa_masks_set(int cpu)
6875 int node = cpu_to_node(cpu);
6877 for (i = 0; i < sched_domains_numa_levels; i++) {
6878 for (j = 0; j < nr_node_ids; j++) {
6879 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6880 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6885 static void sched_domains_numa_masks_clear(int cpu)
6888 for (i = 0; i < sched_domains_numa_levels; i++) {
6889 for (j = 0; j < nr_node_ids; j++)
6890 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6895 * Update sched_domains_numa_masks[level][node] array when new cpus
6898 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6899 unsigned long action,
6902 int cpu = (long)hcpu;
6904 switch (action & ~CPU_TASKS_FROZEN) {
6906 sched_domains_numa_masks_set(cpu);
6910 sched_domains_numa_masks_clear(cpu);
6920 static inline void sched_init_numa(void)
6924 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6925 unsigned long action,
6930 #endif /* CONFIG_NUMA */
6932 static int __sdt_alloc(const struct cpumask *cpu_map)
6934 struct sched_domain_topology_level *tl;
6937 for_each_sd_topology(tl) {
6938 struct sd_data *sdd = &tl->data;
6940 sdd->sd = alloc_percpu(struct sched_domain *);
6944 sdd->sg = alloc_percpu(struct sched_group *);
6948 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6952 for_each_cpu(j, cpu_map) {
6953 struct sched_domain *sd;
6954 struct sched_group *sg;
6955 struct sched_group_capacity *sgc;
6957 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6958 GFP_KERNEL, cpu_to_node(j));
6962 *per_cpu_ptr(sdd->sd, j) = sd;
6964 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6965 GFP_KERNEL, cpu_to_node(j));
6971 *per_cpu_ptr(sdd->sg, j) = sg;
6973 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6974 GFP_KERNEL, cpu_to_node(j));
6978 *per_cpu_ptr(sdd->sgc, j) = sgc;
6985 static void __sdt_free(const struct cpumask *cpu_map)
6987 struct sched_domain_topology_level *tl;
6990 for_each_sd_topology(tl) {
6991 struct sd_data *sdd = &tl->data;
6993 for_each_cpu(j, cpu_map) {
6994 struct sched_domain *sd;
6997 sd = *per_cpu_ptr(sdd->sd, j);
6998 if (sd && (sd->flags & SD_OVERLAP))
6999 free_sched_groups(sd->groups, 0);
7000 kfree(*per_cpu_ptr(sdd->sd, j));
7004 kfree(*per_cpu_ptr(sdd->sg, j));
7006 kfree(*per_cpu_ptr(sdd->sgc, j));
7008 free_percpu(sdd->sd);
7010 free_percpu(sdd->sg);
7012 free_percpu(sdd->sgc);
7017 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7018 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7019 struct sched_domain *child, int cpu)
7021 struct sched_domain *sd = sd_init(tl, cpu);
7025 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7027 sd->level = child->level + 1;
7028 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7032 if (!cpumask_subset(sched_domain_span(child),
7033 sched_domain_span(sd))) {
7034 pr_err("BUG: arch topology borken\n");
7035 #ifdef CONFIG_SCHED_DEBUG
7036 pr_err(" the %s domain not a subset of the %s domain\n",
7037 child->name, sd->name);
7039 /* Fixup, ensure @sd has at least @child cpus. */
7040 cpumask_or(sched_domain_span(sd),
7041 sched_domain_span(sd),
7042 sched_domain_span(child));
7046 set_domain_attribute(sd, attr);
7052 * Build sched domains for a given set of cpus and attach the sched domains
7053 * to the individual cpus
7055 static int build_sched_domains(const struct cpumask *cpu_map,
7056 struct sched_domain_attr *attr)
7058 enum s_alloc alloc_state;
7059 struct sched_domain *sd;
7061 int i, ret = -ENOMEM;
7063 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7064 if (alloc_state != sa_rootdomain)
7067 /* Set up domains for cpus specified by the cpu_map. */
7068 for_each_cpu(i, cpu_map) {
7069 struct sched_domain_topology_level *tl;
7072 for_each_sd_topology(tl) {
7073 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7074 if (tl == sched_domain_topology)
7075 *per_cpu_ptr(d.sd, i) = sd;
7076 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7077 sd->flags |= SD_OVERLAP;
7078 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7083 /* Build the groups for the domains */
7084 for_each_cpu(i, cpu_map) {
7085 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7086 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7087 if (sd->flags & SD_OVERLAP) {
7088 if (build_overlap_sched_groups(sd, i))
7091 if (build_sched_groups(sd, i))
7097 /* Calculate CPU capacity for physical packages and nodes */
7098 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7099 if (!cpumask_test_cpu(i, cpu_map))
7102 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7103 claim_allocations(i, sd);
7104 init_sched_groups_capacity(i, sd);
7108 /* Attach the domains */
7110 for_each_cpu(i, cpu_map) {
7111 sd = *per_cpu_ptr(d.sd, i);
7112 cpu_attach_domain(sd, d.rd, i);
7118 __free_domain_allocs(&d, alloc_state, cpu_map);
7122 static cpumask_var_t *doms_cur; /* current sched domains */
7123 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7124 static struct sched_domain_attr *dattr_cur;
7125 /* attribues of custom domains in 'doms_cur' */
7128 * Special case: If a kmalloc of a doms_cur partition (array of
7129 * cpumask) fails, then fallback to a single sched domain,
7130 * as determined by the single cpumask fallback_doms.
