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/binfmts.h>
74 #include <linux/context_tracking.h>
75 #include <linux/compiler.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 DEFINE_MUTEX(sched_domains_mutex);
93 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
95 static void update_rq_clock_task(struct rq *rq, s64 delta);
97 void update_rq_clock(struct rq *rq)
101 lockdep_assert_held(&rq->lock);
103 if (rq->clock_skip_update & RQCF_ACT_SKIP)
106 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
110 update_rq_clock_task(rq, delta);
114 * Debugging: various feature bits
117 #define SCHED_FEAT(name, enabled) \
118 (1UL << __SCHED_FEAT_##name) * enabled |
120 const_debug unsigned int sysctl_sched_features =
121 #include "features.h"
127 * Number of tasks to iterate in a single balance run.
128 * Limited because this is done with IRQs disabled.
130 const_debug unsigned int sysctl_sched_nr_migrate = 32;
133 * period over which we average the RT time consumption, measured
138 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
141 * period over which we measure -rt task cpu usage in us.
144 unsigned int sysctl_sched_rt_period = 1000000;
146 __read_mostly int scheduler_running;
149 * part of the period that we allow rt tasks to run in us.
152 int sysctl_sched_rt_runtime = 950000;
154 /* cpus with isolated domains */
155 cpumask_var_t cpu_isolated_map;
158 * this_rq_lock - lock this runqueue and disable interrupts.
160 static struct rq *this_rq_lock(void)
167 raw_spin_lock(&rq->lock);
172 #ifdef CONFIG_SCHED_HRTICK
174 * Use HR-timers to deliver accurate preemption points.
177 static void hrtick_clear(struct rq *rq)
179 if (hrtimer_active(&rq->hrtick_timer))
180 hrtimer_cancel(&rq->hrtick_timer);
184 * High-resolution timer tick.
185 * Runs from hardirq context with interrupts disabled.
187 static enum hrtimer_restart hrtick(struct hrtimer *timer)
189 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
191 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
193 raw_spin_lock(&rq->lock);
195 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
196 raw_spin_unlock(&rq->lock);
198 return HRTIMER_NORESTART;
203 static void __hrtick_restart(struct rq *rq)
205 struct hrtimer *timer = &rq->hrtick_timer;
207 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
211 * called from hardirq (IPI) context
213 static void __hrtick_start(void *arg)
217 raw_spin_lock(&rq->lock);
218 __hrtick_restart(rq);
219 rq->hrtick_csd_pending = 0;
220 raw_spin_unlock(&rq->lock);
224 * Called to set the hrtick timer state.
226 * called with rq->lock held and irqs disabled
228 void hrtick_start(struct rq *rq, u64 delay)
230 struct hrtimer *timer = &rq->hrtick_timer;
235 * Don't schedule slices shorter than 10000ns, that just
236 * doesn't make sense and can cause timer DoS.
238 delta = max_t(s64, delay, 10000LL);
239 time = ktime_add_ns(timer->base->get_time(), delta);
241 hrtimer_set_expires(timer, time);
243 if (rq == this_rq()) {
244 __hrtick_restart(rq);
245 } else if (!rq->hrtick_csd_pending) {
246 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
247 rq->hrtick_csd_pending = 1;
252 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
254 int cpu = (int)(long)hcpu;
257 case CPU_UP_CANCELED:
258 case CPU_UP_CANCELED_FROZEN:
259 case CPU_DOWN_PREPARE:
260 case CPU_DOWN_PREPARE_FROZEN:
262 case CPU_DEAD_FROZEN:
263 hrtick_clear(cpu_rq(cpu));
270 static __init void init_hrtick(void)
272 hotcpu_notifier(hotplug_hrtick, 0);
276 * Called to set the hrtick timer state.
278 * called with rq->lock held and irqs disabled
280 void hrtick_start(struct rq *rq, u64 delay)
283 * Don't schedule slices shorter than 10000ns, that just
284 * doesn't make sense. Rely on vruntime for fairness.
286 delay = max_t(u64, delay, 10000LL);
287 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
288 HRTIMER_MODE_REL_PINNED);
291 static inline void init_hrtick(void)
294 #endif /* CONFIG_SMP */
296 static void init_rq_hrtick(struct rq *rq)
299 rq->hrtick_csd_pending = 0;
301 rq->hrtick_csd.flags = 0;
302 rq->hrtick_csd.func = __hrtick_start;
303 rq->hrtick_csd.info = rq;
306 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
307 rq->hrtick_timer.function = hrtick;
309 #else /* CONFIG_SCHED_HRTICK */
310 static inline void hrtick_clear(struct rq *rq)
314 static inline void init_rq_hrtick(struct rq *rq)
318 static inline void init_hrtick(void)
321 #endif /* CONFIG_SCHED_HRTICK */
324 * cmpxchg based fetch_or, macro so it works for different integer types
326 #define fetch_or(ptr, val) \
327 ({ typeof(*(ptr)) __old, __val = *(ptr); \
329 __old = cmpxchg((ptr), __val, __val | (val)); \
330 if (__old == __val) \
337 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
339 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
340 * this avoids any races wrt polling state changes and thereby avoids
343 static bool set_nr_and_not_polling(struct task_struct *p)
345 struct thread_info *ti = task_thread_info(p);
346 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
350 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
352 * If this returns true, then the idle task promises to call
353 * sched_ttwu_pending() and reschedule soon.
355 static bool set_nr_if_polling(struct task_struct *p)
357 struct thread_info *ti = task_thread_info(p);
358 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
361 if (!(val & _TIF_POLLING_NRFLAG))
363 if (val & _TIF_NEED_RESCHED)
365 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
374 static bool set_nr_and_not_polling(struct task_struct *p)
376 set_tsk_need_resched(p);
381 static bool set_nr_if_polling(struct task_struct *p)
388 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
390 struct wake_q_node *node = &task->wake_q;
393 * Atomically grab the task, if ->wake_q is !nil already it means
394 * its already queued (either by us or someone else) and will get the
395 * wakeup due to that.
397 * This cmpxchg() implies a full barrier, which pairs with the write
398 * barrier implied by the wakeup in wake_up_list().
400 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
403 get_task_struct(task);
406 * The head is context local, there can be no concurrency.
409 head->lastp = &node->next;
412 void wake_up_q(struct wake_q_head *head)
414 struct wake_q_node *node = head->first;
416 while (node != WAKE_Q_TAIL) {
417 struct task_struct *task;
419 task = container_of(node, struct task_struct, wake_q);
421 /* task can safely be re-inserted now */
423 task->wake_q.next = NULL;
426 * wake_up_process() implies a wmb() to pair with the queueing
427 * in wake_q_add() so as not to miss wakeups.
429 wake_up_process(task);
430 put_task_struct(task);
435 * resched_curr - mark rq's current task 'to be rescheduled now'.
437 * On UP this means the setting of the need_resched flag, on SMP it
438 * might also involve a cross-CPU call to trigger the scheduler on
441 void resched_curr(struct rq *rq)
443 struct task_struct *curr = rq->curr;
446 lockdep_assert_held(&rq->lock);
448 if (test_tsk_need_resched(curr))
453 if (cpu == smp_processor_id()) {
454 set_tsk_need_resched(curr);
455 set_preempt_need_resched();
459 if (set_nr_and_not_polling(curr))
460 smp_send_reschedule(cpu);
462 trace_sched_wake_idle_without_ipi(cpu);
465 void resched_cpu(int cpu)
467 struct rq *rq = cpu_rq(cpu);
470 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
473 raw_spin_unlock_irqrestore(&rq->lock, flags);
477 #ifdef CONFIG_NO_HZ_COMMON
479 * In the semi idle case, use the nearest busy cpu for migrating timers
480 * from an idle cpu. This is good for power-savings.
482 * We don't do similar optimization for completely idle system, as
483 * selecting an idle cpu will add more delays to the timers than intended
484 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
486 int get_nohz_timer_target(void)
488 int i, cpu = smp_processor_id();
489 struct sched_domain *sd;
491 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
495 for_each_domain(cpu, sd) {
496 for_each_cpu(i, sched_domain_span(sd)) {
497 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
504 if (!is_housekeeping_cpu(cpu))
505 cpu = housekeeping_any_cpu();
511 * When add_timer_on() enqueues a timer into the timer wheel of an
512 * idle CPU then this timer might expire before the next timer event
513 * which is scheduled to wake up that CPU. In case of a completely
514 * idle system the next event might even be infinite time into the
515 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
516 * leaves the inner idle loop so the newly added timer is taken into
517 * account when the CPU goes back to idle and evaluates the timer
518 * wheel for the next timer event.
520 static void wake_up_idle_cpu(int cpu)
522 struct rq *rq = cpu_rq(cpu);
524 if (cpu == smp_processor_id())
527 if (set_nr_and_not_polling(rq->idle))
528 smp_send_reschedule(cpu);
530 trace_sched_wake_idle_without_ipi(cpu);
533 static bool wake_up_full_nohz_cpu(int cpu)
536 * We just need the target to call irq_exit() and re-evaluate
537 * the next tick. The nohz full kick at least implies that.
538 * If needed we can still optimize that later with an
541 if (tick_nohz_full_cpu(cpu)) {
542 if (cpu != smp_processor_id() ||
543 tick_nohz_tick_stopped())
544 tick_nohz_full_kick_cpu(cpu);
551 void wake_up_nohz_cpu(int cpu)
553 if (!wake_up_full_nohz_cpu(cpu))
554 wake_up_idle_cpu(cpu);
557 static inline bool got_nohz_idle_kick(void)
559 int cpu = smp_processor_id();
561 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
564 if (idle_cpu(cpu) && !need_resched())
568 * We can't run Idle Load Balance on this CPU for this time so we
569 * cancel it and clear NOHZ_BALANCE_KICK
571 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
575 #else /* CONFIG_NO_HZ_COMMON */
577 static inline bool got_nohz_idle_kick(void)
582 #endif /* CONFIG_NO_HZ_COMMON */
584 #ifdef CONFIG_NO_HZ_FULL
585 bool sched_can_stop_tick(void)
588 * FIFO realtime policy runs the highest priority task. Other runnable
589 * tasks are of a lower priority. The scheduler tick does nothing.
591 if (current->policy == SCHED_FIFO)
595 * Round-robin realtime tasks time slice with other tasks at the same
596 * realtime priority. Is this task the only one at this priority?
598 if (current->policy == SCHED_RR) {
599 struct sched_rt_entity *rt_se = ¤t->rt;
601 return list_is_singular(&rt_se->run_list);
605 * More than one running task need preemption.
606 * nr_running update is assumed to be visible
607 * after IPI is sent from wakers.
609 if (this_rq()->nr_running > 1)
614 #endif /* CONFIG_NO_HZ_FULL */
616 void sched_avg_update(struct rq *rq)
618 s64 period = sched_avg_period();
620 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
622 * Inline assembly required to prevent the compiler
623 * optimising this loop into a divmod call.
624 * See __iter_div_u64_rem() for another example of this.
626 asm("" : "+rm" (rq->age_stamp));
627 rq->age_stamp += period;
632 #endif /* CONFIG_SMP */
634 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
635 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
637 * Iterate task_group tree rooted at *from, calling @down when first entering a
638 * node and @up when leaving it for the final time.
640 * Caller must hold rcu_lock or sufficient equivalent.
642 int walk_tg_tree_from(struct task_group *from,
643 tg_visitor down, tg_visitor up, void *data)
645 struct task_group *parent, *child;
651 ret = (*down)(parent, data);
654 list_for_each_entry_rcu(child, &parent->children, siblings) {
661 ret = (*up)(parent, data);
662 if (ret || parent == from)
666 parent = parent->parent;
673 int tg_nop(struct task_group *tg, void *data)
679 static void set_load_weight(struct task_struct *p)
681 int prio = p->static_prio - MAX_RT_PRIO;
682 struct load_weight *load = &p->se.load;
685 * SCHED_IDLE tasks get minimal weight:
687 if (idle_policy(p->policy)) {
688 load->weight = scale_load(WEIGHT_IDLEPRIO);
689 load->inv_weight = WMULT_IDLEPRIO;
693 load->weight = scale_load(sched_prio_to_weight[prio]);
694 load->inv_weight = sched_prio_to_wmult[prio];
697 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
700 if (!(flags & ENQUEUE_RESTORE))
701 sched_info_queued(rq, p);
702 p->sched_class->enqueue_task(rq, p, flags);
705 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
708 if (!(flags & DEQUEUE_SAVE))
709 sched_info_dequeued(rq, p);
710 p->sched_class->dequeue_task(rq, p, flags);
713 void activate_task(struct rq *rq, struct task_struct *p, int flags)
715 if (task_contributes_to_load(p))
716 rq->nr_uninterruptible--;
718 enqueue_task(rq, p, flags);
721 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
723 if (task_contributes_to_load(p))
724 rq->nr_uninterruptible++;
726 dequeue_task(rq, p, flags);
729 static void update_rq_clock_task(struct rq *rq, s64 delta)
732 * In theory, the compile should just see 0 here, and optimize out the call
733 * to sched_rt_avg_update. But I don't trust it...
735 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
736 s64 steal = 0, irq_delta = 0;
738 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
739 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
742 * Since irq_time is only updated on {soft,}irq_exit, we might run into
743 * this case when a previous update_rq_clock() happened inside a
746 * When this happens, we stop ->clock_task and only update the
747 * prev_irq_time stamp to account for the part that fit, so that a next
748 * update will consume the rest. This ensures ->clock_task is
751 * It does however cause some slight miss-attribution of {soft,}irq
752 * time, a more accurate solution would be to update the irq_time using
753 * the current rq->clock timestamp, except that would require using
756 if (irq_delta > delta)
759 rq->prev_irq_time += irq_delta;
762 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
763 if (static_key_false((¶virt_steal_rq_enabled))) {
764 steal = paravirt_steal_clock(cpu_of(rq));
765 steal -= rq->prev_steal_time_rq;
767 if (unlikely(steal > delta))
770 rq->prev_steal_time_rq += steal;
775 rq->clock_task += delta;
777 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
778 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
779 sched_rt_avg_update(rq, irq_delta + steal);
783 void sched_set_stop_task(int cpu, struct task_struct *stop)
785 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
786 struct task_struct *old_stop = cpu_rq(cpu)->stop;
790 * Make it appear like a SCHED_FIFO task, its something
791 * userspace knows about and won't get confused about.
793 * Also, it will make PI more or less work without too
794 * much confusion -- but then, stop work should not
795 * rely on PI working anyway.
797 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
799 stop->sched_class = &stop_sched_class;
802 cpu_rq(cpu)->stop = stop;
806 * Reset it back to a normal scheduling class so that
807 * it can die in pieces.
809 old_stop->sched_class = &rt_sched_class;
814 * __normal_prio - return the priority that is based on the static prio
816 static inline int __normal_prio(struct task_struct *p)
818 return p->static_prio;
822 * Calculate the expected normal priority: i.e. priority
823 * without taking RT-inheritance into account. Might be
824 * boosted by interactivity modifiers. Changes upon fork,
825 * setprio syscalls, and whenever the interactivity
826 * estimator recalculates.
828 static inline int normal_prio(struct task_struct *p)
832 if (task_has_dl_policy(p))
833 prio = MAX_DL_PRIO-1;
834 else if (task_has_rt_policy(p))
835 prio = MAX_RT_PRIO-1 - p->rt_priority;
837 prio = __normal_prio(p);
842 * Calculate the current priority, i.e. the priority
843 * taken into account by the scheduler. This value might
844 * be boosted by RT tasks, or might be boosted by
845 * interactivity modifiers. Will be RT if the task got
846 * RT-boosted. If not then it returns p->normal_prio.
848 static int effective_prio(struct task_struct *p)
850 p->normal_prio = normal_prio(p);
852 * If we are RT tasks or we were boosted to RT priority,
853 * keep the priority unchanged. Otherwise, update priority
854 * to the normal priority:
856 if (!rt_prio(p->prio))
857 return p->normal_prio;
862 * task_curr - is this task currently executing on a CPU?
863 * @p: the task in question.
865 * Return: 1 if the task is currently executing. 0 otherwise.
867 inline int task_curr(const struct task_struct *p)
869 return cpu_curr(task_cpu(p)) == p;
873 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
874 * use the balance_callback list if you want balancing.
876 * this means any call to check_class_changed() must be followed by a call to
877 * balance_callback().
879 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
880 const struct sched_class *prev_class,
883 if (prev_class != p->sched_class) {
884 if (prev_class->switched_from)
885 prev_class->switched_from(rq, p);
887 p->sched_class->switched_to(rq, p);
888 } else if (oldprio != p->prio || dl_task(p))
889 p->sched_class->prio_changed(rq, p, oldprio);
892 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
894 const struct sched_class *class;
896 if (p->sched_class == rq->curr->sched_class) {
897 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
899 for_each_class(class) {
900 if (class == rq->curr->sched_class)
902 if (class == p->sched_class) {
910 * A queue event has occurred, and we're going to schedule. In
911 * this case, we can save a useless back to back clock update.
913 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
914 rq_clock_skip_update(rq, true);
919 * This is how migration works:
921 * 1) we invoke migration_cpu_stop() on the target CPU using
923 * 2) stopper starts to run (implicitly forcing the migrated thread
925 * 3) it checks whether the migrated task is still in the wrong runqueue.
926 * 4) if it's in the wrong runqueue then the migration thread removes
927 * it and puts it into the right queue.
928 * 5) stopper completes and stop_one_cpu() returns and the migration
933 * move_queued_task - move a queued task to new rq.
935 * Returns (locked) new rq. Old rq's lock is released.
937 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
939 lockdep_assert_held(&rq->lock);
941 p->on_rq = TASK_ON_RQ_MIGRATING;
942 dequeue_task(rq, p, 0);
943 set_task_cpu(p, new_cpu);
944 raw_spin_unlock(&rq->lock);
946 rq = cpu_rq(new_cpu);
948 raw_spin_lock(&rq->lock);
949 BUG_ON(task_cpu(p) != new_cpu);
950 enqueue_task(rq, p, 0);
951 p->on_rq = TASK_ON_RQ_QUEUED;
952 check_preempt_curr(rq, p, 0);
957 struct migration_arg {
958 struct task_struct *task;
963 * Move (not current) task off this cpu, onto dest cpu. We're doing
964 * this because either it can't run here any more (set_cpus_allowed()
965 * away from this CPU, or CPU going down), or because we're
966 * attempting to rebalance this task on exec (sched_exec).
968 * So we race with normal scheduler movements, but that's OK, as long
969 * as the task is no longer on this CPU.
971 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
973 if (unlikely(!cpu_active(dest_cpu)))
976 /* Affinity changed (again). */
977 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
980 rq = move_queued_task(rq, p, dest_cpu);
986 * migration_cpu_stop - this will be executed by a highprio stopper thread
987 * and performs thread migration by bumping thread off CPU then
988 * 'pushing' onto another runqueue.
990 static int migration_cpu_stop(void *data)
992 struct migration_arg *arg = data;
993 struct task_struct *p = arg->task;
994 struct rq *rq = this_rq();
997 * The original target cpu might have gone down and we might
998 * be on another cpu but it doesn't matter.
1000 local_irq_disable();
1002 * We need to explicitly wake pending tasks before running
1003 * __migrate_task() such that we will not miss enforcing cpus_allowed
1004 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1006 sched_ttwu_pending();
1008 raw_spin_lock(&p->pi_lock);
1009 raw_spin_lock(&rq->lock);
1011 * If task_rq(p) != rq, it cannot be migrated here, because we're
1012 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1013 * we're holding p->pi_lock.
1015 if (task_rq(p) == rq && task_on_rq_queued(p))
1016 rq = __migrate_task(rq, p, arg->dest_cpu);
1017 raw_spin_unlock(&rq->lock);
1018 raw_spin_unlock(&p->pi_lock);
1025 * sched_class::set_cpus_allowed must do the below, but is not required to
1026 * actually call this function.
1028 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1030 cpumask_copy(&p->cpus_allowed, new_mask);
1031 p->nr_cpus_allowed = cpumask_weight(new_mask);
1034 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1036 struct rq *rq = task_rq(p);
1037 bool queued, running;
1039 lockdep_assert_held(&p->pi_lock);
1041 queued = task_on_rq_queued(p);
1042 running = task_current(rq, p);
1046 * Because __kthread_bind() calls this on blocked tasks without
1049 lockdep_assert_held(&rq->lock);
1050 dequeue_task(rq, p, DEQUEUE_SAVE);
1053 put_prev_task(rq, p);
1055 p->sched_class->set_cpus_allowed(p, new_mask);
1058 p->sched_class->set_curr_task(rq);
1060 enqueue_task(rq, p, ENQUEUE_RESTORE);
1064 * Change a given task's CPU affinity. Migrate the thread to a
1065 * proper CPU and schedule it away if the CPU it's executing on
1066 * is removed from the allowed bitmask.
1068 * NOTE: the caller must have a valid reference to the task, the
1069 * task must not exit() & deallocate itself prematurely. The
1070 * call is not atomic; no spinlocks may be held.
1072 static int __set_cpus_allowed_ptr(struct task_struct *p,
1073 const struct cpumask *new_mask, bool check)
1075 unsigned long flags;
1077 unsigned int dest_cpu;
1080 rq = task_rq_lock(p, &flags);
1083 * Must re-check here, to close a race against __kthread_bind(),
1084 * sched_setaffinity() is not guaranteed to observe the flag.
1086 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1091 if (cpumask_equal(&p->cpus_allowed, new_mask))
1094 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1099 do_set_cpus_allowed(p, new_mask);
1101 /* Can the task run on the task's current CPU? If so, we're done */
1102 if (cpumask_test_cpu(task_cpu(p), new_mask))
1105 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1106 if (task_running(rq, p) || p->state == TASK_WAKING) {
1107 struct migration_arg arg = { p, dest_cpu };
1108 /* Need help from migration thread: drop lock and wait. */
1109 task_rq_unlock(rq, p, &flags);
1110 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1111 tlb_migrate_finish(p->mm);
1113 } else if (task_on_rq_queued(p)) {
1115 * OK, since we're going to drop the lock immediately
1116 * afterwards anyway.
1118 lockdep_unpin_lock(&rq->lock);
1119 rq = move_queued_task(rq, p, dest_cpu);
1120 lockdep_pin_lock(&rq->lock);
1123 task_rq_unlock(rq, p, &flags);
1128 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1130 return __set_cpus_allowed_ptr(p, new_mask, false);
1132 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1134 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1136 #ifdef CONFIG_SCHED_DEBUG
1138 * We should never call set_task_cpu() on a blocked task,
1139 * ttwu() will sort out the placement.
1141 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1145 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1146 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1147 * time relying on p->on_rq.
