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 targer 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)
1632 #ifdef CONFIG_SCHEDSTATS
1633 struct rq *rq = this_rq();
1636 int this_cpu = smp_processor_id();
1638 if (cpu == this_cpu) {
1639 schedstat_inc(rq, ttwu_local);
1640 schedstat_inc(p, se.statistics.nr_wakeups_local);
1642 struct sched_domain *sd;
1644 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1646 for_each_domain(this_cpu, sd) {
1647 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1648 schedstat_inc(sd, ttwu_wake_remote);
1655 if (wake_flags & WF_MIGRATED)
1656 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1658 #endif /* CONFIG_SMP */
1660 schedstat_inc(rq, ttwu_count);
1661 schedstat_inc(p, se.statistics.nr_wakeups);
1663 if (wake_flags & WF_SYNC)
1664 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1666 #endif /* CONFIG_SCHEDSTATS */
1669 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1671 activate_task(rq, p, en_flags);
1672 p->on_rq = TASK_ON_RQ_QUEUED;
1674 /* if a worker is waking up, notify workqueue */
1675 if (p->flags & PF_WQ_WORKER)
1676 wq_worker_waking_up(p, cpu_of(rq));
1680 * Mark the task runnable and perform wakeup-preemption.
1682 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1683 struct pin_cookie cookie)
1685 check_preempt_curr(rq, p, wake_flags);
1686 p->state = TASK_RUNNING;
1687 trace_sched_wakeup(p);
1690 if (p->sched_class->task_woken) {
1692 * Our task @p is fully woken up and running; so its safe to
1693 * drop the rq->lock, hereafter rq is only used for statistics.
1695 lockdep_unpin_lock(&rq->lock, cookie);
1696 p->sched_class->task_woken(rq, p);
1697 lockdep_repin_lock(&rq->lock, cookie);
1700 if (rq->idle_stamp) {
1701 u64 delta = rq_clock(rq) - rq->idle_stamp;
1702 u64 max = 2*rq->max_idle_balance_cost;
1704 update_avg(&rq->avg_idle, delta);
1706 if (rq->avg_idle > max)
1715 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1716 struct pin_cookie cookie)
1718 int en_flags = ENQUEUE_WAKEUP;
1720 lockdep_assert_held(&rq->lock);
1723 if (p->sched_contributes_to_load)
1724 rq->nr_uninterruptible--;
1726 if (wake_flags & WF_MIGRATED)
1727 en_flags |= ENQUEUE_MIGRATED;
1730 ttwu_activate(rq, p, en_flags);
1731 ttwu_do_wakeup(rq, p, wake_flags, cookie);
1735 * Called in case the task @p isn't fully descheduled from its runqueue,
1736 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1737 * since all we need to do is flip p->state to TASK_RUNNING, since
1738 * the task is still ->on_rq.
1740 static int ttwu_remote(struct task_struct *p, int wake_flags)
1746 rq = __task_rq_lock(p, &rf);
1747 if (task_on_rq_queued(p)) {
1748 /* check_preempt_curr() may use rq clock */
1749 update_rq_clock(rq);
1750 ttwu_do_wakeup(rq, p, wake_flags, rf.cookie);
1753 __task_rq_unlock(rq, &rf);
1759 void sched_ttwu_pending(void)
1761 struct rq *rq = this_rq();
1762 struct llist_node *llist = llist_del_all(&rq->wake_list);
1763 struct pin_cookie cookie;
1764 struct task_struct *p;
1765 unsigned long flags;
1770 raw_spin_lock_irqsave(&rq->lock, flags);
1771 cookie = lockdep_pin_lock(&rq->lock);
1776 p = llist_entry(llist, struct task_struct, wake_entry);
1777 llist = llist_next(llist);
1779 if (p->sched_remote_wakeup)
1780 wake_flags = WF_MIGRATED;
1782 ttwu_do_activate(rq, p, wake_flags, cookie);
1785 lockdep_unpin_lock(&rq->lock, cookie);
1786 raw_spin_unlock_irqrestore(&rq->lock, flags);
1789 void scheduler_ipi(void)
1792 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1793 * TIF_NEED_RESCHED remotely (for the first time) will also send
1796 preempt_fold_need_resched();
1798 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1802 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1803 * traditionally all their work was done from the interrupt return
1804 * path. Now that we actually do some work, we need to make sure
1807 * Some archs already do call them, luckily irq_enter/exit nest
1810 * Arguably we should visit all archs and update all handlers,
1811 * however a fair share of IPIs are still resched only so this would
1812 * somewhat pessimize the simple resched case.
1815 sched_ttwu_pending();
1818 * Check if someone kicked us for doing the nohz idle load balance.
