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/ctype.h>
70 #include <linux/ftrace.h>
71 #include <linux/slab.h>
72 #include <linux/init_task.h>
73 #include <linux/context_tracking.h>
74 #include <linux/compiler.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_internal.h"
86 #include "../smpboot.h"
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/sched.h>
91 DEFINE_MUTEX(sched_domains_mutex);
92 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
94 static void update_rq_clock_task(struct rq *rq, s64 delta);
96 void update_rq_clock(struct rq *rq)
100 lockdep_assert_held(&rq->lock);
102 if (rq->clock_skip_update & RQCF_ACT_SKIP)
105 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
109 update_rq_clock_task(rq, delta);
113 * Debugging: various feature bits
116 #define SCHED_FEAT(name, enabled) \
117 (1UL << __SCHED_FEAT_##name) * enabled |
119 const_debug unsigned int sysctl_sched_features =
120 #include "features.h"
126 * Number of tasks to iterate in a single balance run.
127 * Limited because this is done with IRQs disabled.
129 const_debug unsigned int sysctl_sched_nr_migrate = 32;
132 * period over which we average the RT time consumption, measured
137 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
140 * period over which we measure -rt task cpu usage in us.
143 unsigned int sysctl_sched_rt_period = 1000000;
145 __read_mostly int scheduler_running;
148 * part of the period that we allow rt tasks to run in us.
151 int sysctl_sched_rt_runtime = 950000;
153 /* cpus with isolated domains */
154 cpumask_var_t cpu_isolated_map;
157 * this_rq_lock - lock this runqueue and disable interrupts.
159 static struct rq *this_rq_lock(void)
166 raw_spin_lock(&rq->lock);
171 #ifdef CONFIG_SCHED_HRTICK
173 * Use HR-timers to deliver accurate preemption points.
176 static void hrtick_clear(struct rq *rq)
178 if (hrtimer_active(&rq->hrtick_timer))
179 hrtimer_cancel(&rq->hrtick_timer);
183 * High-resolution timer tick.
184 * Runs from hardirq context with interrupts disabled.
186 static enum hrtimer_restart hrtick(struct hrtimer *timer)
188 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
190 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
192 raw_spin_lock(&rq->lock);
194 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
195 raw_spin_unlock(&rq->lock);
197 return HRTIMER_NORESTART;
202 static void __hrtick_restart(struct rq *rq)
204 struct hrtimer *timer = &rq->hrtick_timer;
206 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
210 * called from hardirq (IPI) context
212 static void __hrtick_start(void *arg)
216 raw_spin_lock(&rq->lock);
217 __hrtick_restart(rq);
218 rq->hrtick_csd_pending = 0;
219 raw_spin_unlock(&rq->lock);
223 * Called to set the hrtick timer state.
225 * called with rq->lock held and irqs disabled
227 void hrtick_start(struct rq *rq, u64 delay)
229 struct hrtimer *timer = &rq->hrtick_timer;
234 * Don't schedule slices shorter than 10000ns, that just
235 * doesn't make sense and can cause timer DoS.
237 delta = max_t(s64, delay, 10000LL);
238 time = ktime_add_ns(timer->base->get_time(), delta);
240 hrtimer_set_expires(timer, time);
242 if (rq == this_rq()) {
243 __hrtick_restart(rq);
244 } else if (!rq->hrtick_csd_pending) {
245 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
246 rq->hrtick_csd_pending = 1;
251 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
253 int cpu = (int)(long)hcpu;
256 case CPU_UP_CANCELED:
257 case CPU_UP_CANCELED_FROZEN:
258 case CPU_DOWN_PREPARE:
259 case CPU_DOWN_PREPARE_FROZEN:
261 case CPU_DEAD_FROZEN:
262 hrtick_clear(cpu_rq(cpu));
269 static __init void init_hrtick(void)
271 hotcpu_notifier(hotplug_hrtick, 0);
275 * Called to set the hrtick timer state.
277 * called with rq->lock held and irqs disabled
279 void hrtick_start(struct rq *rq, u64 delay)
282 * Don't schedule slices shorter than 10000ns, that just
283 * doesn't make sense. Rely on vruntime for fairness.
285 delay = max_t(u64, delay, 10000LL);
286 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
287 HRTIMER_MODE_REL_PINNED);
290 static inline void init_hrtick(void)
293 #endif /* CONFIG_SMP */
295 static void init_rq_hrtick(struct rq *rq)
298 rq->hrtick_csd_pending = 0;
300 rq->hrtick_csd.flags = 0;
301 rq->hrtick_csd.func = __hrtick_start;
302 rq->hrtick_csd.info = rq;
305 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
306 rq->hrtick_timer.function = hrtick;
308 #else /* CONFIG_SCHED_HRTICK */
309 static inline void hrtick_clear(struct rq *rq)
313 static inline void init_rq_hrtick(struct rq *rq)
317 static inline void init_hrtick(void)
320 #endif /* CONFIG_SCHED_HRTICK */
323 * cmpxchg based fetch_or, macro so it works for different integer types
325 #define fetch_or(ptr, val) \
326 ({ typeof(*(ptr)) __old, __val = *(ptr); \
328 __old = cmpxchg((ptr), __val, __val | (val)); \
329 if (__old == __val) \
336 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
338 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
339 * this avoids any races wrt polling state changes and thereby avoids
342 static bool set_nr_and_not_polling(struct task_struct *p)
344 struct thread_info *ti = task_thread_info(p);
345 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
349 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
351 * If this returns true, then the idle task promises to call
352 * sched_ttwu_pending() and reschedule soon.
354 static bool set_nr_if_polling(struct task_struct *p)
356 struct thread_info *ti = task_thread_info(p);
357 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
360 if (!(val & _TIF_POLLING_NRFLAG))
362 if (val & _TIF_NEED_RESCHED)
364 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
373 static bool set_nr_and_not_polling(struct task_struct *p)
375 set_tsk_need_resched(p);
380 static bool set_nr_if_polling(struct task_struct *p)
387 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
389 struct wake_q_node *node = &task->wake_q;
392 * Atomically grab the task, if ->wake_q is !nil already it means
393 * its already queued (either by us or someone else) and will get the
394 * wakeup due to that.
396 * This cmpxchg() implies a full barrier, which pairs with the write
397 * barrier implied by the wakeup in wake_up_list().
399 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
402 get_task_struct(task);
405 * The head is context local, there can be no concurrency.
408 head->lastp = &node->next;
411 void wake_up_q(struct wake_q_head *head)
413 struct wake_q_node *node = head->first;
415 while (node != WAKE_Q_TAIL) {
416 struct task_struct *task;
418 task = container_of(node, struct task_struct, wake_q);
420 /* task can safely be re-inserted now */
422 task->wake_q.next = NULL;
425 * wake_up_process() implies a wmb() to pair with the queueing
426 * in wake_q_add() so as not to miss wakeups.
428 wake_up_process(task);
429 put_task_struct(task);
434 * resched_curr - mark rq's current task 'to be rescheduled now'.
436 * On UP this means the setting of the need_resched flag, on SMP it
437 * might also involve a cross-CPU call to trigger the scheduler on
440 void resched_curr(struct rq *rq)
442 struct task_struct *curr = rq->curr;
445 lockdep_assert_held(&rq->lock);
447 if (test_tsk_need_resched(curr))
452 if (cpu == smp_processor_id()) {
453 set_tsk_need_resched(curr);
454 set_preempt_need_resched();
458 if (set_nr_and_not_polling(curr))
459 smp_send_reschedule(cpu);
461 trace_sched_wake_idle_without_ipi(cpu);
464 void resched_cpu(int cpu)
466 struct rq *rq = cpu_rq(cpu);
469 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
472 raw_spin_unlock_irqrestore(&rq->lock, flags);
476 #ifdef CONFIG_NO_HZ_COMMON
478 * In the semi idle case, use the nearest busy cpu for migrating timers
479 * from an idle cpu. This is good for power-savings.
481 * We don't do similar optimization for completely idle system, as
482 * selecting an idle cpu will add more delays to the timers than intended
483 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
485 int get_nohz_timer_target(void)
487 int i, cpu = smp_processor_id();
488 struct sched_domain *sd;
490 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
494 for_each_domain(cpu, sd) {
495 for_each_cpu(i, sched_domain_span(sd)) {
496 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
503 if (!is_housekeeping_cpu(cpu))
504 cpu = housekeeping_any_cpu();
510 * When add_timer_on() enqueues a timer into the timer wheel of an
511 * idle CPU then this timer might expire before the next timer event
512 * which is scheduled to wake up that CPU. In case of a completely
513 * idle system the next event might even be infinite time into the
514 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
515 * leaves the inner idle loop so the newly added timer is taken into
516 * account when the CPU goes back to idle and evaluates the timer
517 * wheel for the next timer event.
519 static void wake_up_idle_cpu(int cpu)
521 struct rq *rq = cpu_rq(cpu);
523 if (cpu == smp_processor_id())
526 if (set_nr_and_not_polling(rq->idle))
527 smp_send_reschedule(cpu);
529 trace_sched_wake_idle_without_ipi(cpu);
532 static bool wake_up_full_nohz_cpu(int cpu)
535 * We just need the target to call irq_exit() and re-evaluate
536 * the next tick. The nohz full kick at least implies that.
537 * If needed we can still optimize that later with an
540 if (tick_nohz_full_cpu(cpu)) {
541 if (cpu != smp_processor_id() ||
542 tick_nohz_tick_stopped())
543 tick_nohz_full_kick_cpu(cpu);
550 void wake_up_nohz_cpu(int cpu)
552 if (!wake_up_full_nohz_cpu(cpu))
553 wake_up_idle_cpu(cpu);
556 static inline bool got_nohz_idle_kick(void)
558 int cpu = smp_processor_id();
560 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
563 if (idle_cpu(cpu) && !need_resched())
567 * We can't run Idle Load Balance on this CPU for this time so we
568 * cancel it and clear NOHZ_BALANCE_KICK
570 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
574 #else /* CONFIG_NO_HZ_COMMON */
576 static inline bool got_nohz_idle_kick(void)
581 #endif /* CONFIG_NO_HZ_COMMON */
583 #ifdef CONFIG_NO_HZ_FULL
584 bool sched_can_stop_tick(void)
587 * FIFO realtime policy runs the highest priority task. Other runnable
588 * tasks are of a lower priority. The scheduler tick does nothing.
590 if (current->policy == SCHED_FIFO)
594 * Round-robin realtime tasks time slice with other tasks at the same
595 * realtime priority. Is this task the only one at this priority?
597 if (current->policy == SCHED_RR) {
598 struct sched_rt_entity *rt_se = ¤t->rt;
600 return list_is_singular(&rt_se->run_list);
604 * More than one running task need preemption.
605 * nr_running update is assumed to be visible
606 * after IPI is sent from wakers.
608 if (this_rq()->nr_running > 1)
613 #endif /* CONFIG_NO_HZ_FULL */
615 void sched_avg_update(struct rq *rq)
617 s64 period = sched_avg_period();
619 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
621 * Inline assembly required to prevent the compiler
622 * optimising this loop into a divmod call.
623 * See __iter_div_u64_rem() for another example of this.
625 asm("" : "+rm" (rq->age_stamp));
626 rq->age_stamp += period;
631 #endif /* CONFIG_SMP */
633 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
634 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
636 * Iterate task_group tree rooted at *from, calling @down when first entering a
637 * node and @up when leaving it for the final time.
639 * Caller must hold rcu_lock or sufficient equivalent.
641 int walk_tg_tree_from(struct task_group *from,
642 tg_visitor down, tg_visitor up, void *data)
644 struct task_group *parent, *child;
650 ret = (*down)(parent, data);
653 list_for_each_entry_rcu(child, &parent->children, siblings) {
660 ret = (*up)(parent, data);
661 if (ret || parent == from)
665 parent = parent->parent;
672 int tg_nop(struct task_group *tg, void *data)
678 static void set_load_weight(struct task_struct *p)
680 int prio = p->static_prio - MAX_RT_PRIO;
681 struct load_weight *load = &p->se.load;
684 * SCHED_IDLE tasks get minimal weight:
686 if (idle_policy(p->policy)) {
687 load->weight = scale_load(WEIGHT_IDLEPRIO);
688 load->inv_weight = WMULT_IDLEPRIO;
692 load->weight = scale_load(sched_prio_to_weight[prio]);
693 load->inv_weight = sched_prio_to_wmult[prio];
696 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
699 if (!(flags & ENQUEUE_RESTORE))
700 sched_info_queued(rq, p);
701 p->sched_class->enqueue_task(rq, p, flags);
704 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
707 if (!(flags & DEQUEUE_SAVE))
708 sched_info_dequeued(rq, p);
709 p->sched_class->dequeue_task(rq, p, flags);
712 void activate_task(struct rq *rq, struct task_struct *p, int flags)
714 if (task_contributes_to_load(p))
715 rq->nr_uninterruptible--;
717 enqueue_task(rq, p, flags);
720 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
722 if (task_contributes_to_load(p))
723 rq->nr_uninterruptible++;
725 dequeue_task(rq, p, flags);
728 static void update_rq_clock_task(struct rq *rq, s64 delta)
731 * In theory, the compile should just see 0 here, and optimize out the call
732 * to sched_rt_avg_update. But I don't trust it...
734 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
735 s64 steal = 0, irq_delta = 0;
737 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
738 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
741 * Since irq_time is only updated on {soft,}irq_exit, we might run into
742 * this case when a previous update_rq_clock() happened inside a
745 * When this happens, we stop ->clock_task and only update the
746 * prev_irq_time stamp to account for the part that fit, so that a next
747 * update will consume the rest. This ensures ->clock_task is
750 * It does however cause some slight miss-attribution of {soft,}irq
751 * time, a more accurate solution would be to update the irq_time using
752 * the current rq->clock timestamp, except that would require using
755 if (irq_delta > delta)
758 rq->prev_irq_time += irq_delta;
761 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
762 if (static_key_false((¶virt_steal_rq_enabled))) {
763 steal = paravirt_steal_clock(cpu_of(rq));
764 steal -= rq->prev_steal_time_rq;
766 if (unlikely(steal > delta))
769 rq->prev_steal_time_rq += steal;
774 rq->clock_task += delta;
776 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
777 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
778 sched_rt_avg_update(rq, irq_delta + steal);
782 void sched_set_stop_task(int cpu, struct task_struct *stop)
784 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
785 struct task_struct *old_stop = cpu_rq(cpu)->stop;
789 * Make it appear like a SCHED_FIFO task, its something
790 * userspace knows about and won't get confused about.
792 * Also, it will make PI more or less work without too
793 * much confusion -- but then, stop work should not
794 * rely on PI working anyway.
796 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
798 stop->sched_class = &stop_sched_class;
801 cpu_rq(cpu)->stop = stop;
805 * Reset it back to a normal scheduling class so that
806 * it can die in pieces.
808 old_stop->sched_class = &rt_sched_class;
813 * __normal_prio - return the priority that is based on the static prio
815 static inline int __normal_prio(struct task_struct *p)
817 return p->static_prio;
821 * Calculate the expected normal priority: i.e. priority
822 * without taking RT-inheritance into account. Might be
823 * boosted by interactivity modifiers. Changes upon fork,
824 * setprio syscalls, and whenever the interactivity
825 * estimator recalculates.
827 static inline int normal_prio(struct task_struct *p)
831 if (task_has_dl_policy(p))
832 prio = MAX_DL_PRIO-1;
833 else if (task_has_rt_policy(p))
834 prio = MAX_RT_PRIO-1 - p->rt_priority;
836 prio = __normal_prio(p);
841 * Calculate the current priority, i.e. the priority
842 * taken into account by the scheduler. This value might
843 * be boosted by RT tasks, or might be boosted by
844 * interactivity modifiers. Will be RT if the task got
845 * RT-boosted. If not then it returns p->normal_prio.
847 static int effective_prio(struct task_struct *p)
849 p->normal_prio = normal_prio(p);
851 * If we are RT tasks or we were boosted to RT priority,
852 * keep the priority unchanged. Otherwise, update priority
853 * to the normal priority:
855 if (!rt_prio(p->prio))
856 return p->normal_prio;
861 * task_curr - is this task currently executing on a CPU?
862 * @p: the task in question.
864 * Return: 1 if the task is currently executing. 0 otherwise.
866 inline int task_curr(const struct task_struct *p)
868 return cpu_curr(task_cpu(p)) == p;
872 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
873 * use the balance_callback list if you want balancing.
875 * this means any call to check_class_changed() must be followed by a call to
876 * balance_callback().
878 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
879 const struct sched_class *prev_class,
882 if (prev_class != p->sched_class) {
883 if (prev_class->switched_from)
884 prev_class->switched_from(rq, p);
886 p->sched_class->switched_to(rq, p);
887 } else if (oldprio != p->prio || dl_task(p))
888 p->sched_class->prio_changed(rq, p, oldprio);
891 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
893 const struct sched_class *class;
895 if (p->sched_class == rq->curr->sched_class) {
896 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
898 for_each_class(class) {
899 if (class == rq->curr->sched_class)
901 if (class == p->sched_class) {
909 * A queue event has occurred, and we're going to schedule. In
910 * this case, we can save a useless back to back clock update.
912 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
913 rq_clock_skip_update(rq, true);
918 * This is how migration works:
920 * 1) we invoke migration_cpu_stop() on the target CPU using
922 * 2) stopper starts to run (implicitly forcing the migrated thread
924 * 3) it checks whether the migrated task is still in the wrong runqueue.
925 * 4) if it's in the wrong runqueue then the migration thread removes
926 * it and puts it into the right queue.
927 * 5) stopper completes and stop_one_cpu() returns and the migration
932 * move_queued_task - move a queued task to new rq.
934 * Returns (locked) new rq. Old rq's lock is released.
936 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
938 lockdep_assert_held(&rq->lock);
940 p->on_rq = TASK_ON_RQ_MIGRATING;
941 dequeue_task(rq, p, 0);
942 set_task_cpu(p, new_cpu);
943 raw_spin_unlock(&rq->lock);
945 rq = cpu_rq(new_cpu);
947 raw_spin_lock(&rq->lock);
948 BUG_ON(task_cpu(p) != new_cpu);
949 enqueue_task(rq, p, 0);
950 p->on_rq = TASK_ON_RQ_QUEUED;
951 check_preempt_curr(rq, p, 0);
956 struct migration_arg {
957 struct task_struct *task;
962 * Move (not current) task off this cpu, onto dest cpu. We're doing
963 * this because either it can't run here any more (set_cpus_allowed()
964 * away from this CPU, or CPU going down), or because we're
965 * attempting to rebalance this task on exec (sched_exec).
967 * So we race with normal scheduler movements, but that's OK, as long
968 * as the task is no longer on this CPU.
970 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
972 if (unlikely(!cpu_active(dest_cpu)))
975 /* Affinity changed (again). */
976 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
979 rq = move_queued_task(rq, p, dest_cpu);
985 * migration_cpu_stop - this will be executed by a highprio stopper thread
986 * and performs thread migration by bumping thread off CPU then
987 * 'pushing' onto another runqueue.
989 static int migration_cpu_stop(void *data)
991 struct migration_arg *arg = data;
992 struct task_struct *p = arg->task;
993 struct rq *rq = this_rq();
996 * The original target cpu might have gone down and we might
997 * be on another cpu but it doesn't matter.
1001 * We need to explicitly wake pending tasks before running
1002 * __migrate_task() such that we will not miss enforcing cpus_allowed
1003 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1005 sched_ttwu_pending();
1007 raw_spin_lock(&p->pi_lock);
1008 raw_spin_lock(&rq->lock);
1010 * If task_rq(p) != rq, it cannot be migrated here, because we're
1011 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1012 * we're holding p->pi_lock.
1014 if (task_rq(p) == rq && task_on_rq_queued(p))
1015 rq = __migrate_task(rq, p, arg->dest_cpu);
1016 raw_spin_unlock(&rq->lock);
1017 raw_spin_unlock(&p->pi_lock);
1024 * sched_class::set_cpus_allowed must do the below, but is not required to
1025 * actually call this function.
1027 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1029 cpumask_copy(&p->cpus_allowed, new_mask);
1030 p->nr_cpus_allowed = cpumask_weight(new_mask);
1033 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1035 struct rq *rq = task_rq(p);
1036 bool queued, running;
1038 lockdep_assert_held(&p->pi_lock);
1040 queued = task_on_rq_queued(p);
1041 running = task_current(rq, p);
1045 * Because __kthread_bind() calls this on blocked tasks without
1048 lockdep_assert_held(&rq->lock);
1049 dequeue_task(rq, p, DEQUEUE_SAVE);
1052 put_prev_task(rq, p);
1054 p->sched_class->set_cpus_allowed(p, new_mask);
1057 p->sched_class->set_curr_task(rq);
1059 enqueue_task(rq, p, ENQUEUE_RESTORE);
1063 * Change a given task's CPU affinity. Migrate the thread to a
1064 * proper CPU and schedule it away if the CPU it's executing on
1065 * is removed from the allowed bitmask.
1067 * NOTE: the caller must have a valid reference to the task, the
1068 * task must not exit() & deallocate itself prematurely. The
1069 * call is not atomic; no spinlocks may be held.
1071 static int __set_cpus_allowed_ptr(struct task_struct *p,
1072 const struct cpumask *new_mask, bool check)
1074 unsigned long flags;
1076 unsigned int dest_cpu;
1079 rq = task_rq_lock(p, &flags);
1082 * Must re-check here, to close a race against __kthread_bind(),
1083 * sched_setaffinity() is not guaranteed to observe the flag.
1085 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1090 if (cpumask_equal(&p->cpus_allowed, new_mask))
1093 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1098 do_set_cpus_allowed(p, new_mask);
1100 /* Can the task run on the task's current CPU? If so, we're done */
1101 if (cpumask_test_cpu(task_cpu(p), new_mask))
1104 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1105 if (task_running(rq, p) || p->state == TASK_WAKING) {
1106 struct migration_arg arg = { p, dest_cpu };
1107 /* Need help from migration thread: drop lock and wait. */
1108 task_rq_unlock(rq, p, &flags);
1109 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1110 tlb_migrate_finish(p->mm);
1112 } else if (task_on_rq_queued(p)) {
1114 * OK, since we're going to drop the lock immediately
1115 * afterwards anyway.
1117 lockdep_unpin_lock(&rq->lock);
1118 rq = move_queued_task(rq, p, dest_cpu);
1119 lockdep_pin_lock(&rq->lock);
1122 task_rq_unlock(rq, p, &flags);
1127 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1129 return __set_cpus_allowed_ptr(p, new_mask, false);
1131 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1133 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1135 #ifdef CONFIG_SCHED_DEBUG
1137 * We should never call set_task_cpu() on a blocked task,
1138 * ttwu() will sort out the placement.
1140 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1144 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1145 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1146 * time relying on p->on_rq.
