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
29 #include <linux/kasan.h>
31 #include <linux/module.h>
32 #include <linux/nmi.h>
33 #include <linux/init.h>
34 #include <linux/uaccess.h>
35 #include <linux/highmem.h>
36 #include <linux/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/context_tracking.h>
75 #include <linux/compiler.h>
76 #include <linux/frame.h>
77 #include <linux/prefetch.h>
79 #include <asm/switch_to.h>
81 #include <asm/irq_regs.h>
82 #include <asm/mutex.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
94 DEFINE_MUTEX(sched_domains_mutex);
95 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
97 static void update_rq_clock_task(struct rq *rq, s64 delta);
99 void update_rq_clock(struct rq *rq)
103 lockdep_assert_held(&rq->lock);
105 if (rq->clock_skip_update & RQCF_ACT_SKIP)
108 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
112 update_rq_clock_task(rq, delta);
116 * Debugging: various feature bits
119 #define SCHED_FEAT(name, enabled) \
120 (1UL << __SCHED_FEAT_##name) * enabled |
122 const_debug unsigned int sysctl_sched_features =
123 #include "features.h"
129 * Number of tasks to iterate in a single balance run.
130 * Limited because this is done with IRQs disabled.
132 const_debug unsigned int sysctl_sched_nr_migrate = 32;
135 * period over which we average the RT time consumption, measured
140 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
143 * period over which we measure -rt task cpu usage in us.
146 unsigned int sysctl_sched_rt_period = 1000000;
148 __read_mostly int scheduler_running;
151 * part of the period that we allow rt tasks to run in us.
154 int sysctl_sched_rt_runtime = 950000;
156 /* cpus with isolated domains */
157 cpumask_var_t cpu_isolated_map;
160 * this_rq_lock - lock this runqueue and disable interrupts.
162 static struct rq *this_rq_lock(void)
169 raw_spin_lock(&rq->lock);
175 * __task_rq_lock - lock the rq @p resides on.
177 struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
182 lockdep_assert_held(&p->pi_lock);
186 raw_spin_lock(&rq->lock);
187 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
188 rf->cookie = lockdep_pin_lock(&rq->lock);
191 raw_spin_unlock(&rq->lock);
193 while (unlikely(task_on_rq_migrating(p)))
199 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
201 struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
202 __acquires(p->pi_lock)
208 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
210 raw_spin_lock(&rq->lock);
212 * move_queued_task() task_rq_lock()
215 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
216 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
217 * [S] ->cpu = new_cpu [L] task_rq()
221 * If we observe the old cpu in task_rq_lock, the acquire of
222 * the old rq->lock will fully serialize against the stores.
224 * If we observe the new cpu in task_rq_lock, the acquire will
225 * pair with the WMB to ensure we must then also see migrating.
227 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
228 rf->cookie = lockdep_pin_lock(&rq->lock);
231 raw_spin_unlock(&rq->lock);
232 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
234 while (unlikely(task_on_rq_migrating(p)))
239 #ifdef CONFIG_SCHED_HRTICK
241 * Use HR-timers to deliver accurate preemption points.
244 static void hrtick_clear(struct rq *rq)
246 if (hrtimer_active(&rq->hrtick_timer))
247 hrtimer_cancel(&rq->hrtick_timer);
251 * High-resolution timer tick.
252 * Runs from hardirq context with interrupts disabled.
254 static enum hrtimer_restart hrtick(struct hrtimer *timer)
256 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
258 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
260 raw_spin_lock(&rq->lock);
262 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
263 raw_spin_unlock(&rq->lock);
265 return HRTIMER_NORESTART;
270 static void __hrtick_restart(struct rq *rq)
272 struct hrtimer *timer = &rq->hrtick_timer;
274 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
278 * called from hardirq (IPI) context
280 static void __hrtick_start(void *arg)
284 raw_spin_lock(&rq->lock);
285 __hrtick_restart(rq);
286 rq->hrtick_csd_pending = 0;
287 raw_spin_unlock(&rq->lock);
291 * Called to set the hrtick timer state.
293 * called with rq->lock held and irqs disabled
295 void hrtick_start(struct rq *rq, u64 delay)
297 struct hrtimer *timer = &rq->hrtick_timer;
302 * Don't schedule slices shorter than 10000ns, that just
303 * doesn't make sense and can cause timer DoS.
305 delta = max_t(s64, delay, 10000LL);
306 time = ktime_add_ns(timer->base->get_time(), delta);
308 hrtimer_set_expires(timer, time);
310 if (rq == this_rq()) {
311 __hrtick_restart(rq);
312 } else if (!rq->hrtick_csd_pending) {
313 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
314 rq->hrtick_csd_pending = 1;
320 * Called to set the hrtick timer state.
322 * called with rq->lock held and irqs disabled
324 void hrtick_start(struct rq *rq, u64 delay)
327 * Don't schedule slices shorter than 10000ns, that just
328 * doesn't make sense. Rely on vruntime for fairness.
330 delay = max_t(u64, delay, 10000LL);
331 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
332 HRTIMER_MODE_REL_PINNED);
334 #endif /* CONFIG_SMP */
336 static void init_rq_hrtick(struct rq *rq)
339 rq->hrtick_csd_pending = 0;
341 rq->hrtick_csd.flags = 0;
342 rq->hrtick_csd.func = __hrtick_start;
343 rq->hrtick_csd.info = rq;
346 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
347 rq->hrtick_timer.function = hrtick;
349 #else /* CONFIG_SCHED_HRTICK */
350 static inline void hrtick_clear(struct rq *rq)
354 static inline void init_rq_hrtick(struct rq *rq)
357 #endif /* CONFIG_SCHED_HRTICK */
360 * cmpxchg based fetch_or, macro so it works for different integer types
362 #define fetch_or(ptr, mask) \
364 typeof(ptr) _ptr = (ptr); \
365 typeof(mask) _mask = (mask); \
366 typeof(*_ptr) _old, _val = *_ptr; \
369 _old = cmpxchg(_ptr, _val, _val | _mask); \
377 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
379 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
380 * this avoids any races wrt polling state changes and thereby avoids
383 static bool set_nr_and_not_polling(struct task_struct *p)
385 struct thread_info *ti = task_thread_info(p);
386 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
390 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
392 * If this returns true, then the idle task promises to call
393 * sched_ttwu_pending() and reschedule soon.
395 static bool set_nr_if_polling(struct task_struct *p)
397 struct thread_info *ti = task_thread_info(p);
398 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
401 if (!(val & _TIF_POLLING_NRFLAG))
403 if (val & _TIF_NEED_RESCHED)
405 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
414 static bool set_nr_and_not_polling(struct task_struct *p)
416 set_tsk_need_resched(p);
421 static bool set_nr_if_polling(struct task_struct *p)
428 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
430 struct wake_q_node *node = &task->wake_q;
433 * Atomically grab the task, if ->wake_q is !nil already it means
434 * its already queued (either by us or someone else) and will get the
435 * wakeup due to that.
437 * This cmpxchg() implies a full barrier, which pairs with the write
438 * barrier implied by the wakeup in wake_up_q().
440 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
443 get_task_struct(task);
446 * The head is context local, there can be no concurrency.
449 head->lastp = &node->next;
452 void wake_up_q(struct wake_q_head *head)
454 struct wake_q_node *node = head->first;
456 while (node != WAKE_Q_TAIL) {
457 struct task_struct *task;
459 task = container_of(node, struct task_struct, wake_q);
461 /* task can safely be re-inserted now */
463 task->wake_q.next = NULL;
466 * wake_up_process() implies a wmb() to pair with the queueing
467 * in wake_q_add() so as not to miss wakeups.
469 wake_up_process(task);
470 put_task_struct(task);
475 * resched_curr - mark rq's current task 'to be rescheduled now'.
477 * On UP this means the setting of the need_resched flag, on SMP it
478 * might also involve a cross-CPU call to trigger the scheduler on
481 void resched_curr(struct rq *rq)
483 struct task_struct *curr = rq->curr;
486 lockdep_assert_held(&rq->lock);
488 if (test_tsk_need_resched(curr))
493 if (cpu == smp_processor_id()) {
494 set_tsk_need_resched(curr);
495 set_preempt_need_resched();
499 if (set_nr_and_not_polling(curr))
500 smp_send_reschedule(cpu);
502 trace_sched_wake_idle_without_ipi(cpu);
505 void resched_cpu(int cpu)
507 struct rq *rq = cpu_rq(cpu);
510 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
513 raw_spin_unlock_irqrestore(&rq->lock, flags);
517 #ifdef CONFIG_NO_HZ_COMMON
519 * In the semi idle case, use the nearest busy cpu for migrating timers
520 * from an idle cpu. This is good for power-savings.
522 * We don't do similar optimization for completely idle system, as
523 * selecting an idle cpu will add more delays to the timers than intended
524 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
526 int get_nohz_timer_target(void)
528 int i, cpu = smp_processor_id();
529 struct sched_domain *sd;
531 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
535 for_each_domain(cpu, sd) {
536 for_each_cpu(i, sched_domain_span(sd)) {
540 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
547 if (!is_housekeeping_cpu(cpu))
548 cpu = housekeeping_any_cpu();
554 * When add_timer_on() enqueues a timer into the timer wheel of an
555 * idle CPU then this timer might expire before the next timer event
556 * which is scheduled to wake up that CPU. In case of a completely
557 * idle system the next event might even be infinite time into the
558 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
559 * leaves the inner idle loop so the newly added timer is taken into
560 * account when the CPU goes back to idle and evaluates the timer
561 * wheel for the next timer event.
563 static void wake_up_idle_cpu(int cpu)
565 struct rq *rq = cpu_rq(cpu);
567 if (cpu == smp_processor_id())
570 if (set_nr_and_not_polling(rq->idle))
571 smp_send_reschedule(cpu);
573 trace_sched_wake_idle_without_ipi(cpu);
576 static bool wake_up_full_nohz_cpu(int cpu)
579 * We just need the target to call irq_exit() and re-evaluate
580 * the next tick. The nohz full kick at least implies that.
581 * If needed we can still optimize that later with an
584 if (tick_nohz_full_cpu(cpu)) {
585 if (cpu != smp_processor_id() ||
586 tick_nohz_tick_stopped())
587 tick_nohz_full_kick_cpu(cpu);
594 void wake_up_nohz_cpu(int cpu)
596 if (!wake_up_full_nohz_cpu(cpu))
597 wake_up_idle_cpu(cpu);
600 static inline bool got_nohz_idle_kick(void)
602 int cpu = smp_processor_id();
604 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
607 if (idle_cpu(cpu) && !need_resched())
611 * We can't run Idle Load Balance on this CPU for this time so we
612 * cancel it and clear NOHZ_BALANCE_KICK
614 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
618 #else /* CONFIG_NO_HZ_COMMON */
620 static inline bool got_nohz_idle_kick(void)
625 #endif /* CONFIG_NO_HZ_COMMON */
627 #ifdef CONFIG_NO_HZ_FULL
628 bool sched_can_stop_tick(struct rq *rq)
632 /* Deadline tasks, even if single, need the tick */
633 if (rq->dl.dl_nr_running)
637 * If there are more than one RR tasks, we need the tick to effect the
638 * actual RR behaviour.
640 if (rq->rt.rr_nr_running) {
641 if (rq->rt.rr_nr_running == 1)
648 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
649 * forced preemption between FIFO tasks.
651 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
656 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
657 * if there's more than one we need the tick for involuntary
660 if (rq->nr_running > 1)
665 #endif /* CONFIG_NO_HZ_FULL */
667 void sched_avg_update(struct rq *rq)
669 s64 period = sched_avg_period();
671 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
673 * Inline assembly required to prevent the compiler
674 * optimising this loop into a divmod call.
675 * See __iter_div_u64_rem() for another example of this.
677 asm("" : "+rm" (rq->age_stamp));
678 rq->age_stamp += period;
683 #endif /* CONFIG_SMP */
685 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
686 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
688 * Iterate task_group tree rooted at *from, calling @down when first entering a
689 * node and @up when leaving it for the final time.
691 * Caller must hold rcu_lock or sufficient equivalent.
693 int walk_tg_tree_from(struct task_group *from,
694 tg_visitor down, tg_visitor up, void *data)
696 struct task_group *parent, *child;
702 ret = (*down)(parent, data);
705 list_for_each_entry_rcu(child, &parent->children, siblings) {
712 ret = (*up)(parent, data);
713 if (ret || parent == from)
717 parent = parent->parent;
724 int tg_nop(struct task_group *tg, void *data)
730 static void set_load_weight(struct task_struct *p)
732 int prio = p->static_prio - MAX_RT_PRIO;
733 struct load_weight *load = &p->se.load;
736 * SCHED_IDLE tasks get minimal weight:
738 if (idle_policy(p->policy)) {
739 load->weight = scale_load(WEIGHT_IDLEPRIO);
740 load->inv_weight = WMULT_IDLEPRIO;
744 load->weight = scale_load(sched_prio_to_weight[prio]);
745 load->inv_weight = sched_prio_to_wmult[prio];
748 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
751 if (!(flags & ENQUEUE_RESTORE))
752 sched_info_queued(rq, p);
753 p->sched_class->enqueue_task(rq, p, flags);
756 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
759 if (!(flags & DEQUEUE_SAVE))
760 sched_info_dequeued(rq, p);
761 p->sched_class->dequeue_task(rq, p, flags);
764 void activate_task(struct rq *rq, struct task_struct *p, int flags)
766 if (task_contributes_to_load(p))
767 rq->nr_uninterruptible--;
769 enqueue_task(rq, p, flags);
772 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
774 if (task_contributes_to_load(p))
775 rq->nr_uninterruptible++;
777 dequeue_task(rq, p, flags);
780 static void update_rq_clock_task(struct rq *rq, s64 delta)
783 * In theory, the compile should just see 0 here, and optimize out the call
784 * to sched_rt_avg_update. But I don't trust it...
786 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
787 s64 steal = 0, irq_delta = 0;
789 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
790 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
793 * Since irq_time is only updated on {soft,}irq_exit, we might run into
794 * this case when a previous update_rq_clock() happened inside a
797 * When this happens, we stop ->clock_task and only update the
798 * prev_irq_time stamp to account for the part that fit, so that a next
799 * update will consume the rest. This ensures ->clock_task is
802 * It does however cause some slight miss-attribution of {soft,}irq
803 * time, a more accurate solution would be to update the irq_time using
804 * the current rq->clock timestamp, except that would require using
807 if (irq_delta > delta)
810 rq->prev_irq_time += irq_delta;
813 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
814 if (static_key_false((¶virt_steal_rq_enabled))) {
815 steal = paravirt_steal_clock(cpu_of(rq));
816 steal -= rq->prev_steal_time_rq;
818 if (unlikely(steal > delta))
821 rq->prev_steal_time_rq += steal;
826 rq->clock_task += delta;
828 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
829 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
830 sched_rt_avg_update(rq, irq_delta + steal);
834 void sched_set_stop_task(int cpu, struct task_struct *stop)
836 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
837 struct task_struct *old_stop = cpu_rq(cpu)->stop;
841 * Make it appear like a SCHED_FIFO task, its something
842 * userspace knows about and won't get confused about.
844 * Also, it will make PI more or less work without too
845 * much confusion -- but then, stop work should not
846 * rely on PI working anyway.
848 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
850 stop->sched_class = &stop_sched_class;
853 cpu_rq(cpu)->stop = stop;
857 * Reset it back to a normal scheduling class so that
858 * it can die in pieces.
860 old_stop->sched_class = &rt_sched_class;
865 * __normal_prio - return the priority that is based on the static prio
867 static inline int __normal_prio(struct task_struct *p)
869 return p->static_prio;
873 * Calculate the expected normal priority: i.e. priority
874 * without taking RT-inheritance into account. Might be
875 * boosted by interactivity modifiers. Changes upon fork,
876 * setprio syscalls, and whenever the interactivity
877 * estimator recalculates.
879 static inline int normal_prio(struct task_struct *p)
883 if (task_has_dl_policy(p))
884 prio = MAX_DL_PRIO-1;
885 else if (task_has_rt_policy(p))
886 prio = MAX_RT_PRIO-1 - p->rt_priority;
888 prio = __normal_prio(p);
893 * Calculate the current priority, i.e. the priority
894 * taken into account by the scheduler. This value might
895 * be boosted by RT tasks, or might be boosted by
896 * interactivity modifiers. Will be RT if the task got
897 * RT-boosted. If not then it returns p->normal_prio.
899 static int effective_prio(struct task_struct *p)
901 p->normal_prio = normal_prio(p);
903 * If we are RT tasks or we were boosted to RT priority,
904 * keep the priority unchanged. Otherwise, update priority
905 * to the normal priority:
907 if (!rt_prio(p->prio))
908 return p->normal_prio;
913 * task_curr - is this task currently executing on a CPU?
914 * @p: the task in question.
916 * Return: 1 if the task is currently executing. 0 otherwise.
918 inline int task_curr(const struct task_struct *p)
920 return cpu_curr(task_cpu(p)) == p;
924 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
925 * use the balance_callback list if you want balancing.
927 * this means any call to check_class_changed() must be followed by a call to
928 * balance_callback().
930 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
931 const struct sched_class *prev_class,
934 if (prev_class != p->sched_class) {
935 if (prev_class->switched_from)
936 prev_class->switched_from(rq, p);
938 p->sched_class->switched_to(rq, p);
939 } else if (oldprio != p->prio || dl_task(p))
940 p->sched_class->prio_changed(rq, p, oldprio);
943 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
945 const struct sched_class *class;
947 if (p->sched_class == rq->curr->sched_class) {
948 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
950 for_each_class(class) {
951 if (class == rq->curr->sched_class)
953 if (class == p->sched_class) {
961 * A queue event has occurred, and we're going to schedule. In
962 * this case, we can save a useless back to back clock update.
964 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
965 rq_clock_skip_update(rq, true);
970 * This is how migration works:
972 * 1) we invoke migration_cpu_stop() on the target CPU using
974 * 2) stopper starts to run (implicitly forcing the migrated thread
976 * 3) it checks whether the migrated task is still in the wrong runqueue.
977 * 4) if it's in the wrong runqueue then the migration thread removes
978 * it and puts it into the right queue.
979 * 5) stopper completes and stop_one_cpu() returns and the migration
984 * move_queued_task - move a queued task to new rq.
986 * Returns (locked) new rq. Old rq's lock is released.
988 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
990 lockdep_assert_held(&rq->lock);
992 p->on_rq = TASK_ON_RQ_MIGRATING;
993 dequeue_task(rq, p, 0);
994 set_task_cpu(p, new_cpu);
995 raw_spin_unlock(&rq->lock);
997 rq = cpu_rq(new_cpu);
999 raw_spin_lock(&rq->lock);
1000 BUG_ON(task_cpu(p) != new_cpu);
1001 enqueue_task(rq, p, 0);
1002 p->on_rq = TASK_ON_RQ_QUEUED;
1003 check_preempt_curr(rq, p, 0);
1008 struct migration_arg {
1009 struct task_struct *task;
1014 * Move (not current) task off this cpu, onto dest cpu. We're doing
1015 * this because either it can't run here any more (set_cpus_allowed()
1016 * away from this CPU, or CPU going down), or because we're
1017 * attempting to rebalance this task on exec (sched_exec).
1019 * So we race with normal scheduler movements, but that's OK, as long
1020 * as the task is no longer on this CPU.
1022 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1024 if (unlikely(!cpu_active(dest_cpu)))
1027 /* Affinity changed (again). */
1028 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1031 rq = move_queued_task(rq, p, dest_cpu);
1037 * migration_cpu_stop - this will be executed by a highprio stopper thread
1038 * and performs thread migration by bumping thread off CPU then
1039 * 'pushing' onto another runqueue.
1041 static int migration_cpu_stop(void *data)
1043 struct migration_arg *arg = data;
1044 struct task_struct *p = arg->task;
1045 struct rq *rq = this_rq();
1048 * The original target cpu might have gone down and we might
1049 * be on another cpu but it doesn't matter.
1051 local_irq_disable();
1053 * We need to explicitly wake pending tasks before running
1054 * __migrate_task() such that we will not miss enforcing cpus_allowed
1055 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1057 sched_ttwu_pending();
1059 raw_spin_lock(&p->pi_lock);
1060 raw_spin_lock(&rq->lock);
1062 * If task_rq(p) != rq, it cannot be migrated here, because we're
1063 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1064 * we're holding p->pi_lock.
1066 if (task_rq(p) == rq && task_on_rq_queued(p))
1067 rq = __migrate_task(rq, p, arg->dest_cpu);
1068 raw_spin_unlock(&rq->lock);
1069 raw_spin_unlock(&p->pi_lock);
1076 * sched_class::set_cpus_allowed must do the below, but is not required to
1077 * actually call this function.
1079 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1081 cpumask_copy(&p->cpus_allowed, new_mask);
1082 p->nr_cpus_allowed = cpumask_weight(new_mask);
1085 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1087 struct rq *rq = task_rq(p);
1088 bool queued, running;
1090 lockdep_assert_held(&p->pi_lock);
1092 queued = task_on_rq_queued(p);
1093 running = task_current(rq, p);
1097 * Because __kthread_bind() calls this on blocked tasks without
1100 lockdep_assert_held(&rq->lock);
1101 dequeue_task(rq, p, DEQUEUE_SAVE);
1104 put_prev_task(rq, p);
1106 p->sched_class->set_cpus_allowed(p, new_mask);
1109 p->sched_class->set_curr_task(rq);
1111 enqueue_task(rq, p, ENQUEUE_RESTORE);
1115 * Change a given task's CPU affinity. Migrate the thread to a
1116 * proper CPU and schedule it away if the CPU it's executing on
1117 * is removed from the allowed bitmask.
1119 * NOTE: the caller must have a valid reference to the task, the
1120 * task must not exit() & deallocate itself prematurely. The
1121 * call is not atomic; no spinlocks may be held.
1123 static int __set_cpus_allowed_ptr(struct task_struct *p,
1124 const struct cpumask *new_mask, bool check)
1126 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1127 unsigned int dest_cpu;
1132 rq = task_rq_lock(p, &rf);
1134 if (p->flags & PF_KTHREAD) {
1136 * Kernel threads are allowed on online && !active CPUs
1138 cpu_valid_mask = cpu_online_mask;
1142 * Must re-check here, to close a race against __kthread_bind(),
1143 * sched_setaffinity() is not guaranteed to observe the flag.
1145 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1150 if (cpumask_equal(&p->cpus_allowed, new_mask))
1153 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1158 do_set_cpus_allowed(p, new_mask);
1160 if (p->flags & PF_KTHREAD) {
1162 * For kernel threads that do indeed end up on online &&
1163 * !active we want to ensure they are strict per-cpu threads.
1165 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1166 !cpumask_intersects(new_mask, cpu_active_mask) &&
1167 p->nr_cpus_allowed != 1);
1170 /* Can the task run on the task's current CPU? If so, we're done */
1171 if (cpumask_test_cpu(task_cpu(p), new_mask))
1174 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1175 if (task_running(rq, p) || p->state == TASK_WAKING) {
1176 struct migration_arg arg = { p, dest_cpu };
1177 /* Need help from migration thread: drop lock and wait. */
1178 task_rq_unlock(rq, p, &rf);
1179 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1180 tlb_migrate_finish(p->mm);
1182 } else if (task_on_rq_queued(p)) {
1184 * OK, since we're going to drop the lock immediately
1185 * afterwards anyway.
1187 lockdep_unpin_lock(&rq->lock, rf.cookie);
1188 rq = move_queued_task(rq, p, dest_cpu);
1189 lockdep_repin_lock(&rq->lock, rf.cookie);
1192 task_rq_unlock(rq, p, &rf);
1197 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1199 return __set_cpus_allowed_ptr(p, new_mask, false);
1201 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1203 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1205 #ifdef CONFIG_SCHED_DEBUG
1207 * We should never call set_task_cpu() on a blocked task,
1208 * ttwu() will sort out the placement.
