4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
96 static void update_rq_clock_task(struct rq *rq, s64 delta);
98 void update_rq_clock(struct rq *rq)
102 lockdep_assert_held(&rq->lock);
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
111 update_rq_clock_task(rq, delta);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug unsigned int sysctl_sched_features =
122 #include "features.h"
127 #ifdef CONFIG_SCHED_DEBUG
128 #define SCHED_FEAT(name, enabled) \
131 static const char * const sched_feat_names[] = {
132 #include "features.h"
137 static int sched_feat_show(struct seq_file *m, void *v)
141 for (i = 0; i < __SCHED_FEAT_NR; i++) {
142 if (!(sysctl_sched_features & (1UL << i)))
144 seq_printf(m, "%s ", sched_feat_names[i]);
151 #ifdef HAVE_JUMP_LABEL
153 #define jump_label_key__true STATIC_KEY_INIT_TRUE
154 #define jump_label_key__false STATIC_KEY_INIT_FALSE
156 #define SCHED_FEAT(name, enabled) \
157 jump_label_key__##enabled ,
159 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
160 #include "features.h"
165 static void sched_feat_disable(int i)
167 static_key_disable(&sched_feat_keys[i]);
170 static void sched_feat_enable(int i)
172 static_key_enable(&sched_feat_keys[i]);
175 static void sched_feat_disable(int i) { };
176 static void sched_feat_enable(int i) { };
177 #endif /* HAVE_JUMP_LABEL */
179 static int sched_feat_set(char *cmp)
184 if (strncmp(cmp, "NO_", 3) == 0) {
189 for (i = 0; i < __SCHED_FEAT_NR; i++) {
190 if (strcmp(cmp, sched_feat_names[i]) == 0) {
192 sysctl_sched_features &= ~(1UL << i);
193 sched_feat_disable(i);
195 sysctl_sched_features |= (1UL << i);
196 sched_feat_enable(i);
206 sched_feat_write(struct file *filp, const char __user *ubuf,
207 size_t cnt, loff_t *ppos)
217 if (copy_from_user(&buf, ubuf, cnt))
223 /* Ensure the static_key remains in a consistent state */
224 inode = file_inode(filp);
225 mutex_lock(&inode->i_mutex);
226 i = sched_feat_set(cmp);
227 mutex_unlock(&inode->i_mutex);
228 if (i == __SCHED_FEAT_NR)
236 static int sched_feat_open(struct inode *inode, struct file *filp)
238 return single_open(filp, sched_feat_show, NULL);
241 static const struct file_operations sched_feat_fops = {
242 .open = sched_feat_open,
243 .write = sched_feat_write,
246 .release = single_release,
249 static __init int sched_init_debug(void)
251 debugfs_create_file("sched_features", 0644, NULL, NULL,
256 late_initcall(sched_init_debug);
257 #endif /* CONFIG_SCHED_DEBUG */
260 * Number of tasks to iterate in a single balance run.
261 * Limited because this is done with IRQs disabled.
263 const_debug unsigned int sysctl_sched_nr_migrate = 32;
266 * period over which we average the RT time consumption, measured
271 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
274 * period over which we measure -rt task cpu usage in us.
277 unsigned int sysctl_sched_rt_period = 1000000;
279 __read_mostly int scheduler_running;
282 * part of the period that we allow rt tasks to run in us.
285 int sysctl_sched_rt_runtime = 950000;
287 /* cpus with isolated domains */
288 cpumask_var_t cpu_isolated_map;
291 * this_rq_lock - lock this runqueue and disable interrupts.
293 static struct rq *this_rq_lock(void)
300 raw_spin_lock(&rq->lock);
305 #ifdef CONFIG_SCHED_HRTICK
307 * Use HR-timers to deliver accurate preemption points.
310 static void hrtick_clear(struct rq *rq)
312 if (hrtimer_active(&rq->hrtick_timer))
313 hrtimer_cancel(&rq->hrtick_timer);
317 * High-resolution timer tick.
318 * Runs from hardirq context with interrupts disabled.
320 static enum hrtimer_restart hrtick(struct hrtimer *timer)
322 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
324 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
326 raw_spin_lock(&rq->lock);
328 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
329 raw_spin_unlock(&rq->lock);
331 return HRTIMER_NORESTART;
336 static void __hrtick_restart(struct rq *rq)
338 struct hrtimer *timer = &rq->hrtick_timer;
340 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
344 * called from hardirq (IPI) context
346 static void __hrtick_start(void *arg)
350 raw_spin_lock(&rq->lock);
351 __hrtick_restart(rq);
352 rq->hrtick_csd_pending = 0;
353 raw_spin_unlock(&rq->lock);
357 * Called to set the hrtick timer state.
359 * called with rq->lock held and irqs disabled
361 void hrtick_start(struct rq *rq, u64 delay)
363 struct hrtimer *timer = &rq->hrtick_timer;
368 * Don't schedule slices shorter than 10000ns, that just
369 * doesn't make sense and can cause timer DoS.
371 delta = max_t(s64, delay, 10000LL);
372 time = ktime_add_ns(timer->base->get_time(), delta);
374 hrtimer_set_expires(timer, time);
376 if (rq == this_rq()) {
377 __hrtick_restart(rq);
378 } else if (!rq->hrtick_csd_pending) {
379 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
380 rq->hrtick_csd_pending = 1;
385 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
387 int cpu = (int)(long)hcpu;
390 case CPU_UP_CANCELED:
391 case CPU_UP_CANCELED_FROZEN:
392 case CPU_DOWN_PREPARE:
393 case CPU_DOWN_PREPARE_FROZEN:
395 case CPU_DEAD_FROZEN:
396 hrtick_clear(cpu_rq(cpu));
403 static __init void init_hrtick(void)
405 hotcpu_notifier(hotplug_hrtick, 0);
409 * Called to set the hrtick timer state.
411 * called with rq->lock held and irqs disabled
413 void hrtick_start(struct rq *rq, u64 delay)
416 * Don't schedule slices shorter than 10000ns, that just
417 * doesn't make sense. Rely on vruntime for fairness.
419 delay = max_t(u64, delay, 10000LL);
420 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
421 HRTIMER_MODE_REL_PINNED);
424 static inline void init_hrtick(void)
427 #endif /* CONFIG_SMP */
429 static void init_rq_hrtick(struct rq *rq)
432 rq->hrtick_csd_pending = 0;
434 rq->hrtick_csd.flags = 0;
435 rq->hrtick_csd.func = __hrtick_start;
436 rq->hrtick_csd.info = rq;
439 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
440 rq->hrtick_timer.function = hrtick;
442 #else /* CONFIG_SCHED_HRTICK */
443 static inline void hrtick_clear(struct rq *rq)
447 static inline void init_rq_hrtick(struct rq *rq)
451 static inline void init_hrtick(void)
454 #endif /* CONFIG_SCHED_HRTICK */
457 * cmpxchg based fetch_or, macro so it works for different integer types
459 #define fetch_or(ptr, val) \
460 ({ typeof(*(ptr)) __old, __val = *(ptr); \
462 __old = cmpxchg((ptr), __val, __val | (val)); \
463 if (__old == __val) \
470 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
472 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
473 * this avoids any races wrt polling state changes and thereby avoids
476 static bool set_nr_and_not_polling(struct task_struct *p)
478 struct thread_info *ti = task_thread_info(p);
479 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
483 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
485 * If this returns true, then the idle task promises to call
486 * sched_ttwu_pending() and reschedule soon.
488 static bool set_nr_if_polling(struct task_struct *p)
490 struct thread_info *ti = task_thread_info(p);
491 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
494 if (!(val & _TIF_POLLING_NRFLAG))
496 if (val & _TIF_NEED_RESCHED)
498 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
507 static bool set_nr_and_not_polling(struct task_struct *p)
509 set_tsk_need_resched(p);
514 static bool set_nr_if_polling(struct task_struct *p)
521 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
523 struct wake_q_node *node = &task->wake_q;
526 * Atomically grab the task, if ->wake_q is !nil already it means
527 * its already queued (either by us or someone else) and will get the
528 * wakeup due to that.
530 * This cmpxchg() implies a full barrier, which pairs with the write
531 * barrier implied by the wakeup in wake_up_list().
533 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
536 get_task_struct(task);
539 * The head is context local, there can be no concurrency.
542 head->lastp = &node->next;
545 void wake_up_q(struct wake_q_head *head)
547 struct wake_q_node *node = head->first;
549 while (node != WAKE_Q_TAIL) {
550 struct task_struct *task;
552 task = container_of(node, struct task_struct, wake_q);
554 /* task can safely be re-inserted now */
556 task->wake_q.next = NULL;
559 * wake_up_process() implies a wmb() to pair with the queueing
560 * in wake_q_add() so as not to miss wakeups.
562 wake_up_process(task);
563 put_task_struct(task);
568 * resched_curr - mark rq's current task 'to be rescheduled now'.
570 * On UP this means the setting of the need_resched flag, on SMP it
571 * might also involve a cross-CPU call to trigger the scheduler on
574 void resched_curr(struct rq *rq)
576 struct task_struct *curr = rq->curr;
579 lockdep_assert_held(&rq->lock);
581 if (test_tsk_need_resched(curr))
586 if (cpu == smp_processor_id()) {
587 set_tsk_need_resched(curr);
588 set_preempt_need_resched();
592 if (set_nr_and_not_polling(curr))
593 smp_send_reschedule(cpu);
595 trace_sched_wake_idle_without_ipi(cpu);
598 void resched_cpu(int cpu)
600 struct rq *rq = cpu_rq(cpu);
603 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
606 raw_spin_unlock_irqrestore(&rq->lock, flags);
610 #ifdef CONFIG_NO_HZ_COMMON
612 * In the semi idle case, use the nearest busy cpu for migrating timers
613 * from an idle cpu. This is good for power-savings.
615 * We don't do similar optimization for completely idle system, as
616 * selecting an idle cpu will add more delays to the timers than intended
617 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
619 int get_nohz_timer_target(void)
621 int i, cpu = smp_processor_id();
622 struct sched_domain *sd;
624 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
628 for_each_domain(cpu, sd) {
629 for_each_cpu(i, sched_domain_span(sd)) {
630 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
637 if (!is_housekeeping_cpu(cpu))
638 cpu = housekeeping_any_cpu();
644 * When add_timer_on() enqueues a timer into the timer wheel of an
645 * idle CPU then this timer might expire before the next timer event
646 * which is scheduled to wake up that CPU. In case of a completely
647 * idle system the next event might even be infinite time into the
648 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
649 * leaves the inner idle loop so the newly added timer is taken into
650 * account when the CPU goes back to idle and evaluates the timer
651 * wheel for the next timer event.
653 static void wake_up_idle_cpu(int cpu)
655 struct rq *rq = cpu_rq(cpu);
657 if (cpu == smp_processor_id())
660 if (set_nr_and_not_polling(rq->idle))
661 smp_send_reschedule(cpu);
663 trace_sched_wake_idle_without_ipi(cpu);
666 static bool wake_up_full_nohz_cpu(int cpu)
669 * We just need the target to call irq_exit() and re-evaluate
670 * the next tick. The nohz full kick at least implies that.
671 * If needed we can still optimize that later with an
674 if (tick_nohz_full_cpu(cpu)) {
675 if (cpu != smp_processor_id() ||
676 tick_nohz_tick_stopped())
677 tick_nohz_full_kick_cpu(cpu);
684 void wake_up_nohz_cpu(int cpu)
686 if (!wake_up_full_nohz_cpu(cpu))
687 wake_up_idle_cpu(cpu);
690 static inline bool got_nohz_idle_kick(void)
692 int cpu = smp_processor_id();
694 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
697 if (idle_cpu(cpu) && !need_resched())
701 * We can't run Idle Load Balance on this CPU for this time so we
702 * cancel it and clear NOHZ_BALANCE_KICK
704 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
708 #else /* CONFIG_NO_HZ_COMMON */
710 static inline bool got_nohz_idle_kick(void)
715 #endif /* CONFIG_NO_HZ_COMMON */
717 #ifdef CONFIG_NO_HZ_FULL
718 bool sched_can_stop_tick(void)
721 * FIFO realtime policy runs the highest priority task. Other runnable
722 * tasks are of a lower priority. The scheduler tick does nothing.
724 if (current->policy == SCHED_FIFO)
728 * Round-robin realtime tasks time slice with other tasks at the same
729 * realtime priority. Is this task the only one at this priority?
731 if (current->policy == SCHED_RR) {
732 struct sched_rt_entity *rt_se = ¤t->rt;
734 return rt_se->run_list.prev == rt_se->run_list.next;
738 * More than one running task need preemption.
739 * nr_running update is assumed to be visible
740 * after IPI is sent from wakers.
742 if (this_rq()->nr_running > 1)
747 #endif /* CONFIG_NO_HZ_FULL */
749 void sched_avg_update(struct rq *rq)
751 s64 period = sched_avg_period();
753 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
755 * Inline assembly required to prevent the compiler
756 * optimising this loop into a divmod call.
757 * See __iter_div_u64_rem() for another example of this.
759 asm("" : "+rm" (rq->age_stamp));
760 rq->age_stamp += period;
765 #endif /* CONFIG_SMP */
767 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
768 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
770 * Iterate task_group tree rooted at *from, calling @down when first entering a
771 * node and @up when leaving it for the final time.
773 * Caller must hold rcu_lock or sufficient equivalent.
775 int walk_tg_tree_from(struct task_group *from,
776 tg_visitor down, tg_visitor up, void *data)
778 struct task_group *parent, *child;
784 ret = (*down)(parent, data);
787 list_for_each_entry_rcu(child, &parent->children, siblings) {
794 ret = (*up)(parent, data);
795 if (ret || parent == from)
799 parent = parent->parent;
806 int tg_nop(struct task_group *tg, void *data)
812 static void set_load_weight(struct task_struct *p)
814 int prio = p->static_prio - MAX_RT_PRIO;
815 struct load_weight *load = &p->se.load;
818 * SCHED_IDLE tasks get minimal weight:
820 if (idle_policy(p->policy)) {
821 load->weight = scale_load(WEIGHT_IDLEPRIO);
822 load->inv_weight = WMULT_IDLEPRIO;
826 load->weight = scale_load(prio_to_weight[prio]);
827 load->inv_weight = prio_to_wmult[prio];
830 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
833 if (!(flags & ENQUEUE_RESTORE))
834 sched_info_queued(rq, p);
835 p->sched_class->enqueue_task(rq, p, flags);
838 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
841 if (!(flags & DEQUEUE_SAVE))
842 sched_info_dequeued(rq, p);
843 p->sched_class->dequeue_task(rq, p, flags);
846 void activate_task(struct rq *rq, struct task_struct *p, int flags)
848 if (task_contributes_to_load(p))
849 rq->nr_uninterruptible--;
851 enqueue_task(rq, p, flags);
854 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
856 if (task_contributes_to_load(p))
857 rq->nr_uninterruptible++;
859 dequeue_task(rq, p, flags);
862 static void update_rq_clock_task(struct rq *rq, s64 delta)
865 * In theory, the compile should just see 0 here, and optimize out the call
866 * to sched_rt_avg_update. But I don't trust it...
868 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
869 s64 steal = 0, irq_delta = 0;
871 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
872 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
875 * Since irq_time is only updated on {soft,}irq_exit, we might run into
876 * this case when a previous update_rq_clock() happened inside a
879 * When this happens, we stop ->clock_task and only update the
880 * prev_irq_time stamp to account for the part that fit, so that a next
881 * update will consume the rest. This ensures ->clock_task is
884 * It does however cause some slight miss-attribution of {soft,}irq
885 * time, a more accurate solution would be to update the irq_time using
886 * the current rq->clock timestamp, except that would require using
889 if (irq_delta > delta)
892 rq->prev_irq_time += irq_delta;
895 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
896 if (static_key_false((¶virt_steal_rq_enabled))) {
897 steal = paravirt_steal_clock(cpu_of(rq));
898 steal -= rq->prev_steal_time_rq;
900 if (unlikely(steal > delta))
903 rq->prev_steal_time_rq += steal;
908 rq->clock_task += delta;
910 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
911 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
912 sched_rt_avg_update(rq, irq_delta + steal);
916 void sched_set_stop_task(int cpu, struct task_struct *stop)
918 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
919 struct task_struct *old_stop = cpu_rq(cpu)->stop;
923 * Make it appear like a SCHED_FIFO task, its something
924 * userspace knows about and won't get confused about.
926 * Also, it will make PI more or less work without too
927 * much confusion -- but then, stop work should not
928 * rely on PI working anyway.
930 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
932 stop->sched_class = &stop_sched_class;
935 cpu_rq(cpu)->stop = stop;
939 * Reset it back to a normal scheduling class so that
940 * it can die in pieces.
942 old_stop->sched_class = &rt_sched_class;
947 * __normal_prio - return the priority that is based on the static prio
949 static inline int __normal_prio(struct task_struct *p)
951 return p->static_prio;
955 * Calculate the expected normal priority: i.e. priority
956 * without taking RT-inheritance into account. Might be
957 * boosted by interactivity modifiers. Changes upon fork,
958 * setprio syscalls, and whenever the interactivity
959 * estimator recalculates.
961 static inline int normal_prio(struct task_struct *p)
965 if (task_has_dl_policy(p))
966 prio = MAX_DL_PRIO-1;
967 else if (task_has_rt_policy(p))
968 prio = MAX_RT_PRIO-1 - p->rt_priority;
970 prio = __normal_prio(p);
975 * Calculate the current priority, i.e. the priority
976 * taken into account by the scheduler. This value might
977 * be boosted by RT tasks, or might be boosted by
978 * interactivity modifiers. Will be RT if the task got
979 * RT-boosted. If not then it returns p->normal_prio.
981 static int effective_prio(struct task_struct *p)
983 p->normal_prio = normal_prio(p);
985 * If we are RT tasks or we were boosted to RT priority,
986 * keep the priority unchanged. Otherwise, update priority
987 * to the normal priority:
989 if (!rt_prio(p->prio))
990 return p->normal_prio;
995 * task_curr - is this task currently executing on a CPU?
996 * @p: the task in question.
998 * Return: 1 if the task is currently executing. 0 otherwise.
1000 inline int task_curr(const struct task_struct *p)
1002 return cpu_curr(task_cpu(p)) == p;
1006 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1007 * use the balance_callback list if you want balancing.
1009 * this means any call to check_class_changed() must be followed by a call to
1010 * balance_callback().
1012 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1013 const struct sched_class *prev_class,
1016 if (prev_class != p->sched_class) {
1017 if (prev_class->switched_from)
1018 prev_class->switched_from(rq, p);
1020 p->sched_class->switched_to(rq, p);
1021 } else if (oldprio != p->prio || dl_task(p))
1022 p->sched_class->prio_changed(rq, p, oldprio);
1025 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1027 const struct sched_class *class;
1029 if (p->sched_class == rq->curr->sched_class) {
1030 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1032 for_each_class(class) {
1033 if (class == rq->curr->sched_class)
1035 if (class == p->sched_class) {
1043 * A queue event has occurred, and we're going to schedule. In
1044 * this case, we can save a useless back to back clock update.
1046 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1047 rq_clock_skip_update(rq, true);
1052 * This is how migration works:
1054 * 1) we invoke migration_cpu_stop() on the target CPU using
1056 * 2) stopper starts to run (implicitly forcing the migrated thread
1058 * 3) it checks whether the migrated task is still in the wrong runqueue.
1059 * 4) if it's in the wrong runqueue then the migration thread removes
1060 * it and puts it into the right queue.
1061 * 5) stopper completes and stop_one_cpu() returns and the migration
1066 * move_queued_task - move a queued task to new rq.
1068 * Returns (locked) new rq. Old rq's lock is released.
1070 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1072 lockdep_assert_held(&rq->lock);
1074 dequeue_task(rq, p, 0);
1075 p->on_rq = TASK_ON_RQ_MIGRATING;
1076 set_task_cpu(p, new_cpu);
1077 raw_spin_unlock(&rq->lock);
1079 rq = cpu_rq(new_cpu);
1081 raw_spin_lock(&rq->lock);
1082 BUG_ON(task_cpu(p) != new_cpu);
1083 p->on_rq = TASK_ON_RQ_QUEUED;
1084 enqueue_task(rq, p, 0);
1085 check_preempt_curr(rq, p, 0);
1090 struct migration_arg {
1091 struct task_struct *task;
1096 * Move (not current) task off this cpu, onto dest cpu. We're doing
1097 * this because either it can't run here any more (set_cpus_allowed()
1098 * away from this CPU, or CPU going down), or because we're
1099 * attempting to rebalance this task on exec (sched_exec).
1101 * So we race with normal scheduler movements, but that's OK, as long
1102 * as the task is no longer on this CPU.
1104 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1106 if (unlikely(!cpu_active(dest_cpu)))
1109 /* Affinity changed (again). */
1110 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1113 rq = move_queued_task(rq, p, dest_cpu);
1119 * migration_cpu_stop - this will be executed by a highprio stopper thread
1120 * and performs thread migration by bumping thread off CPU then
1121 * 'pushing' onto another runqueue.
1123 static int migration_cpu_stop(void *data)
1125 struct migration_arg *arg = data;
1126 struct task_struct *p = arg->task;
1127 struct rq *rq = this_rq();
1130 * The original target cpu might have gone down and we might
1131 * be on another cpu but it doesn't matter.
1133 local_irq_disable();
1135 * We need to explicitly wake pending tasks before running
1136 * __migrate_task() such that we will not miss enforcing cpus_allowed
1137 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1139 sched_ttwu_pending();
1141 raw_spin_lock(&p->pi_lock);
1142 raw_spin_lock(&rq->lock);
1144 * If task_rq(p) != rq, it cannot be migrated here, because we're
1145 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1146 * we're holding p->pi_lock.
1148 if (task_rq(p) == rq && task_on_rq_queued(p))
1149 rq = __migrate_task(rq, p, arg->dest_cpu);
1150 raw_spin_unlock(&rq->lock);
1151 raw_spin_unlock(&p->pi_lock);
1158 * sched_class::set_cpus_allowed must do the below, but is not required to
1159 * actually call this function.
1161 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1163 cpumask_copy(&p->cpus_allowed, new_mask);
1164 p->nr_cpus_allowed = cpumask_weight(new_mask);
1167 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1169 struct rq *rq = task_rq(p);
1170 bool queued, running;
1172 lockdep_assert_held(&p->pi_lock);
1174 queued = task_on_rq_queued(p);
1175 running = task_current(rq, p);
1179 * Because __kthread_bind() calls this on blocked tasks without
1182 lockdep_assert_held(&rq->lock);
1183 dequeue_task(rq, p, DEQUEUE_SAVE);
1186 put_prev_task(rq, p);
1188 p->sched_class->set_cpus_allowed(p, new_mask);
1191 p->sched_class->set_curr_task(rq);
1193 enqueue_task(rq, p, ENQUEUE_RESTORE);
1197 * Change a given task's CPU affinity. Migrate the thread to a
1198 * proper CPU and schedule it away if the CPU it's executing on
1199 * is removed from the allowed bitmask.
1201 * NOTE: the caller must have a valid reference to the task, the
1202 * task must not exit() & deallocate itself prematurely. The
1203 * call is not atomic; no spinlocks may be held.
1205 static int __set_cpus_allowed_ptr(struct task_struct *p,
1206 const struct cpumask *new_mask, bool check)
1208 unsigned long flags;
1210 unsigned int dest_cpu;
1213 rq = task_rq_lock(p, &flags);
1216 * Must re-check here, to close a race against __kthread_bind(),
1217 * sched_setaffinity() is not guaranteed to observe the flag.