1148 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1149 p->sched_class == &fair_sched_class &&
1150 (p->on_rq && !task_on_rq_migrating(p)));
1152 #ifdef CONFIG_LOCKDEP
1154 * The caller should hold either p->pi_lock or rq->lock, when changing
1155 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1157 * sched_move_task() holds both and thus holding either pins the cgroup,
1160 * Furthermore, all task_rq users should acquire both locks, see
1163 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1164 lockdep_is_held(&task_rq(p)->lock)));
1168 trace_sched_migrate_task(p, new_cpu);
1170 if (task_cpu(p) != new_cpu) {
1171 if (p->sched_class->migrate_task_rq)
1172 p->sched_class->migrate_task_rq(p);
1173 p->se.nr_migrations++;
1174 perf_event_task_migrate(p);
1177 __set_task_cpu(p, new_cpu);
1180 static void __migrate_swap_task(struct task_struct *p, int cpu)
1182 if (task_on_rq_queued(p)) {
1183 struct rq *src_rq, *dst_rq;
1185 src_rq = task_rq(p);
1186 dst_rq = cpu_rq(cpu);
1188 p->on_rq = TASK_ON_RQ_MIGRATING;
1189 deactivate_task(src_rq, p, 0);
1190 set_task_cpu(p, cpu);
1191 activate_task(dst_rq, p, 0);
1192 p->on_rq = TASK_ON_RQ_QUEUED;
1193 check_preempt_curr(dst_rq, p, 0);
1196 * Task isn't running anymore; make it appear like we migrated
1197 * it before it went to sleep. This means on wakeup we make the
1198 * previous cpu our targer instead of where it really is.
1204 struct migration_swap_arg {
1205 struct task_struct *src_task, *dst_task;
1206 int src_cpu, dst_cpu;
1209 static int migrate_swap_stop(void *data)
1211 struct migration_swap_arg *arg = data;
1212 struct rq *src_rq, *dst_rq;
1215 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1218 src_rq = cpu_rq(arg->src_cpu);
1219 dst_rq = cpu_rq(arg->dst_cpu);
1221 double_raw_lock(&arg->src_task->pi_lock,
1222 &arg->dst_task->pi_lock);
1223 double_rq_lock(src_rq, dst_rq);
1225 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1228 if (task_cpu(arg->src_task) != arg->src_cpu)
1231 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1234 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1237 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1238 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1243 double_rq_unlock(src_rq, dst_rq);
1244 raw_spin_unlock(&arg->dst_task->pi_lock);
1245 raw_spin_unlock(&arg->src_task->pi_lock);
1251 * Cross migrate two tasks
1253 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1255 struct migration_swap_arg arg;
1258 arg = (struct migration_swap_arg){
1260 .src_cpu = task_cpu(cur),
1262 .dst_cpu = task_cpu(p),
1265 if (arg.src_cpu == arg.dst_cpu)
1269 * These three tests are all lockless; this is OK since all of them
1270 * will be re-checked with proper locks held further down the line.
1272 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1275 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1278 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1281 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1282 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1289 * wait_task_inactive - wait for a thread to unschedule.
1291 * If @match_state is nonzero, it's the @p->state value just checked and
1292 * not expected to change. If it changes, i.e. @p might have woken up,
1293 * then return zero. When we succeed in waiting for @p to be off its CPU,
1294 * we return a positive number (its total switch count). If a second call
1295 * a short while later returns the same number, the caller can be sure that
1296 * @p has remained unscheduled the whole time.
1298 * The caller must ensure that the task *will* unschedule sometime soon,
1299 * else this function might spin for a *long* time. This function can't
1300 * be called with interrupts off, or it may introduce deadlock with
1301 * smp_call_function() if an IPI is sent by the same process we are
1302 * waiting to become inactive.
1304 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1306 unsigned long flags;
1307 int running, queued;
1313 * We do the initial early heuristics without holding
1314 * any task-queue locks at all. We'll only try to get
1315 * the runqueue lock when things look like they will
1321 * If the task is actively running on another CPU
1322 * still, just relax and busy-wait without holding
1325 * NOTE! Since we don't hold any locks, it's not
1326 * even sure that "rq" stays as the right runqueue!
1327 * But we don't care, since "task_running()" will
1328 * return false if the runqueue has changed and p
1329 * is actually now running somewhere else!
1331 while (task_running(rq, p)) {
1332 if (match_state && unlikely(p->state != match_state))
1338 * Ok, time to look more closely! We need the rq
1339 * lock now, to be *sure*. If we're wrong, we'll
1340 * just go back and repeat.
1342 rq = task_rq_lock(p, &flags);
1343 trace_sched_wait_task(p);
1344 running = task_running(rq, p);
1345 queued = task_on_rq_queued(p);
1347 if (!match_state || p->state == match_state)
1348 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1349 task_rq_unlock(rq, p, &flags);
1352 * If it changed from the expected state, bail out now.
1354 if (unlikely(!ncsw))
1358 * Was it really running after all now that we
1359 * checked with the proper locks actually held?
1361 * Oops. Go back and try again..
1363 if (unlikely(running)) {
1369 * It's not enough that it's not actively running,
1370 * it must be off the runqueue _entirely_, and not
1373 * So if it was still runnable (but just not actively
1374 * running right now), it's preempted, and we should
1375 * yield - it could be a while.
1377 if (unlikely(queued)) {
1378 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1380 set_current_state(TASK_UNINTERRUPTIBLE);
1381 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1386 * Ahh, all good. It wasn't running, and it wasn't
1387 * runnable, which means that it will never become
1388 * running in the future either. We're all done!
1397 * kick_process - kick a running thread to enter/exit the kernel
1398 * @p: the to-be-kicked thread
1400 * Cause a process which is running on another CPU to enter
1401 * kernel-mode, without any delay. (to get signals handled.)
1403 * NOTE: this function doesn't have to take the runqueue lock,
1404 * because all it wants to ensure is that the remote task enters
1405 * the kernel. If the IPI races and the task has been migrated
1406 * to another CPU then no harm is done and the purpose has been
1409 void kick_process(struct task_struct *p)
1415 if ((cpu != smp_processor_id()) && task_curr(p))
1416 smp_send_reschedule(cpu);
1419 EXPORT_SYMBOL_GPL(kick_process);
1422 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1424 static int select_fallback_rq(int cpu, struct task_struct *p)
1426 int nid = cpu_to_node(cpu);
1427 const struct cpumask *nodemask = NULL;
1428 enum { cpuset, possible, fail } state = cpuset;
1432 * If the node that the cpu is on has been offlined, cpu_to_node()
1433 * will return -1. There is no cpu on the node, and we should
1434 * select the cpu on the other node.
1437 nodemask = cpumask_of_node(nid);
1439 /* Look for allowed, online CPU in same node. */
1440 for_each_cpu(dest_cpu, nodemask) {
1441 if (!cpu_online(dest_cpu))
1443 if (!cpu_active(dest_cpu))
1445 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1451 /* Any allowed, online CPU? */
1452 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1453 if (!cpu_online(dest_cpu))
1455 if (!cpu_active(dest_cpu))
1460 /* No more Mr. Nice Guy. */
1463 if (IS_ENABLED(CONFIG_CPUSETS)) {
1464 cpuset_cpus_allowed_fallback(p);
1470 do_set_cpus_allowed(p, cpu_possible_mask);
1481 if (state != cpuset) {
1483 * Don't tell them about moving exiting tasks or
1484 * kernel threads (both mm NULL), since they never
1487 if (p->mm && printk_ratelimit()) {
1488 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1489 task_pid_nr(p), p->comm, cpu);
1497 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1500 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1502 lockdep_assert_held(&p->pi_lock);
1504 if (p->nr_cpus_allowed > 1)
1505 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1508 * In order not to call set_task_cpu() on a blocking task we need
1509 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1512 * Since this is common to all placement strategies, this lives here.
1514 * [ this allows ->select_task() to simply return task_cpu(p) and
1515 * not worry about this generic constraint ]
1517 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1519 cpu = select_fallback_rq(task_cpu(p), p);
1524 static void update_avg(u64 *avg, u64 sample)
1526 s64 diff = sample - *avg;
1532 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1533 const struct cpumask *new_mask, bool check)
1535 return set_cpus_allowed_ptr(p, new_mask);
1538 #endif /* CONFIG_SMP */
1541 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1543 #ifdef CONFIG_SCHEDSTATS
1544 struct rq *rq = this_rq();
1547 int this_cpu = smp_processor_id();
1549 if (cpu == this_cpu) {
1550 schedstat_inc(rq, ttwu_local);
1551 schedstat_inc(p, se.statistics.nr_wakeups_local);
1553 struct sched_domain *sd;
1555 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1557 for_each_domain(this_cpu, sd) {
1558 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1559 schedstat_inc(sd, ttwu_wake_remote);
1566 if (wake_flags & WF_MIGRATED)
1567 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1569 #endif /* CONFIG_SMP */
1571 schedstat_inc(rq, ttwu_count);
1572 schedstat_inc(p, se.statistics.nr_wakeups);
1574 if (wake_flags & WF_SYNC)
1575 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1577 #endif /* CONFIG_SCHEDSTATS */
1580 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1582 activate_task(rq, p, en_flags);
1583 p->on_rq = TASK_ON_RQ_QUEUED;
1585 /* if a worker is waking up, notify workqueue */
1586 if (p->flags & PF_WQ_WORKER)
1587 wq_worker_waking_up(p, cpu_of(rq));
1591 * Mark the task runnable and perform wakeup-preemption.
1594 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1596 check_preempt_curr(rq, p, wake_flags);
1597 p->state = TASK_RUNNING;
1598 trace_sched_wakeup(p);
1601 if (p->sched_class->task_woken) {
1603 * Our task @p is fully woken up and running; so its safe to
1604 * drop the rq->lock, hereafter rq is only used for statistics.
1606 lockdep_unpin_lock(&rq->lock);
1607 p->sched_class->task_woken(rq, p);
1608 lockdep_pin_lock(&rq->lock);
1611 if (rq->idle_stamp) {
1612 u64 delta = rq_clock(rq) - rq->idle_stamp;
1613 u64 max = 2*rq->max_idle_balance_cost;
1615 update_avg(&rq->avg_idle, delta);
1617 if (rq->avg_idle > max)
1626 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1628 lockdep_assert_held(&rq->lock);
1631 if (p->sched_contributes_to_load)
1632 rq->nr_uninterruptible--;
1635 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1636 ttwu_do_wakeup(rq, p, wake_flags);
1640 * Called in case the task @p isn't fully descheduled from its runqueue,
1641 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1642 * since all we need to do is flip p->state to TASK_RUNNING, since
1643 * the task is still ->on_rq.
1645 static int ttwu_remote(struct task_struct *p, int wake_flags)
1650 rq = __task_rq_lock(p);
1651 if (task_on_rq_queued(p)) {
1652 /* check_preempt_curr() may use rq clock */
1653 update_rq_clock(rq);
1654 ttwu_do_wakeup(rq, p, wake_flags);
1657 __task_rq_unlock(rq);
1663 void sched_ttwu_pending(void)
1665 struct rq *rq = this_rq();
1666 struct llist_node *llist = llist_del_all(&rq->wake_list);
1667 struct task_struct *p;
1668 unsigned long flags;
1673 raw_spin_lock_irqsave(&rq->lock, flags);
1674 lockdep_pin_lock(&rq->lock);
1677 p = llist_entry(llist, struct task_struct, wake_entry);
1678 llist = llist_next(llist);
1679 ttwu_do_activate(rq, p, 0);
1682 lockdep_unpin_lock(&rq->lock);
1683 raw_spin_unlock_irqrestore(&rq->lock, flags);
1686 void scheduler_ipi(void)
1689 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1690 * TIF_NEED_RESCHED remotely (for the first time) will also send
1693 preempt_fold_need_resched();
1695 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1699 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1700 * traditionally all their work was done from the interrupt return
1701 * path. Now that we actually do some work, we need to make sure
1704 * Some archs already do call them, luckily irq_enter/exit nest
1707 * Arguably we should visit all archs and update all handlers,
1708 * however a fair share of IPIs are still resched only so this would
1709 * somewhat pessimize the simple resched case.
1712 sched_ttwu_pending();
1715 * Check if someone kicked us for doing the nohz idle load balance.
1717 if (unlikely(got_nohz_idle_kick())) {
1718 this_rq()->idle_balance = 1;
1719 raise_softirq_irqoff(SCHED_SOFTIRQ);
1724 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1726 struct rq *rq = cpu_rq(cpu);
1728 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1729 if (!set_nr_if_polling(rq->idle))
1730 smp_send_reschedule(cpu);
1732 trace_sched_wake_idle_without_ipi(cpu);
1736 void wake_up_if_idle(int cpu)
1738 struct rq *rq = cpu_rq(cpu);
1739 unsigned long flags;
1743 if (!is_idle_task(rcu_dereference(rq->curr)))
1746 if (set_nr_if_polling(rq->idle)) {
1747 trace_sched_wake_idle_without_ipi(cpu);
1749 raw_spin_lock_irqsave(&rq->lock, flags);
1750 if (is_idle_task(rq->curr))
1751 smp_send_reschedule(cpu);
1752 /* Else cpu is not in idle, do nothing here */
1753 raw_spin_unlock_irqrestore(&rq->lock, flags);
1760 bool cpus_share_cache(int this_cpu, int that_cpu)
1762 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1764 #endif /* CONFIG_SMP */
1766 static void ttwu_queue(struct task_struct *p, int cpu)
1768 struct rq *rq = cpu_rq(cpu);
1770 #if defined(CONFIG_SMP)
1771 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1772 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1773 ttwu_queue_remote(p, cpu);
1778 raw_spin_lock(&rq->lock);
1779 lockdep_pin_lock(&rq->lock);
1780 ttwu_do_activate(rq, p, 0);
1781 lockdep_unpin_lock(&rq->lock);
1782 raw_spin_unlock(&rq->lock);
1786 * Notes on Program-Order guarantees on SMP systems.
1790 * The basic program-order guarantee on SMP systems is that when a task [t]
1791 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1792 * execution on its new cpu [c1].
1794 * For migration (of runnable tasks) this is provided by the following means:
1796 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1797 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1798 * rq(c1)->lock (if not at the same time, then in that order).
1799 * C) LOCK of the rq(c1)->lock scheduling in task
1801 * Transitivity guarantees that B happens after A and C after B.
1802 * Note: we only require RCpc transitivity.
1803 * Note: the cpu doing B need not be c0 or c1
1812 * UNLOCK rq(0)->lock
1814 * LOCK rq(0)->lock // orders against CPU0
1816 * UNLOCK rq(0)->lock
1820 * UNLOCK rq(1)->lock
1822 * LOCK rq(1)->lock // orders against CPU2
1825 * UNLOCK rq(1)->lock
1828 * BLOCKING -- aka. SLEEP + WAKEUP
1830 * For blocking we (obviously) need to provide the same guarantee as for
1831 * migration. However the means are completely different as there is no lock
1832 * chain to provide order. Instead we do:
1834 * 1) smp_store_release(X->on_cpu, 0)
1835 * 2) smp_cond_acquire(!X->on_cpu)
1839 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1841 * LOCK rq(0)->lock LOCK X->pi_lock
1844 * smp_store_release(X->on_cpu, 0);
1846 * smp_cond_acquire(!X->on_cpu);
1852 * X->state = RUNNING
1853 * UNLOCK rq(2)->lock
1855 * LOCK rq(2)->lock // orders against CPU1
1858 * UNLOCK rq(2)->lock
1861 * UNLOCK rq(0)->lock
1864 * However; for wakeups there is a second guarantee we must provide, namely we
1865 * must observe the state that lead to our wakeup. That is, not only must our
1866 * task observe its own prior state, it must also observe the stores prior to
1869 * This means that any means of doing remote wakeups must order the CPU doing
1870 * the wakeup against the CPU the task is going to end up running on. This,
1871 * however, is already required for the regular Program-Order guarantee above,
1872 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1877 * try_to_wake_up - wake up a thread
1878 * @p: the thread to be awakened
1879 * @state: the mask of task states that can be woken
1880 * @wake_flags: wake modifier flags (WF_*)
1882 * Put it on the run-queue if it's not already there. The "current"
1883 * thread is always on the run-queue (except when the actual
1884 * re-schedule is in progress), and as such you're allowed to do
1885 * the simpler "current->state = TASK_RUNNING" to mark yourself
1886 * runnable without the overhead of this.
1888 * Return: %true if @p was woken up, %false if it was already running.
1889 * or @state didn't match @p's state.
1892 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1894 unsigned long flags;
1895 int cpu, success = 0;
1898 * If we are going to wake up a thread waiting for CONDITION we
1899 * need to ensure that CONDITION=1 done by the caller can not be
1900 * reordered with p->state check below. This pairs with mb() in
1901 * set_current_state() the waiting thread does.
1903 smp_mb__before_spinlock();
1904 raw_spin_lock_irqsave(&p->pi_lock, flags);
1905 if (!(p->state & state))
1908 trace_sched_waking(p);
1910 success = 1; /* we're going to change ->state */
1913 if (p->on_rq && ttwu_remote(p, wake_flags))
1918 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1919 * possible to, falsely, observe p->on_cpu == 0.
1921 * One must be running (->on_cpu == 1) in order to remove oneself
1922 * from the runqueue.
1924 * [S] ->on_cpu = 1; [L] ->on_rq
1928 * [S] ->on_rq = 0; [L] ->on_cpu
1930 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1931 * from the consecutive calls to schedule(); the first switching to our
1932 * task, the second putting it to sleep.
1937 * If the owning (remote) cpu is still in the middle of schedule() with
1938 * this task as prev, wait until its done referencing the task.
1940 * Pairs with the smp_store_release() in finish_lock_switch().
1942 * This ensures that tasks getting woken will be fully ordered against
1943 * their previous state and preserve Program Order.
1945 smp_cond_acquire(!p->on_cpu);
1947 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1948 p->state = TASK_WAKING;
1950 if (p->sched_class->task_waking)
1951 p->sched_class->task_waking(p);
1953 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1954 if (task_cpu(p) != cpu) {
1955 wake_flags |= WF_MIGRATED;
1956 set_task_cpu(p, cpu);
1958 #endif /* CONFIG_SMP */
1962 if (schedstat_enabled())
1963 ttwu_stat(p, cpu, wake_flags);
1965 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1971 * try_to_wake_up_local - try to wake up a local task with rq lock held
1972 * @p: the thread to be awakened
1974 * Put @p on the run-queue if it's not already there. The caller must
1975 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1978 static void try_to_wake_up_local(struct task_struct *p)
1980 struct rq *rq = task_rq(p);
1982 if (WARN_ON_ONCE(rq != this_rq()) ||
1983 WARN_ON_ONCE(p == current))
1986 lockdep_assert_held(&rq->lock);
1988 if (!raw_spin_trylock(&p->pi_lock)) {
1990 * This is OK, because current is on_cpu, which avoids it being
1991 * picked for load-balance and preemption/IRQs are still
1992 * disabled avoiding further scheduler activity on it and we've
1993 * not yet picked a replacement task.
1995 lockdep_unpin_lock(&rq->lock);
1996 raw_spin_unlock(&rq->lock);
1997 raw_spin_lock(&p->pi_lock);
1998 raw_spin_lock(&rq->lock);
1999 lockdep_pin_lock(&rq->lock);
2002 if (!(p->state & TASK_NORMAL))
2005 trace_sched_waking(p);
2007 if (!task_on_rq_queued(p))
2008 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2010 ttwu_do_wakeup(rq, p, 0);
2011 if (schedstat_enabled())
2012 ttwu_stat(p, smp_processor_id(), 0);
2014 raw_spin_unlock(&p->pi_lock);
2018 * wake_up_process - Wake up a specific process
2019 * @p: The process to be woken up.
2021 * Attempt to wake up the nominated process and move it to the set of runnable
2024 * Return: 1 if the process was woken up, 0 if it was already running.
2026 * It may be assumed that this function implies a write memory barrier before
2027 * changing the task state if and only if any tasks are woken up.
2029 int wake_up_process(struct task_struct *p)
2031 return try_to_wake_up(p, TASK_NORMAL, 0);
2033 EXPORT_SYMBOL(wake_up_process);
2035 int wake_up_state(struct task_struct *p, unsigned int state)
2037 return try_to_wake_up(p, state, 0);
2041 * This function clears the sched_dl_entity static params.
2043 void __dl_clear_params(struct task_struct *p)
2045 struct sched_dl_entity *dl_se = &p->dl;
2047 dl_se->dl_runtime = 0;
2048 dl_se->dl_deadline = 0;
2049 dl_se->dl_period = 0;
2053 dl_se->dl_throttled = 0;
2054 dl_se->dl_yielded = 0;
2058 * Perform scheduler related setup for a newly forked process p.
2059 * p is forked by current.
2061 * __sched_fork() is basic setup used by init_idle() too:
2063 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2068 p->se.exec_start = 0;
2069 p->se.sum_exec_runtime = 0;
2070 p->se.prev_sum_exec_runtime = 0;
2071 p->se.nr_migrations = 0;
2073 INIT_LIST_HEAD(&p->se.group_node);
2075 #ifdef CONFIG_FAIR_GROUP_SCHED
2076 p->se.cfs_rq = NULL;
2079 #ifdef CONFIG_SCHEDSTATS
2080 /* Even if schedstat is disabled, there should not be garbage */
2081 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2084 RB_CLEAR_NODE(&p->dl.rb_node);
2085 init_dl_task_timer(&p->dl);
2086 __dl_clear_params(p);
2088 INIT_LIST_HEAD(&p->rt.run_list);
2090 p->rt.time_slice = sched_rr_timeslice;
2094 #ifdef CONFIG_PREEMPT_NOTIFIERS
2095 INIT_HLIST_HEAD(&p->preempt_notifiers);
2098 #ifdef CONFIG_NUMA_BALANCING
2099 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2100 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2101 p->mm->numa_scan_seq = 0;
2104 if (clone_flags & CLONE_VM)
2105 p->numa_preferred_nid = current->numa_preferred_nid;
2107 p->numa_preferred_nid = -1;
2109 p->node_stamp = 0ULL;
2110 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2111 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2112 p->numa_work.next = &p->numa_work;
2113 p->numa_faults = NULL;
2114 p->last_task_numa_placement = 0;
2115 p->last_sum_exec_runtime = 0;
2117 p->numa_group = NULL;
2118 #endif /* CONFIG_NUMA_BALANCING */
2121 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2123 #ifdef CONFIG_NUMA_BALANCING
2125 void set_numabalancing_state(bool enabled)
2128 static_branch_enable(&sched_numa_balancing);
2130 static_branch_disable(&sched_numa_balancing);
2133 #ifdef CONFIG_PROC_SYSCTL
2134 int sysctl_numa_balancing(struct ctl_table *table, int write,
2135 void __user *buffer, size_t *lenp, loff_t *ppos)
2139 int state = static_branch_likely(&sched_numa_balancing);
2141 if (write && !capable(CAP_SYS_ADMIN))
2146 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2150 set_numabalancing_state(state);
2156 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2158 #ifdef CONFIG_SCHEDSTATS
2159 static void set_schedstats(bool enabled)
2162 static_branch_enable(&sched_schedstats);
2164 static_branch_disable(&sched_schedstats);
2167 void force_schedstat_enabled(void)
2169 if (!schedstat_enabled()) {
2170 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2171 static_branch_enable(&sched_schedstats);
2175 static int __init setup_schedstats(char *str)
2181 if (!strcmp(str, "enable")) {
2182 set_schedstats(true);
2184 } else if (!strcmp(str, "disable")) {
2185 set_schedstats(false);
2190 pr_warn("Unable to parse schedstats=\n");
2194 __setup("schedstats=", setup_schedstats);
2196 #ifdef CONFIG_PROC_SYSCTL
2197 int sysctl_schedstats(struct ctl_table *table, int write,
2198 void __user *buffer, size_t *lenp, loff_t *ppos)
2202 int state = static_branch_likely(&sched_schedstats);
2204 if (write && !capable(CAP_SYS_ADMIN))
2209 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2213 set_schedstats(state);
2220 * fork()/clone()-time setup:
2222 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2224 unsigned long flags;
2225 int cpu = get_cpu();
2227 __sched_fork(clone_flags, p);
2229 * We mark the process as running here. This guarantees that
2230 * nobody will actually run it, and a signal or other external
2231 * event cannot wake it up and insert it on the runqueue either.