1210 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1214 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1215 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1216 * time relying on p->on_rq.
1218 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1219 p->sched_class == &fair_sched_class &&
1220 (p->on_rq && !task_on_rq_migrating(p)));
1222 #ifdef CONFIG_LOCKDEP
1224 * The caller should hold either p->pi_lock or rq->lock, when changing
1225 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1227 * sched_move_task() holds both and thus holding either pins the cgroup,
1230 * Furthermore, all task_rq users should acquire both locks, see
1233 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1234 lockdep_is_held(&task_rq(p)->lock)));
1238 trace_sched_migrate_task(p, new_cpu);
1240 if (task_cpu(p) != new_cpu) {
1241 if (p->sched_class->migrate_task_rq)
1242 p->sched_class->migrate_task_rq(p);
1243 p->se.nr_migrations++;
1244 perf_event_task_migrate(p);
1247 __set_task_cpu(p, new_cpu);
1250 static void __migrate_swap_task(struct task_struct *p, int cpu)
1252 if (task_on_rq_queued(p)) {
1253 struct rq *src_rq, *dst_rq;
1255 src_rq = task_rq(p);
1256 dst_rq = cpu_rq(cpu);
1258 p->on_rq = TASK_ON_RQ_MIGRATING;
1259 deactivate_task(src_rq, p, 0);
1260 set_task_cpu(p, cpu);
1261 activate_task(dst_rq, p, 0);
1262 p->on_rq = TASK_ON_RQ_QUEUED;
1263 check_preempt_curr(dst_rq, p, 0);
1266 * Task isn't running anymore; make it appear like we migrated
1267 * it before it went to sleep. This means on wakeup we make the
1268 * previous cpu our target instead of where it really is.
1274 struct migration_swap_arg {
1275 struct task_struct *src_task, *dst_task;
1276 int src_cpu, dst_cpu;
1279 static int migrate_swap_stop(void *data)
1281 struct migration_swap_arg *arg = data;
1282 struct rq *src_rq, *dst_rq;
1285 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1288 src_rq = cpu_rq(arg->src_cpu);
1289 dst_rq = cpu_rq(arg->dst_cpu);
1291 double_raw_lock(&arg->src_task->pi_lock,
1292 &arg->dst_task->pi_lock);
1293 double_rq_lock(src_rq, dst_rq);
1295 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1298 if (task_cpu(arg->src_task) != arg->src_cpu)
1301 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1304 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1307 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1308 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1313 double_rq_unlock(src_rq, dst_rq);
1314 raw_spin_unlock(&arg->dst_task->pi_lock);
1315 raw_spin_unlock(&arg->src_task->pi_lock);
1321 * Cross migrate two tasks
1323 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1325 struct migration_swap_arg arg;
1328 arg = (struct migration_swap_arg){
1330 .src_cpu = task_cpu(cur),
1332 .dst_cpu = task_cpu(p),
1335 if (arg.src_cpu == arg.dst_cpu)
1339 * These three tests are all lockless; this is OK since all of them
1340 * will be re-checked with proper locks held further down the line.
1342 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1345 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1348 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1351 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1352 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1359 * wait_task_inactive - wait for a thread to unschedule.
1361 * If @match_state is nonzero, it's the @p->state value just checked and
1362 * not expected to change. If it changes, i.e. @p might have woken up,
1363 * then return zero. When we succeed in waiting for @p to be off its CPU,
1364 * we return a positive number (its total switch count). If a second call
1365 * a short while later returns the same number, the caller can be sure that
1366 * @p has remained unscheduled the whole time.
1368 * The caller must ensure that the task *will* unschedule sometime soon,
1369 * else this function might spin for a *long* time. This function can't
1370 * be called with interrupts off, or it may introduce deadlock with
1371 * smp_call_function() if an IPI is sent by the same process we are
1372 * waiting to become inactive.
1374 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1376 int running, queued;
1383 * We do the initial early heuristics without holding
1384 * any task-queue locks at all. We'll only try to get
1385 * the runqueue lock when things look like they will
1391 * If the task is actively running on another CPU
1392 * still, just relax and busy-wait without holding
1395 * NOTE! Since we don't hold any locks, it's not
1396 * even sure that "rq" stays as the right runqueue!
1397 * But we don't care, since "task_running()" will
1398 * return false if the runqueue has changed and p
1399 * is actually now running somewhere else!
1401 while (task_running(rq, p)) {
1402 if (match_state && unlikely(p->state != match_state))
1408 * Ok, time to look more closely! We need the rq
1409 * lock now, to be *sure*. If we're wrong, we'll
1410 * just go back and repeat.
1412 rq = task_rq_lock(p, &rf);
1413 trace_sched_wait_task(p);
1414 running = task_running(rq, p);
1415 queued = task_on_rq_queued(p);
1417 if (!match_state || p->state == match_state)
1418 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1419 task_rq_unlock(rq, p, &rf);
1422 * If it changed from the expected state, bail out now.
1424 if (unlikely(!ncsw))
1428 * Was it really running after all now that we
1429 * checked with the proper locks actually held?
1431 * Oops. Go back and try again..
1433 if (unlikely(running)) {
1439 * It's not enough that it's not actively running,
1440 * it must be off the runqueue _entirely_, and not
1443 * So if it was still runnable (but just not actively
1444 * running right now), it's preempted, and we should
1445 * yield - it could be a while.
1447 if (unlikely(queued)) {
1448 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1450 set_current_state(TASK_UNINTERRUPTIBLE);
1451 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1456 * Ahh, all good. It wasn't running, and it wasn't
1457 * runnable, which means that it will never become
1458 * running in the future either. We're all done!
1467 * kick_process - kick a running thread to enter/exit the kernel
1468 * @p: the to-be-kicked thread
1470 * Cause a process which is running on another CPU to enter
1471 * kernel-mode, without any delay. (to get signals handled.)
1473 * NOTE: this function doesn't have to take the runqueue lock,
1474 * because all it wants to ensure is that the remote task enters
1475 * the kernel. If the IPI races and the task has been migrated
1476 * to another CPU then no harm is done and the purpose has been
1479 void kick_process(struct task_struct *p)
1485 if ((cpu != smp_processor_id()) && task_curr(p))
1486 smp_send_reschedule(cpu);
1489 EXPORT_SYMBOL_GPL(kick_process);
1492 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1494 * A few notes on cpu_active vs cpu_online:
1496 * - cpu_active must be a subset of cpu_online
1498 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1499 * see __set_cpus_allowed_ptr(). At this point the newly online
1500 * cpu isn't yet part of the sched domains, and balancing will not
1503 * - on cpu-down we clear cpu_active() to mask the sched domains and
1504 * avoid the load balancer to place new tasks on the to be removed
1505 * cpu. Existing tasks will remain running there and will be taken
1508 * This means that fallback selection must not select !active CPUs.
1509 * And can assume that any active CPU must be online. Conversely
1510 * select_task_rq() below may allow selection of !active CPUs in order
1511 * to satisfy the above rules.
1513 static int select_fallback_rq(int cpu, struct task_struct *p)
1515 int nid = cpu_to_node(cpu);
1516 const struct cpumask *nodemask = NULL;
1517 enum { cpuset, possible, fail } state = cpuset;
1521 * If the node that the cpu is on has been offlined, cpu_to_node()
1522 * will return -1. There is no cpu on the node, and we should
1523 * select the cpu on the other node.
1526 nodemask = cpumask_of_node(nid);
1528 /* Look for allowed, online CPU in same node. */
1529 for_each_cpu(dest_cpu, nodemask) {
1530 if (!cpu_active(dest_cpu))
1532 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1538 /* Any allowed, online CPU? */
1539 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1540 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1542 if (!cpu_online(dest_cpu))
1547 /* No more Mr. Nice Guy. */
1550 if (IS_ENABLED(CONFIG_CPUSETS)) {
1551 cpuset_cpus_allowed_fallback(p);
1557 do_set_cpus_allowed(p, cpu_possible_mask);
1568 if (state != cpuset) {
1570 * Don't tell them about moving exiting tasks or
1571 * kernel threads (both mm NULL), since they never
1574 if (p->mm && printk_ratelimit()) {
1575 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1576 task_pid_nr(p), p->comm, cpu);
1584 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1587 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1589 lockdep_assert_held(&p->pi_lock);
1591 if (tsk_nr_cpus_allowed(p) > 1)
1592 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1594 cpu = cpumask_any(tsk_cpus_allowed(p));
1597 * In order not to call set_task_cpu() on a blocking task we need
1598 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1601 * Since this is common to all placement strategies, this lives here.
1603 * [ this allows ->select_task() to simply return task_cpu(p) and
1604 * not worry about this generic constraint ]
1606 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1608 cpu = select_fallback_rq(task_cpu(p), p);
1613 static void update_avg(u64 *avg, u64 sample)
1615 s64 diff = sample - *avg;
1621 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1622 const struct cpumask *new_mask, bool check)
1624 return set_cpus_allowed_ptr(p, new_mask);
1627 #endif /* CONFIG_SMP */
1630 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1634 if (!schedstat_enabled())
1640 if (cpu == rq->cpu) {
1641 schedstat_inc(rq->ttwu_local);
1642 schedstat_inc(p->se.statistics.nr_wakeups_local);
1644 struct sched_domain *sd;
1646 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1648 for_each_domain(rq->cpu, sd) {
1649 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1650 schedstat_inc(sd->ttwu_wake_remote);
1657 if (wake_flags & WF_MIGRATED)
1658 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1659 #endif /* CONFIG_SMP */
1661 schedstat_inc(rq->ttwu_count);
1662 schedstat_inc(p->se.statistics.nr_wakeups);
1664 if (wake_flags & WF_SYNC)
1665 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1668 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1670 activate_task(rq, p, en_flags);
1671 p->on_rq = TASK_ON_RQ_QUEUED;
1673 /* if a worker is waking up, notify workqueue */
1674 if (p->flags & PF_WQ_WORKER)
1675 wq_worker_waking_up(p, cpu_of(rq));
1679 * Mark the task runnable and perform wakeup-preemption.
1681 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1682 struct pin_cookie cookie)
1684 check_preempt_curr(rq, p, wake_flags);
1685 p->state = TASK_RUNNING;
1686 trace_sched_wakeup(p);
1689 if (p->sched_class->task_woken) {
1691 * Our task @p is fully woken up and running; so its safe to
1692 * drop the rq->lock, hereafter rq is only used for statistics.
1694 lockdep_unpin_lock(&rq->lock, cookie);
1695 p->sched_class->task_woken(rq, p);
1696 lockdep_repin_lock(&rq->lock, cookie);
1699 if (rq->idle_stamp) {
1700 u64 delta = rq_clock(rq) - rq->idle_stamp;
1701 u64 max = 2*rq->max_idle_balance_cost;
1703 update_avg(&rq->avg_idle, delta);
1705 if (rq->avg_idle > max)
1714 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1715 struct pin_cookie cookie)
1717 int en_flags = ENQUEUE_WAKEUP;
1719 lockdep_assert_held(&rq->lock);
1722 if (p->sched_contributes_to_load)
1723 rq->nr_uninterruptible--;
1725 if (wake_flags & WF_MIGRATED)
1726 en_flags |= ENQUEUE_MIGRATED;
1729 ttwu_activate(rq, p, en_flags);
1730 ttwu_do_wakeup(rq, p, wake_flags, cookie);
1734 * Called in case the task @p isn't fully descheduled from its runqueue,
1735 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1736 * since all we need to do is flip p->state to TASK_RUNNING, since
1737 * the task is still ->on_rq.
1739 static int ttwu_remote(struct task_struct *p, int wake_flags)
1745 rq = __task_rq_lock(p, &rf);
1746 if (task_on_rq_queued(p)) {
1747 /* check_preempt_curr() may use rq clock */
1748 update_rq_clock(rq);
1749 ttwu_do_wakeup(rq, p, wake_flags, rf.cookie);
1752 __task_rq_unlock(rq, &rf);
1758 void sched_ttwu_pending(void)
1760 struct rq *rq = this_rq();
1761 struct llist_node *llist = llist_del_all(&rq->wake_list);
1762 struct pin_cookie cookie;
1763 struct task_struct *p;
1764 unsigned long flags;
1769 raw_spin_lock_irqsave(&rq->lock, flags);
1770 cookie = lockdep_pin_lock(&rq->lock);
1775 p = llist_entry(llist, struct task_struct, wake_entry);
1776 llist = llist_next(llist);
1778 if (p->sched_remote_wakeup)
1779 wake_flags = WF_MIGRATED;
1781 ttwu_do_activate(rq, p, wake_flags, cookie);
1784 lockdep_unpin_lock(&rq->lock, cookie);
1785 raw_spin_unlock_irqrestore(&rq->lock, flags);
1788 void scheduler_ipi(void)
1791 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1792 * TIF_NEED_RESCHED remotely (for the first time) will also send
1795 preempt_fold_need_resched();
1797 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1801 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1802 * traditionally all their work was done from the interrupt return
1803 * path. Now that we actually do some work, we need to make sure
1806 * Some archs already do call them, luckily irq_enter/exit nest
1809 * Arguably we should visit all archs and update all handlers,
1810 * however a fair share of IPIs are still resched only so this would
1811 * somewhat pessimize the simple resched case.
1814 sched_ttwu_pending();
1817 * Check if someone kicked us for doing the nohz idle load balance.
1819 if (unlikely(got_nohz_idle_kick())) {
1820 this_rq()->idle_balance = 1;
1821 raise_softirq_irqoff(SCHED_SOFTIRQ);
1826 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1828 struct rq *rq = cpu_rq(cpu);
1830 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1832 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1833 if (!set_nr_if_polling(rq->idle))
1834 smp_send_reschedule(cpu);
1836 trace_sched_wake_idle_without_ipi(cpu);
1840 void wake_up_if_idle(int cpu)
1842 struct rq *rq = cpu_rq(cpu);
1843 unsigned long flags;
1847 if (!is_idle_task(rcu_dereference(rq->curr)))
1850 if (set_nr_if_polling(rq->idle)) {
1851 trace_sched_wake_idle_without_ipi(cpu);
1853 raw_spin_lock_irqsave(&rq->lock, flags);
1854 if (is_idle_task(rq->curr))
1855 smp_send_reschedule(cpu);
1856 /* Else cpu is not in idle, do nothing here */
1857 raw_spin_unlock_irqrestore(&rq->lock, flags);
1864 bool cpus_share_cache(int this_cpu, int that_cpu)
1866 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1868 #endif /* CONFIG_SMP */
1870 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1872 struct rq *rq = cpu_rq(cpu);
1873 struct pin_cookie cookie;
1875 #if defined(CONFIG_SMP)
1876 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1877 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1878 ttwu_queue_remote(p, cpu, wake_flags);
1883 raw_spin_lock(&rq->lock);
1884 cookie = lockdep_pin_lock(&rq->lock);
1885 ttwu_do_activate(rq, p, wake_flags, cookie);
1886 lockdep_unpin_lock(&rq->lock, cookie);
1887 raw_spin_unlock(&rq->lock);
1891 * Notes on Program-Order guarantees on SMP systems.
1895 * The basic program-order guarantee on SMP systems is that when a task [t]
1896 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1897 * execution on its new cpu [c1].
1899 * For migration (of runnable tasks) this is provided by the following means:
1901 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1902 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1903 * rq(c1)->lock (if not at the same time, then in that order).
1904 * C) LOCK of the rq(c1)->lock scheduling in task
1906 * Transitivity guarantees that B happens after A and C after B.
1907 * Note: we only require RCpc transitivity.
1908 * Note: the cpu doing B need not be c0 or c1
1917 * UNLOCK rq(0)->lock
1919 * LOCK rq(0)->lock // orders against CPU0
1921 * UNLOCK rq(0)->lock
1925 * UNLOCK rq(1)->lock
1927 * LOCK rq(1)->lock // orders against CPU2
1930 * UNLOCK rq(1)->lock
1933 * BLOCKING -- aka. SLEEP + WAKEUP
1935 * For blocking we (obviously) need to provide the same guarantee as for
1936 * migration. However the means are completely different as there is no lock
1937 * chain to provide order. Instead we do:
1939 * 1) smp_store_release(X->on_cpu, 0)
1940 * 2) smp_cond_load_acquire(!X->on_cpu)
1944 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1946 * LOCK rq(0)->lock LOCK X->pi_lock
1949 * smp_store_release(X->on_cpu, 0);
1951 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1957 * X->state = RUNNING
1958 * UNLOCK rq(2)->lock
1960 * LOCK rq(2)->lock // orders against CPU1
1963 * UNLOCK rq(2)->lock
1966 * UNLOCK rq(0)->lock
1969 * However; for wakeups there is a second guarantee we must provide, namely we
1970 * must observe the state that lead to our wakeup. That is, not only must our
1971 * task observe its own prior state, it must also observe the stores prior to
1974 * This means that any means of doing remote wakeups must order the CPU doing
1975 * the wakeup against the CPU the task is going to end up running on. This,
1976 * however, is already required for the regular Program-Order guarantee above,
1977 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1982 * try_to_wake_up - wake up a thread
1983 * @p: the thread to be awakened
1984 * @state: the mask of task states that can be woken
1985 * @wake_flags: wake modifier flags (WF_*)
1987 * Put it on the run-queue if it's not already there. The "current"
1988 * thread is always on the run-queue (except when the actual
1989 * re-schedule is in progress), and as such you're allowed to do
1990 * the simpler "current->state = TASK_RUNNING" to mark yourself
1991 * runnable without the overhead of this.
1993 * Return: %true if @p was woken up, %false if it was already running.
1994 * or @state didn't match @p's state.
1997 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1999 unsigned long flags;
2000 int cpu, success = 0;
2003 * If we are going to wake up a thread waiting for CONDITION we
2004 * need to ensure that CONDITION=1 done by the caller can not be
2005 * reordered with p->state check below. This pairs with mb() in
2006 * set_current_state() the waiting thread does.
2008 smp_mb__before_spinlock();
2009 raw_spin_lock_irqsave(&p->pi_lock, flags);
2010 if (!(p->state & state))
2013 trace_sched_waking(p);
2015 success = 1; /* we're going to change ->state */
2019 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2020 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2021 * in smp_cond_load_acquire() below.
2023 * sched_ttwu_pending() try_to_wake_up()
2024 * [S] p->on_rq = 1; [L] P->state
2025 * UNLOCK rq->lock -----.
2029 * LOCK rq->lock -----'
2033 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2035 * Pairs with the UNLOCK+LOCK on rq->lock from the
2036 * last wakeup of our task and the schedule that got our task
2040 if (p->on_rq && ttwu_remote(p, wake_flags))
2045 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2046 * possible to, falsely, observe p->on_cpu == 0.
2048 * One must be running (->on_cpu == 1) in order to remove oneself
2049 * from the runqueue.
2051 * [S] ->on_cpu = 1; [L] ->on_rq
2055 * [S] ->on_rq = 0; [L] ->on_cpu
2057 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2058 * from the consecutive calls to schedule(); the first switching to our
2059 * task, the second putting it to sleep.
2064 * If the owning (remote) cpu is still in the middle of schedule() with
2065 * this task as prev, wait until its done referencing the task.
2067 * Pairs with the smp_store_release() in finish_lock_switch().
2069 * This ensures that tasks getting woken will be fully ordered against
2070 * their previous state and preserve Program Order.
2072 smp_cond_load_acquire(&p->on_cpu, !VAL);
2074 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2075 p->state = TASK_WAKING;
2077 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2078 if (task_cpu(p) != cpu) {
2079 wake_flags |= WF_MIGRATED;
2080 set_task_cpu(p, cpu);
2082 #endif /* CONFIG_SMP */
2084 ttwu_queue(p, cpu, wake_flags);
2086 ttwu_stat(p, cpu, wake_flags);
2088 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2094 * try_to_wake_up_local - try to wake up a local task with rq lock held
2095 * @p: the thread to be awakened
2096 * @cookie: context's cookie for pinning
2098 * Put @p on the run-queue if it's not already there. The caller must
2099 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2102 static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie)
2104 struct rq *rq = task_rq(p);
2106 if (WARN_ON_ONCE(rq != this_rq()) ||
2107 WARN_ON_ONCE(p == current))
2110 lockdep_assert_held(&rq->lock);
2112 if (!raw_spin_trylock(&p->pi_lock)) {
2114 * This is OK, because current is on_cpu, which avoids it being
2115 * picked for load-balance and preemption/IRQs are still
2116 * disabled avoiding further scheduler activity on it and we've
2117 * not yet picked a replacement task.
2119 lockdep_unpin_lock(&rq->lock, cookie);
2120 raw_spin_unlock(&rq->lock);
2121 raw_spin_lock(&p->pi_lock);
2122 raw_spin_lock(&rq->lock);
2123 lockdep_repin_lock(&rq->lock, cookie);
2126 if (!(p->state & TASK_NORMAL))
2129 trace_sched_waking(p);
2131 if (!task_on_rq_queued(p))
2132 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2134 ttwu_do_wakeup(rq, p, 0, cookie);
2135 ttwu_stat(p, smp_processor_id(), 0);
2137 raw_spin_unlock(&p->pi_lock);
2141 * wake_up_process - Wake up a specific process
2142 * @p: The process to be woken up.
2144 * Attempt to wake up the nominated process and move it to the set of runnable
2147 * Return: 1 if the process was woken up, 0 if it was already running.
2149 * It may be assumed that this function implies a write memory barrier before
2150 * changing the task state if and only if any tasks are woken up.
2152 int wake_up_process(struct task_struct *p)
2154 return try_to_wake_up(p, TASK_NORMAL, 0);
2156 EXPORT_SYMBOL(wake_up_process);
2158 int wake_up_state(struct task_struct *p, unsigned int state)
2160 return try_to_wake_up(p, state, 0);
2164 * This function clears the sched_dl_entity static params.
2166 void __dl_clear_params(struct task_struct *p)
2168 struct sched_dl_entity *dl_se = &p->dl;
2170 dl_se->dl_runtime = 0;
2171 dl_se->dl_deadline = 0;
2172 dl_se->dl_period = 0;
2176 dl_se->dl_throttled = 0;
2177 dl_se->dl_yielded = 0;
2181 * Perform scheduler related setup for a newly forked process p.
2182 * p is forked by current.