7132 static cpumask_var_t fallback_doms;
7135 * arch_update_cpu_topology lets virtualized architectures update the
7136 * cpu core maps. It is supposed to return 1 if the topology changed
7137 * or 0 if it stayed the same.
7139 int __weak arch_update_cpu_topology(void)
7144 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7147 cpumask_var_t *doms;
7149 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7152 for (i = 0; i < ndoms; i++) {
7153 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7154 free_sched_domains(doms, i);
7161 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7164 for (i = 0; i < ndoms; i++)
7165 free_cpumask_var(doms[i]);
7170 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7171 * For now this just excludes isolated cpus, but could be used to
7172 * exclude other special cases in the future.
7174 static int init_sched_domains(const struct cpumask *cpu_map)
7178 arch_update_cpu_topology();
7180 doms_cur = alloc_sched_domains(ndoms_cur);
7182 doms_cur = &fallback_doms;
7183 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7184 err = build_sched_domains(doms_cur[0], NULL);
7185 register_sched_domain_sysctl();
7191 * Detach sched domains from a group of cpus specified in cpu_map
7192 * These cpus will now be attached to the NULL domain
7194 static void detach_destroy_domains(const struct cpumask *cpu_map)
7199 for_each_cpu(i, cpu_map)
7200 cpu_attach_domain(NULL, &def_root_domain, i);
7204 /* handle null as "default" */
7205 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7206 struct sched_domain_attr *new, int idx_new)
7208 struct sched_domain_attr tmp;
7215 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7216 new ? (new + idx_new) : &tmp,
7217 sizeof(struct sched_domain_attr));
7221 * Partition sched domains as specified by the 'ndoms_new'
7222 * cpumasks in the array doms_new[] of cpumasks. This compares
7223 * doms_new[] to the current sched domain partitioning, doms_cur[].
7224 * It destroys each deleted domain and builds each new domain.
7226 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7227 * The masks don't intersect (don't overlap.) We should setup one
7228 * sched domain for each mask. CPUs not in any of the cpumasks will
7229 * not be load balanced. If the same cpumask appears both in the
7230 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7233 * The passed in 'doms_new' should be allocated using
7234 * alloc_sched_domains. This routine takes ownership of it and will
7235 * free_sched_domains it when done with it. If the caller failed the
7236 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7237 * and partition_sched_domains() will fallback to the single partition
7238 * 'fallback_doms', it also forces the domains to be rebuilt.
7240 * If doms_new == NULL it will be replaced with cpu_online_mask.
7241 * ndoms_new == 0 is a special case for destroying existing domains,
7242 * and it will not create the default domain.
7244 * Call with hotplug lock held
7246 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7247 struct sched_domain_attr *dattr_new)
7252 mutex_lock(&sched_domains_mutex);
7254 /* always unregister in case we don't destroy any domains */
7255 unregister_sched_domain_sysctl();
7257 /* Let architecture update cpu core mappings. */
7258 new_topology = arch_update_cpu_topology();
7260 n = doms_new ? ndoms_new : 0;
7262 /* Destroy deleted domains */
7263 for (i = 0; i < ndoms_cur; i++) {
7264 for (j = 0; j < n && !new_topology; j++) {
7265 if (cpumask_equal(doms_cur[i], doms_new[j])
7266 && dattrs_equal(dattr_cur, i, dattr_new, j))
7269 /* no match - a current sched domain not in new doms_new[] */
7270 detach_destroy_domains(doms_cur[i]);
7276 if (doms_new == NULL) {
7278 doms_new = &fallback_doms;
7279 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7280 WARN_ON_ONCE(dattr_new);
7283 /* Build new domains */
7284 for (i = 0; i < ndoms_new; i++) {
7285 for (j = 0; j < n && !new_topology; j++) {
7286 if (cpumask_equal(doms_new[i], doms_cur[j])
7287 && dattrs_equal(dattr_new, i, dattr_cur, j))
7290 /* no match - add a new doms_new */
7291 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7296 /* Remember the new sched domains */
7297 if (doms_cur != &fallback_doms)
7298 free_sched_domains(doms_cur, ndoms_cur);
7299 kfree(dattr_cur); /* kfree(NULL) is safe */
7300 doms_cur = doms_new;
7301 dattr_cur = dattr_new;
7302 ndoms_cur = ndoms_new;
7304 register_sched_domain_sysctl();
7306 mutex_unlock(&sched_domains_mutex);
7309 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7312 * Update cpusets according to cpu_active mask. If cpusets are
7313 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7314 * around partition_sched_domains().
7316 * If we come here as part of a suspend/resume, don't touch cpusets because we
7317 * want to restore it back to its original state upon resume anyway.
7319 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7323 case CPU_ONLINE_FROZEN:
7324 case CPU_DOWN_FAILED_FROZEN:
7327 * num_cpus_frozen tracks how many CPUs are involved in suspend
7328 * resume sequence. As long as this is not the last online
7329 * operation in the resume sequence, just build a single sched
7330 * domain, ignoring cpusets.