2233 p->state = TASK_RUNNING;
2236 * Make sure we do not leak PI boosting priority to the child.
2238 p->prio = current->normal_prio;
2241 * Revert to default priority/policy on fork if requested.
2243 if (unlikely(p->sched_reset_on_fork)) {
2244 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2245 p->policy = SCHED_NORMAL;
2246 p->static_prio = NICE_TO_PRIO(0);
2248 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2249 p->static_prio = NICE_TO_PRIO(0);
2251 p->prio = p->normal_prio = __normal_prio(p);
2255 * We don't need the reset flag anymore after the fork. It has
2256 * fulfilled its duty:
2258 p->sched_reset_on_fork = 0;
2261 if (dl_prio(p->prio)) {
2264 } else if (rt_prio(p->prio)) {
2265 p->sched_class = &rt_sched_class;
2267 p->sched_class = &fair_sched_class;
2270 if (p->sched_class->task_fork)
2271 p->sched_class->task_fork(p);
2274 * The child is not yet in the pid-hash so no cgroup attach races,
2275 * and the cgroup is pinned to this child due to cgroup_fork()
2276 * is ran before sched_fork().
2278 * Silence PROVE_RCU.
2280 raw_spin_lock_irqsave(&p->pi_lock, flags);
2281 set_task_cpu(p, cpu);
2282 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2284 #ifdef CONFIG_SCHED_INFO
2285 if (likely(sched_info_on()))
2286 memset(&p->sched_info, 0, sizeof(p->sched_info));
2288 #if defined(CONFIG_SMP)
2291 init_task_preempt_count(p);
2293 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2294 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2301 unsigned long to_ratio(u64 period, u64 runtime)
2303 if (runtime == RUNTIME_INF)
2307 * Doing this here saves a lot of checks in all
2308 * the calling paths, and returning zero seems
2309 * safe for them anyway.
2314 return div64_u64(runtime << 20, period);
2318 inline struct dl_bw *dl_bw_of(int i)
2320 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2321 "sched RCU must be held");
2322 return &cpu_rq(i)->rd->dl_bw;
2325 static inline int dl_bw_cpus(int i)
2327 struct root_domain *rd = cpu_rq(i)->rd;
2330 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2331 "sched RCU must be held");
2332 for_each_cpu_and(i, rd->span, cpu_active_mask)
2338 inline struct dl_bw *dl_bw_of(int i)
2340 return &cpu_rq(i)->dl.dl_bw;
2343 static inline int dl_bw_cpus(int i)
2350 * We must be sure that accepting a new task (or allowing changing the
2351 * parameters of an existing one) is consistent with the bandwidth
2352 * constraints. If yes, this function also accordingly updates the currently
2353 * allocated bandwidth to reflect the new situation.
2355 * This function is called while holding p's rq->lock.
2357 * XXX we should delay bw change until the task's 0-lag point, see
2360 static int dl_overflow(struct task_struct *p, int policy,
2361 const struct sched_attr *attr)
2364 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2365 u64 period = attr->sched_period ?: attr->sched_deadline;
2366 u64 runtime = attr->sched_runtime;
2367 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2370 if (new_bw == p->dl.dl_bw)
2374 * Either if a task, enters, leave, or stays -deadline but changes
2375 * its parameters, we may need to update accordingly the total
2376 * allocated bandwidth of the container.
2378 raw_spin_lock(&dl_b->lock);
2379 cpus = dl_bw_cpus(task_cpu(p));
2380 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2381 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2382 __dl_add(dl_b, new_bw);
2384 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2385 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2386 __dl_clear(dl_b, p->dl.dl_bw);
2387 __dl_add(dl_b, new_bw);
2389 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2390 __dl_clear(dl_b, p->dl.dl_bw);
2393 raw_spin_unlock(&dl_b->lock);
2398 extern void init_dl_bw(struct dl_bw *dl_b);
2401 * wake_up_new_task - wake up a newly created task for the first time.
2403 * This function will do some initial scheduler statistics housekeeping
2404 * that must be done for every newly created context, then puts the task
2405 * on the runqueue and wakes it.
2407 void wake_up_new_task(struct task_struct *p)
2409 unsigned long flags;
2412 raw_spin_lock_irqsave(&p->pi_lock, flags);
2413 /* Initialize new task's runnable average */
2414 init_entity_runnable_average(&p->se);
2417 * Fork balancing, do it here and not earlier because:
2418 * - cpus_allowed can change in the fork path
2419 * - any previously selected cpu might disappear through hotplug
2421 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2424 rq = __task_rq_lock(p);
2425 activate_task(rq, p, 0);
2426 p->on_rq = TASK_ON_RQ_QUEUED;
2427 trace_sched_wakeup_new(p);
2428 check_preempt_curr(rq, p, WF_FORK);
2430 if (p->sched_class->task_woken) {
2432 * Nothing relies on rq->lock after this, so its fine to
2435 lockdep_unpin_lock(&rq->lock);
2436 p->sched_class->task_woken(rq, p);
2437 lockdep_pin_lock(&rq->lock);
2440 task_rq_unlock(rq, p, &flags);
2443 #ifdef CONFIG_PREEMPT_NOTIFIERS
2445 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2447 void preempt_notifier_inc(void)
2449 static_key_slow_inc(&preempt_notifier_key);
2451 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2453 void preempt_notifier_dec(void)
2455 static_key_slow_dec(&preempt_notifier_key);
2457 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2460 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2461 * @notifier: notifier struct to register
2463 void preempt_notifier_register(struct preempt_notifier *notifier)
2465 if (!static_key_false(&preempt_notifier_key))
2466 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2468 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2470 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2473 * preempt_notifier_unregister - no longer interested in preemption notifications
2474 * @notifier: notifier struct to unregister
2476 * This is *not* safe to call from within a preemption notifier.
2478 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2480 hlist_del(¬ifier->link);
2482 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2484 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2486 struct preempt_notifier *notifier;
2488 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2489 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2492 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2494 if (static_key_false(&preempt_notifier_key))
2495 __fire_sched_in_preempt_notifiers(curr);
2499 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2500 struct task_struct *next)
2502 struct preempt_notifier *notifier;
2504 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2505 notifier->ops->sched_out(notifier, next);
2508 static __always_inline void
2509 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2510 struct task_struct *next)
2512 if (static_key_false(&preempt_notifier_key))
2513 __fire_sched_out_preempt_notifiers(curr, next);
2516 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2518 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2523 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2524 struct task_struct *next)
2528 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2531 * prepare_task_switch - prepare to switch tasks
2532 * @rq: the runqueue preparing to switch
2533 * @prev: the current task that is being switched out
2534 * @next: the task we are going to switch to.
2536 * This is called with the rq lock held and interrupts off. It must
2537 * be paired with a subsequent finish_task_switch after the context
2540 * prepare_task_switch sets up locking and calls architecture specific
2544 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2545 struct task_struct *next)
2547 sched_info_switch(rq, prev, next);
2548 perf_event_task_sched_out(prev, next);
2549 fire_sched_out_preempt_notifiers(prev, next);
2550 prepare_lock_switch(rq, next);
2551 prepare_arch_switch(next);
2555 * finish_task_switch - clean up after a task-switch
2556 * @prev: the thread we just switched away from.
2558 * finish_task_switch must be called after the context switch, paired
2559 * with a prepare_task_switch call before the context switch.
2560 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2561 * and do any other architecture-specific cleanup actions.
2563 * Note that we may have delayed dropping an mm in context_switch(). If
2564 * so, we finish that here outside of the runqueue lock. (Doing it
2565 * with the lock held can cause deadlocks; see schedule() for
2568 * The context switch have flipped the stack from under us and restored the
2569 * local variables which were saved when this task called schedule() in the
2570 * past. prev == current is still correct but we need to recalculate this_rq
2571 * because prev may have moved to another CPU.
2573 static struct rq *finish_task_switch(struct task_struct *prev)
2574 __releases(rq->lock)
2576 struct rq *rq = this_rq();
2577 struct mm_struct *mm = rq->prev_mm;
2581 * The previous task will have left us with a preempt_count of 2
2582 * because it left us after:
2585 * preempt_disable(); // 1
2587 * raw_spin_lock_irq(&rq->lock) // 2
2589 * Also, see FORK_PREEMPT_COUNT.
2591 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2592 "corrupted preempt_count: %s/%d/0x%x\n",
2593 current->comm, current->pid, preempt_count()))
2594 preempt_count_set(FORK_PREEMPT_COUNT);
2599 * A task struct has one reference for the use as "current".
2600 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2601 * schedule one last time. The schedule call will never return, and
2602 * the scheduled task must drop that reference.
2604 * We must observe prev->state before clearing prev->on_cpu (in
2605 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2606 * running on another CPU and we could rave with its RUNNING -> DEAD
2607 * transition, resulting in a double drop.
2609 prev_state = prev->state;
2610 vtime_task_switch(prev);
2611 perf_event_task_sched_in(prev, current);
2612 finish_lock_switch(rq, prev);
2613 finish_arch_post_lock_switch();
2615 fire_sched_in_preempt_notifiers(current);
2618 if (unlikely(prev_state == TASK_DEAD)) {
2619 if (prev->sched_class->task_dead)
2620 prev->sched_class->task_dead(prev);
2623 * Remove function-return probe instances associated with this
2624 * task and put them back on the free list.
2626 kprobe_flush_task(prev);
2627 put_task_struct(prev);
2630 tick_nohz_task_switch();
2636 /* rq->lock is NOT held, but preemption is disabled */
2637 static void __balance_callback(struct rq *rq)
2639 struct callback_head *head, *next;
2640 void (*func)(struct rq *rq);
2641 unsigned long flags;
2643 raw_spin_lock_irqsave(&rq->lock, flags);
2644 head = rq->balance_callback;
2645 rq->balance_callback = NULL;
2647 func = (void (*)(struct rq *))head->func;
2654 raw_spin_unlock_irqrestore(&rq->lock, flags);
2657 static inline void balance_callback(struct rq *rq)
2659 if (unlikely(rq->balance_callback))
2660 __balance_callback(rq);
2665 static inline void balance_callback(struct rq *rq)
2672 * schedule_tail - first thing a freshly forked thread must call.
2673 * @prev: the thread we just switched away from.
2675 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2676 __releases(rq->lock)
2681 * New tasks start with FORK_PREEMPT_COUNT, see there and
2682 * finish_task_switch() for details.
2684 * finish_task_switch() will drop rq->lock() and lower preempt_count
2685 * and the preempt_enable() will end up enabling preemption (on
2686 * PREEMPT_COUNT kernels).
2689 rq = finish_task_switch(prev);
2690 balance_callback(rq);
2693 if (current->set_child_tid)
2694 put_user(task_pid_vnr(current), current->set_child_tid);
2698 * context_switch - switch to the new MM and the new thread's register state.
2700 static inline struct rq *
2701 context_switch(struct rq *rq, struct task_struct *prev,
2702 struct task_struct *next)
2704 struct mm_struct *mm, *oldmm;
2706 prepare_task_switch(rq, prev, next);
2709 oldmm = prev->active_mm;
2711 * For paravirt, this is coupled with an exit in switch_to to
2712 * combine the page table reload and the switch backend into
2715 arch_start_context_switch(prev);
2718 next->active_mm = oldmm;
2719 atomic_inc(&oldmm->mm_count);
2720 enter_lazy_tlb(oldmm, next);
2722 switch_mm(oldmm, mm, next);
2725 prev->active_mm = NULL;
2726 rq->prev_mm = oldmm;
2729 * Since the runqueue lock will be released by the next
2730 * task (which is an invalid locking op but in the case
2731 * of the scheduler it's an obvious special-case), so we
2732 * do an early lockdep release here:
2734 lockdep_unpin_lock(&rq->lock);
2735 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2737 /* Here we just switch the register state and the stack. */
2738 switch_to(prev, next, prev);
2741 return finish_task_switch(prev);
2745 * nr_running and nr_context_switches:
2747 * externally visible scheduler statistics: current number of runnable
2748 * threads, total number of context switches performed since bootup.
2750 unsigned long nr_running(void)
2752 unsigned long i, sum = 0;
2754 for_each_online_cpu(i)
2755 sum += cpu_rq(i)->nr_running;
2761 * Check if only the current task is running on the cpu.
2763 * Caution: this function does not check that the caller has disabled
2764 * preemption, thus the result might have a time-of-check-to-time-of-use
2765 * race. The caller is responsible to use it correctly, for example:
2767 * - from a non-preemptable section (of course)
2769 * - from a thread that is bound to a single CPU
2771 * - in a loop with very short iterations (e.g. a polling loop)
2773 bool single_task_running(void)
2775 return raw_rq()->nr_running == 1;
2777 EXPORT_SYMBOL(single_task_running);
2779 unsigned long long nr_context_switches(void)
2782 unsigned long long sum = 0;
2784 for_each_possible_cpu(i)
2785 sum += cpu_rq(i)->nr_switches;
2790 unsigned long nr_iowait(void)
2792 unsigned long i, sum = 0;
2794 for_each_possible_cpu(i)
2795 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2800 unsigned long nr_iowait_cpu(int cpu)
2802 struct rq *this = cpu_rq(cpu);
2803 return atomic_read(&this->nr_iowait);
2806 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2808 struct rq *rq = this_rq();
2809 *nr_waiters = atomic_read(&rq->nr_iowait);
2810 *load = rq->load.weight;
2816 * sched_exec - execve() is a valuable balancing opportunity, because at
2817 * this point the task has the smallest effective memory and cache footprint.
2819 void sched_exec(void)
2821 struct task_struct *p = current;
2822 unsigned long flags;
2825 raw_spin_lock_irqsave(&p->pi_lock, flags);
2826 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2827 if (dest_cpu == smp_processor_id())
2830 if (likely(cpu_active(dest_cpu))) {
2831 struct migration_arg arg = { p, dest_cpu };
2833 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2834 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2838 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2843 DEFINE_PER_CPU(struct kernel_stat, kstat);
2844 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2846 EXPORT_PER_CPU_SYMBOL(kstat);
2847 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2850 * Return accounted runtime for the task.
2851 * In case the task is currently running, return the runtime plus current's
2852 * pending runtime that have not been accounted yet.
2854 unsigned long long task_sched_runtime(struct task_struct *p)
2856 unsigned long flags;
2860 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2862 * 64-bit doesn't need locks to atomically read a 64bit value.
2863 * So we have a optimization chance when the task's delta_exec is 0.
2864 * Reading ->on_cpu is racy, but this is ok.
2866 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2867 * If we race with it entering cpu, unaccounted time is 0. This is
2868 * indistinguishable from the read occurring a few cycles earlier.
2869 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2870 * been accounted, so we're correct here as well.
2872 if (!p->on_cpu || !task_on_rq_queued(p))
2873 return p->se.sum_exec_runtime;
2876 rq = task_rq_lock(p, &flags);
2878 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2879 * project cycles that may never be accounted to this
2880 * thread, breaking clock_gettime().
2882 if (task_current(rq, p) && task_on_rq_queued(p)) {
2883 update_rq_clock(rq);
2884 p->sched_class->update_curr(rq);
2886 ns = p->se.sum_exec_runtime;
2887 task_rq_unlock(rq, p, &flags);
2893 * This function gets called by the timer code, with HZ frequency.
2894 * We call it with interrupts disabled.
2896 void scheduler_tick(void)
2898 int cpu = smp_processor_id();
2899 struct rq *rq = cpu_rq(cpu);
2900 struct task_struct *curr = rq->curr;
2904 raw_spin_lock(&rq->lock);
2905 update_rq_clock(rq);
2906 curr->sched_class->task_tick(rq, curr, 0);
2907 update_cpu_load_active(rq);
2908 calc_global_load_tick(rq);
2909 raw_spin_unlock(&rq->lock);
2911 perf_event_task_tick();
2914 rq->idle_balance = idle_cpu(cpu);
2915 trigger_load_balance(rq);
2917 rq_last_tick_reset(rq);
2920 #ifdef CONFIG_NO_HZ_FULL
2922 * scheduler_tick_max_deferment
2924 * Keep at least one tick per second when a single
2925 * active task is running because the scheduler doesn't
2926 * yet completely support full dynticks environment.
2928 * This makes sure that uptime, CFS vruntime, load
2929 * balancing, etc... continue to move forward, even
2930 * with a very low granularity.
2932 * Return: Maximum deferment in nanoseconds.
2934 u64 scheduler_tick_max_deferment(void)
2936 struct rq *rq = this_rq();
2937 unsigned long next, now = READ_ONCE(jiffies);
2939 next = rq->last_sched_tick + HZ;
2941 if (time_before_eq(next, now))
2944 return jiffies_to_nsecs(next - now);
2948 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2949 defined(CONFIG_PREEMPT_TRACER))
2951 void preempt_count_add(int val)
2953 #ifdef CONFIG_DEBUG_PREEMPT
2957 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2960 __preempt_count_add(val);
2961 #ifdef CONFIG_DEBUG_PREEMPT
2963 * Spinlock count overflowing soon?
2965 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2968 if (preempt_count() == val) {
2969 unsigned long ip = get_lock_parent_ip();
2970 #ifdef CONFIG_DEBUG_PREEMPT
2971 current->preempt_disable_ip = ip;
2973 trace_preempt_off(CALLER_ADDR0, ip);
2976 EXPORT_SYMBOL(preempt_count_add);
2977 NOKPROBE_SYMBOL(preempt_count_add);
2979 void preempt_count_sub(int val)
2981 #ifdef CONFIG_DEBUG_PREEMPT
2985 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2988 * Is the spinlock portion underflowing?
2990 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2991 !(preempt_count() & PREEMPT_MASK)))
2995 if (preempt_count() == val)
2996 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
2997 __preempt_count_sub(val);
2999 EXPORT_SYMBOL(preempt_count_sub);
3000 NOKPROBE_SYMBOL(preempt_count_sub);
3005 * Print scheduling while atomic bug:
3007 static noinline void __schedule_bug(struct task_struct *prev)
3009 if (oops_in_progress)
3012 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3013 prev->comm, prev->pid, preempt_count());
3015 debug_show_held_locks(prev);
3017 if (irqs_disabled())
3018 print_irqtrace_events(prev);
3019 #ifdef CONFIG_DEBUG_PREEMPT
3020 if (in_atomic_preempt_off()) {
3021 pr_err("Preemption disabled at:");
3022 print_ip_sym(current->preempt_disable_ip);
3027 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3031 * Various schedule()-time debugging checks and statistics:
3033 static inline void schedule_debug(struct task_struct *prev)
3035 #ifdef CONFIG_SCHED_STACK_END_CHECK
3036 BUG_ON(task_stack_end_corrupted(prev));
3039 if (unlikely(in_atomic_preempt_off())) {
3040 __schedule_bug(prev);
3041 preempt_count_set(PREEMPT_DISABLED);
3045 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3047 schedstat_inc(this_rq(), sched_count);
3051 * Pick up the highest-prio task:
3053 static inline struct task_struct *
3054 pick_next_task(struct rq *rq, struct task_struct *prev)
3056 const struct sched_class *class = &fair_sched_class;
3057 struct task_struct *p;
3060 * Optimization: we know that if all tasks are in
3061 * the fair class we can call that function directly:
3063 if (likely(prev->sched_class == class &&
3064 rq->nr_running == rq->cfs.h_nr_running)) {
3065 p = fair_sched_class.pick_next_task(rq, prev);
3066 if (unlikely(p == RETRY_TASK))
3069 /* assumes fair_sched_class->next == idle_sched_class */
3071 p = idle_sched_class.pick_next_task(rq, prev);
3077 for_each_class(class) {
3078 p = class->pick_next_task(rq, prev);
3080 if (unlikely(p == RETRY_TASK))
3086 BUG(); /* the idle class will always have a runnable task */
3090 * __schedule() is the main scheduler function.
3092 * The main means of driving the scheduler and thus entering this function are:
3094 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3096 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3097 * paths. For example, see arch/x86/entry_64.S.
3099 * To drive preemption between tasks, the scheduler sets the flag in timer
3100 * interrupt handler scheduler_tick().
3102 * 3. Wakeups don't really cause entry into schedule(). They add a
3103 * task to the run-queue and that's it.
3105 * Now, if the new task added to the run-queue preempts the current
3106 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3107 * called on the nearest possible occasion:
3109 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3111 * - in syscall or exception context, at the next outmost
3112 * preempt_enable(). (this might be as soon as the wake_up()'s
3115 * - in IRQ context, return from interrupt-handler to
3116 * preemptible context
3118 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3121 * - cond_resched() call
3122 * - explicit schedule() call
3123 * - return from syscall or exception to user-space
3124 * - return from interrupt-handler to user-space
3126 * WARNING: must be called with preemption disabled!
3128 static void __sched notrace __schedule(bool preempt)
3130 struct task_struct *prev, *next;
3131 unsigned long *switch_count;
3135 cpu = smp_processor_id();
3140 * do_exit() calls schedule() with preemption disabled as an exception;
3141 * however we must fix that up, otherwise the next task will see an
3142 * inconsistent (higher) preempt count.
3144 * It also avoids the below schedule_debug() test from complaining
3147 if (unlikely(prev->state == TASK_DEAD))
3148 preempt_enable_no_resched_notrace();
3150 schedule_debug(prev);
3152 if (sched_feat(HRTICK))
3155 local_irq_disable();
3156 rcu_note_context_switch();
3159 * Make sure that signal_pending_state()->signal_pending() below
3160 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3161 * done by the caller to avoid the race with signal_wake_up().