2184 * __sched_fork() is basic setup used by init_idle() too:
2186 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2191 p->se.exec_start = 0;
2192 p->se.sum_exec_runtime = 0;
2193 p->se.prev_sum_exec_runtime = 0;
2194 p->se.nr_migrations = 0;
2196 INIT_LIST_HEAD(&p->se.group_node);
2198 #ifdef CONFIG_FAIR_GROUP_SCHED
2199 p->se.cfs_rq = NULL;
2202 #ifdef CONFIG_SCHEDSTATS
2203 /* Even if schedstat is disabled, there should not be garbage */
2204 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2207 RB_CLEAR_NODE(&p->dl.rb_node);
2208 init_dl_task_timer(&p->dl);
2209 __dl_clear_params(p);
2211 INIT_LIST_HEAD(&p->rt.run_list);
2213 p->rt.time_slice = sched_rr_timeslice;
2217 #ifdef CONFIG_PREEMPT_NOTIFIERS
2218 INIT_HLIST_HEAD(&p->preempt_notifiers);
2221 #ifdef CONFIG_NUMA_BALANCING
2222 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2223 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2224 p->mm->numa_scan_seq = 0;
2227 if (clone_flags & CLONE_VM)
2228 p->numa_preferred_nid = current->numa_preferred_nid;
2230 p->numa_preferred_nid = -1;
2232 p->node_stamp = 0ULL;
2233 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2234 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2235 p->numa_work.next = &p->numa_work;
2236 p->numa_faults = NULL;
2237 p->last_task_numa_placement = 0;
2238 p->last_sum_exec_runtime = 0;
2240 p->numa_group = NULL;
2241 #endif /* CONFIG_NUMA_BALANCING */
2244 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2246 #ifdef CONFIG_NUMA_BALANCING
2248 void set_numabalancing_state(bool enabled)
2251 static_branch_enable(&sched_numa_balancing);
2253 static_branch_disable(&sched_numa_balancing);
2256 #ifdef CONFIG_PROC_SYSCTL
2257 int sysctl_numa_balancing(struct ctl_table *table, int write,
2258 void __user *buffer, size_t *lenp, loff_t *ppos)
2262 int state = static_branch_likely(&sched_numa_balancing);
2264 if (write && !capable(CAP_SYS_ADMIN))
2269 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2273 set_numabalancing_state(state);
2279 #ifdef CONFIG_SCHEDSTATS
2281 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2282 static bool __initdata __sched_schedstats = false;
2284 static void set_schedstats(bool enabled)
2287 static_branch_enable(&sched_schedstats);
2289 static_branch_disable(&sched_schedstats);
2292 void force_schedstat_enabled(void)
2294 if (!schedstat_enabled()) {
2295 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2296 static_branch_enable(&sched_schedstats);
2300 static int __init setup_schedstats(char *str)
2307 * This code is called before jump labels have been set up, so we can't
2308 * change the static branch directly just yet. Instead set a temporary
2309 * variable so init_schedstats() can do it later.
2311 if (!strcmp(str, "enable")) {
2312 __sched_schedstats = true;
2314 } else if (!strcmp(str, "disable")) {
2315 __sched_schedstats = false;
2320 pr_warn("Unable to parse schedstats=\n");
2324 __setup("schedstats=", setup_schedstats);
2326 static void __init init_schedstats(void)
2328 set_schedstats(__sched_schedstats);
2331 #ifdef CONFIG_PROC_SYSCTL
2332 int sysctl_schedstats(struct ctl_table *table, int write,
2333 void __user *buffer, size_t *lenp, loff_t *ppos)
2337 int state = static_branch_likely(&sched_schedstats);
2339 if (write && !capable(CAP_SYS_ADMIN))
2344 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2348 set_schedstats(state);
2351 #endif /* CONFIG_PROC_SYSCTL */
2352 #else /* !CONFIG_SCHEDSTATS */
2353 static inline void init_schedstats(void) {}
2354 #endif /* CONFIG_SCHEDSTATS */
2357 * fork()/clone()-time setup:
2359 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2361 unsigned long flags;
2362 int cpu = get_cpu();
2364 __sched_fork(clone_flags, p);
2366 * We mark the process as NEW here. This guarantees that
2367 * nobody will actually run it, and a signal or other external
2368 * event cannot wake it up and insert it on the runqueue either.
2370 p->state = TASK_NEW;
2373 * Make sure we do not leak PI boosting priority to the child.
2375 p->prio = current->normal_prio;
2378 * Revert to default priority/policy on fork if requested.
2380 if (unlikely(p->sched_reset_on_fork)) {
2381 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2382 p->policy = SCHED_NORMAL;
2383 p->static_prio = NICE_TO_PRIO(0);
2385 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2386 p->static_prio = NICE_TO_PRIO(0);
2388 p->prio = p->normal_prio = __normal_prio(p);
2392 * We don't need the reset flag anymore after the fork. It has
2393 * fulfilled its duty:
2395 p->sched_reset_on_fork = 0;
2398 if (dl_prio(p->prio)) {
2401 } else if (rt_prio(p->prio)) {
2402 p->sched_class = &rt_sched_class;
2404 p->sched_class = &fair_sched_class;
2407 init_entity_runnable_average(&p->se);
2410 * The child is not yet in the pid-hash so no cgroup attach races,
2411 * and the cgroup is pinned to this child due to cgroup_fork()
2412 * is ran before sched_fork().
2414 * Silence PROVE_RCU.
2416 raw_spin_lock_irqsave(&p->pi_lock, flags);
2418 * We're setting the cpu for the first time, we don't migrate,
2419 * so use __set_task_cpu().
2421 __set_task_cpu(p, cpu);
2422 if (p->sched_class->task_fork)
2423 p->sched_class->task_fork(p);
2424 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2426 #ifdef CONFIG_SCHED_INFO
2427 if (likely(sched_info_on()))
2428 memset(&p->sched_info, 0, sizeof(p->sched_info));
2430 #if defined(CONFIG_SMP)
2433 init_task_preempt_count(p);
2435 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2436 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2443 unsigned long to_ratio(u64 period, u64 runtime)
2445 if (runtime == RUNTIME_INF)
2449 * Doing this here saves a lot of checks in all
2450 * the calling paths, and returning zero seems
2451 * safe for them anyway.
2456 return div64_u64(runtime << 20, period);
2460 inline struct dl_bw *dl_bw_of(int i)
2462 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2463 "sched RCU must be held");
2464 return &cpu_rq(i)->rd->dl_bw;
2467 static inline int dl_bw_cpus(int i)
2469 struct root_domain *rd = cpu_rq(i)->rd;
2472 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2473 "sched RCU must be held");
2474 for_each_cpu_and(i, rd->span, cpu_active_mask)
2480 inline struct dl_bw *dl_bw_of(int i)
2482 return &cpu_rq(i)->dl.dl_bw;
2485 static inline int dl_bw_cpus(int i)
2492 * We must be sure that accepting a new task (or allowing changing the
2493 * parameters of an existing one) is consistent with the bandwidth
2494 * constraints. If yes, this function also accordingly updates the currently
2495 * allocated bandwidth to reflect the new situation.
2497 * This function is called while holding p's rq->lock.
2499 * XXX we should delay bw change until the task's 0-lag point, see
2502 static int dl_overflow(struct task_struct *p, int policy,
2503 const struct sched_attr *attr)
2506 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2507 u64 period = attr->sched_period ?: attr->sched_deadline;
2508 u64 runtime = attr->sched_runtime;
2509 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2512 /* !deadline task may carry old deadline bandwidth */
2513 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2517 * Either if a task, enters, leave, or stays -deadline but changes
2518 * its parameters, we may need to update accordingly the total
2519 * allocated bandwidth of the container.
2521 raw_spin_lock(&dl_b->lock);
2522 cpus = dl_bw_cpus(task_cpu(p));
2523 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2524 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2525 __dl_add(dl_b, new_bw);
2527 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2528 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2529 __dl_clear(dl_b, p->dl.dl_bw);
2530 __dl_add(dl_b, new_bw);
2532 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2533 __dl_clear(dl_b, p->dl.dl_bw);
2536 raw_spin_unlock(&dl_b->lock);
2541 extern void init_dl_bw(struct dl_bw *dl_b);
2544 * wake_up_new_task - wake up a newly created task for the first time.
2546 * This function will do some initial scheduler statistics housekeeping
2547 * that must be done for every newly created context, then puts the task
2548 * on the runqueue and wakes it.
2550 void wake_up_new_task(struct task_struct *p)
2555 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2556 p->state = TASK_RUNNING;
2559 * Fork balancing, do it here and not earlier because:
2560 * - cpus_allowed can change in the fork path
2561 * - any previously selected cpu might disappear through hotplug
2563 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2564 * as we're not fully set-up yet.
2566 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2568 rq = __task_rq_lock(p, &rf);
2569 post_init_entity_util_avg(&p->se);
2571 activate_task(rq, p, 0);
2572 p->on_rq = TASK_ON_RQ_QUEUED;
2573 trace_sched_wakeup_new(p);
2574 check_preempt_curr(rq, p, WF_FORK);
2576 if (p->sched_class->task_woken) {
2578 * Nothing relies on rq->lock after this, so its fine to
2581 lockdep_unpin_lock(&rq->lock, rf.cookie);
2582 p->sched_class->task_woken(rq, p);
2583 lockdep_repin_lock(&rq->lock, rf.cookie);
2586 task_rq_unlock(rq, p, &rf);
2589 #ifdef CONFIG_PREEMPT_NOTIFIERS
2591 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2593 void preempt_notifier_inc(void)
2595 static_key_slow_inc(&preempt_notifier_key);
2597 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2599 void preempt_notifier_dec(void)
2601 static_key_slow_dec(&preempt_notifier_key);
2603 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2606 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2607 * @notifier: notifier struct to register
2609 void preempt_notifier_register(struct preempt_notifier *notifier)
2611 if (!static_key_false(&preempt_notifier_key))
2612 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2614 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2616 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2619 * preempt_notifier_unregister - no longer interested in preemption notifications
2620 * @notifier: notifier struct to unregister
2622 * This is *not* safe to call from within a preemption notifier.
2624 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2626 hlist_del(¬ifier->link);
2628 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2630 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2632 struct preempt_notifier *notifier;
2634 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2635 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2638 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2640 if (static_key_false(&preempt_notifier_key))
2641 __fire_sched_in_preempt_notifiers(curr);
2645 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2646 struct task_struct *next)
2648 struct preempt_notifier *notifier;
2650 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2651 notifier->ops->sched_out(notifier, next);
2654 static __always_inline void
2655 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2656 struct task_struct *next)
2658 if (static_key_false(&preempt_notifier_key))
2659 __fire_sched_out_preempt_notifiers(curr, next);
2662 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2664 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2669 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2670 struct task_struct *next)
2674 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2677 * prepare_task_switch - prepare to switch tasks
2678 * @rq: the runqueue preparing to switch
2679 * @prev: the current task that is being switched out
2680 * @next: the task we are going to switch to.
2682 * This is called with the rq lock held and interrupts off. It must
2683 * be paired with a subsequent finish_task_switch after the context
2686 * prepare_task_switch sets up locking and calls architecture specific
2690 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2691 struct task_struct *next)
2693 sched_info_switch(rq, prev, next);
2694 perf_event_task_sched_out(prev, next);
2695 fire_sched_out_preempt_notifiers(prev, next);
2696 prepare_lock_switch(rq, next);
2697 prepare_arch_switch(next);
2701 * finish_task_switch - clean up after a task-switch
2702 * @prev: the thread we just switched away from.
2704 * finish_task_switch must be called after the context switch, paired
2705 * with a prepare_task_switch call before the context switch.
2706 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2707 * and do any other architecture-specific cleanup actions.
2709 * Note that we may have delayed dropping an mm in context_switch(). If
2710 * so, we finish that here outside of the runqueue lock. (Doing it
2711 * with the lock held can cause deadlocks; see schedule() for
2714 * The context switch have flipped the stack from under us and restored the
2715 * local variables which were saved when this task called schedule() in the
2716 * past. prev == current is still correct but we need to recalculate this_rq
2717 * because prev may have moved to another CPU.
2719 static struct rq *finish_task_switch(struct task_struct *prev)
2720 __releases(rq->lock)
2722 struct rq *rq = this_rq();
2723 struct mm_struct *mm = rq->prev_mm;
2727 * The previous task will have left us with a preempt_count of 2
2728 * because it left us after:
2731 * preempt_disable(); // 1
2733 * raw_spin_lock_irq(&rq->lock) // 2
2735 * Also, see FORK_PREEMPT_COUNT.
2737 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2738 "corrupted preempt_count: %s/%d/0x%x\n",
2739 current->comm, current->pid, preempt_count()))
2740 preempt_count_set(FORK_PREEMPT_COUNT);
2745 * A task struct has one reference for the use as "current".
2746 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2747 * schedule one last time. The schedule call will never return, and
2748 * the scheduled task must drop that reference.
2750 * We must observe prev->state before clearing prev->on_cpu (in
2751 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2752 * running on another CPU and we could rave with its RUNNING -> DEAD
2753 * transition, resulting in a double drop.
2755 prev_state = prev->state;
2756 vtime_task_switch(prev);
2757 perf_event_task_sched_in(prev, current);
2758 finish_lock_switch(rq, prev);
2759 finish_arch_post_lock_switch();
2761 fire_sched_in_preempt_notifiers(current);
2764 if (unlikely(prev_state == TASK_DEAD)) {
2765 if (prev->sched_class->task_dead)
2766 prev->sched_class->task_dead(prev);
2769 * Remove function-return probe instances associated with this
2770 * task and put them back on the free list.
2772 kprobe_flush_task(prev);
2773 put_task_struct(prev);
2776 tick_nohz_task_switch();
2782 /* rq->lock is NOT held, but preemption is disabled */
2783 static void __balance_callback(struct rq *rq)
2785 struct callback_head *head, *next;
2786 void (*func)(struct rq *rq);
2787 unsigned long flags;
2789 raw_spin_lock_irqsave(&rq->lock, flags);
2790 head = rq->balance_callback;
2791 rq->balance_callback = NULL;
2793 func = (void (*)(struct rq *))head->func;
2800 raw_spin_unlock_irqrestore(&rq->lock, flags);
2803 static inline void balance_callback(struct rq *rq)
2805 if (unlikely(rq->balance_callback))
2806 __balance_callback(rq);
2811 static inline void balance_callback(struct rq *rq)
2818 * schedule_tail - first thing a freshly forked thread must call.
2819 * @prev: the thread we just switched away from.
2821 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2822 __releases(rq->lock)
2827 * New tasks start with FORK_PREEMPT_COUNT, see there and
2828 * finish_task_switch() for details.
2830 * finish_task_switch() will drop rq->lock() and lower preempt_count
2831 * and the preempt_enable() will end up enabling preemption (on
2832 * PREEMPT_COUNT kernels).
2835 rq = finish_task_switch(prev);
2836 balance_callback(rq);
2839 if (current->set_child_tid)
2840 put_user(task_pid_vnr(current), current->set_child_tid);
2844 * context_switch - switch to the new MM and the new thread's register state.
2846 static __always_inline struct rq *
2847 context_switch(struct rq *rq, struct task_struct *prev,
2848 struct task_struct *next, struct pin_cookie cookie)
2850 struct mm_struct *mm, *oldmm;
2852 prepare_task_switch(rq, prev, next);
2855 oldmm = prev->active_mm;
2857 * For paravirt, this is coupled with an exit in switch_to to
2858 * combine the page table reload and the switch backend into
2861 arch_start_context_switch(prev);
2864 next->active_mm = oldmm;
2865 atomic_inc(&oldmm->mm_count);
2866 enter_lazy_tlb(oldmm, next);
2868 switch_mm_irqs_off(oldmm, mm, next);
2871 prev->active_mm = NULL;
2872 rq->prev_mm = oldmm;
2875 * Since the runqueue lock will be released by the next
2876 * task (which is an invalid locking op but in the case
2877 * of the scheduler it's an obvious special-case), so we
2878 * do an early lockdep release here:
2880 lockdep_unpin_lock(&rq->lock, cookie);
2881 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2883 /* Here we just switch the register state and the stack. */
2884 switch_to(prev, next, prev);
2887 return finish_task_switch(prev);
2891 * nr_running and nr_context_switches:
2893 * externally visible scheduler statistics: current number of runnable
2894 * threads, total number of context switches performed since bootup.
2896 unsigned long nr_running(void)
2898 unsigned long i, sum = 0;
2900 for_each_online_cpu(i)
2901 sum += cpu_rq(i)->nr_running;
2907 * Check if only the current task is running on the cpu.
2909 * Caution: this function does not check that the caller has disabled
2910 * preemption, thus the result might have a time-of-check-to-time-of-use
2911 * race. The caller is responsible to use it correctly, for example:
2913 * - from a non-preemptable section (of course)
2915 * - from a thread that is bound to a single CPU
2917 * - in a loop with very short iterations (e.g. a polling loop)
2919 bool single_task_running(void)
2921 return raw_rq()->nr_running == 1;
2923 EXPORT_SYMBOL(single_task_running);
2925 unsigned long long nr_context_switches(void)
2928 unsigned long long sum = 0;
2930 for_each_possible_cpu(i)
2931 sum += cpu_rq(i)->nr_switches;
2936 unsigned long nr_iowait(void)
2938 unsigned long i, sum = 0;
2940 for_each_possible_cpu(i)
2941 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2946 unsigned long nr_iowait_cpu(int cpu)
2948 struct rq *this = cpu_rq(cpu);
2949 return atomic_read(&this->nr_iowait);
2952 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2954 struct rq *rq = this_rq();
2955 *nr_waiters = atomic_read(&rq->nr_iowait);
2956 *load = rq->load.weight;
2962 * sched_exec - execve() is a valuable balancing opportunity, because at
2963 * this point the task has the smallest effective memory and cache footprint.
2965 void sched_exec(void)
2967 struct task_struct *p = current;
2968 unsigned long flags;
2971 raw_spin_lock_irqsave(&p->pi_lock, flags);
2972 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2973 if (dest_cpu == smp_processor_id())
2976 if (likely(cpu_active(dest_cpu))) {
2977 struct migration_arg arg = { p, dest_cpu };
2979 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2980 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2984 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2989 DEFINE_PER_CPU(struct kernel_stat, kstat);
2990 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2992 EXPORT_PER_CPU_SYMBOL(kstat);
2993 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2996 * The function fair_sched_class.update_curr accesses the struct curr
2997 * and its field curr->exec_start; when called from task_sched_runtime(),
2998 * we observe a high rate of cache misses in practice.
2999 * Prefetching this data results in improved performance.
3001 static inline void prefetch_curr_exec_start(struct task_struct *p)
3003 #ifdef CONFIG_FAIR_GROUP_SCHED
3004 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3006 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3009 prefetch(&curr->exec_start);
3013 * Return accounted runtime for the task.
3014 * In case the task is currently running, return the runtime plus current's
3015 * pending runtime that have not been accounted yet.
3017 unsigned long long task_sched_runtime(struct task_struct *p)
3023 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3025 * 64-bit doesn't need locks to atomically read a 64bit value.
3026 * So we have a optimization chance when the task's delta_exec is 0.
3027 * Reading ->on_cpu is racy, but this is ok.
3029 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3030 * If we race with it entering cpu, unaccounted time is 0. This is
3031 * indistinguishable from the read occurring a few cycles earlier.
3032 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3033 * been accounted, so we're correct here as well.
3035 if (!p->on_cpu || !task_on_rq_queued(p))
3036 return p->se.sum_exec_runtime;
3039 rq = task_rq_lock(p, &rf);
3041 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3042 * project cycles that may never be accounted to this
3043 * thread, breaking clock_gettime().
3045 if (task_current(rq, p) && task_on_rq_queued(p)) {
3046 prefetch_curr_exec_start(p);
3047 update_rq_clock(rq);
3048 p->sched_class->update_curr(rq);
3050 ns = p->se.sum_exec_runtime;
3051 task_rq_unlock(rq, p, &rf);
3057 * This function gets called by the timer code, with HZ frequency.
3058 * We call it with interrupts disabled.
3060 void scheduler_tick(void)
3062 int cpu = smp_processor_id();
3063 struct rq *rq = cpu_rq(cpu);
3064 struct task_struct *curr = rq->curr;
3068 raw_spin_lock(&rq->lock);
3069 update_rq_clock(rq);
3070 curr->sched_class->task_tick(rq, curr, 0);
3071 cpu_load_update_active(rq);
3072 calc_global_load_tick(rq);
3073 raw_spin_unlock(&rq->lock);
3075 perf_event_task_tick();
3078 rq->idle_balance = idle_cpu(cpu);
3079 trigger_load_balance(rq);
3081 rq_last_tick_reset(rq);
3084 #ifdef CONFIG_NO_HZ_FULL
3086 * scheduler_tick_max_deferment
3088 * Keep at least one tick per second when a single
3089 * active task is running because the scheduler doesn't
3090 * yet completely support full dynticks environment.
3092 * This makes sure that uptime, CFS vruntime, load
3093 * balancing, etc... continue to move forward, even
3094 * with a very low granularity.
3096 * Return: Maximum deferment in nanoseconds.
3098 u64 scheduler_tick_max_deferment(void)
3100 struct rq *rq = this_rq();
3101 unsigned long next, now = READ_ONCE(jiffies);
3103 next = rq->last_sched_tick + HZ;
3105 if (time_before_eq(next, now))
3108 return jiffies_to_nsecs(next - now);
3112 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3113 defined(CONFIG_PREEMPT_TRACER))
3115 * If the value passed in is equal to the current preempt count
3116 * then we just disabled preemption. Start timing the latency.
3118 static inline void preempt_latency_start(int val)
3120 if (preempt_count() == val) {
3121 unsigned long ip = get_lock_parent_ip();
3122 #ifdef CONFIG_DEBUG_PREEMPT
3123 current->preempt_disable_ip = ip;
3125 trace_preempt_off(CALLER_ADDR0, ip);
3129 void preempt_count_add(int val)
3131 #ifdef CONFIG_DEBUG_PREEMPT
3135 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3138 __preempt_count_add(val);
3139 #ifdef CONFIG_DEBUG_PREEMPT
3141 * Spinlock count overflowing soon?
3143 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3146 preempt_latency_start(val);
3148 EXPORT_SYMBOL(preempt_count_add);
3149 NOKPROBE_SYMBOL(preempt_count_add);
3152 * If the value passed in equals to the current preempt count
3153 * then we just enabled preemption. Stop timing the latency.
3155 static inline void preempt_latency_stop(int val)
3157 if (preempt_count() == val)
3158 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3161 void preempt_count_sub(int val)
3163 #ifdef CONFIG_DEBUG_PREEMPT
3167 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3170 * Is the spinlock portion underflowing?
3172 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3173 !(preempt_count() & PREEMPT_MASK)))
3177 preempt_latency_stop(val);
3178 __preempt_count_sub(val);
3180 EXPORT_SYMBOL(preempt_count_sub);
3181 NOKPROBE_SYMBOL(preempt_count_sub);
3184 static inline void preempt_latency_start(int val) { }
3185 static inline void preempt_latency_stop(int val) { }
3189 * Print scheduling while atomic bug:
3191 static noinline void __schedule_bug(struct task_struct *prev)
3193 /* Save this before calling printk(), since that will clobber it */
3194 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3196 if (oops_in_progress)
3199 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3200 prev->comm, prev->pid, preempt_count());
3202 debug_show_held_locks(prev);
3204 if (irqs_disabled())
3205 print_irqtrace_events(prev);
3206 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3207 && in_atomic_preempt_off()) {
3208 pr_err("Preemption disabled at:");
3209 print_ip_sym(preempt_disable_ip);
3213 panic("scheduling while atomic\n");
3216 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3220 * Various schedule()-time debugging checks and statistics:
3222 static inline void schedule_debug(struct task_struct *prev)
3224 #ifdef CONFIG_SCHED_STACK_END_CHECK
3225 if (task_stack_end_corrupted(prev))
3226 panic("corrupted stack end detected inside scheduler\n");
3229 if (unlikely(in_atomic_preempt_off())) {
3230 __schedule_bug(prev);
3231 preempt_count_set(PREEMPT_DISABLED);
3235 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3237 schedstat_inc(this_rq()->sched_count);
3241 * Pick up the highest-prio task:
3243 static inline struct task_struct *
3244 pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
3246 const struct sched_class *class = &fair_sched_class;
3247 struct task_struct *p;
3250 * Optimization: we know that if all tasks are in
3251 * the fair class we can call that function directly:
3253 if (likely(prev->sched_class == class &&
3254 rq->nr_running == rq->cfs.h_nr_running)) {
3255 p = fair_sched_class.pick_next_task(rq, prev, cookie);
3256 if (unlikely(p == RETRY_TASK))
3259 /* assumes fair_sched_class->next == idle_sched_class */
3261 p = idle_sched_class.pick_next_task(rq, prev, cookie);
3267 for_each_class(class) {
3268 p = class->pick_next_task(rq, prev, cookie);
3270 if (unlikely(p == RETRY_TASK))
3276 BUG(); /* the idle class will always have a runnable task */
3280 * __schedule() is the main scheduler function.