1149 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1150 p->sched_class == &fair_sched_class &&
1151 (p->on_rq && !task_on_rq_migrating(p)));
1153 #ifdef CONFIG_LOCKDEP
1155 * The caller should hold either p->pi_lock or rq->lock, when changing
1156 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1158 * sched_move_task() holds both and thus holding either pins the cgroup,
1161 * Furthermore, all task_rq users should acquire both locks, see
1164 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1165 lockdep_is_held(&task_rq(p)->lock)));
1169 trace_sched_migrate_task(p, new_cpu);
1171 if (task_cpu(p) != new_cpu) {
1172 if (p->sched_class->migrate_task_rq)
1173 p->sched_class->migrate_task_rq(p);
1174 p->se.nr_migrations++;
1175 perf_event_task_migrate(p);
1178 __set_task_cpu(p, new_cpu);
1181 static void __migrate_swap_task(struct task_struct *p, int cpu)
1183 if (task_on_rq_queued(p)) {
1184 struct rq *src_rq, *dst_rq;
1186 src_rq = task_rq(p);
1187 dst_rq = cpu_rq(cpu);
1189 p->on_rq = TASK_ON_RQ_MIGRATING;
1190 deactivate_task(src_rq, p, 0);
1191 set_task_cpu(p, cpu);
1192 activate_task(dst_rq, p, 0);
1193 p->on_rq = TASK_ON_RQ_QUEUED;
1194 check_preempt_curr(dst_rq, p, 0);
1197 * Task isn't running anymore; make it appear like we migrated
1198 * it before it went to sleep. This means on wakeup we make the
1199 * previous cpu our targer instead of where it really is.
1205 struct migration_swap_arg {
1206 struct task_struct *src_task, *dst_task;
1207 int src_cpu, dst_cpu;
1210 static int migrate_swap_stop(void *data)
1212 struct migration_swap_arg *arg = data;
1213 struct rq *src_rq, *dst_rq;
1216 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1219 src_rq = cpu_rq(arg->src_cpu);
1220 dst_rq = cpu_rq(arg->dst_cpu);
1222 double_raw_lock(&arg->src_task->pi_lock,
1223 &arg->dst_task->pi_lock);
1224 double_rq_lock(src_rq, dst_rq);
1226 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1229 if (task_cpu(arg->src_task) != arg->src_cpu)
1232 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1235 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1238 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1239 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1244 double_rq_unlock(src_rq, dst_rq);
1245 raw_spin_unlock(&arg->dst_task->pi_lock);
1246 raw_spin_unlock(&arg->src_task->pi_lock);
1252 * Cross migrate two tasks
1254 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1256 struct migration_swap_arg arg;
1259 arg = (struct migration_swap_arg){
1261 .src_cpu = task_cpu(cur),
1263 .dst_cpu = task_cpu(p),
1266 if (arg.src_cpu == arg.dst_cpu)
1270 * These three tests are all lockless; this is OK since all of them
1271 * will be re-checked with proper locks held further down the line.
1273 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1276 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1279 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1282 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1283 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1290 * wait_task_inactive - wait for a thread to unschedule.
1292 * If @match_state is nonzero, it's the @p->state value just checked and
1293 * not expected to change. If it changes, i.e. @p might have woken up,
1294 * then return zero. When we succeed in waiting for @p to be off its CPU,
1295 * we return a positive number (its total switch count). If a second call
1296 * a short while later returns the same number, the caller can be sure that
1297 * @p has remained unscheduled the whole time.
1299 * The caller must ensure that the task *will* unschedule sometime soon,
1300 * else this function might spin for a *long* time. This function can't
1301 * be called with interrupts off, or it may introduce deadlock with
1302 * smp_call_function() if an IPI is sent by the same process we are
1303 * waiting to become inactive.
1305 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1307 unsigned long flags;
1308 int running, queued;
1314 * We do the initial early heuristics without holding
1315 * any task-queue locks at all. We'll only try to get
1316 * the runqueue lock when things look like they will
1322 * If the task is actively running on another CPU
1323 * still, just relax and busy-wait without holding
1326 * NOTE! Since we don't hold any locks, it's not
1327 * even sure that "rq" stays as the right runqueue!
1328 * But we don't care, since "task_running()" will
1329 * return false if the runqueue has changed and p
1330 * is actually now running somewhere else!
1332 while (task_running(rq, p)) {
1333 if (match_state && unlikely(p->state != match_state))
1339 * Ok, time to look more closely! We need the rq
1340 * lock now, to be *sure*. If we're wrong, we'll
1341 * just go back and repeat.
1343 rq = task_rq_lock(p, &flags);
1344 trace_sched_wait_task(p);
1345 running = task_running(rq, p);
1346 queued = task_on_rq_queued(p);
1348 if (!match_state || p->state == match_state)
1349 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1350 task_rq_unlock(rq, p, &flags);
1353 * If it changed from the expected state, bail out now.
1355 if (unlikely(!ncsw))
1359 * Was it really running after all now that we
1360 * checked with the proper locks actually held?
1362 * Oops. Go back and try again..
1364 if (unlikely(running)) {
1370 * It's not enough that it's not actively running,
1371 * it must be off the runqueue _entirely_, and not
1374 * So if it was still runnable (but just not actively
1375 * running right now), it's preempted, and we should
1376 * yield - it could be a while.
1378 if (unlikely(queued)) {
1379 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1381 set_current_state(TASK_UNINTERRUPTIBLE);
1382 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1387 * Ahh, all good. It wasn't running, and it wasn't
1388 * runnable, which means that it will never become
1389 * running in the future either. We're all done!
1398 * kick_process - kick a running thread to enter/exit the kernel
1399 * @p: the to-be-kicked thread
1401 * Cause a process which is running on another CPU to enter
1402 * kernel-mode, without any delay. (to get signals handled.)
1404 * NOTE: this function doesn't have to take the runqueue lock,
1405 * because all it wants to ensure is that the remote task enters
1406 * the kernel. If the IPI races and the task has been migrated
1407 * to another CPU then no harm is done and the purpose has been
1410 void kick_process(struct task_struct *p)
1416 if ((cpu != smp_processor_id()) && task_curr(p))
1417 smp_send_reschedule(cpu);
1420 EXPORT_SYMBOL_GPL(kick_process);
1423 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1425 static int select_fallback_rq(int cpu, struct task_struct *p)
1427 int nid = cpu_to_node(cpu);
1428 const struct cpumask *nodemask = NULL;
1429 enum { cpuset, possible, fail } state = cpuset;
1433 * If the node that the cpu is on has been offlined, cpu_to_node()
1434 * will return -1. There is no cpu on the node, and we should
1435 * select the cpu on the other node.
1438 nodemask = cpumask_of_node(nid);
1440 /* Look for allowed, online CPU in same node. */
1441 for_each_cpu(dest_cpu, nodemask) {
1442 if (!cpu_online(dest_cpu))
1444 if (!cpu_active(dest_cpu))
1446 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1452 /* Any allowed, online CPU? */
1453 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1454 if (!cpu_online(dest_cpu))
1456 if (!cpu_active(dest_cpu))
1461 /* No more Mr. Nice Guy. */
1464 if (IS_ENABLED(CONFIG_CPUSETS)) {
1465 cpuset_cpus_allowed_fallback(p);
1471 do_set_cpus_allowed(p, cpu_possible_mask);
1482 if (state != cpuset) {
1484 * Don't tell them about moving exiting tasks or
1485 * kernel threads (both mm NULL), since they never
1488 if (p->mm && printk_ratelimit()) {
1489 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1490 task_pid_nr(p), p->comm, cpu);
1498 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1501 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1503 lockdep_assert_held(&p->pi_lock);
1505 if (p->nr_cpus_allowed > 1)
1506 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1509 * In order not to call set_task_cpu() on a blocking task we need
1510 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1513 * Since this is common to all placement strategies, this lives here.
1515 * [ this allows ->select_task() to simply return task_cpu(p) and
1516 * not worry about this generic constraint ]
1518 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1520 cpu = select_fallback_rq(task_cpu(p), p);
1525 static void update_avg(u64 *avg, u64 sample)
1527 s64 diff = sample - *avg;
1533 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1534 const struct cpumask *new_mask, bool check)
1536 return set_cpus_allowed_ptr(p, new_mask);
1539 #endif /* CONFIG_SMP */
1542 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1544 #ifdef CONFIG_SCHEDSTATS
1545 struct rq *rq = this_rq();
1548 int this_cpu = smp_processor_id();
1550 if (cpu == this_cpu) {
1551 schedstat_inc(rq, ttwu_local);
1552 schedstat_inc(p, se.statistics.nr_wakeups_local);
1554 struct sched_domain *sd;
1556 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1558 for_each_domain(this_cpu, sd) {
1559 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1560 schedstat_inc(sd, ttwu_wake_remote);
1567 if (wake_flags & WF_MIGRATED)
1568 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1570 #endif /* CONFIG_SMP */
1572 schedstat_inc(rq, ttwu_count);
1573 schedstat_inc(p, se.statistics.nr_wakeups);
1575 if (wake_flags & WF_SYNC)
1576 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1578 #endif /* CONFIG_SCHEDSTATS */
1581 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1583 activate_task(rq, p, en_flags);
1584 p->on_rq = TASK_ON_RQ_QUEUED;
1586 /* if a worker is waking up, notify workqueue */
1587 if (p->flags & PF_WQ_WORKER)
1588 wq_worker_waking_up(p, cpu_of(rq));
1592 * Mark the task runnable and perform wakeup-preemption.
1595 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1597 check_preempt_curr(rq, p, wake_flags);
1598 p->state = TASK_RUNNING;
1599 trace_sched_wakeup(p);
1602 if (p->sched_class->task_woken) {
1604 * Our task @p is fully woken up and running; so its safe to
1605 * drop the rq->lock, hereafter rq is only used for statistics.
1607 lockdep_unpin_lock(&rq->lock);
1608 p->sched_class->task_woken(rq, p);
1609 lockdep_pin_lock(&rq->lock);
1612 if (rq->idle_stamp) {
1613 u64 delta = rq_clock(rq) - rq->idle_stamp;
1614 u64 max = 2*rq->max_idle_balance_cost;
1616 update_avg(&rq->avg_idle, delta);
1618 if (rq->avg_idle > max)
1627 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1629 lockdep_assert_held(&rq->lock);
1632 if (p->sched_contributes_to_load)
1633 rq->nr_uninterruptible--;
1636 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1637 ttwu_do_wakeup(rq, p, wake_flags);
1641 * Called in case the task @p isn't fully descheduled from its runqueue,
1642 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1643 * since all we need to do is flip p->state to TASK_RUNNING, since
1644 * the task is still ->on_rq.
1646 static int ttwu_remote(struct task_struct *p, int wake_flags)
1651 rq = __task_rq_lock(p);
1652 if (task_on_rq_queued(p)) {
1653 /* check_preempt_curr() may use rq clock */
1654 update_rq_clock(rq);
1655 ttwu_do_wakeup(rq, p, wake_flags);
1658 __task_rq_unlock(rq);
1664 void sched_ttwu_pending(void)
1666 struct rq *rq = this_rq();
1667 struct llist_node *llist = llist_del_all(&rq->wake_list);
1668 struct task_struct *p;
1669 unsigned long flags;
1674 raw_spin_lock_irqsave(&rq->lock, flags);
1675 lockdep_pin_lock(&rq->lock);
1678 p = llist_entry(llist, struct task_struct, wake_entry);
1679 llist = llist_next(llist);
1680 ttwu_do_activate(rq, p, 0);
1683 lockdep_unpin_lock(&rq->lock);
1684 raw_spin_unlock_irqrestore(&rq->lock, flags);
1687 void scheduler_ipi(void)
1690 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1691 * TIF_NEED_RESCHED remotely (for the first time) will also send
1694 preempt_fold_need_resched();
1696 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1700 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1701 * traditionally all their work was done from the interrupt return
1702 * path. Now that we actually do some work, we need to make sure
1705 * Some archs already do call them, luckily irq_enter/exit nest
1708 * Arguably we should visit all archs and update all handlers,
1709 * however a fair share of IPIs are still resched only so this would
1710 * somewhat pessimize the simple resched case.
1713 sched_ttwu_pending();
1716 * Check if someone kicked us for doing the nohz idle load balance.
1718 if (unlikely(got_nohz_idle_kick())) {
1719 this_rq()->idle_balance = 1;
1720 raise_softirq_irqoff(SCHED_SOFTIRQ);
1725 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1727 struct rq *rq = cpu_rq(cpu);
1729 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1730 if (!set_nr_if_polling(rq->idle))
1731 smp_send_reschedule(cpu);
1733 trace_sched_wake_idle_without_ipi(cpu);
1737 void wake_up_if_idle(int cpu)
1739 struct rq *rq = cpu_rq(cpu);
1740 unsigned long flags;
1744 if (!is_idle_task(rcu_dereference(rq->curr)))
1747 if (set_nr_if_polling(rq->idle)) {
1748 trace_sched_wake_idle_without_ipi(cpu);
1750 raw_spin_lock_irqsave(&rq->lock, flags);
1751 if (is_idle_task(rq->curr))
1752 smp_send_reschedule(cpu);
1753 /* Else cpu is not in idle, do nothing here */
1754 raw_spin_unlock_irqrestore(&rq->lock, flags);
1761 bool cpus_share_cache(int this_cpu, int that_cpu)
1763 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1765 #endif /* CONFIG_SMP */
1767 static void ttwu_queue(struct task_struct *p, int cpu)
1769 struct rq *rq = cpu_rq(cpu);
1771 #if defined(CONFIG_SMP)
1772 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1773 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1774 ttwu_queue_remote(p, cpu);
1779 raw_spin_lock(&rq->lock);
1780 lockdep_pin_lock(&rq->lock);
1781 ttwu_do_activate(rq, p, 0);
1782 lockdep_unpin_lock(&rq->lock);
1783 raw_spin_unlock(&rq->lock);
1787 * Notes on Program-Order guarantees on SMP systems.
1791 * The basic program-order guarantee on SMP systems is that when a task [t]
1792 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1793 * execution on its new cpu [c1].
1795 * For migration (of runnable tasks) this is provided by the following means:
1797 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1798 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1799 * rq(c1)->lock (if not at the same time, then in that order).
1800 * C) LOCK of the rq(c1)->lock scheduling in task
1802 * Transitivity guarantees that B happens after A and C after B.
1803 * Note: we only require RCpc transitivity.
1804 * Note: the cpu doing B need not be c0 or c1
1813 * UNLOCK rq(0)->lock
1815 * LOCK rq(0)->lock // orders against CPU0
1817 * UNLOCK rq(0)->lock
1821 * UNLOCK rq(1)->lock
1823 * LOCK rq(1)->lock // orders against CPU2
1826 * UNLOCK rq(1)->lock
1829 * BLOCKING -- aka. SLEEP + WAKEUP
1831 * For blocking we (obviously) need to provide the same guarantee as for
1832 * migration. However the means are completely different as there is no lock
1833 * chain to provide order. Instead we do:
1835 * 1) smp_store_release(X->on_cpu, 0)
1836 * 2) smp_cond_acquire(!X->on_cpu)
1840 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1842 * LOCK rq(0)->lock LOCK X->pi_lock
1845 * smp_store_release(X->on_cpu, 0);
1847 * smp_cond_acquire(!X->on_cpu);
1853 * X->state = RUNNING
1854 * UNLOCK rq(2)->lock
1856 * LOCK rq(2)->lock // orders against CPU1
1859 * UNLOCK rq(2)->lock
1862 * UNLOCK rq(0)->lock
1865 * However; for wakeups there is a second guarantee we must provide, namely we
1866 * must observe the state that lead to our wakeup. That is, not only must our
1867 * task observe its own prior state, it must also observe the stores prior to
1870 * This means that any means of doing remote wakeups must order the CPU doing
1871 * the wakeup against the CPU the task is going to end up running on. This,
1872 * however, is already required for the regular Program-Order guarantee above,
1873 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_acquire).
1878 * try_to_wake_up - wake up a thread
1879 * @p: the thread to be awakened
1880 * @state: the mask of task states that can be woken
1881 * @wake_flags: wake modifier flags (WF_*)
1883 * Put it on the run-queue if it's not already there. The "current"
1884 * thread is always on the run-queue (except when the actual
1885 * re-schedule is in progress), and as such you're allowed to do
1886 * the simpler "current->state = TASK_RUNNING" to mark yourself
1887 * runnable without the overhead of this.
1889 * Return: %true if @p was woken up, %false if it was already running.
1890 * or @state didn't match @p's state.
1893 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1895 unsigned long flags;
1896 int cpu, success = 0;
1899 * If we are going to wake up a thread waiting for CONDITION we
1900 * need to ensure that CONDITION=1 done by the caller can not be
1901 * reordered with p->state check below. This pairs with mb() in
1902 * set_current_state() the waiting thread does.
1904 smp_mb__before_spinlock();
1905 raw_spin_lock_irqsave(&p->pi_lock, flags);
1906 if (!(p->state & state))
1909 trace_sched_waking(p);
1911 success = 1; /* we're going to change ->state */
1914 if (p->on_rq && ttwu_remote(p, wake_flags))
1919 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1920 * possible to, falsely, observe p->on_cpu == 0.
1922 * One must be running (->on_cpu == 1) in order to remove oneself
1923 * from the runqueue.
1925 * [S] ->on_cpu = 1; [L] ->on_rq
1929 * [S] ->on_rq = 0; [L] ->on_cpu
1931 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1932 * from the consecutive calls to schedule(); the first switching to our
1933 * task, the second putting it to sleep.
1938 * If the owning (remote) cpu is still in the middle of schedule() with
1939 * this task as prev, wait until its done referencing the task.
1941 * Pairs with the smp_store_release() in finish_lock_switch().
1943 * This ensures that tasks getting woken will be fully ordered against
1944 * their previous state and preserve Program Order.
1946 smp_cond_acquire(!p->on_cpu);
1948 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1949 p->state = TASK_WAKING;
1951 if (p->sched_class->task_waking)
1952 p->sched_class->task_waking(p);
1954 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1955 if (task_cpu(p) != cpu) {
1956 wake_flags |= WF_MIGRATED;
1957 set_task_cpu(p, cpu);
1959 #endif /* CONFIG_SMP */
1963 if (schedstat_enabled())
1964 ttwu_stat(p, cpu, wake_flags);
1966 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1972 * try_to_wake_up_local - try to wake up a local task with rq lock held
1973 * @p: the thread to be awakened
1975 * Put @p on the run-queue if it's not already there. The caller must
1976 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1979 static void try_to_wake_up_local(struct task_struct *p)
1981 struct rq *rq = task_rq(p);
1983 if (WARN_ON_ONCE(rq != this_rq()) ||
1984 WARN_ON_ONCE(p == current))
1987 lockdep_assert_held(&rq->lock);
1989 if (!raw_spin_trylock(&p->pi_lock)) {
1991 * This is OK, because current is on_cpu, which avoids it being
1992 * picked for load-balance and preemption/IRQs are still
1993 * disabled avoiding further scheduler activity on it and we've
1994 * not yet picked a replacement task.
1996 lockdep_unpin_lock(&rq->lock);
1997 raw_spin_unlock(&rq->lock);
1998 raw_spin_lock(&p->pi_lock);
1999 raw_spin_lock(&rq->lock);
2000 lockdep_pin_lock(&rq->lock);
2003 if (!(p->state & TASK_NORMAL))
2006 trace_sched_waking(p);
2008 if (!task_on_rq_queued(p))
2009 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2011 ttwu_do_wakeup(rq, p, 0);
2012 if (schedstat_enabled())
2013 ttwu_stat(p, smp_processor_id(), 0);
2015 raw_spin_unlock(&p->pi_lock);
2019 * wake_up_process - Wake up a specific process
2020 * @p: The process to be woken up.
2022 * Attempt to wake up the nominated process and move it to the set of runnable
2025 * Return: 1 if the process was woken up, 0 if it was already running.
2027 * It may be assumed that this function implies a write memory barrier before
2028 * changing the task state if and only if any tasks are woken up.
2030 int wake_up_process(struct task_struct *p)
2032 return try_to_wake_up(p, TASK_NORMAL, 0);
2034 EXPORT_SYMBOL(wake_up_process);
2036 int wake_up_state(struct task_struct *p, unsigned int state)
2038 return try_to_wake_up(p, state, 0);
2042 * This function clears the sched_dl_entity static params.
2044 void __dl_clear_params(struct task_struct *p)
2046 struct sched_dl_entity *dl_se = &p->dl;
2048 dl_se->dl_runtime = 0;
2049 dl_se->dl_deadline = 0;
2050 dl_se->dl_period = 0;
2054 dl_se->dl_throttled = 0;
2056 dl_se->dl_yielded = 0;
2060 * Perform scheduler related setup for a newly forked process p.
2061 * p is forked by current.
2063 * __sched_fork() is basic setup used by init_idle() too:
2065 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2070 p->se.exec_start = 0;
2071 p->se.sum_exec_runtime = 0;
2072 p->se.prev_sum_exec_runtime = 0;
2073 p->se.nr_migrations = 0;
2075 INIT_LIST_HEAD(&p->se.group_node);
2077 #ifdef CONFIG_FAIR_GROUP_SCHED
2078 p->se.cfs_rq = NULL;
2081 #ifdef CONFIG_SCHEDSTATS
2082 /* Even if schedstat is disabled, there should not be garbage */
2083 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2086 RB_CLEAR_NODE(&p->dl.rb_node);
2087 init_dl_task_timer(&p->dl);
2088 __dl_clear_params(p);
2090 INIT_LIST_HEAD(&p->rt.run_list);
2092 p->rt.time_slice = sched_rr_timeslice;
2096 #ifdef CONFIG_PREEMPT_NOTIFIERS
2097 INIT_HLIST_HEAD(&p->preempt_notifiers);
2100 #ifdef CONFIG_NUMA_BALANCING
2101 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2102 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2103 p->mm->numa_scan_seq = 0;
2106 if (clone_flags & CLONE_VM)
2107 p->numa_preferred_nid = current->numa_preferred_nid;
2109 p->numa_preferred_nid = -1;
2111 p->node_stamp = 0ULL;
2112 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2113 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2114 p->numa_work.next = &p->numa_work;
2115 p->numa_faults = NULL;
2116 p->last_task_numa_placement = 0;
2117 p->last_sum_exec_runtime = 0;
2119 p->numa_group = NULL;
2120 #endif /* CONFIG_NUMA_BALANCING */
2123 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2125 #ifdef CONFIG_NUMA_BALANCING
2127 void set_numabalancing_state(bool enabled)
2130 static_branch_enable(&sched_numa_balancing);
2132 static_branch_disable(&sched_numa_balancing);
2135 #ifdef CONFIG_PROC_SYSCTL
2136 int sysctl_numa_balancing(struct ctl_table *table, int write,
2137 void __user *buffer, size_t *lenp, loff_t *ppos)
2141 int state = static_branch_likely(&sched_numa_balancing);
2143 if (write && !capable(CAP_SYS_ADMIN))
2148 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2152 set_numabalancing_state(state);
2158 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2160 #ifdef CONFIG_SCHEDSTATS
2161 static void set_schedstats(bool enabled)
2164 static_branch_enable(&sched_schedstats);
2166 static_branch_disable(&sched_schedstats);
2169 void force_schedstat_enabled(void)
2171 if (!schedstat_enabled()) {
2172 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2173 static_branch_enable(&sched_schedstats);
2177 static int __init setup_schedstats(char *str)
2183 if (!strcmp(str, "enable")) {
2184 set_schedstats(true);
2186 } else if (!strcmp(str, "disable")) {
2187 set_schedstats(false);
2192 pr_warn("Unable to parse schedstats=\n");
2196 __setup("schedstats=", setup_schedstats);
2198 #ifdef CONFIG_PROC_SYSCTL
2199 int sysctl_schedstats(struct ctl_table *table, int write,
2200 void __user *buffer, size_t *lenp, loff_t *ppos)
2204 int state = static_branch_likely(&sched_schedstats);
2206 if (write && !capable(CAP_SYS_ADMIN))
2211 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2215 set_schedstats(state);
2222 * fork()/clone()-time setup:
2224 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2226 unsigned long flags;
2227 int cpu = get_cpu();
2229 __sched_fork(clone_flags, p);
2231 * We mark the process as running here. This guarantees that
2232 * nobody will actually run it, and a signal or other external
2233 * event cannot wake it up and insert it on the runqueue either.