3163 smp_mb__before_spinlock();
3164 raw_spin_lock(&rq->lock);
3165 lockdep_pin_lock(&rq->lock);
3167 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3169 switch_count = &prev->nivcsw;
3170 if (!preempt && prev->state) {
3171 if (unlikely(signal_pending_state(prev->state, prev))) {
3172 prev->state = TASK_RUNNING;
3174 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3178 * If a worker went to sleep, notify and ask workqueue
3179 * whether it wants to wake up a task to maintain
3182 if (prev->flags & PF_WQ_WORKER) {
3183 struct task_struct *to_wakeup;
3185 to_wakeup = wq_worker_sleeping(prev, cpu);
3187 try_to_wake_up_local(to_wakeup);
3190 switch_count = &prev->nvcsw;
3193 if (task_on_rq_queued(prev))
3194 update_rq_clock(rq);
3196 next = pick_next_task(rq, prev);
3197 clear_tsk_need_resched(prev);
3198 clear_preempt_need_resched();
3199 rq->clock_skip_update = 0;
3201 if (likely(prev != next)) {
3206 trace_sched_switch(preempt, prev, next);
3207 rq = context_switch(rq, prev, next); /* unlocks the rq */
3209 lockdep_unpin_lock(&rq->lock);
3210 raw_spin_unlock_irq(&rq->lock);
3213 balance_callback(rq);
3216 static inline void sched_submit_work(struct task_struct *tsk)
3218 if (!tsk->state || tsk_is_pi_blocked(tsk))
3221 * If we are going to sleep and we have plugged IO queued,
3222 * make sure to submit it to avoid deadlocks.
3224 if (blk_needs_flush_plug(tsk))
3225 blk_schedule_flush_plug(tsk);
3228 asmlinkage __visible void __sched schedule(void)
3230 struct task_struct *tsk = current;
3232 sched_submit_work(tsk);
3236 sched_preempt_enable_no_resched();
3237 } while (need_resched());
3239 EXPORT_SYMBOL(schedule);
3241 #ifdef CONFIG_CONTEXT_TRACKING
3242 asmlinkage __visible void __sched schedule_user(void)
3245 * If we come here after a random call to set_need_resched(),
3246 * or we have been woken up remotely but the IPI has not yet arrived,
3247 * we haven't yet exited the RCU idle mode. Do it here manually until
3248 * we find a better solution.
3250 * NB: There are buggy callers of this function. Ideally we
3251 * should warn if prev_state != CONTEXT_USER, but that will trigger
3252 * too frequently to make sense yet.
3254 enum ctx_state prev_state = exception_enter();
3256 exception_exit(prev_state);
3261 * schedule_preempt_disabled - called with preemption disabled
3263 * Returns with preemption disabled. Note: preempt_count must be 1
3265 void __sched schedule_preempt_disabled(void)
3267 sched_preempt_enable_no_resched();
3272 static void __sched notrace preempt_schedule_common(void)
3275 preempt_disable_notrace();
3277 preempt_enable_no_resched_notrace();
3280 * Check again in case we missed a preemption opportunity
3281 * between schedule and now.
3283 } while (need_resched());
3286 #ifdef CONFIG_PREEMPT
3288 * this is the entry point to schedule() from in-kernel preemption
3289 * off of preempt_enable. Kernel preemptions off return from interrupt
3290 * occur there and call schedule directly.
3292 asmlinkage __visible void __sched notrace preempt_schedule(void)
3295 * If there is a non-zero preempt_count or interrupts are disabled,
3296 * we do not want to preempt the current task. Just return..
3298 if (likely(!preemptible()))
3301 preempt_schedule_common();
3303 NOKPROBE_SYMBOL(preempt_schedule);
3304 EXPORT_SYMBOL(preempt_schedule);
3307 * preempt_schedule_notrace - preempt_schedule called by tracing
3309 * The tracing infrastructure uses preempt_enable_notrace to prevent
3310 * recursion and tracing preempt enabling caused by the tracing
3311 * infrastructure itself. But as tracing can happen in areas coming
3312 * from userspace or just about to enter userspace, a preempt enable
3313 * can occur before user_exit() is called. This will cause the scheduler
3314 * to be called when the system is still in usermode.
3316 * To prevent this, the preempt_enable_notrace will use this function
3317 * instead of preempt_schedule() to exit user context if needed before
3318 * calling the scheduler.
3320 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3322 enum ctx_state prev_ctx;
3324 if (likely(!preemptible()))
3328 preempt_disable_notrace();
3330 * Needs preempt disabled in case user_exit() is traced
3331 * and the tracer calls preempt_enable_notrace() causing
3332 * an infinite recursion.
3334 prev_ctx = exception_enter();
3336 exception_exit(prev_ctx);
3338 preempt_enable_no_resched_notrace();
3339 } while (need_resched());
3341 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3343 #endif /* CONFIG_PREEMPT */
3346 * this is the entry point to schedule() from kernel preemption
3347 * off of irq context.
3348 * Note, that this is called and return with irqs disabled. This will
3349 * protect us against recursive calling from irq.
3351 asmlinkage __visible void __sched preempt_schedule_irq(void)
3353 enum ctx_state prev_state;
3355 /* Catch callers which need to be fixed */
3356 BUG_ON(preempt_count() || !irqs_disabled());
3358 prev_state = exception_enter();
3364 local_irq_disable();
3365 sched_preempt_enable_no_resched();
3366 } while (need_resched());
3368 exception_exit(prev_state);
3371 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3374 return try_to_wake_up(curr->private, mode, wake_flags);
3376 EXPORT_SYMBOL(default_wake_function);
3378 #ifdef CONFIG_RT_MUTEXES
3381 * rt_mutex_setprio - set the current priority of a task
3383 * @prio: prio value (kernel-internal form)
3385 * This function changes the 'effective' priority of a task. It does
3386 * not touch ->normal_prio like __setscheduler().
3388 * Used by the rt_mutex code to implement priority inheritance
3389 * logic. Call site only calls if the priority of the task changed.
3391 void rt_mutex_setprio(struct task_struct *p, int prio)
3393 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3395 const struct sched_class *prev_class;
3397 BUG_ON(prio > MAX_PRIO);
3399 rq = __task_rq_lock(p);
3402 * Idle task boosting is a nono in general. There is one
3403 * exception, when PREEMPT_RT and NOHZ is active:
3405 * The idle task calls get_next_timer_interrupt() and holds
3406 * the timer wheel base->lock on the CPU and another CPU wants
3407 * to access the timer (probably to cancel it). We can safely
3408 * ignore the boosting request, as the idle CPU runs this code
3409 * with interrupts disabled and will complete the lock
3410 * protected section without being interrupted. So there is no
3411 * real need to boost.
3413 if (unlikely(p == rq->idle)) {
3414 WARN_ON(p != rq->curr);
3415 WARN_ON(p->pi_blocked_on);
3419 trace_sched_pi_setprio(p, prio);
3422 if (oldprio == prio)
3423 queue_flag &= ~DEQUEUE_MOVE;
3425 prev_class = p->sched_class;
3426 queued = task_on_rq_queued(p);
3427 running = task_current(rq, p);
3429 dequeue_task(rq, p, queue_flag);
3431 put_prev_task(rq, p);
3434 * Boosting condition are:
3435 * 1. -rt task is running and holds mutex A
3436 * --> -dl task blocks on mutex A
3438 * 2. -dl task is running and holds mutex A
3439 * --> -dl task blocks on mutex A and could preempt the
3442 if (dl_prio(prio)) {
3443 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3444 if (!dl_prio(p->normal_prio) ||
3445 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3446 p->dl.dl_boosted = 1;
3447 queue_flag |= ENQUEUE_REPLENISH;
3449 p->dl.dl_boosted = 0;
3450 p->sched_class = &dl_sched_class;
3451 } else if (rt_prio(prio)) {
3452 if (dl_prio(oldprio))
3453 p->dl.dl_boosted = 0;
3455 queue_flag |= ENQUEUE_HEAD;
3456 p->sched_class = &rt_sched_class;
3458 if (dl_prio(oldprio))
3459 p->dl.dl_boosted = 0;
3460 if (rt_prio(oldprio))
3462 p->sched_class = &fair_sched_class;
3468 p->sched_class->set_curr_task(rq);
3470 enqueue_task(rq, p, queue_flag);
3472 check_class_changed(rq, p, prev_class, oldprio);
3474 preempt_disable(); /* avoid rq from going away on us */
3475 __task_rq_unlock(rq);
3477 balance_callback(rq);
3482 void set_user_nice(struct task_struct *p, long nice)
3484 int old_prio, delta, queued;
3485 unsigned long flags;
3488 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3491 * We have to be careful, if called from sys_setpriority(),
3492 * the task might be in the middle of scheduling on another CPU.
3494 rq = task_rq_lock(p, &flags);
3496 * The RT priorities are set via sched_setscheduler(), but we still
3497 * allow the 'normal' nice value to be set - but as expected
3498 * it wont have any effect on scheduling until the task is
3499 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3501 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3502 p->static_prio = NICE_TO_PRIO(nice);
3505 queued = task_on_rq_queued(p);
3507 dequeue_task(rq, p, DEQUEUE_SAVE);
3509 p->static_prio = NICE_TO_PRIO(nice);
3512 p->prio = effective_prio(p);
3513 delta = p->prio - old_prio;
3516 enqueue_task(rq, p, ENQUEUE_RESTORE);
3518 * If the task increased its priority or is running and
3519 * lowered its priority, then reschedule its CPU:
3521 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3525 task_rq_unlock(rq, p, &flags);
3527 EXPORT_SYMBOL(set_user_nice);
3530 * can_nice - check if a task can reduce its nice value
3534 int can_nice(const struct task_struct *p, const int nice)
3536 /* convert nice value [19,-20] to rlimit style value [1,40] */
3537 int nice_rlim = nice_to_rlimit(nice);
3539 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3540 capable(CAP_SYS_NICE));
3543 #ifdef __ARCH_WANT_SYS_NICE
3546 * sys_nice - change the priority of the current process.
3547 * @increment: priority increment
3549 * sys_setpriority is a more generic, but much slower function that
3550 * does similar things.
3552 SYSCALL_DEFINE1(nice, int, increment)
3557 * Setpriority might change our priority at the same moment.
3558 * We don't have to worry. Conceptually one call occurs first
3559 * and we have a single winner.
3561 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3562 nice = task_nice(current) + increment;
3564 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3565 if (increment < 0 && !can_nice(current, nice))
3568 retval = security_task_setnice(current, nice);
3572 set_user_nice(current, nice);
3579 * task_prio - return the priority value of a given task.
3580 * @p: the task in question.
3582 * Return: The priority value as seen by users in /proc.
3583 * RT tasks are offset by -200. Normal tasks are centered
3584 * around 0, value goes from -16 to +15.
3586 int task_prio(const struct task_struct *p)
3588 return p->prio - MAX_RT_PRIO;
3592 * idle_cpu - is a given cpu idle currently?
3593 * @cpu: the processor in question.
3595 * Return: 1 if the CPU is currently idle. 0 otherwise.
3597 int idle_cpu(int cpu)
3599 struct rq *rq = cpu_rq(cpu);
3601 if (rq->curr != rq->idle)
3608 if (!llist_empty(&rq->wake_list))
3616 * idle_task - return the idle task for a given cpu.
3617 * @cpu: the processor in question.
3619 * Return: The idle task for the cpu @cpu.
3621 struct task_struct *idle_task(int cpu)
3623 return cpu_rq(cpu)->idle;
3627 * find_process_by_pid - find a process with a matching PID value.
3628 * @pid: the pid in question.
3630 * The task of @pid, if found. %NULL otherwise.
3632 static struct task_struct *find_process_by_pid(pid_t pid)
3634 return pid ? find_task_by_vpid(pid) : current;
3638 * This function initializes the sched_dl_entity of a newly becoming
3639 * SCHED_DEADLINE task.
3641 * Only the static values are considered here, the actual runtime and the
3642 * absolute deadline will be properly calculated when the task is enqueued
3643 * for the first time with its new policy.
3646 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3648 struct sched_dl_entity *dl_se = &p->dl;
3650 dl_se->dl_runtime = attr->sched_runtime;
3651 dl_se->dl_deadline = attr->sched_deadline;
3652 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3653 dl_se->flags = attr->sched_flags;
3654 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3657 * Changing the parameters of a task is 'tricky' and we're not doing
3658 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3660 * What we SHOULD do is delay the bandwidth release until the 0-lag
3661 * point. This would include retaining the task_struct until that time
3662 * and change dl_overflow() to not immediately decrement the current
3665 * Instead we retain the current runtime/deadline and let the new
3666 * parameters take effect after the current reservation period lapses.
3667 * This is safe (albeit pessimistic) because the 0-lag point is always
3668 * before the current scheduling deadline.
3670 * We can still have temporary overloads because we do not delay the
3671 * change in bandwidth until that time; so admission control is
3672 * not on the safe side. It does however guarantee tasks will never
3673 * consume more than promised.
3678 * sched_setparam() passes in -1 for its policy, to let the functions
3679 * it calls know not to change it.
3681 #define SETPARAM_POLICY -1
3683 static void __setscheduler_params(struct task_struct *p,
3684 const struct sched_attr *attr)
3686 int policy = attr->sched_policy;
3688 if (policy == SETPARAM_POLICY)
3693 if (dl_policy(policy))
3694 __setparam_dl(p, attr);
3695 else if (fair_policy(policy))
3696 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3699 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3700 * !rt_policy. Always setting this ensures that things like
3701 * getparam()/getattr() don't report silly values for !rt tasks.
3703 p->rt_priority = attr->sched_priority;
3704 p->normal_prio = normal_prio(p);
3708 /* Actually do priority change: must hold pi & rq lock. */
3709 static void __setscheduler(struct rq *rq, struct task_struct *p,
3710 const struct sched_attr *attr, bool keep_boost)
3712 __setscheduler_params(p, attr);
3715 * Keep a potential priority boosting if called from
3716 * sched_setscheduler().
3719 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3721 p->prio = normal_prio(p);
3723 if (dl_prio(p->prio))
3724 p->sched_class = &dl_sched_class;
3725 else if (rt_prio(p->prio))
3726 p->sched_class = &rt_sched_class;
3728 p->sched_class = &fair_sched_class;
3732 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3734 struct sched_dl_entity *dl_se = &p->dl;
3736 attr->sched_priority = p->rt_priority;
3737 attr->sched_runtime = dl_se->dl_runtime;
3738 attr->sched_deadline = dl_se->dl_deadline;
3739 attr->sched_period = dl_se->dl_period;
3740 attr->sched_flags = dl_se->flags;
3744 * This function validates the new parameters of a -deadline task.
3745 * We ask for the deadline not being zero, and greater or equal
3746 * than the runtime, as well as the period of being zero or
3747 * greater than deadline. Furthermore, we have to be sure that
3748 * user parameters are above the internal resolution of 1us (we
3749 * check sched_runtime only since it is always the smaller one) and
3750 * below 2^63 ns (we have to check both sched_deadline and
3751 * sched_period, as the latter can be zero).
3754 __checkparam_dl(const struct sched_attr *attr)
3757 if (attr->sched_deadline == 0)
3761 * Since we truncate DL_SCALE bits, make sure we're at least
3764 if (attr->sched_runtime < (1ULL << DL_SCALE))
3768 * Since we use the MSB for wrap-around and sign issues, make
3769 * sure it's not set (mind that period can be equal to zero).
3771 if (attr->sched_deadline & (1ULL << 63) ||
3772 attr->sched_period & (1ULL << 63))
3775 /* runtime <= deadline <= period (if period != 0) */
3776 if ((attr->sched_period != 0 &&
3777 attr->sched_period < attr->sched_deadline) ||
3778 attr->sched_deadline < attr->sched_runtime)
3785 * check the target process has a UID that matches the current process's
3787 static bool check_same_owner(struct task_struct *p)
3789 const struct cred *cred = current_cred(), *pcred;
3793 pcred = __task_cred(p);
3794 match = (uid_eq(cred->euid, pcred->euid) ||
3795 uid_eq(cred->euid, pcred->uid));
3800 static bool dl_param_changed(struct task_struct *p,
3801 const struct sched_attr *attr)
3803 struct sched_dl_entity *dl_se = &p->dl;
3805 if (dl_se->dl_runtime != attr->sched_runtime ||
3806 dl_se->dl_deadline != attr->sched_deadline ||
3807 dl_se->dl_period != attr->sched_period ||
3808 dl_se->flags != attr->sched_flags)
3814 static int __sched_setscheduler(struct task_struct *p,
3815 const struct sched_attr *attr,
3818 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3819 MAX_RT_PRIO - 1 - attr->sched_priority;
3820 int retval, oldprio, oldpolicy = -1, queued, running;
3821 int new_effective_prio, policy = attr->sched_policy;
3822 unsigned long flags;
3823 const struct sched_class *prev_class;
3826 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
3828 /* may grab non-irq protected spin_locks */
3829 BUG_ON(in_interrupt());
3831 /* double check policy once rq lock held */
3833 reset_on_fork = p->sched_reset_on_fork;
3834 policy = oldpolicy = p->policy;
3836 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3838 if (!valid_policy(policy))
3842 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3846 * Valid priorities for SCHED_FIFO and SCHED_RR are
3847 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3848 * SCHED_BATCH and SCHED_IDLE is 0.
3850 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3851 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3853 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3854 (rt_policy(policy) != (attr->sched_priority != 0)))
3858 * Allow unprivileged RT tasks to decrease priority:
3860 if (user && !capable(CAP_SYS_NICE)) {
3861 if (fair_policy(policy)) {
3862 if (attr->sched_nice < task_nice(p) &&
3863 !can_nice(p, attr->sched_nice))
3867 if (rt_policy(policy)) {
3868 unsigned long rlim_rtprio =
3869 task_rlimit(p, RLIMIT_RTPRIO);
3871 /* can't set/change the rt policy */
3872 if (policy != p->policy && !rlim_rtprio)
3875 /* can't increase priority */
3876 if (attr->sched_priority > p->rt_priority &&
3877 attr->sched_priority > rlim_rtprio)
3882 * Can't set/change SCHED_DEADLINE policy at all for now
3883 * (safest behavior); in the future we would like to allow
3884 * unprivileged DL tasks to increase their relative deadline
3885 * or reduce their runtime (both ways reducing utilization)
3887 if (dl_policy(policy))
3891 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3892 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3894 if (idle_policy(p->policy) && !idle_policy(policy)) {
3895 if (!can_nice(p, task_nice(p)))
3899 /* can't change other user's priorities */
3900 if (!check_same_owner(p))
3903 /* Normal users shall not reset the sched_reset_on_fork flag */
3904 if (p->sched_reset_on_fork && !reset_on_fork)
3909 retval = security_task_setscheduler(p);
3915 * make sure no PI-waiters arrive (or leave) while we are
3916 * changing the priority of the task:
3918 * To be able to change p->policy safely, the appropriate
3919 * runqueue lock must be held.
3921 rq = task_rq_lock(p, &flags);
3924 * Changing the policy of the stop threads its a very bad idea
3926 if (p == rq->stop) {
3927 task_rq_unlock(rq, p, &flags);
3932 * If not changing anything there's no need to proceed further,
3933 * but store a possible modification of reset_on_fork.
3935 if (unlikely(policy == p->policy)) {
3936 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3938 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3940 if (dl_policy(policy) && dl_param_changed(p, attr))
3943 p->sched_reset_on_fork = reset_on_fork;
3944 task_rq_unlock(rq, p, &flags);
3950 #ifdef CONFIG_RT_GROUP_SCHED
3952 * Do not allow realtime tasks into groups that have no runtime
3955 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3956 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3957 !task_group_is_autogroup(task_group(p))) {
3958 task_rq_unlock(rq, p, &flags);
3963 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3964 cpumask_t *span = rq->rd->span;
3967 * Don't allow tasks with an affinity mask smaller than
3968 * the entire root_domain to become SCHED_DEADLINE. We
3969 * will also fail if there's no bandwidth available.
3971 if (!cpumask_subset(span, &p->cpus_allowed) ||
3972 rq->rd->dl_bw.bw == 0) {
3973 task_rq_unlock(rq, p, &flags);
3980 /* recheck policy now with rq lock held */
3981 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3982 policy = oldpolicy = -1;
3983 task_rq_unlock(rq, p, &flags);
3988 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3989 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3992 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3993 task_rq_unlock(rq, p, &flags);
3997 p->sched_reset_on_fork = reset_on_fork;
4002 * Take priority boosted tasks into account. If the new
4003 * effective priority is unchanged, we just store the new
4004 * normal parameters and do not touch the scheduler class and
4005 * the runqueue. This will be done when the task deboost
4008 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4009 if (new_effective_prio == oldprio)
4010 queue_flags &= ~DEQUEUE_MOVE;
4013 queued = task_on_rq_queued(p);
4014 running = task_current(rq, p);
4016 dequeue_task(rq, p, queue_flags);
4018 put_prev_task(rq, p);
4020 prev_class = p->sched_class;
4021 __setscheduler(rq, p, attr, pi);
4024 p->sched_class->set_curr_task(rq);
4027 * We enqueue to tail when the priority of a task is
4028 * increased (user space view).
4030 if (oldprio < p->prio)
4031 queue_flags |= ENQUEUE_HEAD;
4033 enqueue_task(rq, p, queue_flags);
4036 check_class_changed(rq, p, prev_class, oldprio);
4037 preempt_disable(); /* avoid rq from going away on us */
4038 task_rq_unlock(rq, p, &flags);
4041 rt_mutex_adjust_pi(p);
4044 * Run balance callbacks after we've adjusted the PI chain.
4046 balance_callback(rq);
4052 static int _sched_setscheduler(struct task_struct *p, int policy,
4053 const struct sched_param *param, bool check)
4055 struct sched_attr attr = {
4056 .sched_policy = policy,
4057 .sched_priority = param->sched_priority,
4058 .sched_nice = PRIO_TO_NICE(p->static_prio),
4061 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4062 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4063 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4064 policy &= ~SCHED_RESET_ON_FORK;
4065 attr.sched_policy = policy;
4068 return __sched_setscheduler(p, &attr, check, true);
4071 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4072 * @p: the task in question.
4073 * @policy: new policy.
4074 * @param: structure containing the new RT priority.
4076 * Return: 0 on success. An error code otherwise.
4078 * NOTE that the task may be already dead.
4080 int sched_setscheduler(struct task_struct *p, int policy,
4081 const struct sched_param *param)
4083 return _sched_setscheduler(p, policy, param, true);
4085 EXPORT_SYMBOL_GPL(sched_setscheduler);
4087 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4089 return __sched_setscheduler(p, attr, true, true);
4091 EXPORT_SYMBOL_GPL(sched_setattr);
4094 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4095 * @p: the task in question.
4096 * @policy: new policy.