1219 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1224 if (cpumask_equal(&p->cpus_allowed, new_mask))
1227 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1232 do_set_cpus_allowed(p, new_mask);
1234 /* Can the task run on the task's current CPU? If so, we're done */
1235 if (cpumask_test_cpu(task_cpu(p), new_mask))
1238 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1239 if (task_running(rq, p) || p->state == TASK_WAKING) {
1240 struct migration_arg arg = { p, dest_cpu };
1241 /* Need help from migration thread: drop lock and wait. */
1242 task_rq_unlock(rq, p, &flags);
1243 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1244 tlb_migrate_finish(p->mm);
1246 } else if (task_on_rq_queued(p)) {
1248 * OK, since we're going to drop the lock immediately
1249 * afterwards anyway.
1251 lockdep_unpin_lock(&rq->lock);
1252 rq = move_queued_task(rq, p, dest_cpu);
1253 lockdep_pin_lock(&rq->lock);
1256 task_rq_unlock(rq, p, &flags);
1261 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1263 return __set_cpus_allowed_ptr(p, new_mask, false);
1265 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1267 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1269 #ifdef CONFIG_SCHED_DEBUG
1271 * We should never call set_task_cpu() on a blocked task,
1272 * ttwu() will sort out the placement.
1274 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1277 #ifdef CONFIG_LOCKDEP
1279 * The caller should hold either p->pi_lock or rq->lock, when changing
1280 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1282 * sched_move_task() holds both and thus holding either pins the cgroup,
1285 * Furthermore, all task_rq users should acquire both locks, see
1288 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1289 lockdep_is_held(&task_rq(p)->lock)));
1293 trace_sched_migrate_task(p, new_cpu);
1295 if (task_cpu(p) != new_cpu) {
1296 if (p->sched_class->migrate_task_rq)
1297 p->sched_class->migrate_task_rq(p);
1298 p->se.nr_migrations++;
1299 perf_event_task_migrate(p);
1302 __set_task_cpu(p, new_cpu);
1305 static void __migrate_swap_task(struct task_struct *p, int cpu)
1307 if (task_on_rq_queued(p)) {
1308 struct rq *src_rq, *dst_rq;
1310 src_rq = task_rq(p);
1311 dst_rq = cpu_rq(cpu);
1313 deactivate_task(src_rq, p, 0);
1314 set_task_cpu(p, cpu);
1315 activate_task(dst_rq, p, 0);
1316 check_preempt_curr(dst_rq, p, 0);
1319 * Task isn't running anymore; make it appear like we migrated
1320 * it before it went to sleep. This means on wakeup we make the
1321 * previous cpu our targer instead of where it really is.
1327 struct migration_swap_arg {
1328 struct task_struct *src_task, *dst_task;
1329 int src_cpu, dst_cpu;
1332 static int migrate_swap_stop(void *data)
1334 struct migration_swap_arg *arg = data;
1335 struct rq *src_rq, *dst_rq;
1338 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1341 src_rq = cpu_rq(arg->src_cpu);
1342 dst_rq = cpu_rq(arg->dst_cpu);
1344 double_raw_lock(&arg->src_task->pi_lock,
1345 &arg->dst_task->pi_lock);
1346 double_rq_lock(src_rq, dst_rq);
1348 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1351 if (task_cpu(arg->src_task) != arg->src_cpu)
1354 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1357 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1360 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1361 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1366 double_rq_unlock(src_rq, dst_rq);
1367 raw_spin_unlock(&arg->dst_task->pi_lock);
1368 raw_spin_unlock(&arg->src_task->pi_lock);
1374 * Cross migrate two tasks
1376 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1378 struct migration_swap_arg arg;
1381 arg = (struct migration_swap_arg){
1383 .src_cpu = task_cpu(cur),
1385 .dst_cpu = task_cpu(p),
1388 if (arg.src_cpu == arg.dst_cpu)
1392 * These three tests are all lockless; this is OK since all of them
1393 * will be re-checked with proper locks held further down the line.
1395 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1398 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1401 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1404 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1405 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1412 * wait_task_inactive - wait for a thread to unschedule.
1414 * If @match_state is nonzero, it's the @p->state value just checked and
1415 * not expected to change. If it changes, i.e. @p might have woken up,
1416 * then return zero. When we succeed in waiting for @p to be off its CPU,
1417 * we return a positive number (its total switch count). If a second call
1418 * a short while later returns the same number, the caller can be sure that
1419 * @p has remained unscheduled the whole time.
1421 * The caller must ensure that the task *will* unschedule sometime soon,
1422 * else this function might spin for a *long* time. This function can't
1423 * be called with interrupts off, or it may introduce deadlock with
1424 * smp_call_function() if an IPI is sent by the same process we are
1425 * waiting to become inactive.
1427 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1429 unsigned long flags;
1430 int running, queued;
1436 * We do the initial early heuristics without holding
1437 * any task-queue locks at all. We'll only try to get
1438 * the runqueue lock when things look like they will
1444 * If the task is actively running on another CPU
1445 * still, just relax and busy-wait without holding
1448 * NOTE! Since we don't hold any locks, it's not
1449 * even sure that "rq" stays as the right runqueue!
1450 * But we don't care, since "task_running()" will
1451 * return false if the runqueue has changed and p
1452 * is actually now running somewhere else!
1454 while (task_running(rq, p)) {
1455 if (match_state && unlikely(p->state != match_state))
1461 * Ok, time to look more closely! We need the rq
1462 * lock now, to be *sure*. If we're wrong, we'll
1463 * just go back and repeat.
1465 rq = task_rq_lock(p, &flags);
1466 trace_sched_wait_task(p);
1467 running = task_running(rq, p);
1468 queued = task_on_rq_queued(p);
1470 if (!match_state || p->state == match_state)
1471 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1472 task_rq_unlock(rq, p, &flags);
1475 * If it changed from the expected state, bail out now.
1477 if (unlikely(!ncsw))
1481 * Was it really running after all now that we
1482 * checked with the proper locks actually held?
1484 * Oops. Go back and try again..
1486 if (unlikely(running)) {
1492 * It's not enough that it's not actively running,
1493 * it must be off the runqueue _entirely_, and not
1496 * So if it was still runnable (but just not actively
1497 * running right now), it's preempted, and we should
1498 * yield - it could be a while.
1500 if (unlikely(queued)) {
1501 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1503 set_current_state(TASK_UNINTERRUPTIBLE);
1504 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1509 * Ahh, all good. It wasn't running, and it wasn't
1510 * runnable, which means that it will never become
1511 * running in the future either. We're all done!
1520 * kick_process - kick a running thread to enter/exit the kernel
1521 * @p: the to-be-kicked thread
1523 * Cause a process which is running on another CPU to enter
1524 * kernel-mode, without any delay. (to get signals handled.)
1526 * NOTE: this function doesn't have to take the runqueue lock,
1527 * because all it wants to ensure is that the remote task enters
1528 * the kernel. If the IPI races and the task has been migrated
1529 * to another CPU then no harm is done and the purpose has been
1532 void kick_process(struct task_struct *p)
1538 if ((cpu != smp_processor_id()) && task_curr(p))
1539 smp_send_reschedule(cpu);
1542 EXPORT_SYMBOL_GPL(kick_process);
1545 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1547 static int select_fallback_rq(int cpu, struct task_struct *p)
1549 int nid = cpu_to_node(cpu);
1550 const struct cpumask *nodemask = NULL;
1551 enum { cpuset, possible, fail } state = cpuset;
1555 * If the node that the cpu is on has been offlined, cpu_to_node()
1556 * will return -1. There is no cpu on the node, and we should
1557 * select the cpu on the other node.
1560 nodemask = cpumask_of_node(nid);
1562 /* Look for allowed, online CPU in same node. */
1563 for_each_cpu(dest_cpu, nodemask) {
1564 if (!cpu_online(dest_cpu))
1566 if (!cpu_active(dest_cpu))
1568 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1574 /* Any allowed, online CPU? */
1575 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1576 if (!cpu_online(dest_cpu))
1578 if (!cpu_active(dest_cpu))
1583 /* No more Mr. Nice Guy. */
1586 if (IS_ENABLED(CONFIG_CPUSETS)) {
1587 cpuset_cpus_allowed_fallback(p);
1593 do_set_cpus_allowed(p, cpu_possible_mask);
1604 if (state != cpuset) {
1606 * Don't tell them about moving exiting tasks or
1607 * kernel threads (both mm NULL), since they never
1610 if (p->mm && printk_ratelimit()) {
1611 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1612 task_pid_nr(p), p->comm, cpu);
1620 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1623 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1625 lockdep_assert_held(&p->pi_lock);
1627 if (p->nr_cpus_allowed > 1)
1628 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1631 * In order not to call set_task_cpu() on a blocking task we need
1632 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1635 * Since this is common to all placement strategies, this lives here.
1637 * [ this allows ->select_task() to simply return task_cpu(p) and
1638 * not worry about this generic constraint ]
1640 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1642 cpu = select_fallback_rq(task_cpu(p), p);
1647 static void update_avg(u64 *avg, u64 sample)
1649 s64 diff = sample - *avg;
1655 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1656 const struct cpumask *new_mask, bool check)
1658 return set_cpus_allowed_ptr(p, new_mask);
1661 #endif /* CONFIG_SMP */
1664 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1666 #ifdef CONFIG_SCHEDSTATS
1667 struct rq *rq = this_rq();
1670 int this_cpu = smp_processor_id();
1672 if (cpu == this_cpu) {
1673 schedstat_inc(rq, ttwu_local);
1674 schedstat_inc(p, se.statistics.nr_wakeups_local);
1676 struct sched_domain *sd;
1678 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1680 for_each_domain(this_cpu, sd) {
1681 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1682 schedstat_inc(sd, ttwu_wake_remote);
1689 if (wake_flags & WF_MIGRATED)
1690 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1692 #endif /* CONFIG_SMP */
1694 schedstat_inc(rq, ttwu_count);
1695 schedstat_inc(p, se.statistics.nr_wakeups);
1697 if (wake_flags & WF_SYNC)
1698 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1700 #endif /* CONFIG_SCHEDSTATS */
1703 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1705 activate_task(rq, p, en_flags);
1706 p->on_rq = TASK_ON_RQ_QUEUED;
1708 /* if a worker is waking up, notify workqueue */
1709 if (p->flags & PF_WQ_WORKER)
1710 wq_worker_waking_up(p, cpu_of(rq));
1714 * Mark the task runnable and perform wakeup-preemption.
1717 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1719 check_preempt_curr(rq, p, wake_flags);
1720 p->state = TASK_RUNNING;
1721 trace_sched_wakeup(p);
1724 if (p->sched_class->task_woken) {
1726 * Our task @p is fully woken up and running; so its safe to
1727 * drop the rq->lock, hereafter rq is only used for statistics.
1729 lockdep_unpin_lock(&rq->lock);
1730 p->sched_class->task_woken(rq, p);
1731 lockdep_pin_lock(&rq->lock);
1734 if (rq->idle_stamp) {
1735 u64 delta = rq_clock(rq) - rq->idle_stamp;
1736 u64 max = 2*rq->max_idle_balance_cost;
1738 update_avg(&rq->avg_idle, delta);
1740 if (rq->avg_idle > max)
1749 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1751 lockdep_assert_held(&rq->lock);
1754 if (p->sched_contributes_to_load)
1755 rq->nr_uninterruptible--;
1758 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1759 ttwu_do_wakeup(rq, p, wake_flags);
1763 * Called in case the task @p isn't fully descheduled from its runqueue,
1764 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1765 * since all we need to do is flip p->state to TASK_RUNNING, since
1766 * the task is still ->on_rq.
1768 static int ttwu_remote(struct task_struct *p, int wake_flags)
1773 rq = __task_rq_lock(p);
1774 if (task_on_rq_queued(p)) {
1775 /* check_preempt_curr() may use rq clock */
1776 update_rq_clock(rq);
1777 ttwu_do_wakeup(rq, p, wake_flags);
1780 __task_rq_unlock(rq);
1786 void sched_ttwu_pending(void)
1788 struct rq *rq = this_rq();
1789 struct llist_node *llist = llist_del_all(&rq->wake_list);
1790 struct task_struct *p;
1791 unsigned long flags;
1796 raw_spin_lock_irqsave(&rq->lock, flags);
1797 lockdep_pin_lock(&rq->lock);
1800 p = llist_entry(llist, struct task_struct, wake_entry);
1801 llist = llist_next(llist);
1802 ttwu_do_activate(rq, p, 0);
1805 lockdep_unpin_lock(&rq->lock);
1806 raw_spin_unlock_irqrestore(&rq->lock, flags);
1809 void scheduler_ipi(void)
1812 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1813 * TIF_NEED_RESCHED remotely (for the first time) will also send
1816 preempt_fold_need_resched();
1818 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1822 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1823 * traditionally all their work was done from the interrupt return
1824 * path. Now that we actually do some work, we need to make sure
1827 * Some archs already do call them, luckily irq_enter/exit nest
1830 * Arguably we should visit all archs and update all handlers,
1831 * however a fair share of IPIs are still resched only so this would
1832 * somewhat pessimize the simple resched case.
1835 sched_ttwu_pending();
1838 * Check if someone kicked us for doing the nohz idle load balance.
1840 if (unlikely(got_nohz_idle_kick())) {
1841 this_rq()->idle_balance = 1;
1842 raise_softirq_irqoff(SCHED_SOFTIRQ);
1847 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1849 struct rq *rq = cpu_rq(cpu);
1851 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1852 if (!set_nr_if_polling(rq->idle))
1853 smp_send_reschedule(cpu);
1855 trace_sched_wake_idle_without_ipi(cpu);
1859 void wake_up_if_idle(int cpu)
1861 struct rq *rq = cpu_rq(cpu);
1862 unsigned long flags;
1866 if (!is_idle_task(rcu_dereference(rq->curr)))
1869 if (set_nr_if_polling(rq->idle)) {
1870 trace_sched_wake_idle_without_ipi(cpu);
1872 raw_spin_lock_irqsave(&rq->lock, flags);
1873 if (is_idle_task(rq->curr))
1874 smp_send_reschedule(cpu);
1875 /* Else cpu is not in idle, do nothing here */
1876 raw_spin_unlock_irqrestore(&rq->lock, flags);
1883 bool cpus_share_cache(int this_cpu, int that_cpu)
1885 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1887 #endif /* CONFIG_SMP */
1889 static void ttwu_queue(struct task_struct *p, int cpu)
1891 struct rq *rq = cpu_rq(cpu);
1893 #if defined(CONFIG_SMP)
1894 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1895 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1896 ttwu_queue_remote(p, cpu);
1901 raw_spin_lock(&rq->lock);
1902 lockdep_pin_lock(&rq->lock);
1903 ttwu_do_activate(rq, p, 0);
1904 lockdep_unpin_lock(&rq->lock);
1905 raw_spin_unlock(&rq->lock);
1909 * try_to_wake_up - wake up a thread
1910 * @p: the thread to be awakened
1911 * @state: the mask of task states that can be woken
1912 * @wake_flags: wake modifier flags (WF_*)
1914 * Put it on the run-queue if it's not already there. The "current"
1915 * thread is always on the run-queue (except when the actual
1916 * re-schedule is in progress), and as such you're allowed to do
1917 * the simpler "current->state = TASK_RUNNING" to mark yourself
1918 * runnable without the overhead of this.
1920 * Return: %true if @p was woken up, %false if it was already running.
1921 * or @state didn't match @p's state.
1924 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1926 unsigned long flags;
1927 int cpu, success = 0;
1930 * If we are going to wake up a thread waiting for CONDITION we
1931 * need to ensure that CONDITION=1 done by the caller can not be
1932 * reordered with p->state check below. This pairs with mb() in
1933 * set_current_state() the waiting thread does.
1935 smp_mb__before_spinlock();
1936 raw_spin_lock_irqsave(&p->pi_lock, flags);
1937 if (!(p->state & state))
1940 trace_sched_waking(p);
1942 success = 1; /* we're going to change ->state */
1945 if (p->on_rq && ttwu_remote(p, wake_flags))
1950 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1951 * possible to, falsely, observe p->on_cpu == 0.
1953 * One must be running (->on_cpu == 1) in order to remove oneself
1954 * from the runqueue.
1956 * [S] ->on_cpu = 1; [L] ->on_rq
1960 * [S] ->on_rq = 0; [L] ->on_cpu
1962 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1963 * from the consecutive calls to schedule(); the first switching to our
1964 * task, the second putting it to sleep.
1969 * If the owning (remote) cpu is still in the middle of schedule() with
1970 * this task as prev, wait until its done referencing the task.
1975 * Combined with the control dependency above, we have an effective
1976 * smp_load_acquire() without the need for full barriers.
1978 * Pairs with the smp_store_release() in finish_lock_switch().
1980 * This ensures that tasks getting woken will be fully ordered against
1981 * their previous state and preserve Program Order.
1985 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1986 p->state = TASK_WAKING;
1988 if (p->sched_class->task_waking)
1989 p->sched_class->task_waking(p);
1991 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1992 if (task_cpu(p) != cpu) {
1993 wake_flags |= WF_MIGRATED;
1994 set_task_cpu(p, cpu);
1996 #endif /* CONFIG_SMP */
2000 ttwu_stat(p, cpu, wake_flags);
2002 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2008 * try_to_wake_up_local - try to wake up a local task with rq lock held
2009 * @p: the thread to be awakened
2011 * Put @p on the run-queue if it's not already there. The caller must
2012 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2015 static void try_to_wake_up_local(struct task_struct *p)
2017 struct rq *rq = task_rq(p);
2019 if (WARN_ON_ONCE(rq != this_rq()) ||
2020 WARN_ON_ONCE(p == current))
2023 lockdep_assert_held(&rq->lock);
2025 if (!raw_spin_trylock(&p->pi_lock)) {
2027 * This is OK, because current is on_cpu, which avoids it being
2028 * picked for load-balance and preemption/IRQs are still
2029 * disabled avoiding further scheduler activity on it and we've
2030 * not yet picked a replacement task.
2032 lockdep_unpin_lock(&rq->lock);
2033 raw_spin_unlock(&rq->lock);
2034 raw_spin_lock(&p->pi_lock);
2035 raw_spin_lock(&rq->lock);
2036 lockdep_pin_lock(&rq->lock);
2039 if (!(p->state & TASK_NORMAL))
2042 trace_sched_waking(p);
2044 if (!task_on_rq_queued(p))
2045 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2047 ttwu_do_wakeup(rq, p, 0);
2048 ttwu_stat(p, smp_processor_id(), 0);
2050 raw_spin_unlock(&p->pi_lock);
2054 * wake_up_process - Wake up a specific process
2055 * @p: The process to be woken up.
2057 * Attempt to wake up the nominated process and move it to the set of runnable
2060 * Return: 1 if the process was woken up, 0 if it was already running.
2062 * It may be assumed that this function implies a write memory barrier before
2063 * changing the task state if and only if any tasks are woken up.
2065 int wake_up_process(struct task_struct *p)
2067 return try_to_wake_up(p, TASK_NORMAL, 0);
2069 EXPORT_SYMBOL(wake_up_process);
2071 int wake_up_state(struct task_struct *p, unsigned int state)
2073 return try_to_wake_up(p, state, 0);
2077 * This function clears the sched_dl_entity static params.
2079 void __dl_clear_params(struct task_struct *p)
2081 struct sched_dl_entity *dl_se = &p->dl;
2083 dl_se->dl_runtime = 0;
2084 dl_se->dl_deadline = 0;
2085 dl_se->dl_period = 0;
2089 dl_se->dl_throttled = 0;
2091 dl_se->dl_yielded = 0;
2095 * Perform scheduler related setup for a newly forked process p.
2096 * p is forked by current.
2098 * __sched_fork() is basic setup used by init_idle() too:
2100 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2105 p->se.exec_start = 0;
2106 p->se.sum_exec_runtime = 0;
2107 p->se.prev_sum_exec_runtime = 0;
2108 p->se.nr_migrations = 0;
2110 INIT_LIST_HEAD(&p->se.group_node);
2112 #ifdef CONFIG_SCHEDSTATS
2113 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2116 RB_CLEAR_NODE(&p->dl.rb_node);
2117 init_dl_task_timer(&p->dl);
2118 __dl_clear_params(p);
2120 INIT_LIST_HEAD(&p->rt.run_list);
2122 #ifdef CONFIG_PREEMPT_NOTIFIERS
2123 INIT_HLIST_HEAD(&p->preempt_notifiers);
2126 #ifdef CONFIG_NUMA_BALANCING
2127 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2128 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2129 p->mm->numa_scan_seq = 0;
2132 if (clone_flags & CLONE_VM)
2133 p->numa_preferred_nid = current->numa_preferred_nid;
2135 p->numa_preferred_nid = -1;
2137 p->node_stamp = 0ULL;
2138 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2139 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2140 p->numa_work.next = &p->numa_work;
2141 p->numa_faults = NULL;
2142 p->last_task_numa_placement = 0;
2143 p->last_sum_exec_runtime = 0;
2145 p->numa_group = NULL;
2146 #endif /* CONFIG_NUMA_BALANCING */
2149 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2151 #ifdef CONFIG_NUMA_BALANCING
2153 void set_numabalancing_state(bool enabled)
2156 static_branch_enable(&sched_numa_balancing);
2158 static_branch_disable(&sched_numa_balancing);
2161 #ifdef CONFIG_PROC_SYSCTL
2162 int sysctl_numa_balancing(struct ctl_table *table, int write,
2163 void __user *buffer, size_t *lenp, loff_t *ppos)
2167 int state = static_branch_likely(&sched_numa_balancing);
2169 if (write && !capable(CAP_SYS_ADMIN))
2174 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2178 set_numabalancing_state(state);
2185 * fork()/clone()-time setup:
2187 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2189 unsigned long flags;
2190 int cpu = get_cpu();
2192 __sched_fork(clone_flags, p);
2194 * We mark the process as running here. This guarantees that
2195 * nobody will actually run it, and a signal or other external
2196 * event cannot wake it up and insert it on the runqueue either.
2198 p->state = TASK_RUNNING;
2201 * Make sure we do not leak PI boosting priority to the child.
2203 p->prio = current->normal_prio;
2206 * Revert to default priority/policy on fork if requested.
2208 if (unlikely(p->sched_reset_on_fork)) {
2209 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2210 p->policy = SCHED_NORMAL;
2211 p->static_prio = NICE_TO_PRIO(0);
2213 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2214 p->static_prio = NICE_TO_PRIO(0);
2216 p->prio = p->normal_prio = __normal_prio(p);
2220 * We don't need the reset flag anymore after the fork. It has
2221 * fulfilled its duty:
2223 p->sched_reset_on_fork = 0;
2226 if (dl_prio(p->prio)) {
2229 } else if (rt_prio(p->prio)) {
2230 p->sched_class = &rt_sched_class;
2232 p->sched_class = &fair_sched_class;
2235 if (p->sched_class->task_fork)
2236 p->sched_class->task_fork(p);
2239 * The child is not yet in the pid-hash so no cgroup attach races,
2240 * and the cgroup is pinned to this child due to cgroup_fork()
2241 * is ran before sched_fork().
2243 * Silence PROVE_RCU.
2245 raw_spin_lock_irqsave(&p->pi_lock, flags);
2246 set_task_cpu(p, cpu);
2247 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2249 #ifdef CONFIG_SCHED_INFO
2250 if (likely(sched_info_on()))
2251 memset(&p->sched_info, 0, sizeof(p->sched_info));
2253 #if defined(CONFIG_SMP)
2256 init_task_preempt_count(p);
2258 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2259 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2266 unsigned long to_ratio(u64 period, u64 runtime)
2268 if (runtime == RUNTIME_INF)
2272 * Doing this here saves a lot of checks in all
2273 * the calling paths, and returning zero seems
2274 * safe for them anyway.