2235 p->state = TASK_RUNNING;
2238 * Make sure we do not leak PI boosting priority to the child.
2240 p->prio = current->normal_prio;
2243 * Revert to default priority/policy on fork if requested.
2245 if (unlikely(p->sched_reset_on_fork)) {
2246 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2247 p->policy = SCHED_NORMAL;
2248 p->static_prio = NICE_TO_PRIO(0);
2250 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2251 p->static_prio = NICE_TO_PRIO(0);
2253 p->prio = p->normal_prio = __normal_prio(p);
2257 * We don't need the reset flag anymore after the fork. It has
2258 * fulfilled its duty:
2260 p->sched_reset_on_fork = 0;
2263 if (dl_prio(p->prio)) {
2266 } else if (rt_prio(p->prio)) {
2267 p->sched_class = &rt_sched_class;
2269 p->sched_class = &fair_sched_class;
2272 if (p->sched_class->task_fork)
2273 p->sched_class->task_fork(p);
2276 * The child is not yet in the pid-hash so no cgroup attach races,
2277 * and the cgroup is pinned to this child due to cgroup_fork()
2278 * is ran before sched_fork().
2280 * Silence PROVE_RCU.
2282 raw_spin_lock_irqsave(&p->pi_lock, flags);
2283 set_task_cpu(p, cpu);
2284 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2286 #ifdef CONFIG_SCHED_INFO
2287 if (likely(sched_info_on()))
2288 memset(&p->sched_info, 0, sizeof(p->sched_info));
2290 #if defined(CONFIG_SMP)
2293 init_task_preempt_count(p);
2295 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2296 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2303 unsigned long to_ratio(u64 period, u64 runtime)
2305 if (runtime == RUNTIME_INF)
2309 * Doing this here saves a lot of checks in all
2310 * the calling paths, and returning zero seems
2311 * safe for them anyway.
2316 return div64_u64(runtime << 20, period);
2320 inline struct dl_bw *dl_bw_of(int i)
2322 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2323 "sched RCU must be held");
2324 return &cpu_rq(i)->rd->dl_bw;
2327 static inline int dl_bw_cpus(int i)
2329 struct root_domain *rd = cpu_rq(i)->rd;
2332 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2333 "sched RCU must be held");
2334 for_each_cpu_and(i, rd->span, cpu_active_mask)
2340 inline struct dl_bw *dl_bw_of(int i)
2342 return &cpu_rq(i)->dl.dl_bw;
2345 static inline int dl_bw_cpus(int i)
2352 * We must be sure that accepting a new task (or allowing changing the
2353 * parameters of an existing one) is consistent with the bandwidth
2354 * constraints. If yes, this function also accordingly updates the currently
2355 * allocated bandwidth to reflect the new situation.
2357 * This function is called while holding p's rq->lock.
2359 * XXX we should delay bw change until the task's 0-lag point, see
2362 static int dl_overflow(struct task_struct *p, int policy,
2363 const struct sched_attr *attr)
2366 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2367 u64 period = attr->sched_period ?: attr->sched_deadline;
2368 u64 runtime = attr->sched_runtime;
2369 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2372 if (new_bw == p->dl.dl_bw)
2376 * Either if a task, enters, leave, or stays -deadline but changes
2377 * its parameters, we may need to update accordingly the total
2378 * allocated bandwidth of the container.
2380 raw_spin_lock(&dl_b->lock);
2381 cpus = dl_bw_cpus(task_cpu(p));
2382 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2383 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2384 __dl_add(dl_b, new_bw);
2386 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2387 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2388 __dl_clear(dl_b, p->dl.dl_bw);
2389 __dl_add(dl_b, new_bw);
2391 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2392 __dl_clear(dl_b, p->dl.dl_bw);
2395 raw_spin_unlock(&dl_b->lock);
2400 extern void init_dl_bw(struct dl_bw *dl_b);
2403 * wake_up_new_task - wake up a newly created task for the first time.
2405 * This function will do some initial scheduler statistics housekeeping
2406 * that must be done for every newly created context, then puts the task
2407 * on the runqueue and wakes it.
2409 void wake_up_new_task(struct task_struct *p)
2411 unsigned long flags;
2414 raw_spin_lock_irqsave(&p->pi_lock, flags);
2415 /* Initialize new task's runnable average */
2416 init_entity_runnable_average(&p->se);
2419 * Fork balancing, do it here and not earlier because:
2420 * - cpus_allowed can change in the fork path
2421 * - any previously selected cpu might disappear through hotplug
2423 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2426 rq = __task_rq_lock(p);
2427 activate_task(rq, p, 0);
2428 p->on_rq = TASK_ON_RQ_QUEUED;
2429 trace_sched_wakeup_new(p);
2430 check_preempt_curr(rq, p, WF_FORK);
2432 if (p->sched_class->task_woken) {
2434 * Nothing relies on rq->lock after this, so its fine to
2437 lockdep_unpin_lock(&rq->lock);
2438 p->sched_class->task_woken(rq, p);
2439 lockdep_pin_lock(&rq->lock);
2442 task_rq_unlock(rq, p, &flags);
2445 #ifdef CONFIG_PREEMPT_NOTIFIERS
2447 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2449 void preempt_notifier_inc(void)
2451 static_key_slow_inc(&preempt_notifier_key);
2453 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2455 void preempt_notifier_dec(void)
2457 static_key_slow_dec(&preempt_notifier_key);
2459 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2462 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2463 * @notifier: notifier struct to register
2465 void preempt_notifier_register(struct preempt_notifier *notifier)
2467 if (!static_key_false(&preempt_notifier_key))
2468 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2470 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2472 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2475 * preempt_notifier_unregister - no longer interested in preemption notifications
2476 * @notifier: notifier struct to unregister
2478 * This is *not* safe to call from within a preemption notifier.
2480 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2482 hlist_del(¬ifier->link);
2484 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2486 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2488 struct preempt_notifier *notifier;
2490 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2491 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2494 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2496 if (static_key_false(&preempt_notifier_key))
2497 __fire_sched_in_preempt_notifiers(curr);
2501 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2502 struct task_struct *next)
2504 struct preempt_notifier *notifier;
2506 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2507 notifier->ops->sched_out(notifier, next);
2510 static __always_inline void
2511 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2512 struct task_struct *next)
2514 if (static_key_false(&preempt_notifier_key))
2515 __fire_sched_out_preempt_notifiers(curr, next);
2518 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2520 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2525 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2526 struct task_struct *next)
2530 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2533 * prepare_task_switch - prepare to switch tasks
2534 * @rq: the runqueue preparing to switch
2535 * @prev: the current task that is being switched out
2536 * @next: the task we are going to switch to.
2538 * This is called with the rq lock held and interrupts off. It must
2539 * be paired with a subsequent finish_task_switch after the context
2542 * prepare_task_switch sets up locking and calls architecture specific
2546 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2547 struct task_struct *next)
2549 sched_info_switch(rq, prev, next);
2550 perf_event_task_sched_out(prev, next);
2551 fire_sched_out_preempt_notifiers(prev, next);
2552 prepare_lock_switch(rq, next);
2553 prepare_arch_switch(next);
2557 * finish_task_switch - clean up after a task-switch
2558 * @prev: the thread we just switched away from.
2560 * finish_task_switch must be called after the context switch, paired
2561 * with a prepare_task_switch call before the context switch.
2562 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2563 * and do any other architecture-specific cleanup actions.
2565 * Note that we may have delayed dropping an mm in context_switch(). If
2566 * so, we finish that here outside of the runqueue lock. (Doing it
2567 * with the lock held can cause deadlocks; see schedule() for
2570 * The context switch have flipped the stack from under us and restored the
2571 * local variables which were saved when this task called schedule() in the
2572 * past. prev == current is still correct but we need to recalculate this_rq
2573 * because prev may have moved to another CPU.
2575 static struct rq *finish_task_switch(struct task_struct *prev)
2576 __releases(rq->lock)
2578 struct rq *rq = this_rq();
2579 struct mm_struct *mm = rq->prev_mm;
2583 * The previous task will have left us with a preempt_count of 2
2584 * because it left us after:
2587 * preempt_disable(); // 1
2589 * raw_spin_lock_irq(&rq->lock) // 2
2591 * Also, see FORK_PREEMPT_COUNT.
2593 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2594 "corrupted preempt_count: %s/%d/0x%x\n",
2595 current->comm, current->pid, preempt_count()))
2596 preempt_count_set(FORK_PREEMPT_COUNT);
2601 * A task struct has one reference for the use as "current".
2602 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2603 * schedule one last time. The schedule call will never return, and
2604 * the scheduled task must drop that reference.
2606 * We must observe prev->state before clearing prev->on_cpu (in
2607 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2608 * running on another CPU and we could rave with its RUNNING -> DEAD
2609 * transition, resulting in a double drop.
2611 prev_state = prev->state;
2612 vtime_task_switch(prev);
2613 perf_event_task_sched_in(prev, current);
2614 finish_lock_switch(rq, prev);
2615 finish_arch_post_lock_switch();
2617 fire_sched_in_preempt_notifiers(current);
2620 if (unlikely(prev_state == TASK_DEAD)) {
2621 if (prev->sched_class->task_dead)
2622 prev->sched_class->task_dead(prev);
2625 * Remove function-return probe instances associated with this
2626 * task and put them back on the free list.
2628 kprobe_flush_task(prev);
2629 put_task_struct(prev);
2632 tick_nohz_task_switch();
2638 /* rq->lock is NOT held, but preemption is disabled */
2639 static void __balance_callback(struct rq *rq)
2641 struct callback_head *head, *next;
2642 void (*func)(struct rq *rq);
2643 unsigned long flags;
2645 raw_spin_lock_irqsave(&rq->lock, flags);
2646 head = rq->balance_callback;
2647 rq->balance_callback = NULL;
2649 func = (void (*)(struct rq *))head->func;
2656 raw_spin_unlock_irqrestore(&rq->lock, flags);
2659 static inline void balance_callback(struct rq *rq)
2661 if (unlikely(rq->balance_callback))
2662 __balance_callback(rq);
2667 static inline void balance_callback(struct rq *rq)
2674 * schedule_tail - first thing a freshly forked thread must call.
2675 * @prev: the thread we just switched away from.
2677 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2678 __releases(rq->lock)
2683 * New tasks start with FORK_PREEMPT_COUNT, see there and
2684 * finish_task_switch() for details.
2686 * finish_task_switch() will drop rq->lock() and lower preempt_count
2687 * and the preempt_enable() will end up enabling preemption (on
2688 * PREEMPT_COUNT kernels).
2691 rq = finish_task_switch(prev);
2692 balance_callback(rq);
2695 if (current->set_child_tid)
2696 put_user(task_pid_vnr(current), current->set_child_tid);
2700 * context_switch - switch to the new MM and the new thread's register state.
2702 static inline struct rq *
2703 context_switch(struct rq *rq, struct task_struct *prev,
2704 struct task_struct *next)
2706 struct mm_struct *mm, *oldmm;
2708 prepare_task_switch(rq, prev, next);
2711 oldmm = prev->active_mm;
2713 * For paravirt, this is coupled with an exit in switch_to to
2714 * combine the page table reload and the switch backend into
2717 arch_start_context_switch(prev);
2720 next->active_mm = oldmm;
2721 atomic_inc(&oldmm->mm_count);
2722 enter_lazy_tlb(oldmm, next);
2724 switch_mm(oldmm, mm, next);
2727 prev->active_mm = NULL;
2728 rq->prev_mm = oldmm;
2731 * Since the runqueue lock will be released by the next
2732 * task (which is an invalid locking op but in the case
2733 * of the scheduler it's an obvious special-case), so we
2734 * do an early lockdep release here:
2736 lockdep_unpin_lock(&rq->lock);
2737 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2739 /* Here we just switch the register state and the stack. */
2740 switch_to(prev, next, prev);
2743 return finish_task_switch(prev);
2747 * nr_running and nr_context_switches:
2749 * externally visible scheduler statistics: current number of runnable
2750 * threads, total number of context switches performed since bootup.
2752 unsigned long nr_running(void)
2754 unsigned long i, sum = 0;
2756 for_each_online_cpu(i)
2757 sum += cpu_rq(i)->nr_running;
2763 * Check if only the current task is running on the cpu.
2765 * Caution: this function does not check that the caller has disabled
2766 * preemption, thus the result might have a time-of-check-to-time-of-use
2767 * race. The caller is responsible to use it correctly, for example:
2769 * - from a non-preemptable section (of course)
2771 * - from a thread that is bound to a single CPU
2773 * - in a loop with very short iterations (e.g. a polling loop)
2775 bool single_task_running(void)
2777 return raw_rq()->nr_running == 1;
2779 EXPORT_SYMBOL(single_task_running);
2781 unsigned long long nr_context_switches(void)
2784 unsigned long long sum = 0;
2786 for_each_possible_cpu(i)
2787 sum += cpu_rq(i)->nr_switches;
2792 unsigned long nr_iowait(void)
2794 unsigned long i, sum = 0;
2796 for_each_possible_cpu(i)
2797 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2802 unsigned long nr_iowait_cpu(int cpu)
2804 struct rq *this = cpu_rq(cpu);
2805 return atomic_read(&this->nr_iowait);
2808 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2810 struct rq *rq = this_rq();
2811 *nr_waiters = atomic_read(&rq->nr_iowait);
2812 *load = rq->load.weight;
2818 * sched_exec - execve() is a valuable balancing opportunity, because at
2819 * this point the task has the smallest effective memory and cache footprint.
2821 void sched_exec(void)
2823 struct task_struct *p = current;
2824 unsigned long flags;
2827 raw_spin_lock_irqsave(&p->pi_lock, flags);
2828 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2829 if (dest_cpu == smp_processor_id())
2832 if (likely(cpu_active(dest_cpu))) {
2833 struct migration_arg arg = { p, dest_cpu };
2835 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2836 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2840 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2845 DEFINE_PER_CPU(struct kernel_stat, kstat);
2846 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2848 EXPORT_PER_CPU_SYMBOL(kstat);
2849 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2852 * Return accounted runtime for the task.
2853 * In case the task is currently running, return the runtime plus current's
2854 * pending runtime that have not been accounted yet.
2856 unsigned long long task_sched_runtime(struct task_struct *p)
2858 unsigned long flags;
2862 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2864 * 64-bit doesn't need locks to atomically read a 64bit value.
2865 * So we have a optimization chance when the task's delta_exec is 0.
2866 * Reading ->on_cpu is racy, but this is ok.
2868 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2869 * If we race with it entering cpu, unaccounted time is 0. This is
2870 * indistinguishable from the read occurring a few cycles earlier.
2871 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2872 * been accounted, so we're correct here as well.
2874 if (!p->on_cpu || !task_on_rq_queued(p))
2875 return p->se.sum_exec_runtime;
2878 rq = task_rq_lock(p, &flags);
2880 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2881 * project cycles that may never be accounted to this
2882 * thread, breaking clock_gettime().
2884 if (task_current(rq, p) && task_on_rq_queued(p)) {
2885 update_rq_clock(rq);
2886 p->sched_class->update_curr(rq);
2888 ns = p->se.sum_exec_runtime;
2889 task_rq_unlock(rq, p, &flags);
2895 * This function gets called by the timer code, with HZ frequency.
2896 * We call it with interrupts disabled.
2898 void scheduler_tick(void)
2900 int cpu = smp_processor_id();
2901 struct rq *rq = cpu_rq(cpu);
2902 struct task_struct *curr = rq->curr;
2906 raw_spin_lock(&rq->lock);
2907 update_rq_clock(rq);
2908 curr->sched_class->task_tick(rq, curr, 0);
2909 update_cpu_load_active(rq);
2910 calc_global_load_tick(rq);
2911 raw_spin_unlock(&rq->lock);
2913 perf_event_task_tick();
2916 rq->idle_balance = idle_cpu(cpu);
2917 trigger_load_balance(rq);
2919 rq_last_tick_reset(rq);
2922 #ifdef CONFIG_NO_HZ_FULL
2924 * scheduler_tick_max_deferment
2926 * Keep at least one tick per second when a single
2927 * active task is running because the scheduler doesn't
2928 * yet completely support full dynticks environment.
2930 * This makes sure that uptime, CFS vruntime, load
2931 * balancing, etc... continue to move forward, even
2932 * with a very low granularity.
2934 * Return: Maximum deferment in nanoseconds.
2936 u64 scheduler_tick_max_deferment(void)
2938 struct rq *rq = this_rq();
2939 unsigned long next, now = READ_ONCE(jiffies);
2941 next = rq->last_sched_tick + HZ;
2943 if (time_before_eq(next, now))
2946 return jiffies_to_nsecs(next - now);
2950 notrace unsigned long get_parent_ip(unsigned long addr)
2952 if (in_lock_functions(addr)) {
2953 addr = CALLER_ADDR2;
2954 if (in_lock_functions(addr))
2955 addr = CALLER_ADDR3;
2960 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2961 defined(CONFIG_PREEMPT_TRACER))
2963 void preempt_count_add(int val)
2965 #ifdef CONFIG_DEBUG_PREEMPT
2969 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2972 __preempt_count_add(val);
2973 #ifdef CONFIG_DEBUG_PREEMPT
2975 * Spinlock count overflowing soon?
2977 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2980 if (preempt_count() == val) {
2981 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2982 #ifdef CONFIG_DEBUG_PREEMPT
2983 current->preempt_disable_ip = ip;
2985 trace_preempt_off(CALLER_ADDR0, ip);
2988 EXPORT_SYMBOL(preempt_count_add);
2989 NOKPROBE_SYMBOL(preempt_count_add);
2991 void preempt_count_sub(int val)
2993 #ifdef CONFIG_DEBUG_PREEMPT
2997 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3000 * Is the spinlock portion underflowing?
3002 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3003 !(preempt_count() & PREEMPT_MASK)))
3007 if (preempt_count() == val)
3008 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3009 __preempt_count_sub(val);
3011 EXPORT_SYMBOL(preempt_count_sub);
3012 NOKPROBE_SYMBOL(preempt_count_sub);
3017 * Print scheduling while atomic bug:
3019 static noinline void __schedule_bug(struct task_struct *prev)
3021 if (oops_in_progress)
3024 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3025 prev->comm, prev->pid, preempt_count());
3027 debug_show_held_locks(prev);
3029 if (irqs_disabled())
3030 print_irqtrace_events(prev);
3031 #ifdef CONFIG_DEBUG_PREEMPT
3032 if (in_atomic_preempt_off()) {
3033 pr_err("Preemption disabled at:");
3034 print_ip_sym(current->preempt_disable_ip);
3039 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3043 * Various schedule()-time debugging checks and statistics:
3045 static inline void schedule_debug(struct task_struct *prev)
3047 #ifdef CONFIG_SCHED_STACK_END_CHECK
3048 BUG_ON(task_stack_end_corrupted(prev));
3051 if (unlikely(in_atomic_preempt_off())) {
3052 __schedule_bug(prev);
3053 preempt_count_set(PREEMPT_DISABLED);
3057 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3059 schedstat_inc(this_rq(), sched_count);
3063 * Pick up the highest-prio task:
3065 static inline struct task_struct *
3066 pick_next_task(struct rq *rq, struct task_struct *prev)
3068 const struct sched_class *class = &fair_sched_class;
3069 struct task_struct *p;
3072 * Optimization: we know that if all tasks are in
3073 * the fair class we can call that function directly:
3075 if (likely(prev->sched_class == class &&
3076 rq->nr_running == rq->cfs.h_nr_running)) {
3077 p = fair_sched_class.pick_next_task(rq, prev);
3078 if (unlikely(p == RETRY_TASK))
3081 /* assumes fair_sched_class->next == idle_sched_class */
3083 p = idle_sched_class.pick_next_task(rq, prev);
3089 for_each_class(class) {
3090 p = class->pick_next_task(rq, prev);
3092 if (unlikely(p == RETRY_TASK))
3098 BUG(); /* the idle class will always have a runnable task */
3102 * __schedule() is the main scheduler function.
3104 * The main means of driving the scheduler and thus entering this function are:
3106 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3108 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3109 * paths. For example, see arch/x86/entry_64.S.
3111 * To drive preemption between tasks, the scheduler sets the flag in timer
3112 * interrupt handler scheduler_tick().
3114 * 3. Wakeups don't really cause entry into schedule(). They add a
3115 * task to the run-queue and that's it.
3117 * Now, if the new task added to the run-queue preempts the current
3118 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3119 * called on the nearest possible occasion:
3121 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3123 * - in syscall or exception context, at the next outmost
3124 * preempt_enable(). (this might be as soon as the wake_up()'s
3127 * - in IRQ context, return from interrupt-handler to
3128 * preemptible context
3130 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3133 * - cond_resched() call
3134 * - explicit schedule() call
3135 * - return from syscall or exception to user-space
3136 * - return from interrupt-handler to user-space
3138 * WARNING: must be called with preemption disabled!
3140 static void __sched notrace __schedule(bool preempt)
3142 struct task_struct *prev, *next;
3143 unsigned long *switch_count;
3147 cpu = smp_processor_id();
3152 * do_exit() calls schedule() with preemption disabled as an exception;
3153 * however we must fix that up, otherwise the next task will see an
3154 * inconsistent (higher) preempt count.
3156 * It also avoids the below schedule_debug() test from complaining
3159 if (unlikely(prev->state == TASK_DEAD))
3160 preempt_enable_no_resched_notrace();
3162 schedule_debug(prev);
3164 if (sched_feat(HRTICK))
3167 local_irq_disable();
3168 rcu_note_context_switch();
3171 * Make sure that signal_pending_state()->signal_pending() below
3172 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3173 * done by the caller to avoid the race with signal_wake_up().
3175 smp_mb__before_spinlock();
3176 raw_spin_lock(&rq->lock);
3177 lockdep_pin_lock(&rq->lock);
3179 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3181 switch_count = &prev->nivcsw;
3182 if (!preempt && prev->state) {
3183 if (unlikely(signal_pending_state(prev->state, prev))) {
3184 prev->state = TASK_RUNNING;
3186 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3190 * If a worker went to sleep, notify and ask workqueue
3191 * whether it wants to wake up a task to maintain
3194 if (prev->flags & PF_WQ_WORKER) {
3195 struct task_struct *to_wakeup;
3197 to_wakeup = wq_worker_sleeping(prev, cpu);
3199 try_to_wake_up_local(to_wakeup);
3202 switch_count = &prev->nvcsw;
3205 if (task_on_rq_queued(prev))
3206 update_rq_clock(rq);
3208 next = pick_next_task(rq, prev);
3209 clear_tsk_need_resched(prev);
3210 clear_preempt_need_resched();
3211 rq->clock_skip_update = 0;
3213 if (likely(prev != next)) {
3218 trace_sched_switch(preempt, prev, next);
3219 rq = context_switch(rq, prev, next); /* unlocks the rq */
3221 lockdep_unpin_lock(&rq->lock);
3222 raw_spin_unlock_irq(&rq->lock);
3225 balance_callback(rq);
3228 static inline void sched_submit_work(struct task_struct *tsk)
3230 if (!tsk->state || tsk_is_pi_blocked(tsk))
3233 * If we are going to sleep and we have plugged IO queued,
3234 * make sure to submit it to avoid deadlocks.