3282 * The main means of driving the scheduler and thus entering this function are:
3284 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3286 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3287 * paths. For example, see arch/x86/entry_64.S.
3289 * To drive preemption between tasks, the scheduler sets the flag in timer
3290 * interrupt handler scheduler_tick().
3292 * 3. Wakeups don't really cause entry into schedule(). They add a
3293 * task to the run-queue and that's it.
3295 * Now, if the new task added to the run-queue preempts the current
3296 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3297 * called on the nearest possible occasion:
3299 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3301 * - in syscall or exception context, at the next outmost
3302 * preempt_enable(). (this might be as soon as the wake_up()'s
3305 * - in IRQ context, return from interrupt-handler to
3306 * preemptible context
3308 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3311 * - cond_resched() call
3312 * - explicit schedule() call
3313 * - return from syscall or exception to user-space
3314 * - return from interrupt-handler to user-space
3316 * WARNING: must be called with preemption disabled!
3318 static void __sched notrace __schedule(bool preempt)
3320 struct task_struct *prev, *next;
3321 unsigned long *switch_count;
3322 struct pin_cookie cookie;
3326 cpu = smp_processor_id();
3331 * do_exit() calls schedule() with preemption disabled as an exception;
3332 * however we must fix that up, otherwise the next task will see an
3333 * inconsistent (higher) preempt count.
3335 * It also avoids the below schedule_debug() test from complaining
3338 if (unlikely(prev->state == TASK_DEAD))
3339 preempt_enable_no_resched_notrace();
3341 schedule_debug(prev);
3343 if (sched_feat(HRTICK))
3346 local_irq_disable();
3347 rcu_note_context_switch();
3350 * Make sure that signal_pending_state()->signal_pending() below
3351 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3352 * done by the caller to avoid the race with signal_wake_up().
3354 smp_mb__before_spinlock();
3355 raw_spin_lock(&rq->lock);
3356 cookie = lockdep_pin_lock(&rq->lock);
3358 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3360 switch_count = &prev->nivcsw;
3361 if (!preempt && prev->state) {
3362 if (unlikely(signal_pending_state(prev->state, prev))) {
3363 prev->state = TASK_RUNNING;
3365 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3369 * If a worker went to sleep, notify and ask workqueue
3370 * whether it wants to wake up a task to maintain
3373 if (prev->flags & PF_WQ_WORKER) {
3374 struct task_struct *to_wakeup;
3376 to_wakeup = wq_worker_sleeping(prev);
3378 try_to_wake_up_local(to_wakeup, cookie);
3381 switch_count = &prev->nvcsw;
3384 if (task_on_rq_queued(prev))
3385 update_rq_clock(rq);
3387 next = pick_next_task(rq, prev, cookie);
3388 clear_tsk_need_resched(prev);
3389 clear_preempt_need_resched();
3390 rq->clock_skip_update = 0;
3392 if (likely(prev != next)) {
3397 trace_sched_switch(preempt, prev, next);
3398 rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */
3400 lockdep_unpin_lock(&rq->lock, cookie);
3401 raw_spin_unlock_irq(&rq->lock);
3404 balance_callback(rq);
3406 STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3408 static inline void sched_submit_work(struct task_struct *tsk)
3410 if (!tsk->state || tsk_is_pi_blocked(tsk))
3413 * If we are going to sleep and we have plugged IO queued,
3414 * make sure to submit it to avoid deadlocks.
3416 if (blk_needs_flush_plug(tsk))
3417 blk_schedule_flush_plug(tsk);
3420 asmlinkage __visible void __sched schedule(void)
3422 struct task_struct *tsk = current;
3424 sched_submit_work(tsk);
3428 sched_preempt_enable_no_resched();
3429 } while (need_resched());
3431 EXPORT_SYMBOL(schedule);
3433 #ifdef CONFIG_CONTEXT_TRACKING
3434 asmlinkage __visible void __sched schedule_user(void)
3437 * If we come here after a random call to set_need_resched(),
3438 * or we have been woken up remotely but the IPI has not yet arrived,
3439 * we haven't yet exited the RCU idle mode. Do it here manually until
3440 * we find a better solution.
3442 * NB: There are buggy callers of this function. Ideally we
3443 * should warn if prev_state != CONTEXT_USER, but that will trigger
3444 * too frequently to make sense yet.
3446 enum ctx_state prev_state = exception_enter();
3448 exception_exit(prev_state);
3453 * schedule_preempt_disabled - called with preemption disabled
3455 * Returns with preemption disabled. Note: preempt_count must be 1
3457 void __sched schedule_preempt_disabled(void)
3459 sched_preempt_enable_no_resched();
3464 static void __sched notrace preempt_schedule_common(void)
3468 * Because the function tracer can trace preempt_count_sub()
3469 * and it also uses preempt_enable/disable_notrace(), if
3470 * NEED_RESCHED is set, the preempt_enable_notrace() called
3471 * by the function tracer will call this function again and
3472 * cause infinite recursion.
3474 * Preemption must be disabled here before the function
3475 * tracer can trace. Break up preempt_disable() into two
3476 * calls. One to disable preemption without fear of being
3477 * traced. The other to still record the preemption latency,
3478 * which can also be traced by the function tracer.
3480 preempt_disable_notrace();
3481 preempt_latency_start(1);
3483 preempt_latency_stop(1);
3484 preempt_enable_no_resched_notrace();
3487 * Check again in case we missed a preemption opportunity
3488 * between schedule and now.
3490 } while (need_resched());
3493 #ifdef CONFIG_PREEMPT
3495 * this is the entry point to schedule() from in-kernel preemption
3496 * off of preempt_enable. Kernel preemptions off return from interrupt
3497 * occur there and call schedule directly.
3499 asmlinkage __visible void __sched notrace preempt_schedule(void)
3502 * If there is a non-zero preempt_count or interrupts are disabled,
3503 * we do not want to preempt the current task. Just return..
3505 if (likely(!preemptible()))
3508 preempt_schedule_common();
3510 NOKPROBE_SYMBOL(preempt_schedule);
3511 EXPORT_SYMBOL(preempt_schedule);
3514 * preempt_schedule_notrace - preempt_schedule called by tracing
3516 * The tracing infrastructure uses preempt_enable_notrace to prevent
3517 * recursion and tracing preempt enabling caused by the tracing
3518 * infrastructure itself. But as tracing can happen in areas coming
3519 * from userspace or just about to enter userspace, a preempt enable
3520 * can occur before user_exit() is called. This will cause the scheduler
3521 * to be called when the system is still in usermode.
3523 * To prevent this, the preempt_enable_notrace will use this function
3524 * instead of preempt_schedule() to exit user context if needed before
3525 * calling the scheduler.
3527 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3529 enum ctx_state prev_ctx;
3531 if (likely(!preemptible()))
3536 * Because the function tracer can trace preempt_count_sub()
3537 * and it also uses preempt_enable/disable_notrace(), if
3538 * NEED_RESCHED is set, the preempt_enable_notrace() called
3539 * by the function tracer will call this function again and
3540 * cause infinite recursion.
3542 * Preemption must be disabled here before the function
3543 * tracer can trace. Break up preempt_disable() into two
3544 * calls. One to disable preemption without fear of being
3545 * traced. The other to still record the preemption latency,
3546 * which can also be traced by the function tracer.
3548 preempt_disable_notrace();
3549 preempt_latency_start(1);
3551 * Needs preempt disabled in case user_exit() is traced
3552 * and the tracer calls preempt_enable_notrace() causing
3553 * an infinite recursion.
3555 prev_ctx = exception_enter();
3557 exception_exit(prev_ctx);
3559 preempt_latency_stop(1);
3560 preempt_enable_no_resched_notrace();
3561 } while (need_resched());
3563 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3565 #endif /* CONFIG_PREEMPT */
3568 * this is the entry point to schedule() from kernel preemption
3569 * off of irq context.
3570 * Note, that this is called and return with irqs disabled. This will
3571 * protect us against recursive calling from irq.
3573 asmlinkage __visible void __sched preempt_schedule_irq(void)
3575 enum ctx_state prev_state;
3577 /* Catch callers which need to be fixed */
3578 BUG_ON(preempt_count() || !irqs_disabled());
3580 prev_state = exception_enter();
3586 local_irq_disable();
3587 sched_preempt_enable_no_resched();
3588 } while (need_resched());
3590 exception_exit(prev_state);
3593 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3596 return try_to_wake_up(curr->private, mode, wake_flags);
3598 EXPORT_SYMBOL(default_wake_function);
3600 #ifdef CONFIG_RT_MUTEXES
3603 * rt_mutex_setprio - set the current priority of a task
3605 * @prio: prio value (kernel-internal form)
3607 * This function changes the 'effective' priority of a task. It does
3608 * not touch ->normal_prio like __setscheduler().
3610 * Used by the rt_mutex code to implement priority inheritance
3611 * logic. Call site only calls if the priority of the task changed.
3613 void rt_mutex_setprio(struct task_struct *p, int prio)
3615 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3616 const struct sched_class *prev_class;
3620 BUG_ON(prio > MAX_PRIO);
3622 rq = __task_rq_lock(p, &rf);
3625 * Idle task boosting is a nono in general. There is one
3626 * exception, when PREEMPT_RT and NOHZ is active:
3628 * The idle task calls get_next_timer_interrupt() and holds
3629 * the timer wheel base->lock on the CPU and another CPU wants
3630 * to access the timer (probably to cancel it). We can safely
3631 * ignore the boosting request, as the idle CPU runs this code
3632 * with interrupts disabled and will complete the lock
3633 * protected section without being interrupted. So there is no
3634 * real need to boost.
3636 if (unlikely(p == rq->idle)) {
3637 WARN_ON(p != rq->curr);
3638 WARN_ON(p->pi_blocked_on);
3642 trace_sched_pi_setprio(p, prio);
3645 if (oldprio == prio)
3646 queue_flag &= ~DEQUEUE_MOVE;
3648 prev_class = p->sched_class;
3649 queued = task_on_rq_queued(p);
3650 running = task_current(rq, p);
3652 dequeue_task(rq, p, queue_flag);
3654 put_prev_task(rq, p);
3657 * Boosting condition are:
3658 * 1. -rt task is running and holds mutex A
3659 * --> -dl task blocks on mutex A
3661 * 2. -dl task is running and holds mutex A
3662 * --> -dl task blocks on mutex A and could preempt the
3665 if (dl_prio(prio)) {
3666 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3667 if (!dl_prio(p->normal_prio) ||
3668 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3669 p->dl.dl_boosted = 1;
3670 queue_flag |= ENQUEUE_REPLENISH;
3672 p->dl.dl_boosted = 0;
3673 p->sched_class = &dl_sched_class;
3674 } else if (rt_prio(prio)) {
3675 if (dl_prio(oldprio))
3676 p->dl.dl_boosted = 0;
3678 queue_flag |= ENQUEUE_HEAD;
3679 p->sched_class = &rt_sched_class;
3681 if (dl_prio(oldprio))
3682 p->dl.dl_boosted = 0;
3683 if (rt_prio(oldprio))
3685 p->sched_class = &fair_sched_class;
3691 p->sched_class->set_curr_task(rq);
3693 enqueue_task(rq, p, queue_flag);
3695 check_class_changed(rq, p, prev_class, oldprio);
3697 preempt_disable(); /* avoid rq from going away on us */
3698 __task_rq_unlock(rq, &rf);
3700 balance_callback(rq);
3705 void set_user_nice(struct task_struct *p, long nice)
3707 int old_prio, delta, queued;
3711 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3714 * We have to be careful, if called from sys_setpriority(),
3715 * the task might be in the middle of scheduling on another CPU.
3717 rq = task_rq_lock(p, &rf);
3719 * The RT priorities are set via sched_setscheduler(), but we still
3720 * allow the 'normal' nice value to be set - but as expected
3721 * it wont have any effect on scheduling until the task is
3722 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3724 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3725 p->static_prio = NICE_TO_PRIO(nice);
3728 queued = task_on_rq_queued(p);
3730 dequeue_task(rq, p, DEQUEUE_SAVE);
3732 p->static_prio = NICE_TO_PRIO(nice);
3735 p->prio = effective_prio(p);
3736 delta = p->prio - old_prio;
3739 enqueue_task(rq, p, ENQUEUE_RESTORE);
3741 * If the task increased its priority or is running and
3742 * lowered its priority, then reschedule its CPU:
3744 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3748 task_rq_unlock(rq, p, &rf);
3750 EXPORT_SYMBOL(set_user_nice);
3753 * can_nice - check if a task can reduce its nice value
3757 int can_nice(const struct task_struct *p, const int nice)
3759 /* convert nice value [19,-20] to rlimit style value [1,40] */
3760 int nice_rlim = nice_to_rlimit(nice);
3762 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3763 capable(CAP_SYS_NICE));
3766 #ifdef __ARCH_WANT_SYS_NICE
3769 * sys_nice - change the priority of the current process.
3770 * @increment: priority increment
3772 * sys_setpriority is a more generic, but much slower function that
3773 * does similar things.
3775 SYSCALL_DEFINE1(nice, int, increment)
3780 * Setpriority might change our priority at the same moment.
3781 * We don't have to worry. Conceptually one call occurs first
3782 * and we have a single winner.
3784 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3785 nice = task_nice(current) + increment;
3787 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3788 if (increment < 0 && !can_nice(current, nice))
3791 retval = security_task_setnice(current, nice);
3795 set_user_nice(current, nice);
3802 * task_prio - return the priority value of a given task.
3803 * @p: the task in question.
3805 * Return: The priority value as seen by users in /proc.
3806 * RT tasks are offset by -200. Normal tasks are centered
3807 * around 0, value goes from -16 to +15.
3809 int task_prio(const struct task_struct *p)
3811 return p->prio - MAX_RT_PRIO;
3815 * idle_cpu - is a given cpu idle currently?
3816 * @cpu: the processor in question.
3818 * Return: 1 if the CPU is currently idle. 0 otherwise.
3820 int idle_cpu(int cpu)
3822 struct rq *rq = cpu_rq(cpu);
3824 if (rq->curr != rq->idle)
3831 if (!llist_empty(&rq->wake_list))
3839 * idle_task - return the idle task for a given cpu.
3840 * @cpu: the processor in question.
3842 * Return: The idle task for the cpu @cpu.
3844 struct task_struct *idle_task(int cpu)
3846 return cpu_rq(cpu)->idle;
3850 * find_process_by_pid - find a process with a matching PID value.
3851 * @pid: the pid in question.
3853 * The task of @pid, if found. %NULL otherwise.
3855 static struct task_struct *find_process_by_pid(pid_t pid)
3857 return pid ? find_task_by_vpid(pid) : current;
3861 * This function initializes the sched_dl_entity of a newly becoming
3862 * SCHED_DEADLINE task.
3864 * Only the static values are considered here, the actual runtime and the
3865 * absolute deadline will be properly calculated when the task is enqueued
3866 * for the first time with its new policy.
3869 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3871 struct sched_dl_entity *dl_se = &p->dl;
3873 dl_se->dl_runtime = attr->sched_runtime;
3874 dl_se->dl_deadline = attr->sched_deadline;
3875 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3876 dl_se->flags = attr->sched_flags;
3877 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3880 * Changing the parameters of a task is 'tricky' and we're not doing
3881 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3883 * What we SHOULD do is delay the bandwidth release until the 0-lag
3884 * point. This would include retaining the task_struct until that time
3885 * and change dl_overflow() to not immediately decrement the current
3888 * Instead we retain the current runtime/deadline and let the new
3889 * parameters take effect after the current reservation period lapses.
3890 * This is safe (albeit pessimistic) because the 0-lag point is always
3891 * before the current scheduling deadline.
3893 * We can still have temporary overloads because we do not delay the
3894 * change in bandwidth until that time; so admission control is
3895 * not on the safe side. It does however guarantee tasks will never
3896 * consume more than promised.
3901 * sched_setparam() passes in -1 for its policy, to let the functions
3902 * it calls know not to change it.
3904 #define SETPARAM_POLICY -1
3906 static void __setscheduler_params(struct task_struct *p,
3907 const struct sched_attr *attr)
3909 int policy = attr->sched_policy;
3911 if (policy == SETPARAM_POLICY)
3916 if (dl_policy(policy))
3917 __setparam_dl(p, attr);
3918 else if (fair_policy(policy))
3919 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3922 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3923 * !rt_policy. Always setting this ensures that things like
3924 * getparam()/getattr() don't report silly values for !rt tasks.
3926 p->rt_priority = attr->sched_priority;
3927 p->normal_prio = normal_prio(p);
3931 /* Actually do priority change: must hold pi & rq lock. */
3932 static void __setscheduler(struct rq *rq, struct task_struct *p,
3933 const struct sched_attr *attr, bool keep_boost)
3935 __setscheduler_params(p, attr);
3938 * Keep a potential priority boosting if called from
3939 * sched_setscheduler().
3942 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3944 p->prio = normal_prio(p);
3946 if (dl_prio(p->prio))
3947 p->sched_class = &dl_sched_class;
3948 else if (rt_prio(p->prio))
3949 p->sched_class = &rt_sched_class;
3951 p->sched_class = &fair_sched_class;
3955 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3957 struct sched_dl_entity *dl_se = &p->dl;
3959 attr->sched_priority = p->rt_priority;
3960 attr->sched_runtime = dl_se->dl_runtime;
3961 attr->sched_deadline = dl_se->dl_deadline;
3962 attr->sched_period = dl_se->dl_period;
3963 attr->sched_flags = dl_se->flags;
3967 * This function validates the new parameters of a -deadline task.
3968 * We ask for the deadline not being zero, and greater or equal
3969 * than the runtime, as well as the period of being zero or
3970 * greater than deadline. Furthermore, we have to be sure that
3971 * user parameters are above the internal resolution of 1us (we
3972 * check sched_runtime only since it is always the smaller one) and
3973 * below 2^63 ns (we have to check both sched_deadline and
3974 * sched_period, as the latter can be zero).
3977 __checkparam_dl(const struct sched_attr *attr)
3980 if (attr->sched_deadline == 0)
3984 * Since we truncate DL_SCALE bits, make sure we're at least
3987 if (attr->sched_runtime < (1ULL << DL_SCALE))
3991 * Since we use the MSB for wrap-around and sign issues, make
3992 * sure it's not set (mind that period can be equal to zero).
3994 if (attr->sched_deadline & (1ULL << 63) ||
3995 attr->sched_period & (1ULL << 63))
3998 /* runtime <= deadline <= period (if period != 0) */
3999 if ((attr->sched_period != 0 &&
4000 attr->sched_period < attr->sched_deadline) ||
4001 attr->sched_deadline < attr->sched_runtime)
4008 * check the target process has a UID that matches the current process's
4010 static bool check_same_owner(struct task_struct *p)
4012 const struct cred *cred = current_cred(), *pcred;
4016 pcred = __task_cred(p);
4017 match = (uid_eq(cred->euid, pcred->euid) ||
4018 uid_eq(cred->euid, pcred->uid));
4023 static bool dl_param_changed(struct task_struct *p,
4024 const struct sched_attr *attr)
4026 struct sched_dl_entity *dl_se = &p->dl;
4028 if (dl_se->dl_runtime != attr->sched_runtime ||
4029 dl_se->dl_deadline != attr->sched_deadline ||
4030 dl_se->dl_period != attr->sched_period ||
4031 dl_se->flags != attr->sched_flags)
4037 static int __sched_setscheduler(struct task_struct *p,
4038 const struct sched_attr *attr,
4041 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4042 MAX_RT_PRIO - 1 - attr->sched_priority;
4043 int retval, oldprio, oldpolicy = -1, queued, running;
4044 int new_effective_prio, policy = attr->sched_policy;
4045 const struct sched_class *prev_class;
4048 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4051 /* may grab non-irq protected spin_locks */
4052 BUG_ON(in_interrupt());
4054 /* double check policy once rq lock held */
4056 reset_on_fork = p->sched_reset_on_fork;
4057 policy = oldpolicy = p->policy;
4059 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4061 if (!valid_policy(policy))
4065 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4069 * Valid priorities for SCHED_FIFO and SCHED_RR are
4070 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4071 * SCHED_BATCH and SCHED_IDLE is 0.
4073 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4074 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4076 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4077 (rt_policy(policy) != (attr->sched_priority != 0)))
4081 * Allow unprivileged RT tasks to decrease priority:
4083 if (user && !capable(CAP_SYS_NICE)) {
4084 if (fair_policy(policy)) {
4085 if (attr->sched_nice < task_nice(p) &&
4086 !can_nice(p, attr->sched_nice))
4090 if (rt_policy(policy)) {
4091 unsigned long rlim_rtprio =
4092 task_rlimit(p, RLIMIT_RTPRIO);
4094 /* can't set/change the rt policy */
4095 if (policy != p->policy && !rlim_rtprio)
4098 /* can't increase priority */
4099 if (attr->sched_priority > p->rt_priority &&
4100 attr->sched_priority > rlim_rtprio)
4105 * Can't set/change SCHED_DEADLINE policy at all for now
4106 * (safest behavior); in the future we would like to allow
4107 * unprivileged DL tasks to increase their relative deadline
4108 * or reduce their runtime (both ways reducing utilization)
4110 if (dl_policy(policy))
4114 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4115 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4117 if (idle_policy(p->policy) && !idle_policy(policy)) {
4118 if (!can_nice(p, task_nice(p)))
4122 /* can't change other user's priorities */
4123 if (!check_same_owner(p))
4126 /* Normal users shall not reset the sched_reset_on_fork flag */
4127 if (p->sched_reset_on_fork && !reset_on_fork)
4132 retval = security_task_setscheduler(p);
4138 * make sure no PI-waiters arrive (or leave) while we are
4139 * changing the priority of the task:
4141 * To be able to change p->policy safely, the appropriate
4142 * runqueue lock must be held.
4144 rq = task_rq_lock(p, &rf);
4147 * Changing the policy of the stop threads its a very bad idea
4149 if (p == rq->stop) {
4150 task_rq_unlock(rq, p, &rf);
4155 * If not changing anything there's no need to proceed further,
4156 * but store a possible modification of reset_on_fork.
4158 if (unlikely(policy == p->policy)) {
4159 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4161 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4163 if (dl_policy(policy) && dl_param_changed(p, attr))
4166 p->sched_reset_on_fork = reset_on_fork;
4167 task_rq_unlock(rq, p, &rf);
4173 #ifdef CONFIG_RT_GROUP_SCHED
4175 * Do not allow realtime tasks into groups that have no runtime
4178 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4179 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4180 !task_group_is_autogroup(task_group(p))) {
4181 task_rq_unlock(rq, p, &rf);
4186 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4187 cpumask_t *span = rq->rd->span;
4190 * Don't allow tasks with an affinity mask smaller than
4191 * the entire root_domain to become SCHED_DEADLINE. We
4192 * will also fail if there's no bandwidth available.
4194 if (!cpumask_subset(span, &p->cpus_allowed) ||
4195 rq->rd->dl_bw.bw == 0) {
4196 task_rq_unlock(rq, p, &rf);
4203 /* recheck policy now with rq lock held */
4204 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4205 policy = oldpolicy = -1;
4206 task_rq_unlock(rq, p, &rf);
4211 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4212 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4215 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4216 task_rq_unlock(rq, p, &rf);
4220 p->sched_reset_on_fork = reset_on_fork;
4225 * Take priority boosted tasks into account. If the new
4226 * effective priority is unchanged, we just store the new
4227 * normal parameters and do not touch the scheduler class and
4228 * the runqueue. This will be done when the task deboost
4231 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4232 if (new_effective_prio == oldprio)
4233 queue_flags &= ~DEQUEUE_MOVE;
4236 queued = task_on_rq_queued(p);
4237 running = task_current(rq, p);
4239 dequeue_task(rq, p, queue_flags);
4241 put_prev_task(rq, p);
4243 prev_class = p->sched_class;
4244 __setscheduler(rq, p, attr, pi);
4247 p->sched_class->set_curr_task(rq);
4250 * We enqueue to tail when the priority of a task is
4251 * increased (user space view).