7333 if (likely(num_cpus_frozen)) {
7334 partition_sched_domains(1, NULL, NULL);
7339 * This is the last CPU online operation. So fall through and
7340 * restore the original sched domains by considering the
7341 * cpuset configurations.
7345 cpuset_update_active_cpus(true);
7353 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7356 unsigned long flags;
7357 long cpu = (long)hcpu;
7363 case CPU_DOWN_PREPARE:
7364 rcu_read_lock_sched();
7365 dl_b = dl_bw_of(cpu);
7367 raw_spin_lock_irqsave(&dl_b->lock, flags);
7368 cpus = dl_bw_cpus(cpu);
7369 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7370 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7372 rcu_read_unlock_sched();
7375 return notifier_from_errno(-EBUSY);
7376 cpuset_update_active_cpus(false);
7378 case CPU_DOWN_PREPARE_FROZEN:
7380 partition_sched_domains(1, NULL, NULL);
7388 void __init sched_init_smp(void)
7390 cpumask_var_t non_isolated_cpus;
7392 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7393 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7398 * There's no userspace yet to cause hotplug operations; hence all the
7399 * cpu masks are stable and all blatant races in the below code cannot
7402 mutex_lock(&sched_domains_mutex);
7403 init_sched_domains(cpu_active_mask);
7404 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7405 if (cpumask_empty(non_isolated_cpus))
7406 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7407 mutex_unlock(&sched_domains_mutex);
7409 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7410 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7411 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7415 /* Move init over to a non-isolated CPU */
7416 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7418 sched_init_granularity();
7419 free_cpumask_var(non_isolated_cpus);
7421 init_sched_rt_class();
7422 init_sched_dl_class();
7425 void __init sched_init_smp(void)
7427 sched_init_granularity();
7429 #endif /* CONFIG_SMP */
7431 int in_sched_functions(unsigned long addr)
7433 return in_lock_functions(addr) ||
7434 (addr >= (unsigned long)__sched_text_start
7435 && addr < (unsigned long)__sched_text_end);
7438 #ifdef CONFIG_CGROUP_SCHED
7440 * Default task group.
7441 * Every task in system belongs to this group at bootup.
7443 struct task_group root_task_group;
7444 LIST_HEAD(task_groups);
7447 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7449 void __init sched_init(void)
7452 unsigned long alloc_size = 0, ptr;
7454 #ifdef CONFIG_FAIR_GROUP_SCHED
7455 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7457 #ifdef CONFIG_RT_GROUP_SCHED
7458 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7461 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7463 #ifdef CONFIG_FAIR_GROUP_SCHED
7464 root_task_group.se = (struct sched_entity **)ptr;
7465 ptr += nr_cpu_ids * sizeof(void **);
7467 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7468 ptr += nr_cpu_ids * sizeof(void **);
7470 #endif /* CONFIG_FAIR_GROUP_SCHED */
7471 #ifdef CONFIG_RT_GROUP_SCHED
7472 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7473 ptr += nr_cpu_ids * sizeof(void **);
7475 root_task_group.rt_rq = (struct rt_rq **)ptr;
7476 ptr += nr_cpu_ids * sizeof(void **);
7478 #endif /* CONFIG_RT_GROUP_SCHED */
7480 #ifdef CONFIG_CPUMASK_OFFSTACK
7481 for_each_possible_cpu(i) {
7482 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7483 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7485 #endif /* CONFIG_CPUMASK_OFFSTACK */
7487 init_rt_bandwidth(&def_rt_bandwidth,
7488 global_rt_period(), global_rt_runtime());
7489 init_dl_bandwidth(&def_dl_bandwidth,
7490 global_rt_period(), global_rt_runtime());
7493 init_defrootdomain();
7496 #ifdef CONFIG_RT_GROUP_SCHED
7497 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7498 global_rt_period(), global_rt_runtime());
7499 #endif /* CONFIG_RT_GROUP_SCHED */
7501 #ifdef CONFIG_CGROUP_SCHED
7502 list_add(&root_task_group.list, &task_groups);
7503 INIT_LIST_HEAD(&root_task_group.children);
7504 INIT_LIST_HEAD(&root_task_group.siblings);
7505 autogroup_init(&init_task);
7507 #endif /* CONFIG_CGROUP_SCHED */
7509 for_each_possible_cpu(i) {
7513 raw_spin_lock_init(&rq->lock);
7515 rq->calc_load_active = 0;
7516 rq->calc_load_update = jiffies + LOAD_FREQ;
7517 init_cfs_rq(&rq->cfs);
7518 init_rt_rq(&rq->rt);
7519 init_dl_rq(&rq->dl);
7520 #ifdef CONFIG_FAIR_GROUP_SCHED
7521 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7522 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7524 * How much cpu bandwidth does root_task_group get?
7526 * In case of task-groups formed thr' the cgroup filesystem, it
7527 * gets 100% of the cpu resources in the system. This overall
7528 * system cpu resource is divided among the tasks of
7529 * root_task_group and its child task-groups in a fair manner,
7530 * based on each entity's (task or task-group's) weight
7531 * (se->load.weight).