2279 return div64_u64(runtime << 20, period);
2283 inline struct dl_bw *dl_bw_of(int i)
2285 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2286 "sched RCU must be held");
2287 return &cpu_rq(i)->rd->dl_bw;
2290 static inline int dl_bw_cpus(int i)
2292 struct root_domain *rd = cpu_rq(i)->rd;
2295 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2296 "sched RCU must be held");
2297 for_each_cpu_and(i, rd->span, cpu_active_mask)
2303 inline struct dl_bw *dl_bw_of(int i)
2305 return &cpu_rq(i)->dl.dl_bw;
2308 static inline int dl_bw_cpus(int i)
2315 * We must be sure that accepting a new task (or allowing changing the
2316 * parameters of an existing one) is consistent with the bandwidth
2317 * constraints. If yes, this function also accordingly updates the currently
2318 * allocated bandwidth to reflect the new situation.
2320 * This function is called while holding p's rq->lock.
2322 * XXX we should delay bw change until the task's 0-lag point, see
2325 static int dl_overflow(struct task_struct *p, int policy,
2326 const struct sched_attr *attr)
2329 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2330 u64 period = attr->sched_period ?: attr->sched_deadline;
2331 u64 runtime = attr->sched_runtime;
2332 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2335 if (new_bw == p->dl.dl_bw)
2339 * Either if a task, enters, leave, or stays -deadline but changes
2340 * its parameters, we may need to update accordingly the total
2341 * allocated bandwidth of the container.
2343 raw_spin_lock(&dl_b->lock);
2344 cpus = dl_bw_cpus(task_cpu(p));
2345 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2346 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2347 __dl_add(dl_b, new_bw);
2349 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2350 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2351 __dl_clear(dl_b, p->dl.dl_bw);
2352 __dl_add(dl_b, new_bw);
2354 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2355 __dl_clear(dl_b, p->dl.dl_bw);
2358 raw_spin_unlock(&dl_b->lock);
2363 extern void init_dl_bw(struct dl_bw *dl_b);
2366 * wake_up_new_task - wake up a newly created task for the first time.
2368 * This function will do some initial scheduler statistics housekeeping
2369 * that must be done for every newly created context, then puts the task
2370 * on the runqueue and wakes it.
2372 void wake_up_new_task(struct task_struct *p)
2374 unsigned long flags;
2377 raw_spin_lock_irqsave(&p->pi_lock, flags);
2378 /* Initialize new task's runnable average */
2379 init_entity_runnable_average(&p->se);
2382 * Fork balancing, do it here and not earlier because:
2383 * - cpus_allowed can change in the fork path
2384 * - any previously selected cpu might disappear through hotplug
2386 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2389 rq = __task_rq_lock(p);
2390 activate_task(rq, p, ENQUEUE_WAKEUP_NEW);
2391 p->on_rq = TASK_ON_RQ_QUEUED;
2392 trace_sched_wakeup_new(p);
2393 check_preempt_curr(rq, p, WF_FORK);
2395 if (p->sched_class->task_woken) {
2397 * Nothing relies on rq->lock after this, so its fine to
2400 lockdep_unpin_lock(&rq->lock);
2401 p->sched_class->task_woken(rq, p);
2402 lockdep_pin_lock(&rq->lock);
2405 task_rq_unlock(rq, p, &flags);
2408 #ifdef CONFIG_PREEMPT_NOTIFIERS
2410 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2412 void preempt_notifier_inc(void)
2414 static_key_slow_inc(&preempt_notifier_key);
2416 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2418 void preempt_notifier_dec(void)
2420 static_key_slow_dec(&preempt_notifier_key);
2422 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2425 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2426 * @notifier: notifier struct to register
2428 void preempt_notifier_register(struct preempt_notifier *notifier)
2430 if (!static_key_false(&preempt_notifier_key))
2431 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2433 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2435 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2438 * preempt_notifier_unregister - no longer interested in preemption notifications
2439 * @notifier: notifier struct to unregister
2441 * This is *not* safe to call from within a preemption notifier.
2443 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2445 hlist_del(¬ifier->link);
2447 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2449 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2451 struct preempt_notifier *notifier;
2453 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2454 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2457 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2459 if (static_key_false(&preempt_notifier_key))
2460 __fire_sched_in_preempt_notifiers(curr);
2464 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2465 struct task_struct *next)
2467 struct preempt_notifier *notifier;
2469 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2470 notifier->ops->sched_out(notifier, next);
2473 static __always_inline void
2474 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2475 struct task_struct *next)
2477 if (static_key_false(&preempt_notifier_key))
2478 __fire_sched_out_preempt_notifiers(curr, next);
2481 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2483 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2488 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2489 struct task_struct *next)
2493 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2496 * prepare_task_switch - prepare to switch tasks
2497 * @rq: the runqueue preparing to switch
2498 * @prev: the current task that is being switched out
2499 * @next: the task we are going to switch to.
2501 * This is called with the rq lock held and interrupts off. It must
2502 * be paired with a subsequent finish_task_switch after the context
2505 * prepare_task_switch sets up locking and calls architecture specific
2509 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2510 struct task_struct *next)
2512 sched_info_switch(rq, prev, next);
2513 perf_event_task_sched_out(prev, next);
2514 fire_sched_out_preempt_notifiers(prev, next);
2515 prepare_lock_switch(rq, next);
2516 prepare_arch_switch(next);
2520 * finish_task_switch - clean up after a task-switch
2521 * @prev: the thread we just switched away from.
2523 * finish_task_switch must be called after the context switch, paired
2524 * with a prepare_task_switch call before the context switch.
2525 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2526 * and do any other architecture-specific cleanup actions.
2528 * Note that we may have delayed dropping an mm in context_switch(). If
2529 * so, we finish that here outside of the runqueue lock. (Doing it
2530 * with the lock held can cause deadlocks; see schedule() for
2533 * The context switch have flipped the stack from under us and restored the
2534 * local variables which were saved when this task called schedule() in the
2535 * past. prev == current is still correct but we need to recalculate this_rq
2536 * because prev may have moved to another CPU.
2538 static struct rq *finish_task_switch(struct task_struct *prev)
2539 __releases(rq->lock)
2541 struct rq *rq = this_rq();
2542 struct mm_struct *mm = rq->prev_mm;
2546 * The previous task will have left us with a preempt_count of 2
2547 * because it left us after:
2550 * preempt_disable(); // 1
2552 * raw_spin_lock_irq(&rq->lock) // 2
2554 * Also, see FORK_PREEMPT_COUNT.
2556 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2557 "corrupted preempt_count: %s/%d/0x%x\n",
2558 current->comm, current->pid, preempt_count()))
2559 preempt_count_set(FORK_PREEMPT_COUNT);
2564 * A task struct has one reference for the use as "current".
2565 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2566 * schedule one last time. The schedule call will never return, and
2567 * the scheduled task must drop that reference.
2569 * We must observe prev->state before clearing prev->on_cpu (in
2570 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2571 * running on another CPU and we could rave with its RUNNING -> DEAD
2572 * transition, resulting in a double drop.
2574 prev_state = prev->state;
2575 vtime_task_switch(prev);
2576 perf_event_task_sched_in(prev, current);
2577 finish_lock_switch(rq, prev);
2578 finish_arch_post_lock_switch();
2580 fire_sched_in_preempt_notifiers(current);
2583 if (unlikely(prev_state == TASK_DEAD)) {
2584 if (prev->sched_class->task_dead)
2585 prev->sched_class->task_dead(prev);
2588 * Remove function-return probe instances associated with this
2589 * task and put them back on the free list.
2591 kprobe_flush_task(prev);
2592 put_task_struct(prev);
2595 tick_nohz_task_switch();
2601 /* rq->lock is NOT held, but preemption is disabled */
2602 static void __balance_callback(struct rq *rq)
2604 struct callback_head *head, *next;
2605 void (*func)(struct rq *rq);
2606 unsigned long flags;
2608 raw_spin_lock_irqsave(&rq->lock, flags);
2609 head = rq->balance_callback;
2610 rq->balance_callback = NULL;
2612 func = (void (*)(struct rq *))head->func;
2619 raw_spin_unlock_irqrestore(&rq->lock, flags);
2622 static inline void balance_callback(struct rq *rq)
2624 if (unlikely(rq->balance_callback))
2625 __balance_callback(rq);
2630 static inline void balance_callback(struct rq *rq)
2637 * schedule_tail - first thing a freshly forked thread must call.
2638 * @prev: the thread we just switched away from.
2640 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2641 __releases(rq->lock)
2646 * New tasks start with FORK_PREEMPT_COUNT, see there and
2647 * finish_task_switch() for details.
2649 * finish_task_switch() will drop rq->lock() and lower preempt_count
2650 * and the preempt_enable() will end up enabling preemption (on
2651 * PREEMPT_COUNT kernels).
2654 rq = finish_task_switch(prev);
2655 balance_callback(rq);
2658 if (current->set_child_tid)
2659 put_user(task_pid_vnr(current), current->set_child_tid);
2663 * context_switch - switch to the new MM and the new thread's register state.
2665 static inline struct rq *
2666 context_switch(struct rq *rq, struct task_struct *prev,
2667 struct task_struct *next)
2669 struct mm_struct *mm, *oldmm;
2671 prepare_task_switch(rq, prev, next);
2674 oldmm = prev->active_mm;
2676 * For paravirt, this is coupled with an exit in switch_to to
2677 * combine the page table reload and the switch backend into
2680 arch_start_context_switch(prev);
2683 next->active_mm = oldmm;
2684 atomic_inc(&oldmm->mm_count);
2685 enter_lazy_tlb(oldmm, next);
2687 switch_mm(oldmm, mm, next);
2690 prev->active_mm = NULL;
2691 rq->prev_mm = oldmm;
2694 * Since the runqueue lock will be released by the next
2695 * task (which is an invalid locking op but in the case
2696 * of the scheduler it's an obvious special-case), so we
2697 * do an early lockdep release here:
2699 lockdep_unpin_lock(&rq->lock);
2700 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2702 /* Here we just switch the register state and the stack. */
2703 switch_to(prev, next, prev);
2706 return finish_task_switch(prev);
2710 * nr_running and nr_context_switches:
2712 * externally visible scheduler statistics: current number of runnable
2713 * threads, total number of context switches performed since bootup.
2715 unsigned long nr_running(void)
2717 unsigned long i, sum = 0;
2719 for_each_online_cpu(i)
2720 sum += cpu_rq(i)->nr_running;
2726 * Check if only the current task is running on the cpu.
2728 * Caution: this function does not check that the caller has disabled
2729 * preemption, thus the result might have a time-of-check-to-time-of-use
2730 * race. The caller is responsible to use it correctly, for example:
2732 * - from a non-preemptable section (of course)
2734 * - from a thread that is bound to a single CPU
2736 * - in a loop with very short iterations (e.g. a polling loop)
2738 bool single_task_running(void)
2740 return raw_rq()->nr_running == 1;
2742 EXPORT_SYMBOL(single_task_running);
2744 unsigned long long nr_context_switches(void)
2747 unsigned long long sum = 0;
2749 for_each_possible_cpu(i)
2750 sum += cpu_rq(i)->nr_switches;
2755 unsigned long nr_iowait(void)
2757 unsigned long i, sum = 0;
2759 for_each_possible_cpu(i)
2760 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2765 unsigned long nr_iowait_cpu(int cpu)
2767 struct rq *this = cpu_rq(cpu);
2768 return atomic_read(&this->nr_iowait);
2771 #ifdef CONFIG_CPU_QUIET
2772 u64 nr_running_integral(unsigned int cpu)
2774 unsigned int seqcnt;
2778 if (cpu >= nr_cpu_ids)
2784 * Update average to avoid reading stalled value if there were
2785 * no run-queue changes for a long time. On the other hand if
2786 * the changes are happening right now, just read current value
2790 seqcnt = read_seqcount_begin(&q->ave_seqcnt);
2791 integral = do_nr_running_integral(q);
2792 if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
2793 read_seqcount_begin(&q->ave_seqcnt);
2794 integral = q->nr_running_integral;
2801 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2803 struct rq *rq = this_rq();
2804 *nr_waiters = atomic_read(&rq->nr_iowait);
2805 *load = rq->load.weight;
2811 * sched_exec - execve() is a valuable balancing opportunity, because at
2812 * this point the task has the smallest effective memory and cache footprint.
2814 void sched_exec(void)
2816 struct task_struct *p = current;
2817 unsigned long flags;
2820 raw_spin_lock_irqsave(&p->pi_lock, flags);
2821 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2822 if (dest_cpu == smp_processor_id())
2825 if (likely(cpu_active(dest_cpu))) {
2826 struct migration_arg arg = { p, dest_cpu };
2828 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2829 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2833 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2838 DEFINE_PER_CPU(struct kernel_stat, kstat);
2839 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2841 EXPORT_PER_CPU_SYMBOL(kstat);
2842 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2845 * Return accounted runtime for the task.
2846 * In case the task is currently running, return the runtime plus current's
2847 * pending runtime that have not been accounted yet.
2849 unsigned long long task_sched_runtime(struct task_struct *p)
2851 unsigned long flags;
2855 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2857 * 64-bit doesn't need locks to atomically read a 64bit value.
2858 * So we have a optimization chance when the task's delta_exec is 0.
2859 * Reading ->on_cpu is racy, but this is ok.
2861 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2862 * If we race with it entering cpu, unaccounted time is 0. This is
2863 * indistinguishable from the read occurring a few cycles earlier.
2864 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2865 * been accounted, so we're correct here as well.
2867 if (!p->on_cpu || !task_on_rq_queued(p))
2868 return p->se.sum_exec_runtime;
2871 rq = task_rq_lock(p, &flags);
2873 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2874 * project cycles that may never be accounted to this
2875 * thread, breaking clock_gettime().
2877 if (task_current(rq, p) && task_on_rq_queued(p)) {
2878 update_rq_clock(rq);
2879 p->sched_class->update_curr(rq);
2881 ns = p->se.sum_exec_runtime;
2882 task_rq_unlock(rq, p, &flags);
2887 #ifdef CONFIG_CPU_FREQ_GOV_SCHED
2888 static unsigned long sum_capacity_reqs(unsigned long cfs_cap,
2889 struct sched_capacity_reqs *scr)
2891 unsigned long total = cfs_cap + scr->rt;
2893 total = total * capacity_margin;
2894 total /= SCHED_CAPACITY_SCALE;
2899 static void sched_freq_tick(int cpu)
2901 struct sched_capacity_reqs *scr;
2902 unsigned long capacity_orig, capacity_curr;
2907 capacity_orig = capacity_orig_of(cpu);
2908 capacity_curr = capacity_curr_of(cpu);
2909 if (capacity_curr == capacity_orig)
2913 * To make free room for a task that is building up its "real"
2914 * utilization and to harm its performance the least, request
2915 * a jump to max OPP as soon as the margin of free capacity is
2916 * impacted (specified by capacity_margin).
2918 scr = &per_cpu(cpu_sched_capacity_reqs, cpu);
2919 if (capacity_curr < sum_capacity_reqs(cpu_util(cpu), scr))
2920 set_cfs_cpu_capacity(cpu, true, capacity_max);
2923 static inline void sched_freq_tick(int cpu) { }
2927 * This function gets called by the timer code, with HZ frequency.
2928 * We call it with interrupts disabled.
2930 void scheduler_tick(void)
2932 int cpu = smp_processor_id();
2933 struct rq *rq = cpu_rq(cpu);
2934 struct task_struct *curr = rq->curr;
2938 raw_spin_lock(&rq->lock);
2939 update_rq_clock(rq);
2940 curr->sched_class->task_tick(rq, curr, 0);
2941 update_cpu_load_active(rq);
2942 calc_global_load_tick(rq);
2943 sched_freq_tick(cpu);
2944 raw_spin_unlock(&rq->lock);
2946 perf_event_task_tick();
2949 rq->idle_balance = idle_cpu(cpu);
2950 trigger_load_balance(rq);
2952 rq_last_tick_reset(rq);
2955 #ifdef CONFIG_NO_HZ_FULL
2957 * scheduler_tick_max_deferment
2959 * Keep at least one tick per second when a single
2960 * active task is running because the scheduler doesn't
2961 * yet completely support full dynticks environment.
2963 * This makes sure that uptime, CFS vruntime, load
2964 * balancing, etc... continue to move forward, even
2965 * with a very low granularity.
2967 * Return: Maximum deferment in nanoseconds.
2969 u64 scheduler_tick_max_deferment(void)
2971 struct rq *rq = this_rq();
2972 unsigned long next, now = READ_ONCE(jiffies);
2974 next = rq->last_sched_tick + HZ;
2976 if (time_before_eq(next, now))
2979 return jiffies_to_nsecs(next - now);
2983 notrace unsigned long get_parent_ip(unsigned long addr)
2985 if (in_lock_functions(addr)) {
2986 addr = CALLER_ADDR2;
2987 if (in_lock_functions(addr))
2988 addr = CALLER_ADDR3;
2993 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2994 defined(CONFIG_PREEMPT_TRACER))
2996 void preempt_count_add(int val)
2998 #ifdef CONFIG_DEBUG_PREEMPT
3002 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3005 __preempt_count_add(val);
3006 #ifdef CONFIG_DEBUG_PREEMPT
3008 * Spinlock count overflowing soon?
3010 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3013 if (preempt_count() == val) {
3014 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3015 #ifdef CONFIG_DEBUG_PREEMPT
3016 current->preempt_disable_ip = ip;
3018 trace_preempt_off(CALLER_ADDR0, ip);
3021 EXPORT_SYMBOL(preempt_count_add);
3022 NOKPROBE_SYMBOL(preempt_count_add);
3024 void preempt_count_sub(int val)
3026 #ifdef CONFIG_DEBUG_PREEMPT
3030 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3033 * Is the spinlock portion underflowing?
3035 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3036 !(preempt_count() & PREEMPT_MASK)))
3040 if (preempt_count() == val)
3041 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3042 __preempt_count_sub(val);
3044 EXPORT_SYMBOL(preempt_count_sub);
3045 NOKPROBE_SYMBOL(preempt_count_sub);
3050 * Print scheduling while atomic bug:
3052 static noinline void __schedule_bug(struct task_struct *prev)
3054 if (oops_in_progress)
3057 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3058 prev->comm, prev->pid, preempt_count());
3060 debug_show_held_locks(prev);
3062 if (irqs_disabled())
3063 print_irqtrace_events(prev);
3064 #ifdef CONFIG_DEBUG_PREEMPT
3065 if (in_atomic_preempt_off()) {
3066 pr_err("Preemption disabled at:");
3067 print_ip_sym(current->preempt_disable_ip);
3072 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3076 * Various schedule()-time debugging checks and statistics:
3078 static inline void schedule_debug(struct task_struct *prev)
3080 #ifdef CONFIG_SCHED_STACK_END_CHECK
3081 if (task_stack_end_corrupted(prev))
3082 panic("corrupted stack end detected inside scheduler\n");
3085 if (unlikely(in_atomic_preempt_off())) {
3086 __schedule_bug(prev);
3087 preempt_count_set(PREEMPT_DISABLED);
3091 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3093 schedstat_inc(this_rq(), sched_count);
3097 * Pick up the highest-prio task:
3099 static inline struct task_struct *
3100 pick_next_task(struct rq *rq, struct task_struct *prev)
3102 const struct sched_class *class = &fair_sched_class;
3103 struct task_struct *p;
3106 * Optimization: we know that if all tasks are in
3107 * the fair class we can call that function directly:
3109 if (likely(prev->sched_class == class &&
3110 rq->nr_running == rq->cfs.h_nr_running)) {
3111 p = fair_sched_class.pick_next_task(rq, prev);
3112 if (unlikely(p == RETRY_TASK))
3115 /* assumes fair_sched_class->next == idle_sched_class */
3117 p = idle_sched_class.pick_next_task(rq, prev);
3123 for_each_class(class) {
3124 p = class->pick_next_task(rq, prev);
3126 if (unlikely(p == RETRY_TASK))
3132 BUG(); /* the idle class will always have a runnable task */
3136 * __schedule() is the main scheduler function.
3138 * The main means of driving the scheduler and thus entering this function are:
3140 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3142 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3143 * paths. For example, see arch/x86/entry_64.S.
3145 * To drive preemption between tasks, the scheduler sets the flag in timer
3146 * interrupt handler scheduler_tick().
3148 * 3. Wakeups don't really cause entry into schedule(). They add a
3149 * task to the run-queue and that's it.
3151 * Now, if the new task added to the run-queue preempts the current
3152 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3153 * called on the nearest possible occasion:
3155 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3157 * - in syscall or exception context, at the next outmost
3158 * preempt_enable(). (this might be as soon as the wake_up()'s
3161 * - in IRQ context, return from interrupt-handler to
3162 * preemptible context
3164 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3167 * - cond_resched() call
3168 * - explicit schedule() call
3169 * - return from syscall or exception to user-space
3170 * - return from interrupt-handler to user-space
3172 * WARNING: must be called with preemption disabled!
3174 static void __sched notrace __schedule(bool preempt)
3176 struct task_struct *prev, *next;
3177 unsigned long *switch_count;
3181 cpu = smp_processor_id();
3183 rcu_note_context_switch();
3187 * do_exit() calls schedule() with preemption disabled as an exception;
3188 * however we must fix that up, otherwise the next task will see an
3189 * inconsistent (higher) preempt count.
3191 * It also avoids the below schedule_debug() test from complaining
3194 if (unlikely(prev->state == TASK_DEAD))
3195 preempt_enable_no_resched_notrace();
3197 schedule_debug(prev);
3199 if (sched_feat(HRTICK))
3203 * Make sure that signal_pending_state()->signal_pending() below
3204 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3205 * done by the caller to avoid the race with signal_wake_up().
3207 smp_mb__before_spinlock();
3208 raw_spin_lock_irq(&rq->lock);
3209 lockdep_pin_lock(&rq->lock);
3211 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3213 switch_count = &prev->nivcsw;
3214 if (!preempt && prev->state) {
3215 if (unlikely(signal_pending_state(prev->state, prev))) {
3216 prev->state = TASK_RUNNING;
3218 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3222 * If a worker went to sleep, notify and ask workqueue
3223 * whether it wants to wake up a task to maintain
3226 if (prev->flags & PF_WQ_WORKER) {
3227 struct task_struct *to_wakeup;
3229 to_wakeup = wq_worker_sleeping(prev, cpu);
3231 try_to_wake_up_local(to_wakeup);
3234 switch_count = &prev->nvcsw;
3237 if (task_on_rq_queued(prev))
3238 update_rq_clock(rq);
3240 next = pick_next_task(rq, prev);
3241 clear_tsk_need_resched(prev);
3242 clear_preempt_need_resched();
3243 rq->clock_skip_update = 0;
3245 if (likely(prev != next)) {
3250 trace_sched_switch(preempt, prev, next);
3251 rq = context_switch(rq, prev, next); /* unlocks the rq */
3254 lockdep_unpin_lock(&rq->lock);
3255 raw_spin_unlock_irq(&rq->lock);
3258 balance_callback(rq);
3261 static inline void sched_submit_work(struct task_struct *tsk)
3263 if (!tsk->state || tsk_is_pi_blocked(tsk))
3266 * If we are going to sleep and we have plugged IO queued,
3267 * make sure to submit it to avoid deadlocks.