4253 if (oldprio < p->prio)
4254 queue_flags |= ENQUEUE_HEAD;
4256 enqueue_task(rq, p, queue_flags);
4259 check_class_changed(rq, p, prev_class, oldprio);
4260 preempt_disable(); /* avoid rq from going away on us */
4261 task_rq_unlock(rq, p, &rf);
4264 rt_mutex_adjust_pi(p);
4267 * Run balance callbacks after we've adjusted the PI chain.
4269 balance_callback(rq);
4275 static int _sched_setscheduler(struct task_struct *p, int policy,
4276 const struct sched_param *param, bool check)
4278 struct sched_attr attr = {
4279 .sched_policy = policy,
4280 .sched_priority = param->sched_priority,
4281 .sched_nice = PRIO_TO_NICE(p->static_prio),
4284 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4285 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4286 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4287 policy &= ~SCHED_RESET_ON_FORK;
4288 attr.sched_policy = policy;
4291 return __sched_setscheduler(p, &attr, check, true);
4294 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4295 * @p: the task in question.
4296 * @policy: new policy.
4297 * @param: structure containing the new RT priority.
4299 * Return: 0 on success. An error code otherwise.
4301 * NOTE that the task may be already dead.
4303 int sched_setscheduler(struct task_struct *p, int policy,
4304 const struct sched_param *param)
4306 return _sched_setscheduler(p, policy, param, true);
4308 EXPORT_SYMBOL_GPL(sched_setscheduler);
4310 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4312 return __sched_setscheduler(p, attr, true, true);
4314 EXPORT_SYMBOL_GPL(sched_setattr);
4317 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4318 * @p: the task in question.
4319 * @policy: new policy.
4320 * @param: structure containing the new RT priority.
4322 * Just like sched_setscheduler, only don't bother checking if the
4323 * current context has permission. For example, this is needed in
4324 * stop_machine(): we create temporary high priority worker threads,
4325 * but our caller might not have that capability.
4327 * Return: 0 on success. An error code otherwise.
4329 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4330 const struct sched_param *param)
4332 return _sched_setscheduler(p, policy, param, false);
4334 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4337 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4339 struct sched_param lparam;
4340 struct task_struct *p;
4343 if (!param || pid < 0)
4345 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4350 p = find_process_by_pid(pid);
4352 retval = sched_setscheduler(p, policy, &lparam);
4359 * Mimics kernel/events/core.c perf_copy_attr().
4361 static int sched_copy_attr(struct sched_attr __user *uattr,
4362 struct sched_attr *attr)
4367 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4371 * zero the full structure, so that a short copy will be nice.
4373 memset(attr, 0, sizeof(*attr));
4375 ret = get_user(size, &uattr->size);
4379 if (size > PAGE_SIZE) /* silly large */
4382 if (!size) /* abi compat */
4383 size = SCHED_ATTR_SIZE_VER0;
4385 if (size < SCHED_ATTR_SIZE_VER0)
4389 * If we're handed a bigger struct than we know of,
4390 * ensure all the unknown bits are 0 - i.e. new
4391 * user-space does not rely on any kernel feature
4392 * extensions we dont know about yet.
4394 if (size > sizeof(*attr)) {
4395 unsigned char __user *addr;
4396 unsigned char __user *end;
4399 addr = (void __user *)uattr + sizeof(*attr);
4400 end = (void __user *)uattr + size;
4402 for (; addr < end; addr++) {
4403 ret = get_user(val, addr);
4409 size = sizeof(*attr);
4412 ret = copy_from_user(attr, uattr, size);
4417 * XXX: do we want to be lenient like existing syscalls; or do we want
4418 * to be strict and return an error on out-of-bounds values?
4420 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4425 put_user(sizeof(*attr), &uattr->size);
4430 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4431 * @pid: the pid in question.
4432 * @policy: new policy.
4433 * @param: structure containing the new RT priority.
4435 * Return: 0 on success. An error code otherwise.
4437 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4438 struct sched_param __user *, param)
4440 /* negative values for policy are not valid */
4444 return do_sched_setscheduler(pid, policy, param);
4448 * sys_sched_setparam - set/change the RT priority of a thread
4449 * @pid: the pid in question.
4450 * @param: structure containing the new RT priority.
4452 * Return: 0 on success. An error code otherwise.
4454 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4456 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4460 * sys_sched_setattr - same as above, but with extended sched_attr
4461 * @pid: the pid in question.
4462 * @uattr: structure containing the extended parameters.
4463 * @flags: for future extension.
4465 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4466 unsigned int, flags)
4468 struct sched_attr attr;
4469 struct task_struct *p;
4472 if (!uattr || pid < 0 || flags)
4475 retval = sched_copy_attr(uattr, &attr);
4479 if ((int)attr.sched_policy < 0)
4484 p = find_process_by_pid(pid);
4486 retval = sched_setattr(p, &attr);
4493 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4494 * @pid: the pid in question.
4496 * Return: On success, the policy of the thread. Otherwise, a negative error
4499 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4501 struct task_struct *p;
4509 p = find_process_by_pid(pid);
4511 retval = security_task_getscheduler(p);
4514 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4521 * sys_sched_getparam - get the RT priority of a thread
4522 * @pid: the pid in question.
4523 * @param: structure containing the RT priority.
4525 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4528 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4530 struct sched_param lp = { .sched_priority = 0 };
4531 struct task_struct *p;
4534 if (!param || pid < 0)
4538 p = find_process_by_pid(pid);
4543 retval = security_task_getscheduler(p);
4547 if (task_has_rt_policy(p))
4548 lp.sched_priority = p->rt_priority;
4552 * This one might sleep, we cannot do it with a spinlock held ...
4554 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4563 static int sched_read_attr(struct sched_attr __user *uattr,
4564 struct sched_attr *attr,
4569 if (!access_ok(VERIFY_WRITE, uattr, usize))
4573 * If we're handed a smaller struct than we know of,
4574 * ensure all the unknown bits are 0 - i.e. old
4575 * user-space does not get uncomplete information.
4577 if (usize < sizeof(*attr)) {
4578 unsigned char *addr;
4581 addr = (void *)attr + usize;
4582 end = (void *)attr + sizeof(*attr);
4584 for (; addr < end; addr++) {
4592 ret = copy_to_user(uattr, attr, attr->size);
4600 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4601 * @pid: the pid in question.
4602 * @uattr: structure containing the extended parameters.
4603 * @size: sizeof(attr) for fwd/bwd comp.
4604 * @flags: for future extension.
4606 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4607 unsigned int, size, unsigned int, flags)
4609 struct sched_attr attr = {
4610 .size = sizeof(struct sched_attr),
4612 struct task_struct *p;
4615 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4616 size < SCHED_ATTR_SIZE_VER0 || flags)
4620 p = find_process_by_pid(pid);
4625 retval = security_task_getscheduler(p);
4629 attr.sched_policy = p->policy;
4630 if (p->sched_reset_on_fork)
4631 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4632 if (task_has_dl_policy(p))
4633 __getparam_dl(p, &attr);
4634 else if (task_has_rt_policy(p))
4635 attr.sched_priority = p->rt_priority;
4637 attr.sched_nice = task_nice(p);
4641 retval = sched_read_attr(uattr, &attr, size);
4649 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4651 cpumask_var_t cpus_allowed, new_mask;
4652 struct task_struct *p;
4657 p = find_process_by_pid(pid);
4663 /* Prevent p going away */
4667 if (p->flags & PF_NO_SETAFFINITY) {
4671 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4675 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4677 goto out_free_cpus_allowed;
4680 if (!check_same_owner(p)) {
4682 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4684 goto out_free_new_mask;
4689 retval = security_task_setscheduler(p);
4691 goto out_free_new_mask;
4694 cpuset_cpus_allowed(p, cpus_allowed);
4695 cpumask_and(new_mask, in_mask, cpus_allowed);
4698 * Since bandwidth control happens on root_domain basis,
4699 * if admission test is enabled, we only admit -deadline
4700 * tasks allowed to run on all the CPUs in the task's
4704 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4706 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4709 goto out_free_new_mask;
4715 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4718 cpuset_cpus_allowed(p, cpus_allowed);
4719 if (!cpumask_subset(new_mask, cpus_allowed)) {
4721 * We must have raced with a concurrent cpuset
4722 * update. Just reset the cpus_allowed to the
4723 * cpuset's cpus_allowed
4725 cpumask_copy(new_mask, cpus_allowed);
4730 free_cpumask_var(new_mask);
4731 out_free_cpus_allowed:
4732 free_cpumask_var(cpus_allowed);
4738 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4739 struct cpumask *new_mask)
4741 if (len < cpumask_size())
4742 cpumask_clear(new_mask);
4743 else if (len > cpumask_size())
4744 len = cpumask_size();
4746 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4750 * sys_sched_setaffinity - set the cpu affinity of a process
4751 * @pid: pid of the process
4752 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4753 * @user_mask_ptr: user-space pointer to the new cpu mask
4755 * Return: 0 on success. An error code otherwise.
4757 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4758 unsigned long __user *, user_mask_ptr)
4760 cpumask_var_t new_mask;
4763 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4766 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4768 retval = sched_setaffinity(pid, new_mask);
4769 free_cpumask_var(new_mask);
4773 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4775 struct task_struct *p;
4776 unsigned long flags;
4782 p = find_process_by_pid(pid);
4786 retval = security_task_getscheduler(p);
4790 raw_spin_lock_irqsave(&p->pi_lock, flags);
4791 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4792 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4801 * sys_sched_getaffinity - get the cpu affinity of a process
4802 * @pid: pid of the process
4803 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4804 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4806 * Return: size of CPU mask copied to user_mask_ptr on success. An
4807 * error code otherwise.
4809 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4810 unsigned long __user *, user_mask_ptr)
4815 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4817 if (len & (sizeof(unsigned long)-1))
4820 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4823 ret = sched_getaffinity(pid, mask);
4825 size_t retlen = min_t(size_t, len, cpumask_size());
4827 if (copy_to_user(user_mask_ptr, mask, retlen))
4832 free_cpumask_var(mask);
4838 * sys_sched_yield - yield the current processor to other threads.
4840 * This function yields the current CPU to other tasks. If there are no
4841 * other threads running on this CPU then this function will return.
4845 SYSCALL_DEFINE0(sched_yield)
4847 struct rq *rq = this_rq_lock();
4849 schedstat_inc(rq->yld_count);
4850 current->sched_class->yield_task(rq);
4853 * Since we are going to call schedule() anyway, there's
4854 * no need to preempt or enable interrupts:
4856 __release(rq->lock);
4857 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4858 do_raw_spin_unlock(&rq->lock);
4859 sched_preempt_enable_no_resched();
4866 int __sched _cond_resched(void)
4868 if (should_resched(0)) {
4869 preempt_schedule_common();
4874 EXPORT_SYMBOL(_cond_resched);
4877 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4878 * call schedule, and on return reacquire the lock.
4880 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4881 * operations here to prevent schedule() from being called twice (once via
4882 * spin_unlock(), once by hand).
4884 int __cond_resched_lock(spinlock_t *lock)
4886 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4889 lockdep_assert_held(lock);
4891 if (spin_needbreak(lock) || resched) {
4894 preempt_schedule_common();
4902 EXPORT_SYMBOL(__cond_resched_lock);
4904 int __sched __cond_resched_softirq(void)
4906 BUG_ON(!in_softirq());
4908 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4910 preempt_schedule_common();
4916 EXPORT_SYMBOL(__cond_resched_softirq);
4919 * yield - yield the current processor to other threads.
4921 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4923 * The scheduler is at all times free to pick the calling task as the most
4924 * eligible task to run, if removing the yield() call from your code breaks
4925 * it, its already broken.
4927 * Typical broken usage is:
4932 * where one assumes that yield() will let 'the other' process run that will
4933 * make event true. If the current task is a SCHED_FIFO task that will never
4934 * happen. Never use yield() as a progress guarantee!!
4936 * If you want to use yield() to wait for something, use wait_event().
4937 * If you want to use yield() to be 'nice' for others, use cond_resched().
4938 * If you still want to use yield(), do not!
4940 void __sched yield(void)
4942 set_current_state(TASK_RUNNING);
4945 EXPORT_SYMBOL(yield);
4948 * yield_to - yield the current processor to another thread in
4949 * your thread group, or accelerate that thread toward the
4950 * processor it's on.
4952 * @preempt: whether task preemption is allowed or not
4954 * It's the caller's job to ensure that the target task struct
4955 * can't go away on us before we can do any checks.
4958 * true (>0) if we indeed boosted the target task.
4959 * false (0) if we failed to boost the target.
4960 * -ESRCH if there's no task to yield to.
4962 int __sched yield_to(struct task_struct *p, bool preempt)
4964 struct task_struct *curr = current;
4965 struct rq *rq, *p_rq;
4966 unsigned long flags;
4969 local_irq_save(flags);
4975 * If we're the only runnable task on the rq and target rq also
4976 * has only one task, there's absolutely no point in yielding.
4978 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4983 double_rq_lock(rq, p_rq);
4984 if (task_rq(p) != p_rq) {
4985 double_rq_unlock(rq, p_rq);
4989 if (!curr->sched_class->yield_to_task)
4992 if (curr->sched_class != p->sched_class)
4995 if (task_running(p_rq, p) || p->state)
4998 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5000 schedstat_inc(rq->yld_count);
5002 * Make p's CPU reschedule; pick_next_entity takes care of
5005 if (preempt && rq != p_rq)
5010 double_rq_unlock(rq, p_rq);
5012 local_irq_restore(flags);
5019 EXPORT_SYMBOL_GPL(yield_to);
5022 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5023 * that process accounting knows that this is a task in IO wait state.
5025 long __sched io_schedule_timeout(long timeout)
5027 int old_iowait = current->in_iowait;
5031 current->in_iowait = 1;
5032 blk_schedule_flush_plug(current);
5034 delayacct_blkio_start();
5036 atomic_inc(&rq->nr_iowait);
5037 ret = schedule_timeout(timeout);
5038 current->in_iowait = old_iowait;
5039 atomic_dec(&rq->nr_iowait);
5040 delayacct_blkio_end();
5044 EXPORT_SYMBOL(io_schedule_timeout);
5047 * sys_sched_get_priority_max - return maximum RT priority.
5048 * @policy: scheduling class.
5050 * Return: On success, this syscall returns the maximum
5051 * rt_priority that can be used by a given scheduling class.
5052 * On failure, a negative error code is returned.
5054 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5061 ret = MAX_USER_RT_PRIO-1;
5063 case SCHED_DEADLINE:
5074 * sys_sched_get_priority_min - return minimum RT priority.
5075 * @policy: scheduling class.
5077 * Return: On success, this syscall returns the minimum
5078 * rt_priority that can be used by a given scheduling class.
5079 * On failure, a negative error code is returned.
5081 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5090 case SCHED_DEADLINE:
5100 * sys_sched_rr_get_interval - return the default timeslice of a process.
5101 * @pid: pid of the process.
5102 * @interval: userspace pointer to the timeslice value.
5104 * this syscall writes the default timeslice value of a given process
5105 * into the user-space timespec buffer. A value of '0' means infinity.
5107 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5110 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5111 struct timespec __user *, interval)
5113 struct task_struct *p;
5114 unsigned int time_slice;
5125 p = find_process_by_pid(pid);
5129 retval = security_task_getscheduler(p);
5133 rq = task_rq_lock(p, &rf);
5135 if (p->sched_class->get_rr_interval)
5136 time_slice = p->sched_class->get_rr_interval(rq, p);
5137 task_rq_unlock(rq, p, &rf);
5140 jiffies_to_timespec(time_slice, &t);
5141 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5149 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5151 void sched_show_task(struct task_struct *p)
5153 unsigned long free = 0;
5155 unsigned long state = p->state;
5158 state = __ffs(state) + 1;
5159 printk(KERN_INFO "%-15.15s %c", p->comm,
5160 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5161 #if BITS_PER_LONG == 32
5162 if (state == TASK_RUNNING)
5163 printk(KERN_CONT " running ");
5165 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5167 if (state == TASK_RUNNING)
5168 printk(KERN_CONT " running task ");
5170 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5172 #ifdef CONFIG_DEBUG_STACK_USAGE
5173 free = stack_not_used(p);
5178 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5180 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5181 task_pid_nr(p), ppid,
5182 (unsigned long)task_thread_info(p)->flags);
5184 print_worker_info(KERN_INFO, p);
5185 show_stack(p, NULL);
5188 void show_state_filter(unsigned long state_filter)
5190 struct task_struct *g, *p;
5192 #if BITS_PER_LONG == 32
5194 " task PC stack pid father\n");
5197 " task PC stack pid father\n");
5200 for_each_process_thread(g, p) {
5202 * reset the NMI-timeout, listing all files on a slow
5203 * console might take a lot of time:
5204 * Also, reset softlockup watchdogs on all CPUs, because
5205 * another CPU might be blocked waiting for us to process
5208 touch_nmi_watchdog();
5209 touch_all_softlockup_watchdogs();
5210 if (!state_filter || (p->state & state_filter))
5214 #ifdef CONFIG_SCHED_DEBUG
5216 sysrq_sched_debug_show();
5220 * Only show locks if all tasks are dumped:
5223 debug_show_all_locks();
5226 void init_idle_bootup_task(struct task_struct *idle)
5228 idle->sched_class = &idle_sched_class;
5232 * init_idle - set up an idle thread for a given CPU
5233 * @idle: task in question
5234 * @cpu: cpu the idle task belongs to
5236 * NOTE: this function does not set the idle thread's NEED_RESCHED
5237 * flag, to make booting more robust.
5239 void init_idle(struct task_struct *idle, int cpu)
5241 struct rq *rq = cpu_rq(cpu);
5242 unsigned long flags;
5244 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5245 raw_spin_lock(&rq->lock);
5247 __sched_fork(0, idle);
5248 idle->state = TASK_RUNNING;
5249 idle->se.exec_start = sched_clock();
5251 kasan_unpoison_task_stack(idle);
5255 * Its possible that init_idle() gets called multiple times on a task,
5256 * in that case do_set_cpus_allowed() will not do the right thing.
5258 * And since this is boot we can forgo the serialization.
5260 set_cpus_allowed_common(idle, cpumask_of(cpu));
5263 * We're having a chicken and egg problem, even though we are
5264 * holding rq->lock, the cpu isn't yet set to this cpu so the
5265 * lockdep check in task_group() will fail.
5267 * Similar case to sched_fork(). / Alternatively we could
5268 * use task_rq_lock() here and obtain the other rq->lock.
5273 __set_task_cpu(idle, cpu);
5276 rq->curr = rq->idle = idle;
5277 idle->on_rq = TASK_ON_RQ_QUEUED;
5281 raw_spin_unlock(&rq->lock);
5282 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5284 /* Set the preempt count _outside_ the spinlocks! */
5285 init_idle_preempt_count(idle, cpu);
5288 * The idle tasks have their own, simple scheduling class:
5290 idle->sched_class = &idle_sched_class;
5291 ftrace_graph_init_idle_task(idle, cpu);
5292 vtime_init_idle(idle, cpu);
5294 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5298 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5299 const struct cpumask *trial)
5301 int ret = 1, trial_cpus;
5302 struct dl_bw *cur_dl_b;
5303 unsigned long flags;
5305 if (!cpumask_weight(cur))
5308 rcu_read_lock_sched();
5309 cur_dl_b = dl_bw_of(cpumask_any(cur));
5310 trial_cpus = cpumask_weight(trial);
5312 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5313 if (cur_dl_b->bw != -1 &&
5314 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5316 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5317 rcu_read_unlock_sched();
5322 int task_can_attach(struct task_struct *p,
5323 const struct cpumask *cs_cpus_allowed)
5328 * Kthreads which disallow setaffinity shouldn't be moved
5329 * to a new cpuset; we don't want to change their cpu
5330 * affinity and isolating such threads by their set of
5331 * allowed nodes is unnecessary. Thus, cpusets are not
5332 * applicable for such threads. This prevents checking for
5333 * success of set_cpus_allowed_ptr() on all attached tasks
5334 * before cpus_allowed may be changed.
5336 if (p->flags & PF_NO_SETAFFINITY) {
5342 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5344 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5349 unsigned long flags;
5351 rcu_read_lock_sched();
5352 dl_b = dl_bw_of(dest_cpu);
5353 raw_spin_lock_irqsave(&dl_b->lock, flags);
5354 cpus = dl_bw_cpus(dest_cpu);
5355 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5360 * We reserve space for this task in the destination
5361 * root_domain, as we can't fail after this point.
5362 * We will free resources in the source root_domain
5363 * later on (see set_cpus_allowed_dl()).
5365 __dl_add(dl_b, p->dl.dl_bw);
5367 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5368 rcu_read_unlock_sched();
5378 static bool sched_smp_initialized __read_mostly;
5380 #ifdef CONFIG_NUMA_BALANCING
5381 /* Migrate current task p to target_cpu */
5382 int migrate_task_to(struct task_struct *p, int target_cpu)
5384 struct migration_arg arg = { p, target_cpu };
5385 int curr_cpu = task_cpu(p);
5387 if (curr_cpu == target_cpu)
5390 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5393 /* TODO: This is not properly updating schedstats */
5395 trace_sched_move_numa(p, curr_cpu, target_cpu);
5396 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5400 * Requeue a task on a given node and accurately track the number of NUMA
5401 * tasks on the runqueues
5403 void sched_setnuma(struct task_struct *p, int nid)
5405 bool queued, running;
5409 rq = task_rq_lock(p, &rf);
5410 queued = task_on_rq_queued(p);
5411 running = task_current(rq, p);
5414 dequeue_task(rq, p, DEQUEUE_SAVE);
5416 put_prev_task(rq, p);
5418 p->numa_preferred_nid = nid;
5421 p->sched_class->set_curr_task(rq);
5423 enqueue_task(rq, p, ENQUEUE_RESTORE);
5424 task_rq_unlock(rq, p, &rf);
5426 #endif /* CONFIG_NUMA_BALANCING */
5428 #ifdef CONFIG_HOTPLUG_CPU
5430 * Ensures that the idle task is using init_mm right before its cpu goes
5433 void idle_task_exit(void)
5435 struct mm_struct *mm = current->active_mm;
5437 BUG_ON(cpu_online(smp_processor_id()));
5439 if (mm != &init_mm) {
5440 switch_mm_irqs_off(mm, &init_mm, current);
5441 finish_arch_post_lock_switch();
5447 * Since this CPU is going 'away' for a while, fold any nr_active delta
5448 * we might have. Assumes we're called after migrate_tasks() so that the
5449 * nr_active count is stable. We need to take the teardown thread which
5450 * is calling this into account, so we hand in adjust = 1 to the load
5453 * Also see the comment "Global load-average calculations".
5455 static void calc_load_migrate(struct rq *rq)
5457 long delta = calc_load_fold_active(rq, 1);
5459 atomic_long_add(delta, &calc_load_tasks);
5462 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5466 static const struct sched_class fake_sched_class = {
5467 .put_prev_task = put_prev_task_fake,
5470 static struct task_struct fake_task = {
5472 * Avoid pull_{rt,dl}_task()
5474 .prio = MAX_PRIO + 1,
5475 .sched_class = &fake_sched_class,
5479 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5480 * try_to_wake_up()->select_task_rq().