7533 * In other words, if root_task_group has 10 tasks of weight
7534 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7535 * then A0's share of the cpu resource is:
7537 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7539 * We achieve this by letting root_task_group's tasks sit
7540 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7542 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7543 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7544 #endif /* CONFIG_FAIR_GROUP_SCHED */
7546 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7547 #ifdef CONFIG_RT_GROUP_SCHED
7548 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7551 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7552 rq->cpu_load[j] = 0;
7554 rq->last_load_update_tick = jiffies;
7559 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7560 rq->balance_callback = NULL;
7561 rq->active_balance = 0;
7562 rq->next_balance = jiffies;
7567 rq->avg_idle = 2*sysctl_sched_migration_cost;
7568 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7570 INIT_LIST_HEAD(&rq->cfs_tasks);
7572 rq_attach_root(rq, &def_root_domain);
7573 #ifdef CONFIG_NO_HZ_COMMON
7576 #ifdef CONFIG_NO_HZ_FULL
7577 rq->last_sched_tick = 0;
7581 atomic_set(&rq->nr_iowait, 0);
7584 set_load_weight(&init_task);
7586 #ifdef CONFIG_PREEMPT_NOTIFIERS
7587 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7591 * The boot idle thread does lazy MMU switching as well:
7593 atomic_inc(&init_mm.mm_count);
7594 enter_lazy_tlb(&init_mm, current);
7597 * During early bootup we pretend to be a normal task:
7599 current->sched_class = &fair_sched_class;
7602 * Make us the idle thread. Technically, schedule() should not be
7603 * called from this thread, however somewhere below it might be,
7604 * but because we are the idle thread, we just pick up running again
7605 * when this runqueue becomes "idle".
7607 init_idle(current, smp_processor_id());
7609 calc_load_update = jiffies + LOAD_FREQ;
7612 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7613 /* May be allocated at isolcpus cmdline parse time */
7614 if (cpu_isolated_map == NULL)
7615 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7616 idle_thread_set_boot_cpu();
7617 set_cpu_rq_start_time();
7619 init_sched_fair_class();
7621 scheduler_running = 1;
7624 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7625 static inline int preempt_count_equals(int preempt_offset)
7627 int nested = preempt_count() + rcu_preempt_depth();
7629 return (nested == preempt_offset);
7632 void __might_sleep(const char *file, int line, int preempt_offset)
7635 * Blocking primitives will set (and therefore destroy) current->state,
7636 * since we will exit with TASK_RUNNING make sure we enter with it,
7637 * otherwise we will destroy state.
7639 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7640 "do not call blocking ops when !TASK_RUNNING; "
7641 "state=%lx set at [<%p>] %pS\n",
7643 (void *)current->task_state_change,
7644 (void *)current->task_state_change);
7646 ___might_sleep(file, line, preempt_offset);
7648 EXPORT_SYMBOL(__might_sleep);
7650 void ___might_sleep(const char *file, int line, int preempt_offset)
7652 static unsigned long prev_jiffy; /* ratelimiting */
7654 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7655 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7656 !is_idle_task(current)) ||
7657 system_state != SYSTEM_RUNNING || oops_in_progress)
7659 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7661 prev_jiffy = jiffies;
7664 "BUG: sleeping function called from invalid context at %s:%d\n",
7667 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7668 in_atomic(), irqs_disabled(),
7669 current->pid, current->comm);
7671 if (task_stack_end_corrupted(current))
7672 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7674 debug_show_held_locks(current);
7675 if (irqs_disabled())
7676 print_irqtrace_events(current);
7677 #ifdef CONFIG_DEBUG_PREEMPT
7678 if (!preempt_count_equals(preempt_offset)) {
7679 pr_err("Preemption disabled at:");
7680 print_ip_sym(current->preempt_disable_ip);
7686 EXPORT_SYMBOL(___might_sleep);
7689 #ifdef CONFIG_MAGIC_SYSRQ
7690 void normalize_rt_tasks(void)
7692 struct task_struct *g, *p;
7693 struct sched_attr attr = {
7694 .sched_policy = SCHED_NORMAL,
7697 read_lock(&tasklist_lock);
7698 for_each_process_thread(g, p) {
7700 * Only normalize user tasks:
7702 if (p->flags & PF_KTHREAD)
7705 p->se.exec_start = 0;
7706 #ifdef CONFIG_SCHEDSTATS
7707 p->se.statistics.wait_start = 0;
7708 p->se.statistics.sleep_start = 0;
7709 p->se.statistics.block_start = 0;
7712 if (!dl_task(p) && !rt_task(p)) {
7714 * Renice negative nice level userspace
7717 if (task_nice(p) < 0)
7718 set_user_nice(p, 0);
7722 __sched_setscheduler(p, &attr, false, false);
7724 read_unlock(&tasklist_lock);
7727 #endif /* CONFIG_MAGIC_SYSRQ */
7729 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7731 * These functions are only useful for the IA64 MCA handling, or kdb.
7733 * They can only be called when the whole system has been
7734 * stopped - every CPU needs to be quiescent, and no scheduling
7735 * activity can take place. Using them for anything else would
7736 * be a serious bug, and as a result, they aren't even visible
7737 * under any other configuration.