3236 if (blk_needs_flush_plug(tsk))
3237 blk_schedule_flush_plug(tsk);
3240 asmlinkage __visible void __sched schedule(void)
3242 struct task_struct *tsk = current;
3244 sched_submit_work(tsk);
3248 sched_preempt_enable_no_resched();
3249 } while (need_resched());
3251 EXPORT_SYMBOL(schedule);
3253 #ifdef CONFIG_CONTEXT_TRACKING
3254 asmlinkage __visible void __sched schedule_user(void)
3257 * If we come here after a random call to set_need_resched(),
3258 * or we have been woken up remotely but the IPI has not yet arrived,
3259 * we haven't yet exited the RCU idle mode. Do it here manually until
3260 * we find a better solution.
3262 * NB: There are buggy callers of this function. Ideally we
3263 * should warn if prev_state != CONTEXT_USER, but that will trigger
3264 * too frequently to make sense yet.
3266 enum ctx_state prev_state = exception_enter();
3268 exception_exit(prev_state);
3273 * schedule_preempt_disabled - called with preemption disabled
3275 * Returns with preemption disabled. Note: preempt_count must be 1
3277 void __sched schedule_preempt_disabled(void)
3279 sched_preempt_enable_no_resched();
3284 static void __sched notrace preempt_schedule_common(void)
3287 preempt_disable_notrace();
3289 preempt_enable_no_resched_notrace();
3292 * Check again in case we missed a preemption opportunity
3293 * between schedule and now.
3295 } while (need_resched());
3298 #ifdef CONFIG_PREEMPT
3300 * this is the entry point to schedule() from in-kernel preemption
3301 * off of preempt_enable. Kernel preemptions off return from interrupt
3302 * occur there and call schedule directly.
3304 asmlinkage __visible void __sched notrace preempt_schedule(void)
3307 * If there is a non-zero preempt_count or interrupts are disabled,
3308 * we do not want to preempt the current task. Just return..
3310 if (likely(!preemptible()))
3313 preempt_schedule_common();
3315 NOKPROBE_SYMBOL(preempt_schedule);
3316 EXPORT_SYMBOL(preempt_schedule);
3319 * preempt_schedule_notrace - preempt_schedule called by tracing
3321 * The tracing infrastructure uses preempt_enable_notrace to prevent
3322 * recursion and tracing preempt enabling caused by the tracing
3323 * infrastructure itself. But as tracing can happen in areas coming
3324 * from userspace or just about to enter userspace, a preempt enable
3325 * can occur before user_exit() is called. This will cause the scheduler
3326 * to be called when the system is still in usermode.
3328 * To prevent this, the preempt_enable_notrace will use this function
3329 * instead of preempt_schedule() to exit user context if needed before
3330 * calling the scheduler.
3332 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3334 enum ctx_state prev_ctx;
3336 if (likely(!preemptible()))
3340 preempt_disable_notrace();
3342 * Needs preempt disabled in case user_exit() is traced
3343 * and the tracer calls preempt_enable_notrace() causing
3344 * an infinite recursion.
3346 prev_ctx = exception_enter();
3348 exception_exit(prev_ctx);
3350 preempt_enable_no_resched_notrace();
3351 } while (need_resched());
3353 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3355 #endif /* CONFIG_PREEMPT */
3358 * this is the entry point to schedule() from kernel preemption
3359 * off of irq context.
3360 * Note, that this is called and return with irqs disabled. This will
3361 * protect us against recursive calling from irq.
3363 asmlinkage __visible void __sched preempt_schedule_irq(void)
3365 enum ctx_state prev_state;
3367 /* Catch callers which need to be fixed */
3368 BUG_ON(preempt_count() || !irqs_disabled());
3370 prev_state = exception_enter();
3376 local_irq_disable();
3377 sched_preempt_enable_no_resched();
3378 } while (need_resched());
3380 exception_exit(prev_state);
3383 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3386 return try_to_wake_up(curr->private, mode, wake_flags);
3388 EXPORT_SYMBOL(default_wake_function);
3390 #ifdef CONFIG_RT_MUTEXES
3393 * rt_mutex_setprio - set the current priority of a task
3395 * @prio: prio value (kernel-internal form)
3397 * This function changes the 'effective' priority of a task. It does
3398 * not touch ->normal_prio like __setscheduler().
3400 * Used by the rt_mutex code to implement priority inheritance
3401 * logic. Call site only calls if the priority of the task changed.
3403 void rt_mutex_setprio(struct task_struct *p, int prio)
3405 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3407 const struct sched_class *prev_class;
3409 BUG_ON(prio > MAX_PRIO);
3411 rq = __task_rq_lock(p);
3414 * Idle task boosting is a nono in general. There is one
3415 * exception, when PREEMPT_RT and NOHZ is active:
3417 * The idle task calls get_next_timer_interrupt() and holds
3418 * the timer wheel base->lock on the CPU and another CPU wants
3419 * to access the timer (probably to cancel it). We can safely
3420 * ignore the boosting request, as the idle CPU runs this code
3421 * with interrupts disabled and will complete the lock
3422 * protected section without being interrupted. So there is no
3423 * real need to boost.
3425 if (unlikely(p == rq->idle)) {
3426 WARN_ON(p != rq->curr);
3427 WARN_ON(p->pi_blocked_on);
3431 trace_sched_pi_setprio(p, prio);
3434 if (oldprio == prio)
3435 queue_flag &= ~DEQUEUE_MOVE;
3437 prev_class = p->sched_class;
3438 queued = task_on_rq_queued(p);
3439 running = task_current(rq, p);
3441 dequeue_task(rq, p, queue_flag);
3443 put_prev_task(rq, p);
3446 * Boosting condition are:
3447 * 1. -rt task is running and holds mutex A
3448 * --> -dl task blocks on mutex A
3450 * 2. -dl task is running and holds mutex A
3451 * --> -dl task blocks on mutex A and could preempt the
3454 if (dl_prio(prio)) {
3455 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3456 if (!dl_prio(p->normal_prio) ||
3457 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3458 p->dl.dl_boosted = 1;
3459 queue_flag |= ENQUEUE_REPLENISH;
3461 p->dl.dl_boosted = 0;
3462 p->sched_class = &dl_sched_class;
3463 } else if (rt_prio(prio)) {
3464 if (dl_prio(oldprio))
3465 p->dl.dl_boosted = 0;
3467 queue_flag |= ENQUEUE_HEAD;
3468 p->sched_class = &rt_sched_class;
3470 if (dl_prio(oldprio))
3471 p->dl.dl_boosted = 0;
3472 if (rt_prio(oldprio))
3474 p->sched_class = &fair_sched_class;
3480 p->sched_class->set_curr_task(rq);
3482 enqueue_task(rq, p, queue_flag);
3484 check_class_changed(rq, p, prev_class, oldprio);
3486 preempt_disable(); /* avoid rq from going away on us */
3487 __task_rq_unlock(rq);
3489 balance_callback(rq);
3494 void set_user_nice(struct task_struct *p, long nice)
3496 int old_prio, delta, queued;
3497 unsigned long flags;
3500 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3503 * We have to be careful, if called from sys_setpriority(),
3504 * the task might be in the middle of scheduling on another CPU.
3506 rq = task_rq_lock(p, &flags);
3508 * The RT priorities are set via sched_setscheduler(), but we still
3509 * allow the 'normal' nice value to be set - but as expected
3510 * it wont have any effect on scheduling until the task is
3511 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3513 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3514 p->static_prio = NICE_TO_PRIO(nice);
3517 queued = task_on_rq_queued(p);
3519 dequeue_task(rq, p, DEQUEUE_SAVE);
3521 p->static_prio = NICE_TO_PRIO(nice);
3524 p->prio = effective_prio(p);
3525 delta = p->prio - old_prio;
3528 enqueue_task(rq, p, ENQUEUE_RESTORE);
3530 * If the task increased its priority or is running and
3531 * lowered its priority, then reschedule its CPU:
3533 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3537 task_rq_unlock(rq, p, &flags);
3539 EXPORT_SYMBOL(set_user_nice);
3542 * can_nice - check if a task can reduce its nice value
3546 int can_nice(const struct task_struct *p, const int nice)
3548 /* convert nice value [19,-20] to rlimit style value [1,40] */
3549 int nice_rlim = nice_to_rlimit(nice);
3551 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3552 capable(CAP_SYS_NICE));
3555 #ifdef __ARCH_WANT_SYS_NICE
3558 * sys_nice - change the priority of the current process.
3559 * @increment: priority increment
3561 * sys_setpriority is a more generic, but much slower function that
3562 * does similar things.
3564 SYSCALL_DEFINE1(nice, int, increment)
3569 * Setpriority might change our priority at the same moment.
3570 * We don't have to worry. Conceptually one call occurs first
3571 * and we have a single winner.
3573 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3574 nice = task_nice(current) + increment;
3576 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3577 if (increment < 0 && !can_nice(current, nice))
3580 retval = security_task_setnice(current, nice);
3584 set_user_nice(current, nice);
3591 * task_prio - return the priority value of a given task.
3592 * @p: the task in question.
3594 * Return: The priority value as seen by users in /proc.
3595 * RT tasks are offset by -200. Normal tasks are centered
3596 * around 0, value goes from -16 to +15.
3598 int task_prio(const struct task_struct *p)
3600 return p->prio - MAX_RT_PRIO;
3604 * idle_cpu - is a given cpu idle currently?
3605 * @cpu: the processor in question.
3607 * Return: 1 if the CPU is currently idle. 0 otherwise.
3609 int idle_cpu(int cpu)
3611 struct rq *rq = cpu_rq(cpu);
3613 if (rq->curr != rq->idle)
3620 if (!llist_empty(&rq->wake_list))
3628 * idle_task - return the idle task for a given cpu.
3629 * @cpu: the processor in question.
3631 * Return: The idle task for the cpu @cpu.
3633 struct task_struct *idle_task(int cpu)
3635 return cpu_rq(cpu)->idle;
3639 * find_process_by_pid - find a process with a matching PID value.
3640 * @pid: the pid in question.
3642 * The task of @pid, if found. %NULL otherwise.
3644 static struct task_struct *find_process_by_pid(pid_t pid)
3646 return pid ? find_task_by_vpid(pid) : current;
3650 * This function initializes the sched_dl_entity of a newly becoming
3651 * SCHED_DEADLINE task.
3653 * Only the static values are considered here, the actual runtime and the
3654 * absolute deadline will be properly calculated when the task is enqueued
3655 * for the first time with its new policy.
3658 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3660 struct sched_dl_entity *dl_se = &p->dl;
3662 dl_se->dl_runtime = attr->sched_runtime;
3663 dl_se->dl_deadline = attr->sched_deadline;
3664 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3665 dl_se->flags = attr->sched_flags;
3666 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3669 * Changing the parameters of a task is 'tricky' and we're not doing
3670 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3672 * What we SHOULD do is delay the bandwidth release until the 0-lag
3673 * point. This would include retaining the task_struct until that time
3674 * and change dl_overflow() to not immediately decrement the current
3677 * Instead we retain the current runtime/deadline and let the new
3678 * parameters take effect after the current reservation period lapses.
3679 * This is safe (albeit pessimistic) because the 0-lag point is always
3680 * before the current scheduling deadline.
3682 * We can still have temporary overloads because we do not delay the
3683 * change in bandwidth until that time; so admission control is
3684 * not on the safe side. It does however guarantee tasks will never
3685 * consume more than promised.
3690 * sched_setparam() passes in -1 for its policy, to let the functions
3691 * it calls know not to change it.
3693 #define SETPARAM_POLICY -1
3695 static void __setscheduler_params(struct task_struct *p,
3696 const struct sched_attr *attr)
3698 int policy = attr->sched_policy;
3700 if (policy == SETPARAM_POLICY)
3705 if (dl_policy(policy))
3706 __setparam_dl(p, attr);
3707 else if (fair_policy(policy))
3708 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3711 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3712 * !rt_policy. Always setting this ensures that things like
3713 * getparam()/getattr() don't report silly values for !rt tasks.
3715 p->rt_priority = attr->sched_priority;
3716 p->normal_prio = normal_prio(p);
3720 /* Actually do priority change: must hold pi & rq lock. */
3721 static void __setscheduler(struct rq *rq, struct task_struct *p,
3722 const struct sched_attr *attr, bool keep_boost)
3724 __setscheduler_params(p, attr);
3727 * Keep a potential priority boosting if called from
3728 * sched_setscheduler().
3731 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3733 p->prio = normal_prio(p);
3735 if (dl_prio(p->prio))
3736 p->sched_class = &dl_sched_class;
3737 else if (rt_prio(p->prio))
3738 p->sched_class = &rt_sched_class;
3740 p->sched_class = &fair_sched_class;
3744 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3746 struct sched_dl_entity *dl_se = &p->dl;
3748 attr->sched_priority = p->rt_priority;
3749 attr->sched_runtime = dl_se->dl_runtime;
3750 attr->sched_deadline = dl_se->dl_deadline;
3751 attr->sched_period = dl_se->dl_period;
3752 attr->sched_flags = dl_se->flags;
3756 * This function validates the new parameters of a -deadline task.
3757 * We ask for the deadline not being zero, and greater or equal
3758 * than the runtime, as well as the period of being zero or
3759 * greater than deadline. Furthermore, we have to be sure that
3760 * user parameters are above the internal resolution of 1us (we
3761 * check sched_runtime only since it is always the smaller one) and
3762 * below 2^63 ns (we have to check both sched_deadline and
3763 * sched_period, as the latter can be zero).
3766 __checkparam_dl(const struct sched_attr *attr)
3769 if (attr->sched_deadline == 0)
3773 * Since we truncate DL_SCALE bits, make sure we're at least
3776 if (attr->sched_runtime < (1ULL << DL_SCALE))
3780 * Since we use the MSB for wrap-around and sign issues, make
3781 * sure it's not set (mind that period can be equal to zero).
3783 if (attr->sched_deadline & (1ULL << 63) ||
3784 attr->sched_period & (1ULL << 63))
3787 /* runtime <= deadline <= period (if period != 0) */
3788 if ((attr->sched_period != 0 &&
3789 attr->sched_period < attr->sched_deadline) ||
3790 attr->sched_deadline < attr->sched_runtime)
3797 * check the target process has a UID that matches the current process's
3799 static bool check_same_owner(struct task_struct *p)
3801 const struct cred *cred = current_cred(), *pcred;
3805 pcred = __task_cred(p);
3806 match = (uid_eq(cred->euid, pcred->euid) ||
3807 uid_eq(cred->euid, pcred->uid));
3812 static bool dl_param_changed(struct task_struct *p,
3813 const struct sched_attr *attr)
3815 struct sched_dl_entity *dl_se = &p->dl;
3817 if (dl_se->dl_runtime != attr->sched_runtime ||
3818 dl_se->dl_deadline != attr->sched_deadline ||
3819 dl_se->dl_period != attr->sched_period ||
3820 dl_se->flags != attr->sched_flags)
3826 static int __sched_setscheduler(struct task_struct *p,
3827 const struct sched_attr *attr,
3830 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3831 MAX_RT_PRIO - 1 - attr->sched_priority;
3832 int retval, oldprio, oldpolicy = -1, queued, running;
3833 int new_effective_prio, policy = attr->sched_policy;
3834 unsigned long flags;
3835 const struct sched_class *prev_class;
3838 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
3840 /* may grab non-irq protected spin_locks */
3841 BUG_ON(in_interrupt());
3843 /* double check policy once rq lock held */
3845 reset_on_fork = p->sched_reset_on_fork;
3846 policy = oldpolicy = p->policy;
3848 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3850 if (!valid_policy(policy))
3854 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3858 * Valid priorities for SCHED_FIFO and SCHED_RR are
3859 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3860 * SCHED_BATCH and SCHED_IDLE is 0.
3862 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3863 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3865 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3866 (rt_policy(policy) != (attr->sched_priority != 0)))
3870 * Allow unprivileged RT tasks to decrease priority:
3872 if (user && !capable(CAP_SYS_NICE)) {
3873 if (fair_policy(policy)) {
3874 if (attr->sched_nice < task_nice(p) &&
3875 !can_nice(p, attr->sched_nice))
3879 if (rt_policy(policy)) {
3880 unsigned long rlim_rtprio =
3881 task_rlimit(p, RLIMIT_RTPRIO);
3883 /* can't set/change the rt policy */
3884 if (policy != p->policy && !rlim_rtprio)
3887 /* can't increase priority */
3888 if (attr->sched_priority > p->rt_priority &&
3889 attr->sched_priority > rlim_rtprio)
3894 * Can't set/change SCHED_DEADLINE policy at all for now
3895 * (safest behavior); in the future we would like to allow
3896 * unprivileged DL tasks to increase their relative deadline
3897 * or reduce their runtime (both ways reducing utilization)
3899 if (dl_policy(policy))
3903 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3904 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3906 if (idle_policy(p->policy) && !idle_policy(policy)) {
3907 if (!can_nice(p, task_nice(p)))
3911 /* can't change other user's priorities */
3912 if (!check_same_owner(p))
3915 /* Normal users shall not reset the sched_reset_on_fork flag */
3916 if (p->sched_reset_on_fork && !reset_on_fork)
3921 retval = security_task_setscheduler(p);
3927 * make sure no PI-waiters arrive (or leave) while we are
3928 * changing the priority of the task:
3930 * To be able to change p->policy safely, the appropriate
3931 * runqueue lock must be held.
3933 rq = task_rq_lock(p, &flags);
3936 * Changing the policy of the stop threads its a very bad idea
3938 if (p == rq->stop) {
3939 task_rq_unlock(rq, p, &flags);
3944 * If not changing anything there's no need to proceed further,
3945 * but store a possible modification of reset_on_fork.
3947 if (unlikely(policy == p->policy)) {
3948 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3950 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3952 if (dl_policy(policy) && dl_param_changed(p, attr))
3955 p->sched_reset_on_fork = reset_on_fork;
3956 task_rq_unlock(rq, p, &flags);
3962 #ifdef CONFIG_RT_GROUP_SCHED
3964 * Do not allow realtime tasks into groups that have no runtime
3967 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3968 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3969 !task_group_is_autogroup(task_group(p))) {
3970 task_rq_unlock(rq, p, &flags);
3975 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3976 cpumask_t *span = rq->rd->span;
3979 * Don't allow tasks with an affinity mask smaller than
3980 * the entire root_domain to become SCHED_DEADLINE. We
3981 * will also fail if there's no bandwidth available.
3983 if (!cpumask_subset(span, &p->cpus_allowed) ||
3984 rq->rd->dl_bw.bw == 0) {
3985 task_rq_unlock(rq, p, &flags);
3992 /* recheck policy now with rq lock held */
3993 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3994 policy = oldpolicy = -1;
3995 task_rq_unlock(rq, p, &flags);
4000 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4001 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4004 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4005 task_rq_unlock(rq, p, &flags);
4009 p->sched_reset_on_fork = reset_on_fork;
4014 * Take priority boosted tasks into account. If the new
4015 * effective priority is unchanged, we just store the new
4016 * normal parameters and do not touch the scheduler class and
4017 * the runqueue. This will be done when the task deboost
4020 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4021 if (new_effective_prio == oldprio)
4022 queue_flags &= ~DEQUEUE_MOVE;
4025 queued = task_on_rq_queued(p);
4026 running = task_current(rq, p);
4028 dequeue_task(rq, p, queue_flags);
4030 put_prev_task(rq, p);
4032 prev_class = p->sched_class;
4033 __setscheduler(rq, p, attr, pi);
4036 p->sched_class->set_curr_task(rq);
4039 * We enqueue to tail when the priority of a task is
4040 * increased (user space view).
4042 if (oldprio < p->prio)
4043 queue_flags |= ENQUEUE_HEAD;
4045 enqueue_task(rq, p, queue_flags);
4048 check_class_changed(rq, p, prev_class, oldprio);
4049 preempt_disable(); /* avoid rq from going away on us */
4050 task_rq_unlock(rq, p, &flags);
4053 rt_mutex_adjust_pi(p);
4056 * Run balance callbacks after we've adjusted the PI chain.
4058 balance_callback(rq);
4064 static int _sched_setscheduler(struct task_struct *p, int policy,
4065 const struct sched_param *param, bool check)
4067 struct sched_attr attr = {
4068 .sched_policy = policy,
4069 .sched_priority = param->sched_priority,
4070 .sched_nice = PRIO_TO_NICE(p->static_prio),
4073 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4074 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4075 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4076 policy &= ~SCHED_RESET_ON_FORK;
4077 attr.sched_policy = policy;
4080 return __sched_setscheduler(p, &attr, check, true);
4083 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4084 * @p: the task in question.
4085 * @policy: new policy.
4086 * @param: structure containing the new RT priority.
4088 * Return: 0 on success. An error code otherwise.
4090 * NOTE that the task may be already dead.
4092 int sched_setscheduler(struct task_struct *p, int policy,
4093 const struct sched_param *param)
4095 return _sched_setscheduler(p, policy, param, true);
4097 EXPORT_SYMBOL_GPL(sched_setscheduler);
4099 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4101 return __sched_setscheduler(p, attr, true, true);
4103 EXPORT_SYMBOL_GPL(sched_setattr);
4106 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4107 * @p: the task in question.
4108 * @policy: new policy.
4109 * @param: structure containing the new RT priority.
4111 * Just like sched_setscheduler, only don't bother checking if the
4112 * current context has permission. For example, this is needed in
4113 * stop_machine(): we create temporary high priority worker threads,
4114 * but our caller might not have that capability.
4116 * Return: 0 on success. An error code otherwise.
4118 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4119 const struct sched_param *param)
4121 return _sched_setscheduler(p, policy, param, false);
4123 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4126 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4128 struct sched_param lparam;
4129 struct task_struct *p;
4132 if (!param || pid < 0)
4134 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4139 p = find_process_by_pid(pid);
4141 retval = sched_setscheduler(p, policy, &lparam);
4148 * Mimics kernel/events/core.c perf_copy_attr().
4150 static int sched_copy_attr(struct sched_attr __user *uattr,
4151 struct sched_attr *attr)
4156 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4160 * zero the full structure, so that a short copy will be nice.
4162 memset(attr, 0, sizeof(*attr));
4164 ret = get_user(size, &uattr->size);
4168 if (size > PAGE_SIZE) /* silly large */
4171 if (!size) /* abi compat */
4172 size = SCHED_ATTR_SIZE_VER0;
4174 if (size < SCHED_ATTR_SIZE_VER0)
4178 * If we're handed a bigger struct than we know of,
4179 * ensure all the unknown bits are 0 - i.e. new
4180 * user-space does not rely on any kernel feature
4181 * extensions we dont know about yet.
4183 if (size > sizeof(*attr)) {
4184 unsigned char __user *addr;
4185 unsigned char __user *end;
4188 addr = (void __user *)uattr + sizeof(*attr);
4189 end = (void __user *)uattr + size;
4191 for (; addr < end; addr++) {
4192 ret = get_user(val, addr);
4198 size = sizeof(*attr);
4201 ret = copy_from_user(attr, uattr, size);
4206 * XXX: do we want to be lenient like existing syscalls; or do we want
4207 * to be strict and return an error on out-of-bounds values?
4209 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4214 put_user(sizeof(*attr), &uattr->size);
4219 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4220 * @pid: the pid in question.