3269 if (blk_needs_flush_plug(tsk))
3270 blk_schedule_flush_plug(tsk);
3273 asmlinkage __visible void __sched schedule(void)
3275 struct task_struct *tsk = current;
3277 sched_submit_work(tsk);
3281 sched_preempt_enable_no_resched();
3282 } while (need_resched());
3284 EXPORT_SYMBOL(schedule);
3286 #ifdef CONFIG_CONTEXT_TRACKING
3287 asmlinkage __visible void __sched schedule_user(void)
3290 * If we come here after a random call to set_need_resched(),
3291 * or we have been woken up remotely but the IPI has not yet arrived,
3292 * we haven't yet exited the RCU idle mode. Do it here manually until
3293 * we find a better solution.
3295 * NB: There are buggy callers of this function. Ideally we
3296 * should warn if prev_state != CONTEXT_USER, but that will trigger
3297 * too frequently to make sense yet.
3299 enum ctx_state prev_state = exception_enter();
3301 exception_exit(prev_state);
3306 * schedule_preempt_disabled - called with preemption disabled
3308 * Returns with preemption disabled. Note: preempt_count must be 1
3310 void __sched schedule_preempt_disabled(void)
3312 sched_preempt_enable_no_resched();
3317 static void __sched notrace preempt_schedule_common(void)
3320 preempt_disable_notrace();
3322 preempt_enable_no_resched_notrace();
3325 * Check again in case we missed a preemption opportunity
3326 * between schedule and now.
3328 } while (need_resched());
3331 #ifdef CONFIG_PREEMPT
3333 * this is the entry point to schedule() from in-kernel preemption
3334 * off of preempt_enable. Kernel preemptions off return from interrupt
3335 * occur there and call schedule directly.
3337 asmlinkage __visible void __sched notrace preempt_schedule(void)
3340 * If there is a non-zero preempt_count or interrupts are disabled,
3341 * we do not want to preempt the current task. Just return..
3343 if (likely(!preemptible()))
3346 preempt_schedule_common();
3348 NOKPROBE_SYMBOL(preempt_schedule);
3349 EXPORT_SYMBOL(preempt_schedule);
3352 * preempt_schedule_notrace - preempt_schedule called by tracing
3354 * The tracing infrastructure uses preempt_enable_notrace to prevent
3355 * recursion and tracing preempt enabling caused by the tracing
3356 * infrastructure itself. But as tracing can happen in areas coming
3357 * from userspace or just about to enter userspace, a preempt enable
3358 * can occur before user_exit() is called. This will cause the scheduler
3359 * to be called when the system is still in usermode.
3361 * To prevent this, the preempt_enable_notrace will use this function
3362 * instead of preempt_schedule() to exit user context if needed before
3363 * calling the scheduler.
3365 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3367 enum ctx_state prev_ctx;
3369 if (likely(!preemptible()))
3373 preempt_disable_notrace();
3375 * Needs preempt disabled in case user_exit() is traced
3376 * and the tracer calls preempt_enable_notrace() causing
3377 * an infinite recursion.
3379 prev_ctx = exception_enter();
3381 exception_exit(prev_ctx);
3383 preempt_enable_no_resched_notrace();
3384 } while (need_resched());
3386 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3388 #endif /* CONFIG_PREEMPT */
3391 * this is the entry point to schedule() from kernel preemption
3392 * off of irq context.
3393 * Note, that this is called and return with irqs disabled. This will
3394 * protect us against recursive calling from irq.
3396 asmlinkage __visible void __sched preempt_schedule_irq(void)
3398 enum ctx_state prev_state;
3400 /* Catch callers which need to be fixed */
3401 BUG_ON(preempt_count() || !irqs_disabled());
3403 prev_state = exception_enter();
3409 local_irq_disable();
3410 sched_preempt_enable_no_resched();
3411 } while (need_resched());
3413 exception_exit(prev_state);
3416 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3419 return try_to_wake_up(curr->private, mode, wake_flags);
3421 EXPORT_SYMBOL(default_wake_function);
3423 #ifdef CONFIG_RT_MUTEXES
3426 * rt_mutex_setprio - set the current priority of a task
3428 * @prio: prio value (kernel-internal form)
3430 * This function changes the 'effective' priority of a task. It does
3431 * not touch ->normal_prio like __setscheduler().
3433 * Used by the rt_mutex code to implement priority inheritance
3434 * logic. Call site only calls if the priority of the task changed.
3436 void rt_mutex_setprio(struct task_struct *p, int prio)
3438 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3440 const struct sched_class *prev_class;
3442 BUG_ON(prio > MAX_PRIO);
3444 rq = __task_rq_lock(p);
3447 * Idle task boosting is a nono in general. There is one
3448 * exception, when PREEMPT_RT and NOHZ is active:
3450 * The idle task calls get_next_timer_interrupt() and holds
3451 * the timer wheel base->lock on the CPU and another CPU wants
3452 * to access the timer (probably to cancel it). We can safely
3453 * ignore the boosting request, as the idle CPU runs this code
3454 * with interrupts disabled and will complete the lock
3455 * protected section without being interrupted. So there is no
3456 * real need to boost.
3458 if (unlikely(p == rq->idle)) {
3459 WARN_ON(p != rq->curr);
3460 WARN_ON(p->pi_blocked_on);
3464 trace_sched_pi_setprio(p, prio);
3466 prev_class = p->sched_class;
3467 queued = task_on_rq_queued(p);
3468 running = task_current(rq, p);
3470 dequeue_task(rq, p, DEQUEUE_SAVE);
3472 put_prev_task(rq, p);
3475 * Boosting condition are:
3476 * 1. -rt task is running and holds mutex A
3477 * --> -dl task blocks on mutex A
3479 * 2. -dl task is running and holds mutex A
3480 * --> -dl task blocks on mutex A and could preempt the
3483 if (dl_prio(prio)) {
3484 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3485 if (!dl_prio(p->normal_prio) ||
3486 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3487 p->dl.dl_boosted = 1;
3488 enqueue_flag |= ENQUEUE_REPLENISH;
3490 p->dl.dl_boosted = 0;
3491 p->sched_class = &dl_sched_class;
3492 } else if (rt_prio(prio)) {
3493 if (dl_prio(oldprio))
3494 p->dl.dl_boosted = 0;
3496 enqueue_flag |= ENQUEUE_HEAD;
3497 p->sched_class = &rt_sched_class;
3499 if (dl_prio(oldprio))
3500 p->dl.dl_boosted = 0;
3501 if (rt_prio(oldprio))
3503 p->sched_class = &fair_sched_class;
3509 p->sched_class->set_curr_task(rq);
3511 enqueue_task(rq, p, enqueue_flag);
3513 check_class_changed(rq, p, prev_class, oldprio);
3515 preempt_disable(); /* avoid rq from going away on us */
3516 __task_rq_unlock(rq);
3518 balance_callback(rq);
3523 void set_user_nice(struct task_struct *p, long nice)
3525 int old_prio, delta, queued;
3526 unsigned long flags;
3529 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3532 * We have to be careful, if called from sys_setpriority(),
3533 * the task might be in the middle of scheduling on another CPU.
3535 rq = task_rq_lock(p, &flags);
3537 * The RT priorities are set via sched_setscheduler(), but we still
3538 * allow the 'normal' nice value to be set - but as expected
3539 * it wont have any effect on scheduling until the task is
3540 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3542 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3543 p->static_prio = NICE_TO_PRIO(nice);
3546 queued = task_on_rq_queued(p);
3548 dequeue_task(rq, p, DEQUEUE_SAVE);
3550 p->static_prio = NICE_TO_PRIO(nice);
3553 p->prio = effective_prio(p);
3554 delta = p->prio - old_prio;
3557 enqueue_task(rq, p, ENQUEUE_RESTORE);
3559 * If the task increased its priority or is running and
3560 * lowered its priority, then reschedule its CPU:
3562 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3566 task_rq_unlock(rq, p, &flags);
3568 EXPORT_SYMBOL(set_user_nice);
3571 * can_nice - check if a task can reduce its nice value
3575 int can_nice(const struct task_struct *p, const int nice)
3577 /* convert nice value [19,-20] to rlimit style value [1,40] */
3578 int nice_rlim = nice_to_rlimit(nice);
3580 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3581 capable(CAP_SYS_NICE));
3584 #ifdef __ARCH_WANT_SYS_NICE
3587 * sys_nice - change the priority of the current process.
3588 * @increment: priority increment
3590 * sys_setpriority is a more generic, but much slower function that
3591 * does similar things.
3593 SYSCALL_DEFINE1(nice, int, increment)
3598 * Setpriority might change our priority at the same moment.
3599 * We don't have to worry. Conceptually one call occurs first
3600 * and we have a single winner.
3602 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3603 nice = task_nice(current) + increment;
3605 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3606 if (increment < 0 && !can_nice(current, nice))
3609 retval = security_task_setnice(current, nice);
3613 set_user_nice(current, nice);
3620 * task_prio - return the priority value of a given task.
3621 * @p: the task in question.
3623 * Return: The priority value as seen by users in /proc.
3624 * RT tasks are offset by -200. Normal tasks are centered
3625 * around 0, value goes from -16 to +15.
3627 int task_prio(const struct task_struct *p)
3629 return p->prio - MAX_RT_PRIO;
3633 * idle_cpu - is a given cpu idle currently?
3634 * @cpu: the processor in question.
3636 * Return: 1 if the CPU is currently idle. 0 otherwise.
3638 int idle_cpu(int cpu)
3640 struct rq *rq = cpu_rq(cpu);
3642 if (rq->curr != rq->idle)
3649 if (!llist_empty(&rq->wake_list))
3657 * idle_task - return the idle task for a given cpu.
3658 * @cpu: the processor in question.
3660 * Return: The idle task for the cpu @cpu.
3662 struct task_struct *idle_task(int cpu)
3664 return cpu_rq(cpu)->idle;
3668 * find_process_by_pid - find a process with a matching PID value.
3669 * @pid: the pid in question.
3671 * The task of @pid, if found. %NULL otherwise.
3673 static struct task_struct *find_process_by_pid(pid_t pid)
3675 return pid ? find_task_by_vpid(pid) : current;
3679 * This function initializes the sched_dl_entity of a newly becoming
3680 * SCHED_DEADLINE task.
3682 * Only the static values are considered here, the actual runtime and the
3683 * absolute deadline will be properly calculated when the task is enqueued
3684 * for the first time with its new policy.
3687 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3689 struct sched_dl_entity *dl_se = &p->dl;
3691 dl_se->dl_runtime = attr->sched_runtime;
3692 dl_se->dl_deadline = attr->sched_deadline;
3693 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3694 dl_se->flags = attr->sched_flags;
3695 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3698 * Changing the parameters of a task is 'tricky' and we're not doing
3699 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3701 * What we SHOULD do is delay the bandwidth release until the 0-lag
3702 * point. This would include retaining the task_struct until that time
3703 * and change dl_overflow() to not immediately decrement the current
3706 * Instead we retain the current runtime/deadline and let the new
3707 * parameters take effect after the current reservation period lapses.
3708 * This is safe (albeit pessimistic) because the 0-lag point is always
3709 * before the current scheduling deadline.
3711 * We can still have temporary overloads because we do not delay the
3712 * change in bandwidth until that time; so admission control is
3713 * not on the safe side. It does however guarantee tasks will never
3714 * consume more than promised.
3719 * sched_setparam() passes in -1 for its policy, to let the functions
3720 * it calls know not to change it.
3722 #define SETPARAM_POLICY -1
3724 static void __setscheduler_params(struct task_struct *p,
3725 const struct sched_attr *attr)
3727 int policy = attr->sched_policy;
3729 if (policy == SETPARAM_POLICY)
3734 if (dl_policy(policy))
3735 __setparam_dl(p, attr);
3736 else if (fair_policy(policy))
3737 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3740 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3741 * !rt_policy. Always setting this ensures that things like
3742 * getparam()/getattr() don't report silly values for !rt tasks.
3744 p->rt_priority = attr->sched_priority;
3745 p->normal_prio = normal_prio(p);
3749 /* Actually do priority change: must hold pi & rq lock. */
3750 static void __setscheduler(struct rq *rq, struct task_struct *p,
3751 const struct sched_attr *attr, bool keep_boost)
3753 __setscheduler_params(p, attr);
3756 * Keep a potential priority boosting if called from
3757 * sched_setscheduler().
3760 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3762 p->prio = normal_prio(p);
3764 if (dl_prio(p->prio))
3765 p->sched_class = &dl_sched_class;
3766 else if (rt_prio(p->prio))
3767 p->sched_class = &rt_sched_class;
3769 p->sched_class = &fair_sched_class;
3773 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3775 struct sched_dl_entity *dl_se = &p->dl;
3777 attr->sched_priority = p->rt_priority;
3778 attr->sched_runtime = dl_se->dl_runtime;
3779 attr->sched_deadline = dl_se->dl_deadline;
3780 attr->sched_period = dl_se->dl_period;
3781 attr->sched_flags = dl_se->flags;
3785 * This function validates the new parameters of a -deadline task.
3786 * We ask for the deadline not being zero, and greater or equal
3787 * than the runtime, as well as the period of being zero or
3788 * greater than deadline. Furthermore, we have to be sure that
3789 * user parameters are above the internal resolution of 1us (we
3790 * check sched_runtime only since it is always the smaller one) and
3791 * below 2^63 ns (we have to check both sched_deadline and
3792 * sched_period, as the latter can be zero).
3795 __checkparam_dl(const struct sched_attr *attr)
3798 if (attr->sched_deadline == 0)
3802 * Since we truncate DL_SCALE bits, make sure we're at least
3805 if (attr->sched_runtime < (1ULL << DL_SCALE))
3809 * Since we use the MSB for wrap-around and sign issues, make
3810 * sure it's not set (mind that period can be equal to zero).
3812 if (attr->sched_deadline & (1ULL << 63) ||
3813 attr->sched_period & (1ULL << 63))
3816 /* runtime <= deadline <= period (if period != 0) */
3817 if ((attr->sched_period != 0 &&
3818 attr->sched_period < attr->sched_deadline) ||
3819 attr->sched_deadline < attr->sched_runtime)
3826 * check the target process has a UID that matches the current process's
3828 static bool check_same_owner(struct task_struct *p)
3830 const struct cred *cred = current_cred(), *pcred;
3834 pcred = __task_cred(p);
3835 match = (uid_eq(cred->euid, pcred->euid) ||
3836 uid_eq(cred->euid, pcred->uid));
3841 static bool dl_param_changed(struct task_struct *p,
3842 const struct sched_attr *attr)
3844 struct sched_dl_entity *dl_se = &p->dl;
3846 if (dl_se->dl_runtime != attr->sched_runtime ||
3847 dl_se->dl_deadline != attr->sched_deadline ||
3848 dl_se->dl_period != attr->sched_period ||
3849 dl_se->flags != attr->sched_flags)
3855 static int __sched_setscheduler(struct task_struct *p,
3856 const struct sched_attr *attr,
3859 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3860 MAX_RT_PRIO - 1 - attr->sched_priority;
3861 int retval, oldprio, oldpolicy = -1, queued, running;
3862 int new_effective_prio, policy = attr->sched_policy;
3863 unsigned long flags;
3864 const struct sched_class *prev_class;
3868 /* may grab non-irq protected spin_locks */
3869 BUG_ON(in_interrupt());
3871 /* double check policy once rq lock held */
3873 reset_on_fork = p->sched_reset_on_fork;
3874 policy = oldpolicy = p->policy;
3876 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3878 if (!valid_policy(policy))
3882 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3886 * Valid priorities for SCHED_FIFO and SCHED_RR are
3887 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3888 * SCHED_BATCH and SCHED_IDLE is 0.
3890 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3891 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3893 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3894 (rt_policy(policy) != (attr->sched_priority != 0)))
3898 * Allow unprivileged RT tasks to decrease priority:
3900 if (user && !capable(CAP_SYS_NICE)) {
3901 if (fair_policy(policy)) {
3902 if (attr->sched_nice < task_nice(p) &&
3903 !can_nice(p, attr->sched_nice))
3907 if (rt_policy(policy)) {
3908 unsigned long rlim_rtprio =
3909 task_rlimit(p, RLIMIT_RTPRIO);
3911 /* can't set/change the rt policy */
3912 if (policy != p->policy && !rlim_rtprio)
3915 /* can't increase priority */
3916 if (attr->sched_priority > p->rt_priority &&
3917 attr->sched_priority > rlim_rtprio)
3922 * Can't set/change SCHED_DEADLINE policy at all for now
3923 * (safest behavior); in the future we would like to allow
3924 * unprivileged DL tasks to increase their relative deadline
3925 * or reduce their runtime (both ways reducing utilization)
3927 if (dl_policy(policy))
3931 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3932 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3934 if (idle_policy(p->policy) && !idle_policy(policy)) {
3935 if (!can_nice(p, task_nice(p)))
3939 /* can't change other user's priorities */
3940 if (!check_same_owner(p))
3943 /* Normal users shall not reset the sched_reset_on_fork flag */
3944 if (p->sched_reset_on_fork && !reset_on_fork)
3949 retval = security_task_setscheduler(p);
3955 * make sure no PI-waiters arrive (or leave) while we are
3956 * changing the priority of the task:
3958 * To be able to change p->policy safely, the appropriate
3959 * runqueue lock must be held.
3961 rq = task_rq_lock(p, &flags);
3964 * Changing the policy of the stop threads its a very bad idea
3966 if (p == rq->stop) {
3967 task_rq_unlock(rq, p, &flags);
3972 * If not changing anything there's no need to proceed further,
3973 * but store a possible modification of reset_on_fork.
3975 if (unlikely(policy == p->policy)) {
3976 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3978 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3980 if (dl_policy(policy) && dl_param_changed(p, attr))
3983 p->sched_reset_on_fork = reset_on_fork;
3984 task_rq_unlock(rq, p, &flags);
3990 #ifdef CONFIG_RT_GROUP_SCHED
3992 * Do not allow realtime tasks into groups that have no runtime
3995 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3996 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3997 !task_group_is_autogroup(task_group(p))) {
3998 task_rq_unlock(rq, p, &flags);
4003 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4004 cpumask_t *span = rq->rd->span;
4007 * Don't allow tasks with an affinity mask smaller than
4008 * the entire root_domain to become SCHED_DEADLINE. We
4009 * will also fail if there's no bandwidth available.
4011 if (!cpumask_subset(span, &p->cpus_allowed) ||
4012 rq->rd->dl_bw.bw == 0) {
4013 task_rq_unlock(rq, p, &flags);
4020 /* recheck policy now with rq lock held */
4021 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4022 policy = oldpolicy = -1;
4023 task_rq_unlock(rq, p, &flags);
4028 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4029 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4032 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4033 task_rq_unlock(rq, p, &flags);
4037 p->sched_reset_on_fork = reset_on_fork;
4042 * Take priority boosted tasks into account. If the new
4043 * effective priority is unchanged, we just store the new
4044 * normal parameters and do not touch the scheduler class and
4045 * the runqueue. This will be done when the task deboost
4048 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4049 if (new_effective_prio == oldprio) {
4050 __setscheduler_params(p, attr);
4051 task_rq_unlock(rq, p, &flags);
4056 queued = task_on_rq_queued(p);
4057 running = task_current(rq, p);
4059 dequeue_task(rq, p, DEQUEUE_SAVE);
4061 put_prev_task(rq, p);
4063 prev_class = p->sched_class;
4064 __setscheduler(rq, p, attr, pi);
4067 p->sched_class->set_curr_task(rq);
4069 int enqueue_flags = ENQUEUE_RESTORE;
4071 * We enqueue to tail when the priority of a task is
4072 * increased (user space view).
4074 if (oldprio <= p->prio)
4075 enqueue_flags |= ENQUEUE_HEAD;
4077 enqueue_task(rq, p, enqueue_flags);
4080 check_class_changed(rq, p, prev_class, oldprio);
4081 preempt_disable(); /* avoid rq from going away on us */
4082 task_rq_unlock(rq, p, &flags);
4085 rt_mutex_adjust_pi(p);
4088 * Run balance callbacks after we've adjusted the PI chain.
4090 balance_callback(rq);
4096 static int _sched_setscheduler(struct task_struct *p, int policy,
4097 const struct sched_param *param, bool check)
4099 struct sched_attr attr = {
4100 .sched_policy = policy,
4101 .sched_priority = param->sched_priority,
4102 .sched_nice = PRIO_TO_NICE(p->static_prio),
4105 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4106 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4107 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4108 policy &= ~SCHED_RESET_ON_FORK;
4109 attr.sched_policy = policy;
4112 return __sched_setscheduler(p, &attr, check, true);
4115 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4116 * @p: the task in question.
4117 * @policy: new policy.
4118 * @param: structure containing the new RT priority.
4120 * Return: 0 on success. An error code otherwise.
4122 * NOTE that the task may be already dead.
4124 int sched_setscheduler(struct task_struct *p, int policy,
4125 const struct sched_param *param)
4127 return _sched_setscheduler(p, policy, param, true);
4129 EXPORT_SYMBOL_GPL(sched_setscheduler);
4131 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4133 return __sched_setscheduler(p, attr, true, true);
4135 EXPORT_SYMBOL_GPL(sched_setattr);
4138 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4139 * @p: the task in question.
4140 * @policy: new policy.
4141 * @param: structure containing the new RT priority.
4143 * Just like sched_setscheduler, only don't bother checking if the
4144 * current context has permission. For example, this is needed in
4145 * stop_machine(): we create temporary high priority worker threads,
4146 * but our caller might not have that capability.
4148 * Return: 0 on success. An error code otherwise.
4150 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4151 const struct sched_param *param)
4153 return _sched_setscheduler(p, policy, param, false);
4155 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4158 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4160 struct sched_param lparam;
4161 struct task_struct *p;
4164 if (!param || pid < 0)
4166 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4171 p = find_process_by_pid(pid);
4173 retval = sched_setscheduler(p, policy, &lparam);
4180 * Mimics kernel/events/core.c perf_copy_attr().
4182 static int sched_copy_attr(struct sched_attr __user *uattr,
4183 struct sched_attr *attr)
4188 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4192 * zero the full structure, so that a short copy will be nice.
4194 memset(attr, 0, sizeof(*attr));
4196 ret = get_user(size, &uattr->size);
4200 if (size > PAGE_SIZE) /* silly large */
4203 if (!size) /* abi compat */
4204 size = SCHED_ATTR_SIZE_VER0;
4206 if (size < SCHED_ATTR_SIZE_VER0)
4210 * If we're handed a bigger struct than we know of,
4211 * ensure all the unknown bits are 0 - i.e. new
4212 * user-space does not rely on any kernel feature
4213 * extensions we dont know about yet.
4215 if (size > sizeof(*attr)) {
4216 unsigned char __user *addr;
4217 unsigned char __user *end;
4220 addr = (void __user *)uattr + sizeof(*attr);
4221 end = (void __user *)uattr + size;
4223 for (; addr < end; addr++) {
4224 ret = get_user(val, addr);
4230 size = sizeof(*attr);
4233 ret = copy_from_user(attr, uattr, size);
4238 * XXX: do we want to be lenient like existing syscalls; or do we want
4239 * to be strict and return an error on out-of-bounds values?
4241 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4246 put_user(sizeof(*attr), &uattr->size);
4251 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4252 * @pid: the pid in question.
4253 * @policy: new policy.
4254 * @param: structure containing the new RT priority.
4256 * Return: 0 on success. An error code otherwise.
4258 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4259 struct sched_param __user *, param)
4261 /* negative values for policy are not valid */
4265 return do_sched_setscheduler(pid, policy, param);
4269 * sys_sched_setparam - set/change the RT priority of a thread
4270 * @pid: the pid in question.
4271 * @param: structure containing the new RT priority.
4273 * Return: 0 on success. An error code otherwise.
4275 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4277 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4281 * sys_sched_setattr - same as above, but with extended sched_attr
4282 * @pid: the pid in question.
4283 * @uattr: structure containing the extended parameters.
4284 * @flags: for future extension.