7741 * curr_task - return the current task for a given cpu.
7742 * @cpu: the processor in question.
7744 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7746 * Return: The current task for @cpu.
7748 struct task_struct *curr_task(int cpu)
7750 return cpu_curr(cpu);
7753 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7757 * set_curr_task - set the current task for a given cpu.
7758 * @cpu: the processor in question.
7759 * @p: the task pointer to set.
7761 * Description: This function must only be used when non-maskable interrupts
7762 * are serviced on a separate stack. It allows the architecture to switch the
7763 * notion of the current task on a cpu in a non-blocking manner. This function
7764 * must be called with all CPU's synchronized, and interrupts disabled, the
7765 * and caller must save the original value of the current task (see
7766 * curr_task() above) and restore that value before reenabling interrupts and
7767 * re-starting the system.
7769 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7771 void set_curr_task(int cpu, struct task_struct *p)
7778 #ifdef CONFIG_CGROUP_SCHED
7779 /* task_group_lock serializes the addition/removal of task groups */
7780 static DEFINE_SPINLOCK(task_group_lock);
7782 static void free_sched_group(struct task_group *tg)
7784 free_fair_sched_group(tg);
7785 free_rt_sched_group(tg);
7790 /* allocate runqueue etc for a new task group */
7791 struct task_group *sched_create_group(struct task_group *parent)
7793 struct task_group *tg;
7795 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7797 return ERR_PTR(-ENOMEM);
7799 if (!alloc_fair_sched_group(tg, parent))
7802 if (!alloc_rt_sched_group(tg, parent))
7808 free_sched_group(tg);
7809 return ERR_PTR(-ENOMEM);
7812 void sched_online_group(struct task_group *tg, struct task_group *parent)
7814 unsigned long flags;
7816 spin_lock_irqsave(&task_group_lock, flags);
7817 list_add_rcu(&tg->list, &task_groups);
7819 WARN_ON(!parent); /* root should already exist */
7821 tg->parent = parent;
7822 INIT_LIST_HEAD(&tg->children);
7823 list_add_rcu(&tg->siblings, &parent->children);
7824 spin_unlock_irqrestore(&task_group_lock, flags);
7827 /* rcu callback to free various structures associated with a task group */
7828 static void free_sched_group_rcu(struct rcu_head *rhp)
7830 /* now it should be safe to free those cfs_rqs */
7831 free_sched_group(container_of(rhp, struct task_group, rcu));
7834 /* Destroy runqueue etc associated with a task group */
7835 void sched_destroy_group(struct task_group *tg)
7837 /* wait for possible concurrent references to cfs_rqs complete */
7838 call_rcu(&tg->rcu, free_sched_group_rcu);
7841 void sched_offline_group(struct task_group *tg)
7843 unsigned long flags;
7846 /* end participation in shares distribution */
7847 for_each_possible_cpu(i)
7848 unregister_fair_sched_group(tg, i);
7850 spin_lock_irqsave(&task_group_lock, flags);
7851 list_del_rcu(&tg->list);
7852 list_del_rcu(&tg->siblings);
7853 spin_unlock_irqrestore(&task_group_lock, flags);
7856 /* change task's runqueue when it moves between groups.
7857 * The caller of this function should have put the task in its new group
7858 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7859 * reflect its new group.
7861 void sched_move_task(struct task_struct *tsk)
7863 struct task_group *tg;
7864 int queued, running;
7865 unsigned long flags;
7868 rq = task_rq_lock(tsk, &flags);
7870 running = task_current(rq, tsk);
7871 queued = task_on_rq_queued(tsk);
7874 dequeue_task(rq, tsk, DEQUEUE_SAVE);
7875 if (unlikely(running))
7876 put_prev_task(rq, tsk);
7879 * All callers are synchronized by task_rq_lock(); we do not use RCU
7880 * which is pointless here. Thus, we pass "true" to task_css_check()
7881 * to prevent lockdep warnings.
7883 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7884 struct task_group, css);
7885 tg = autogroup_task_group(tsk, tg);
7886 tsk->sched_task_group = tg;
7888 #ifdef CONFIG_FAIR_GROUP_SCHED
7889 if (tsk->sched_class->task_move_group)
7890 tsk->sched_class->task_move_group(tsk);
7893 set_task_rq(tsk, task_cpu(tsk));
7895 if (unlikely(running))
7896 tsk->sched_class->set_curr_task(rq);
7898 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
7900 task_rq_unlock(rq, tsk, &flags);
7902 #endif /* CONFIG_CGROUP_SCHED */
7904 #ifdef CONFIG_RT_GROUP_SCHED
7906 * Ensure that the real time constraints are schedulable.
7908 static DEFINE_MUTEX(rt_constraints_mutex);
7910 /* Must be called with tasklist_lock held */
7911 static inline int tg_has_rt_tasks(struct task_group *tg)
7913 struct task_struct *g, *p;
7916 * Autogroups do not have RT tasks; see autogroup_create().