4221 * @policy: new policy.
4222 * @param: structure containing the new RT priority.
4224 * Return: 0 on success. An error code otherwise.
4226 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4227 struct sched_param __user *, param)
4229 /* negative values for policy are not valid */
4233 return do_sched_setscheduler(pid, policy, param);
4237 * sys_sched_setparam - set/change the RT priority of a thread
4238 * @pid: the pid in question.
4239 * @param: structure containing the new RT priority.
4241 * Return: 0 on success. An error code otherwise.
4243 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4245 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4249 * sys_sched_setattr - same as above, but with extended sched_attr
4250 * @pid: the pid in question.
4251 * @uattr: structure containing the extended parameters.
4252 * @flags: for future extension.
4254 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4255 unsigned int, flags)
4257 struct sched_attr attr;
4258 struct task_struct *p;
4261 if (!uattr || pid < 0 || flags)
4264 retval = sched_copy_attr(uattr, &attr);
4268 if ((int)attr.sched_policy < 0)
4273 p = find_process_by_pid(pid);
4275 retval = sched_setattr(p, &attr);
4282 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4283 * @pid: the pid in question.
4285 * Return: On success, the policy of the thread. Otherwise, a negative error
4288 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4290 struct task_struct *p;
4298 p = find_process_by_pid(pid);
4300 retval = security_task_getscheduler(p);
4303 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4310 * sys_sched_getparam - get the RT priority of a thread
4311 * @pid: the pid in question.
4312 * @param: structure containing the RT priority.
4314 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4317 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4319 struct sched_param lp = { .sched_priority = 0 };
4320 struct task_struct *p;
4323 if (!param || pid < 0)
4327 p = find_process_by_pid(pid);
4332 retval = security_task_getscheduler(p);
4336 if (task_has_rt_policy(p))
4337 lp.sched_priority = p->rt_priority;
4341 * This one might sleep, we cannot do it with a spinlock held ...
4343 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4352 static int sched_read_attr(struct sched_attr __user *uattr,
4353 struct sched_attr *attr,
4358 if (!access_ok(VERIFY_WRITE, uattr, usize))
4362 * If we're handed a smaller struct than we know of,
4363 * ensure all the unknown bits are 0 - i.e. old
4364 * user-space does not get uncomplete information.
4366 if (usize < sizeof(*attr)) {
4367 unsigned char *addr;
4370 addr = (void *)attr + usize;
4371 end = (void *)attr + sizeof(*attr);
4373 for (; addr < end; addr++) {
4381 ret = copy_to_user(uattr, attr, attr->size);
4389 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4390 * @pid: the pid in question.
4391 * @uattr: structure containing the extended parameters.
4392 * @size: sizeof(attr) for fwd/bwd comp.
4393 * @flags: for future extension.
4395 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4396 unsigned int, size, unsigned int, flags)
4398 struct sched_attr attr = {
4399 .size = sizeof(struct sched_attr),
4401 struct task_struct *p;
4404 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4405 size < SCHED_ATTR_SIZE_VER0 || flags)
4409 p = find_process_by_pid(pid);
4414 retval = security_task_getscheduler(p);
4418 attr.sched_policy = p->policy;
4419 if (p->sched_reset_on_fork)
4420 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4421 if (task_has_dl_policy(p))
4422 __getparam_dl(p, &attr);
4423 else if (task_has_rt_policy(p))
4424 attr.sched_priority = p->rt_priority;
4426 attr.sched_nice = task_nice(p);
4430 retval = sched_read_attr(uattr, &attr, size);
4438 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4440 cpumask_var_t cpus_allowed, new_mask;
4441 struct task_struct *p;
4446 p = find_process_by_pid(pid);
4452 /* Prevent p going away */
4456 if (p->flags & PF_NO_SETAFFINITY) {
4460 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4464 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4466 goto out_free_cpus_allowed;
4469 if (!check_same_owner(p)) {
4471 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4473 goto out_free_new_mask;
4478 retval = security_task_setscheduler(p);
4480 goto out_free_new_mask;
4483 cpuset_cpus_allowed(p, cpus_allowed);
4484 cpumask_and(new_mask, in_mask, cpus_allowed);
4487 * Since bandwidth control happens on root_domain basis,
4488 * if admission test is enabled, we only admit -deadline
4489 * tasks allowed to run on all the CPUs in the task's
4493 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4495 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4498 goto out_free_new_mask;
4504 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4507 cpuset_cpus_allowed(p, cpus_allowed);
4508 if (!cpumask_subset(new_mask, cpus_allowed)) {
4510 * We must have raced with a concurrent cpuset
4511 * update. Just reset the cpus_allowed to the
4512 * cpuset's cpus_allowed
4514 cpumask_copy(new_mask, cpus_allowed);
4519 free_cpumask_var(new_mask);
4520 out_free_cpus_allowed:
4521 free_cpumask_var(cpus_allowed);
4527 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4528 struct cpumask *new_mask)
4530 if (len < cpumask_size())
4531 cpumask_clear(new_mask);
4532 else if (len > cpumask_size())
4533 len = cpumask_size();
4535 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4539 * sys_sched_setaffinity - set the cpu affinity of a process
4540 * @pid: pid of the process
4541 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4542 * @user_mask_ptr: user-space pointer to the new cpu mask
4544 * Return: 0 on success. An error code otherwise.
4546 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4547 unsigned long __user *, user_mask_ptr)
4549 cpumask_var_t new_mask;
4552 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4555 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4557 retval = sched_setaffinity(pid, new_mask);
4558 free_cpumask_var(new_mask);
4562 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4564 struct task_struct *p;
4565 unsigned long flags;
4571 p = find_process_by_pid(pid);
4575 retval = security_task_getscheduler(p);
4579 raw_spin_lock_irqsave(&p->pi_lock, flags);
4580 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4581 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4590 * sys_sched_getaffinity - get the cpu affinity of a process
4591 * @pid: pid of the process
4592 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4593 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4595 * Return: 0 on success. An error code otherwise.
4597 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4598 unsigned long __user *, user_mask_ptr)
4603 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4605 if (len & (sizeof(unsigned long)-1))
4608 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4611 ret = sched_getaffinity(pid, mask);
4613 size_t retlen = min_t(size_t, len, cpumask_size());
4615 if (copy_to_user(user_mask_ptr, mask, retlen))
4620 free_cpumask_var(mask);
4626 * sys_sched_yield - yield the current processor to other threads.
4628 * This function yields the current CPU to other tasks. If there are no
4629 * other threads running on this CPU then this function will return.
4633 SYSCALL_DEFINE0(sched_yield)
4635 struct rq *rq = this_rq_lock();
4637 schedstat_inc(rq, yld_count);
4638 current->sched_class->yield_task(rq);
4641 * Since we are going to call schedule() anyway, there's
4642 * no need to preempt or enable interrupts:
4644 __release(rq->lock);
4645 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4646 do_raw_spin_unlock(&rq->lock);
4647 sched_preempt_enable_no_resched();
4654 int __sched _cond_resched(void)
4656 if (should_resched(0)) {
4657 preempt_schedule_common();
4662 EXPORT_SYMBOL(_cond_resched);
4665 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4666 * call schedule, and on return reacquire the lock.
4668 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4669 * operations here to prevent schedule() from being called twice (once via
4670 * spin_unlock(), once by hand).
4672 int __cond_resched_lock(spinlock_t *lock)
4674 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4677 lockdep_assert_held(lock);
4679 if (spin_needbreak(lock) || resched) {
4682 preempt_schedule_common();
4690 EXPORT_SYMBOL(__cond_resched_lock);
4692 int __sched __cond_resched_softirq(void)
4694 BUG_ON(!in_softirq());
4696 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4698 preempt_schedule_common();
4704 EXPORT_SYMBOL(__cond_resched_softirq);
4707 * yield - yield the current processor to other threads.
4709 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4711 * The scheduler is at all times free to pick the calling task as the most
4712 * eligible task to run, if removing the yield() call from your code breaks
4713 * it, its already broken.
4715 * Typical broken usage is:
4720 * where one assumes that yield() will let 'the other' process run that will
4721 * make event true. If the current task is a SCHED_FIFO task that will never
4722 * happen. Never use yield() as a progress guarantee!!
4724 * If you want to use yield() to wait for something, use wait_event().
4725 * If you want to use yield() to be 'nice' for others, use cond_resched().
4726 * If you still want to use yield(), do not!
4728 void __sched yield(void)
4730 set_current_state(TASK_RUNNING);
4733 EXPORT_SYMBOL(yield);
4736 * yield_to - yield the current processor to another thread in
4737 * your thread group, or accelerate that thread toward the
4738 * processor it's on.
4740 * @preempt: whether task preemption is allowed or not
4742 * It's the caller's job to ensure that the target task struct
4743 * can't go away on us before we can do any checks.
4746 * true (>0) if we indeed boosted the target task.
4747 * false (0) if we failed to boost the target.
4748 * -ESRCH if there's no task to yield to.
4750 int __sched yield_to(struct task_struct *p, bool preempt)
4752 struct task_struct *curr = current;
4753 struct rq *rq, *p_rq;
4754 unsigned long flags;
4757 local_irq_save(flags);
4763 * If we're the only runnable task on the rq and target rq also
4764 * has only one task, there's absolutely no point in yielding.
4766 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4771 double_rq_lock(rq, p_rq);
4772 if (task_rq(p) != p_rq) {
4773 double_rq_unlock(rq, p_rq);
4777 if (!curr->sched_class->yield_to_task)
4780 if (curr->sched_class != p->sched_class)
4783 if (task_running(p_rq, p) || p->state)
4786 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4788 schedstat_inc(rq, yld_count);
4790 * Make p's CPU reschedule; pick_next_entity takes care of
4793 if (preempt && rq != p_rq)
4798 double_rq_unlock(rq, p_rq);
4800 local_irq_restore(flags);
4807 EXPORT_SYMBOL_GPL(yield_to);
4810 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4811 * that process accounting knows that this is a task in IO wait state.
4813 long __sched io_schedule_timeout(long timeout)
4815 int old_iowait = current->in_iowait;
4819 current->in_iowait = 1;
4820 blk_schedule_flush_plug(current);
4822 delayacct_blkio_start();
4824 atomic_inc(&rq->nr_iowait);
4825 ret = schedule_timeout(timeout);
4826 current->in_iowait = old_iowait;
4827 atomic_dec(&rq->nr_iowait);
4828 delayacct_blkio_end();
4832 EXPORT_SYMBOL(io_schedule_timeout);
4835 * sys_sched_get_priority_max - return maximum RT priority.
4836 * @policy: scheduling class.
4838 * Return: On success, this syscall returns the maximum
4839 * rt_priority that can be used by a given scheduling class.
4840 * On failure, a negative error code is returned.
4842 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4849 ret = MAX_USER_RT_PRIO-1;
4851 case SCHED_DEADLINE:
4862 * sys_sched_get_priority_min - return minimum RT priority.
4863 * @policy: scheduling class.
4865 * Return: On success, this syscall returns the minimum
4866 * rt_priority that can be used by a given scheduling class.
4867 * On failure, a negative error code is returned.
4869 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4878 case SCHED_DEADLINE:
4888 * sys_sched_rr_get_interval - return the default timeslice of a process.
4889 * @pid: pid of the process.
4890 * @interval: userspace pointer to the timeslice value.
4892 * this syscall writes the default timeslice value of a given process
4893 * into the user-space timespec buffer. A value of '0' means infinity.
4895 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4898 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4899 struct timespec __user *, interval)
4901 struct task_struct *p;
4902 unsigned int time_slice;
4903 unsigned long flags;
4913 p = find_process_by_pid(pid);
4917 retval = security_task_getscheduler(p);
4921 rq = task_rq_lock(p, &flags);
4923 if (p->sched_class->get_rr_interval)
4924 time_slice = p->sched_class->get_rr_interval(rq, p);
4925 task_rq_unlock(rq, p, &flags);
4928 jiffies_to_timespec(time_slice, &t);
4929 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4937 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4939 void sched_show_task(struct task_struct *p)
4941 unsigned long free = 0;
4943 unsigned long state = p->state;
4946 state = __ffs(state) + 1;
4947 printk(KERN_INFO "%-15.15s %c", p->comm,
4948 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4949 #if BITS_PER_LONG == 32
4950 if (state == TASK_RUNNING)
4951 printk(KERN_CONT " running ");
4953 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4955 if (state == TASK_RUNNING)
4956 printk(KERN_CONT " running task ");
4958 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4960 #ifdef CONFIG_DEBUG_STACK_USAGE
4961 free = stack_not_used(p);
4966 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4968 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4969 task_pid_nr(p), ppid,
4970 (unsigned long)task_thread_info(p)->flags);
4972 print_worker_info(KERN_INFO, p);
4973 show_stack(p, NULL);
4976 void show_state_filter(unsigned long state_filter)
4978 struct task_struct *g, *p;
4980 #if BITS_PER_LONG == 32
4982 " task PC stack pid father\n");
4985 " task PC stack pid father\n");
4988 for_each_process_thread(g, p) {
4990 * reset the NMI-timeout, listing all files on a slow
4991 * console might take a lot of time:
4993 touch_nmi_watchdog();
4994 if (!state_filter || (p->state & state_filter))
4998 touch_all_softlockup_watchdogs();
5000 #ifdef CONFIG_SCHED_DEBUG
5001 sysrq_sched_debug_show();
5005 * Only show locks if all tasks are dumped:
5008 debug_show_all_locks();
5011 void init_idle_bootup_task(struct task_struct *idle)
5013 idle->sched_class = &idle_sched_class;
5017 * init_idle - set up an idle thread for a given CPU
5018 * @idle: task in question
5019 * @cpu: cpu the idle task belongs to
5021 * NOTE: this function does not set the idle thread's NEED_RESCHED
5022 * flag, to make booting more robust.
5024 void init_idle(struct task_struct *idle, int cpu)
5026 struct rq *rq = cpu_rq(cpu);
5027 unsigned long flags;
5029 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5030 raw_spin_lock(&rq->lock);
5032 __sched_fork(0, idle);
5033 idle->state = TASK_RUNNING;
5034 idle->se.exec_start = sched_clock();
5038 * Its possible that init_idle() gets called multiple times on a task,
5039 * in that case do_set_cpus_allowed() will not do the right thing.
5041 * And since this is boot we can forgo the serialization.
5043 set_cpus_allowed_common(idle, cpumask_of(cpu));
5046 * We're having a chicken and egg problem, even though we are
5047 * holding rq->lock, the cpu isn't yet set to this cpu so the
5048 * lockdep check in task_group() will fail.
5050 * Similar case to sched_fork(). / Alternatively we could
5051 * use task_rq_lock() here and obtain the other rq->lock.
5056 __set_task_cpu(idle, cpu);
5059 rq->curr = rq->idle = idle;
5060 idle->on_rq = TASK_ON_RQ_QUEUED;
5064 raw_spin_unlock(&rq->lock);
5065 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5067 /* Set the preempt count _outside_ the spinlocks! */
5068 init_idle_preempt_count(idle, cpu);
5071 * The idle tasks have their own, simple scheduling class:
5073 idle->sched_class = &idle_sched_class;
5074 ftrace_graph_init_idle_task(idle, cpu);
5075 vtime_init_idle(idle, cpu);
5077 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5081 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5082 const struct cpumask *trial)
5084 int ret = 1, trial_cpus;
5085 struct dl_bw *cur_dl_b;
5086 unsigned long flags;
5088 if (!cpumask_weight(cur))
5091 rcu_read_lock_sched();
5092 cur_dl_b = dl_bw_of(cpumask_any(cur));
5093 trial_cpus = cpumask_weight(trial);
5095 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5096 if (cur_dl_b->bw != -1 &&
5097 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5099 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5100 rcu_read_unlock_sched();
5105 int task_can_attach(struct task_struct *p,
5106 const struct cpumask *cs_cpus_allowed)
5111 * Kthreads which disallow setaffinity shouldn't be moved
5112 * to a new cpuset; we don't want to change their cpu
5113 * affinity and isolating such threads by their set of
5114 * allowed nodes is unnecessary. Thus, cpusets are not
5115 * applicable for such threads. This prevents checking for
5116 * success of set_cpus_allowed_ptr() on all attached tasks
5117 * before cpus_allowed may be changed.
5119 if (p->flags & PF_NO_SETAFFINITY) {
5125 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5127 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5132 unsigned long flags;
5134 rcu_read_lock_sched();
5135 dl_b = dl_bw_of(dest_cpu);
5136 raw_spin_lock_irqsave(&dl_b->lock, flags);
5137 cpus = dl_bw_cpus(dest_cpu);
5138 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5143 * We reserve space for this task in the destination
5144 * root_domain, as we can't fail after this point.
5145 * We will free resources in the source root_domain
5146 * later on (see set_cpus_allowed_dl()).
5148 __dl_add(dl_b, p->dl.dl_bw);
5150 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5151 rcu_read_unlock_sched();
5161 #ifdef CONFIG_NUMA_BALANCING
5162 /* Migrate current task p to target_cpu */
5163 int migrate_task_to(struct task_struct *p, int target_cpu)
5165 struct migration_arg arg = { p, target_cpu };
5166 int curr_cpu = task_cpu(p);
5168 if (curr_cpu == target_cpu)
5171 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5174 /* TODO: This is not properly updating schedstats */
5176 trace_sched_move_numa(p, curr_cpu, target_cpu);
5177 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5181 * Requeue a task on a given node and accurately track the number of NUMA
5182 * tasks on the runqueues
5184 void sched_setnuma(struct task_struct *p, int nid)
5187 unsigned long flags;
5188 bool queued, running;
5190 rq = task_rq_lock(p, &flags);
5191 queued = task_on_rq_queued(p);
5192 running = task_current(rq, p);
5195 dequeue_task(rq, p, DEQUEUE_SAVE);
5197 put_prev_task(rq, p);
5199 p->numa_preferred_nid = nid;
5202 p->sched_class->set_curr_task(rq);
5204 enqueue_task(rq, p, ENQUEUE_RESTORE);
5205 task_rq_unlock(rq, p, &flags);
5207 #endif /* CONFIG_NUMA_BALANCING */
5209 #ifdef CONFIG_HOTPLUG_CPU
5211 * Ensures that the idle task is using init_mm right before its cpu goes
5214 void idle_task_exit(void)
5216 struct mm_struct *mm = current->active_mm;
5218 BUG_ON(cpu_online(smp_processor_id()));
5220 if (mm != &init_mm) {
5221 switch_mm(mm, &init_mm, current);
5222 finish_arch_post_lock_switch();
5228 * Since this CPU is going 'away' for a while, fold any nr_active delta
5229 * we might have. Assumes we're called after migrate_tasks() so that the
5230 * nr_active count is stable.
5232 * Also see the comment "Global load-average calculations".
5234 static void calc_load_migrate(struct rq *rq)
5236 long delta = calc_load_fold_active(rq);
5238 atomic_long_add(delta, &calc_load_tasks);
5241 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5245 static const struct sched_class fake_sched_class = {
5246 .put_prev_task = put_prev_task_fake,
5249 static struct task_struct fake_task = {
5251 * Avoid pull_{rt,dl}_task()
5253 .prio = MAX_PRIO + 1,
5254 .sched_class = &fake_sched_class,
5258 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5259 * try_to_wake_up()->select_task_rq().
5261 * Called with rq->lock held even though we'er in stop_machine() and
5262 * there's no concurrency possible, we hold the required locks anyway
5263 * because of lock validation efforts.
5265 static void migrate_tasks(struct rq *dead_rq)
5267 struct rq *rq = dead_rq;
5268 struct task_struct *next, *stop = rq->stop;
5272 * Fudge the rq selection such that the below task selection loop
5273 * doesn't get stuck on the currently eligible stop task.
5275 * We're currently inside stop_machine() and the rq is either stuck
5276 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5277 * either way we should never end up calling schedule() until we're
5283 * put_prev_task() and pick_next_task() sched
5284 * class method both need to have an up-to-date
5285 * value of rq->clock[_task]
5287 update_rq_clock(rq);
5291 * There's this thread running, bail when that's the only
5294 if (rq->nr_running == 1)
5298 * pick_next_task assumes pinned rq->lock.
5300 lockdep_pin_lock(&rq->lock);
5301 next = pick_next_task(rq, &fake_task);
5303 next->sched_class->put_prev_task(rq, next);
5306 * Rules for changing task_struct::cpus_allowed are holding
5307 * both pi_lock and rq->lock, such that holding either
5308 * stabilizes the mask.
5310 * Drop rq->lock is not quite as disastrous as it usually is
5311 * because !cpu_active at this point, which means load-balance
5312 * will not interfere. Also, stop-machine.
5314 lockdep_unpin_lock(&rq->lock);
5315 raw_spin_unlock(&rq->lock);
5316 raw_spin_lock(&next->pi_lock);
5317 raw_spin_lock(&rq->lock);
5320 * Since we're inside stop-machine, _nothing_ should have
5321 * changed the task, WARN if weird stuff happened, because in
5322 * that case the above rq->lock drop is a fail too.
5324 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5325 raw_spin_unlock(&next->pi_lock);
5329 /* Find suitable destination for @next, with force if needed. */
5330 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5332 rq = __migrate_task(rq, next, dest_cpu);
5333 if (rq != dead_rq) {
5334 raw_spin_unlock(&rq->lock);
5336 raw_spin_lock(&rq->lock);
5338 raw_spin_unlock(&next->pi_lock);
5343 #endif /* CONFIG_HOTPLUG_CPU */
5345 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5347 static struct ctl_table sd_ctl_dir[] = {
5349 .procname = "sched_domain",
5355 static struct ctl_table sd_ctl_root[] = {
5357 .procname = "kernel",
5359 .child = sd_ctl_dir,
5364 static struct ctl_table *sd_alloc_ctl_entry(int n)
5366 struct ctl_table *entry =
5367 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5372 static void sd_free_ctl_entry(struct ctl_table **tablep)
5374 struct ctl_table *entry;
5377 * In the intermediate directories, both the child directory and
5378 * procname are dynamically allocated and could fail but the mode
5379 * will always be set. In the lowest directory the names are
5380 * static strings and all have proc handlers.