4286 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4287 unsigned int, flags)
4289 struct sched_attr attr;
4290 struct task_struct *p;
4293 if (!uattr || pid < 0 || flags)
4296 retval = sched_copy_attr(uattr, &attr);
4300 if ((int)attr.sched_policy < 0)
4305 p = find_process_by_pid(pid);
4307 retval = sched_setattr(p, &attr);
4314 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4315 * @pid: the pid in question.
4317 * Return: On success, the policy of the thread. Otherwise, a negative error
4320 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4322 struct task_struct *p;
4330 p = find_process_by_pid(pid);
4332 retval = security_task_getscheduler(p);
4335 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4342 * sys_sched_getparam - get the RT priority of a thread
4343 * @pid: the pid in question.
4344 * @param: structure containing the RT priority.
4346 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4349 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4351 struct sched_param lp = { .sched_priority = 0 };
4352 struct task_struct *p;
4355 if (!param || pid < 0)
4359 p = find_process_by_pid(pid);
4364 retval = security_task_getscheduler(p);
4368 if (task_has_rt_policy(p))
4369 lp.sched_priority = p->rt_priority;
4373 * This one might sleep, we cannot do it with a spinlock held ...
4375 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4384 static int sched_read_attr(struct sched_attr __user *uattr,
4385 struct sched_attr *attr,
4390 if (!access_ok(VERIFY_WRITE, uattr, usize))
4394 * If we're handed a smaller struct than we know of,
4395 * ensure all the unknown bits are 0 - i.e. old
4396 * user-space does not get uncomplete information.
4398 if (usize < sizeof(*attr)) {
4399 unsigned char *addr;
4402 addr = (void *)attr + usize;
4403 end = (void *)attr + sizeof(*attr);
4405 for (; addr < end; addr++) {
4413 ret = copy_to_user(uattr, attr, attr->size);
4421 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4422 * @pid: the pid in question.
4423 * @uattr: structure containing the extended parameters.
4424 * @size: sizeof(attr) for fwd/bwd comp.
4425 * @flags: for future extension.
4427 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4428 unsigned int, size, unsigned int, flags)
4430 struct sched_attr attr = {
4431 .size = sizeof(struct sched_attr),
4433 struct task_struct *p;
4436 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4437 size < SCHED_ATTR_SIZE_VER0 || flags)
4441 p = find_process_by_pid(pid);
4446 retval = security_task_getscheduler(p);
4450 attr.sched_policy = p->policy;
4451 if (p->sched_reset_on_fork)
4452 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4453 if (task_has_dl_policy(p))
4454 __getparam_dl(p, &attr);
4455 else if (task_has_rt_policy(p))
4456 attr.sched_priority = p->rt_priority;
4458 attr.sched_nice = task_nice(p);
4462 retval = sched_read_attr(uattr, &attr, size);
4470 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4472 cpumask_var_t cpus_allowed, new_mask;
4473 struct task_struct *p;
4478 p = find_process_by_pid(pid);
4484 /* Prevent p going away */
4488 if (p->flags & PF_NO_SETAFFINITY) {
4492 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4496 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4498 goto out_free_cpus_allowed;
4501 if (!check_same_owner(p)) {
4503 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4505 goto out_free_new_mask;
4510 retval = security_task_setscheduler(p);
4512 goto out_free_new_mask;
4515 cpuset_cpus_allowed(p, cpus_allowed);
4516 cpumask_and(new_mask, in_mask, cpus_allowed);
4519 * Since bandwidth control happens on root_domain basis,
4520 * if admission test is enabled, we only admit -deadline
4521 * tasks allowed to run on all the CPUs in the task's
4525 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4527 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4530 goto out_free_new_mask;
4536 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4539 cpuset_cpus_allowed(p, cpus_allowed);
4540 if (!cpumask_subset(new_mask, cpus_allowed)) {
4542 * We must have raced with a concurrent cpuset
4543 * update. Just reset the cpus_allowed to the
4544 * cpuset's cpus_allowed
4546 cpumask_copy(new_mask, cpus_allowed);
4551 free_cpumask_var(new_mask);
4552 out_free_cpus_allowed:
4553 free_cpumask_var(cpus_allowed);
4559 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4560 struct cpumask *new_mask)
4562 if (len < cpumask_size())
4563 cpumask_clear(new_mask);
4564 else if (len > cpumask_size())
4565 len = cpumask_size();
4567 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4571 * sys_sched_setaffinity - set the cpu affinity of a process
4572 * @pid: pid of the process
4573 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4574 * @user_mask_ptr: user-space pointer to the new cpu mask
4576 * Return: 0 on success. An error code otherwise.
4578 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4579 unsigned long __user *, user_mask_ptr)
4581 cpumask_var_t new_mask;
4584 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4587 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4589 retval = sched_setaffinity(pid, new_mask);
4590 free_cpumask_var(new_mask);
4594 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4596 struct task_struct *p;
4597 unsigned long flags;
4603 p = find_process_by_pid(pid);
4607 retval = security_task_getscheduler(p);
4611 raw_spin_lock_irqsave(&p->pi_lock, flags);
4612 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4613 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4622 * sys_sched_getaffinity - get the cpu affinity of a process
4623 * @pid: pid of the process
4624 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4625 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4627 * Return: 0 on success. An error code otherwise.
4629 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4630 unsigned long __user *, user_mask_ptr)
4635 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4637 if (len & (sizeof(unsigned long)-1))
4640 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4643 ret = sched_getaffinity(pid, mask);
4645 size_t retlen = min_t(size_t, len, cpumask_size());
4647 if (copy_to_user(user_mask_ptr, mask, retlen))
4652 free_cpumask_var(mask);
4658 * sys_sched_yield - yield the current processor to other threads.
4660 * This function yields the current CPU to other tasks. If there are no
4661 * other threads running on this CPU then this function will return.
4665 SYSCALL_DEFINE0(sched_yield)
4667 struct rq *rq = this_rq_lock();
4669 schedstat_inc(rq, yld_count);
4670 current->sched_class->yield_task(rq);
4673 * Since we are going to call schedule() anyway, there's
4674 * no need to preempt or enable interrupts:
4676 __release(rq->lock);
4677 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4678 do_raw_spin_unlock(&rq->lock);
4679 sched_preempt_enable_no_resched();
4686 int __sched _cond_resched(void)
4688 if (should_resched(0)) {
4689 preempt_schedule_common();
4694 EXPORT_SYMBOL(_cond_resched);
4697 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4698 * call schedule, and on return reacquire the lock.
4700 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4701 * operations here to prevent schedule() from being called twice (once via
4702 * spin_unlock(), once by hand).
4704 int __cond_resched_lock(spinlock_t *lock)
4706 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4709 lockdep_assert_held(lock);
4711 if (spin_needbreak(lock) || resched) {
4714 preempt_schedule_common();
4722 EXPORT_SYMBOL(__cond_resched_lock);
4724 int __sched __cond_resched_softirq(void)
4726 BUG_ON(!in_softirq());
4728 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4730 preempt_schedule_common();
4736 EXPORT_SYMBOL(__cond_resched_softirq);
4739 * yield - yield the current processor to other threads.
4741 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4743 * The scheduler is at all times free to pick the calling task as the most
4744 * eligible task to run, if removing the yield() call from your code breaks
4745 * it, its already broken.
4747 * Typical broken usage is:
4752 * where one assumes that yield() will let 'the other' process run that will
4753 * make event true. If the current task is a SCHED_FIFO task that will never
4754 * happen. Never use yield() as a progress guarantee!!
4756 * If you want to use yield() to wait for something, use wait_event().
4757 * If you want to use yield() to be 'nice' for others, use cond_resched().
4758 * If you still want to use yield(), do not!
4760 void __sched yield(void)
4762 set_current_state(TASK_RUNNING);
4765 EXPORT_SYMBOL(yield);
4768 * yield_to - yield the current processor to another thread in
4769 * your thread group, or accelerate that thread toward the
4770 * processor it's on.
4772 * @preempt: whether task preemption is allowed or not
4774 * It's the caller's job to ensure that the target task struct
4775 * can't go away on us before we can do any checks.
4778 * true (>0) if we indeed boosted the target task.
4779 * false (0) if we failed to boost the target.
4780 * -ESRCH if there's no task to yield to.
4782 int __sched yield_to(struct task_struct *p, bool preempt)
4784 struct task_struct *curr = current;
4785 struct rq *rq, *p_rq;
4786 unsigned long flags;
4789 local_irq_save(flags);
4795 * If we're the only runnable task on the rq and target rq also
4796 * has only one task, there's absolutely no point in yielding.
4798 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4803 double_rq_lock(rq, p_rq);
4804 if (task_rq(p) != p_rq) {
4805 double_rq_unlock(rq, p_rq);
4809 if (!curr->sched_class->yield_to_task)
4812 if (curr->sched_class != p->sched_class)
4815 if (task_running(p_rq, p) || p->state)
4818 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4820 schedstat_inc(rq, yld_count);
4822 * Make p's CPU reschedule; pick_next_entity takes care of
4825 if (preempt && rq != p_rq)
4830 double_rq_unlock(rq, p_rq);
4832 local_irq_restore(flags);
4839 EXPORT_SYMBOL_GPL(yield_to);
4842 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4843 * that process accounting knows that this is a task in IO wait state.
4845 long __sched io_schedule_timeout(long timeout)
4847 int old_iowait = current->in_iowait;
4851 current->in_iowait = 1;
4852 blk_schedule_flush_plug(current);
4854 delayacct_blkio_start();
4856 atomic_inc(&rq->nr_iowait);
4857 ret = schedule_timeout(timeout);
4858 current->in_iowait = old_iowait;
4859 atomic_dec(&rq->nr_iowait);
4860 delayacct_blkio_end();
4864 EXPORT_SYMBOL(io_schedule_timeout);
4867 * sys_sched_get_priority_max - return maximum RT priority.
4868 * @policy: scheduling class.
4870 * Return: On success, this syscall returns the maximum
4871 * rt_priority that can be used by a given scheduling class.
4872 * On failure, a negative error code is returned.
4874 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4881 ret = MAX_USER_RT_PRIO-1;
4883 case SCHED_DEADLINE:
4894 * sys_sched_get_priority_min - return minimum RT priority.
4895 * @policy: scheduling class.
4897 * Return: On success, this syscall returns the minimum
4898 * rt_priority that can be used by a given scheduling class.
4899 * On failure, a negative error code is returned.
4901 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4910 case SCHED_DEADLINE:
4920 * sys_sched_rr_get_interval - return the default timeslice of a process.
4921 * @pid: pid of the process.
4922 * @interval: userspace pointer to the timeslice value.
4924 * this syscall writes the default timeslice value of a given process
4925 * into the user-space timespec buffer. A value of '0' means infinity.
4927 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4930 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4931 struct timespec __user *, interval)
4933 struct task_struct *p;
4934 unsigned int time_slice;
4935 unsigned long flags;
4945 p = find_process_by_pid(pid);
4949 retval = security_task_getscheduler(p);
4953 rq = task_rq_lock(p, &flags);
4955 if (p->sched_class->get_rr_interval)
4956 time_slice = p->sched_class->get_rr_interval(rq, p);
4957 task_rq_unlock(rq, p, &flags);
4960 jiffies_to_timespec(time_slice, &t);
4961 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4969 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4971 void sched_show_task(struct task_struct *p)
4973 unsigned long free = 0;
4975 unsigned long state = p->state;
4978 state = __ffs(state) + 1;
4979 printk(KERN_INFO "%-15.15s %c", p->comm,
4980 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4981 #if BITS_PER_LONG == 32
4982 if (state == TASK_RUNNING)
4983 printk(KERN_CONT " running ");
4985 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4987 if (state == TASK_RUNNING)
4988 printk(KERN_CONT " running task ");
4990 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4992 #ifdef CONFIG_DEBUG_STACK_USAGE
4993 free = stack_not_used(p);
4998 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5000 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5001 task_pid_nr(p), ppid,
5002 (unsigned long)task_thread_info(p)->flags);
5004 print_worker_info(KERN_INFO, p);
5005 show_stack(p, NULL);
5008 void show_state_filter(unsigned long state_filter)
5010 struct task_struct *g, *p;
5012 #if BITS_PER_LONG == 32
5014 " task PC stack pid father\n");
5017 " task PC stack pid father\n");
5020 for_each_process_thread(g, p) {
5022 * reset the NMI-timeout, listing all files on a slow
5023 * console might take a lot of time:
5024 * Also, reset softlockup watchdogs on all CPUs, because
5025 * another CPU might be blocked waiting for us to process
5028 touch_nmi_watchdog();
5029 touch_all_softlockup_watchdogs();
5030 if (!state_filter || (p->state & state_filter))
5034 #ifdef CONFIG_SCHED_DEBUG
5035 sysrq_sched_debug_show();
5039 * Only show locks if all tasks are dumped:
5042 debug_show_all_locks();
5045 void init_idle_bootup_task(struct task_struct *idle)
5047 idle->sched_class = &idle_sched_class;
5051 * init_idle - set up an idle thread for a given CPU
5052 * @idle: task in question
5053 * @cpu: cpu the idle task belongs to
5055 * NOTE: this function does not set the idle thread's NEED_RESCHED
5056 * flag, to make booting more robust.
5058 void init_idle(struct task_struct *idle, int cpu)
5060 struct rq *rq = cpu_rq(cpu);
5061 unsigned long flags;
5063 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5064 raw_spin_lock(&rq->lock);
5066 __sched_fork(0, idle);
5067 idle->state = TASK_RUNNING;
5068 idle->se.exec_start = sched_clock();
5072 * Its possible that init_idle() gets called multiple times on a task,
5073 * in that case do_set_cpus_allowed() will not do the right thing.
5075 * And since this is boot we can forgo the serialization.
5077 set_cpus_allowed_common(idle, cpumask_of(cpu));
5080 * We're having a chicken and egg problem, even though we are
5081 * holding rq->lock, the cpu isn't yet set to this cpu so the
5082 * lockdep check in task_group() will fail.
5084 * Similar case to sched_fork(). / Alternatively we could
5085 * use task_rq_lock() here and obtain the other rq->lock.
5090 __set_task_cpu(idle, cpu);
5093 rq->curr = rq->idle = idle;
5094 idle->on_rq = TASK_ON_RQ_QUEUED;
5098 raw_spin_unlock(&rq->lock);
5099 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5101 /* Set the preempt count _outside_ the spinlocks! */
5102 init_idle_preempt_count(idle, cpu);
5105 * The idle tasks have their own, simple scheduling class:
5107 idle->sched_class = &idle_sched_class;
5108 ftrace_graph_init_idle_task(idle, cpu);
5109 vtime_init_idle(idle, cpu);
5111 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5115 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5116 const struct cpumask *trial)
5118 int ret = 1, trial_cpus;
5119 struct dl_bw *cur_dl_b;
5120 unsigned long flags;
5122 if (!cpumask_weight(cur))
5125 rcu_read_lock_sched();
5126 cur_dl_b = dl_bw_of(cpumask_any(cur));
5127 trial_cpus = cpumask_weight(trial);
5129 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5130 if (cur_dl_b->bw != -1 &&
5131 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5133 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5134 rcu_read_unlock_sched();
5139 int task_can_attach(struct task_struct *p,
5140 const struct cpumask *cs_cpus_allowed)
5145 * Kthreads which disallow setaffinity shouldn't be moved
5146 * to a new cpuset; we don't want to change their cpu
5147 * affinity and isolating such threads by their set of
5148 * allowed nodes is unnecessary. Thus, cpusets are not
5149 * applicable for such threads. This prevents checking for
5150 * success of set_cpus_allowed_ptr() on all attached tasks
5151 * before cpus_allowed may be changed.
5153 if (p->flags & PF_NO_SETAFFINITY) {
5159 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5161 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5166 unsigned long flags;
5168 rcu_read_lock_sched();
5169 dl_b = dl_bw_of(dest_cpu);
5170 raw_spin_lock_irqsave(&dl_b->lock, flags);
5171 cpus = dl_bw_cpus(dest_cpu);
5172 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5177 * We reserve space for this task in the destination
5178 * root_domain, as we can't fail after this point.
5179 * We will free resources in the source root_domain
5180 * later on (see set_cpus_allowed_dl()).
5182 __dl_add(dl_b, p->dl.dl_bw);
5184 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5185 rcu_read_unlock_sched();
5195 #ifdef CONFIG_NUMA_BALANCING
5196 /* Migrate current task p to target_cpu */
5197 int migrate_task_to(struct task_struct *p, int target_cpu)
5199 struct migration_arg arg = { p, target_cpu };
5200 int curr_cpu = task_cpu(p);
5202 if (curr_cpu == target_cpu)
5205 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5208 /* TODO: This is not properly updating schedstats */
5210 trace_sched_move_numa(p, curr_cpu, target_cpu);
5211 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5215 * Requeue a task on a given node and accurately track the number of NUMA
5216 * tasks on the runqueues
5218 void sched_setnuma(struct task_struct *p, int nid)
5221 unsigned long flags;
5222 bool queued, running;
5224 rq = task_rq_lock(p, &flags);
5225 queued = task_on_rq_queued(p);
5226 running = task_current(rq, p);
5229 dequeue_task(rq, p, DEQUEUE_SAVE);
5231 put_prev_task(rq, p);
5233 p->numa_preferred_nid = nid;
5236 p->sched_class->set_curr_task(rq);
5238 enqueue_task(rq, p, ENQUEUE_RESTORE);
5239 task_rq_unlock(rq, p, &flags);
5241 #endif /* CONFIG_NUMA_BALANCING */
5243 #ifdef CONFIG_HOTPLUG_CPU
5245 * Ensures that the idle task is using init_mm right before its cpu goes
5248 void idle_task_exit(void)
5250 struct mm_struct *mm = current->active_mm;
5252 BUG_ON(cpu_online(smp_processor_id()));
5254 if (mm != &init_mm) {
5255 switch_mm(mm, &init_mm, current);
5256 finish_arch_post_lock_switch();
5262 * Since this CPU is going 'away' for a while, fold any nr_active delta
5263 * we might have. Assumes we're called after migrate_tasks() so that the
5264 * nr_active count is stable.
5266 * Also see the comment "Global load-average calculations".
5268 static void calc_load_migrate(struct rq *rq)
5270 long delta = calc_load_fold_active(rq);
5272 atomic_long_add(delta, &calc_load_tasks);
5275 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5279 static const struct sched_class fake_sched_class = {
5280 .put_prev_task = put_prev_task_fake,
5283 static struct task_struct fake_task = {
5285 * Avoid pull_{rt,dl}_task()
5287 .prio = MAX_PRIO + 1,
5288 .sched_class = &fake_sched_class,
5292 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5293 * try_to_wake_up()->select_task_rq().
5295 * Called with rq->lock held even though we'er in stop_machine() and
5296 * there's no concurrency possible, we hold the required locks anyway
5297 * because of lock validation efforts.
5299 static void migrate_tasks(struct rq *dead_rq)
5301 struct rq *rq = dead_rq;
5302 struct task_struct *next, *stop = rq->stop;
5306 * Fudge the rq selection such that the below task selection loop
5307 * doesn't get stuck on the currently eligible stop task.
5309 * We're currently inside stop_machine() and the rq is either stuck
5310 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5311 * either way we should never end up calling schedule() until we're
5317 * put_prev_task() and pick_next_task() sched
5318 * class method both need to have an up-to-date
5319 * value of rq->clock[_task]
5321 update_rq_clock(rq);
5325 * There's this thread running, bail when that's the only
5328 if (rq->nr_running == 1)
5332 * pick_next_task assumes pinned rq->lock.
5334 lockdep_pin_lock(&rq->lock);
5335 next = pick_next_task(rq, &fake_task);
5337 next->sched_class->put_prev_task(rq, next);
5340 * Rules for changing task_struct::cpus_allowed are holding
5341 * both pi_lock and rq->lock, such that holding either
5342 * stabilizes the mask.
5344 * Drop rq->lock is not quite as disastrous as it usually is
5345 * because !cpu_active at this point, which means load-balance
5346 * will not interfere. Also, stop-machine.
5348 lockdep_unpin_lock(&rq->lock);
5349 raw_spin_unlock(&rq->lock);
5350 raw_spin_lock(&next->pi_lock);
5351 raw_spin_lock(&rq->lock);
5354 * Since we're inside stop-machine, _nothing_ should have
5355 * changed the task, WARN if weird stuff happened, because in
5356 * that case the above rq->lock drop is a fail too.
5358 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5359 raw_spin_unlock(&next->pi_lock);
5363 /* Find suitable destination for @next, with force if needed. */
5364 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5366 rq = __migrate_task(rq, next, dest_cpu);
5367 if (rq != dead_rq) {
5368 raw_spin_unlock(&rq->lock);
5370 raw_spin_lock(&rq->lock);
5372 raw_spin_unlock(&next->pi_lock);
5377 #endif /* CONFIG_HOTPLUG_CPU */
5379 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5381 static struct ctl_table sd_ctl_dir[] = {
5383 .procname = "sched_domain",
5389 static struct ctl_table sd_ctl_root[] = {
5391 .procname = "kernel",
5393 .child = sd_ctl_dir,
5398 static struct ctl_table *sd_alloc_ctl_entry(int n)
5400 struct ctl_table *entry =
5401 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5406 static void sd_free_ctl_entry(struct ctl_table **tablep)
5408 struct ctl_table *entry;
5411 * In the intermediate directories, both the child directory and
5412 * procname are dynamically allocated and could fail but the mode
5413 * will always be set. In the lowest directory the names are
5414 * static strings and all have proc handlers.