5482 * Called with rq->lock held even though we'er in stop_machine() and
5483 * there's no concurrency possible, we hold the required locks anyway
5484 * because of lock validation efforts.
5486 static void migrate_tasks(struct rq *dead_rq)
5488 struct rq *rq = dead_rq;
5489 struct task_struct *next, *stop = rq->stop;
5490 struct pin_cookie cookie;
5494 * Fudge the rq selection such that the below task selection loop
5495 * doesn't get stuck on the currently eligible stop task.
5497 * We're currently inside stop_machine() and the rq is either stuck
5498 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5499 * either way we should never end up calling schedule() until we're
5505 * put_prev_task() and pick_next_task() sched
5506 * class method both need to have an up-to-date
5507 * value of rq->clock[_task]
5509 update_rq_clock(rq);
5513 * There's this thread running, bail when that's the only
5516 if (rq->nr_running == 1)
5520 * pick_next_task assumes pinned rq->lock.
5522 cookie = lockdep_pin_lock(&rq->lock);
5523 next = pick_next_task(rq, &fake_task, cookie);
5525 next->sched_class->put_prev_task(rq, next);
5528 * Rules for changing task_struct::cpus_allowed are holding
5529 * both pi_lock and rq->lock, such that holding either
5530 * stabilizes the mask.
5532 * Drop rq->lock is not quite as disastrous as it usually is
5533 * because !cpu_active at this point, which means load-balance
5534 * will not interfere. Also, stop-machine.
5536 lockdep_unpin_lock(&rq->lock, cookie);
5537 raw_spin_unlock(&rq->lock);
5538 raw_spin_lock(&next->pi_lock);
5539 raw_spin_lock(&rq->lock);
5542 * Since we're inside stop-machine, _nothing_ should have
5543 * changed the task, WARN if weird stuff happened, because in
5544 * that case the above rq->lock drop is a fail too.
5546 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5547 raw_spin_unlock(&next->pi_lock);
5551 /* Find suitable destination for @next, with force if needed. */
5552 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5554 rq = __migrate_task(rq, next, dest_cpu);
5555 if (rq != dead_rq) {
5556 raw_spin_unlock(&rq->lock);
5558 raw_spin_lock(&rq->lock);
5560 raw_spin_unlock(&next->pi_lock);
5565 #endif /* CONFIG_HOTPLUG_CPU */
5567 static void set_rq_online(struct rq *rq)
5570 const struct sched_class *class;
5572 cpumask_set_cpu(rq->cpu, rq->rd->online);
5575 for_each_class(class) {
5576 if (class->rq_online)
5577 class->rq_online(rq);
5582 static void set_rq_offline(struct rq *rq)
5585 const struct sched_class *class;
5587 for_each_class(class) {
5588 if (class->rq_offline)
5589 class->rq_offline(rq);
5592 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5597 static void set_cpu_rq_start_time(unsigned int cpu)
5599 struct rq *rq = cpu_rq(cpu);
5601 rq->age_stamp = sched_clock_cpu(cpu);
5604 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5606 #ifdef CONFIG_SCHED_DEBUG
5608 static __read_mostly int sched_debug_enabled;
5610 static int __init sched_debug_setup(char *str)
5612 sched_debug_enabled = 1;
5616 early_param("sched_debug", sched_debug_setup);
5618 static inline bool sched_debug(void)
5620 return sched_debug_enabled;
5623 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5624 struct cpumask *groupmask)
5626 struct sched_group *group = sd->groups;
5628 cpumask_clear(groupmask);
5630 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5632 if (!(sd->flags & SD_LOAD_BALANCE)) {
5633 printk("does not load-balance\n");
5635 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5640 printk(KERN_CONT "span %*pbl level %s\n",
5641 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5643 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5644 printk(KERN_ERR "ERROR: domain->span does not contain "
5647 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5648 printk(KERN_ERR "ERROR: domain->groups does not contain"
5652 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5656 printk(KERN_ERR "ERROR: group is NULL\n");
5660 if (!cpumask_weight(sched_group_cpus(group))) {
5661 printk(KERN_CONT "\n");
5662 printk(KERN_ERR "ERROR: empty group\n");
5666 if (!(sd->flags & SD_OVERLAP) &&
5667 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5668 printk(KERN_CONT "\n");
5669 printk(KERN_ERR "ERROR: repeated CPUs\n");
5673 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5675 printk(KERN_CONT " %*pbl",
5676 cpumask_pr_args(sched_group_cpus(group)));
5677 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5678 printk(KERN_CONT " (cpu_capacity = %d)",
5679 group->sgc->capacity);
5682 group = group->next;
5683 } while (group != sd->groups);
5684 printk(KERN_CONT "\n");
5686 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5687 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5690 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5691 printk(KERN_ERR "ERROR: parent span is not a superset "
5692 "of domain->span\n");
5696 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5700 if (!sched_debug_enabled)
5704 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5708 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5711 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5719 #else /* !CONFIG_SCHED_DEBUG */
5720 # define sched_domain_debug(sd, cpu) do { } while (0)
5721 static inline bool sched_debug(void)
5725 #endif /* CONFIG_SCHED_DEBUG */
5727 static int sd_degenerate(struct sched_domain *sd)
5729 if (cpumask_weight(sched_domain_span(sd)) == 1)
5732 /* Following flags need at least 2 groups */
5733 if (sd->flags & (SD_LOAD_BALANCE |
5734 SD_BALANCE_NEWIDLE |
5737 SD_SHARE_CPUCAPACITY |
5738 SD_ASYM_CPUCAPACITY |
5739 SD_SHARE_PKG_RESOURCES |
5740 SD_SHARE_POWERDOMAIN)) {
5741 if (sd->groups != sd->groups->next)
5745 /* Following flags don't use groups */
5746 if (sd->flags & (SD_WAKE_AFFINE))
5753 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5755 unsigned long cflags = sd->flags, pflags = parent->flags;
5757 if (sd_degenerate(parent))
5760 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5763 /* Flags needing groups don't count if only 1 group in parent */
5764 if (parent->groups == parent->groups->next) {
5765 pflags &= ~(SD_LOAD_BALANCE |
5766 SD_BALANCE_NEWIDLE |
5769 SD_ASYM_CPUCAPACITY |
5770 SD_SHARE_CPUCAPACITY |
5771 SD_SHARE_PKG_RESOURCES |
5773 SD_SHARE_POWERDOMAIN);
5774 if (nr_node_ids == 1)
5775 pflags &= ~SD_SERIALIZE;
5777 if (~cflags & pflags)
5783 static void free_rootdomain(struct rcu_head *rcu)
5785 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5787 cpupri_cleanup(&rd->cpupri);
5788 cpudl_cleanup(&rd->cpudl);
5789 free_cpumask_var(rd->dlo_mask);
5790 free_cpumask_var(rd->rto_mask);
5791 free_cpumask_var(rd->online);
5792 free_cpumask_var(rd->span);
5796 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5798 struct root_domain *old_rd = NULL;
5799 unsigned long flags;
5801 raw_spin_lock_irqsave(&rq->lock, flags);
5806 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5809 cpumask_clear_cpu(rq->cpu, old_rd->span);
5812 * If we dont want to free the old_rd yet then
5813 * set old_rd to NULL to skip the freeing later
5816 if (!atomic_dec_and_test(&old_rd->refcount))
5820 atomic_inc(&rd->refcount);
5823 cpumask_set_cpu(rq->cpu, rd->span);
5824 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5827 raw_spin_unlock_irqrestore(&rq->lock, flags);
5830 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5833 static int init_rootdomain(struct root_domain *rd)
5835 memset(rd, 0, sizeof(*rd));
5837 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5839 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5841 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5843 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5846 init_dl_bw(&rd->dl_bw);
5847 if (cpudl_init(&rd->cpudl) != 0)
5850 if (cpupri_init(&rd->cpupri) != 0)
5855 free_cpumask_var(rd->rto_mask);
5857 free_cpumask_var(rd->dlo_mask);
5859 free_cpumask_var(rd->online);
5861 free_cpumask_var(rd->span);
5867 * By default the system creates a single root-domain with all cpus as
5868 * members (mimicking the global state we have today).
5870 struct root_domain def_root_domain;
5872 static void init_defrootdomain(void)
5874 init_rootdomain(&def_root_domain);
5876 atomic_set(&def_root_domain.refcount, 1);
5879 static struct root_domain *alloc_rootdomain(void)
5881 struct root_domain *rd;
5883 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5887 if (init_rootdomain(rd) != 0) {
5895 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5897 struct sched_group *tmp, *first;
5906 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5911 } while (sg != first);
5914 static void free_sched_domain(struct rcu_head *rcu)
5916 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5919 * If its an overlapping domain it has private groups, iterate and
5922 if (sd->flags & SD_OVERLAP) {
5923 free_sched_groups(sd->groups, 1);
5924 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5925 kfree(sd->groups->sgc);
5931 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5933 call_rcu(&sd->rcu, free_sched_domain);
5936 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5938 for (; sd; sd = sd->parent)
5939 destroy_sched_domain(sd, cpu);
5943 * Keep a special pointer to the highest sched_domain that has
5944 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5945 * allows us to avoid some pointer chasing select_idle_sibling().
5947 * Also keep a unique ID per domain (we use the first cpu number in
5948 * the cpumask of the domain), this allows us to quickly tell if
5949 * two cpus are in the same cache domain, see cpus_share_cache().
5951 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5952 DEFINE_PER_CPU(int, sd_llc_size);
5953 DEFINE_PER_CPU(int, sd_llc_id);
5954 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5955 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5956 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5958 static void update_top_cache_domain(int cpu)
5960 struct sched_domain *sd;
5961 struct sched_domain *busy_sd = NULL;
5965 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5967 id = cpumask_first(sched_domain_span(sd));
5968 size = cpumask_weight(sched_domain_span(sd));
5969 busy_sd = sd->parent; /* sd_busy */
5971 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5973 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5974 per_cpu(sd_llc_size, cpu) = size;
5975 per_cpu(sd_llc_id, cpu) = id;
5977 sd = lowest_flag_domain(cpu, SD_NUMA);
5978 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5980 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5981 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5985 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5986 * hold the hotplug lock.
5989 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5991 struct rq *rq = cpu_rq(cpu);
5992 struct sched_domain *tmp;
5994 /* Remove the sched domains which do not contribute to scheduling. */
5995 for (tmp = sd; tmp; ) {
5996 struct sched_domain *parent = tmp->parent;
6000 if (sd_parent_degenerate(tmp, parent)) {
6001 tmp->parent = parent->parent;
6003 parent->parent->child = tmp;
6005 * Transfer SD_PREFER_SIBLING down in case of a
6006 * degenerate parent; the spans match for this
6007 * so the property transfers.
6009 if (parent->flags & SD_PREFER_SIBLING)
6010 tmp->flags |= SD_PREFER_SIBLING;
6011 destroy_sched_domain(parent, cpu);
6016 if (sd && sd_degenerate(sd)) {
6019 destroy_sched_domain(tmp, cpu);
6024 sched_domain_debug(sd, cpu);
6026 rq_attach_root(rq, rd);
6028 rcu_assign_pointer(rq->sd, sd);
6029 destroy_sched_domains(tmp, cpu);
6031 update_top_cache_domain(cpu);
6034 /* Setup the mask of cpus configured for isolated domains */
6035 static int __init isolated_cpu_setup(char *str)
6039 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6040 ret = cpulist_parse(str, cpu_isolated_map);
6042 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6047 __setup("isolcpus=", isolated_cpu_setup);
6050 struct sched_domain ** __percpu sd;
6051 struct root_domain *rd;
6062 * Build an iteration mask that can exclude certain CPUs from the upwards
6065 * Asymmetric node setups can result in situations where the domain tree is of
6066 * unequal depth, make sure to skip domains that already cover the entire
6069 * In that case build_sched_domains() will have terminated the iteration early
6070 * and our sibling sd spans will be empty. Domains should always include the
6071 * cpu they're built on, so check that.
6074 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6076 const struct cpumask *span = sched_domain_span(sd);
6077 struct sd_data *sdd = sd->private;
6078 struct sched_domain *sibling;
6081 for_each_cpu(i, span) {
6082 sibling = *per_cpu_ptr(sdd->sd, i);
6083 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6086 cpumask_set_cpu(i, sched_group_mask(sg));
6091 * Return the canonical balance cpu for this group, this is the first cpu
6092 * of this group that's also in the iteration mask.
6094 int group_balance_cpu(struct sched_group *sg)
6096 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6100 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6102 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6103 const struct cpumask *span = sched_domain_span(sd);
6104 struct cpumask *covered = sched_domains_tmpmask;
6105 struct sd_data *sdd = sd->private;
6106 struct sched_domain *sibling;
6109 cpumask_clear(covered);
6111 for_each_cpu(i, span) {
6112 struct cpumask *sg_span;
6114 if (cpumask_test_cpu(i, covered))
6117 sibling = *per_cpu_ptr(sdd->sd, i);
6119 /* See the comment near build_group_mask(). */
6120 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6123 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6124 GFP_KERNEL, cpu_to_node(cpu));
6129 sg_span = sched_group_cpus(sg);
6131 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6133 cpumask_set_cpu(i, sg_span);
6135 cpumask_or(covered, covered, sg_span);
6137 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6138 if (atomic_inc_return(&sg->sgc->ref) == 1)
6139 build_group_mask(sd, sg);
6142 * Initialize sgc->capacity such that even if we mess up the
6143 * domains and no possible iteration will get us here, we won't
6146 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6149 * Make sure the first group of this domain contains the
6150 * canonical balance cpu. Otherwise the sched_domain iteration
6151 * breaks. See update_sg_lb_stats().
6153 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6154 group_balance_cpu(sg) == cpu)
6164 sd->groups = groups;
6169 free_sched_groups(first, 0);
6174 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6176 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6177 struct sched_domain *child = sd->child;
6180 cpu = cpumask_first(sched_domain_span(child));
6183 *sg = *per_cpu_ptr(sdd->sg, cpu);
6184 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6185 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6192 * build_sched_groups will build a circular linked list of the groups
6193 * covered by the given span, and will set each group's ->cpumask correctly,
6194 * and ->cpu_capacity to 0.
6196 * Assumes the sched_domain tree is fully constructed
6199 build_sched_groups(struct sched_domain *sd, int cpu)
6201 struct sched_group *first = NULL, *last = NULL;
6202 struct sd_data *sdd = sd->private;
6203 const struct cpumask *span = sched_domain_span(sd);
6204 struct cpumask *covered;
6207 get_group(cpu, sdd, &sd->groups);
6208 atomic_inc(&sd->groups->ref);
6210 if (cpu != cpumask_first(span))
6213 lockdep_assert_held(&sched_domains_mutex);
6214 covered = sched_domains_tmpmask;
6216 cpumask_clear(covered);
6218 for_each_cpu(i, span) {
6219 struct sched_group *sg;
6222 if (cpumask_test_cpu(i, covered))
6225 group = get_group(i, sdd, &sg);
6226 cpumask_setall(sched_group_mask(sg));
6228 for_each_cpu(j, span) {
6229 if (get_group(j, sdd, NULL) != group)
6232 cpumask_set_cpu(j, covered);
6233 cpumask_set_cpu(j, sched_group_cpus(sg));
6248 * Initialize sched groups cpu_capacity.
6250 * cpu_capacity indicates the capacity of sched group, which is used while
6251 * distributing the load between different sched groups in a sched domain.
6252 * Typically cpu_capacity for all the groups in a sched domain will be same
6253 * unless there are asymmetries in the topology. If there are asymmetries,
6254 * group having more cpu_capacity will pickup more load compared to the
6255 * group having less cpu_capacity.
6257 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6259 struct sched_group *sg = sd->groups;
6264 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6266 } while (sg != sd->groups);
6268 if (cpu != group_balance_cpu(sg))
6271 update_group_capacity(sd, cpu);
6272 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6276 * Initializers for schedule domains
6277 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6280 static int default_relax_domain_level = -1;
6281 int sched_domain_level_max;
6283 static int __init setup_relax_domain_level(char *str)
6285 if (kstrtoint(str, 0, &default_relax_domain_level))
6286 pr_warn("Unable to set relax_domain_level\n");
6290 __setup("relax_domain_level=", setup_relax_domain_level);
6292 static void set_domain_attribute(struct sched_domain *sd,
6293 struct sched_domain_attr *attr)
6297 if (!attr || attr->relax_domain_level < 0) {
6298 if (default_relax_domain_level < 0)
6301 request = default_relax_domain_level;
6303 request = attr->relax_domain_level;
6304 if (request < sd->level) {
6305 /* turn off idle balance on this domain */
6306 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6308 /* turn on idle balance on this domain */
6309 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6313 static void __sdt_free(const struct cpumask *cpu_map);
6314 static int __sdt_alloc(const struct cpumask *cpu_map);
6316 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6317 const struct cpumask *cpu_map)
6321 if (!atomic_read(&d->rd->refcount))
6322 free_rootdomain(&d->rd->rcu); /* fall through */
6324 free_percpu(d->sd); /* fall through */
6326 __sdt_free(cpu_map); /* fall through */
6332 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6333 const struct cpumask *cpu_map)
6335 memset(d, 0, sizeof(*d));
6337 if (__sdt_alloc(cpu_map))
6338 return sa_sd_storage;
6339 d->sd = alloc_percpu(struct sched_domain *);
6341 return sa_sd_storage;
6342 d->rd = alloc_rootdomain();
6345 return sa_rootdomain;
6349 * NULL the sd_data elements we've used to build the sched_domain and
6350 * sched_group structure so that the subsequent __free_domain_allocs()
6351 * will not free the data we're using.
6353 static void claim_allocations(int cpu, struct sched_domain *sd)
6355 struct sd_data *sdd = sd->private;
6357 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6358 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6360 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6361 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6363 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6364 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6368 static int sched_domains_numa_levels;
6369 enum numa_topology_type sched_numa_topology_type;
6370 static int *sched_domains_numa_distance;
6371 int sched_max_numa_distance;
6372 static struct cpumask ***sched_domains_numa_masks;
6373 static int sched_domains_curr_level;
6377 * SD_flags allowed in topology descriptions.
6379 * These flags are purely descriptive of the topology and do not prescribe
6380 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6383 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6384 * SD_SHARE_PKG_RESOURCES - describes shared caches
6385 * SD_NUMA - describes NUMA topologies
6386 * SD_SHARE_POWERDOMAIN - describes shared power domain
6387 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6389 * Odd one out, which beside describing the topology has a quirk also
6390 * prescribes the desired behaviour that goes along with it:
6392 * SD_ASYM_PACKING - describes SMT quirks
6394 #define TOPOLOGY_SD_FLAGS \
6395 (SD_SHARE_CPUCAPACITY | \
6396 SD_SHARE_PKG_RESOURCES | \
6399 SD_ASYM_CPUCAPACITY | \
6400 SD_SHARE_POWERDOMAIN)
6402 static struct sched_domain *
6403 sd_init(struct sched_domain_topology_level *tl,
6404 struct sched_domain *child, int cpu)
6406 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6407 int sd_weight, sd_flags = 0;
6411 * Ugly hack to pass state to sd_numa_mask()...
6413 sched_domains_curr_level = tl->numa_level;
6416 sd_weight = cpumask_weight(tl->mask(cpu));
6419 sd_flags = (*tl->sd_flags)();
6420 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6421 "wrong sd_flags in topology description\n"))
6422 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6424 *sd = (struct sched_domain){
6425 .min_interval = sd_weight,
6426 .max_interval = 2*sd_weight,
6428 .imbalance_pct = 125,
6430 .cache_nice_tries = 0,
6437 .flags = 1*SD_LOAD_BALANCE
6438 | 1*SD_BALANCE_NEWIDLE
6443 | 0*SD_SHARE_CPUCAPACITY
6444 | 0*SD_SHARE_PKG_RESOURCES
6446 | 0*SD_PREFER_SIBLING
6451 .last_balance = jiffies,
6452 .balance_interval = sd_weight,
6454 .max_newidle_lb_cost = 0,
6455 .next_decay_max_lb_cost = jiffies,
6457 #ifdef CONFIG_SCHED_DEBUG
6463 * Convert topological properties into behaviour.
6466 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6467 struct sched_domain *t = sd;
6469 for_each_lower_domain(t)
6470 t->flags |= SD_BALANCE_WAKE;
6473 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6474 sd->flags |= SD_PREFER_SIBLING;
6475 sd->imbalance_pct = 110;
6476 sd->smt_gain = 1178; /* ~15% */
6478 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6479 sd->imbalance_pct = 117;
6480 sd->cache_nice_tries = 1;
6484 } else if (sd->flags & SD_NUMA) {
6485 sd->cache_nice_tries = 2;
6489 sd->flags |= SD_SERIALIZE;
6490 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6491 sd->flags &= ~(SD_BALANCE_EXEC |
6498 sd->flags |= SD_PREFER_SIBLING;
6499 sd->cache_nice_tries = 1;
6504 sd->private = &tl->data;
6510 * Topology list, bottom-up.
6512 static struct sched_domain_topology_level default_topology[] = {
6513 #ifdef CONFIG_SCHED_SMT
6514 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6516 #ifdef CONFIG_SCHED_MC
6517 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6519 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6523 static struct sched_domain_topology_level *sched_domain_topology =
6526 #define for_each_sd_topology(tl) \
6527 for (tl = sched_domain_topology; tl->mask; tl++)
6529 void set_sched_topology(struct sched_domain_topology_level *tl)
6531 sched_domain_topology = tl;
6536 static const struct cpumask *sd_numa_mask(int cpu)
6538 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6541 static void sched_numa_warn(const char *str)
6543 static int done = false;
6551 printk(KERN_WARNING "ERROR: %s\n\n", str);
6553 for (i = 0; i < nr_node_ids; i++) {
6554 printk(KERN_WARNING " ");
6555 for (j = 0; j < nr_node_ids; j++)
6556 printk(KERN_CONT "%02d ", node_distance(i,j));
6557 printk(KERN_CONT "\n");
6559 printk(KERN_WARNING "\n");
6562 bool find_numa_distance(int distance)
6566 if (distance == node_distance(0, 0))
6569 for (i = 0; i < sched_domains_numa_levels; i++) {
6570 if (sched_domains_numa_distance[i] == distance)
6578 * A system can have three types of NUMA topology:
6579 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6580 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6581 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6583 * The difference between a glueless mesh topology and a backplane
6584 * topology lies in whether communication between not directly
6585 * connected nodes goes through intermediary nodes (where programs
6586 * could run), or through backplane controllers. This affects
6587 * placement of programs.
6589 * The type of topology can be discerned with the following tests:
6590 * - If the maximum distance between any nodes is 1 hop, the system
6591 * is directly connected.
6592 * - If for two nodes A and B, located N > 1 hops away from each other,
6593 * there is an intermediary node C, which is < N hops away from both
6594 * nodes A and B, the system is a glueless mesh.
6596 static void init_numa_topology_type(void)
6600 n = sched_max_numa_distance;
6602 if (sched_domains_numa_levels <= 1) {
6603 sched_numa_topology_type = NUMA_DIRECT;
6607 for_each_online_node(a) {
6608 for_each_online_node(b) {
6609 /* Find two nodes furthest removed from each other. */
6610 if (node_distance(a, b) < n)
6613 /* Is there an intermediary node between a and b? */
6614 for_each_online_node(c) {
6615 if (node_distance(a, c) < n &&
6616 node_distance(b, c) < n) {
6617 sched_numa_topology_type =
6623 sched_numa_topology_type = NUMA_BACKPLANE;
6629 static void sched_init_numa(void)
6631 int next_distance, curr_distance = node_distance(0, 0);
6632 struct sched_domain_topology_level *tl;
6636 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6637 if (!sched_domains_numa_distance)
6641 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6642 * unique distances in the node_distance() table.