7918 if (task_group_is_autogroup(tg))
7921 for_each_process_thread(g, p) {
7922 if (rt_task(p) && task_group(p) == tg)
7929 struct rt_schedulable_data {
7930 struct task_group *tg;
7935 static int tg_rt_schedulable(struct task_group *tg, void *data)
7937 struct rt_schedulable_data *d = data;
7938 struct task_group *child;
7939 unsigned long total, sum = 0;
7940 u64 period, runtime;
7942 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7943 runtime = tg->rt_bandwidth.rt_runtime;
7946 period = d->rt_period;
7947 runtime = d->rt_runtime;
7951 * Cannot have more runtime than the period.
7953 if (runtime > period && runtime != RUNTIME_INF)
7957 * Ensure we don't starve existing RT tasks.
7959 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7962 total = to_ratio(period, runtime);
7965 * Nobody can have more than the global setting allows.
7967 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7971 * The sum of our children's runtime should not exceed our own.
7973 list_for_each_entry_rcu(child, &tg->children, siblings) {
7974 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7975 runtime = child->rt_bandwidth.rt_runtime;
7977 if (child == d->tg) {
7978 period = d->rt_period;
7979 runtime = d->rt_runtime;
7982 sum += to_ratio(period, runtime);
7991 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7995 struct rt_schedulable_data data = {
7997 .rt_period = period,
7998 .rt_runtime = runtime,
8002 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8008 static int tg_set_rt_bandwidth(struct task_group *tg,
8009 u64 rt_period, u64 rt_runtime)
8014 * Disallowing the root group RT runtime is BAD, it would disallow the
8015 * kernel creating (and or operating) RT threads.
8017 if (tg == &root_task_group && rt_runtime == 0)
8020 /* No period doesn't make any sense. */
8024 mutex_lock(&rt_constraints_mutex);
8025 read_lock(&tasklist_lock);
8026 err = __rt_schedulable(tg, rt_period, rt_runtime);
8030 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8031 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8032 tg->rt_bandwidth.rt_runtime = rt_runtime;
8034 for_each_possible_cpu(i) {
8035 struct rt_rq *rt_rq = tg->rt_rq[i];
8037 raw_spin_lock(&rt_rq->rt_runtime_lock);
8038 rt_rq->rt_runtime = rt_runtime;
8039 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8041 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8043 read_unlock(&tasklist_lock);
8044 mutex_unlock(&rt_constraints_mutex);
8049 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8051 u64 rt_runtime, rt_period;
8053 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8054 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8055 if (rt_runtime_us < 0)
8056 rt_runtime = RUNTIME_INF;
8058 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8061 static long sched_group_rt_runtime(struct task_group *tg)
8065 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8068 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8069 do_div(rt_runtime_us, NSEC_PER_USEC);
8070 return rt_runtime_us;
8073 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8075 u64 rt_runtime, rt_period;
8077 rt_period = rt_period_us * NSEC_PER_USEC;
8078 rt_runtime = tg->rt_bandwidth.rt_runtime;
8080 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8083 static long sched_group_rt_period(struct task_group *tg)
8087 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8088 do_div(rt_period_us, NSEC_PER_USEC);
8089 return rt_period_us;
8091 #endif /* CONFIG_RT_GROUP_SCHED */
8093 #ifdef CONFIG_RT_GROUP_SCHED
8094 static int sched_rt_global_constraints(void)
8098 mutex_lock(&rt_constraints_mutex);
8099 read_lock(&tasklist_lock);
8100 ret = __rt_schedulable(NULL, 0, 0);
8101 read_unlock(&tasklist_lock);
8102 mutex_unlock(&rt_constraints_mutex);
8107 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8109 /* Don't accept realtime tasks when there is no way for them to run */
8110 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8116 #else /* !CONFIG_RT_GROUP_SCHED */
8117 static int sched_rt_global_constraints(void)
8119 unsigned long flags;
8122 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8123 for_each_possible_cpu(i) {
8124 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8126 raw_spin_lock(&rt_rq->rt_runtime_lock);
8127 rt_rq->rt_runtime = global_rt_runtime();
8128 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8130 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8134 #endif /* CONFIG_RT_GROUP_SCHED */
8136 static int sched_dl_global_validate(void)
8138 u64 runtime = global_rt_runtime();
8139 u64 period = global_rt_period();
8140 u64 new_bw = to_ratio(period, runtime);
8143 unsigned long flags;
8146 * Here we want to check the bandwidth not being set to some
8147 * value smaller than the currently allocated bandwidth in
8148 * any of the root_domains.
8150 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8151 * cycling on root_domains... Discussion on different/better
8152 * solutions is welcome!
8154 for_each_possible_cpu(cpu) {
8155 rcu_read_lock_sched();
8156 dl_b = dl_bw_of(cpu);
8158 raw_spin_lock_irqsave(&dl_b->lock, flags);
8159 if (new_bw < dl_b->total_bw)
8161 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8163 rcu_read_unlock_sched();
8172 static void sched_dl_do_global(void)
8177 unsigned long flags;
8179 def_dl_bandwidth.dl_period = global_rt_period();
8180 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8182 if (global_rt_runtime() != RUNTIME_INF)
8183 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8186 * FIXME: As above...