4097 * @param: structure containing the new RT priority.
4099 * Just like sched_setscheduler, only don't bother checking if the
4100 * current context has permission. For example, this is needed in
4101 * stop_machine(): we create temporary high priority worker threads,
4102 * but our caller might not have that capability.
4104 * Return: 0 on success. An error code otherwise.
4106 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4107 const struct sched_param *param)
4109 return _sched_setscheduler(p, policy, param, false);
4111 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4114 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4116 struct sched_param lparam;
4117 struct task_struct *p;
4120 if (!param || pid < 0)
4122 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4127 p = find_process_by_pid(pid);
4129 retval = sched_setscheduler(p, policy, &lparam);
4136 * Mimics kernel/events/core.c perf_copy_attr().
4138 static int sched_copy_attr(struct sched_attr __user *uattr,
4139 struct sched_attr *attr)
4144 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4148 * zero the full structure, so that a short copy will be nice.
4150 memset(attr, 0, sizeof(*attr));
4152 ret = get_user(size, &uattr->size);
4156 if (size > PAGE_SIZE) /* silly large */
4159 if (!size) /* abi compat */
4160 size = SCHED_ATTR_SIZE_VER0;
4162 if (size < SCHED_ATTR_SIZE_VER0)
4166 * If we're handed a bigger struct than we know of,
4167 * ensure all the unknown bits are 0 - i.e. new
4168 * user-space does not rely on any kernel feature
4169 * extensions we dont know about yet.
4171 if (size > sizeof(*attr)) {
4172 unsigned char __user *addr;
4173 unsigned char __user *end;
4176 addr = (void __user *)uattr + sizeof(*attr);
4177 end = (void __user *)uattr + size;
4179 for (; addr < end; addr++) {
4180 ret = get_user(val, addr);
4186 size = sizeof(*attr);
4189 ret = copy_from_user(attr, uattr, size);
4194 * XXX: do we want to be lenient like existing syscalls; or do we want
4195 * to be strict and return an error on out-of-bounds values?
4197 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4202 put_user(sizeof(*attr), &uattr->size);
4207 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4208 * @pid: the pid in question.
4209 * @policy: new policy.
4210 * @param: structure containing the new RT priority.
4212 * Return: 0 on success. An error code otherwise.
4214 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4215 struct sched_param __user *, param)
4217 /* negative values for policy are not valid */
4221 return do_sched_setscheduler(pid, policy, param);
4225 * sys_sched_setparam - set/change the RT priority of a thread
4226 * @pid: the pid in question.
4227 * @param: structure containing the new RT priority.
4229 * Return: 0 on success. An error code otherwise.
4231 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4233 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4237 * sys_sched_setattr - same as above, but with extended sched_attr
4238 * @pid: the pid in question.
4239 * @uattr: structure containing the extended parameters.
4240 * @flags: for future extension.
4242 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4243 unsigned int, flags)
4245 struct sched_attr attr;
4246 struct task_struct *p;
4249 if (!uattr || pid < 0 || flags)
4252 retval = sched_copy_attr(uattr, &attr);
4256 if ((int)attr.sched_policy < 0)
4261 p = find_process_by_pid(pid);
4263 retval = sched_setattr(p, &attr);
4270 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4271 * @pid: the pid in question.
4273 * Return: On success, the policy of the thread. Otherwise, a negative error
4276 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4278 struct task_struct *p;
4286 p = find_process_by_pid(pid);
4288 retval = security_task_getscheduler(p);
4291 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4298 * sys_sched_getparam - get the RT priority of a thread
4299 * @pid: the pid in question.
4300 * @param: structure containing the RT priority.
4302 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4305 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4307 struct sched_param lp = { .sched_priority = 0 };
4308 struct task_struct *p;
4311 if (!param || pid < 0)
4315 p = find_process_by_pid(pid);
4320 retval = security_task_getscheduler(p);
4324 if (task_has_rt_policy(p))
4325 lp.sched_priority = p->rt_priority;
4329 * This one might sleep, we cannot do it with a spinlock held ...
4331 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4340 static int sched_read_attr(struct sched_attr __user *uattr,
4341 struct sched_attr *attr,
4346 if (!access_ok(VERIFY_WRITE, uattr, usize))
4350 * If we're handed a smaller struct than we know of,
4351 * ensure all the unknown bits are 0 - i.e. old
4352 * user-space does not get uncomplete information.
4354 if (usize < sizeof(*attr)) {
4355 unsigned char *addr;
4358 addr = (void *)attr + usize;
4359 end = (void *)attr + sizeof(*attr);
4361 for (; addr < end; addr++) {
4369 ret = copy_to_user(uattr, attr, attr->size);
4377 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4378 * @pid: the pid in question.
4379 * @uattr: structure containing the extended parameters.
4380 * @size: sizeof(attr) for fwd/bwd comp.
4381 * @flags: for future extension.
4383 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4384 unsigned int, size, unsigned int, flags)
4386 struct sched_attr attr = {
4387 .size = sizeof(struct sched_attr),
4389 struct task_struct *p;
4392 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4393 size < SCHED_ATTR_SIZE_VER0 || flags)
4397 p = find_process_by_pid(pid);
4402 retval = security_task_getscheduler(p);
4406 attr.sched_policy = p->policy;
4407 if (p->sched_reset_on_fork)
4408 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4409 if (task_has_dl_policy(p))
4410 __getparam_dl(p, &attr);
4411 else if (task_has_rt_policy(p))
4412 attr.sched_priority = p->rt_priority;
4414 attr.sched_nice = task_nice(p);
4418 retval = sched_read_attr(uattr, &attr, size);
4426 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4428 cpumask_var_t cpus_allowed, new_mask;
4429 struct task_struct *p;
4434 p = find_process_by_pid(pid);
4440 /* Prevent p going away */
4444 if (p->flags & PF_NO_SETAFFINITY) {
4448 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4452 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4454 goto out_free_cpus_allowed;
4457 if (!check_same_owner(p)) {
4459 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4461 goto out_free_new_mask;
4466 retval = security_task_setscheduler(p);
4468 goto out_free_new_mask;
4471 cpuset_cpus_allowed(p, cpus_allowed);
4472 cpumask_and(new_mask, in_mask, cpus_allowed);
4475 * Since bandwidth control happens on root_domain basis,
4476 * if admission test is enabled, we only admit -deadline
4477 * tasks allowed to run on all the CPUs in the task's
4481 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4483 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4486 goto out_free_new_mask;
4492 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4495 cpuset_cpus_allowed(p, cpus_allowed);
4496 if (!cpumask_subset(new_mask, cpus_allowed)) {
4498 * We must have raced with a concurrent cpuset
4499 * update. Just reset the cpus_allowed to the
4500 * cpuset's cpus_allowed
4502 cpumask_copy(new_mask, cpus_allowed);
4507 free_cpumask_var(new_mask);
4508 out_free_cpus_allowed:
4509 free_cpumask_var(cpus_allowed);
4515 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4516 struct cpumask *new_mask)
4518 if (len < cpumask_size())
4519 cpumask_clear(new_mask);
4520 else if (len > cpumask_size())
4521 len = cpumask_size();
4523 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4527 * sys_sched_setaffinity - set the cpu affinity of a process
4528 * @pid: pid of the process
4529 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4530 * @user_mask_ptr: user-space pointer to the new cpu mask
4532 * Return: 0 on success. An error code otherwise.
4534 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4535 unsigned long __user *, user_mask_ptr)
4537 cpumask_var_t new_mask;
4540 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4543 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4545 retval = sched_setaffinity(pid, new_mask);
4546 free_cpumask_var(new_mask);
4550 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4552 struct task_struct *p;
4553 unsigned long flags;
4559 p = find_process_by_pid(pid);
4563 retval = security_task_getscheduler(p);
4567 raw_spin_lock_irqsave(&p->pi_lock, flags);
4568 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4569 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4578 * sys_sched_getaffinity - get the cpu affinity of a process
4579 * @pid: pid of the process
4580 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4581 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4583 * Return: 0 on success. An error code otherwise.
4585 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4586 unsigned long __user *, user_mask_ptr)
4591 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4593 if (len & (sizeof(unsigned long)-1))
4596 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4599 ret = sched_getaffinity(pid, mask);
4601 size_t retlen = min_t(size_t, len, cpumask_size());
4603 if (copy_to_user(user_mask_ptr, mask, retlen))
4608 free_cpumask_var(mask);
4614 * sys_sched_yield - yield the current processor to other threads.
4616 * This function yields the current CPU to other tasks. If there are no
4617 * other threads running on this CPU then this function will return.
4621 SYSCALL_DEFINE0(sched_yield)
4623 struct rq *rq = this_rq_lock();
4625 schedstat_inc(rq, yld_count);
4626 current->sched_class->yield_task(rq);
4629 * Since we are going to call schedule() anyway, there's
4630 * no need to preempt or enable interrupts:
4632 __release(rq->lock);
4633 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4634 do_raw_spin_unlock(&rq->lock);
4635 sched_preempt_enable_no_resched();
4642 int __sched _cond_resched(void)
4644 if (should_resched(0)) {
4645 preempt_schedule_common();
4650 EXPORT_SYMBOL(_cond_resched);
4653 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4654 * call schedule, and on return reacquire the lock.
4656 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4657 * operations here to prevent schedule() from being called twice (once via
4658 * spin_unlock(), once by hand).
4660 int __cond_resched_lock(spinlock_t *lock)
4662 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4665 lockdep_assert_held(lock);
4667 if (spin_needbreak(lock) || resched) {
4670 preempt_schedule_common();
4678 EXPORT_SYMBOL(__cond_resched_lock);
4680 int __sched __cond_resched_softirq(void)
4682 BUG_ON(!in_softirq());
4684 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4686 preempt_schedule_common();
4692 EXPORT_SYMBOL(__cond_resched_softirq);
4695 * yield - yield the current processor to other threads.
4697 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4699 * The scheduler is at all times free to pick the calling task as the most
4700 * eligible task to run, if removing the yield() call from your code breaks
4701 * it, its already broken.
4703 * Typical broken usage is:
4708 * where one assumes that yield() will let 'the other' process run that will
4709 * make event true. If the current task is a SCHED_FIFO task that will never
4710 * happen. Never use yield() as a progress guarantee!!
4712 * If you want to use yield() to wait for something, use wait_event().
4713 * If you want to use yield() to be 'nice' for others, use cond_resched().
4714 * If you still want to use yield(), do not!
4716 void __sched yield(void)
4718 set_current_state(TASK_RUNNING);
4721 EXPORT_SYMBOL(yield);
4724 * yield_to - yield the current processor to another thread in
4725 * your thread group, or accelerate that thread toward the
4726 * processor it's on.
4728 * @preempt: whether task preemption is allowed or not
4730 * It's the caller's job to ensure that the target task struct
4731 * can't go away on us before we can do any checks.
4734 * true (>0) if we indeed boosted the target task.
4735 * false (0) if we failed to boost the target.
4736 * -ESRCH if there's no task to yield to.
4738 int __sched yield_to(struct task_struct *p, bool preempt)
4740 struct task_struct *curr = current;
4741 struct rq *rq, *p_rq;
4742 unsigned long flags;
4745 local_irq_save(flags);
4751 * If we're the only runnable task on the rq and target rq also
4752 * has only one task, there's absolutely no point in yielding.
4754 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4759 double_rq_lock(rq, p_rq);
4760 if (task_rq(p) != p_rq) {
4761 double_rq_unlock(rq, p_rq);
4765 if (!curr->sched_class->yield_to_task)
4768 if (curr->sched_class != p->sched_class)
4771 if (task_running(p_rq, p) || p->state)
4774 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4776 schedstat_inc(rq, yld_count);
4778 * Make p's CPU reschedule; pick_next_entity takes care of
4781 if (preempt && rq != p_rq)
4786 double_rq_unlock(rq, p_rq);
4788 local_irq_restore(flags);
4795 EXPORT_SYMBOL_GPL(yield_to);
4798 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4799 * that process accounting knows that this is a task in IO wait state.
4801 long __sched io_schedule_timeout(long timeout)
4803 int old_iowait = current->in_iowait;
4807 current->in_iowait = 1;
4808 blk_schedule_flush_plug(current);
4810 delayacct_blkio_start();
4812 atomic_inc(&rq->nr_iowait);
4813 ret = schedule_timeout(timeout);
4814 current->in_iowait = old_iowait;
4815 atomic_dec(&rq->nr_iowait);
4816 delayacct_blkio_end();
4820 EXPORT_SYMBOL(io_schedule_timeout);
4823 * sys_sched_get_priority_max - return maximum RT priority.
4824 * @policy: scheduling class.
4826 * Return: On success, this syscall returns the maximum
4827 * rt_priority that can be used by a given scheduling class.
4828 * On failure, a negative error code is returned.
4830 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4837 ret = MAX_USER_RT_PRIO-1;
4839 case SCHED_DEADLINE:
4850 * sys_sched_get_priority_min - return minimum RT priority.
4851 * @policy: scheduling class.
4853 * Return: On success, this syscall returns the minimum
4854 * rt_priority that can be used by a given scheduling class.
4855 * On failure, a negative error code is returned.
4857 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4866 case SCHED_DEADLINE:
4876 * sys_sched_rr_get_interval - return the default timeslice of a process.
4877 * @pid: pid of the process.
4878 * @interval: userspace pointer to the timeslice value.
4880 * this syscall writes the default timeslice value of a given process
4881 * into the user-space timespec buffer. A value of '0' means infinity.
4883 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4886 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4887 struct timespec __user *, interval)
4889 struct task_struct *p;
4890 unsigned int time_slice;
4891 unsigned long flags;
4901 p = find_process_by_pid(pid);
4905 retval = security_task_getscheduler(p);
4909 rq = task_rq_lock(p, &flags);
4911 if (p->sched_class->get_rr_interval)
4912 time_slice = p->sched_class->get_rr_interval(rq, p);
4913 task_rq_unlock(rq, p, &flags);
4916 jiffies_to_timespec(time_slice, &t);
4917 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4925 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4927 void sched_show_task(struct task_struct *p)
4929 unsigned long free = 0;
4931 unsigned long state = p->state;
4934 state = __ffs(state) + 1;
4935 printk(KERN_INFO "%-15.15s %c", p->comm,
4936 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4937 #if BITS_PER_LONG == 32
4938 if (state == TASK_RUNNING)
4939 printk(KERN_CONT " running ");
4941 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4943 if (state == TASK_RUNNING)
4944 printk(KERN_CONT " running task ");
4946 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4948 #ifdef CONFIG_DEBUG_STACK_USAGE
4949 free = stack_not_used(p);
4954 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4956 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4957 task_pid_nr(p), ppid,
4958 (unsigned long)task_thread_info(p)->flags);
4960 print_worker_info(KERN_INFO, p);
4961 show_stack(p, NULL);
4964 void show_state_filter(unsigned long state_filter)
4966 struct task_struct *g, *p;
4968 #if BITS_PER_LONG == 32
4970 " task PC stack pid father\n");
4973 " task PC stack pid father\n");
4976 for_each_process_thread(g, p) {
4978 * reset the NMI-timeout, listing all files on a slow
4979 * console might take a lot of time:
4981 touch_nmi_watchdog();
4982 if (!state_filter || (p->state & state_filter))
4986 touch_all_softlockup_watchdogs();
4988 #ifdef CONFIG_SCHED_DEBUG
4989 sysrq_sched_debug_show();
4993 * Only show locks if all tasks are dumped:
4996 debug_show_all_locks();
4999 void init_idle_bootup_task(struct task_struct *idle)
5001 idle->sched_class = &idle_sched_class;
5005 * init_idle - set up an idle thread for a given CPU
5006 * @idle: task in question
5007 * @cpu: cpu the idle task belongs to
5009 * NOTE: this function does not set the idle thread's NEED_RESCHED
5010 * flag, to make booting more robust.
5012 void init_idle(struct task_struct *idle, int cpu)
5014 struct rq *rq = cpu_rq(cpu);
5015 unsigned long flags;
5017 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5018 raw_spin_lock(&rq->lock);
5020 __sched_fork(0, idle);
5021 idle->state = TASK_RUNNING;
5022 idle->se.exec_start = sched_clock();
5026 * Its possible that init_idle() gets called multiple times on a task,
5027 * in that case do_set_cpus_allowed() will not do the right thing.
5029 * And since this is boot we can forgo the serialization.
5031 set_cpus_allowed_common(idle, cpumask_of(cpu));
5034 * We're having a chicken and egg problem, even though we are
5035 * holding rq->lock, the cpu isn't yet set to this cpu so the
5036 * lockdep check in task_group() will fail.
5038 * Similar case to sched_fork(). / Alternatively we could
5039 * use task_rq_lock() here and obtain the other rq->lock.
5044 __set_task_cpu(idle, cpu);
5047 rq->curr = rq->idle = idle;
5048 idle->on_rq = TASK_ON_RQ_QUEUED;
5052 raw_spin_unlock(&rq->lock);
5053 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5055 /* Set the preempt count _outside_ the spinlocks! */
5056 init_idle_preempt_count(idle, cpu);
5059 * The idle tasks have their own, simple scheduling class:
5061 idle->sched_class = &idle_sched_class;
5062 ftrace_graph_init_idle_task(idle, cpu);
5063 vtime_init_idle(idle, cpu);
5065 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5069 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5070 const struct cpumask *trial)
5072 int ret = 1, trial_cpus;
5073 struct dl_bw *cur_dl_b;
5074 unsigned long flags;
5076 if (!cpumask_weight(cur))
5079 rcu_read_lock_sched();
5080 cur_dl_b = dl_bw_of(cpumask_any(cur));
5081 trial_cpus = cpumask_weight(trial);
5083 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5084 if (cur_dl_b->bw != -1 &&
5085 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5087 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5088 rcu_read_unlock_sched();
5093 int task_can_attach(struct task_struct *p,
5094 const struct cpumask *cs_cpus_allowed)
5099 * Kthreads which disallow setaffinity shouldn't be moved
5100 * to a new cpuset; we don't want to change their cpu
5101 * affinity and isolating such threads by their set of
5102 * allowed nodes is unnecessary. Thus, cpusets are not
5103 * applicable for such threads. This prevents checking for
5104 * success of set_cpus_allowed_ptr() on all attached tasks
5105 * before cpus_allowed may be changed.
5107 if (p->flags & PF_NO_SETAFFINITY) {
5113 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5115 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5120 unsigned long flags;
5122 rcu_read_lock_sched();
5123 dl_b = dl_bw_of(dest_cpu);
5124 raw_spin_lock_irqsave(&dl_b->lock, flags);
5125 cpus = dl_bw_cpus(dest_cpu);
5126 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5131 * We reserve space for this task in the destination
5132 * root_domain, as we can't fail after this point.
5133 * We will free resources in the source root_domain
5134 * later on (see set_cpus_allowed_dl()).
5136 __dl_add(dl_b, p->dl.dl_bw);
5138 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5139 rcu_read_unlock_sched();
5149 #ifdef CONFIG_NUMA_BALANCING
5150 /* Migrate current task p to target_cpu */
5151 int migrate_task_to(struct task_struct *p, int target_cpu)
5153 struct migration_arg arg = { p, target_cpu };
5154 int curr_cpu = task_cpu(p);
5156 if (curr_cpu == target_cpu)
5159 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5162 /* TODO: This is not properly updating schedstats */
5164 trace_sched_move_numa(p, curr_cpu, target_cpu);
5165 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5169 * Requeue a task on a given node and accurately track the number of NUMA
5170 * tasks on the runqueues
5172 void sched_setnuma(struct task_struct *p, int nid)
5175 unsigned long flags;
5176 bool queued, running;
5178 rq = task_rq_lock(p, &flags);
5179 queued = task_on_rq_queued(p);
5180 running = task_current(rq, p);
5183 dequeue_task(rq, p, DEQUEUE_SAVE);
5185 put_prev_task(rq, p);
5187 p->numa_preferred_nid = nid;
5190 p->sched_class->set_curr_task(rq);
5192 enqueue_task(rq, p, ENQUEUE_RESTORE);
5193 task_rq_unlock(rq, p, &flags);
5195 #endif /* CONFIG_NUMA_BALANCING */
5197 #ifdef CONFIG_HOTPLUG_CPU
5199 * Ensures that the idle task is using init_mm right before its cpu goes
5202 void idle_task_exit(void)
5204 struct mm_struct *mm = current->active_mm;
5206 BUG_ON(cpu_online(smp_processor_id()));
5208 if (mm != &init_mm) {
5209 switch_mm(mm, &init_mm, current);
5210 finish_arch_post_lock_switch();
5216 * Since this CPU is going 'away' for a while, fold any nr_active delta
5217 * we might have. Assumes we're called after migrate_tasks() so that the
5218 * nr_active count is stable.
5220 * Also see the comment "Global load-average calculations".
5222 static void calc_load_migrate(struct rq *rq)
5224 long delta = calc_load_fold_active(rq);
5226 atomic_long_add(delta, &calc_load_tasks);
5229 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5233 static const struct sched_class fake_sched_class = {
5234 .put_prev_task = put_prev_task_fake,
5237 static struct task_struct fake_task = {
5239 * Avoid pull_{rt,dl}_task()
5241 .prio = MAX_PRIO + 1,
5242 .sched_class = &fake_sched_class,
5246 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5247 * try_to_wake_up()->select_task_rq().
5249 * Called with rq->lock held even though we'er in stop_machine() and
5250 * there's no concurrency possible, we hold the required locks anyway
5251 * because of lock validation efforts.
5253 static void migrate_tasks(struct rq *dead_rq)
5255 struct rq *rq = dead_rq;
5256 struct task_struct *next, *stop = rq->stop;
5260 * Fudge the rq selection such that the below task selection loop
5261 * doesn't get stuck on the currently eligible stop task.
5263 * We're currently inside stop_machine() and the rq is either stuck
5264 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5265 * either way we should never end up calling schedule() until we're
5271 * put_prev_task() and pick_next_task() sched
5272 * class method both need to have an up-to-date
5273 * value of rq->clock[_task]
5275 update_rq_clock(rq);
5279 * There's this thread running, bail when that's the only
5282 if (rq->nr_running == 1)
5286 * pick_next_task assumes pinned rq->lock.
5288 lockdep_pin_lock(&rq->lock);
5289 next = pick_next_task(rq, &fake_task);
5291 next->sched_class->put_prev_task(rq, next);
5294 * Rules for changing task_struct::cpus_allowed are holding
5295 * both pi_lock and rq->lock, such that holding either
5296 * stabilizes the mask.
5298 * Drop rq->lock is not quite as disastrous as it usually is
5299 * because !cpu_active at this point, which means load-balance
5300 * will not interfere. Also, stop-machine.
5302 lockdep_unpin_lock(&rq->lock);
5303 raw_spin_unlock(&rq->lock);
5304 raw_spin_lock(&next->pi_lock);
5305 raw_spin_lock(&rq->lock);
5308 * Since we're inside stop-machine, _nothing_ should have
5309 * changed the task, WARN if weird stuff happened, because in
5310 * that case the above rq->lock drop is a fail too.