5416 for (entry = *tablep; entry->mode; entry++) {
5418 sd_free_ctl_entry(&entry->child);
5419 if (entry->proc_handler == NULL)
5420 kfree(entry->procname);
5427 static int min_load_idx = 0;
5428 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5431 set_table_entry(struct ctl_table *entry,
5432 const char *procname, void *data, int maxlen,
5433 umode_t mode, proc_handler *proc_handler,
5436 entry->procname = procname;
5438 entry->maxlen = maxlen;
5440 entry->proc_handler = proc_handler;
5443 entry->extra1 = &min_load_idx;
5444 entry->extra2 = &max_load_idx;
5448 static struct ctl_table *
5449 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5451 struct ctl_table *table = sd_alloc_ctl_entry(5);
5456 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5457 sizeof(int), 0644, proc_dointvec_minmax, false);
5458 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5459 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5460 proc_doulongvec_minmax, false);
5461 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5462 sizeof(int), 0644, proc_dointvec_minmax, false);
5463 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5464 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5465 proc_doulongvec_minmax, false);
5470 static struct ctl_table *
5471 sd_alloc_ctl_group_table(struct sched_group *sg)
5473 struct ctl_table *table = sd_alloc_ctl_entry(2);
5478 table->procname = kstrdup("energy", GFP_KERNEL);
5480 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5485 static struct ctl_table *
5486 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5488 struct ctl_table *table;
5489 unsigned int nr_entries = 14;
5492 struct sched_group *sg = sd->groups;
5497 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5499 nr_entries += nr_sgs;
5502 table = sd_alloc_ctl_entry(nr_entries);
5507 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5508 sizeof(long), 0644, proc_doulongvec_minmax, false);
5509 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5510 sizeof(long), 0644, proc_doulongvec_minmax, false);
5511 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5512 sizeof(int), 0644, proc_dointvec_minmax, true);
5513 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5514 sizeof(int), 0644, proc_dointvec_minmax, true);
5515 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5516 sizeof(int), 0644, proc_dointvec_minmax, true);
5517 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5518 sizeof(int), 0644, proc_dointvec_minmax, true);
5519 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5520 sizeof(int), 0644, proc_dointvec_minmax, true);
5521 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5522 sizeof(int), 0644, proc_dointvec_minmax, false);
5523 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5524 sizeof(int), 0644, proc_dointvec_minmax, false);
5525 set_table_entry(&table[9], "cache_nice_tries",
5526 &sd->cache_nice_tries,
5527 sizeof(int), 0644, proc_dointvec_minmax, false);
5528 set_table_entry(&table[10], "flags", &sd->flags,
5529 sizeof(int), 0644, proc_dointvec_minmax, false);
5530 set_table_entry(&table[11], "max_newidle_lb_cost",
5531 &sd->max_newidle_lb_cost,
5532 sizeof(long), 0644, proc_doulongvec_minmax, false);
5533 set_table_entry(&table[12], "name", sd->name,
5534 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5538 struct ctl_table *entry = &table[13];
5541 snprintf(buf, 32, "group%d", i);
5542 entry->procname = kstrdup(buf, GFP_KERNEL);
5544 entry->child = sd_alloc_ctl_group_table(sg);
5545 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5547 /* &table[nr_entries-1] is terminator */
5552 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5554 struct ctl_table *entry, *table;
5555 struct sched_domain *sd;
5556 int domain_num = 0, i;
5559 for_each_domain(cpu, sd)
5561 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5566 for_each_domain(cpu, sd) {
5567 snprintf(buf, 32, "domain%d", i);
5568 entry->procname = kstrdup(buf, GFP_KERNEL);
5570 entry->child = sd_alloc_ctl_domain_table(sd);
5577 static struct ctl_table_header *sd_sysctl_header;
5578 static void register_sched_domain_sysctl(void)
5580 int i, cpu_num = num_possible_cpus();
5581 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5584 WARN_ON(sd_ctl_dir[0].child);
5585 sd_ctl_dir[0].child = entry;
5590 for_each_possible_cpu(i) {
5591 snprintf(buf, 32, "cpu%d", i);
5592 entry->procname = kstrdup(buf, GFP_KERNEL);
5594 entry->child = sd_alloc_ctl_cpu_table(i);
5598 WARN_ON(sd_sysctl_header);
5599 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5602 /* may be called multiple times per register */
5603 static void unregister_sched_domain_sysctl(void)
5605 unregister_sysctl_table(sd_sysctl_header);
5606 sd_sysctl_header = NULL;
5607 if (sd_ctl_dir[0].child)
5608 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5611 static void register_sched_domain_sysctl(void)
5614 static void unregister_sched_domain_sysctl(void)
5617 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5619 static void set_rq_online(struct rq *rq)
5622 const struct sched_class *class;
5624 cpumask_set_cpu(rq->cpu, rq->rd->online);
5627 for_each_class(class) {
5628 if (class->rq_online)
5629 class->rq_online(rq);
5634 static void set_rq_offline(struct rq *rq)
5637 const struct sched_class *class;
5639 for_each_class(class) {
5640 if (class->rq_offline)
5641 class->rq_offline(rq);
5644 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5650 * migration_call - callback that gets triggered when a CPU is added.
5651 * Here we can start up the necessary migration thread for the new CPU.
5654 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5656 int cpu = (long)hcpu;
5657 unsigned long flags;
5658 struct rq *rq = cpu_rq(cpu);
5660 switch (action & ~CPU_TASKS_FROZEN) {
5662 case CPU_UP_PREPARE:
5663 rq->calc_load_update = calc_load_update;
5664 account_reset_rq(rq);
5668 /* Update our root-domain */
5669 raw_spin_lock_irqsave(&rq->lock, flags);
5671 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5675 raw_spin_unlock_irqrestore(&rq->lock, flags);
5678 #ifdef CONFIG_HOTPLUG_CPU
5680 sched_ttwu_pending();
5681 /* Update our root-domain */
5682 raw_spin_lock_irqsave(&rq->lock, flags);
5684 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5688 BUG_ON(rq->nr_running != 1); /* the migration thread */
5689 raw_spin_unlock_irqrestore(&rq->lock, flags);
5693 calc_load_migrate(rq);
5698 update_max_interval();
5704 * Register at high priority so that task migration (migrate_all_tasks)
5705 * happens before everything else. This has to be lower priority than
5706 * the notifier in the perf_event subsystem, though.
5708 static struct notifier_block migration_notifier = {
5709 .notifier_call = migration_call,
5710 .priority = CPU_PRI_MIGRATION,
5713 static void set_cpu_rq_start_time(void)
5715 int cpu = smp_processor_id();
5716 struct rq *rq = cpu_rq(cpu);
5717 rq->age_stamp = sched_clock_cpu(cpu);
5720 static int sched_cpu_active(struct notifier_block *nfb,
5721 unsigned long action, void *hcpu)
5723 int cpu = (long)hcpu;
5725 switch (action & ~CPU_TASKS_FROZEN) {
5727 set_cpu_rq_start_time();
5732 * At this point a starting CPU has marked itself as online via
5733 * set_cpu_online(). But it might not yet have marked itself
5734 * as active, which is essential from here on.
5736 set_cpu_active(cpu, true);
5737 stop_machine_unpark(cpu);
5740 case CPU_DOWN_FAILED:
5741 set_cpu_active(cpu, true);
5749 static int sched_cpu_inactive(struct notifier_block *nfb,
5750 unsigned long action, void *hcpu)
5752 switch (action & ~CPU_TASKS_FROZEN) {
5753 case CPU_DOWN_PREPARE:
5754 set_cpu_active((long)hcpu, false);
5761 static int __init migration_init(void)
5763 void *cpu = (void *)(long)smp_processor_id();
5766 /* Initialize migration for the boot CPU */
5767 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5768 BUG_ON(err == NOTIFY_BAD);
5769 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5770 register_cpu_notifier(&migration_notifier);
5772 /* Register cpu active notifiers */
5773 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5774 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5778 early_initcall(migration_init);
5780 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5782 #ifdef CONFIG_SCHED_DEBUG
5784 static __read_mostly int sched_debug_enabled;
5786 static int __init sched_debug_setup(char *str)
5788 sched_debug_enabled = 1;
5792 early_param("sched_debug", sched_debug_setup);
5794 static inline bool sched_debug(void)
5796 return sched_debug_enabled;
5799 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5800 struct cpumask *groupmask)
5802 struct sched_group *group = sd->groups;
5804 cpumask_clear(groupmask);
5806 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5808 if (!(sd->flags & SD_LOAD_BALANCE)) {
5809 printk("does not load-balance\n");
5811 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5816 printk(KERN_CONT "span %*pbl level %s\n",
5817 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5819 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5820 printk(KERN_ERR "ERROR: domain->span does not contain "
5823 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5824 printk(KERN_ERR "ERROR: domain->groups does not contain"
5828 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5832 printk(KERN_ERR "ERROR: group is NULL\n");
5836 if (!cpumask_weight(sched_group_cpus(group))) {
5837 printk(KERN_CONT "\n");
5838 printk(KERN_ERR "ERROR: empty group\n");
5842 if (!(sd->flags & SD_OVERLAP) &&
5843 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5844 printk(KERN_CONT "\n");
5845 printk(KERN_ERR "ERROR: repeated CPUs\n");
5849 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5851 printk(KERN_CONT " %*pbl",
5852 cpumask_pr_args(sched_group_cpus(group)));
5853 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5854 printk(KERN_CONT " (cpu_capacity = %lu)",
5855 group->sgc->capacity);
5858 group = group->next;
5859 } while (group != sd->groups);
5860 printk(KERN_CONT "\n");
5862 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5863 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5866 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5867 printk(KERN_ERR "ERROR: parent span is not a superset "
5868 "of domain->span\n");
5872 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5876 if (!sched_debug_enabled)
5880 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5884 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5887 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5895 #else /* !CONFIG_SCHED_DEBUG */
5896 # define sched_domain_debug(sd, cpu) do { } while (0)
5897 static inline bool sched_debug(void)
5901 #endif /* CONFIG_SCHED_DEBUG */
5903 static int sd_degenerate(struct sched_domain *sd)
5905 if (cpumask_weight(sched_domain_span(sd)) == 1)
5908 /* Following flags need at least 2 groups */
5909 if (sd->flags & (SD_LOAD_BALANCE |
5910 SD_BALANCE_NEWIDLE |
5913 SD_SHARE_CPUCAPACITY |
5914 SD_SHARE_PKG_RESOURCES |
5915 SD_SHARE_POWERDOMAIN |
5916 SD_SHARE_CAP_STATES)) {
5917 if (sd->groups != sd->groups->next)
5921 /* Following flags don't use groups */
5922 if (sd->flags & (SD_WAKE_AFFINE))
5929 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5931 unsigned long cflags = sd->flags, pflags = parent->flags;
5933 if (sd_degenerate(parent))
5936 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5939 /* Flags needing groups don't count if only 1 group in parent */
5940 if (parent->groups == parent->groups->next) {
5941 pflags &= ~(SD_LOAD_BALANCE |
5942 SD_BALANCE_NEWIDLE |
5945 SD_SHARE_CPUCAPACITY |
5946 SD_SHARE_PKG_RESOURCES |
5948 SD_SHARE_POWERDOMAIN |
5949 SD_SHARE_CAP_STATES);
5950 if (nr_node_ids == 1)
5951 pflags &= ~SD_SERIALIZE;
5953 if (~cflags & pflags)
5959 static void free_rootdomain(struct rcu_head *rcu)
5961 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5963 cpupri_cleanup(&rd->cpupri);
5964 cpudl_cleanup(&rd->cpudl);
5965 free_cpumask_var(rd->dlo_mask);
5966 free_cpumask_var(rd->rto_mask);
5967 free_cpumask_var(rd->online);
5968 free_cpumask_var(rd->span);
5972 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5974 struct root_domain *old_rd = NULL;
5975 unsigned long flags;
5977 raw_spin_lock_irqsave(&rq->lock, flags);
5982 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5985 cpumask_clear_cpu(rq->cpu, old_rd->span);
5988 * If we dont want to free the old_rd yet then
5989 * set old_rd to NULL to skip the freeing later
5992 if (!atomic_dec_and_test(&old_rd->refcount))
5996 atomic_inc(&rd->refcount);
5999 cpumask_set_cpu(rq->cpu, rd->span);
6000 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6003 raw_spin_unlock_irqrestore(&rq->lock, flags);
6006 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6009 static int init_rootdomain(struct root_domain *rd)
6011 memset(rd, 0, sizeof(*rd));
6013 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6015 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6017 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6019 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6022 init_dl_bw(&rd->dl_bw);
6023 if (cpudl_init(&rd->cpudl) != 0)
6026 if (cpupri_init(&rd->cpupri) != 0)
6029 init_max_cpu_capacity(&rd->max_cpu_capacity);
6033 free_cpumask_var(rd->rto_mask);
6035 free_cpumask_var(rd->dlo_mask);
6037 free_cpumask_var(rd->online);
6039 free_cpumask_var(rd->span);
6045 * By default the system creates a single root-domain with all cpus as
6046 * members (mimicking the global state we have today).
6048 struct root_domain def_root_domain;
6050 static void init_defrootdomain(void)
6052 init_rootdomain(&def_root_domain);
6054 atomic_set(&def_root_domain.refcount, 1);
6057 static struct root_domain *alloc_rootdomain(void)
6059 struct root_domain *rd;
6061 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6065 if (init_rootdomain(rd) != 0) {
6073 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6075 struct sched_group *tmp, *first;
6084 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6089 } while (sg != first);
6092 static void free_sched_domain(struct rcu_head *rcu)
6094 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6097 * If its an overlapping domain it has private groups, iterate and
6100 if (sd->flags & SD_OVERLAP) {
6101 free_sched_groups(sd->groups, 1);
6102 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6103 kfree(sd->groups->sgc);
6109 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6111 call_rcu(&sd->rcu, free_sched_domain);
6114 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6116 for (; sd; sd = sd->parent)
6117 destroy_sched_domain(sd, cpu);
6121 * Keep a special pointer to the highest sched_domain that has
6122 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6123 * allows us to avoid some pointer chasing select_idle_sibling().
6125 * Also keep a unique ID per domain (we use the first cpu number in
6126 * the cpumask of the domain), this allows us to quickly tell if
6127 * two cpus are in the same cache domain, see cpus_share_cache().
6129 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6130 DEFINE_PER_CPU(int, sd_llc_size);
6131 DEFINE_PER_CPU(int, sd_llc_id);
6132 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6133 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6134 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6135 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6136 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6138 static void update_top_cache_domain(int cpu)
6140 struct sched_domain *sd;
6141 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6145 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6147 id = cpumask_first(sched_domain_span(sd));
6148 size = cpumask_weight(sched_domain_span(sd));
6149 busy_sd = sd->parent; /* sd_busy */
6151 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6153 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6154 per_cpu(sd_llc_size, cpu) = size;
6155 per_cpu(sd_llc_id, cpu) = id;
6157 sd = lowest_flag_domain(cpu, SD_NUMA);
6158 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6160 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6161 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6163 for_each_domain(cpu, sd) {
6164 if (sd->groups->sge)
6169 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6171 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6172 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6176 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6177 * hold the hotplug lock.
6180 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6182 struct rq *rq = cpu_rq(cpu);
6183 struct sched_domain *tmp;
6185 /* Remove the sched domains which do not contribute to scheduling. */
6186 for (tmp = sd; tmp; ) {
6187 struct sched_domain *parent = tmp->parent;
6191 if (sd_parent_degenerate(tmp, parent)) {
6192 tmp->parent = parent->parent;
6194 parent->parent->child = tmp;
6196 * Transfer SD_PREFER_SIBLING down in case of a
6197 * degenerate parent; the spans match for this
6198 * so the property transfers.
6200 if (parent->flags & SD_PREFER_SIBLING)
6201 tmp->flags |= SD_PREFER_SIBLING;
6202 destroy_sched_domain(parent, cpu);
6207 if (sd && sd_degenerate(sd)) {
6210 destroy_sched_domain(tmp, cpu);
6215 sched_domain_debug(sd, cpu);
6217 rq_attach_root(rq, rd);
6219 rcu_assign_pointer(rq->sd, sd);
6220 destroy_sched_domains(tmp, cpu);
6222 update_top_cache_domain(cpu);
6225 /* Setup the mask of cpus configured for isolated domains */
6226 static int __init isolated_cpu_setup(char *str)
6228 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6229 cpulist_parse(str, cpu_isolated_map);
6233 __setup("isolcpus=", isolated_cpu_setup);
6236 struct sched_domain ** __percpu sd;
6237 struct root_domain *rd;
6248 * Build an iteration mask that can exclude certain CPUs from the upwards
6251 * Asymmetric node setups can result in situations where the domain tree is of
6252 * unequal depth, make sure to skip domains that already cover the entire
6255 * In that case build_sched_domains() will have terminated the iteration early
6256 * and our sibling sd spans will be empty. Domains should always include the
6257 * cpu they're built on, so check that.
6260 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6262 const struct cpumask *span = sched_domain_span(sd);
6263 struct sd_data *sdd = sd->private;
6264 struct sched_domain *sibling;
6267 for_each_cpu(i, span) {
6268 sibling = *per_cpu_ptr(sdd->sd, i);
6269 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6272 cpumask_set_cpu(i, sched_group_mask(sg));
6277 * Return the canonical balance cpu for this group, this is the first cpu
6278 * of this group that's also in the iteration mask.
6280 int group_balance_cpu(struct sched_group *sg)
6282 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6286 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6288 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6289 const struct cpumask *span = sched_domain_span(sd);
6290 struct cpumask *covered = sched_domains_tmpmask;
6291 struct sd_data *sdd = sd->private;
6292 struct sched_domain *sibling;
6295 cpumask_clear(covered);
6297 for_each_cpu(i, span) {
6298 struct cpumask *sg_span;
6300 if (cpumask_test_cpu(i, covered))
6303 sibling = *per_cpu_ptr(sdd->sd, i);
6305 /* See the comment near build_group_mask(). */
6306 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6309 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6310 GFP_KERNEL, cpu_to_node(cpu));
6315 sg_span = sched_group_cpus(sg);
6317 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6319 cpumask_set_cpu(i, sg_span);
6321 cpumask_or(covered, covered, sg_span);
6323 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6324 if (atomic_inc_return(&sg->sgc->ref) == 1)
6325 build_group_mask(sd, sg);
6328 * Initialize sgc->capacity such that even if we mess up the
6329 * domains and no possible iteration will get us here, we won't
6332 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6333 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6336 * Make sure the first group of this domain contains the
6337 * canonical balance cpu. Otherwise the sched_domain iteration
6338 * breaks. See update_sg_lb_stats().
6340 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6341 group_balance_cpu(sg) == cpu)
6351 sd->groups = groups;
6356 free_sched_groups(first, 0);
6361 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6363 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6364 struct sched_domain *child = sd->child;
6367 cpu = cpumask_first(sched_domain_span(child));
6370 *sg = *per_cpu_ptr(sdd->sg, cpu);
6371 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6372 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6379 * build_sched_groups will build a circular linked list of the groups
6380 * covered by the given span, and will set each group's ->cpumask correctly,
6381 * and ->cpu_capacity to 0.
6383 * Assumes the sched_domain tree is fully constructed
6386 build_sched_groups(struct sched_domain *sd, int cpu)
6388 struct sched_group *first = NULL, *last = NULL;
6389 struct sd_data *sdd = sd->private;
6390 const struct cpumask *span = sched_domain_span(sd);
6391 struct cpumask *covered;
6394 get_group(cpu, sdd, &sd->groups);
6395 atomic_inc(&sd->groups->ref);
6397 if (cpu != cpumask_first(span))
6400 lockdep_assert_held(&sched_domains_mutex);
6401 covered = sched_domains_tmpmask;
6403 cpumask_clear(covered);
6405 for_each_cpu(i, span) {
6406 struct sched_group *sg;
6409 if (cpumask_test_cpu(i, covered))
6412 group = get_group(i, sdd, &sg);
6413 cpumask_setall(sched_group_mask(sg));
6415 for_each_cpu(j, span) {
6416 if (get_group(j, sdd, NULL) != group)
6419 cpumask_set_cpu(j, covered);
6420 cpumask_set_cpu(j, sched_group_cpus(sg));
6435 * Initialize sched groups cpu_capacity.
6437 * cpu_capacity indicates the capacity of sched group, which is used while
6438 * distributing the load between different sched groups in a sched domain.
6439 * Typically cpu_capacity for all the groups in a sched domain will be same
6440 * unless there are asymmetries in the topology. If there are asymmetries,
6441 * group having more cpu_capacity will pickup more load compared to the
6442 * group having less cpu_capacity.
6444 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6446 struct sched_group *sg = sd->groups;
6451 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6453 } while (sg != sd->groups);
6455 if (cpu != group_balance_cpu(sg))
6458 update_group_capacity(sd, cpu);
6459 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6463 * Check that the per-cpu provided sd energy data is consistent for all cpus
6466 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6467 const struct cpumask *cpumask)
6469 const struct sched_group_energy * const sge = fn(cpu);
6470 struct cpumask mask;
6473 if (cpumask_weight(cpumask) <= 1)
6476 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6478 for_each_cpu(i, &mask) {
6479 const struct sched_group_energy * const e = fn(i);
6482 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6484 for (y = 0; y < (e->nr_idle_states); y++) {
6485 BUG_ON(e->idle_states[y].power !=
6486 sge->idle_states[y].power);
6489 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6491 for (y = 0; y < (e->nr_cap_states); y++) {
6492 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6493 BUG_ON(e->cap_states[y].power !=
6494 sge->cap_states[y].power);
6499 static void init_sched_energy(int cpu, struct sched_domain *sd,
6500 sched_domain_energy_f fn)
6502 if (!(fn && fn(cpu)))
6505 if (cpu != group_balance_cpu(sd->groups))
6508 if (sd->child && !sd->child->groups->sge) {
6509 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6510 #ifdef CONFIG_SCHED_DEBUG
6511 pr_err(" energy data on %s but not on %s domain\n",
6512 sd->name, sd->child->name);
6517 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6519 sd->groups->sge = fn(cpu);
6523 * Initializers for schedule domains
6524 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6527 static int default_relax_domain_level = -1;
6528 int sched_domain_level_max;
6530 static int __init setup_relax_domain_level(char *str)
6532 if (kstrtoint(str, 0, &default_relax_domain_level))
6533 pr_warn("Unable to set relax_domain_level\n");
6537 __setup("relax_domain_level=", setup_relax_domain_level);
6539 static void set_domain_attribute(struct sched_domain *sd,
6540 struct sched_domain_attr *attr)
6544 if (!attr || attr->relax_domain_level < 0) {
6545 if (default_relax_domain_level < 0)
6548 request = default_relax_domain_level;
6550 request = attr->relax_domain_level;
6551 if (request < sd->level) {
6552 /* turn off idle balance on this domain */
6553 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6555 /* turn on idle balance on this domain */
6556 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6560 static void __sdt_free(const struct cpumask *cpu_map);
6561 static int __sdt_alloc(const struct cpumask *cpu_map);
6563 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6564 const struct cpumask *cpu_map)
6568 if (!atomic_read(&d->rd->refcount))
6569 free_rootdomain(&d->rd->rcu); /* fall through */
6571 free_percpu(d->sd); /* fall through */
6573 __sdt_free(cpu_map); /* fall through */
6579 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6580 const struct cpumask *cpu_map)
6582 memset(d, 0, sizeof(*d));
6584 if (__sdt_alloc(cpu_map))
6585 return sa_sd_storage;
6586 d->sd = alloc_percpu(struct sched_domain *);
6588 return sa_sd_storage;
6589 d->rd = alloc_rootdomain();
6592 return sa_rootdomain;
6596 * NULL the sd_data elements we've used to build the sched_domain and
6597 * sched_group structure so that the subsequent __free_domain_allocs()
6598 * will not free the data we're using.
6600 static void claim_allocations(int cpu, struct sched_domain *sd)
6602 struct sd_data *sdd = sd->private;
6604 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6605 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6607 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6608 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6610 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6611 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6615 static int sched_domains_numa_levels;
6616 enum numa_topology_type sched_numa_topology_type;
6617 static int *sched_domains_numa_distance;
6618 int sched_max_numa_distance;
6619 static struct cpumask ***sched_domains_numa_masks;
6620 static int sched_domains_curr_level;
6624 * SD_flags allowed in topology descriptions.
6626 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6627 * SD_SHARE_PKG_RESOURCES - describes shared caches
6628 * SD_NUMA - describes NUMA topologies
6629 * SD_SHARE_POWERDOMAIN - describes shared power domain
6630 * SD_SHARE_CAP_STATES - describes shared capacity states
6633 * SD_ASYM_PACKING - describes SMT quirks
6635 #define TOPOLOGY_SD_FLAGS \
6636 (SD_SHARE_CPUCAPACITY | \
6637 SD_SHARE_PKG_RESOURCES | \
6640 SD_SHARE_POWERDOMAIN | \
6641 SD_SHARE_CAP_STATES)
6643 static struct sched_domain *
6644 sd_init(struct sched_domain_topology_level *tl, int cpu)
6646 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6647 int sd_weight, sd_flags = 0;
6651 * Ugly hack to pass state to sd_numa_mask()...
6653 sched_domains_curr_level = tl->numa_level;
6656 sd_weight = cpumask_weight(tl->mask(cpu));
6659 sd_flags = (*tl->sd_flags)();
6660 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6661 "wrong sd_flags in topology description\n"))
6662 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6664 *sd = (struct sched_domain){
6665 .min_interval = sd_weight,
6666 .max_interval = 2*sd_weight,
6668 .imbalance_pct = 125,
6670 .cache_nice_tries = 0,
6677 .flags = 1*SD_LOAD_BALANCE
6678 | 1*SD_BALANCE_NEWIDLE
6683 | 0*SD_SHARE_CPUCAPACITY
6684 | 0*SD_SHARE_PKG_RESOURCES
6686 | 0*SD_PREFER_SIBLING
6691 .last_balance = jiffies,
6692 .balance_interval = sd_weight,
6694 .max_newidle_lb_cost = 0,
6695 .next_decay_max_lb_cost = jiffies,
6696 #ifdef CONFIG_SCHED_DEBUG
6702 * Convert topological properties into behaviour.