8188 for_each_possible_cpu(cpu) {
8189 rcu_read_lock_sched();
8190 dl_b = dl_bw_of(cpu);
8192 raw_spin_lock_irqsave(&dl_b->lock, flags);
8194 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8196 rcu_read_unlock_sched();
8200 static int sched_rt_global_validate(void)
8202 if (sysctl_sched_rt_period <= 0)
8205 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8206 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8212 static void sched_rt_do_global(void)
8214 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8215 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8218 int sched_rt_handler(struct ctl_table *table, int write,
8219 void __user *buffer, size_t *lenp,
8222 int old_period, old_runtime;
8223 static DEFINE_MUTEX(mutex);
8227 old_period = sysctl_sched_rt_period;
8228 old_runtime = sysctl_sched_rt_runtime;
8230 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8232 if (!ret && write) {
8233 ret = sched_rt_global_validate();
8237 ret = sched_dl_global_validate();
8241 ret = sched_rt_global_constraints();
8245 sched_rt_do_global();
8246 sched_dl_do_global();
8250 sysctl_sched_rt_period = old_period;
8251 sysctl_sched_rt_runtime = old_runtime;
8253 mutex_unlock(&mutex);
8258 int sched_rr_handler(struct ctl_table *table, int write,
8259 void __user *buffer, size_t *lenp,
8263 static DEFINE_MUTEX(mutex);
8266 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8267 /* make sure that internally we keep jiffies */
8268 /* also, writing zero resets timeslice to default */
8269 if (!ret && write) {
8270 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8271 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8273 mutex_unlock(&mutex);
8277 #ifdef CONFIG_CGROUP_SCHED
8279 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8281 return css ? container_of(css, struct task_group, css) : NULL;
8284 static struct cgroup_subsys_state *
8285 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8287 struct task_group *parent = css_tg(parent_css);
8288 struct task_group *tg;
8291 /* This is early initialization for the top cgroup */
8292 return &root_task_group.css;
8295 tg = sched_create_group(parent);
8297 return ERR_PTR(-ENOMEM);
8302 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8304 struct task_group *tg = css_tg(css);
8305 struct task_group *parent = css_tg(css->parent);
8308 sched_online_group(tg, parent);
8312 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8314 struct task_group *tg = css_tg(css);
8316 sched_destroy_group(tg);
8319 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8321 struct task_group *tg = css_tg(css);
8323 sched_offline_group(tg);
8326 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8328 sched_move_task(task);
8331 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8333 struct task_struct *task;
8334 struct cgroup_subsys_state *css;
8336 cgroup_taskset_for_each(task, css, tset) {
8337 #ifdef CONFIG_RT_GROUP_SCHED
8338 if (!sched_rt_can_attach(css_tg(css), task))
8341 /* We don't support RT-tasks being in separate groups */
8342 if (task->sched_class != &fair_sched_class)
8349 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8351 struct task_struct *task;
8352 struct cgroup_subsys_state *css;
8354 cgroup_taskset_for_each(task, css, tset)
8355 sched_move_task(task);
8358 #ifdef CONFIG_FAIR_GROUP_SCHED
8359 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8360 struct cftype *cftype, u64 shareval)
8362 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8365 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8368 struct task_group *tg = css_tg(css);
8370 return (u64) scale_load_down(tg->shares);
8373 #ifdef CONFIG_CFS_BANDWIDTH
8374 static DEFINE_MUTEX(cfs_constraints_mutex);
8376 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8377 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8379 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8381 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8383 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8384 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8386 if (tg == &root_task_group)
8390 * Ensure we have at some amount of bandwidth every period. This is
8391 * to prevent reaching a state of large arrears when throttled via
8392 * entity_tick() resulting in prolonged exit starvation.
8394 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8398 * Likewise, bound things on the otherside by preventing insane quota
8399 * periods. This also allows us to normalize in computing quota
8402 if (period > max_cfs_quota_period)
8406 * Prevent race between setting of cfs_rq->runtime_enabled and
8407 * unthrottle_offline_cfs_rqs().