5312 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5313 raw_spin_unlock(&next->pi_lock);
5317 /* Find suitable destination for @next, with force if needed. */
5318 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5320 rq = __migrate_task(rq, next, dest_cpu);
5321 if (rq != dead_rq) {
5322 raw_spin_unlock(&rq->lock);
5324 raw_spin_lock(&rq->lock);
5326 raw_spin_unlock(&next->pi_lock);
5331 #endif /* CONFIG_HOTPLUG_CPU */
5333 static void set_rq_online(struct rq *rq)
5336 const struct sched_class *class;
5338 cpumask_set_cpu(rq->cpu, rq->rd->online);
5341 for_each_class(class) {
5342 if (class->rq_online)
5343 class->rq_online(rq);
5348 static void set_rq_offline(struct rq *rq)
5351 const struct sched_class *class;
5353 for_each_class(class) {
5354 if (class->rq_offline)
5355 class->rq_offline(rq);
5358 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5364 * migration_call - callback that gets triggered when a CPU is added.
5365 * Here we can start up the necessary migration thread for the new CPU.
5368 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5370 int cpu = (long)hcpu;
5371 unsigned long flags;
5372 struct rq *rq = cpu_rq(cpu);
5374 switch (action & ~CPU_TASKS_FROZEN) {
5376 case CPU_UP_PREPARE:
5377 rq->calc_load_update = calc_load_update;
5381 /* Update our root-domain */
5382 raw_spin_lock_irqsave(&rq->lock, flags);
5384 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5388 raw_spin_unlock_irqrestore(&rq->lock, flags);
5391 #ifdef CONFIG_HOTPLUG_CPU
5393 sched_ttwu_pending();
5394 /* Update our root-domain */
5395 raw_spin_lock_irqsave(&rq->lock, flags);
5397 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5401 BUG_ON(rq->nr_running != 1); /* the migration thread */
5402 raw_spin_unlock_irqrestore(&rq->lock, flags);
5406 calc_load_migrate(rq);
5411 update_max_interval();
5417 * Register at high priority so that task migration (migrate_all_tasks)
5418 * happens before everything else. This has to be lower priority than
5419 * the notifier in the perf_event subsystem, though.
5421 static struct notifier_block migration_notifier = {
5422 .notifier_call = migration_call,
5423 .priority = CPU_PRI_MIGRATION,
5426 static void set_cpu_rq_start_time(void)
5428 int cpu = smp_processor_id();
5429 struct rq *rq = cpu_rq(cpu);
5430 rq->age_stamp = sched_clock_cpu(cpu);
5433 static int sched_cpu_active(struct notifier_block *nfb,
5434 unsigned long action, void *hcpu)
5436 int cpu = (long)hcpu;
5438 switch (action & ~CPU_TASKS_FROZEN) {
5440 set_cpu_rq_start_time();
5445 * At this point a starting CPU has marked itself as online via
5446 * set_cpu_online(). But it might not yet have marked itself
5447 * as active, which is essential from here on.
5449 set_cpu_active(cpu, true);
5450 stop_machine_unpark(cpu);
5453 case CPU_DOWN_FAILED:
5454 set_cpu_active(cpu, true);
5462 static int sched_cpu_inactive(struct notifier_block *nfb,
5463 unsigned long action, void *hcpu)
5465 switch (action & ~CPU_TASKS_FROZEN) {
5466 case CPU_DOWN_PREPARE:
5467 set_cpu_active((long)hcpu, false);
5474 static int __init migration_init(void)
5476 void *cpu = (void *)(long)smp_processor_id();
5479 /* Initialize migration for the boot CPU */
5480 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5481 BUG_ON(err == NOTIFY_BAD);
5482 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5483 register_cpu_notifier(&migration_notifier);
5485 /* Register cpu active notifiers */
5486 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5487 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5491 early_initcall(migration_init);
5493 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5495 #ifdef CONFIG_SCHED_DEBUG
5497 static __read_mostly int sched_debug_enabled;
5499 static int __init sched_debug_setup(char *str)
5501 sched_debug_enabled = 1;
5505 early_param("sched_debug", sched_debug_setup);
5507 static inline bool sched_debug(void)
5509 return sched_debug_enabled;
5512 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5513 struct cpumask *groupmask)
5515 struct sched_group *group = sd->groups;
5517 cpumask_clear(groupmask);
5519 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5521 if (!(sd->flags & SD_LOAD_BALANCE)) {
5522 printk("does not load-balance\n");
5524 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5529 printk(KERN_CONT "span %*pbl level %s\n",
5530 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5532 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5533 printk(KERN_ERR "ERROR: domain->span does not contain "
5536 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5537 printk(KERN_ERR "ERROR: domain->groups does not contain"
5541 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5545 printk(KERN_ERR "ERROR: group is NULL\n");
5549 if (!cpumask_weight(sched_group_cpus(group))) {
5550 printk(KERN_CONT "\n");
5551 printk(KERN_ERR "ERROR: empty group\n");
5555 if (!(sd->flags & SD_OVERLAP) &&
5556 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5557 printk(KERN_CONT "\n");
5558 printk(KERN_ERR "ERROR: repeated CPUs\n");
5562 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5564 printk(KERN_CONT " %*pbl",
5565 cpumask_pr_args(sched_group_cpus(group)));
5566 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5567 printk(KERN_CONT " (cpu_capacity = %d)",
5568 group->sgc->capacity);
5571 group = group->next;
5572 } while (group != sd->groups);
5573 printk(KERN_CONT "\n");
5575 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5576 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5579 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5580 printk(KERN_ERR "ERROR: parent span is not a superset "
5581 "of domain->span\n");
5585 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5589 if (!sched_debug_enabled)
5593 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5597 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5600 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5608 #else /* !CONFIG_SCHED_DEBUG */
5609 # define sched_domain_debug(sd, cpu) do { } while (0)
5610 static inline bool sched_debug(void)
5614 #endif /* CONFIG_SCHED_DEBUG */
5616 static int sd_degenerate(struct sched_domain *sd)
5618 if (cpumask_weight(sched_domain_span(sd)) == 1)
5621 /* Following flags need at least 2 groups */
5622 if (sd->flags & (SD_LOAD_BALANCE |
5623 SD_BALANCE_NEWIDLE |
5626 SD_SHARE_CPUCAPACITY |
5627 SD_SHARE_PKG_RESOURCES |
5628 SD_SHARE_POWERDOMAIN)) {
5629 if (sd->groups != sd->groups->next)
5633 /* Following flags don't use groups */
5634 if (sd->flags & (SD_WAKE_AFFINE))
5641 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5643 unsigned long cflags = sd->flags, pflags = parent->flags;
5645 if (sd_degenerate(parent))
5648 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5651 /* Flags needing groups don't count if only 1 group in parent */
5652 if (parent->groups == parent->groups->next) {
5653 pflags &= ~(SD_LOAD_BALANCE |
5654 SD_BALANCE_NEWIDLE |
5657 SD_SHARE_CPUCAPACITY |
5658 SD_SHARE_PKG_RESOURCES |
5660 SD_SHARE_POWERDOMAIN);
5661 if (nr_node_ids == 1)
5662 pflags &= ~SD_SERIALIZE;
5664 if (~cflags & pflags)
5670 static void free_rootdomain(struct rcu_head *rcu)
5672 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5674 cpupri_cleanup(&rd->cpupri);
5675 cpudl_cleanup(&rd->cpudl);
5676 free_cpumask_var(rd->dlo_mask);
5677 free_cpumask_var(rd->rto_mask);
5678 free_cpumask_var(rd->online);
5679 free_cpumask_var(rd->span);
5683 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5685 struct root_domain *old_rd = NULL;
5686 unsigned long flags;
5688 raw_spin_lock_irqsave(&rq->lock, flags);
5693 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5696 cpumask_clear_cpu(rq->cpu, old_rd->span);
5699 * If we dont want to free the old_rd yet then
5700 * set old_rd to NULL to skip the freeing later
5703 if (!atomic_dec_and_test(&old_rd->refcount))
5707 atomic_inc(&rd->refcount);
5710 cpumask_set_cpu(rq->cpu, rd->span);
5711 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5714 raw_spin_unlock_irqrestore(&rq->lock, flags);
5717 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5720 static int init_rootdomain(struct root_domain *rd)
5722 memset(rd, 0, sizeof(*rd));
5724 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5726 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5728 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5730 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5733 init_dl_bw(&rd->dl_bw);
5734 if (cpudl_init(&rd->cpudl) != 0)
5737 if (cpupri_init(&rd->cpupri) != 0)
5742 free_cpumask_var(rd->rto_mask);
5744 free_cpumask_var(rd->dlo_mask);
5746 free_cpumask_var(rd->online);
5748 free_cpumask_var(rd->span);
5754 * By default the system creates a single root-domain with all cpus as
5755 * members (mimicking the global state we have today).
5757 struct root_domain def_root_domain;
5759 static void init_defrootdomain(void)
5761 init_rootdomain(&def_root_domain);
5763 atomic_set(&def_root_domain.refcount, 1);
5766 static struct root_domain *alloc_rootdomain(void)
5768 struct root_domain *rd;
5770 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5774 if (init_rootdomain(rd) != 0) {
5782 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5784 struct sched_group *tmp, *first;
5793 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5798 } while (sg != first);
5801 static void free_sched_domain(struct rcu_head *rcu)
5803 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5806 * If its an overlapping domain it has private groups, iterate and
5809 if (sd->flags & SD_OVERLAP) {
5810 free_sched_groups(sd->groups, 1);
5811 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5812 kfree(sd->groups->sgc);
5818 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5820 call_rcu(&sd->rcu, free_sched_domain);
5823 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5825 for (; sd; sd = sd->parent)
5826 destroy_sched_domain(sd, cpu);
5830 * Keep a special pointer to the highest sched_domain that has
5831 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5832 * allows us to avoid some pointer chasing select_idle_sibling().
5834 * Also keep a unique ID per domain (we use the first cpu number in
5835 * the cpumask of the domain), this allows us to quickly tell if
5836 * two cpus are in the same cache domain, see cpus_share_cache().
5838 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5839 DEFINE_PER_CPU(int, sd_llc_size);
5840 DEFINE_PER_CPU(int, sd_llc_id);
5841 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5842 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5843 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5845 static void update_top_cache_domain(int cpu)
5847 struct sched_domain *sd;
5848 struct sched_domain *busy_sd = NULL;
5852 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5854 id = cpumask_first(sched_domain_span(sd));
5855 size = cpumask_weight(sched_domain_span(sd));
5856 busy_sd = sd->parent; /* sd_busy */
5858 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5860 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5861 per_cpu(sd_llc_size, cpu) = size;
5862 per_cpu(sd_llc_id, cpu) = id;
5864 sd = lowest_flag_domain(cpu, SD_NUMA);
5865 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5867 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5868 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5872 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5873 * hold the hotplug lock.
5876 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5878 struct rq *rq = cpu_rq(cpu);
5879 struct sched_domain *tmp;
5881 /* Remove the sched domains which do not contribute to scheduling. */
5882 for (tmp = sd; tmp; ) {
5883 struct sched_domain *parent = tmp->parent;
5887 if (sd_parent_degenerate(tmp, parent)) {
5888 tmp->parent = parent->parent;
5890 parent->parent->child = tmp;
5892 * Transfer SD_PREFER_SIBLING down in case of a
5893 * degenerate parent; the spans match for this
5894 * so the property transfers.
5896 if (parent->flags & SD_PREFER_SIBLING)
5897 tmp->flags |= SD_PREFER_SIBLING;
5898 destroy_sched_domain(parent, cpu);
5903 if (sd && sd_degenerate(sd)) {
5906 destroy_sched_domain(tmp, cpu);
5911 sched_domain_debug(sd, cpu);
5913 rq_attach_root(rq, rd);
5915 rcu_assign_pointer(rq->sd, sd);
5916 destroy_sched_domains(tmp, cpu);
5918 update_top_cache_domain(cpu);
5921 /* Setup the mask of cpus configured for isolated domains */
5922 static int __init isolated_cpu_setup(char *str)
5926 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5927 ret = cpulist_parse(str, cpu_isolated_map);
5929 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
5934 __setup("isolcpus=", isolated_cpu_setup);
5937 struct sched_domain ** __percpu sd;
5938 struct root_domain *rd;
5949 * Build an iteration mask that can exclude certain CPUs from the upwards
5952 * Asymmetric node setups can result in situations where the domain tree is of
5953 * unequal depth, make sure to skip domains that already cover the entire
5956 * In that case build_sched_domains() will have terminated the iteration early
5957 * and our sibling sd spans will be empty. Domains should always include the
5958 * cpu they're built on, so check that.
5961 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5963 const struct cpumask *span = sched_domain_span(sd);
5964 struct sd_data *sdd = sd->private;
5965 struct sched_domain *sibling;
5968 for_each_cpu(i, span) {
5969 sibling = *per_cpu_ptr(sdd->sd, i);
5970 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5973 cpumask_set_cpu(i, sched_group_mask(sg));
5978 * Return the canonical balance cpu for this group, this is the first cpu
5979 * of this group that's also in the iteration mask.
5981 int group_balance_cpu(struct sched_group *sg)
5983 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5987 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5989 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5990 const struct cpumask *span = sched_domain_span(sd);
5991 struct cpumask *covered = sched_domains_tmpmask;
5992 struct sd_data *sdd = sd->private;
5993 struct sched_domain *sibling;
5996 cpumask_clear(covered);
5998 for_each_cpu(i, span) {
5999 struct cpumask *sg_span;
6001 if (cpumask_test_cpu(i, covered))
6004 sibling = *per_cpu_ptr(sdd->sd, i);
6006 /* See the comment near build_group_mask(). */
6007 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6010 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6011 GFP_KERNEL, cpu_to_node(cpu));
6016 sg_span = sched_group_cpus(sg);
6018 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6020 cpumask_set_cpu(i, sg_span);
6022 cpumask_or(covered, covered, sg_span);
6024 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6025 if (atomic_inc_return(&sg->sgc->ref) == 1)
6026 build_group_mask(sd, sg);
6029 * Initialize sgc->capacity such that even if we mess up the
6030 * domains and no possible iteration will get us here, we won't
6033 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6036 * Make sure the first group of this domain contains the
6037 * canonical balance cpu. Otherwise the sched_domain iteration
6038 * breaks. See update_sg_lb_stats().
6040 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6041 group_balance_cpu(sg) == cpu)
6051 sd->groups = groups;
6056 free_sched_groups(first, 0);
6061 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6063 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6064 struct sched_domain *child = sd->child;
6067 cpu = cpumask_first(sched_domain_span(child));
6070 *sg = *per_cpu_ptr(sdd->sg, cpu);
6071 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6072 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6079 * build_sched_groups will build a circular linked list of the groups
6080 * covered by the given span, and will set each group's ->cpumask correctly,
6081 * and ->cpu_capacity to 0.
6083 * Assumes the sched_domain tree is fully constructed
6086 build_sched_groups(struct sched_domain *sd, int cpu)
6088 struct sched_group *first = NULL, *last = NULL;
6089 struct sd_data *sdd = sd->private;
6090 const struct cpumask *span = sched_domain_span(sd);
6091 struct cpumask *covered;
6094 get_group(cpu, sdd, &sd->groups);
6095 atomic_inc(&sd->groups->ref);
6097 if (cpu != cpumask_first(span))
6100 lockdep_assert_held(&sched_domains_mutex);
6101 covered = sched_domains_tmpmask;
6103 cpumask_clear(covered);
6105 for_each_cpu(i, span) {
6106 struct sched_group *sg;
6109 if (cpumask_test_cpu(i, covered))
6112 group = get_group(i, sdd, &sg);
6113 cpumask_setall(sched_group_mask(sg));
6115 for_each_cpu(j, span) {
6116 if (get_group(j, sdd, NULL) != group)
6119 cpumask_set_cpu(j, covered);
6120 cpumask_set_cpu(j, sched_group_cpus(sg));
6135 * Initialize sched groups cpu_capacity.
6137 * cpu_capacity indicates the capacity of sched group, which is used while
6138 * distributing the load between different sched groups in a sched domain.
6139 * Typically cpu_capacity for all the groups in a sched domain will be same
6140 * unless there are asymmetries in the topology. If there are asymmetries,
6141 * group having more cpu_capacity will pickup more load compared to the
6142 * group having less cpu_capacity.
6144 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6146 struct sched_group *sg = sd->groups;
6151 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6153 } while (sg != sd->groups);
6155 if (cpu != group_balance_cpu(sg))
6158 update_group_capacity(sd, cpu);
6159 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6163 * Initializers for schedule domains
6164 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6167 static int default_relax_domain_level = -1;
6168 int sched_domain_level_max;
6170 static int __init setup_relax_domain_level(char *str)
6172 if (kstrtoint(str, 0, &default_relax_domain_level))
6173 pr_warn("Unable to set relax_domain_level\n");
6177 __setup("relax_domain_level=", setup_relax_domain_level);
6179 static void set_domain_attribute(struct sched_domain *sd,
6180 struct sched_domain_attr *attr)
6184 if (!attr || attr->relax_domain_level < 0) {
6185 if (default_relax_domain_level < 0)
6188 request = default_relax_domain_level;
6190 request = attr->relax_domain_level;
6191 if (request < sd->level) {
6192 /* turn off idle balance on this domain */
6193 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6195 /* turn on idle balance on this domain */
6196 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6200 static void __sdt_free(const struct cpumask *cpu_map);
6201 static int __sdt_alloc(const struct cpumask *cpu_map);
6203 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6204 const struct cpumask *cpu_map)
6208 if (!atomic_read(&d->rd->refcount))
6209 free_rootdomain(&d->rd->rcu); /* fall through */
6211 free_percpu(d->sd); /* fall through */
6213 __sdt_free(cpu_map); /* fall through */
6219 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6220 const struct cpumask *cpu_map)
6222 memset(d, 0, sizeof(*d));
6224 if (__sdt_alloc(cpu_map))
6225 return sa_sd_storage;
6226 d->sd = alloc_percpu(struct sched_domain *);
6228 return sa_sd_storage;
6229 d->rd = alloc_rootdomain();
6232 return sa_rootdomain;
6236 * NULL the sd_data elements we've used to build the sched_domain and
6237 * sched_group structure so that the subsequent __free_domain_allocs()
6238 * will not free the data we're using.
6240 static void claim_allocations(int cpu, struct sched_domain *sd)
6242 struct sd_data *sdd = sd->private;
6244 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6245 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6247 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6248 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6250 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6251 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6255 static int sched_domains_numa_levels;
6256 enum numa_topology_type sched_numa_topology_type;
6257 static int *sched_domains_numa_distance;
6258 int sched_max_numa_distance;
6259 static struct cpumask ***sched_domains_numa_masks;
6260 static int sched_domains_curr_level;
6264 * SD_flags allowed in topology descriptions.
6266 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6267 * SD_SHARE_PKG_RESOURCES - describes shared caches
6268 * SD_NUMA - describes NUMA topologies
6269 * SD_SHARE_POWERDOMAIN - describes shared power domain
6272 * SD_ASYM_PACKING - describes SMT quirks
6274 #define TOPOLOGY_SD_FLAGS \
6275 (SD_SHARE_CPUCAPACITY | \
6276 SD_SHARE_PKG_RESOURCES | \
6279 SD_SHARE_POWERDOMAIN)
6281 static struct sched_domain *
6282 sd_init(struct sched_domain_topology_level *tl, int cpu)
6284 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6285 int sd_weight, sd_flags = 0;
6289 * Ugly hack to pass state to sd_numa_mask()...
6291 sched_domains_curr_level = tl->numa_level;
6294 sd_weight = cpumask_weight(tl->mask(cpu));
6297 sd_flags = (*tl->sd_flags)();
6298 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6299 "wrong sd_flags in topology description\n"))
6300 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6302 *sd = (struct sched_domain){
6303 .min_interval = sd_weight,
6304 .max_interval = 2*sd_weight,
6306 .imbalance_pct = 125,
6308 .cache_nice_tries = 0,
6315 .flags = 1*SD_LOAD_BALANCE
6316 | 1*SD_BALANCE_NEWIDLE
6321 | 0*SD_SHARE_CPUCAPACITY
6322 | 0*SD_SHARE_PKG_RESOURCES
6324 | 0*SD_PREFER_SIBLING
6329 .last_balance = jiffies,
6330 .balance_interval = sd_weight,
6332 .max_newidle_lb_cost = 0,
6333 .next_decay_max_lb_cost = jiffies,
6334 #ifdef CONFIG_SCHED_DEBUG
6340 * Convert topological properties into behaviour.
6343 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6344 sd->flags |= SD_PREFER_SIBLING;
6345 sd->imbalance_pct = 110;
6346 sd->smt_gain = 1178; /* ~15% */
6348 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6349 sd->imbalance_pct = 117;
6350 sd->cache_nice_tries = 1;
6354 } else if (sd->flags & SD_NUMA) {
6355 sd->cache_nice_tries = 2;
6359 sd->flags |= SD_SERIALIZE;
6360 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6361 sd->flags &= ~(SD_BALANCE_EXEC |
6368 sd->flags |= SD_PREFER_SIBLING;
6369 sd->cache_nice_tries = 1;
6374 sd->private = &tl->data;
6380 * Topology list, bottom-up.
6382 static struct sched_domain_topology_level default_topology[] = {
6383 #ifdef CONFIG_SCHED_SMT
6384 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6386 #ifdef CONFIG_SCHED_MC
6387 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6389 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6393 static struct sched_domain_topology_level *sched_domain_topology =
6396 #define for_each_sd_topology(tl) \
6397 for (tl = sched_domain_topology; tl->mask; tl++)
6399 void set_sched_topology(struct sched_domain_topology_level *tl)
6401 sched_domain_topology = tl;
6406 static const struct cpumask *sd_numa_mask(int cpu)
6408 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6411 static void sched_numa_warn(const char *str)
6413 static int done = false;
6421 printk(KERN_WARNING "ERROR: %s\n\n", str);
6423 for (i = 0; i < nr_node_ids; i++) {
6424 printk(KERN_WARNING " ");
6425 for (j = 0; j < nr_node_ids; j++)
6426 printk(KERN_CONT "%02d ", node_distance(i,j));
6427 printk(KERN_CONT "\n");
6429 printk(KERN_WARNING "\n");
6432 bool find_numa_distance(int distance)
6436 if (distance == node_distance(0, 0))
6439 for (i = 0; i < sched_domains_numa_levels; i++) {
6440 if (sched_domains_numa_distance[i] == distance)
6448 * A system can have three types of NUMA topology:
6449 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6450 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6451 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6453 * The difference between a glueless mesh topology and a backplane
6454 * topology lies in whether communication between not directly
6455 * connected nodes goes through intermediary nodes (where programs
6456 * could run), or through backplane controllers. This affects
6457 * placement of programs.