5382 for (entry = *tablep; entry->mode; entry++) {
5384 sd_free_ctl_entry(&entry->child);
5385 if (entry->proc_handler == NULL)
5386 kfree(entry->procname);
5393 static int min_load_idx = 0;
5394 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5397 set_table_entry(struct ctl_table *entry,
5398 const char *procname, void *data, int maxlen,
5399 umode_t mode, proc_handler *proc_handler,
5402 entry->procname = procname;
5404 entry->maxlen = maxlen;
5406 entry->proc_handler = proc_handler;
5409 entry->extra1 = &min_load_idx;
5410 entry->extra2 = &max_load_idx;
5414 static struct ctl_table *
5415 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5417 struct ctl_table *table = sd_alloc_ctl_entry(14);
5422 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5423 sizeof(long), 0644, proc_doulongvec_minmax, false);
5424 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5425 sizeof(long), 0644, proc_doulongvec_minmax, false);
5426 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5427 sizeof(int), 0644, proc_dointvec_minmax, true);
5428 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5429 sizeof(int), 0644, proc_dointvec_minmax, true);
5430 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5431 sizeof(int), 0644, proc_dointvec_minmax, true);
5432 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5433 sizeof(int), 0644, proc_dointvec_minmax, true);
5434 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5435 sizeof(int), 0644, proc_dointvec_minmax, true);
5436 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5437 sizeof(int), 0644, proc_dointvec_minmax, false);
5438 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5439 sizeof(int), 0644, proc_dointvec_minmax, false);
5440 set_table_entry(&table[9], "cache_nice_tries",
5441 &sd->cache_nice_tries,
5442 sizeof(int), 0644, proc_dointvec_minmax, false);
5443 set_table_entry(&table[10], "flags", &sd->flags,
5444 sizeof(int), 0644, proc_dointvec_minmax, false);
5445 set_table_entry(&table[11], "max_newidle_lb_cost",
5446 &sd->max_newidle_lb_cost,
5447 sizeof(long), 0644, proc_doulongvec_minmax, false);
5448 set_table_entry(&table[12], "name", sd->name,
5449 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5450 /* &table[13] is terminator */
5455 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5457 struct ctl_table *entry, *table;
5458 struct sched_domain *sd;
5459 int domain_num = 0, i;
5462 for_each_domain(cpu, sd)
5464 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5469 for_each_domain(cpu, sd) {
5470 snprintf(buf, 32, "domain%d", i);
5471 entry->procname = kstrdup(buf, GFP_KERNEL);
5473 entry->child = sd_alloc_ctl_domain_table(sd);
5480 static struct ctl_table_header *sd_sysctl_header;
5481 static void register_sched_domain_sysctl(void)
5483 int i, cpu_num = num_possible_cpus();
5484 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5487 WARN_ON(sd_ctl_dir[0].child);
5488 sd_ctl_dir[0].child = entry;
5493 for_each_possible_cpu(i) {
5494 snprintf(buf, 32, "cpu%d", i);
5495 entry->procname = kstrdup(buf, GFP_KERNEL);
5497 entry->child = sd_alloc_ctl_cpu_table(i);
5501 WARN_ON(sd_sysctl_header);
5502 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5505 /* may be called multiple times per register */
5506 static void unregister_sched_domain_sysctl(void)
5508 unregister_sysctl_table(sd_sysctl_header);
5509 sd_sysctl_header = NULL;
5510 if (sd_ctl_dir[0].child)
5511 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5514 static void register_sched_domain_sysctl(void)
5517 static void unregister_sched_domain_sysctl(void)
5520 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5522 static void set_rq_online(struct rq *rq)
5525 const struct sched_class *class;
5527 cpumask_set_cpu(rq->cpu, rq->rd->online);
5530 for_each_class(class) {
5531 if (class->rq_online)
5532 class->rq_online(rq);
5537 static void set_rq_offline(struct rq *rq)
5540 const struct sched_class *class;
5542 for_each_class(class) {
5543 if (class->rq_offline)
5544 class->rq_offline(rq);
5547 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5553 * migration_call - callback that gets triggered when a CPU is added.
5554 * Here we can start up the necessary migration thread for the new CPU.
5557 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5559 int cpu = (long)hcpu;
5560 unsigned long flags;
5561 struct rq *rq = cpu_rq(cpu);
5563 switch (action & ~CPU_TASKS_FROZEN) {
5565 case CPU_UP_PREPARE:
5566 rq->calc_load_update = calc_load_update;
5570 /* Update our root-domain */
5571 raw_spin_lock_irqsave(&rq->lock, flags);
5573 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5577 raw_spin_unlock_irqrestore(&rq->lock, flags);
5580 #ifdef CONFIG_HOTPLUG_CPU
5582 sched_ttwu_pending();
5583 /* Update our root-domain */
5584 raw_spin_lock_irqsave(&rq->lock, flags);
5586 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5590 BUG_ON(rq->nr_running != 1); /* the migration thread */
5591 raw_spin_unlock_irqrestore(&rq->lock, flags);
5595 calc_load_migrate(rq);
5600 update_max_interval();
5606 * Register at high priority so that task migration (migrate_all_tasks)
5607 * happens before everything else. This has to be lower priority than
5608 * the notifier in the perf_event subsystem, though.
5610 static struct notifier_block migration_notifier = {
5611 .notifier_call = migration_call,
5612 .priority = CPU_PRI_MIGRATION,
5615 static void set_cpu_rq_start_time(void)
5617 int cpu = smp_processor_id();
5618 struct rq *rq = cpu_rq(cpu);
5619 rq->age_stamp = sched_clock_cpu(cpu);
5622 static int sched_cpu_active(struct notifier_block *nfb,
5623 unsigned long action, void *hcpu)
5625 int cpu = (long)hcpu;
5627 switch (action & ~CPU_TASKS_FROZEN) {
5629 set_cpu_rq_start_time();
5634 * At this point a starting CPU has marked itself as online via
5635 * set_cpu_online(). But it might not yet have marked itself
5636 * as active, which is essential from here on.
5638 set_cpu_active(cpu, true);
5639 stop_machine_unpark(cpu);
5642 case CPU_DOWN_FAILED:
5643 set_cpu_active(cpu, true);
5651 static int sched_cpu_inactive(struct notifier_block *nfb,
5652 unsigned long action, void *hcpu)
5654 switch (action & ~CPU_TASKS_FROZEN) {
5655 case CPU_DOWN_PREPARE:
5656 set_cpu_active((long)hcpu, false);
5663 static int __init migration_init(void)
5665 void *cpu = (void *)(long)smp_processor_id();
5668 /* Initialize migration for the boot CPU */
5669 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5670 BUG_ON(err == NOTIFY_BAD);
5671 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5672 register_cpu_notifier(&migration_notifier);
5674 /* Register cpu active notifiers */
5675 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5676 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5680 early_initcall(migration_init);
5682 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5684 #ifdef CONFIG_SCHED_DEBUG
5686 static __read_mostly int sched_debug_enabled;
5688 static int __init sched_debug_setup(char *str)
5690 sched_debug_enabled = 1;
5694 early_param("sched_debug", sched_debug_setup);
5696 static inline bool sched_debug(void)
5698 return sched_debug_enabled;
5701 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5702 struct cpumask *groupmask)
5704 struct sched_group *group = sd->groups;
5706 cpumask_clear(groupmask);
5708 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5710 if (!(sd->flags & SD_LOAD_BALANCE)) {
5711 printk("does not load-balance\n");
5713 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5718 printk(KERN_CONT "span %*pbl level %s\n",
5719 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5721 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5722 printk(KERN_ERR "ERROR: domain->span does not contain "
5725 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5726 printk(KERN_ERR "ERROR: domain->groups does not contain"
5730 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5734 printk(KERN_ERR "ERROR: group is NULL\n");
5738 if (!cpumask_weight(sched_group_cpus(group))) {
5739 printk(KERN_CONT "\n");
5740 printk(KERN_ERR "ERROR: empty group\n");
5744 if (!(sd->flags & SD_OVERLAP) &&
5745 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5746 printk(KERN_CONT "\n");
5747 printk(KERN_ERR "ERROR: repeated CPUs\n");
5751 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5753 printk(KERN_CONT " %*pbl",
5754 cpumask_pr_args(sched_group_cpus(group)));
5755 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5756 printk(KERN_CONT " (cpu_capacity = %d)",
5757 group->sgc->capacity);
5760 group = group->next;
5761 } while (group != sd->groups);
5762 printk(KERN_CONT "\n");
5764 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5765 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5768 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5769 printk(KERN_ERR "ERROR: parent span is not a superset "
5770 "of domain->span\n");
5774 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5778 if (!sched_debug_enabled)
5782 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5786 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5789 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5797 #else /* !CONFIG_SCHED_DEBUG */
5798 # define sched_domain_debug(sd, cpu) do { } while (0)
5799 static inline bool sched_debug(void)
5803 #endif /* CONFIG_SCHED_DEBUG */
5805 static int sd_degenerate(struct sched_domain *sd)
5807 if (cpumask_weight(sched_domain_span(sd)) == 1)
5810 /* Following flags need at least 2 groups */
5811 if (sd->flags & (SD_LOAD_BALANCE |
5812 SD_BALANCE_NEWIDLE |
5815 SD_SHARE_CPUCAPACITY |
5816 SD_SHARE_PKG_RESOURCES |
5817 SD_SHARE_POWERDOMAIN)) {
5818 if (sd->groups != sd->groups->next)
5822 /* Following flags don't use groups */
5823 if (sd->flags & (SD_WAKE_AFFINE))
5830 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5832 unsigned long cflags = sd->flags, pflags = parent->flags;
5834 if (sd_degenerate(parent))
5837 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5840 /* Flags needing groups don't count if only 1 group in parent */
5841 if (parent->groups == parent->groups->next) {
5842 pflags &= ~(SD_LOAD_BALANCE |
5843 SD_BALANCE_NEWIDLE |
5846 SD_SHARE_CPUCAPACITY |
5847 SD_SHARE_PKG_RESOURCES |
5849 SD_SHARE_POWERDOMAIN);
5850 if (nr_node_ids == 1)
5851 pflags &= ~SD_SERIALIZE;
5853 if (~cflags & pflags)
5859 static void free_rootdomain(struct rcu_head *rcu)
5861 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5863 cpupri_cleanup(&rd->cpupri);
5864 cpudl_cleanup(&rd->cpudl);
5865 free_cpumask_var(rd->dlo_mask);
5866 free_cpumask_var(rd->rto_mask);
5867 free_cpumask_var(rd->online);
5868 free_cpumask_var(rd->span);
5872 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5874 struct root_domain *old_rd = NULL;
5875 unsigned long flags;
5877 raw_spin_lock_irqsave(&rq->lock, flags);
5882 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5885 cpumask_clear_cpu(rq->cpu, old_rd->span);
5888 * If we dont want to free the old_rd yet then
5889 * set old_rd to NULL to skip the freeing later
5892 if (!atomic_dec_and_test(&old_rd->refcount))
5896 atomic_inc(&rd->refcount);
5899 cpumask_set_cpu(rq->cpu, rd->span);
5900 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5903 raw_spin_unlock_irqrestore(&rq->lock, flags);
5906 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5909 static int init_rootdomain(struct root_domain *rd)
5911 memset(rd, 0, sizeof(*rd));
5913 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5915 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5917 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5919 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5922 init_dl_bw(&rd->dl_bw);
5923 if (cpudl_init(&rd->cpudl) != 0)
5926 if (cpupri_init(&rd->cpupri) != 0)
5931 free_cpumask_var(rd->rto_mask);
5933 free_cpumask_var(rd->dlo_mask);
5935 free_cpumask_var(rd->online);
5937 free_cpumask_var(rd->span);
5943 * By default the system creates a single root-domain with all cpus as
5944 * members (mimicking the global state we have today).
5946 struct root_domain def_root_domain;
5948 static void init_defrootdomain(void)
5950 init_rootdomain(&def_root_domain);
5952 atomic_set(&def_root_domain.refcount, 1);
5955 static struct root_domain *alloc_rootdomain(void)
5957 struct root_domain *rd;
5959 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5963 if (init_rootdomain(rd) != 0) {
5971 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5973 struct sched_group *tmp, *first;
5982 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5987 } while (sg != first);
5990 static void free_sched_domain(struct rcu_head *rcu)
5992 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5995 * If its an overlapping domain it has private groups, iterate and
5998 if (sd->flags & SD_OVERLAP) {
5999 free_sched_groups(sd->groups, 1);
6000 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6001 kfree(sd->groups->sgc);
6007 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6009 call_rcu(&sd->rcu, free_sched_domain);
6012 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6014 for (; sd; sd = sd->parent)
6015 destroy_sched_domain(sd, cpu);
6019 * Keep a special pointer to the highest sched_domain that has
6020 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6021 * allows us to avoid some pointer chasing select_idle_sibling().
6023 * Also keep a unique ID per domain (we use the first cpu number in
6024 * the cpumask of the domain), this allows us to quickly tell if
6025 * two cpus are in the same cache domain, see cpus_share_cache().
6027 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6028 DEFINE_PER_CPU(int, sd_llc_size);
6029 DEFINE_PER_CPU(int, sd_llc_id);
6030 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6031 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6032 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6034 static void update_top_cache_domain(int cpu)
6036 struct sched_domain *sd;
6037 struct sched_domain *busy_sd = NULL;
6041 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6043 id = cpumask_first(sched_domain_span(sd));
6044 size = cpumask_weight(sched_domain_span(sd));
6045 busy_sd = sd->parent; /* sd_busy */
6047 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6049 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6050 per_cpu(sd_llc_size, cpu) = size;
6051 per_cpu(sd_llc_id, cpu) = id;
6053 sd = lowest_flag_domain(cpu, SD_NUMA);
6054 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6056 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6057 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6061 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6062 * hold the hotplug lock.
6065 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6067 struct rq *rq = cpu_rq(cpu);
6068 struct sched_domain *tmp;
6070 /* Remove the sched domains which do not contribute to scheduling. */
6071 for (tmp = sd; tmp; ) {
6072 struct sched_domain *parent = tmp->parent;
6076 if (sd_parent_degenerate(tmp, parent)) {
6077 tmp->parent = parent->parent;
6079 parent->parent->child = tmp;
6081 * Transfer SD_PREFER_SIBLING down in case of a
6082 * degenerate parent; the spans match for this
6083 * so the property transfers.
6085 if (parent->flags & SD_PREFER_SIBLING)
6086 tmp->flags |= SD_PREFER_SIBLING;
6087 destroy_sched_domain(parent, cpu);
6092 if (sd && sd_degenerate(sd)) {
6095 destroy_sched_domain(tmp, cpu);
6100 sched_domain_debug(sd, cpu);
6102 rq_attach_root(rq, rd);
6104 rcu_assign_pointer(rq->sd, sd);
6105 destroy_sched_domains(tmp, cpu);
6107 update_top_cache_domain(cpu);
6110 /* Setup the mask of cpus configured for isolated domains */
6111 static int __init isolated_cpu_setup(char *str)
6115 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6116 ret = cpulist_parse(str, cpu_isolated_map);
6118 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6123 __setup("isolcpus=", isolated_cpu_setup);
6126 struct sched_domain ** __percpu sd;
6127 struct root_domain *rd;
6138 * Build an iteration mask that can exclude certain CPUs from the upwards
6141 * Asymmetric node setups can result in situations where the domain tree is of
6142 * unequal depth, make sure to skip domains that already cover the entire
6145 * In that case build_sched_domains() will have terminated the iteration early
6146 * and our sibling sd spans will be empty. Domains should always include the
6147 * cpu they're built on, so check that.
6150 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6152 const struct cpumask *span = sched_domain_span(sd);
6153 struct sd_data *sdd = sd->private;
6154 struct sched_domain *sibling;
6157 for_each_cpu(i, span) {
6158 sibling = *per_cpu_ptr(sdd->sd, i);
6159 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6162 cpumask_set_cpu(i, sched_group_mask(sg));
6167 * Return the canonical balance cpu for this group, this is the first cpu
6168 * of this group that's also in the iteration mask.
6170 int group_balance_cpu(struct sched_group *sg)
6172 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6176 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6178 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6179 const struct cpumask *span = sched_domain_span(sd);
6180 struct cpumask *covered = sched_domains_tmpmask;
6181 struct sd_data *sdd = sd->private;
6182 struct sched_domain *sibling;
6185 cpumask_clear(covered);
6187 for_each_cpu(i, span) {
6188 struct cpumask *sg_span;
6190 if (cpumask_test_cpu(i, covered))
6193 sibling = *per_cpu_ptr(sdd->sd, i);
6195 /* See the comment near build_group_mask(). */
6196 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6199 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6200 GFP_KERNEL, cpu_to_node(cpu));
6205 sg_span = sched_group_cpus(sg);
6207 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6209 cpumask_set_cpu(i, sg_span);
6211 cpumask_or(covered, covered, sg_span);
6213 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6214 if (atomic_inc_return(&sg->sgc->ref) == 1)
6215 build_group_mask(sd, sg);
6218 * Initialize sgc->capacity such that even if we mess up the
6219 * domains and no possible iteration will get us here, we won't
6222 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6225 * Make sure the first group of this domain contains the
6226 * canonical balance cpu. Otherwise the sched_domain iteration
6227 * breaks. See update_sg_lb_stats().
6229 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6230 group_balance_cpu(sg) == cpu)
6240 sd->groups = groups;
6245 free_sched_groups(first, 0);
6250 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6252 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6253 struct sched_domain *child = sd->child;
6256 cpu = cpumask_first(sched_domain_span(child));
6259 *sg = *per_cpu_ptr(sdd->sg, cpu);
6260 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6261 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6268 * build_sched_groups will build a circular linked list of the groups
6269 * covered by the given span, and will set each group's ->cpumask correctly,
6270 * and ->cpu_capacity to 0.
6272 * Assumes the sched_domain tree is fully constructed
6275 build_sched_groups(struct sched_domain *sd, int cpu)
6277 struct sched_group *first = NULL, *last = NULL;
6278 struct sd_data *sdd = sd->private;
6279 const struct cpumask *span = sched_domain_span(sd);
6280 struct cpumask *covered;
6283 get_group(cpu, sdd, &sd->groups);
6284 atomic_inc(&sd->groups->ref);
6286 if (cpu != cpumask_first(span))
6289 lockdep_assert_held(&sched_domains_mutex);
6290 covered = sched_domains_tmpmask;
6292 cpumask_clear(covered);
6294 for_each_cpu(i, span) {
6295 struct sched_group *sg;
6298 if (cpumask_test_cpu(i, covered))
6301 group = get_group(i, sdd, &sg);
6302 cpumask_setall(sched_group_mask(sg));
6304 for_each_cpu(j, span) {
6305 if (get_group(j, sdd, NULL) != group)
6308 cpumask_set_cpu(j, covered);
6309 cpumask_set_cpu(j, sched_group_cpus(sg));
6324 * Initialize sched groups cpu_capacity.
6326 * cpu_capacity indicates the capacity of sched group, which is used while
6327 * distributing the load between different sched groups in a sched domain.
6328 * Typically cpu_capacity for all the groups in a sched domain will be same
6329 * unless there are asymmetries in the topology. If there are asymmetries,
6330 * group having more cpu_capacity will pickup more load compared to the
6331 * group having less cpu_capacity.
6333 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6335 struct sched_group *sg = sd->groups;
6340 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6342 } while (sg != sd->groups);
6344 if (cpu != group_balance_cpu(sg))
6347 update_group_capacity(sd, cpu);
6348 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6352 * Initializers for schedule domains
6353 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6356 static int default_relax_domain_level = -1;
6357 int sched_domain_level_max;
6359 static int __init setup_relax_domain_level(char *str)
6361 if (kstrtoint(str, 0, &default_relax_domain_level))
6362 pr_warn("Unable to set relax_domain_level\n");
6366 __setup("relax_domain_level=", setup_relax_domain_level);
6368 static void set_domain_attribute(struct sched_domain *sd,
6369 struct sched_domain_attr *attr)
6373 if (!attr || attr->relax_domain_level < 0) {
6374 if (default_relax_domain_level < 0)
6377 request = default_relax_domain_level;
6379 request = attr->relax_domain_level;
6380 if (request < sd->level) {
6381 /* turn off idle balance on this domain */
6382 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6384 /* turn on idle balance on this domain */
6385 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6389 static void __sdt_free(const struct cpumask *cpu_map);
6390 static int __sdt_alloc(const struct cpumask *cpu_map);
6392 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6393 const struct cpumask *cpu_map)
6397 if (!atomic_read(&d->rd->refcount))
6398 free_rootdomain(&d->rd->rcu); /* fall through */
6400 free_percpu(d->sd); /* fall through */
6402 __sdt_free(cpu_map); /* fall through */
6408 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6409 const struct cpumask *cpu_map)
6411 memset(d, 0, sizeof(*d));
6413 if (__sdt_alloc(cpu_map))
6414 return sa_sd_storage;
6415 d->sd = alloc_percpu(struct sched_domain *);
6417 return sa_sd_storage;
6418 d->rd = alloc_rootdomain();
6421 return sa_rootdomain;
6425 * NULL the sd_data elements we've used to build the sched_domain and
6426 * sched_group structure so that the subsequent __free_domain_allocs()
6427 * will not free the data we're using.
6429 static void claim_allocations(int cpu, struct sched_domain *sd)
6431 struct sd_data *sdd = sd->private;
6433 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6434 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6436 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6437 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6439 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6440 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6444 static int sched_domains_numa_levels;
6445 enum numa_topology_type sched_numa_topology_type;
6446 static int *sched_domains_numa_distance;
6447 int sched_max_numa_distance;
6448 static struct cpumask ***sched_domains_numa_masks;
6449 static int sched_domains_curr_level;
6453 * SD_flags allowed in topology descriptions.
6455 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6456 * SD_SHARE_PKG_RESOURCES - describes shared caches
6457 * SD_NUMA - describes NUMA topologies
6458 * SD_SHARE_POWERDOMAIN - describes shared power domain
6461 * SD_ASYM_PACKING - describes SMT quirks
6463 #define TOPOLOGY_SD_FLAGS \
6464 (SD_SHARE_CPUCAPACITY | \
6465 SD_SHARE_PKG_RESOURCES | \
6468 SD_SHARE_POWERDOMAIN)
6470 static struct sched_domain *
6471 sd_init(struct sched_domain_topology_level *tl, int cpu)
6473 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6474 int sd_weight, sd_flags = 0;
6478 * Ugly hack to pass state to sd_numa_mask()...
6480 sched_domains_curr_level = tl->numa_level;
6483 sd_weight = cpumask_weight(tl->mask(cpu));
6486 sd_flags = (*tl->sd_flags)();
6487 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6488 "wrong sd_flags in topology description\n"))
6489 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6491 *sd = (struct sched_domain){
6492 .min_interval = sd_weight,
6493 .max_interval = 2*sd_weight,
6495 .imbalance_pct = 125,
6497 .cache_nice_tries = 0,
6504 .flags = 1*SD_LOAD_BALANCE
6505 | 1*SD_BALANCE_NEWIDLE
6510 | 0*SD_SHARE_CPUCAPACITY
6511 | 0*SD_SHARE_PKG_RESOURCES
6513 | 0*SD_PREFER_SIBLING
6518 .last_balance = jiffies,
6519 .balance_interval = sd_weight,
6521 .max_newidle_lb_cost = 0,
6522 .next_decay_max_lb_cost = jiffies,
6523 #ifdef CONFIG_SCHED_DEBUG
6529 * Convert topological properties into behaviour.
6532 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6533 sd->flags |= SD_PREFER_SIBLING;
6534 sd->imbalance_pct = 110;
6535 sd->smt_gain = 1178; /* ~15% */
6537 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6538 sd->imbalance_pct = 117;
6539 sd->cache_nice_tries = 1;
6543 } else if (sd->flags & SD_NUMA) {
6544 sd->cache_nice_tries = 2;
6548 sd->flags |= SD_SERIALIZE;
6549 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6550 sd->flags &= ~(SD_BALANCE_EXEC |
6557 sd->flags |= SD_PREFER_SIBLING;
6558 sd->cache_nice_tries = 1;
6563 sd->private = &tl->data;
6569 * Topology list, bottom-up.