6705 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6706 sd->flags |= SD_PREFER_SIBLING;
6707 sd->imbalance_pct = 110;
6708 sd->smt_gain = 1178; /* ~15% */
6710 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6711 sd->imbalance_pct = 117;
6712 sd->cache_nice_tries = 1;
6716 } else if (sd->flags & SD_NUMA) {
6717 sd->cache_nice_tries = 2;
6721 sd->flags |= SD_SERIALIZE;
6722 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6723 sd->flags &= ~(SD_BALANCE_EXEC |
6730 sd->flags |= SD_PREFER_SIBLING;
6731 sd->cache_nice_tries = 1;
6736 sd->private = &tl->data;
6742 * Topology list, bottom-up.
6744 static struct sched_domain_topology_level default_topology[] = {
6745 #ifdef CONFIG_SCHED_SMT
6746 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6748 #ifdef CONFIG_SCHED_MC
6749 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6751 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6755 static struct sched_domain_topology_level *sched_domain_topology =
6758 #define for_each_sd_topology(tl) \
6759 for (tl = sched_domain_topology; tl->mask; tl++)
6761 void set_sched_topology(struct sched_domain_topology_level *tl)
6763 sched_domain_topology = tl;
6768 static const struct cpumask *sd_numa_mask(int cpu)
6770 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6773 static void sched_numa_warn(const char *str)
6775 static int done = false;
6783 printk(KERN_WARNING "ERROR: %s\n\n", str);
6785 for (i = 0; i < nr_node_ids; i++) {
6786 printk(KERN_WARNING " ");
6787 for (j = 0; j < nr_node_ids; j++)
6788 printk(KERN_CONT "%02d ", node_distance(i,j));
6789 printk(KERN_CONT "\n");
6791 printk(KERN_WARNING "\n");
6794 bool find_numa_distance(int distance)
6798 if (distance == node_distance(0, 0))
6801 for (i = 0; i < sched_domains_numa_levels; i++) {
6802 if (sched_domains_numa_distance[i] == distance)
6810 * A system can have three types of NUMA topology:
6811 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6812 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6813 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6815 * The difference between a glueless mesh topology and a backplane
6816 * topology lies in whether communication between not directly
6817 * connected nodes goes through intermediary nodes (where programs
6818 * could run), or through backplane controllers. This affects
6819 * placement of programs.
6821 * The type of topology can be discerned with the following tests:
6822 * - If the maximum distance between any nodes is 1 hop, the system
6823 * is directly connected.
6824 * - If for two nodes A and B, located N > 1 hops away from each other,
6825 * there is an intermediary node C, which is < N hops away from both
6826 * nodes A and B, the system is a glueless mesh.
6828 static void init_numa_topology_type(void)
6832 n = sched_max_numa_distance;
6834 if (sched_domains_numa_levels <= 1) {
6835 sched_numa_topology_type = NUMA_DIRECT;
6839 for_each_online_node(a) {
6840 for_each_online_node(b) {
6841 /* Find two nodes furthest removed from each other. */
6842 if (node_distance(a, b) < n)
6845 /* Is there an intermediary node between a and b? */
6846 for_each_online_node(c) {
6847 if (node_distance(a, c) < n &&
6848 node_distance(b, c) < n) {
6849 sched_numa_topology_type =
6855 sched_numa_topology_type = NUMA_BACKPLANE;
6861 static void sched_init_numa(void)
6863 int next_distance, curr_distance = node_distance(0, 0);
6864 struct sched_domain_topology_level *tl;
6868 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6869 if (!sched_domains_numa_distance)
6873 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6874 * unique distances in the node_distance() table.
6876 * Assumes node_distance(0,j) includes all distances in
6877 * node_distance(i,j) in order to avoid cubic time.
6879 next_distance = curr_distance;
6880 for (i = 0; i < nr_node_ids; i++) {
6881 for (j = 0; j < nr_node_ids; j++) {
6882 for (k = 0; k < nr_node_ids; k++) {
6883 int distance = node_distance(i, k);
6885 if (distance > curr_distance &&
6886 (distance < next_distance ||
6887 next_distance == curr_distance))
6888 next_distance = distance;
6891 * While not a strong assumption it would be nice to know
6892 * about cases where if node A is connected to B, B is not
6893 * equally connected to A.
6895 if (sched_debug() && node_distance(k, i) != distance)
6896 sched_numa_warn("Node-distance not symmetric");
6898 if (sched_debug() && i && !find_numa_distance(distance))
6899 sched_numa_warn("Node-0 not representative");
6901 if (next_distance != curr_distance) {
6902 sched_domains_numa_distance[level++] = next_distance;
6903 sched_domains_numa_levels = level;
6904 curr_distance = next_distance;
6909 * In case of sched_debug() we verify the above assumption.
6919 * 'level' contains the number of unique distances, excluding the
6920 * identity distance node_distance(i,i).
6922 * The sched_domains_numa_distance[] array includes the actual distance
6927 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6928 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6929 * the array will contain less then 'level' members. This could be
6930 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6931 * in other functions.
6933 * We reset it to 'level' at the end of this function.
6935 sched_domains_numa_levels = 0;
6937 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6938 if (!sched_domains_numa_masks)
6942 * Now for each level, construct a mask per node which contains all
6943 * cpus of nodes that are that many hops away from us.
6945 for (i = 0; i < level; i++) {
6946 sched_domains_numa_masks[i] =
6947 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6948 if (!sched_domains_numa_masks[i])
6951 for (j = 0; j < nr_node_ids; j++) {
6952 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6956 sched_domains_numa_masks[i][j] = mask;
6959 if (node_distance(j, k) > sched_domains_numa_distance[i])
6962 cpumask_or(mask, mask, cpumask_of_node(k));
6967 /* Compute default topology size */
6968 for (i = 0; sched_domain_topology[i].mask; i++);
6970 tl = kzalloc((i + level + 1) *
6971 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6976 * Copy the default topology bits..
6978 for (i = 0; sched_domain_topology[i].mask; i++)
6979 tl[i] = sched_domain_topology[i];
6982 * .. and append 'j' levels of NUMA goodness.
6984 for (j = 0; j < level; i++, j++) {
6985 tl[i] = (struct sched_domain_topology_level){
6986 .mask = sd_numa_mask,
6987 .sd_flags = cpu_numa_flags,
6988 .flags = SDTL_OVERLAP,
6994 sched_domain_topology = tl;
6996 sched_domains_numa_levels = level;
6997 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6999 init_numa_topology_type();
7002 static void sched_domains_numa_masks_set(int cpu)
7005 int node = cpu_to_node(cpu);
7007 for (i = 0; i < sched_domains_numa_levels; i++) {
7008 for (j = 0; j < nr_node_ids; j++) {
7009 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7010 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7015 static void sched_domains_numa_masks_clear(int cpu)
7018 for (i = 0; i < sched_domains_numa_levels; i++) {
7019 for (j = 0; j < nr_node_ids; j++)
7020 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7025 * Update sched_domains_numa_masks[level][node] array when new cpus
7028 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7029 unsigned long action,
7032 int cpu = (long)hcpu;
7034 switch (action & ~CPU_TASKS_FROZEN) {
7036 sched_domains_numa_masks_set(cpu);
7040 sched_domains_numa_masks_clear(cpu);
7050 static inline void sched_init_numa(void)
7054 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7055 unsigned long action,
7060 #endif /* CONFIG_NUMA */
7062 static int __sdt_alloc(const struct cpumask *cpu_map)
7064 struct sched_domain_topology_level *tl;
7067 for_each_sd_topology(tl) {
7068 struct sd_data *sdd = &tl->data;
7070 sdd->sd = alloc_percpu(struct sched_domain *);
7074 sdd->sg = alloc_percpu(struct sched_group *);
7078 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7082 for_each_cpu(j, cpu_map) {
7083 struct sched_domain *sd;
7084 struct sched_group *sg;
7085 struct sched_group_capacity *sgc;
7087 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7088 GFP_KERNEL, cpu_to_node(j));
7092 *per_cpu_ptr(sdd->sd, j) = sd;
7094 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7095 GFP_KERNEL, cpu_to_node(j));
7101 *per_cpu_ptr(sdd->sg, j) = sg;
7103 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7104 GFP_KERNEL, cpu_to_node(j));
7108 *per_cpu_ptr(sdd->sgc, j) = sgc;
7115 static void __sdt_free(const struct cpumask *cpu_map)
7117 struct sched_domain_topology_level *tl;
7120 for_each_sd_topology(tl) {
7121 struct sd_data *sdd = &tl->data;
7123 for_each_cpu(j, cpu_map) {
7124 struct sched_domain *sd;
7127 sd = *per_cpu_ptr(sdd->sd, j);
7128 if (sd && (sd->flags & SD_OVERLAP))
7129 free_sched_groups(sd->groups, 0);
7130 kfree(*per_cpu_ptr(sdd->sd, j));
7134 kfree(*per_cpu_ptr(sdd->sg, j));
7136 kfree(*per_cpu_ptr(sdd->sgc, j));
7138 free_percpu(sdd->sd);
7140 free_percpu(sdd->sg);
7142 free_percpu(sdd->sgc);
7147 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7148 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7149 struct sched_domain *child, int cpu)
7151 struct sched_domain *sd = sd_init(tl, cpu);
7155 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7157 sd->level = child->level + 1;
7158 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7162 if (!cpumask_subset(sched_domain_span(child),
7163 sched_domain_span(sd))) {
7164 pr_err("BUG: arch topology borken\n");
7165 #ifdef CONFIG_SCHED_DEBUG
7166 pr_err(" the %s domain not a subset of the %s domain\n",
7167 child->name, sd->name);
7169 /* Fixup, ensure @sd has at least @child cpus. */
7170 cpumask_or(sched_domain_span(sd),
7171 sched_domain_span(sd),
7172 sched_domain_span(child));
7176 set_domain_attribute(sd, attr);
7182 * Build sched domains for a given set of cpus and attach the sched domains
7183 * to the individual cpus
7185 static int build_sched_domains(const struct cpumask *cpu_map,
7186 struct sched_domain_attr *attr)
7188 enum s_alloc alloc_state;
7189 struct sched_domain *sd;
7191 struct rq *rq = NULL;
7192 int i, ret = -ENOMEM;
7194 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7195 if (alloc_state != sa_rootdomain)
7198 /* Set up domains for cpus specified by the cpu_map. */
7199 for_each_cpu(i, cpu_map) {
7200 struct sched_domain_topology_level *tl;
7203 for_each_sd_topology(tl) {
7204 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7205 if (tl == sched_domain_topology)
7206 *per_cpu_ptr(d.sd, i) = sd;
7207 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7208 sd->flags |= SD_OVERLAP;
7209 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7214 /* Build the groups for the domains */
7215 for_each_cpu(i, cpu_map) {
7216 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7217 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7218 if (sd->flags & SD_OVERLAP) {
7219 if (build_overlap_sched_groups(sd, i))
7222 if (build_sched_groups(sd, i))
7228 /* Calculate CPU capacity for physical packages and nodes */
7229 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7230 struct sched_domain_topology_level *tl = sched_domain_topology;
7232 if (!cpumask_test_cpu(i, cpu_map))
7235 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7236 init_sched_energy(i, sd, tl->energy);
7237 claim_allocations(i, sd);
7238 init_sched_groups_capacity(i, sd);
7242 /* Attach the domains */
7244 for_each_cpu(i, cpu_map) {
7246 sd = *per_cpu_ptr(d.sd, i);
7247 cpu_attach_domain(sd, d.rd, i);
7253 __free_domain_allocs(&d, alloc_state, cpu_map);
7257 static cpumask_var_t *doms_cur; /* current sched domains */
7258 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7259 static struct sched_domain_attr *dattr_cur;
7260 /* attribues of custom domains in 'doms_cur' */
7263 * Special case: If a kmalloc of a doms_cur partition (array of
7264 * cpumask) fails, then fallback to a single sched domain,
7265 * as determined by the single cpumask fallback_doms.
7267 static cpumask_var_t fallback_doms;
7270 * arch_update_cpu_topology lets virtualized architectures update the
7271 * cpu core maps. It is supposed to return 1 if the topology changed
7272 * or 0 if it stayed the same.
7274 int __weak arch_update_cpu_topology(void)
7279 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7282 cpumask_var_t *doms;
7284 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7287 for (i = 0; i < ndoms; i++) {
7288 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7289 free_sched_domains(doms, i);
7296 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7299 for (i = 0; i < ndoms; i++)
7300 free_cpumask_var(doms[i]);
7305 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7306 * For now this just excludes isolated cpus, but could be used to
7307 * exclude other special cases in the future.
7309 static int init_sched_domains(const struct cpumask *cpu_map)
7313 arch_update_cpu_topology();
7315 doms_cur = alloc_sched_domains(ndoms_cur);
7317 doms_cur = &fallback_doms;
7318 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7319 err = build_sched_domains(doms_cur[0], NULL);
7320 register_sched_domain_sysctl();
7326 * Detach sched domains from a group of cpus specified in cpu_map
7327 * These cpus will now be attached to the NULL domain
7329 static void detach_destroy_domains(const struct cpumask *cpu_map)
7334 for_each_cpu(i, cpu_map)
7335 cpu_attach_domain(NULL, &def_root_domain, i);
7339 /* handle null as "default" */
7340 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7341 struct sched_domain_attr *new, int idx_new)
7343 struct sched_domain_attr tmp;
7350 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7351 new ? (new + idx_new) : &tmp,
7352 sizeof(struct sched_domain_attr));
7356 * Partition sched domains as specified by the 'ndoms_new'
7357 * cpumasks in the array doms_new[] of cpumasks. This compares
7358 * doms_new[] to the current sched domain partitioning, doms_cur[].
7359 * It destroys each deleted domain and builds each new domain.
7361 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7362 * The masks don't intersect (don't overlap.) We should setup one
7363 * sched domain for each mask. CPUs not in any of the cpumasks will
7364 * not be load balanced. If the same cpumask appears both in the
7365 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7368 * The passed in 'doms_new' should be allocated using
7369 * alloc_sched_domains. This routine takes ownership of it and will
7370 * free_sched_domains it when done with it. If the caller failed the
7371 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7372 * and partition_sched_domains() will fallback to the single partition
7373 * 'fallback_doms', it also forces the domains to be rebuilt.
7375 * If doms_new == NULL it will be replaced with cpu_online_mask.
7376 * ndoms_new == 0 is a special case for destroying existing domains,
7377 * and it will not create the default domain.
7379 * Call with hotplug lock held
7381 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7382 struct sched_domain_attr *dattr_new)
7387 mutex_lock(&sched_domains_mutex);
7389 /* always unregister in case we don't destroy any domains */
7390 unregister_sched_domain_sysctl();
7392 /* Let architecture update cpu core mappings. */
7393 new_topology = arch_update_cpu_topology();
7395 n = doms_new ? ndoms_new : 0;
7397 /* Destroy deleted domains */
7398 for (i = 0; i < ndoms_cur; i++) {
7399 for (j = 0; j < n && !new_topology; j++) {
7400 if (cpumask_equal(doms_cur[i], doms_new[j])
7401 && dattrs_equal(dattr_cur, i, dattr_new, j))
7404 /* no match - a current sched domain not in new doms_new[] */
7405 detach_destroy_domains(doms_cur[i]);
7411 if (doms_new == NULL) {
7413 doms_new = &fallback_doms;
7414 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7415 WARN_ON_ONCE(dattr_new);
7418 /* Build new domains */
7419 for (i = 0; i < ndoms_new; i++) {
7420 for (j = 0; j < n && !new_topology; j++) {
7421 if (cpumask_equal(doms_new[i], doms_cur[j])
7422 && dattrs_equal(dattr_new, i, dattr_cur, j))
7425 /* no match - add a new doms_new */
7426 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7431 /* Remember the new sched domains */
7432 if (doms_cur != &fallback_doms)
7433 free_sched_domains(doms_cur, ndoms_cur);
7434 kfree(dattr_cur); /* kfree(NULL) is safe */
7435 doms_cur = doms_new;
7436 dattr_cur = dattr_new;
7437 ndoms_cur = ndoms_new;
7439 register_sched_domain_sysctl();
7441 mutex_unlock(&sched_domains_mutex);
7444 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7447 * Update cpusets according to cpu_active mask. If cpusets are
7448 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7449 * around partition_sched_domains().
7451 * If we come here as part of a suspend/resume, don't touch cpusets because we
7452 * want to restore it back to its original state upon resume anyway.
7454 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7458 case CPU_ONLINE_FROZEN:
7459 case CPU_DOWN_FAILED_FROZEN:
7462 * num_cpus_frozen tracks how many CPUs are involved in suspend
7463 * resume sequence. As long as this is not the last online
7464 * operation in the resume sequence, just build a single sched
7465 * domain, ignoring cpusets.
7468 if (likely(num_cpus_frozen)) {
7469 partition_sched_domains(1, NULL, NULL);
7474 * This is the last CPU online operation. So fall through and
7475 * restore the original sched domains by considering the
7476 * cpuset configurations.
7480 cpuset_update_active_cpus(true);
7488 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7491 unsigned long flags;
7492 long cpu = (long)hcpu;
7498 case CPU_DOWN_PREPARE:
7499 rcu_read_lock_sched();
7500 dl_b = dl_bw_of(cpu);
7502 raw_spin_lock_irqsave(&dl_b->lock, flags);
7503 cpus = dl_bw_cpus(cpu);
7504 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7505 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7507 rcu_read_unlock_sched();
7510 return notifier_from_errno(-EBUSY);
7511 cpuset_update_active_cpus(false);
7513 case CPU_DOWN_PREPARE_FROZEN:
7515 partition_sched_domains(1, NULL, NULL);
7523 void __init sched_init_smp(void)
7525 cpumask_var_t non_isolated_cpus;
7527 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7528 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7533 * There's no userspace yet to cause hotplug operations; hence all the
7534 * cpu masks are stable and all blatant races in the below code cannot
7537 mutex_lock(&sched_domains_mutex);
7538 init_sched_domains(cpu_active_mask);
7539 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7540 if (cpumask_empty(non_isolated_cpus))
7541 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7542 mutex_unlock(&sched_domains_mutex);
7544 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7545 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7546 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7550 /* Move init over to a non-isolated CPU */
7551 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7553 sched_init_granularity();
7554 free_cpumask_var(non_isolated_cpus);
7556 init_sched_rt_class();
7557 init_sched_dl_class();
7560 void __init sched_init_smp(void)
7562 sched_init_granularity();
7564 #endif /* CONFIG_SMP */
7566 int in_sched_functions(unsigned long addr)
7568 return in_lock_functions(addr) ||
7569 (addr >= (unsigned long)__sched_text_start
7570 && addr < (unsigned long)__sched_text_end);
7573 #ifdef CONFIG_CGROUP_SCHED
7575 * Default task group.
7576 * Every task in system belongs to this group at bootup.
7578 struct task_group root_task_group;
7579 LIST_HEAD(task_groups);
7582 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7584 void __init sched_init(void)
7587 unsigned long alloc_size = 0, ptr;
7589 #ifdef CONFIG_FAIR_GROUP_SCHED
7590 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7592 #ifdef CONFIG_RT_GROUP_SCHED
7593 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7596 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7598 #ifdef CONFIG_FAIR_GROUP_SCHED
7599 root_task_group.se = (struct sched_entity **)ptr;
7600 ptr += nr_cpu_ids * sizeof(void **);
7602 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7603 ptr += nr_cpu_ids * sizeof(void **);
7605 #endif /* CONFIG_FAIR_GROUP_SCHED */
7606 #ifdef CONFIG_RT_GROUP_SCHED
7607 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7608 ptr += nr_cpu_ids * sizeof(void **);
7610 root_task_group.rt_rq = (struct rt_rq **)ptr;
7611 ptr += nr_cpu_ids * sizeof(void **);
7613 #endif /* CONFIG_RT_GROUP_SCHED */
7615 #ifdef CONFIG_CPUMASK_OFFSTACK
7616 for_each_possible_cpu(i) {
7617 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7618 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7620 #endif /* CONFIG_CPUMASK_OFFSTACK */
7622 init_rt_bandwidth(&def_rt_bandwidth,
7623 global_rt_period(), global_rt_runtime());
7624 init_dl_bandwidth(&def_dl_bandwidth,
7625 global_rt_period(), global_rt_runtime());
7628 init_defrootdomain();
7631 #ifdef CONFIG_RT_GROUP_SCHED
7632 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7633 global_rt_period(), global_rt_runtime());
7634 #endif /* CONFIG_RT_GROUP_SCHED */
7636 #ifdef CONFIG_CGROUP_SCHED
7637 list_add(&root_task_group.list, &task_groups);
7638 INIT_LIST_HEAD(&root_task_group.children);
7639 INIT_LIST_HEAD(&root_task_group.siblings);
7640 autogroup_init(&init_task);
7642 #endif /* CONFIG_CGROUP_SCHED */
7644 for_each_possible_cpu(i) {
7648 raw_spin_lock_init(&rq->lock);
7650 rq->calc_load_active = 0;
7651 rq->calc_load_update = jiffies + LOAD_FREQ;
7652 init_cfs_rq(&rq->cfs);
7653 init_rt_rq(&rq->rt);
7654 init_dl_rq(&rq->dl);
7655 #ifdef CONFIG_FAIR_GROUP_SCHED
7656 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7657 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7659 * How much cpu bandwidth does root_task_group get?
7661 * In case of task-groups formed thr' the cgroup filesystem, it
7662 * gets 100% of the cpu resources in the system. This overall
7663 * system cpu resource is divided among the tasks of
7664 * root_task_group and its child task-groups in a fair manner,
7665 * based on each entity's (task or task-group's) weight
7666 * (se->load.weight).
7668 * In other words, if root_task_group has 10 tasks of weight
7669 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7670 * then A0's share of the cpu resource is:
7672 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7674 * We achieve this by letting root_task_group's tasks sit
7675 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7677 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7678 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7679 #endif /* CONFIG_FAIR_GROUP_SCHED */
7681 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7682 #ifdef CONFIG_RT_GROUP_SCHED
7683 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7686 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7687 rq->cpu_load[j] = 0;
7689 rq->last_load_update_tick = jiffies;
7694 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7695 rq->balance_callback = NULL;
7696 rq->active_balance = 0;
7697 rq->next_balance = jiffies;
7702 rq->avg_idle = 2*sysctl_sched_migration_cost;
7703 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7705 INIT_LIST_HEAD(&rq->cfs_tasks);
7707 rq_attach_root(rq, &def_root_domain);
7708 #ifdef CONFIG_NO_HZ_COMMON
7711 #ifdef CONFIG_NO_HZ_FULL
7712 rq->last_sched_tick = 0;
7716 atomic_set(&rq->nr_iowait, 0);
7719 set_load_weight(&init_task);
7721 #ifdef CONFIG_PREEMPT_NOTIFIERS
7722 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7726 * The boot idle thread does lazy MMU switching as well:
7728 atomic_inc(&init_mm.mm_count);
7729 enter_lazy_tlb(&init_mm, current);
7732 * During early bootup we pretend to be a normal task:
7734 current->sched_class = &fair_sched_class;
7737 * Make us the idle thread. Technically, schedule() should not be
7738 * called from this thread, however somewhere below it might be,
7739 * but because we are the idle thread, we just pick up running again
7740 * when this runqueue becomes "idle".