8410 mutex_lock(&cfs_constraints_mutex);
8411 ret = __cfs_schedulable(tg, period, quota);
8415 runtime_enabled = quota != RUNTIME_INF;
8416 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8418 * If we need to toggle cfs_bandwidth_used, off->on must occur
8419 * before making related changes, and on->off must occur afterwards
8421 if (runtime_enabled && !runtime_was_enabled)
8422 cfs_bandwidth_usage_inc();
8423 raw_spin_lock_irq(&cfs_b->lock);
8424 cfs_b->period = ns_to_ktime(period);
8425 cfs_b->quota = quota;
8427 __refill_cfs_bandwidth_runtime(cfs_b);
8428 /* restart the period timer (if active) to handle new period expiry */
8429 if (runtime_enabled)
8430 start_cfs_bandwidth(cfs_b);
8431 raw_spin_unlock_irq(&cfs_b->lock);
8433 for_each_online_cpu(i) {
8434 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8435 struct rq *rq = cfs_rq->rq;
8437 raw_spin_lock_irq(&rq->lock);
8438 cfs_rq->runtime_enabled = runtime_enabled;
8439 cfs_rq->runtime_remaining = 0;
8441 if (cfs_rq->throttled)
8442 unthrottle_cfs_rq(cfs_rq);
8443 raw_spin_unlock_irq(&rq->lock);
8445 if (runtime_was_enabled && !runtime_enabled)
8446 cfs_bandwidth_usage_dec();
8448 mutex_unlock(&cfs_constraints_mutex);
8454 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8458 period = ktime_to_ns(tg->cfs_bandwidth.period);
8459 if (cfs_quota_us < 0)
8460 quota = RUNTIME_INF;
8462 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8464 return tg_set_cfs_bandwidth(tg, period, quota);
8467 long tg_get_cfs_quota(struct task_group *tg)
8471 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8474 quota_us = tg->cfs_bandwidth.quota;
8475 do_div(quota_us, NSEC_PER_USEC);
8480 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8484 period = (u64)cfs_period_us * NSEC_PER_USEC;
8485 quota = tg->cfs_bandwidth.quota;
8487 return tg_set_cfs_bandwidth(tg, period, quota);
8490 long tg_get_cfs_period(struct task_group *tg)
8494 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8495 do_div(cfs_period_us, NSEC_PER_USEC);
8497 return cfs_period_us;
8500 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8503 return tg_get_cfs_quota(css_tg(css));
8506 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8507 struct cftype *cftype, s64 cfs_quota_us)
8509 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8512 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8515 return tg_get_cfs_period(css_tg(css));
8518 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8519 struct cftype *cftype, u64 cfs_period_us)
8521 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8524 struct cfs_schedulable_data {
8525 struct task_group *tg;
8530 * normalize group quota/period to be quota/max_period
8531 * note: units are usecs
8533 static u64 normalize_cfs_quota(struct task_group *tg,
8534 struct cfs_schedulable_data *d)
8542 period = tg_get_cfs_period(tg);
8543 quota = tg_get_cfs_quota(tg);
8546 /* note: these should typically be equivalent */
8547 if (quota == RUNTIME_INF || quota == -1)
8550 return to_ratio(period, quota);
8553 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8555 struct cfs_schedulable_data *d = data;
8556 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8557 s64 quota = 0, parent_quota = -1;
8560 quota = RUNTIME_INF;
8562 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8564 quota = normalize_cfs_quota(tg, d);
8565 parent_quota = parent_b->hierarchical_quota;
8568 * ensure max(child_quota) <= parent_quota, inherit when no
8571 if (quota == RUNTIME_INF)
8572 quota = parent_quota;
8573 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8576 cfs_b->hierarchical_quota = quota;
8581 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8584 struct cfs_schedulable_data data = {
8590 if (quota != RUNTIME_INF) {
8591 do_div(data.period, NSEC_PER_USEC);
8592 do_div(data.quota, NSEC_PER_USEC);
8596 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8602 static int cpu_stats_show(struct seq_file *sf, void *v)
8604 struct task_group *tg = css_tg(seq_css(sf));
8605 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8607 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8608 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8609 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8613 #endif /* CONFIG_CFS_BANDWIDTH */
8614 #endif /* CONFIG_FAIR_GROUP_SCHED */
8616 #ifdef CONFIG_RT_GROUP_SCHED
8617 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8618 struct cftype *cft, s64 val)
8620 return sched_group_set_rt_runtime(css_tg(css), val);
8623 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8626 return sched_group_rt_runtime(css_tg(css));
8629 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8630 struct cftype *cftype, u64 rt_period_us)
8632 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8635 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8638 return sched_group_rt_period(css_tg(css));
8640 #endif /* CONFIG_RT_GROUP_SCHED */
8642 static struct cftype cpu_files[] = {
8643 #ifdef CONFIG_FAIR_GROUP_SCHED
8646 .read_u64 = cpu_shares_read_u64,
8647 .write_u64 = cpu_shares_write_u64,
8650 #ifdef CONFIG_CFS_BANDWIDTH
8652 .name = "cfs_quota_us",
8653 .read_s64 = cpu_cfs_quota_read_s64,
8654 .write_s64 = cpu_cfs_quota_write_s64,
8657 .name = "cfs_period_us",
8658 .read_u64 = cpu_cfs_period_read_u64,
8659 .write_u64 = cpu_cfs_period_write_u64,
8663 .seq_show = cpu_stats_show,
8666 #ifdef CONFIG_RT_GROUP_SCHED
8668 .name = "rt_runtime_us",
8669 .read_s64 = cpu_rt_runtime_read,
8670 .write_s64 = cpu_rt_runtime_write,
8673 .name = "rt_period_us",
8674 .read_u64 = cpu_rt_period_read_uint,
8675 .write_u64 = cpu_rt_period_write_uint,
8681 struct cgroup_subsys cpu_cgrp_subsys = {
8682 .css_alloc = cpu_cgroup_css_alloc,
8683 .css_free = cpu_cgroup_css_free,
8684 .css_online = cpu_cgroup_css_online,
8685 .css_offline = cpu_cgroup_css_offline,
8686 .fork = cpu_cgroup_fork,
8687 .can_attach = cpu_cgroup_can_attach,
8688 .attach = cpu_cgroup_attach,
8689 .legacy_cftypes = cpu_files,
8693 #endif /* CONFIG_CGROUP_SCHED */
8695 void dump_cpu_task(int cpu)
8697 pr_info("Task dump for CPU %d:\n", cpu);
8698 sched_show_task(cpu_curr(cpu));