6571 static struct sched_domain_topology_level default_topology[] = {
6572 #ifdef CONFIG_SCHED_SMT
6573 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6575 #ifdef CONFIG_SCHED_MC
6576 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6578 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6582 static struct sched_domain_topology_level *sched_domain_topology =
6585 #define for_each_sd_topology(tl) \
6586 for (tl = sched_domain_topology; tl->mask; tl++)
6588 void set_sched_topology(struct sched_domain_topology_level *tl)
6590 sched_domain_topology = tl;
6595 static const struct cpumask *sd_numa_mask(int cpu)
6597 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6600 static void sched_numa_warn(const char *str)
6602 static int done = false;
6610 printk(KERN_WARNING "ERROR: %s\n\n", str);
6612 for (i = 0; i < nr_node_ids; i++) {
6613 printk(KERN_WARNING " ");
6614 for (j = 0; j < nr_node_ids; j++)
6615 printk(KERN_CONT "%02d ", node_distance(i,j));
6616 printk(KERN_CONT "\n");
6618 printk(KERN_WARNING "\n");
6621 bool find_numa_distance(int distance)
6625 if (distance == node_distance(0, 0))
6628 for (i = 0; i < sched_domains_numa_levels; i++) {
6629 if (sched_domains_numa_distance[i] == distance)
6637 * A system can have three types of NUMA topology:
6638 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6639 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6640 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6642 * The difference between a glueless mesh topology and a backplane
6643 * topology lies in whether communication between not directly
6644 * connected nodes goes through intermediary nodes (where programs
6645 * could run), or through backplane controllers. This affects
6646 * placement of programs.
6648 * The type of topology can be discerned with the following tests:
6649 * - If the maximum distance between any nodes is 1 hop, the system
6650 * is directly connected.
6651 * - If for two nodes A and B, located N > 1 hops away from each other,
6652 * there is an intermediary node C, which is < N hops away from both
6653 * nodes A and B, the system is a glueless mesh.
6655 static void init_numa_topology_type(void)
6659 n = sched_max_numa_distance;
6661 if (sched_domains_numa_levels <= 1) {
6662 sched_numa_topology_type = NUMA_DIRECT;
6666 for_each_online_node(a) {
6667 for_each_online_node(b) {
6668 /* Find two nodes furthest removed from each other. */
6669 if (node_distance(a, b) < n)
6672 /* Is there an intermediary node between a and b? */
6673 for_each_online_node(c) {
6674 if (node_distance(a, c) < n &&
6675 node_distance(b, c) < n) {
6676 sched_numa_topology_type =
6682 sched_numa_topology_type = NUMA_BACKPLANE;
6688 static void sched_init_numa(void)
6690 int next_distance, curr_distance = node_distance(0, 0);
6691 struct sched_domain_topology_level *tl;
6695 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6696 if (!sched_domains_numa_distance)
6700 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6701 * unique distances in the node_distance() table.
6703 * Assumes node_distance(0,j) includes all distances in
6704 * node_distance(i,j) in order to avoid cubic time.
6706 next_distance = curr_distance;
6707 for (i = 0; i < nr_node_ids; i++) {
6708 for (j = 0; j < nr_node_ids; j++) {
6709 for (k = 0; k < nr_node_ids; k++) {
6710 int distance = node_distance(i, k);
6712 if (distance > curr_distance &&
6713 (distance < next_distance ||
6714 next_distance == curr_distance))
6715 next_distance = distance;
6718 * While not a strong assumption it would be nice to know
6719 * about cases where if node A is connected to B, B is not
6720 * equally connected to A.
6722 if (sched_debug() && node_distance(k, i) != distance)
6723 sched_numa_warn("Node-distance not symmetric");
6725 if (sched_debug() && i && !find_numa_distance(distance))
6726 sched_numa_warn("Node-0 not representative");
6728 if (next_distance != curr_distance) {
6729 sched_domains_numa_distance[level++] = next_distance;
6730 sched_domains_numa_levels = level;
6731 curr_distance = next_distance;
6736 * In case of sched_debug() we verify the above assumption.
6746 * 'level' contains the number of unique distances, excluding the
6747 * identity distance node_distance(i,i).
6749 * The sched_domains_numa_distance[] array includes the actual distance
6754 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6755 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6756 * the array will contain less then 'level' members. This could be
6757 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6758 * in other functions.
6760 * We reset it to 'level' at the end of this function.
6762 sched_domains_numa_levels = 0;
6764 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6765 if (!sched_domains_numa_masks)
6769 * Now for each level, construct a mask per node which contains all
6770 * cpus of nodes that are that many hops away from us.
6772 for (i = 0; i < level; i++) {
6773 sched_domains_numa_masks[i] =
6774 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6775 if (!sched_domains_numa_masks[i])
6778 for (j = 0; j < nr_node_ids; j++) {
6779 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6783 sched_domains_numa_masks[i][j] = mask;
6786 if (node_distance(j, k) > sched_domains_numa_distance[i])
6789 cpumask_or(mask, mask, cpumask_of_node(k));
6794 /* Compute default topology size */
6795 for (i = 0; sched_domain_topology[i].mask; i++);
6797 tl = kzalloc((i + level + 1) *
6798 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6803 * Copy the default topology bits..
6805 for (i = 0; sched_domain_topology[i].mask; i++)
6806 tl[i] = sched_domain_topology[i];
6809 * .. and append 'j' levels of NUMA goodness.
6811 for (j = 0; j < level; i++, j++) {
6812 tl[i] = (struct sched_domain_topology_level){
6813 .mask = sd_numa_mask,
6814 .sd_flags = cpu_numa_flags,
6815 .flags = SDTL_OVERLAP,
6821 sched_domain_topology = tl;
6823 sched_domains_numa_levels = level;
6824 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6826 init_numa_topology_type();
6829 static void sched_domains_numa_masks_set(int cpu)
6832 int node = cpu_to_node(cpu);
6834 for (i = 0; i < sched_domains_numa_levels; i++) {
6835 for (j = 0; j < nr_node_ids; j++) {
6836 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6837 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6842 static void sched_domains_numa_masks_clear(int cpu)
6845 for (i = 0; i < sched_domains_numa_levels; i++) {
6846 for (j = 0; j < nr_node_ids; j++)
6847 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6852 * Update sched_domains_numa_masks[level][node] array when new cpus
6855 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6856 unsigned long action,
6859 int cpu = (long)hcpu;
6861 switch (action & ~CPU_TASKS_FROZEN) {
6863 sched_domains_numa_masks_set(cpu);
6867 sched_domains_numa_masks_clear(cpu);
6877 static inline void sched_init_numa(void)
6881 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6882 unsigned long action,
6887 #endif /* CONFIG_NUMA */
6889 static int __sdt_alloc(const struct cpumask *cpu_map)
6891 struct sched_domain_topology_level *tl;
6894 for_each_sd_topology(tl) {
6895 struct sd_data *sdd = &tl->data;
6897 sdd->sd = alloc_percpu(struct sched_domain *);
6901 sdd->sg = alloc_percpu(struct sched_group *);
6905 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6909 for_each_cpu(j, cpu_map) {
6910 struct sched_domain *sd;
6911 struct sched_group *sg;
6912 struct sched_group_capacity *sgc;
6914 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6915 GFP_KERNEL, cpu_to_node(j));
6919 *per_cpu_ptr(sdd->sd, j) = sd;
6921 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6922 GFP_KERNEL, cpu_to_node(j));
6928 *per_cpu_ptr(sdd->sg, j) = sg;
6930 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6931 GFP_KERNEL, cpu_to_node(j));
6935 *per_cpu_ptr(sdd->sgc, j) = sgc;
6942 static void __sdt_free(const struct cpumask *cpu_map)
6944 struct sched_domain_topology_level *tl;
6947 for_each_sd_topology(tl) {
6948 struct sd_data *sdd = &tl->data;
6950 for_each_cpu(j, cpu_map) {
6951 struct sched_domain *sd;
6954 sd = *per_cpu_ptr(sdd->sd, j);
6955 if (sd && (sd->flags & SD_OVERLAP))
6956 free_sched_groups(sd->groups, 0);
6957 kfree(*per_cpu_ptr(sdd->sd, j));
6961 kfree(*per_cpu_ptr(sdd->sg, j));
6963 kfree(*per_cpu_ptr(sdd->sgc, j));
6965 free_percpu(sdd->sd);
6967 free_percpu(sdd->sg);
6969 free_percpu(sdd->sgc);
6974 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6975 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6976 struct sched_domain *child, int cpu)
6978 struct sched_domain *sd = sd_init(tl, cpu);
6982 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6984 sd->level = child->level + 1;
6985 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6989 if (!cpumask_subset(sched_domain_span(child),
6990 sched_domain_span(sd))) {
6991 pr_err("BUG: arch topology borken\n");
6992 #ifdef CONFIG_SCHED_DEBUG
6993 pr_err(" the %s domain not a subset of the %s domain\n",
6994 child->name, sd->name);
6996 /* Fixup, ensure @sd has at least @child cpus. */
6997 cpumask_or(sched_domain_span(sd),
6998 sched_domain_span(sd),
6999 sched_domain_span(child));
7003 set_domain_attribute(sd, attr);
7009 * Build sched domains for a given set of cpus and attach the sched domains
7010 * to the individual cpus
7012 static int build_sched_domains(const struct cpumask *cpu_map,
7013 struct sched_domain_attr *attr)
7015 enum s_alloc alloc_state;
7016 struct sched_domain *sd;
7018 int i, ret = -ENOMEM;
7020 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7021 if (alloc_state != sa_rootdomain)
7024 /* Set up domains for cpus specified by the cpu_map. */
7025 for_each_cpu(i, cpu_map) {
7026 struct sched_domain_topology_level *tl;
7029 for_each_sd_topology(tl) {
7030 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7031 if (tl == sched_domain_topology)
7032 *per_cpu_ptr(d.sd, i) = sd;
7033 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7034 sd->flags |= SD_OVERLAP;
7035 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7040 /* Build the groups for the domains */
7041 for_each_cpu(i, cpu_map) {
7042 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7043 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7044 if (sd->flags & SD_OVERLAP) {
7045 if (build_overlap_sched_groups(sd, i))
7048 if (build_sched_groups(sd, i))
7054 /* Calculate CPU capacity for physical packages and nodes */
7055 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7056 if (!cpumask_test_cpu(i, cpu_map))
7059 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7060 claim_allocations(i, sd);
7061 init_sched_groups_capacity(i, sd);
7065 /* Attach the domains */
7067 for_each_cpu(i, cpu_map) {
7068 sd = *per_cpu_ptr(d.sd, i);
7069 cpu_attach_domain(sd, d.rd, i);
7075 __free_domain_allocs(&d, alloc_state, cpu_map);
7079 static cpumask_var_t *doms_cur; /* current sched domains */
7080 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7081 static struct sched_domain_attr *dattr_cur;
7082 /* attribues of custom domains in 'doms_cur' */
7085 * Special case: If a kmalloc of a doms_cur partition (array of
7086 * cpumask) fails, then fallback to a single sched domain,
7087 * as determined by the single cpumask fallback_doms.
7089 static cpumask_var_t fallback_doms;
7092 * arch_update_cpu_topology lets virtualized architectures update the
7093 * cpu core maps. It is supposed to return 1 if the topology changed
7094 * or 0 if it stayed the same.
7096 int __weak arch_update_cpu_topology(void)
7101 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7104 cpumask_var_t *doms;
7106 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7109 for (i = 0; i < ndoms; i++) {
7110 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7111 free_sched_domains(doms, i);
7118 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7121 for (i = 0; i < ndoms; i++)
7122 free_cpumask_var(doms[i]);
7127 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7128 * For now this just excludes isolated cpus, but could be used to
7129 * exclude other special cases in the future.
7131 static int init_sched_domains(const struct cpumask *cpu_map)
7135 arch_update_cpu_topology();
7137 doms_cur = alloc_sched_domains(ndoms_cur);
7139 doms_cur = &fallback_doms;
7140 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7141 err = build_sched_domains(doms_cur[0], NULL);
7142 register_sched_domain_sysctl();
7148 * Detach sched domains from a group of cpus specified in cpu_map
7149 * These cpus will now be attached to the NULL domain
7151 static void detach_destroy_domains(const struct cpumask *cpu_map)
7156 for_each_cpu(i, cpu_map)
7157 cpu_attach_domain(NULL, &def_root_domain, i);
7161 /* handle null as "default" */
7162 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7163 struct sched_domain_attr *new, int idx_new)
7165 struct sched_domain_attr tmp;
7172 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7173 new ? (new + idx_new) : &tmp,
7174 sizeof(struct sched_domain_attr));
7178 * Partition sched domains as specified by the 'ndoms_new'
7179 * cpumasks in the array doms_new[] of cpumasks. This compares
7180 * doms_new[] to the current sched domain partitioning, doms_cur[].
7181 * It destroys each deleted domain and builds each new domain.
7183 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7184 * The masks don't intersect (don't overlap.) We should setup one
7185 * sched domain for each mask. CPUs not in any of the cpumasks will
7186 * not be load balanced. If the same cpumask appears both in the
7187 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7190 * The passed in 'doms_new' should be allocated using
7191 * alloc_sched_domains. This routine takes ownership of it and will
7192 * free_sched_domains it when done with it. If the caller failed the
7193 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7194 * and partition_sched_domains() will fallback to the single partition
7195 * 'fallback_doms', it also forces the domains to be rebuilt.
7197 * If doms_new == NULL it will be replaced with cpu_online_mask.
7198 * ndoms_new == 0 is a special case for destroying existing domains,
7199 * and it will not create the default domain.
7201 * Call with hotplug lock held
7203 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7204 struct sched_domain_attr *dattr_new)
7209 mutex_lock(&sched_domains_mutex);
7211 /* always unregister in case we don't destroy any domains */
7212 unregister_sched_domain_sysctl();
7214 /* Let architecture update cpu core mappings. */
7215 new_topology = arch_update_cpu_topology();
7217 n = doms_new ? ndoms_new : 0;
7219 /* Destroy deleted domains */
7220 for (i = 0; i < ndoms_cur; i++) {
7221 for (j = 0; j < n && !new_topology; j++) {
7222 if (cpumask_equal(doms_cur[i], doms_new[j])
7223 && dattrs_equal(dattr_cur, i, dattr_new, j))
7226 /* no match - a current sched domain not in new doms_new[] */
7227 detach_destroy_domains(doms_cur[i]);
7233 if (doms_new == NULL) {
7235 doms_new = &fallback_doms;
7236 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7237 WARN_ON_ONCE(dattr_new);
7240 /* Build new domains */
7241 for (i = 0; i < ndoms_new; i++) {
7242 for (j = 0; j < n && !new_topology; j++) {
7243 if (cpumask_equal(doms_new[i], doms_cur[j])
7244 && dattrs_equal(dattr_new, i, dattr_cur, j))
7247 /* no match - add a new doms_new */
7248 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7253 /* Remember the new sched domains */
7254 if (doms_cur != &fallback_doms)
7255 free_sched_domains(doms_cur, ndoms_cur);
7256 kfree(dattr_cur); /* kfree(NULL) is safe */
7257 doms_cur = doms_new;
7258 dattr_cur = dattr_new;
7259 ndoms_cur = ndoms_new;
7261 register_sched_domain_sysctl();
7263 mutex_unlock(&sched_domains_mutex);
7266 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7269 * Update cpusets according to cpu_active mask. If cpusets are
7270 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7271 * around partition_sched_domains().
7273 * If we come here as part of a suspend/resume, don't touch cpusets because we
7274 * want to restore it back to its original state upon resume anyway.
7276 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7280 case CPU_ONLINE_FROZEN:
7281 case CPU_DOWN_FAILED_FROZEN:
7284 * num_cpus_frozen tracks how many CPUs are involved in suspend
7285 * resume sequence. As long as this is not the last online
7286 * operation in the resume sequence, just build a single sched
7287 * domain, ignoring cpusets.
7290 if (likely(num_cpus_frozen)) {
7291 partition_sched_domains(1, NULL, NULL);
7296 * This is the last CPU online operation. So fall through and
7297 * restore the original sched domains by considering the
7298 * cpuset configurations.
7302 cpuset_update_active_cpus(true);
7310 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7313 unsigned long flags;
7314 long cpu = (long)hcpu;
7320 case CPU_DOWN_PREPARE:
7321 rcu_read_lock_sched();
7322 dl_b = dl_bw_of(cpu);
7324 raw_spin_lock_irqsave(&dl_b->lock, flags);
7325 cpus = dl_bw_cpus(cpu);
7326 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7327 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7329 rcu_read_unlock_sched();
7332 return notifier_from_errno(-EBUSY);
7333 cpuset_update_active_cpus(false);
7335 case CPU_DOWN_PREPARE_FROZEN:
7337 partition_sched_domains(1, NULL, NULL);
7345 void __init sched_init_smp(void)
7347 cpumask_var_t non_isolated_cpus;
7349 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7350 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7355 * There's no userspace yet to cause hotplug operations; hence all the
7356 * cpu masks are stable and all blatant races in the below code cannot
7359 mutex_lock(&sched_domains_mutex);
7360 init_sched_domains(cpu_active_mask);
7361 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7362 if (cpumask_empty(non_isolated_cpus))
7363 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7364 mutex_unlock(&sched_domains_mutex);
7366 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7367 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7368 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7372 /* Move init over to a non-isolated CPU */
7373 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7375 sched_init_granularity();
7376 free_cpumask_var(non_isolated_cpus);
7378 init_sched_rt_class();
7379 init_sched_dl_class();
7382 void __init sched_init_smp(void)
7384 sched_init_granularity();
7386 #endif /* CONFIG_SMP */
7388 int in_sched_functions(unsigned long addr)
7390 return in_lock_functions(addr) ||
7391 (addr >= (unsigned long)__sched_text_start
7392 && addr < (unsigned long)__sched_text_end);
7395 #ifdef CONFIG_CGROUP_SCHED
7397 * Default task group.
7398 * Every task in system belongs to this group at bootup.
7400 struct task_group root_task_group;
7401 LIST_HEAD(task_groups);
7403 /* Cacheline aligned slab cache for task_group */
7404 static struct kmem_cache *task_group_cache __read_mostly;
7407 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7409 void __init sched_init(void)
7412 unsigned long alloc_size = 0, ptr;
7414 #ifdef CONFIG_FAIR_GROUP_SCHED
7415 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7417 #ifdef CONFIG_RT_GROUP_SCHED
7418 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7421 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7423 #ifdef CONFIG_FAIR_GROUP_SCHED
7424 root_task_group.se = (struct sched_entity **)ptr;
7425 ptr += nr_cpu_ids * sizeof(void **);
7427 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7428 ptr += nr_cpu_ids * sizeof(void **);
7430 #endif /* CONFIG_FAIR_GROUP_SCHED */
7431 #ifdef CONFIG_RT_GROUP_SCHED
7432 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7433 ptr += nr_cpu_ids * sizeof(void **);
7435 root_task_group.rt_rq = (struct rt_rq **)ptr;
7436 ptr += nr_cpu_ids * sizeof(void **);
7438 #endif /* CONFIG_RT_GROUP_SCHED */
7440 #ifdef CONFIG_CPUMASK_OFFSTACK
7441 for_each_possible_cpu(i) {
7442 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7443 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7445 #endif /* CONFIG_CPUMASK_OFFSTACK */
7447 init_rt_bandwidth(&def_rt_bandwidth,
7448 global_rt_period(), global_rt_runtime());
7449 init_dl_bandwidth(&def_dl_bandwidth,
7450 global_rt_period(), global_rt_runtime());
7453 init_defrootdomain();
7456 #ifdef CONFIG_RT_GROUP_SCHED
7457 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7458 global_rt_period(), global_rt_runtime());
7459 #endif /* CONFIG_RT_GROUP_SCHED */
7461 #ifdef CONFIG_CGROUP_SCHED
7462 task_group_cache = KMEM_CACHE(task_group, 0);
7464 list_add(&root_task_group.list, &task_groups);
7465 INIT_LIST_HEAD(&root_task_group.children);
7466 INIT_LIST_HEAD(&root_task_group.siblings);
7467 autogroup_init(&init_task);
7468 #endif /* CONFIG_CGROUP_SCHED */
7470 for_each_possible_cpu(i) {
7474 raw_spin_lock_init(&rq->lock);
7476 rq->calc_load_active = 0;
7477 rq->calc_load_update = jiffies + LOAD_FREQ;
7478 init_cfs_rq(&rq->cfs);
7479 init_rt_rq(&rq->rt);
7480 init_dl_rq(&rq->dl);
7481 #ifdef CONFIG_FAIR_GROUP_SCHED
7482 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7483 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7485 * How much cpu bandwidth does root_task_group get?
7487 * In case of task-groups formed thr' the cgroup filesystem, it
7488 * gets 100% of the cpu resources in the system. This overall
7489 * system cpu resource is divided among the tasks of
7490 * root_task_group and its child task-groups in a fair manner,
7491 * based on each entity's (task or task-group's) weight
7492 * (se->load.weight).
7494 * In other words, if root_task_group has 10 tasks of weight
7495 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7496 * then A0's share of the cpu resource is:
7498 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7500 * We achieve this by letting root_task_group's tasks sit
7501 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7503 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7504 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7505 #endif /* CONFIG_FAIR_GROUP_SCHED */
7507 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7508 #ifdef CONFIG_RT_GROUP_SCHED
7509 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7512 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7513 rq->cpu_load[j] = 0;
7515 rq->last_load_update_tick = jiffies;
7520 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7521 rq->balance_callback = NULL;
7522 rq->active_balance = 0;
7523 rq->next_balance = jiffies;
7528 rq->avg_idle = 2*sysctl_sched_migration_cost;
7529 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7531 INIT_LIST_HEAD(&rq->cfs_tasks);
7533 rq_attach_root(rq, &def_root_domain);
7534 #ifdef CONFIG_NO_HZ_COMMON
7537 #ifdef CONFIG_NO_HZ_FULL
7538 rq->last_sched_tick = 0;
7542 atomic_set(&rq->nr_iowait, 0);
7545 set_load_weight(&init_task);
7547 #ifdef CONFIG_PREEMPT_NOTIFIERS
7548 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7552 * The boot idle thread does lazy MMU switching as well:
7554 atomic_inc(&init_mm.mm_count);
7555 enter_lazy_tlb(&init_mm, current);
7558 * During early bootup we pretend to be a normal task:
7560 current->sched_class = &fair_sched_class;
7563 * Make us the idle thread. Technically, schedule() should not be
7564 * called from this thread, however somewhere below it might be,
7565 * but because we are the idle thread, we just pick up running again
7566 * when this runqueue becomes "idle".
7568 init_idle(current, smp_processor_id());
7570 calc_load_update = jiffies + LOAD_FREQ;
7573 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7574 /* May be allocated at isolcpus cmdline parse time */
7575 if (cpu_isolated_map == NULL)
7576 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7577 idle_thread_set_boot_cpu();
7578 set_cpu_rq_start_time();
7580 init_sched_fair_class();
7582 scheduler_running = 1;
7585 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7586 static inline int preempt_count_equals(int preempt_offset)
7588 int nested = preempt_count() + rcu_preempt_depth();
7590 return (nested == preempt_offset);
7593 void __might_sleep(const char *file, int line, int preempt_offset)
7596 * Blocking primitives will set (and therefore destroy) current->state,
7597 * since we will exit with TASK_RUNNING make sure we enter with it,
7598 * otherwise we will destroy state.