7742 init_idle(current, smp_processor_id());
7744 calc_load_update = jiffies + LOAD_FREQ;
7747 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7748 /* May be allocated at isolcpus cmdline parse time */
7749 if (cpu_isolated_map == NULL)
7750 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7751 idle_thread_set_boot_cpu();
7752 set_cpu_rq_start_time();
7754 init_sched_fair_class();
7756 scheduler_running = 1;
7759 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7760 static inline int preempt_count_equals(int preempt_offset)
7762 int nested = preempt_count() + rcu_preempt_depth();
7764 return (nested == preempt_offset);
7767 static int __might_sleep_init_called;
7768 int __init __might_sleep_init(void)
7770 __might_sleep_init_called = 1;
7773 early_initcall(__might_sleep_init);
7775 void __might_sleep(const char *file, int line, int preempt_offset)
7778 * Blocking primitives will set (and therefore destroy) current->state,
7779 * since we will exit with TASK_RUNNING make sure we enter with it,
7780 * otherwise we will destroy state.
7782 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7783 "do not call blocking ops when !TASK_RUNNING; "
7784 "state=%lx set at [<%p>] %pS\n",
7786 (void *)current->task_state_change,
7787 (void *)current->task_state_change);
7789 ___might_sleep(file, line, preempt_offset);
7791 EXPORT_SYMBOL(__might_sleep);
7793 void ___might_sleep(const char *file, int line, int preempt_offset)
7795 static unsigned long prev_jiffy; /* ratelimiting */
7797 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7798 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7799 !is_idle_task(current)) || oops_in_progress)
7801 if (system_state != SYSTEM_RUNNING &&
7802 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
7804 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7806 prev_jiffy = jiffies;
7809 "BUG: sleeping function called from invalid context at %s:%d\n",
7812 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7813 in_atomic(), irqs_disabled(),
7814 current->pid, current->comm);
7816 if (task_stack_end_corrupted(current))
7817 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7819 debug_show_held_locks(current);
7820 if (irqs_disabled())
7821 print_irqtrace_events(current);
7822 #ifdef CONFIG_DEBUG_PREEMPT
7823 if (!preempt_count_equals(preempt_offset)) {
7824 pr_err("Preemption disabled at:");
7825 print_ip_sym(current->preempt_disable_ip);
7831 EXPORT_SYMBOL(___might_sleep);
7834 #ifdef CONFIG_MAGIC_SYSRQ
7835 void normalize_rt_tasks(void)
7837 struct task_struct *g, *p;
7838 struct sched_attr attr = {
7839 .sched_policy = SCHED_NORMAL,
7842 read_lock(&tasklist_lock);
7843 for_each_process_thread(g, p) {
7845 * Only normalize user tasks:
7847 if (p->flags & PF_KTHREAD)
7850 p->se.exec_start = 0;
7851 #ifdef CONFIG_SCHEDSTATS
7852 p->se.statistics.wait_start = 0;
7853 p->se.statistics.sleep_start = 0;
7854 p->se.statistics.block_start = 0;
7857 if (!dl_task(p) && !rt_task(p)) {
7859 * Renice negative nice level userspace
7862 if (task_nice(p) < 0)
7863 set_user_nice(p, 0);
7867 __sched_setscheduler(p, &attr, false, false);
7869 read_unlock(&tasklist_lock);
7872 #endif /* CONFIG_MAGIC_SYSRQ */
7874 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7876 * These functions are only useful for the IA64 MCA handling, or kdb.
7878 * They can only be called when the whole system has been
7879 * stopped - every CPU needs to be quiescent, and no scheduling
7880 * activity can take place. Using them for anything else would
7881 * be a serious bug, and as a result, they aren't even visible
7882 * under any other configuration.
7886 * curr_task - return the current task for a given cpu.
7887 * @cpu: the processor in question.
7889 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7891 * Return: The current task for @cpu.
7893 struct task_struct *curr_task(int cpu)
7895 return cpu_curr(cpu);
7898 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7902 * set_curr_task - set the current task for a given cpu.
7903 * @cpu: the processor in question.
7904 * @p: the task pointer to set.
7906 * Description: This function must only be used when non-maskable interrupts
7907 * are serviced on a separate stack. It allows the architecture to switch the
7908 * notion of the current task on a cpu in a non-blocking manner. This function
7909 * must be called with all CPU's synchronized, and interrupts disabled, the
7910 * and caller must save the original value of the current task (see
7911 * curr_task() above) and restore that value before reenabling interrupts and
7912 * re-starting the system.
7914 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7916 void set_curr_task(int cpu, struct task_struct *p)
7923 #ifdef CONFIG_CGROUP_SCHED
7924 /* task_group_lock serializes the addition/removal of task groups */
7925 static DEFINE_SPINLOCK(task_group_lock);
7927 static void sched_free_group(struct task_group *tg)
7929 free_fair_sched_group(tg);
7930 free_rt_sched_group(tg);
7935 /* allocate runqueue etc for a new task group */
7936 struct task_group *sched_create_group(struct task_group *parent)
7938 struct task_group *tg;
7940 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7942 return ERR_PTR(-ENOMEM);
7944 if (!alloc_fair_sched_group(tg, parent))
7947 if (!alloc_rt_sched_group(tg, parent))
7953 sched_free_group(tg);
7954 return ERR_PTR(-ENOMEM);
7957 void sched_online_group(struct task_group *tg, struct task_group *parent)
7959 unsigned long flags;
7961 spin_lock_irqsave(&task_group_lock, flags);
7962 list_add_rcu(&tg->list, &task_groups);
7964 WARN_ON(!parent); /* root should already exist */
7966 tg->parent = parent;
7967 INIT_LIST_HEAD(&tg->children);
7968 list_add_rcu(&tg->siblings, &parent->children);
7969 spin_unlock_irqrestore(&task_group_lock, flags);
7972 /* rcu callback to free various structures associated with a task group */
7973 static void sched_free_group_rcu(struct rcu_head *rhp)
7975 /* now it should be safe to free those cfs_rqs */
7976 sched_free_group(container_of(rhp, struct task_group, rcu));
7979 void sched_destroy_group(struct task_group *tg)
7981 /* wait for possible concurrent references to cfs_rqs complete */
7982 call_rcu(&tg->rcu, sched_free_group_rcu);
7985 void sched_offline_group(struct task_group *tg)
7987 unsigned long flags;
7990 /* end participation in shares distribution */
7991 for_each_possible_cpu(i)
7992 unregister_fair_sched_group(tg, i);
7994 spin_lock_irqsave(&task_group_lock, flags);
7995 list_del_rcu(&tg->list);
7996 list_del_rcu(&tg->siblings);
7997 spin_unlock_irqrestore(&task_group_lock, flags);
8000 /* change task's runqueue when it moves between groups.
8001 * The caller of this function should have put the task in its new group
8002 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8003 * reflect its new group.
8005 void sched_move_task(struct task_struct *tsk)
8007 struct task_group *tg;
8008 int queued, running;
8009 unsigned long flags;
8012 rq = task_rq_lock(tsk, &flags);
8014 running = task_current(rq, tsk);
8015 queued = task_on_rq_queued(tsk);
8018 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8019 if (unlikely(running))
8020 put_prev_task(rq, tsk);
8023 * All callers are synchronized by task_rq_lock(); we do not use RCU
8024 * which is pointless here. Thus, we pass "true" to task_css_check()
8025 * to prevent lockdep warnings.
8027 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8028 struct task_group, css);
8029 tg = autogroup_task_group(tsk, tg);
8030 tsk->sched_task_group = tg;
8032 #ifdef CONFIG_FAIR_GROUP_SCHED
8033 if (tsk->sched_class->task_move_group)
8034 tsk->sched_class->task_move_group(tsk);
8037 set_task_rq(tsk, task_cpu(tsk));
8039 if (unlikely(running))
8040 tsk->sched_class->set_curr_task(rq);
8042 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8044 task_rq_unlock(rq, tsk, &flags);
8046 #endif /* CONFIG_CGROUP_SCHED */
8048 #ifdef CONFIG_RT_GROUP_SCHED
8050 * Ensure that the real time constraints are schedulable.
8052 static DEFINE_MUTEX(rt_constraints_mutex);
8054 /* Must be called with tasklist_lock held */
8055 static inline int tg_has_rt_tasks(struct task_group *tg)
8057 struct task_struct *g, *p;
8060 * Autogroups do not have RT tasks; see autogroup_create().
8062 if (task_group_is_autogroup(tg))
8065 for_each_process_thread(g, p) {
8066 if (rt_task(p) && task_group(p) == tg)
8073 struct rt_schedulable_data {
8074 struct task_group *tg;
8079 static int tg_rt_schedulable(struct task_group *tg, void *data)
8081 struct rt_schedulable_data *d = data;
8082 struct task_group *child;
8083 unsigned long total, sum = 0;
8084 u64 period, runtime;
8086 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8087 runtime = tg->rt_bandwidth.rt_runtime;
8090 period = d->rt_period;
8091 runtime = d->rt_runtime;
8095 * Cannot have more runtime than the period.
8097 if (runtime > period && runtime != RUNTIME_INF)
8101 * Ensure we don't starve existing RT tasks.
8103 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8106 total = to_ratio(period, runtime);
8109 * Nobody can have more than the global setting allows.
8111 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8115 * The sum of our children's runtime should not exceed our own.
8117 list_for_each_entry_rcu(child, &tg->children, siblings) {
8118 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8119 runtime = child->rt_bandwidth.rt_runtime;
8121 if (child == d->tg) {
8122 period = d->rt_period;
8123 runtime = d->rt_runtime;
8126 sum += to_ratio(period, runtime);
8135 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8139 struct rt_schedulable_data data = {
8141 .rt_period = period,
8142 .rt_runtime = runtime,
8146 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8152 static int tg_set_rt_bandwidth(struct task_group *tg,
8153 u64 rt_period, u64 rt_runtime)
8158 * Disallowing the root group RT runtime is BAD, it would disallow the
8159 * kernel creating (and or operating) RT threads.
8161 if (tg == &root_task_group && rt_runtime == 0)
8164 /* No period doesn't make any sense. */
8168 mutex_lock(&rt_constraints_mutex);
8169 read_lock(&tasklist_lock);
8170 err = __rt_schedulable(tg, rt_period, rt_runtime);
8174 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8175 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8176 tg->rt_bandwidth.rt_runtime = rt_runtime;
8178 for_each_possible_cpu(i) {
8179 struct rt_rq *rt_rq = tg->rt_rq[i];
8181 raw_spin_lock(&rt_rq->rt_runtime_lock);
8182 rt_rq->rt_runtime = rt_runtime;
8183 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8185 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8187 read_unlock(&tasklist_lock);
8188 mutex_unlock(&rt_constraints_mutex);
8193 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8195 u64 rt_runtime, rt_period;
8197 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8198 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8199 if (rt_runtime_us < 0)
8200 rt_runtime = RUNTIME_INF;
8202 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8205 static long sched_group_rt_runtime(struct task_group *tg)
8209 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8212 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8213 do_div(rt_runtime_us, NSEC_PER_USEC);
8214 return rt_runtime_us;
8217 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8219 u64 rt_runtime, rt_period;
8221 rt_period = rt_period_us * NSEC_PER_USEC;
8222 rt_runtime = tg->rt_bandwidth.rt_runtime;
8224 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8227 static long sched_group_rt_period(struct task_group *tg)
8231 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8232 do_div(rt_period_us, NSEC_PER_USEC);
8233 return rt_period_us;
8235 #endif /* CONFIG_RT_GROUP_SCHED */
8237 #ifdef CONFIG_RT_GROUP_SCHED
8238 static int sched_rt_global_constraints(void)
8242 mutex_lock(&rt_constraints_mutex);
8243 read_lock(&tasklist_lock);
8244 ret = __rt_schedulable(NULL, 0, 0);
8245 read_unlock(&tasklist_lock);
8246 mutex_unlock(&rt_constraints_mutex);
8251 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8253 /* Don't accept realtime tasks when there is no way for them to run */
8254 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8260 #else /* !CONFIG_RT_GROUP_SCHED */
8261 static int sched_rt_global_constraints(void)
8263 unsigned long flags;
8266 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8267 for_each_possible_cpu(i) {
8268 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8270 raw_spin_lock(&rt_rq->rt_runtime_lock);
8271 rt_rq->rt_runtime = global_rt_runtime();
8272 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8274 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8278 #endif /* CONFIG_RT_GROUP_SCHED */
8280 static int sched_dl_global_validate(void)
8282 u64 runtime = global_rt_runtime();
8283 u64 period = global_rt_period();
8284 u64 new_bw = to_ratio(period, runtime);
8287 unsigned long flags;
8290 * Here we want to check the bandwidth not being set to some
8291 * value smaller than the currently allocated bandwidth in
8292 * any of the root_domains.
8294 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8295 * cycling on root_domains... Discussion on different/better
8296 * solutions is welcome!
8298 for_each_possible_cpu(cpu) {
8299 rcu_read_lock_sched();
8300 dl_b = dl_bw_of(cpu);
8302 raw_spin_lock_irqsave(&dl_b->lock, flags);
8303 if (new_bw < dl_b->total_bw)
8305 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8307 rcu_read_unlock_sched();
8316 static void sched_dl_do_global(void)
8321 unsigned long flags;
8323 def_dl_bandwidth.dl_period = global_rt_period();
8324 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8326 if (global_rt_runtime() != RUNTIME_INF)
8327 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8330 * FIXME: As above...
8332 for_each_possible_cpu(cpu) {
8333 rcu_read_lock_sched();
8334 dl_b = dl_bw_of(cpu);
8336 raw_spin_lock_irqsave(&dl_b->lock, flags);
8338 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8340 rcu_read_unlock_sched();
8344 static int sched_rt_global_validate(void)
8346 if (sysctl_sched_rt_period <= 0)
8349 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8350 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8356 static void sched_rt_do_global(void)
8358 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8359 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8362 int sched_rt_handler(struct ctl_table *table, int write,
8363 void __user *buffer, size_t *lenp,
8366 int old_period, old_runtime;
8367 static DEFINE_MUTEX(mutex);
8371 old_period = sysctl_sched_rt_period;
8372 old_runtime = sysctl_sched_rt_runtime;
8374 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8376 if (!ret && write) {
8377 ret = sched_rt_global_validate();
8381 ret = sched_dl_global_validate();
8385 ret = sched_rt_global_constraints();
8389 sched_rt_do_global();
8390 sched_dl_do_global();
8394 sysctl_sched_rt_period = old_period;
8395 sysctl_sched_rt_runtime = old_runtime;
8397 mutex_unlock(&mutex);
8402 int sched_rr_handler(struct ctl_table *table, int write,
8403 void __user *buffer, size_t *lenp,
8407 static DEFINE_MUTEX(mutex);
8410 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8411 /* make sure that internally we keep jiffies */
8412 /* also, writing zero resets timeslice to default */
8413 if (!ret && write) {
8414 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8415 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8417 mutex_unlock(&mutex);
8421 #ifdef CONFIG_CGROUP_SCHED
8423 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8425 return css ? container_of(css, struct task_group, css) : NULL;
8428 static struct cgroup_subsys_state *
8429 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8431 struct task_group *parent = css_tg(parent_css);
8432 struct task_group *tg;
8435 /* This is early initialization for the top cgroup */
8436 return &root_task_group.css;
8439 tg = sched_create_group(parent);
8441 return ERR_PTR(-ENOMEM);
8443 sched_online_group(tg, parent);
8448 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8450 struct task_group *tg = css_tg(css);
8452 sched_offline_group(tg);
8455 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8457 struct task_group *tg = css_tg(css);
8460 * Relies on the RCU grace period between css_released() and this.
8462 sched_free_group(tg);
8465 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8467 sched_move_task(task);
8470 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8472 struct task_struct *task;
8473 struct cgroup_subsys_state *css;
8475 cgroup_taskset_for_each(task, css, tset) {
8476 #ifdef CONFIG_RT_GROUP_SCHED
8477 if (!sched_rt_can_attach(css_tg(css), task))
8480 /* We don't support RT-tasks being in separate groups */
8481 if (task->sched_class != &fair_sched_class)
8488 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8490 struct task_struct *task;
8491 struct cgroup_subsys_state *css;
8493 cgroup_taskset_for_each(task, css, tset)
8494 sched_move_task(task);
8497 #ifdef CONFIG_FAIR_GROUP_SCHED
8498 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8499 struct cftype *cftype, u64 shareval)
8501 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8504 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8507 struct task_group *tg = css_tg(css);
8509 return (u64) scale_load_down(tg->shares);
8512 #ifdef CONFIG_CFS_BANDWIDTH
8513 static DEFINE_MUTEX(cfs_constraints_mutex);
8515 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8516 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8518 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8520 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8522 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8523 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8525 if (tg == &root_task_group)
8529 * Ensure we have at some amount of bandwidth every period. This is
8530 * to prevent reaching a state of large arrears when throttled via
8531 * entity_tick() resulting in prolonged exit starvation.
8533 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8537 * Likewise, bound things on the otherside by preventing insane quota
8538 * periods. This also allows us to normalize in computing quota
8541 if (period > max_cfs_quota_period)
8545 * Prevent race between setting of cfs_rq->runtime_enabled and
8546 * unthrottle_offline_cfs_rqs().
8549 mutex_lock(&cfs_constraints_mutex);
8550 ret = __cfs_schedulable(tg, period, quota);
8554 runtime_enabled = quota != RUNTIME_INF;
8555 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8557 * If we need to toggle cfs_bandwidth_used, off->on must occur
8558 * before making related changes, and on->off must occur afterwards
8560 if (runtime_enabled && !runtime_was_enabled)
8561 cfs_bandwidth_usage_inc();
8562 raw_spin_lock_irq(&cfs_b->lock);
8563 cfs_b->period = ns_to_ktime(period);
8564 cfs_b->quota = quota;
8566 __refill_cfs_bandwidth_runtime(cfs_b);
8567 /* restart the period timer (if active) to handle new period expiry */
8568 if (runtime_enabled)
8569 start_cfs_bandwidth(cfs_b);
8570 raw_spin_unlock_irq(&cfs_b->lock);
8572 for_each_online_cpu(i) {
8573 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8574 struct rq *rq = cfs_rq->rq;
8576 raw_spin_lock_irq(&rq->lock);
8577 cfs_rq->runtime_enabled = runtime_enabled;
8578 cfs_rq->runtime_remaining = 0;
8580 if (cfs_rq->throttled)
8581 unthrottle_cfs_rq(cfs_rq);
8582 raw_spin_unlock_irq(&rq->lock);
8584 if (runtime_was_enabled && !runtime_enabled)
8585 cfs_bandwidth_usage_dec();
8587 mutex_unlock(&cfs_constraints_mutex);
8593 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8597 period = ktime_to_ns(tg->cfs_bandwidth.period);
8598 if (cfs_quota_us < 0)
8599 quota = RUNTIME_INF;
8601 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8603 return tg_set_cfs_bandwidth(tg, period, quota);
8606 long tg_get_cfs_quota(struct task_group *tg)
8610 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8613 quota_us = tg->cfs_bandwidth.quota;
8614 do_div(quota_us, NSEC_PER_USEC);
8619 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8623 period = (u64)cfs_period_us * NSEC_PER_USEC;
8624 quota = tg->cfs_bandwidth.quota;
8626 return tg_set_cfs_bandwidth(tg, period, quota);
8629 long tg_get_cfs_period(struct task_group *tg)
8633 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8634 do_div(cfs_period_us, NSEC_PER_USEC);
8636 return cfs_period_us;
8639 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8642 return tg_get_cfs_quota(css_tg(css));
8645 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8646 struct cftype *cftype, s64 cfs_quota_us)
8648 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8651 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8654 return tg_get_cfs_period(css_tg(css));
8657 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8658 struct cftype *cftype, u64 cfs_period_us)
8660 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8663 struct cfs_schedulable_data {
8664 struct task_group *tg;
8669 * normalize group quota/period to be quota/max_period
8670 * note: units are usecs
8672 static u64 normalize_cfs_quota(struct task_group *tg,
8673 struct cfs_schedulable_data *d)
8681 period = tg_get_cfs_period(tg);
8682 quota = tg_get_cfs_quota(tg);
8685 /* note: these should typically be equivalent */
8686 if (quota == RUNTIME_INF || quota == -1)
8689 return to_ratio(period, quota);
8692 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8694 struct cfs_schedulable_data *d = data;
8695 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8696 s64 quota = 0, parent_quota = -1;
8699 quota = RUNTIME_INF;
8701 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8703 quota = normalize_cfs_quota(tg, d);
8704 parent_quota = parent_b->hierarchical_quota;
8707 * ensure max(child_quota) <= parent_quota, inherit when no
8710 if (quota == RUNTIME_INF)
8711 quota = parent_quota;
8712 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8715 cfs_b->hierarchical_quota = quota;
8720 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8723 struct cfs_schedulable_data data = {
8729 if (quota != RUNTIME_INF) {
8730 do_div(data.period, NSEC_PER_USEC);
8731 do_div(data.quota, NSEC_PER_USEC);
8735 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8741 static int cpu_stats_show(struct seq_file *sf, void *v)
8743 struct task_group *tg = css_tg(seq_css(sf));
8744 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8746 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8747 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8748 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8752 #endif /* CONFIG_CFS_BANDWIDTH */
8753 #endif /* CONFIG_FAIR_GROUP_SCHED */
8755 #ifdef CONFIG_RT_GROUP_SCHED
8756 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8757 struct cftype *cft, s64 val)
8759 return sched_group_set_rt_runtime(css_tg(css), val);
8762 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8765 return sched_group_rt_runtime(css_tg(css));
8768 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8769 struct cftype *cftype, u64 rt_period_us)
8771 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8774 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8777 return sched_group_rt_period(css_tg(css));
8779 #endif /* CONFIG_RT_GROUP_SCHED */
8781 static struct cftype cpu_files[] = {
8782 #ifdef CONFIG_FAIR_GROUP_SCHED
8785 .read_u64 = cpu_shares_read_u64,
8786 .write_u64 = cpu_shares_write_u64,
8789 #ifdef CONFIG_CFS_BANDWIDTH
8791 .name = "cfs_quota_us",
8792 .read_s64 = cpu_cfs_quota_read_s64,
8793 .write_s64 = cpu_cfs_quota_write_s64,
8796 .name = "cfs_period_us",
8797 .read_u64 = cpu_cfs_period_read_u64,
8798 .write_u64 = cpu_cfs_period_write_u64,
8802 .seq_show = cpu_stats_show,
8805 #ifdef CONFIG_RT_GROUP_SCHED
8807 .name = "rt_runtime_us",
8808 .read_s64 = cpu_rt_runtime_read,
8809 .write_s64 = cpu_rt_runtime_write,
8812 .name = "rt_period_us",
8813 .read_u64 = cpu_rt_period_read_uint,
8814 .write_u64 = cpu_rt_period_write_uint,
8820 struct cgroup_subsys cpu_cgrp_subsys = {
8821 .css_alloc = cpu_cgroup_css_alloc,
8822 .css_released = cpu_cgroup_css_released,
8823 .css_free = cpu_cgroup_css_free,
8824 .fork = cpu_cgroup_fork,
8825 .can_attach = cpu_cgroup_can_attach,
8826 .attach = cpu_cgroup_attach,
8827 .allow_attach = subsys_cgroup_allow_attach,
8828 .legacy_cftypes = cpu_files,
8832 #endif /* CONFIG_CGROUP_SCHED */
8834 void dump_cpu_task(int cpu)
8836 pr_info("Task dump for CPU %d:\n", cpu);
8837 sched_show_task(cpu_curr(cpu));