6644 * Assumes node_distance(0,j) includes all distances in
6645 * node_distance(i,j) in order to avoid cubic time.
6647 next_distance = curr_distance;
6648 for (i = 0; i < nr_node_ids; i++) {
6649 for (j = 0; j < nr_node_ids; j++) {
6650 for (k = 0; k < nr_node_ids; k++) {
6651 int distance = node_distance(i, k);
6653 if (distance > curr_distance &&
6654 (distance < next_distance ||
6655 next_distance == curr_distance))
6656 next_distance = distance;
6659 * While not a strong assumption it would be nice to know
6660 * about cases where if node A is connected to B, B is not
6661 * equally connected to A.
6663 if (sched_debug() && node_distance(k, i) != distance)
6664 sched_numa_warn("Node-distance not symmetric");
6666 if (sched_debug() && i && !find_numa_distance(distance))
6667 sched_numa_warn("Node-0 not representative");
6669 if (next_distance != curr_distance) {
6670 sched_domains_numa_distance[level++] = next_distance;
6671 sched_domains_numa_levels = level;
6672 curr_distance = next_distance;
6677 * In case of sched_debug() we verify the above assumption.
6687 * 'level' contains the number of unique distances, excluding the
6688 * identity distance node_distance(i,i).
6690 * The sched_domains_numa_distance[] array includes the actual distance
6695 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6696 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6697 * the array will contain less then 'level' members. This could be
6698 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6699 * in other functions.
6701 * We reset it to 'level' at the end of this function.
6703 sched_domains_numa_levels = 0;
6705 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6706 if (!sched_domains_numa_masks)
6710 * Now for each level, construct a mask per node which contains all
6711 * cpus of nodes that are that many hops away from us.
6713 for (i = 0; i < level; i++) {
6714 sched_domains_numa_masks[i] =
6715 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6716 if (!sched_domains_numa_masks[i])
6719 for (j = 0; j < nr_node_ids; j++) {
6720 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6724 sched_domains_numa_masks[i][j] = mask;
6727 if (node_distance(j, k) > sched_domains_numa_distance[i])
6730 cpumask_or(mask, mask, cpumask_of_node(k));
6735 /* Compute default topology size */
6736 for (i = 0; sched_domain_topology[i].mask; i++);
6738 tl = kzalloc((i + level + 1) *
6739 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6744 * Copy the default topology bits..
6746 for (i = 0; sched_domain_topology[i].mask; i++)
6747 tl[i] = sched_domain_topology[i];
6750 * .. and append 'j' levels of NUMA goodness.
6752 for (j = 0; j < level; i++, j++) {
6753 tl[i] = (struct sched_domain_topology_level){
6754 .mask = sd_numa_mask,
6755 .sd_flags = cpu_numa_flags,
6756 .flags = SDTL_OVERLAP,
6762 sched_domain_topology = tl;
6764 sched_domains_numa_levels = level;
6765 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6767 init_numa_topology_type();
6770 static void sched_domains_numa_masks_set(unsigned int cpu)
6772 int node = cpu_to_node(cpu);
6775 for (i = 0; i < sched_domains_numa_levels; i++) {
6776 for (j = 0; j < nr_node_ids; j++) {
6777 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6778 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6783 static void sched_domains_numa_masks_clear(unsigned int cpu)
6787 for (i = 0; i < sched_domains_numa_levels; i++) {
6788 for (j = 0; j < nr_node_ids; j++)
6789 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6794 static inline void sched_init_numa(void) { }
6795 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6796 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6797 #endif /* CONFIG_NUMA */
6799 static int __sdt_alloc(const struct cpumask *cpu_map)
6801 struct sched_domain_topology_level *tl;
6804 for_each_sd_topology(tl) {
6805 struct sd_data *sdd = &tl->data;
6807 sdd->sd = alloc_percpu(struct sched_domain *);
6811 sdd->sg = alloc_percpu(struct sched_group *);
6815 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6819 for_each_cpu(j, cpu_map) {
6820 struct sched_domain *sd;
6821 struct sched_group *sg;
6822 struct sched_group_capacity *sgc;
6824 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6825 GFP_KERNEL, cpu_to_node(j));
6829 *per_cpu_ptr(sdd->sd, j) = sd;
6831 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6832 GFP_KERNEL, cpu_to_node(j));
6838 *per_cpu_ptr(sdd->sg, j) = sg;
6840 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6841 GFP_KERNEL, cpu_to_node(j));
6845 *per_cpu_ptr(sdd->sgc, j) = sgc;
6852 static void __sdt_free(const struct cpumask *cpu_map)
6854 struct sched_domain_topology_level *tl;
6857 for_each_sd_topology(tl) {
6858 struct sd_data *sdd = &tl->data;
6860 for_each_cpu(j, cpu_map) {
6861 struct sched_domain *sd;
6864 sd = *per_cpu_ptr(sdd->sd, j);
6865 if (sd && (sd->flags & SD_OVERLAP))
6866 free_sched_groups(sd->groups, 0);
6867 kfree(*per_cpu_ptr(sdd->sd, j));
6871 kfree(*per_cpu_ptr(sdd->sg, j));
6873 kfree(*per_cpu_ptr(sdd->sgc, j));
6875 free_percpu(sdd->sd);
6877 free_percpu(sdd->sg);
6879 free_percpu(sdd->sgc);
6884 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6885 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6886 struct sched_domain *child, int cpu)
6888 struct sched_domain *sd = sd_init(tl, child, cpu);
6890 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6892 sd->level = child->level + 1;
6893 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6896 if (!cpumask_subset(sched_domain_span(child),
6897 sched_domain_span(sd))) {
6898 pr_err("BUG: arch topology borken\n");
6899 #ifdef CONFIG_SCHED_DEBUG
6900 pr_err(" the %s domain not a subset of the %s domain\n",
6901 child->name, sd->name);
6903 /* Fixup, ensure @sd has at least @child cpus. */
6904 cpumask_or(sched_domain_span(sd),
6905 sched_domain_span(sd),
6906 sched_domain_span(child));
6910 set_domain_attribute(sd, attr);
6916 * Build sched domains for a given set of cpus and attach the sched domains
6917 * to the individual cpus
6919 static int build_sched_domains(const struct cpumask *cpu_map,
6920 struct sched_domain_attr *attr)
6922 enum s_alloc alloc_state;
6923 struct sched_domain *sd;
6925 struct rq *rq = NULL;
6926 int i, ret = -ENOMEM;
6928 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6929 if (alloc_state != sa_rootdomain)
6932 /* Set up domains for cpus specified by the cpu_map. */
6933 for_each_cpu(i, cpu_map) {
6934 struct sched_domain_topology_level *tl;
6937 for_each_sd_topology(tl) {
6938 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6939 if (tl == sched_domain_topology)
6940 *per_cpu_ptr(d.sd, i) = sd;
6941 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6942 sd->flags |= SD_OVERLAP;
6943 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6948 /* Build the groups for the domains */
6949 for_each_cpu(i, cpu_map) {
6950 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6951 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6952 if (sd->flags & SD_OVERLAP) {
6953 if (build_overlap_sched_groups(sd, i))
6956 if (build_sched_groups(sd, i))
6962 /* Calculate CPU capacity for physical packages and nodes */
6963 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6964 if (!cpumask_test_cpu(i, cpu_map))
6967 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6968 claim_allocations(i, sd);
6969 init_sched_groups_capacity(i, sd);
6973 /* Attach the domains */
6975 for_each_cpu(i, cpu_map) {
6977 sd = *per_cpu_ptr(d.sd, i);
6979 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
6980 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
6981 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
6983 cpu_attach_domain(sd, d.rd, i);
6988 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
6989 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
6994 __free_domain_allocs(&d, alloc_state, cpu_map);
6998 static cpumask_var_t *doms_cur; /* current sched domains */
6999 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7000 static struct sched_domain_attr *dattr_cur;
7001 /* attribues of custom domains in 'doms_cur' */
7004 * Special case: If a kmalloc of a doms_cur partition (array of
7005 * cpumask) fails, then fallback to a single sched domain,
7006 * as determined by the single cpumask fallback_doms.
7008 static cpumask_var_t fallback_doms;
7011 * arch_update_cpu_topology lets virtualized architectures update the
7012 * cpu core maps. It is supposed to return 1 if the topology changed
7013 * or 0 if it stayed the same.
7015 int __weak arch_update_cpu_topology(void)
7020 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7023 cpumask_var_t *doms;
7025 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7028 for (i = 0; i < ndoms; i++) {
7029 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7030 free_sched_domains(doms, i);
7037 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7040 for (i = 0; i < ndoms; i++)
7041 free_cpumask_var(doms[i]);
7046 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7047 * For now this just excludes isolated cpus, but could be used to
7048 * exclude other special cases in the future.
7050 static int init_sched_domains(const struct cpumask *cpu_map)
7054 arch_update_cpu_topology();
7056 doms_cur = alloc_sched_domains(ndoms_cur);
7058 doms_cur = &fallback_doms;
7059 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7060 err = build_sched_domains(doms_cur[0], NULL);
7061 register_sched_domain_sysctl();
7067 * Detach sched domains from a group of cpus specified in cpu_map
7068 * These cpus will now be attached to the NULL domain
7070 static void detach_destroy_domains(const struct cpumask *cpu_map)
7075 for_each_cpu(i, cpu_map)
7076 cpu_attach_domain(NULL, &def_root_domain, i);
7080 /* handle null as "default" */
7081 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7082 struct sched_domain_attr *new, int idx_new)
7084 struct sched_domain_attr tmp;
7091 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7092 new ? (new + idx_new) : &tmp,
7093 sizeof(struct sched_domain_attr));
7097 * Partition sched domains as specified by the 'ndoms_new'
7098 * cpumasks in the array doms_new[] of cpumasks. This compares
7099 * doms_new[] to the current sched domain partitioning, doms_cur[].
7100 * It destroys each deleted domain and builds each new domain.
7102 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7103 * The masks don't intersect (don't overlap.) We should setup one
7104 * sched domain for each mask. CPUs not in any of the cpumasks will
7105 * not be load balanced. If the same cpumask appears both in the
7106 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7109 * The passed in 'doms_new' should be allocated using
7110 * alloc_sched_domains. This routine takes ownership of it and will
7111 * free_sched_domains it when done with it. If the caller failed the
7112 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7113 * and partition_sched_domains() will fallback to the single partition
7114 * 'fallback_doms', it also forces the domains to be rebuilt.
7116 * If doms_new == NULL it will be replaced with cpu_online_mask.
7117 * ndoms_new == 0 is a special case for destroying existing domains,
7118 * and it will not create the default domain.
7120 * Call with hotplug lock held
7122 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7123 struct sched_domain_attr *dattr_new)
7128 mutex_lock(&sched_domains_mutex);
7130 /* always unregister in case we don't destroy any domains */
7131 unregister_sched_domain_sysctl();
7133 /* Let architecture update cpu core mappings. */
7134 new_topology = arch_update_cpu_topology();
7136 n = doms_new ? ndoms_new : 0;
7138 /* Destroy deleted domains */
7139 for (i = 0; i < ndoms_cur; i++) {
7140 for (j = 0; j < n && !new_topology; j++) {
7141 if (cpumask_equal(doms_cur[i], doms_new[j])
7142 && dattrs_equal(dattr_cur, i, dattr_new, j))
7145 /* no match - a current sched domain not in new doms_new[] */
7146 detach_destroy_domains(doms_cur[i]);
7152 if (doms_new == NULL) {
7154 doms_new = &fallback_doms;
7155 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7156 WARN_ON_ONCE(dattr_new);
7159 /* Build new domains */
7160 for (i = 0; i < ndoms_new; i++) {
7161 for (j = 0; j < n && !new_topology; j++) {
7162 if (cpumask_equal(doms_new[i], doms_cur[j])
7163 && dattrs_equal(dattr_new, i, dattr_cur, j))
7166 /* no match - add a new doms_new */
7167 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7172 /* Remember the new sched domains */
7173 if (doms_cur != &fallback_doms)
7174 free_sched_domains(doms_cur, ndoms_cur);
7175 kfree(dattr_cur); /* kfree(NULL) is safe */
7176 doms_cur = doms_new;
7177 dattr_cur = dattr_new;
7178 ndoms_cur = ndoms_new;
7180 register_sched_domain_sysctl();
7182 mutex_unlock(&sched_domains_mutex);
7185 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7188 * Update cpusets according to cpu_active mask. If cpusets are
7189 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7190 * around partition_sched_domains().
7192 * If we come here as part of a suspend/resume, don't touch cpusets because we
7193 * want to restore it back to its original state upon resume anyway.
7195 static void cpuset_cpu_active(void)
7197 if (cpuhp_tasks_frozen) {
7199 * num_cpus_frozen tracks how many CPUs are involved in suspend
7200 * resume sequence. As long as this is not the last online
7201 * operation in the resume sequence, just build a single sched
7202 * domain, ignoring cpusets.
7205 if (likely(num_cpus_frozen)) {
7206 partition_sched_domains(1, NULL, NULL);
7210 * This is the last CPU online operation. So fall through and
7211 * restore the original sched domains by considering the
7212 * cpuset configurations.
7215 cpuset_update_active_cpus(true);
7218 static int cpuset_cpu_inactive(unsigned int cpu)
7220 unsigned long flags;
7225 if (!cpuhp_tasks_frozen) {
7226 rcu_read_lock_sched();
7227 dl_b = dl_bw_of(cpu);
7229 raw_spin_lock_irqsave(&dl_b->lock, flags);
7230 cpus = dl_bw_cpus(cpu);
7231 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7232 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7234 rcu_read_unlock_sched();
7238 cpuset_update_active_cpus(false);
7241 partition_sched_domains(1, NULL, NULL);
7246 int sched_cpu_activate(unsigned int cpu)
7248 struct rq *rq = cpu_rq(cpu);
7249 unsigned long flags;
7251 set_cpu_active(cpu, true);
7253 if (sched_smp_initialized) {
7254 sched_domains_numa_masks_set(cpu);
7255 cpuset_cpu_active();
7259 * Put the rq online, if not already. This happens:
7261 * 1) In the early boot process, because we build the real domains
7262 * after all cpus have been brought up.
7264 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7267 raw_spin_lock_irqsave(&rq->lock, flags);
7269 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7272 raw_spin_unlock_irqrestore(&rq->lock, flags);
7274 update_max_interval();
7279 int sched_cpu_deactivate(unsigned int cpu)
7283 set_cpu_active(cpu, false);
7285 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7286 * users of this state to go away such that all new such users will
7289 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7290 * not imply sync_sched(), so wait for both.
7292 * Do sync before park smpboot threads to take care the rcu boost case.
7294 if (IS_ENABLED(CONFIG_PREEMPT))
7295 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7299 if (!sched_smp_initialized)
7302 ret = cpuset_cpu_inactive(cpu);
7304 set_cpu_active(cpu, true);
7307 sched_domains_numa_masks_clear(cpu);
7311 static void sched_rq_cpu_starting(unsigned int cpu)
7313 struct rq *rq = cpu_rq(cpu);
7315 rq->calc_load_update = calc_load_update;
7316 update_max_interval();
7319 int sched_cpu_starting(unsigned int cpu)
7321 set_cpu_rq_start_time(cpu);
7322 sched_rq_cpu_starting(cpu);
7326 #ifdef CONFIG_HOTPLUG_CPU
7327 int sched_cpu_dying(unsigned int cpu)
7329 struct rq *rq = cpu_rq(cpu);
7330 unsigned long flags;
7332 /* Handle pending wakeups and then migrate everything off */
7333 sched_ttwu_pending();
7334 raw_spin_lock_irqsave(&rq->lock, flags);
7336 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7340 BUG_ON(rq->nr_running != 1);
7341 raw_spin_unlock_irqrestore(&rq->lock, flags);
7342 calc_load_migrate(rq);
7343 update_max_interval();
7344 nohz_balance_exit_idle(cpu);
7350 void __init sched_init_smp(void)
7352 cpumask_var_t non_isolated_cpus;
7354 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7355 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7360 * There's no userspace yet to cause hotplug operations; hence all the
7361 * cpu masks are stable and all blatant races in the below code cannot
7364 mutex_lock(&sched_domains_mutex);
7365 init_sched_domains(cpu_active_mask);
7366 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7367 if (cpumask_empty(non_isolated_cpus))
7368 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7369 mutex_unlock(&sched_domains_mutex);
7371 /* Move init over to a non-isolated CPU */
7372 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7374 sched_init_granularity();
7375 free_cpumask_var(non_isolated_cpus);
7377 init_sched_rt_class();
7378 init_sched_dl_class();
7379 sched_smp_initialized = true;
7382 static int __init migration_init(void)
7384 sched_rq_cpu_starting(smp_processor_id());
7387 early_initcall(migration_init);
7390 void __init sched_init_smp(void)
7392 sched_init_granularity();
7394 #endif /* CONFIG_SMP */
7396 int in_sched_functions(unsigned long addr)
7398 return in_lock_functions(addr) ||
7399 (addr >= (unsigned long)__sched_text_start
7400 && addr < (unsigned long)__sched_text_end);
7403 #ifdef CONFIG_CGROUP_SCHED
7405 * Default task group.
7406 * Every task in system belongs to this group at bootup.
7408 struct task_group root_task_group;
7409 LIST_HEAD(task_groups);
7411 /* Cacheline aligned slab cache for task_group */
7412 static struct kmem_cache *task_group_cache __read_mostly;
7415 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7417 void __init sched_init(void)
7420 unsigned long alloc_size = 0, ptr;
7422 #ifdef CONFIG_FAIR_GROUP_SCHED
7423 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7425 #ifdef CONFIG_RT_GROUP_SCHED
7426 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7429 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7431 #ifdef CONFIG_FAIR_GROUP_SCHED
7432 root_task_group.se = (struct sched_entity **)ptr;
7433 ptr += nr_cpu_ids * sizeof(void **);
7435 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7436 ptr += nr_cpu_ids * sizeof(void **);
7438 #endif /* CONFIG_FAIR_GROUP_SCHED */
7439 #ifdef CONFIG_RT_GROUP_SCHED
7440 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7441 ptr += nr_cpu_ids * sizeof(void **);
7443 root_task_group.rt_rq = (struct rt_rq **)ptr;
7444 ptr += nr_cpu_ids * sizeof(void **);
7446 #endif /* CONFIG_RT_GROUP_SCHED */
7448 #ifdef CONFIG_CPUMASK_OFFSTACK
7449 for_each_possible_cpu(i) {
7450 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7451 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7453 #endif /* CONFIG_CPUMASK_OFFSTACK */
7455 init_rt_bandwidth(&def_rt_bandwidth,
7456 global_rt_period(), global_rt_runtime());
7457 init_dl_bandwidth(&def_dl_bandwidth,
7458 global_rt_period(), global_rt_runtime());
7461 init_defrootdomain();
7464 #ifdef CONFIG_RT_GROUP_SCHED
7465 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7466 global_rt_period(), global_rt_runtime());
7467 #endif /* CONFIG_RT_GROUP_SCHED */
7469 #ifdef CONFIG_CGROUP_SCHED
7470 task_group_cache = KMEM_CACHE(task_group, 0);
7472 list_add(&root_task_group.list, &task_groups);
7473 INIT_LIST_HEAD(&root_task_group.children);
7474 INIT_LIST_HEAD(&root_task_group.siblings);
7475 autogroup_init(&init_task);
7476 #endif /* CONFIG_CGROUP_SCHED */
7478 for_each_possible_cpu(i) {
7482 raw_spin_lock_init(&rq->lock);
7484 rq->calc_load_active = 0;
7485 rq->calc_load_update = jiffies + LOAD_FREQ;
7486 init_cfs_rq(&rq->cfs);
7487 init_rt_rq(&rq->rt);
7488 init_dl_rq(&rq->dl);
7489 #ifdef CONFIG_FAIR_GROUP_SCHED
7490 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7491 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7493 * How much cpu bandwidth does root_task_group get?
7495 * In case of task-groups formed thr' the cgroup filesystem, it
7496 * gets 100% of the cpu resources in the system. This overall
7497 * system cpu resource is divided among the tasks of
7498 * root_task_group and its child task-groups in a fair manner,
7499 * based on each entity's (task or task-group's) weight
7500 * (se->load.weight).
7502 * In other words, if root_task_group has 10 tasks of weight
7503 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7504 * then A0's share of the cpu resource is:
7506 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7508 * We achieve this by letting root_task_group's tasks sit
7509 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7511 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7512 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7513 #endif /* CONFIG_FAIR_GROUP_SCHED */
7515 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7516 #ifdef CONFIG_RT_GROUP_SCHED
7517 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7520 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7521 rq->cpu_load[j] = 0;
7526 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7527 rq->balance_callback = NULL;
7528 rq->active_balance = 0;
7529 rq->next_balance = jiffies;
7534 rq->avg_idle = 2*sysctl_sched_migration_cost;
7535 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7537 INIT_LIST_HEAD(&rq->cfs_tasks);
7539 rq_attach_root(rq, &def_root_domain);
7540 #ifdef CONFIG_NO_HZ_COMMON
7541 rq->last_load_update_tick = jiffies;
7544 #ifdef CONFIG_NO_HZ_FULL
7545 rq->last_sched_tick = 0;
7547 #endif /* CONFIG_SMP */
7549 atomic_set(&rq->nr_iowait, 0);
7552 set_load_weight(&init_task);
7555 * The boot idle thread does lazy MMU switching as well:
7557 atomic_inc(&init_mm.mm_count);
7558 enter_lazy_tlb(&init_mm, current);
7561 * During early bootup we pretend to be a normal task:
7563 current->sched_class = &fair_sched_class;
7566 * Make us the idle thread. Technically, schedule() should not be
7567 * called from this thread, however somewhere below it might be,
7568 * but because we are the idle thread, we just pick up running again
7569 * when this runqueue becomes "idle".
7571 init_idle(current, smp_processor_id());
7573 calc_load_update = jiffies + LOAD_FREQ;
7576 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7577 /* May be allocated at isolcpus cmdline parse time */
7578 if (cpu_isolated_map == NULL)
7579 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7580 idle_thread_set_boot_cpu();
7581 set_cpu_rq_start_time(smp_processor_id());
7583 init_sched_fair_class();
7587 scheduler_running = 1;
7590 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7591 static inline int preempt_count_equals(int preempt_offset)
7593 int nested = preempt_count() + rcu_preempt_depth();
7595 return (nested == preempt_offset);
7598 void __might_sleep(const char *file, int line, int preempt_offset)
7601 * Blocking primitives will set (and therefore destroy) current->state,
7602 * since we will exit with TASK_RUNNING make sure we enter with it,
7603 * otherwise we will destroy state.