6459 * The type of topology can be discerned with the following tests:
6460 * - If the maximum distance between any nodes is 1 hop, the system
6461 * is directly connected.
6462 * - If for two nodes A and B, located N > 1 hops away from each other,
6463 * there is an intermediary node C, which is < N hops away from both
6464 * nodes A and B, the system is a glueless mesh.
6466 static void init_numa_topology_type(void)
6470 n = sched_max_numa_distance;
6472 if (sched_domains_numa_levels <= 1) {
6473 sched_numa_topology_type = NUMA_DIRECT;
6477 for_each_online_node(a) {
6478 for_each_online_node(b) {
6479 /* Find two nodes furthest removed from each other. */
6480 if (node_distance(a, b) < n)
6483 /* Is there an intermediary node between a and b? */
6484 for_each_online_node(c) {
6485 if (node_distance(a, c) < n &&
6486 node_distance(b, c) < n) {
6487 sched_numa_topology_type =
6493 sched_numa_topology_type = NUMA_BACKPLANE;
6499 static void sched_init_numa(void)
6501 int next_distance, curr_distance = node_distance(0, 0);
6502 struct sched_domain_topology_level *tl;
6506 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6507 if (!sched_domains_numa_distance)
6511 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6512 * unique distances in the node_distance() table.
6514 * Assumes node_distance(0,j) includes all distances in
6515 * node_distance(i,j) in order to avoid cubic time.
6517 next_distance = curr_distance;
6518 for (i = 0; i < nr_node_ids; i++) {
6519 for (j = 0; j < nr_node_ids; j++) {
6520 for (k = 0; k < nr_node_ids; k++) {
6521 int distance = node_distance(i, k);
6523 if (distance > curr_distance &&
6524 (distance < next_distance ||
6525 next_distance == curr_distance))
6526 next_distance = distance;
6529 * While not a strong assumption it would be nice to know
6530 * about cases where if node A is connected to B, B is not
6531 * equally connected to A.
6533 if (sched_debug() && node_distance(k, i) != distance)
6534 sched_numa_warn("Node-distance not symmetric");
6536 if (sched_debug() && i && !find_numa_distance(distance))
6537 sched_numa_warn("Node-0 not representative");
6539 if (next_distance != curr_distance) {
6540 sched_domains_numa_distance[level++] = next_distance;
6541 sched_domains_numa_levels = level;
6542 curr_distance = next_distance;
6547 * In case of sched_debug() we verify the above assumption.
6557 * 'level' contains the number of unique distances, excluding the
6558 * identity distance node_distance(i,i).
6560 * The sched_domains_numa_distance[] array includes the actual distance
6565 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6566 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6567 * the array will contain less then 'level' members. This could be
6568 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6569 * in other functions.
6571 * We reset it to 'level' at the end of this function.
6573 sched_domains_numa_levels = 0;
6575 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6576 if (!sched_domains_numa_masks)
6580 * Now for each level, construct a mask per node which contains all
6581 * cpus of nodes that are that many hops away from us.
6583 for (i = 0; i < level; i++) {
6584 sched_domains_numa_masks[i] =
6585 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6586 if (!sched_domains_numa_masks[i])
6589 for (j = 0; j < nr_node_ids; j++) {
6590 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6594 sched_domains_numa_masks[i][j] = mask;
6597 if (node_distance(j, k) > sched_domains_numa_distance[i])
6600 cpumask_or(mask, mask, cpumask_of_node(k));
6605 /* Compute default topology size */
6606 for (i = 0; sched_domain_topology[i].mask; i++);
6608 tl = kzalloc((i + level + 1) *
6609 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6614 * Copy the default topology bits..
6616 for (i = 0; sched_domain_topology[i].mask; i++)
6617 tl[i] = sched_domain_topology[i];
6620 * .. and append 'j' levels of NUMA goodness.
6622 for (j = 0; j < level; i++, j++) {
6623 tl[i] = (struct sched_domain_topology_level){
6624 .mask = sd_numa_mask,
6625 .sd_flags = cpu_numa_flags,
6626 .flags = SDTL_OVERLAP,
6632 sched_domain_topology = tl;
6634 sched_domains_numa_levels = level;
6635 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6637 init_numa_topology_type();
6640 static void sched_domains_numa_masks_set(int cpu)
6643 int node = cpu_to_node(cpu);
6645 for (i = 0; i < sched_domains_numa_levels; i++) {
6646 for (j = 0; j < nr_node_ids; j++) {
6647 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6648 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6653 static void sched_domains_numa_masks_clear(int cpu)
6656 for (i = 0; i < sched_domains_numa_levels; i++) {
6657 for (j = 0; j < nr_node_ids; j++)
6658 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6663 * Update sched_domains_numa_masks[level][node] array when new cpus
6666 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6667 unsigned long action,
6670 int cpu = (long)hcpu;
6672 switch (action & ~CPU_TASKS_FROZEN) {
6674 sched_domains_numa_masks_set(cpu);
6678 sched_domains_numa_masks_clear(cpu);
6688 static inline void sched_init_numa(void)
6692 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6693 unsigned long action,
6698 #endif /* CONFIG_NUMA */
6700 static int __sdt_alloc(const struct cpumask *cpu_map)
6702 struct sched_domain_topology_level *tl;
6705 for_each_sd_topology(tl) {
6706 struct sd_data *sdd = &tl->data;
6708 sdd->sd = alloc_percpu(struct sched_domain *);
6712 sdd->sg = alloc_percpu(struct sched_group *);
6716 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6720 for_each_cpu(j, cpu_map) {
6721 struct sched_domain *sd;
6722 struct sched_group *sg;
6723 struct sched_group_capacity *sgc;
6725 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6726 GFP_KERNEL, cpu_to_node(j));
6730 *per_cpu_ptr(sdd->sd, j) = sd;
6732 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6733 GFP_KERNEL, cpu_to_node(j));
6739 *per_cpu_ptr(sdd->sg, j) = sg;
6741 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6742 GFP_KERNEL, cpu_to_node(j));
6746 *per_cpu_ptr(sdd->sgc, j) = sgc;
6753 static void __sdt_free(const struct cpumask *cpu_map)
6755 struct sched_domain_topology_level *tl;
6758 for_each_sd_topology(tl) {
6759 struct sd_data *sdd = &tl->data;
6761 for_each_cpu(j, cpu_map) {
6762 struct sched_domain *sd;
6765 sd = *per_cpu_ptr(sdd->sd, j);
6766 if (sd && (sd->flags & SD_OVERLAP))
6767 free_sched_groups(sd->groups, 0);
6768 kfree(*per_cpu_ptr(sdd->sd, j));
6772 kfree(*per_cpu_ptr(sdd->sg, j));
6774 kfree(*per_cpu_ptr(sdd->sgc, j));
6776 free_percpu(sdd->sd);
6778 free_percpu(sdd->sg);
6780 free_percpu(sdd->sgc);
6785 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6786 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6787 struct sched_domain *child, int cpu)
6789 struct sched_domain *sd = sd_init(tl, cpu);
6793 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6795 sd->level = child->level + 1;
6796 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6800 if (!cpumask_subset(sched_domain_span(child),
6801 sched_domain_span(sd))) {
6802 pr_err("BUG: arch topology borken\n");
6803 #ifdef CONFIG_SCHED_DEBUG
6804 pr_err(" the %s domain not a subset of the %s domain\n",
6805 child->name, sd->name);
6807 /* Fixup, ensure @sd has at least @child cpus. */
6808 cpumask_or(sched_domain_span(sd),
6809 sched_domain_span(sd),
6810 sched_domain_span(child));
6814 set_domain_attribute(sd, attr);
6820 * Build sched domains for a given set of cpus and attach the sched domains
6821 * to the individual cpus
6823 static int build_sched_domains(const struct cpumask *cpu_map,
6824 struct sched_domain_attr *attr)
6826 enum s_alloc alloc_state;
6827 struct sched_domain *sd;
6829 int i, ret = -ENOMEM;
6831 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6832 if (alloc_state != sa_rootdomain)
6835 /* Set up domains for cpus specified by the cpu_map. */
6836 for_each_cpu(i, cpu_map) {
6837 struct sched_domain_topology_level *tl;
6840 for_each_sd_topology(tl) {
6841 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6842 if (tl == sched_domain_topology)
6843 *per_cpu_ptr(d.sd, i) = sd;
6844 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6845 sd->flags |= SD_OVERLAP;
6846 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6851 /* Build the groups for the domains */
6852 for_each_cpu(i, cpu_map) {
6853 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6854 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6855 if (sd->flags & SD_OVERLAP) {
6856 if (build_overlap_sched_groups(sd, i))
6859 if (build_sched_groups(sd, i))
6865 /* Calculate CPU capacity for physical packages and nodes */
6866 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6867 if (!cpumask_test_cpu(i, cpu_map))
6870 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6871 claim_allocations(i, sd);
6872 init_sched_groups_capacity(i, sd);
6876 /* Attach the domains */
6878 for_each_cpu(i, cpu_map) {
6879 sd = *per_cpu_ptr(d.sd, i);
6880 cpu_attach_domain(sd, d.rd, i);
6886 __free_domain_allocs(&d, alloc_state, cpu_map);
6890 static cpumask_var_t *doms_cur; /* current sched domains */
6891 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6892 static struct sched_domain_attr *dattr_cur;
6893 /* attribues of custom domains in 'doms_cur' */
6896 * Special case: If a kmalloc of a doms_cur partition (array of
6897 * cpumask) fails, then fallback to a single sched domain,
6898 * as determined by the single cpumask fallback_doms.
6900 static cpumask_var_t fallback_doms;
6903 * arch_update_cpu_topology lets virtualized architectures update the
6904 * cpu core maps. It is supposed to return 1 if the topology changed
6905 * or 0 if it stayed the same.
6907 int __weak arch_update_cpu_topology(void)
6912 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6915 cpumask_var_t *doms;
6917 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6920 for (i = 0; i < ndoms; i++) {
6921 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6922 free_sched_domains(doms, i);
6929 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6932 for (i = 0; i < ndoms; i++)
6933 free_cpumask_var(doms[i]);
6938 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6939 * For now this just excludes isolated cpus, but could be used to
6940 * exclude other special cases in the future.
6942 static int init_sched_domains(const struct cpumask *cpu_map)
6946 arch_update_cpu_topology();
6948 doms_cur = alloc_sched_domains(ndoms_cur);
6950 doms_cur = &fallback_doms;
6951 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6952 err = build_sched_domains(doms_cur[0], NULL);
6953 register_sched_domain_sysctl();
6959 * Detach sched domains from a group of cpus specified in cpu_map
6960 * These cpus will now be attached to the NULL domain
6962 static void detach_destroy_domains(const struct cpumask *cpu_map)
6967 for_each_cpu(i, cpu_map)
6968 cpu_attach_domain(NULL, &def_root_domain, i);
6972 /* handle null as "default" */
6973 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6974 struct sched_domain_attr *new, int idx_new)
6976 struct sched_domain_attr tmp;
6983 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6984 new ? (new + idx_new) : &tmp,
6985 sizeof(struct sched_domain_attr));
6989 * Partition sched domains as specified by the 'ndoms_new'
6990 * cpumasks in the array doms_new[] of cpumasks. This compares
6991 * doms_new[] to the current sched domain partitioning, doms_cur[].
6992 * It destroys each deleted domain and builds each new domain.
6994 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6995 * The masks don't intersect (don't overlap.) We should setup one
6996 * sched domain for each mask. CPUs not in any of the cpumasks will
6997 * not be load balanced. If the same cpumask appears both in the
6998 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7001 * The passed in 'doms_new' should be allocated using
7002 * alloc_sched_domains. This routine takes ownership of it and will
7003 * free_sched_domains it when done with it. If the caller failed the
7004 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7005 * and partition_sched_domains() will fallback to the single partition
7006 * 'fallback_doms', it also forces the domains to be rebuilt.
7008 * If doms_new == NULL it will be replaced with cpu_online_mask.
7009 * ndoms_new == 0 is a special case for destroying existing domains,
7010 * and it will not create the default domain.
7012 * Call with hotplug lock held
7014 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7015 struct sched_domain_attr *dattr_new)
7020 mutex_lock(&sched_domains_mutex);
7022 /* always unregister in case we don't destroy any domains */
7023 unregister_sched_domain_sysctl();
7025 /* Let architecture update cpu core mappings. */
7026 new_topology = arch_update_cpu_topology();
7028 n = doms_new ? ndoms_new : 0;
7030 /* Destroy deleted domains */
7031 for (i = 0; i < ndoms_cur; i++) {
7032 for (j = 0; j < n && !new_topology; j++) {
7033 if (cpumask_equal(doms_cur[i], doms_new[j])
7034 && dattrs_equal(dattr_cur, i, dattr_new, j))
7037 /* no match - a current sched domain not in new doms_new[] */
7038 detach_destroy_domains(doms_cur[i]);
7044 if (doms_new == NULL) {
7046 doms_new = &fallback_doms;
7047 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7048 WARN_ON_ONCE(dattr_new);
7051 /* Build new domains */
7052 for (i = 0; i < ndoms_new; i++) {
7053 for (j = 0; j < n && !new_topology; j++) {
7054 if (cpumask_equal(doms_new[i], doms_cur[j])
7055 && dattrs_equal(dattr_new, i, dattr_cur, j))
7058 /* no match - add a new doms_new */
7059 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7064 /* Remember the new sched domains */
7065 if (doms_cur != &fallback_doms)
7066 free_sched_domains(doms_cur, ndoms_cur);
7067 kfree(dattr_cur); /* kfree(NULL) is safe */
7068 doms_cur = doms_new;
7069 dattr_cur = dattr_new;
7070 ndoms_cur = ndoms_new;
7072 register_sched_domain_sysctl();
7074 mutex_unlock(&sched_domains_mutex);
7077 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7080 * Update cpusets according to cpu_active mask. If cpusets are
7081 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7082 * around partition_sched_domains().
7084 * If we come here as part of a suspend/resume, don't touch cpusets because we
7085 * want to restore it back to its original state upon resume anyway.
7087 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7091 case CPU_ONLINE_FROZEN:
7092 case CPU_DOWN_FAILED_FROZEN:
7095 * num_cpus_frozen tracks how many CPUs are involved in suspend
7096 * resume sequence. As long as this is not the last online
7097 * operation in the resume sequence, just build a single sched
7098 * domain, ignoring cpusets.
7101 if (likely(num_cpus_frozen)) {
7102 partition_sched_domains(1, NULL, NULL);
7107 * This is the last CPU online operation. So fall through and
7108 * restore the original sched domains by considering the
7109 * cpuset configurations.
7113 cpuset_update_active_cpus(true);
7121 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7124 unsigned long flags;
7125 long cpu = (long)hcpu;
7131 case CPU_DOWN_PREPARE:
7132 rcu_read_lock_sched();
7133 dl_b = dl_bw_of(cpu);
7135 raw_spin_lock_irqsave(&dl_b->lock, flags);
7136 cpus = dl_bw_cpus(cpu);
7137 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7138 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7140 rcu_read_unlock_sched();
7143 return notifier_from_errno(-EBUSY);
7144 cpuset_update_active_cpus(false);
7146 case CPU_DOWN_PREPARE_FROZEN:
7148 partition_sched_domains(1, NULL, NULL);
7156 void __init sched_init_smp(void)
7158 cpumask_var_t non_isolated_cpus;
7160 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7161 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7166 * There's no userspace yet to cause hotplug operations; hence all the
7167 * cpu masks are stable and all blatant races in the below code cannot
7170 mutex_lock(&sched_domains_mutex);
7171 init_sched_domains(cpu_active_mask);
7172 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7173 if (cpumask_empty(non_isolated_cpus))
7174 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7175 mutex_unlock(&sched_domains_mutex);
7177 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7178 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7179 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7183 /* Move init over to a non-isolated CPU */
7184 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7186 sched_init_granularity();
7187 free_cpumask_var(non_isolated_cpus);
7189 init_sched_rt_class();
7190 init_sched_dl_class();
7193 void __init sched_init_smp(void)
7195 sched_init_granularity();
7197 #endif /* CONFIG_SMP */
7199 int in_sched_functions(unsigned long addr)
7201 return in_lock_functions(addr) ||
7202 (addr >= (unsigned long)__sched_text_start
7203 && addr < (unsigned long)__sched_text_end);
7206 #ifdef CONFIG_CGROUP_SCHED
7208 * Default task group.
7209 * Every task in system belongs to this group at bootup.
7211 struct task_group root_task_group;
7212 LIST_HEAD(task_groups);
7214 /* Cacheline aligned slab cache for task_group */
7215 static struct kmem_cache *task_group_cache __read_mostly;
7218 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7220 void __init sched_init(void)
7223 unsigned long alloc_size = 0, ptr;
7225 #ifdef CONFIG_FAIR_GROUP_SCHED
7226 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7228 #ifdef CONFIG_RT_GROUP_SCHED
7229 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7232 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7234 #ifdef CONFIG_FAIR_GROUP_SCHED
7235 root_task_group.se = (struct sched_entity **)ptr;
7236 ptr += nr_cpu_ids * sizeof(void **);
7238 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7239 ptr += nr_cpu_ids * sizeof(void **);
7241 #endif /* CONFIG_FAIR_GROUP_SCHED */
7242 #ifdef CONFIG_RT_GROUP_SCHED
7243 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7244 ptr += nr_cpu_ids * sizeof(void **);
7246 root_task_group.rt_rq = (struct rt_rq **)ptr;
7247 ptr += nr_cpu_ids * sizeof(void **);
7249 #endif /* CONFIG_RT_GROUP_SCHED */
7251 #ifdef CONFIG_CPUMASK_OFFSTACK
7252 for_each_possible_cpu(i) {
7253 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7254 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7256 #endif /* CONFIG_CPUMASK_OFFSTACK */
7258 init_rt_bandwidth(&def_rt_bandwidth,
7259 global_rt_period(), global_rt_runtime());
7260 init_dl_bandwidth(&def_dl_bandwidth,
7261 global_rt_period(), global_rt_runtime());
7264 init_defrootdomain();
7267 #ifdef CONFIG_RT_GROUP_SCHED
7268 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7269 global_rt_period(), global_rt_runtime());
7270 #endif /* CONFIG_RT_GROUP_SCHED */
7272 #ifdef CONFIG_CGROUP_SCHED
7273 task_group_cache = KMEM_CACHE(task_group, 0);
7275 list_add(&root_task_group.list, &task_groups);
7276 INIT_LIST_HEAD(&root_task_group.children);
7277 INIT_LIST_HEAD(&root_task_group.siblings);
7278 autogroup_init(&init_task);
7279 #endif /* CONFIG_CGROUP_SCHED */
7281 for_each_possible_cpu(i) {
7285 raw_spin_lock_init(&rq->lock);
7287 rq->calc_load_active = 0;
7288 rq->calc_load_update = jiffies + LOAD_FREQ;
7289 init_cfs_rq(&rq->cfs);
7290 init_rt_rq(&rq->rt);
7291 init_dl_rq(&rq->dl);
7292 #ifdef CONFIG_FAIR_GROUP_SCHED
7293 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7294 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7296 * How much cpu bandwidth does root_task_group get?
7298 * In case of task-groups formed thr' the cgroup filesystem, it
7299 * gets 100% of the cpu resources in the system. This overall
7300 * system cpu resource is divided among the tasks of
7301 * root_task_group and its child task-groups in a fair manner,
7302 * based on each entity's (task or task-group's) weight
7303 * (se->load.weight).
7305 * In other words, if root_task_group has 10 tasks of weight
7306 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7307 * then A0's share of the cpu resource is:
7309 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7311 * We achieve this by letting root_task_group's tasks sit
7312 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7314 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7315 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7316 #endif /* CONFIG_FAIR_GROUP_SCHED */
7318 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7319 #ifdef CONFIG_RT_GROUP_SCHED
7320 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7323 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7324 rq->cpu_load[j] = 0;
7326 rq->last_load_update_tick = jiffies;
7331 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7332 rq->balance_callback = NULL;
7333 rq->active_balance = 0;
7334 rq->next_balance = jiffies;
7339 rq->avg_idle = 2*sysctl_sched_migration_cost;
7340 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7342 INIT_LIST_HEAD(&rq->cfs_tasks);
7344 rq_attach_root(rq, &def_root_domain);
7345 #ifdef CONFIG_NO_HZ_COMMON
7348 #ifdef CONFIG_NO_HZ_FULL
7349 rq->last_sched_tick = 0;
7353 atomic_set(&rq->nr_iowait, 0);
7356 set_load_weight(&init_task);
7358 #ifdef CONFIG_PREEMPT_NOTIFIERS
7359 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7363 * The boot idle thread does lazy MMU switching as well:
7365 atomic_inc(&init_mm.mm_count);
7366 enter_lazy_tlb(&init_mm, current);
7369 * During early bootup we pretend to be a normal task:
7371 current->sched_class = &fair_sched_class;
7374 * Make us the idle thread. Technically, schedule() should not be
7375 * called from this thread, however somewhere below it might be,
7376 * but because we are the idle thread, we just pick up running again
7377 * when this runqueue becomes "idle".
7379 init_idle(current, smp_processor_id());
7381 calc_load_update = jiffies + LOAD_FREQ;
7384 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7385 /* May be allocated at isolcpus cmdline parse time */
7386 if (cpu_isolated_map == NULL)
7387 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7388 idle_thread_set_boot_cpu();
7389 set_cpu_rq_start_time();
7391 init_sched_fair_class();
7393 scheduler_running = 1;
7396 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7397 static inline int preempt_count_equals(int preempt_offset)
7399 int nested = preempt_count() + rcu_preempt_depth();
7401 return (nested == preempt_offset);
7404 void __might_sleep(const char *file, int line, int preempt_offset)
7407 * Blocking primitives will set (and therefore destroy) current->state,
7408 * since we will exit with TASK_RUNNING make sure we enter with it,
7409 * otherwise we will destroy state.