7600 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7601 "do not call blocking ops when !TASK_RUNNING; "
7602 "state=%lx set at [<%p>] %pS\n",
7604 (void *)current->task_state_change,
7605 (void *)current->task_state_change);
7607 ___might_sleep(file, line, preempt_offset);
7609 EXPORT_SYMBOL(__might_sleep);
7611 void ___might_sleep(const char *file, int line, int preempt_offset)
7613 static unsigned long prev_jiffy; /* ratelimiting */
7615 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7616 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7617 !is_idle_task(current)) ||
7618 system_state != SYSTEM_RUNNING || oops_in_progress)
7620 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7622 prev_jiffy = jiffies;
7625 "BUG: sleeping function called from invalid context at %s:%d\n",
7628 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7629 in_atomic(), irqs_disabled(),
7630 current->pid, current->comm);
7632 if (task_stack_end_corrupted(current))
7633 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7635 debug_show_held_locks(current);
7636 if (irqs_disabled())
7637 print_irqtrace_events(current);
7638 #ifdef CONFIG_DEBUG_PREEMPT
7639 if (!preempt_count_equals(preempt_offset)) {
7640 pr_err("Preemption disabled at:");
7641 print_ip_sym(current->preempt_disable_ip);
7647 EXPORT_SYMBOL(___might_sleep);
7650 #ifdef CONFIG_MAGIC_SYSRQ
7651 void normalize_rt_tasks(void)
7653 struct task_struct *g, *p;
7654 struct sched_attr attr = {
7655 .sched_policy = SCHED_NORMAL,
7658 read_lock(&tasklist_lock);
7659 for_each_process_thread(g, p) {
7661 * Only normalize user tasks:
7663 if (p->flags & PF_KTHREAD)
7666 p->se.exec_start = 0;
7667 #ifdef CONFIG_SCHEDSTATS
7668 p->se.statistics.wait_start = 0;
7669 p->se.statistics.sleep_start = 0;
7670 p->se.statistics.block_start = 0;
7673 if (!dl_task(p) && !rt_task(p)) {
7675 * Renice negative nice level userspace
7678 if (task_nice(p) < 0)
7679 set_user_nice(p, 0);
7683 __sched_setscheduler(p, &attr, false, false);
7685 read_unlock(&tasklist_lock);
7688 #endif /* CONFIG_MAGIC_SYSRQ */
7690 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7692 * These functions are only useful for the IA64 MCA handling, or kdb.
7694 * They can only be called when the whole system has been
7695 * stopped - every CPU needs to be quiescent, and no scheduling
7696 * activity can take place. Using them for anything else would
7697 * be a serious bug, and as a result, they aren't even visible
7698 * under any other configuration.
7702 * curr_task - return the current task for a given cpu.
7703 * @cpu: the processor in question.
7705 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7707 * Return: The current task for @cpu.
7709 struct task_struct *curr_task(int cpu)
7711 return cpu_curr(cpu);
7714 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7718 * set_curr_task - set the current task for a given cpu.
7719 * @cpu: the processor in question.
7720 * @p: the task pointer to set.
7722 * Description: This function must only be used when non-maskable interrupts
7723 * are serviced on a separate stack. It allows the architecture to switch the
7724 * notion of the current task on a cpu in a non-blocking manner. This function
7725 * must be called with all CPU's synchronized, and interrupts disabled, the
7726 * and caller must save the original value of the current task (see
7727 * curr_task() above) and restore that value before reenabling interrupts and
7728 * re-starting the system.
7730 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7732 void set_curr_task(int cpu, struct task_struct *p)
7739 #ifdef CONFIG_CGROUP_SCHED
7740 /* task_group_lock serializes the addition/removal of task groups */
7741 static DEFINE_SPINLOCK(task_group_lock);
7743 static void free_sched_group(struct task_group *tg)
7745 free_fair_sched_group(tg);
7746 free_rt_sched_group(tg);
7748 kmem_cache_free(task_group_cache, tg);
7751 /* allocate runqueue etc for a new task group */
7752 struct task_group *sched_create_group(struct task_group *parent)
7754 struct task_group *tg;
7756 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7758 return ERR_PTR(-ENOMEM);
7760 if (!alloc_fair_sched_group(tg, parent))
7763 if (!alloc_rt_sched_group(tg, parent))
7769 free_sched_group(tg);
7770 return ERR_PTR(-ENOMEM);
7773 void sched_online_group(struct task_group *tg, struct task_group *parent)
7775 unsigned long flags;
7777 spin_lock_irqsave(&task_group_lock, flags);
7778 list_add_rcu(&tg->list, &task_groups);
7780 WARN_ON(!parent); /* root should already exist */
7782 tg->parent = parent;
7783 INIT_LIST_HEAD(&tg->children);
7784 list_add_rcu(&tg->siblings, &parent->children);
7785 spin_unlock_irqrestore(&task_group_lock, flags);
7788 /* rcu callback to free various structures associated with a task group */
7789 static void free_sched_group_rcu(struct rcu_head *rhp)
7791 /* now it should be safe to free those cfs_rqs */
7792 free_sched_group(container_of(rhp, struct task_group, rcu));
7795 /* Destroy runqueue etc associated with a task group */
7796 void sched_destroy_group(struct task_group *tg)
7798 /* wait for possible concurrent references to cfs_rqs complete */
7799 call_rcu(&tg->rcu, free_sched_group_rcu);
7802 void sched_offline_group(struct task_group *tg)
7804 unsigned long flags;
7806 /* end participation in shares distribution */
7807 unregister_fair_sched_group(tg);
7809 spin_lock_irqsave(&task_group_lock, flags);
7810 list_del_rcu(&tg->list);
7811 list_del_rcu(&tg->siblings);
7812 spin_unlock_irqrestore(&task_group_lock, flags);
7815 /* change task's runqueue when it moves between groups.
7816 * The caller of this function should have put the task in its new group
7817 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7818 * reflect its new group.
7820 void sched_move_task(struct task_struct *tsk)
7822 struct task_group *tg;
7823 int queued, running;
7824 unsigned long flags;
7827 rq = task_rq_lock(tsk, &flags);
7829 running = task_current(rq, tsk);
7830 queued = task_on_rq_queued(tsk);
7833 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7834 if (unlikely(running))
7835 put_prev_task(rq, tsk);
7838 * All callers are synchronized by task_rq_lock(); we do not use RCU
7839 * which is pointless here. Thus, we pass "true" to task_css_check()
7840 * to prevent lockdep warnings.
7842 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7843 struct task_group, css);
7844 tg = autogroup_task_group(tsk, tg);
7845 tsk->sched_task_group = tg;
7847 #ifdef CONFIG_FAIR_GROUP_SCHED
7848 if (tsk->sched_class->task_move_group)
7849 tsk->sched_class->task_move_group(tsk);
7852 set_task_rq(tsk, task_cpu(tsk));
7854 if (unlikely(running))
7855 tsk->sched_class->set_curr_task(rq);
7857 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7859 task_rq_unlock(rq, tsk, &flags);
7861 #endif /* CONFIG_CGROUP_SCHED */
7863 #ifdef CONFIG_RT_GROUP_SCHED
7865 * Ensure that the real time constraints are schedulable.
7867 static DEFINE_MUTEX(rt_constraints_mutex);
7869 /* Must be called with tasklist_lock held */
7870 static inline int tg_has_rt_tasks(struct task_group *tg)
7872 struct task_struct *g, *p;
7875 * Autogroups do not have RT tasks; see autogroup_create().
7877 if (task_group_is_autogroup(tg))
7880 for_each_process_thread(g, p) {
7881 if (rt_task(p) && task_group(p) == tg)
7888 struct rt_schedulable_data {
7889 struct task_group *tg;
7894 static int tg_rt_schedulable(struct task_group *tg, void *data)
7896 struct rt_schedulable_data *d = data;
7897 struct task_group *child;
7898 unsigned long total, sum = 0;
7899 u64 period, runtime;
7901 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7902 runtime = tg->rt_bandwidth.rt_runtime;
7905 period = d->rt_period;
7906 runtime = d->rt_runtime;
7910 * Cannot have more runtime than the period.
7912 if (runtime > period && runtime != RUNTIME_INF)
7916 * Ensure we don't starve existing RT tasks.
7918 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7921 total = to_ratio(period, runtime);
7924 * Nobody can have more than the global setting allows.
7926 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7930 * The sum of our children's runtime should not exceed our own.
7932 list_for_each_entry_rcu(child, &tg->children, siblings) {
7933 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7934 runtime = child->rt_bandwidth.rt_runtime;
7936 if (child == d->tg) {
7937 period = d->rt_period;
7938 runtime = d->rt_runtime;
7941 sum += to_ratio(period, runtime);
7950 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7954 struct rt_schedulable_data data = {
7956 .rt_period = period,
7957 .rt_runtime = runtime,
7961 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7967 static int tg_set_rt_bandwidth(struct task_group *tg,
7968 u64 rt_period, u64 rt_runtime)
7973 * Disallowing the root group RT runtime is BAD, it would disallow the
7974 * kernel creating (and or operating) RT threads.
7976 if (tg == &root_task_group && rt_runtime == 0)
7979 /* No period doesn't make any sense. */
7983 mutex_lock(&rt_constraints_mutex);
7984 read_lock(&tasklist_lock);
7985 err = __rt_schedulable(tg, rt_period, rt_runtime);
7989 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7990 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7991 tg->rt_bandwidth.rt_runtime = rt_runtime;
7993 for_each_possible_cpu(i) {
7994 struct rt_rq *rt_rq = tg->rt_rq[i];
7996 raw_spin_lock(&rt_rq->rt_runtime_lock);
7997 rt_rq->rt_runtime = rt_runtime;
7998 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8000 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8002 read_unlock(&tasklist_lock);
8003 mutex_unlock(&rt_constraints_mutex);
8008 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8010 u64 rt_runtime, rt_period;
8012 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8013 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8014 if (rt_runtime_us < 0)
8015 rt_runtime = RUNTIME_INF;
8017 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8020 static long sched_group_rt_runtime(struct task_group *tg)
8024 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8027 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8028 do_div(rt_runtime_us, NSEC_PER_USEC);
8029 return rt_runtime_us;
8032 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8034 u64 rt_runtime, rt_period;
8036 rt_period = rt_period_us * NSEC_PER_USEC;
8037 rt_runtime = tg->rt_bandwidth.rt_runtime;
8039 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8042 static long sched_group_rt_period(struct task_group *tg)
8046 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8047 do_div(rt_period_us, NSEC_PER_USEC);
8048 return rt_period_us;
8050 #endif /* CONFIG_RT_GROUP_SCHED */
8052 #ifdef CONFIG_RT_GROUP_SCHED
8053 static int sched_rt_global_constraints(void)
8057 mutex_lock(&rt_constraints_mutex);
8058 read_lock(&tasklist_lock);
8059 ret = __rt_schedulable(NULL, 0, 0);
8060 read_unlock(&tasklist_lock);
8061 mutex_unlock(&rt_constraints_mutex);
8066 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8068 /* Don't accept realtime tasks when there is no way for them to run */
8069 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8075 #else /* !CONFIG_RT_GROUP_SCHED */
8076 static int sched_rt_global_constraints(void)
8078 unsigned long flags;
8081 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8082 for_each_possible_cpu(i) {
8083 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8085 raw_spin_lock(&rt_rq->rt_runtime_lock);
8086 rt_rq->rt_runtime = global_rt_runtime();
8087 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8089 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8093 #endif /* CONFIG_RT_GROUP_SCHED */
8095 static int sched_dl_global_validate(void)
8097 u64 runtime = global_rt_runtime();
8098 u64 period = global_rt_period();
8099 u64 new_bw = to_ratio(period, runtime);
8102 unsigned long flags;
8105 * Here we want to check the bandwidth not being set to some
8106 * value smaller than the currently allocated bandwidth in
8107 * any of the root_domains.
8109 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8110 * cycling on root_domains... Discussion on different/better
8111 * solutions is welcome!
8113 for_each_possible_cpu(cpu) {
8114 rcu_read_lock_sched();
8115 dl_b = dl_bw_of(cpu);
8117 raw_spin_lock_irqsave(&dl_b->lock, flags);
8118 if (new_bw < dl_b->total_bw)
8120 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8122 rcu_read_unlock_sched();
8131 static void sched_dl_do_global(void)
8136 unsigned long flags;
8138 def_dl_bandwidth.dl_period = global_rt_period();
8139 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8141 if (global_rt_runtime() != RUNTIME_INF)
8142 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8145 * FIXME: As above...
8147 for_each_possible_cpu(cpu) {
8148 rcu_read_lock_sched();
8149 dl_b = dl_bw_of(cpu);
8151 raw_spin_lock_irqsave(&dl_b->lock, flags);
8153 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8155 rcu_read_unlock_sched();
8159 static int sched_rt_global_validate(void)
8161 if (sysctl_sched_rt_period <= 0)
8164 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8165 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8171 static void sched_rt_do_global(void)
8173 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8174 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8177 int sched_rt_handler(struct ctl_table *table, int write,
8178 void __user *buffer, size_t *lenp,
8181 int old_period, old_runtime;
8182 static DEFINE_MUTEX(mutex);
8186 old_period = sysctl_sched_rt_period;
8187 old_runtime = sysctl_sched_rt_runtime;
8189 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8191 if (!ret && write) {
8192 ret = sched_rt_global_validate();
8196 ret = sched_dl_global_validate();
8200 ret = sched_rt_global_constraints();
8204 sched_rt_do_global();
8205 sched_dl_do_global();
8209 sysctl_sched_rt_period = old_period;
8210 sysctl_sched_rt_runtime = old_runtime;
8212 mutex_unlock(&mutex);
8217 int sched_rr_handler(struct ctl_table *table, int write,
8218 void __user *buffer, size_t *lenp,
8222 static DEFINE_MUTEX(mutex);
8225 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8226 /* make sure that internally we keep jiffies */
8227 /* also, writing zero resets timeslice to default */
8228 if (!ret && write) {
8229 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8230 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8232 mutex_unlock(&mutex);
8236 #ifdef CONFIG_CGROUP_SCHED
8238 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8240 return css ? container_of(css, struct task_group, css) : NULL;
8243 static struct cgroup_subsys_state *
8244 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8246 struct task_group *parent = css_tg(parent_css);
8247 struct task_group *tg;
8250 /* This is early initialization for the top cgroup */
8251 return &root_task_group.css;
8254 tg = sched_create_group(parent);
8256 return ERR_PTR(-ENOMEM);
8261 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8263 struct task_group *tg = css_tg(css);
8264 struct task_group *parent = css_tg(css->parent);
8267 sched_online_group(tg, parent);
8271 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8273 struct task_group *tg = css_tg(css);
8275 sched_destroy_group(tg);
8278 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8280 struct task_group *tg = css_tg(css);
8282 sched_offline_group(tg);
8285 static void cpu_cgroup_fork(struct task_struct *task)
8287 sched_move_task(task);
8290 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8292 struct task_struct *task;
8293 struct cgroup_subsys_state *css;
8295 cgroup_taskset_for_each(task, css, tset) {
8296 #ifdef CONFIG_RT_GROUP_SCHED
8297 if (!sched_rt_can_attach(css_tg(css), task))
8300 /* We don't support RT-tasks being in separate groups */
8301 if (task->sched_class != &fair_sched_class)
8308 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8310 struct task_struct *task;
8311 struct cgroup_subsys_state *css;
8313 cgroup_taskset_for_each(task, css, tset)
8314 sched_move_task(task);
8317 #ifdef CONFIG_FAIR_GROUP_SCHED
8318 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8319 struct cftype *cftype, u64 shareval)
8321 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8324 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8327 struct task_group *tg = css_tg(css);
8329 return (u64) scale_load_down(tg->shares);
8332 #ifdef CONFIG_CFS_BANDWIDTH
8333 static DEFINE_MUTEX(cfs_constraints_mutex);
8335 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8336 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8338 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8340 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8342 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8343 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8345 if (tg == &root_task_group)
8349 * Ensure we have at some amount of bandwidth every period. This is
8350 * to prevent reaching a state of large arrears when throttled via
8351 * entity_tick() resulting in prolonged exit starvation.
8353 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8357 * Likewise, bound things on the otherside by preventing insane quota
8358 * periods. This also allows us to normalize in computing quota
8361 if (period > max_cfs_quota_period)
8365 * Prevent race between setting of cfs_rq->runtime_enabled and
8366 * unthrottle_offline_cfs_rqs().
8369 mutex_lock(&cfs_constraints_mutex);
8370 ret = __cfs_schedulable(tg, period, quota);
8374 runtime_enabled = quota != RUNTIME_INF;
8375 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8377 * If we need to toggle cfs_bandwidth_used, off->on must occur
8378 * before making related changes, and on->off must occur afterwards
8380 if (runtime_enabled && !runtime_was_enabled)
8381 cfs_bandwidth_usage_inc();
8382 raw_spin_lock_irq(&cfs_b->lock);
8383 cfs_b->period = ns_to_ktime(period);
8384 cfs_b->quota = quota;
8386 __refill_cfs_bandwidth_runtime(cfs_b);
8387 /* restart the period timer (if active) to handle new period expiry */
8388 if (runtime_enabled)
8389 start_cfs_bandwidth(cfs_b);
8390 raw_spin_unlock_irq(&cfs_b->lock);
8392 for_each_online_cpu(i) {
8393 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8394 struct rq *rq = cfs_rq->rq;
8396 raw_spin_lock_irq(&rq->lock);
8397 cfs_rq->runtime_enabled = runtime_enabled;
8398 cfs_rq->runtime_remaining = 0;
8400 if (cfs_rq->throttled)
8401 unthrottle_cfs_rq(cfs_rq);
8402 raw_spin_unlock_irq(&rq->lock);
8404 if (runtime_was_enabled && !runtime_enabled)
8405 cfs_bandwidth_usage_dec();
8407 mutex_unlock(&cfs_constraints_mutex);
8413 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8417 period = ktime_to_ns(tg->cfs_bandwidth.period);
8418 if (cfs_quota_us < 0)
8419 quota = RUNTIME_INF;
8421 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8423 return tg_set_cfs_bandwidth(tg, period, quota);
8426 long tg_get_cfs_quota(struct task_group *tg)
8430 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8433 quota_us = tg->cfs_bandwidth.quota;
8434 do_div(quota_us, NSEC_PER_USEC);
8439 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8443 period = (u64)cfs_period_us * NSEC_PER_USEC;
8444 quota = tg->cfs_bandwidth.quota;
8446 return tg_set_cfs_bandwidth(tg, period, quota);
8449 long tg_get_cfs_period(struct task_group *tg)
8453 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8454 do_div(cfs_period_us, NSEC_PER_USEC);
8456 return cfs_period_us;
8459 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8462 return tg_get_cfs_quota(css_tg(css));
8465 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8466 struct cftype *cftype, s64 cfs_quota_us)
8468 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8471 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8474 return tg_get_cfs_period(css_tg(css));
8477 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8478 struct cftype *cftype, u64 cfs_period_us)
8480 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8483 struct cfs_schedulable_data {
8484 struct task_group *tg;
8489 * normalize group quota/period to be quota/max_period
8490 * note: units are usecs
8492 static u64 normalize_cfs_quota(struct task_group *tg,
8493 struct cfs_schedulable_data *d)
8501 period = tg_get_cfs_period(tg);
8502 quota = tg_get_cfs_quota(tg);
8505 /* note: these should typically be equivalent */
8506 if (quota == RUNTIME_INF || quota == -1)
8509 return to_ratio(period, quota);
8512 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8514 struct cfs_schedulable_data *d = data;
8515 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8516 s64 quota = 0, parent_quota = -1;
8519 quota = RUNTIME_INF;
8521 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8523 quota = normalize_cfs_quota(tg, d);
8524 parent_quota = parent_b->hierarchical_quota;
8527 * ensure max(child_quota) <= parent_quota, inherit when no
8530 if (quota == RUNTIME_INF)
8531 quota = parent_quota;
8532 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8535 cfs_b->hierarchical_quota = quota;
8540 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8543 struct cfs_schedulable_data data = {
8549 if (quota != RUNTIME_INF) {
8550 do_div(data.period, NSEC_PER_USEC);
8551 do_div(data.quota, NSEC_PER_USEC);
8555 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8561 static int cpu_stats_show(struct seq_file *sf, void *v)
8563 struct task_group *tg = css_tg(seq_css(sf));
8564 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8566 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8567 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8568 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8572 #endif /* CONFIG_CFS_BANDWIDTH */
8573 #endif /* CONFIG_FAIR_GROUP_SCHED */
8575 #ifdef CONFIG_RT_GROUP_SCHED
8576 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8577 struct cftype *cft, s64 val)
8579 return sched_group_set_rt_runtime(css_tg(css), val);
8582 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8585 return sched_group_rt_runtime(css_tg(css));
8588 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8589 struct cftype *cftype, u64 rt_period_us)
8591 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8594 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8597 return sched_group_rt_period(css_tg(css));
8599 #endif /* CONFIG_RT_GROUP_SCHED */
8601 static struct cftype cpu_files[] = {
8602 #ifdef CONFIG_FAIR_GROUP_SCHED
8605 .read_u64 = cpu_shares_read_u64,
8606 .write_u64 = cpu_shares_write_u64,
8609 #ifdef CONFIG_CFS_BANDWIDTH
8611 .name = "cfs_quota_us",
8612 .read_s64 = cpu_cfs_quota_read_s64,
8613 .write_s64 = cpu_cfs_quota_write_s64,
8616 .name = "cfs_period_us",
8617 .read_u64 = cpu_cfs_period_read_u64,
8618 .write_u64 = cpu_cfs_period_write_u64,
8622 .seq_show = cpu_stats_show,
8625 #ifdef CONFIG_RT_GROUP_SCHED
8627 .name = "rt_runtime_us",
8628 .read_s64 = cpu_rt_runtime_read,
8629 .write_s64 = cpu_rt_runtime_write,
8632 .name = "rt_period_us",
8633 .read_u64 = cpu_rt_period_read_uint,
8634 .write_u64 = cpu_rt_period_write_uint,
8640 struct cgroup_subsys cpu_cgrp_subsys = {
8641 .css_alloc = cpu_cgroup_css_alloc,
8642 .css_free = cpu_cgroup_css_free,
8643 .css_online = cpu_cgroup_css_online,
8644 .css_offline = cpu_cgroup_css_offline,
8645 .fork = cpu_cgroup_fork,
8646 .can_attach = cpu_cgroup_can_attach,
8647 .attach = cpu_cgroup_attach,
8648 .legacy_cftypes = cpu_files,
8652 #endif /* CONFIG_CGROUP_SCHED */
8654 void dump_cpu_task(int cpu)
8656 pr_info("Task dump for CPU %d:\n", cpu);
8657 sched_show_task(cpu_curr(cpu));
8661 * Nice levels are multiplicative, with a gentle 10% change for every
8662 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8663 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8664 * that remained on nice 0.
8666 * The "10% effect" is relative and cumulative: from _any_ nice level,
8667 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8668 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8669 * If a task goes up by ~10% and another task goes down by ~10% then
8670 * the relative distance between them is ~25%.)
8672 const int sched_prio_to_weight[40] = {
8673 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8674 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8675 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8676 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8677 /* 0 */ 1024, 820, 655, 526, 423,
8678 /* 5 */ 335, 272, 215, 172, 137,
8679 /* 10 */ 110, 87, 70, 56, 45,
8680 /* 15 */ 36, 29, 23, 18, 15,
8684 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8686 * In cases where the weight does not change often, we can use the
8687 * precalculated inverse to speed up arithmetics by turning divisions
8688 * into multiplications:
8690 const u32 sched_prio_to_wmult[40] = {
8691 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8692 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8693 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8694 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8695 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8696 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8697 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8698 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,