7605 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7606 "do not call blocking ops when !TASK_RUNNING; "
7607 "state=%lx set at [<%p>] %pS\n",
7609 (void *)current->task_state_change,
7610 (void *)current->task_state_change);
7612 ___might_sleep(file, line, preempt_offset);
7614 EXPORT_SYMBOL(__might_sleep);
7616 void ___might_sleep(const char *file, int line, int preempt_offset)
7618 static unsigned long prev_jiffy; /* ratelimiting */
7619 unsigned long preempt_disable_ip;
7621 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7622 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7623 !is_idle_task(current)) ||
7624 system_state != SYSTEM_RUNNING || oops_in_progress)
7626 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7628 prev_jiffy = jiffies;
7630 /* Save this before calling printk(), since that will clobber it */
7631 preempt_disable_ip = get_preempt_disable_ip(current);
7634 "BUG: sleeping function called from invalid context at %s:%d\n",
7637 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7638 in_atomic(), irqs_disabled(),
7639 current->pid, current->comm);
7641 if (task_stack_end_corrupted(current))
7642 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7644 debug_show_held_locks(current);
7645 if (irqs_disabled())
7646 print_irqtrace_events(current);
7647 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7648 && !preempt_count_equals(preempt_offset)) {
7649 pr_err("Preemption disabled at:");
7650 print_ip_sym(preempt_disable_ip);
7654 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7656 EXPORT_SYMBOL(___might_sleep);
7659 #ifdef CONFIG_MAGIC_SYSRQ
7660 void normalize_rt_tasks(void)
7662 struct task_struct *g, *p;
7663 struct sched_attr attr = {
7664 .sched_policy = SCHED_NORMAL,
7667 read_lock(&tasklist_lock);
7668 for_each_process_thread(g, p) {
7670 * Only normalize user tasks:
7672 if (p->flags & PF_KTHREAD)
7675 p->se.exec_start = 0;
7676 schedstat_set(p->se.statistics.wait_start, 0);
7677 schedstat_set(p->se.statistics.sleep_start, 0);
7678 schedstat_set(p->se.statistics.block_start, 0);
7680 if (!dl_task(p) && !rt_task(p)) {
7682 * Renice negative nice level userspace
7685 if (task_nice(p) < 0)
7686 set_user_nice(p, 0);
7690 __sched_setscheduler(p, &attr, false, false);
7692 read_unlock(&tasklist_lock);
7695 #endif /* CONFIG_MAGIC_SYSRQ */
7697 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7699 * These functions are only useful for the IA64 MCA handling, or kdb.
7701 * They can only be called when the whole system has been
7702 * stopped - every CPU needs to be quiescent, and no scheduling
7703 * activity can take place. Using them for anything else would
7704 * be a serious bug, and as a result, they aren't even visible
7705 * under any other configuration.
7709 * curr_task - return the current task for a given cpu.
7710 * @cpu: the processor in question.
7712 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7714 * Return: The current task for @cpu.
7716 struct task_struct *curr_task(int cpu)
7718 return cpu_curr(cpu);
7721 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7725 * set_curr_task - set the current task for a given cpu.
7726 * @cpu: the processor in question.
7727 * @p: the task pointer to set.
7729 * Description: This function must only be used when non-maskable interrupts
7730 * are serviced on a separate stack. It allows the architecture to switch the
7731 * notion of the current task on a cpu in a non-blocking manner. This function
7732 * must be called with all CPU's synchronized, and interrupts disabled, the
7733 * and caller must save the original value of the current task (see
7734 * curr_task() above) and restore that value before reenabling interrupts and
7735 * re-starting the system.
7737 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7739 void set_curr_task(int cpu, struct task_struct *p)
7746 #ifdef CONFIG_CGROUP_SCHED
7747 /* task_group_lock serializes the addition/removal of task groups */
7748 static DEFINE_SPINLOCK(task_group_lock);
7750 static void sched_free_group(struct task_group *tg)
7752 free_fair_sched_group(tg);
7753 free_rt_sched_group(tg);
7755 kmem_cache_free(task_group_cache, tg);
7758 /* allocate runqueue etc for a new task group */
7759 struct task_group *sched_create_group(struct task_group *parent)
7761 struct task_group *tg;
7763 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7765 return ERR_PTR(-ENOMEM);
7767 if (!alloc_fair_sched_group(tg, parent))
7770 if (!alloc_rt_sched_group(tg, parent))
7776 sched_free_group(tg);
7777 return ERR_PTR(-ENOMEM);
7780 void sched_online_group(struct task_group *tg, struct task_group *parent)
7782 unsigned long flags;
7784 spin_lock_irqsave(&task_group_lock, flags);
7785 list_add_rcu(&tg->list, &task_groups);
7787 WARN_ON(!parent); /* root should already exist */
7789 tg->parent = parent;
7790 INIT_LIST_HEAD(&tg->children);
7791 list_add_rcu(&tg->siblings, &parent->children);
7792 spin_unlock_irqrestore(&task_group_lock, flags);
7794 online_fair_sched_group(tg);
7797 /* rcu callback to free various structures associated with a task group */
7798 static void sched_free_group_rcu(struct rcu_head *rhp)
7800 /* now it should be safe to free those cfs_rqs */
7801 sched_free_group(container_of(rhp, struct task_group, rcu));
7804 void sched_destroy_group(struct task_group *tg)
7806 /* wait for possible concurrent references to cfs_rqs complete */
7807 call_rcu(&tg->rcu, sched_free_group_rcu);
7810 void sched_offline_group(struct task_group *tg)
7812 unsigned long flags;
7814 /* end participation in shares distribution */
7815 unregister_fair_sched_group(tg);
7817 spin_lock_irqsave(&task_group_lock, flags);
7818 list_del_rcu(&tg->list);
7819 list_del_rcu(&tg->siblings);
7820 spin_unlock_irqrestore(&task_group_lock, flags);
7823 static void sched_change_group(struct task_struct *tsk, int type)
7825 struct task_group *tg;
7828 * All callers are synchronized by task_rq_lock(); we do not use RCU
7829 * which is pointless here. Thus, we pass "true" to task_css_check()
7830 * to prevent lockdep warnings.
7832 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7833 struct task_group, css);
7834 tg = autogroup_task_group(tsk, tg);
7835 tsk->sched_task_group = tg;
7837 #ifdef CONFIG_FAIR_GROUP_SCHED
7838 if (tsk->sched_class->task_change_group)
7839 tsk->sched_class->task_change_group(tsk, type);
7842 set_task_rq(tsk, task_cpu(tsk));
7846 * Change task's runqueue when it moves between groups.
7848 * The caller of this function should have put the task in its new group by
7849 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7852 void sched_move_task(struct task_struct *tsk)
7854 int queued, running;
7858 rq = task_rq_lock(tsk, &rf);
7860 running = task_current(rq, tsk);
7861 queued = task_on_rq_queued(tsk);
7864 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7865 if (unlikely(running))
7866 put_prev_task(rq, tsk);
7868 sched_change_group(tsk, TASK_MOVE_GROUP);
7870 if (unlikely(running))
7871 tsk->sched_class->set_curr_task(rq);
7873 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7875 task_rq_unlock(rq, tsk, &rf);
7877 #endif /* CONFIG_CGROUP_SCHED */
7879 #ifdef CONFIG_RT_GROUP_SCHED
7881 * Ensure that the real time constraints are schedulable.
7883 static DEFINE_MUTEX(rt_constraints_mutex);
7885 /* Must be called with tasklist_lock held */
7886 static inline int tg_has_rt_tasks(struct task_group *tg)
7888 struct task_struct *g, *p;
7891 * Autogroups do not have RT tasks; see autogroup_create().
7893 if (task_group_is_autogroup(tg))
7896 for_each_process_thread(g, p) {
7897 if (rt_task(p) && task_group(p) == tg)
7904 struct rt_schedulable_data {
7905 struct task_group *tg;
7910 static int tg_rt_schedulable(struct task_group *tg, void *data)
7912 struct rt_schedulable_data *d = data;
7913 struct task_group *child;
7914 unsigned long total, sum = 0;
7915 u64 period, runtime;
7917 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7918 runtime = tg->rt_bandwidth.rt_runtime;
7921 period = d->rt_period;
7922 runtime = d->rt_runtime;
7926 * Cannot have more runtime than the period.
7928 if (runtime > period && runtime != RUNTIME_INF)
7932 * Ensure we don't starve existing RT tasks.
7934 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7937 total = to_ratio(period, runtime);
7940 * Nobody can have more than the global setting allows.
7942 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7946 * The sum of our children's runtime should not exceed our own.
7948 list_for_each_entry_rcu(child, &tg->children, siblings) {
7949 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7950 runtime = child->rt_bandwidth.rt_runtime;
7952 if (child == d->tg) {
7953 period = d->rt_period;
7954 runtime = d->rt_runtime;
7957 sum += to_ratio(period, runtime);
7966 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7970 struct rt_schedulable_data data = {
7972 .rt_period = period,
7973 .rt_runtime = runtime,
7977 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7983 static int tg_set_rt_bandwidth(struct task_group *tg,
7984 u64 rt_period, u64 rt_runtime)
7989 * Disallowing the root group RT runtime is BAD, it would disallow the
7990 * kernel creating (and or operating) RT threads.
7992 if (tg == &root_task_group && rt_runtime == 0)
7995 /* No period doesn't make any sense. */
7999 mutex_lock(&rt_constraints_mutex);
8000 read_lock(&tasklist_lock);
8001 err = __rt_schedulable(tg, rt_period, rt_runtime);
8005 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8006 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8007 tg->rt_bandwidth.rt_runtime = rt_runtime;
8009 for_each_possible_cpu(i) {
8010 struct rt_rq *rt_rq = tg->rt_rq[i];
8012 raw_spin_lock(&rt_rq->rt_runtime_lock);
8013 rt_rq->rt_runtime = rt_runtime;
8014 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8016 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8018 read_unlock(&tasklist_lock);
8019 mutex_unlock(&rt_constraints_mutex);
8024 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8026 u64 rt_runtime, rt_period;
8028 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8029 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8030 if (rt_runtime_us < 0)
8031 rt_runtime = RUNTIME_INF;
8033 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8036 static long sched_group_rt_runtime(struct task_group *tg)
8040 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8043 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8044 do_div(rt_runtime_us, NSEC_PER_USEC);
8045 return rt_runtime_us;
8048 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8050 u64 rt_runtime, rt_period;
8052 rt_period = rt_period_us * NSEC_PER_USEC;
8053 rt_runtime = tg->rt_bandwidth.rt_runtime;
8055 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8058 static long sched_group_rt_period(struct task_group *tg)
8062 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8063 do_div(rt_period_us, NSEC_PER_USEC);
8064 return rt_period_us;
8066 #endif /* CONFIG_RT_GROUP_SCHED */
8068 #ifdef CONFIG_RT_GROUP_SCHED
8069 static int sched_rt_global_constraints(void)
8073 mutex_lock(&rt_constraints_mutex);
8074 read_lock(&tasklist_lock);
8075 ret = __rt_schedulable(NULL, 0, 0);
8076 read_unlock(&tasklist_lock);
8077 mutex_unlock(&rt_constraints_mutex);
8082 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8084 /* Don't accept realtime tasks when there is no way for them to run */
8085 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8091 #else /* !CONFIG_RT_GROUP_SCHED */
8092 static int sched_rt_global_constraints(void)
8094 unsigned long flags;
8097 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8098 for_each_possible_cpu(i) {
8099 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8101 raw_spin_lock(&rt_rq->rt_runtime_lock);
8102 rt_rq->rt_runtime = global_rt_runtime();
8103 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8105 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8109 #endif /* CONFIG_RT_GROUP_SCHED */
8111 static int sched_dl_global_validate(void)
8113 u64 runtime = global_rt_runtime();
8114 u64 period = global_rt_period();
8115 u64 new_bw = to_ratio(period, runtime);
8118 unsigned long flags;
8121 * Here we want to check the bandwidth not being set to some
8122 * value smaller than the currently allocated bandwidth in
8123 * any of the root_domains.
8125 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8126 * cycling on root_domains... Discussion on different/better
8127 * solutions is welcome!
8129 for_each_possible_cpu(cpu) {
8130 rcu_read_lock_sched();
8131 dl_b = dl_bw_of(cpu);
8133 raw_spin_lock_irqsave(&dl_b->lock, flags);
8134 if (new_bw < dl_b->total_bw)
8136 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8138 rcu_read_unlock_sched();
8147 static void sched_dl_do_global(void)
8152 unsigned long flags;
8154 def_dl_bandwidth.dl_period = global_rt_period();
8155 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8157 if (global_rt_runtime() != RUNTIME_INF)
8158 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8161 * FIXME: As above...
8163 for_each_possible_cpu(cpu) {
8164 rcu_read_lock_sched();
8165 dl_b = dl_bw_of(cpu);
8167 raw_spin_lock_irqsave(&dl_b->lock, flags);
8169 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8171 rcu_read_unlock_sched();
8175 static int sched_rt_global_validate(void)
8177 if (sysctl_sched_rt_period <= 0)
8180 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8181 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8187 static void sched_rt_do_global(void)
8189 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8190 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8193 int sched_rt_handler(struct ctl_table *table, int write,
8194 void __user *buffer, size_t *lenp,
8197 int old_period, old_runtime;
8198 static DEFINE_MUTEX(mutex);
8202 old_period = sysctl_sched_rt_period;
8203 old_runtime = sysctl_sched_rt_runtime;
8205 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8207 if (!ret && write) {
8208 ret = sched_rt_global_validate();
8212 ret = sched_dl_global_validate();
8216 ret = sched_rt_global_constraints();
8220 sched_rt_do_global();
8221 sched_dl_do_global();
8225 sysctl_sched_rt_period = old_period;
8226 sysctl_sched_rt_runtime = old_runtime;
8228 mutex_unlock(&mutex);
8233 int sched_rr_handler(struct ctl_table *table, int write,
8234 void __user *buffer, size_t *lenp,
8238 static DEFINE_MUTEX(mutex);
8241 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8242 /* make sure that internally we keep jiffies */
8243 /* also, writing zero resets timeslice to default */
8244 if (!ret && write) {
8245 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8246 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8248 mutex_unlock(&mutex);
8252 #ifdef CONFIG_CGROUP_SCHED
8254 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8256 return css ? container_of(css, struct task_group, css) : NULL;
8259 static struct cgroup_subsys_state *
8260 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8262 struct task_group *parent = css_tg(parent_css);
8263 struct task_group *tg;
8266 /* This is early initialization for the top cgroup */
8267 return &root_task_group.css;
8270 tg = sched_create_group(parent);
8272 return ERR_PTR(-ENOMEM);
8274 sched_online_group(tg, parent);
8279 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8281 struct task_group *tg = css_tg(css);
8283 sched_offline_group(tg);
8286 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8288 struct task_group *tg = css_tg(css);
8291 * Relies on the RCU grace period between css_released() and this.
8293 sched_free_group(tg);
8297 * This is called before wake_up_new_task(), therefore we really only
8298 * have to set its group bits, all the other stuff does not apply.
8300 static void cpu_cgroup_fork(struct task_struct *task)
8305 rq = task_rq_lock(task, &rf);
8307 sched_change_group(task, TASK_SET_GROUP);
8309 task_rq_unlock(rq, task, &rf);
8312 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8314 struct task_struct *task;
8315 struct cgroup_subsys_state *css;
8318 cgroup_taskset_for_each(task, css, tset) {
8319 #ifdef CONFIG_RT_GROUP_SCHED
8320 if (!sched_rt_can_attach(css_tg(css), task))
8323 /* We don't support RT-tasks being in separate groups */
8324 if (task->sched_class != &fair_sched_class)
8328 * Serialize against wake_up_new_task() such that if its
8329 * running, we're sure to observe its full state.
8331 raw_spin_lock_irq(&task->pi_lock);
8333 * Avoid calling sched_move_task() before wake_up_new_task()
8334 * has happened. This would lead to problems with PELT, due to
8335 * move wanting to detach+attach while we're not attached yet.
8337 if (task->state == TASK_NEW)
8339 raw_spin_unlock_irq(&task->pi_lock);
8347 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8349 struct task_struct *task;
8350 struct cgroup_subsys_state *css;
8352 cgroup_taskset_for_each(task, css, tset)
8353 sched_move_task(task);
8356 #ifdef CONFIG_FAIR_GROUP_SCHED
8357 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8358 struct cftype *cftype, u64 shareval)
8360 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8363 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8366 struct task_group *tg = css_tg(css);
8368 return (u64) scale_load_down(tg->shares);
8371 #ifdef CONFIG_CFS_BANDWIDTH
8372 static DEFINE_MUTEX(cfs_constraints_mutex);
8374 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8375 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8377 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8379 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8381 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8382 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8384 if (tg == &root_task_group)
8388 * Ensure we have at some amount of bandwidth every period. This is
8389 * to prevent reaching a state of large arrears when throttled via
8390 * entity_tick() resulting in prolonged exit starvation.
8392 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8396 * Likewise, bound things on the otherside by preventing insane quota
8397 * periods. This also allows us to normalize in computing quota
8400 if (period > max_cfs_quota_period)
8404 * Prevent race between setting of cfs_rq->runtime_enabled and
8405 * unthrottle_offline_cfs_rqs().
8408 mutex_lock(&cfs_constraints_mutex);
8409 ret = __cfs_schedulable(tg, period, quota);
8413 runtime_enabled = quota != RUNTIME_INF;
8414 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8416 * If we need to toggle cfs_bandwidth_used, off->on must occur
8417 * before making related changes, and on->off must occur afterwards
8419 if (runtime_enabled && !runtime_was_enabled)
8420 cfs_bandwidth_usage_inc();
8421 raw_spin_lock_irq(&cfs_b->lock);
8422 cfs_b->period = ns_to_ktime(period);
8423 cfs_b->quota = quota;
8425 __refill_cfs_bandwidth_runtime(cfs_b);
8426 /* restart the period timer (if active) to handle new period expiry */
8427 if (runtime_enabled)
8428 start_cfs_bandwidth(cfs_b);
8429 raw_spin_unlock_irq(&cfs_b->lock);
8431 for_each_online_cpu(i) {
8432 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8433 struct rq *rq = cfs_rq->rq;
8435 raw_spin_lock_irq(&rq->lock);
8436 cfs_rq->runtime_enabled = runtime_enabled;
8437 cfs_rq->runtime_remaining = 0;
8439 if (cfs_rq->throttled)
8440 unthrottle_cfs_rq(cfs_rq);
8441 raw_spin_unlock_irq(&rq->lock);
8443 if (runtime_was_enabled && !runtime_enabled)
8444 cfs_bandwidth_usage_dec();
8446 mutex_unlock(&cfs_constraints_mutex);
8452 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8456 period = ktime_to_ns(tg->cfs_bandwidth.period);
8457 if (cfs_quota_us < 0)
8458 quota = RUNTIME_INF;
8460 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8462 return tg_set_cfs_bandwidth(tg, period, quota);
8465 long tg_get_cfs_quota(struct task_group *tg)
8469 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8472 quota_us = tg->cfs_bandwidth.quota;
8473 do_div(quota_us, NSEC_PER_USEC);
8478 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8482 period = (u64)cfs_period_us * NSEC_PER_USEC;
8483 quota = tg->cfs_bandwidth.quota;
8485 return tg_set_cfs_bandwidth(tg, period, quota);
8488 long tg_get_cfs_period(struct task_group *tg)
8492 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8493 do_div(cfs_period_us, NSEC_PER_USEC);
8495 return cfs_period_us;
8498 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8501 return tg_get_cfs_quota(css_tg(css));
8504 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8505 struct cftype *cftype, s64 cfs_quota_us)
8507 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8510 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8513 return tg_get_cfs_period(css_tg(css));
8516 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8517 struct cftype *cftype, u64 cfs_period_us)
8519 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8522 struct cfs_schedulable_data {
8523 struct task_group *tg;
8528 * normalize group quota/period to be quota/max_period
8529 * note: units are usecs
8531 static u64 normalize_cfs_quota(struct task_group *tg,
8532 struct cfs_schedulable_data *d)
8540 period = tg_get_cfs_period(tg);
8541 quota = tg_get_cfs_quota(tg);
8544 /* note: these should typically be equivalent */
8545 if (quota == RUNTIME_INF || quota == -1)
8548 return to_ratio(period, quota);
8551 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8553 struct cfs_schedulable_data *d = data;
8554 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8555 s64 quota = 0, parent_quota = -1;
8558 quota = RUNTIME_INF;
8560 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8562 quota = normalize_cfs_quota(tg, d);
8563 parent_quota = parent_b->hierarchical_quota;
8566 * ensure max(child_quota) <= parent_quota, inherit when no
8569 if (quota == RUNTIME_INF)
8570 quota = parent_quota;
8571 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8574 cfs_b->hierarchical_quota = quota;
8579 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8582 struct cfs_schedulable_data data = {
8588 if (quota != RUNTIME_INF) {
8589 do_div(data.period, NSEC_PER_USEC);
8590 do_div(data.quota, NSEC_PER_USEC);
8594 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8600 static int cpu_stats_show(struct seq_file *sf, void *v)
8602 struct task_group *tg = css_tg(seq_css(sf));
8603 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8605 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8606 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8607 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8611 #endif /* CONFIG_CFS_BANDWIDTH */
8612 #endif /* CONFIG_FAIR_GROUP_SCHED */
8614 #ifdef CONFIG_RT_GROUP_SCHED
8615 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8616 struct cftype *cft, s64 val)
8618 return sched_group_set_rt_runtime(css_tg(css), val);
8621 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8624 return sched_group_rt_runtime(css_tg(css));
8627 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8628 struct cftype *cftype, u64 rt_period_us)
8630 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8633 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8636 return sched_group_rt_period(css_tg(css));
8638 #endif /* CONFIG_RT_GROUP_SCHED */
8640 static struct cftype cpu_files[] = {
8641 #ifdef CONFIG_FAIR_GROUP_SCHED
8644 .read_u64 = cpu_shares_read_u64,
8645 .write_u64 = cpu_shares_write_u64,
8648 #ifdef CONFIG_CFS_BANDWIDTH
8650 .name = "cfs_quota_us",
8651 .read_s64 = cpu_cfs_quota_read_s64,
8652 .write_s64 = cpu_cfs_quota_write_s64,
8655 .name = "cfs_period_us",
8656 .read_u64 = cpu_cfs_period_read_u64,
8657 .write_u64 = cpu_cfs_period_write_u64,
8661 .seq_show = cpu_stats_show,
8664 #ifdef CONFIG_RT_GROUP_SCHED
8666 .name = "rt_runtime_us",
8667 .read_s64 = cpu_rt_runtime_read,
8668 .write_s64 = cpu_rt_runtime_write,
8671 .name = "rt_period_us",
8672 .read_u64 = cpu_rt_period_read_uint,
8673 .write_u64 = cpu_rt_period_write_uint,
8679 struct cgroup_subsys cpu_cgrp_subsys = {
8680 .css_alloc = cpu_cgroup_css_alloc,
8681 .css_released = cpu_cgroup_css_released,
8682 .css_free = cpu_cgroup_css_free,
8683 .fork = cpu_cgroup_fork,
8684 .can_attach = cpu_cgroup_can_attach,
8685 .attach = cpu_cgroup_attach,
8686 .legacy_cftypes = cpu_files,
8690 #endif /* CONFIG_CGROUP_SCHED */
8692 void dump_cpu_task(int cpu)
8694 pr_info("Task dump for CPU %d:\n", cpu);
8695 sched_show_task(cpu_curr(cpu));
8699 * Nice levels are multiplicative, with a gentle 10% change for every
8700 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8701 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8702 * that remained on nice 0.
8704 * The "10% effect" is relative and cumulative: from _any_ nice level,
8705 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8706 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8707 * If a task goes up by ~10% and another task goes down by ~10% then
8708 * the relative distance between them is ~25%.)
8710 const int sched_prio_to_weight[40] = {
8711 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8712 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8713 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8714 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8715 /* 0 */ 1024, 820, 655, 526, 423,
8716 /* 5 */ 335, 272, 215, 172, 137,
8717 /* 10 */ 110, 87, 70, 56, 45,
8718 /* 15 */ 36, 29, 23, 18, 15,
8722 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8724 * In cases where the weight does not change often, we can use the
8725 * precalculated inverse to speed up arithmetics by turning divisions
8726 * into multiplications:
8728 const u32 sched_prio_to_wmult[40] = {
8729 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8730 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8731 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8732 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8733 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8734 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8735 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8736 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,