7411 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7412 "do not call blocking ops when !TASK_RUNNING; "
7413 "state=%lx set at [<%p>] %pS\n",
7415 (void *)current->task_state_change,
7416 (void *)current->task_state_change);
7418 ___might_sleep(file, line, preempt_offset);
7420 EXPORT_SYMBOL(__might_sleep);
7422 void ___might_sleep(const char *file, int line, int preempt_offset)
7424 static unsigned long prev_jiffy; /* ratelimiting */
7426 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7427 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7428 !is_idle_task(current)) ||
7429 system_state != SYSTEM_RUNNING || oops_in_progress)
7431 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7433 prev_jiffy = jiffies;
7436 "BUG: sleeping function called from invalid context at %s:%d\n",
7439 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7440 in_atomic(), irqs_disabled(),
7441 current->pid, current->comm);
7443 if (task_stack_end_corrupted(current))
7444 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7446 debug_show_held_locks(current);
7447 if (irqs_disabled())
7448 print_irqtrace_events(current);
7449 #ifdef CONFIG_DEBUG_PREEMPT
7450 if (!preempt_count_equals(preempt_offset)) {
7451 pr_err("Preemption disabled at:");
7452 print_ip_sym(current->preempt_disable_ip);
7458 EXPORT_SYMBOL(___might_sleep);
7461 #ifdef CONFIG_MAGIC_SYSRQ
7462 void normalize_rt_tasks(void)
7464 struct task_struct *g, *p;
7465 struct sched_attr attr = {
7466 .sched_policy = SCHED_NORMAL,
7469 read_lock(&tasklist_lock);
7470 for_each_process_thread(g, p) {
7472 * Only normalize user tasks:
7474 if (p->flags & PF_KTHREAD)
7477 p->se.exec_start = 0;
7478 #ifdef CONFIG_SCHEDSTATS
7479 p->se.statistics.wait_start = 0;
7480 p->se.statistics.sleep_start = 0;
7481 p->se.statistics.block_start = 0;
7484 if (!dl_task(p) && !rt_task(p)) {
7486 * Renice negative nice level userspace
7489 if (task_nice(p) < 0)
7490 set_user_nice(p, 0);
7494 __sched_setscheduler(p, &attr, false, false);
7496 read_unlock(&tasklist_lock);
7499 #endif /* CONFIG_MAGIC_SYSRQ */
7501 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7503 * These functions are only useful for the IA64 MCA handling, or kdb.
7505 * They can only be called when the whole system has been
7506 * stopped - every CPU needs to be quiescent, and no scheduling
7507 * activity can take place. Using them for anything else would
7508 * be a serious bug, and as a result, they aren't even visible
7509 * under any other configuration.
7513 * curr_task - return the current task for a given cpu.
7514 * @cpu: the processor in question.
7516 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7518 * Return: The current task for @cpu.
7520 struct task_struct *curr_task(int cpu)
7522 return cpu_curr(cpu);
7525 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7529 * set_curr_task - set the current task for a given cpu.
7530 * @cpu: the processor in question.
7531 * @p: the task pointer to set.
7533 * Description: This function must only be used when non-maskable interrupts
7534 * are serviced on a separate stack. It allows the architecture to switch the
7535 * notion of the current task on a cpu in a non-blocking manner. This function
7536 * must be called with all CPU's synchronized, and interrupts disabled, the
7537 * and caller must save the original value of the current task (see
7538 * curr_task() above) and restore that value before reenabling interrupts and
7539 * re-starting the system.
7541 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7543 void set_curr_task(int cpu, struct task_struct *p)
7550 #ifdef CONFIG_CGROUP_SCHED
7551 /* task_group_lock serializes the addition/removal of task groups */
7552 static DEFINE_SPINLOCK(task_group_lock);
7554 static void free_sched_group(struct task_group *tg)
7556 free_fair_sched_group(tg);
7557 free_rt_sched_group(tg);
7559 kmem_cache_free(task_group_cache, tg);
7562 /* allocate runqueue etc for a new task group */
7563 struct task_group *sched_create_group(struct task_group *parent)
7565 struct task_group *tg;
7567 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7569 return ERR_PTR(-ENOMEM);
7571 if (!alloc_fair_sched_group(tg, parent))
7574 if (!alloc_rt_sched_group(tg, parent))
7580 free_sched_group(tg);
7581 return ERR_PTR(-ENOMEM);
7584 void sched_online_group(struct task_group *tg, struct task_group *parent)
7586 unsigned long flags;
7588 spin_lock_irqsave(&task_group_lock, flags);
7589 list_add_rcu(&tg->list, &task_groups);
7591 WARN_ON(!parent); /* root should already exist */
7593 tg->parent = parent;
7594 INIT_LIST_HEAD(&tg->children);
7595 list_add_rcu(&tg->siblings, &parent->children);
7596 spin_unlock_irqrestore(&task_group_lock, flags);
7599 /* rcu callback to free various structures associated with a task group */
7600 static void free_sched_group_rcu(struct rcu_head *rhp)
7602 /* now it should be safe to free those cfs_rqs */
7603 free_sched_group(container_of(rhp, struct task_group, rcu));
7606 /* Destroy runqueue etc associated with a task group */
7607 void sched_destroy_group(struct task_group *tg)
7609 /* wait for possible concurrent references to cfs_rqs complete */
7610 call_rcu(&tg->rcu, free_sched_group_rcu);
7613 void sched_offline_group(struct task_group *tg)
7615 unsigned long flags;
7617 /* end participation in shares distribution */
7618 unregister_fair_sched_group(tg);
7620 spin_lock_irqsave(&task_group_lock, flags);
7621 list_del_rcu(&tg->list);
7622 list_del_rcu(&tg->siblings);
7623 spin_unlock_irqrestore(&task_group_lock, flags);
7626 /* change task's runqueue when it moves between groups.
7627 * The caller of this function should have put the task in its new group
7628 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7629 * reflect its new group.
7631 void sched_move_task(struct task_struct *tsk)
7633 struct task_group *tg;
7634 int queued, running;
7635 unsigned long flags;
7638 rq = task_rq_lock(tsk, &flags);
7640 running = task_current(rq, tsk);
7641 queued = task_on_rq_queued(tsk);
7644 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7645 if (unlikely(running))
7646 put_prev_task(rq, tsk);
7649 * All callers are synchronized by task_rq_lock(); we do not use RCU
7650 * which is pointless here. Thus, we pass "true" to task_css_check()
7651 * to prevent lockdep warnings.
7653 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7654 struct task_group, css);
7655 tg = autogroup_task_group(tsk, tg);
7656 tsk->sched_task_group = tg;
7658 #ifdef CONFIG_FAIR_GROUP_SCHED
7659 if (tsk->sched_class->task_move_group)
7660 tsk->sched_class->task_move_group(tsk);
7663 set_task_rq(tsk, task_cpu(tsk));
7665 if (unlikely(running))
7666 tsk->sched_class->set_curr_task(rq);
7668 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7670 task_rq_unlock(rq, tsk, &flags);
7672 #endif /* CONFIG_CGROUP_SCHED */
7674 #ifdef CONFIG_RT_GROUP_SCHED
7676 * Ensure that the real time constraints are schedulable.
7678 static DEFINE_MUTEX(rt_constraints_mutex);
7680 /* Must be called with tasklist_lock held */
7681 static inline int tg_has_rt_tasks(struct task_group *tg)
7683 struct task_struct *g, *p;
7686 * Autogroups do not have RT tasks; see autogroup_create().
7688 if (task_group_is_autogroup(tg))
7691 for_each_process_thread(g, p) {
7692 if (rt_task(p) && task_group(p) == tg)
7699 struct rt_schedulable_data {
7700 struct task_group *tg;
7705 static int tg_rt_schedulable(struct task_group *tg, void *data)
7707 struct rt_schedulable_data *d = data;
7708 struct task_group *child;
7709 unsigned long total, sum = 0;
7710 u64 period, runtime;
7712 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7713 runtime = tg->rt_bandwidth.rt_runtime;
7716 period = d->rt_period;
7717 runtime = d->rt_runtime;
7721 * Cannot have more runtime than the period.
7723 if (runtime > period && runtime != RUNTIME_INF)
7727 * Ensure we don't starve existing RT tasks.
7729 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7732 total = to_ratio(period, runtime);
7735 * Nobody can have more than the global setting allows.
7737 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7741 * The sum of our children's runtime should not exceed our own.
7743 list_for_each_entry_rcu(child, &tg->children, siblings) {
7744 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7745 runtime = child->rt_bandwidth.rt_runtime;
7747 if (child == d->tg) {
7748 period = d->rt_period;
7749 runtime = d->rt_runtime;
7752 sum += to_ratio(period, runtime);
7761 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7765 struct rt_schedulable_data data = {
7767 .rt_period = period,
7768 .rt_runtime = runtime,
7772 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7778 static int tg_set_rt_bandwidth(struct task_group *tg,
7779 u64 rt_period, u64 rt_runtime)
7784 * Disallowing the root group RT runtime is BAD, it would disallow the
7785 * kernel creating (and or operating) RT threads.
7787 if (tg == &root_task_group && rt_runtime == 0)
7790 /* No period doesn't make any sense. */
7794 mutex_lock(&rt_constraints_mutex);
7795 read_lock(&tasklist_lock);
7796 err = __rt_schedulable(tg, rt_period, rt_runtime);
7800 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7801 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7802 tg->rt_bandwidth.rt_runtime = rt_runtime;
7804 for_each_possible_cpu(i) {
7805 struct rt_rq *rt_rq = tg->rt_rq[i];
7807 raw_spin_lock(&rt_rq->rt_runtime_lock);
7808 rt_rq->rt_runtime = rt_runtime;
7809 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7811 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7813 read_unlock(&tasklist_lock);
7814 mutex_unlock(&rt_constraints_mutex);
7819 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7821 u64 rt_runtime, rt_period;
7823 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7824 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7825 if (rt_runtime_us < 0)
7826 rt_runtime = RUNTIME_INF;
7828 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7831 static long sched_group_rt_runtime(struct task_group *tg)
7835 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7838 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7839 do_div(rt_runtime_us, NSEC_PER_USEC);
7840 return rt_runtime_us;
7843 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7845 u64 rt_runtime, rt_period;
7847 rt_period = rt_period_us * NSEC_PER_USEC;
7848 rt_runtime = tg->rt_bandwidth.rt_runtime;
7850 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7853 static long sched_group_rt_period(struct task_group *tg)
7857 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7858 do_div(rt_period_us, NSEC_PER_USEC);
7859 return rt_period_us;
7861 #endif /* CONFIG_RT_GROUP_SCHED */
7863 #ifdef CONFIG_RT_GROUP_SCHED
7864 static int sched_rt_global_constraints(void)
7868 mutex_lock(&rt_constraints_mutex);
7869 read_lock(&tasklist_lock);
7870 ret = __rt_schedulable(NULL, 0, 0);
7871 read_unlock(&tasklist_lock);
7872 mutex_unlock(&rt_constraints_mutex);
7877 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7879 /* Don't accept realtime tasks when there is no way for them to run */
7880 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7886 #else /* !CONFIG_RT_GROUP_SCHED */
7887 static int sched_rt_global_constraints(void)
7889 unsigned long flags;
7892 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7893 for_each_possible_cpu(i) {
7894 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7896 raw_spin_lock(&rt_rq->rt_runtime_lock);
7897 rt_rq->rt_runtime = global_rt_runtime();
7898 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7900 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7904 #endif /* CONFIG_RT_GROUP_SCHED */
7906 static int sched_dl_global_validate(void)
7908 u64 runtime = global_rt_runtime();
7909 u64 period = global_rt_period();
7910 u64 new_bw = to_ratio(period, runtime);
7913 unsigned long flags;
7916 * Here we want to check the bandwidth not being set to some
7917 * value smaller than the currently allocated bandwidth in
7918 * any of the root_domains.
7920 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7921 * cycling on root_domains... Discussion on different/better
7922 * solutions is welcome!
7924 for_each_possible_cpu(cpu) {
7925 rcu_read_lock_sched();
7926 dl_b = dl_bw_of(cpu);
7928 raw_spin_lock_irqsave(&dl_b->lock, flags);
7929 if (new_bw < dl_b->total_bw)
7931 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7933 rcu_read_unlock_sched();
7942 static void sched_dl_do_global(void)
7947 unsigned long flags;
7949 def_dl_bandwidth.dl_period = global_rt_period();
7950 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7952 if (global_rt_runtime() != RUNTIME_INF)
7953 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7956 * FIXME: As above...
7958 for_each_possible_cpu(cpu) {
7959 rcu_read_lock_sched();
7960 dl_b = dl_bw_of(cpu);
7962 raw_spin_lock_irqsave(&dl_b->lock, flags);
7964 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7966 rcu_read_unlock_sched();
7970 static int sched_rt_global_validate(void)
7972 if (sysctl_sched_rt_period <= 0)
7975 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7976 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7982 static void sched_rt_do_global(void)
7984 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7985 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7988 int sched_rt_handler(struct ctl_table *table, int write,
7989 void __user *buffer, size_t *lenp,
7992 int old_period, old_runtime;
7993 static DEFINE_MUTEX(mutex);
7997 old_period = sysctl_sched_rt_period;
7998 old_runtime = sysctl_sched_rt_runtime;
8000 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8002 if (!ret && write) {
8003 ret = sched_rt_global_validate();
8007 ret = sched_dl_global_validate();
8011 ret = sched_rt_global_constraints();
8015 sched_rt_do_global();
8016 sched_dl_do_global();
8020 sysctl_sched_rt_period = old_period;
8021 sysctl_sched_rt_runtime = old_runtime;
8023 mutex_unlock(&mutex);
8028 int sched_rr_handler(struct ctl_table *table, int write,
8029 void __user *buffer, size_t *lenp,
8033 static DEFINE_MUTEX(mutex);
8036 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8037 /* make sure that internally we keep jiffies */
8038 /* also, writing zero resets timeslice to default */
8039 if (!ret && write) {
8040 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8041 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8043 mutex_unlock(&mutex);
8047 #ifdef CONFIG_CGROUP_SCHED
8049 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8051 return css ? container_of(css, struct task_group, css) : NULL;
8054 static struct cgroup_subsys_state *
8055 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8057 struct task_group *parent = css_tg(parent_css);
8058 struct task_group *tg;
8061 /* This is early initialization for the top cgroup */
8062 return &root_task_group.css;
8065 tg = sched_create_group(parent);
8067 return ERR_PTR(-ENOMEM);
8072 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8074 struct task_group *tg = css_tg(css);
8075 struct task_group *parent = css_tg(css->parent);
8078 sched_online_group(tg, parent);
8082 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8084 struct task_group *tg = css_tg(css);
8086 sched_destroy_group(tg);
8089 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8091 struct task_group *tg = css_tg(css);
8093 sched_offline_group(tg);
8096 static void cpu_cgroup_fork(struct task_struct *task)
8098 sched_move_task(task);
8101 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8103 struct task_struct *task;
8104 struct cgroup_subsys_state *css;
8106 cgroup_taskset_for_each(task, css, tset) {
8107 #ifdef CONFIG_RT_GROUP_SCHED
8108 if (!sched_rt_can_attach(css_tg(css), task))
8111 /* We don't support RT-tasks being in separate groups */
8112 if (task->sched_class != &fair_sched_class)
8119 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8121 struct task_struct *task;
8122 struct cgroup_subsys_state *css;
8124 cgroup_taskset_for_each(task, css, tset)
8125 sched_move_task(task);
8128 #ifdef CONFIG_FAIR_GROUP_SCHED
8129 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8130 struct cftype *cftype, u64 shareval)
8132 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8135 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8138 struct task_group *tg = css_tg(css);
8140 return (u64) scale_load_down(tg->shares);
8143 #ifdef CONFIG_CFS_BANDWIDTH
8144 static DEFINE_MUTEX(cfs_constraints_mutex);
8146 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8147 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8149 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8151 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8153 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8154 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8156 if (tg == &root_task_group)
8160 * Ensure we have at some amount of bandwidth every period. This is
8161 * to prevent reaching a state of large arrears when throttled via
8162 * entity_tick() resulting in prolonged exit starvation.
8164 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8168 * Likewise, bound things on the otherside by preventing insane quota
8169 * periods. This also allows us to normalize in computing quota
8172 if (period > max_cfs_quota_period)
8176 * Prevent race between setting of cfs_rq->runtime_enabled and
8177 * unthrottle_offline_cfs_rqs().
8180 mutex_lock(&cfs_constraints_mutex);
8181 ret = __cfs_schedulable(tg, period, quota);
8185 runtime_enabled = quota != RUNTIME_INF;
8186 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8188 * If we need to toggle cfs_bandwidth_used, off->on must occur
8189 * before making related changes, and on->off must occur afterwards
8191 if (runtime_enabled && !runtime_was_enabled)
8192 cfs_bandwidth_usage_inc();
8193 raw_spin_lock_irq(&cfs_b->lock);
8194 cfs_b->period = ns_to_ktime(period);
8195 cfs_b->quota = quota;
8197 __refill_cfs_bandwidth_runtime(cfs_b);
8198 /* restart the period timer (if active) to handle new period expiry */
8199 if (runtime_enabled)
8200 start_cfs_bandwidth(cfs_b);
8201 raw_spin_unlock_irq(&cfs_b->lock);
8203 for_each_online_cpu(i) {
8204 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8205 struct rq *rq = cfs_rq->rq;
8207 raw_spin_lock_irq(&rq->lock);
8208 cfs_rq->runtime_enabled = runtime_enabled;
8209 cfs_rq->runtime_remaining = 0;
8211 if (cfs_rq->throttled)
8212 unthrottle_cfs_rq(cfs_rq);
8213 raw_spin_unlock_irq(&rq->lock);
8215 if (runtime_was_enabled && !runtime_enabled)
8216 cfs_bandwidth_usage_dec();
8218 mutex_unlock(&cfs_constraints_mutex);
8224 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8228 period = ktime_to_ns(tg->cfs_bandwidth.period);
8229 if (cfs_quota_us < 0)
8230 quota = RUNTIME_INF;
8232 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8234 return tg_set_cfs_bandwidth(tg, period, quota);
8237 long tg_get_cfs_quota(struct task_group *tg)
8241 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8244 quota_us = tg->cfs_bandwidth.quota;
8245 do_div(quota_us, NSEC_PER_USEC);
8250 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8254 period = (u64)cfs_period_us * NSEC_PER_USEC;
8255 quota = tg->cfs_bandwidth.quota;
8257 return tg_set_cfs_bandwidth(tg, period, quota);
8260 long tg_get_cfs_period(struct task_group *tg)
8264 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8265 do_div(cfs_period_us, NSEC_PER_USEC);
8267 return cfs_period_us;
8270 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8273 return tg_get_cfs_quota(css_tg(css));
8276 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8277 struct cftype *cftype, s64 cfs_quota_us)
8279 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8282 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8285 return tg_get_cfs_period(css_tg(css));
8288 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8289 struct cftype *cftype, u64 cfs_period_us)
8291 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8294 struct cfs_schedulable_data {
8295 struct task_group *tg;
8300 * normalize group quota/period to be quota/max_period
8301 * note: units are usecs
8303 static u64 normalize_cfs_quota(struct task_group *tg,
8304 struct cfs_schedulable_data *d)
8312 period = tg_get_cfs_period(tg);
8313 quota = tg_get_cfs_quota(tg);
8316 /* note: these should typically be equivalent */
8317 if (quota == RUNTIME_INF || quota == -1)
8320 return to_ratio(period, quota);
8323 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8325 struct cfs_schedulable_data *d = data;
8326 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8327 s64 quota = 0, parent_quota = -1;
8330 quota = RUNTIME_INF;
8332 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8334 quota = normalize_cfs_quota(tg, d);
8335 parent_quota = parent_b->hierarchical_quota;
8338 * ensure max(child_quota) <= parent_quota, inherit when no
8341 if (quota == RUNTIME_INF)
8342 quota = parent_quota;
8343 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8346 cfs_b->hierarchical_quota = quota;
8351 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8354 struct cfs_schedulable_data data = {
8360 if (quota != RUNTIME_INF) {
8361 do_div(data.period, NSEC_PER_USEC);
8362 do_div(data.quota, NSEC_PER_USEC);
8366 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8372 static int cpu_stats_show(struct seq_file *sf, void *v)
8374 struct task_group *tg = css_tg(seq_css(sf));
8375 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8377 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8378 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8379 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8383 #endif /* CONFIG_CFS_BANDWIDTH */
8384 #endif /* CONFIG_FAIR_GROUP_SCHED */
8386 #ifdef CONFIG_RT_GROUP_SCHED
8387 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8388 struct cftype *cft, s64 val)
8390 return sched_group_set_rt_runtime(css_tg(css), val);
8393 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8396 return sched_group_rt_runtime(css_tg(css));
8399 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8400 struct cftype *cftype, u64 rt_period_us)
8402 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8405 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8408 return sched_group_rt_period(css_tg(css));
8410 #endif /* CONFIG_RT_GROUP_SCHED */
8412 static struct cftype cpu_files[] = {
8413 #ifdef CONFIG_FAIR_GROUP_SCHED
8416 .read_u64 = cpu_shares_read_u64,
8417 .write_u64 = cpu_shares_write_u64,
8420 #ifdef CONFIG_CFS_BANDWIDTH
8422 .name = "cfs_quota_us",
8423 .read_s64 = cpu_cfs_quota_read_s64,
8424 .write_s64 = cpu_cfs_quota_write_s64,
8427 .name = "cfs_period_us",
8428 .read_u64 = cpu_cfs_period_read_u64,
8429 .write_u64 = cpu_cfs_period_write_u64,
8433 .seq_show = cpu_stats_show,
8436 #ifdef CONFIG_RT_GROUP_SCHED
8438 .name = "rt_runtime_us",
8439 .read_s64 = cpu_rt_runtime_read,
8440 .write_s64 = cpu_rt_runtime_write,
8443 .name = "rt_period_us",
8444 .read_u64 = cpu_rt_period_read_uint,
8445 .write_u64 = cpu_rt_period_write_uint,
8451 struct cgroup_subsys cpu_cgrp_subsys = {
8452 .css_alloc = cpu_cgroup_css_alloc,
8453 .css_free = cpu_cgroup_css_free,
8454 .css_online = cpu_cgroup_css_online,
8455 .css_offline = cpu_cgroup_css_offline,
8456 .fork = cpu_cgroup_fork,
8457 .can_attach = cpu_cgroup_can_attach,
8458 .attach = cpu_cgroup_attach,
8459 .legacy_cftypes = cpu_files,
8463 #endif /* CONFIG_CGROUP_SCHED */
8465 void dump_cpu_task(int cpu)
8467 pr_info("Task dump for CPU %d:\n", cpu);
8468 sched_show_task(cpu_curr(cpu));
8472 * Nice levels are multiplicative, with a gentle 10% change for every
8473 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8474 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8475 * that remained on nice 0.
8477 * The "10% effect" is relative and cumulative: from _any_ nice level,
8478 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8479 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8480 * If a task goes up by ~10% and another task goes down by ~10% then
8481 * the relative distance between them is ~25%.)
8483 const int sched_prio_to_weight[40] = {
8484 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8485 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8486 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8487 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8488 /* 0 */ 1024, 820, 655, 526, 423,
8489 /* 5 */ 335, 272, 215, 172, 137,
8490 /* 10 */ 110, 87, 70, 56, 45,
8491 /* 15 */ 36, 29, 23, 18, 15,
8495 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8497 * In cases where the weight does not change often, we can use the
8498 * precalculated inverse to speed up arithmetics by turning divisions
8499 * into multiplications:
8501 const u32 sched_prio_to_wmult[40] = {
8502 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8503 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8504 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8505 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8506 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8507 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8508 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8509 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,