1820 if (unlikely(got_nohz_idle_kick())) {
1821 this_rq()->idle_balance = 1;
1822 raise_softirq_irqoff(SCHED_SOFTIRQ);
1827 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1829 struct rq *rq = cpu_rq(cpu);
1831 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1833 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1834 if (!set_nr_if_polling(rq->idle))
1835 smp_send_reschedule(cpu);
1837 trace_sched_wake_idle_without_ipi(cpu);
1841 void wake_up_if_idle(int cpu)
1843 struct rq *rq = cpu_rq(cpu);
1844 unsigned long flags;
1848 if (!is_idle_task(rcu_dereference(rq->curr)))
1851 if (set_nr_if_polling(rq->idle)) {
1852 trace_sched_wake_idle_without_ipi(cpu);
1854 raw_spin_lock_irqsave(&rq->lock, flags);
1855 if (is_idle_task(rq->curr))
1856 smp_send_reschedule(cpu);
1857 /* Else cpu is not in idle, do nothing here */
1858 raw_spin_unlock_irqrestore(&rq->lock, flags);
1865 bool cpus_share_cache(int this_cpu, int that_cpu)
1867 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1869 #endif /* CONFIG_SMP */
1871 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1873 struct rq *rq = cpu_rq(cpu);
1874 struct pin_cookie cookie;
1876 #if defined(CONFIG_SMP)
1877 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1878 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1879 ttwu_queue_remote(p, cpu, wake_flags);
1884 raw_spin_lock(&rq->lock);
1885 cookie = lockdep_pin_lock(&rq->lock);
1886 ttwu_do_activate(rq, p, wake_flags, cookie);
1887 lockdep_unpin_lock(&rq->lock, cookie);
1888 raw_spin_unlock(&rq->lock);
1892 * Notes on Program-Order guarantees on SMP systems.
1896 * The basic program-order guarantee on SMP systems is that when a task [t]
1897 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1898 * execution on its new cpu [c1].
1900 * For migration (of runnable tasks) this is provided by the following means:
1902 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1903 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1904 * rq(c1)->lock (if not at the same time, then in that order).
1905 * C) LOCK of the rq(c1)->lock scheduling in task
1907 * Transitivity guarantees that B happens after A and C after B.
1908 * Note: we only require RCpc transitivity.
1909 * Note: the cpu doing B need not be c0 or c1
1918 * UNLOCK rq(0)->lock
1920 * LOCK rq(0)->lock // orders against CPU0
1922 * UNLOCK rq(0)->lock
1926 * UNLOCK rq(1)->lock
1928 * LOCK rq(1)->lock // orders against CPU2
1931 * UNLOCK rq(1)->lock
1934 * BLOCKING -- aka. SLEEP + WAKEUP
1936 * For blocking we (obviously) need to provide the same guarantee as for
1937 * migration. However the means are completely different as there is no lock
1938 * chain to provide order. Instead we do:
1940 * 1) smp_store_release(X->on_cpu, 0)
1941 * 2) smp_cond_load_acquire(!X->on_cpu)
1945 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1947 * LOCK rq(0)->lock LOCK X->pi_lock
1950 * smp_store_release(X->on_cpu, 0);
1952 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1958 * X->state = RUNNING
1959 * UNLOCK rq(2)->lock
1961 * LOCK rq(2)->lock // orders against CPU1
1964 * UNLOCK rq(2)->lock
1967 * UNLOCK rq(0)->lock
1970 * However; for wakeups there is a second guarantee we must provide, namely we
1971 * must observe the state that lead to our wakeup. That is, not only must our
1972 * task observe its own prior state, it must also observe the stores prior to
1975 * This means that any means of doing remote wakeups must order the CPU doing
1976 * the wakeup against the CPU the task is going to end up running on. This,
1977 * however, is already required for the regular Program-Order guarantee above,
1978 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1983 * try_to_wake_up - wake up a thread
1984 * @p: the thread to be awakened
1985 * @state: the mask of task states that can be woken
1986 * @wake_flags: wake modifier flags (WF_*)
1988 * Put it on the run-queue if it's not already there. The "current"
1989 * thread is always on the run-queue (except when the actual
1990 * re-schedule is in progress), and as such you're allowed to do
1991 * the simpler "current->state = TASK_RUNNING" to mark yourself
1992 * runnable without the overhead of this.
1994 * Return: %true if @p was woken up, %false if it was already running.
1995 * or @state didn't match @p's state.
1998 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2000 unsigned long flags;
2001 int cpu, success = 0;
2004 * If we are going to wake up a thread waiting for CONDITION we
2005 * need to ensure that CONDITION=1 done by the caller can not be
2006 * reordered with p->state check below. This pairs with mb() in
2007 * set_current_state() the waiting thread does.
2009 smp_mb__before_spinlock();
2010 raw_spin_lock_irqsave(&p->pi_lock, flags);
2011 if (!(p->state & state))
2014 trace_sched_waking(p);
2016 success = 1; /* we're going to change ->state */
2019 if (p->on_rq && ttwu_remote(p, wake_flags))
2024 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2025 * possible to, falsely, observe p->on_cpu == 0.
2027 * One must be running (->on_cpu == 1) in order to remove oneself
2028 * from the runqueue.
2030 * [S] ->on_cpu = 1; [L] ->on_rq
2034 * [S] ->on_rq = 0; [L] ->on_cpu
2036 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2037 * from the consecutive calls to schedule(); the first switching to our
2038 * task, the second putting it to sleep.
2043 * If the owning (remote) cpu is still in the middle of schedule() with
2044 * this task as prev, wait until its done referencing the task.
2046 * Pairs with the smp_store_release() in finish_lock_switch().
2048 * This ensures that tasks getting woken will be fully ordered against
2049 * their previous state and preserve Program Order.
2051 smp_cond_load_acquire(&p->on_cpu, !VAL);
2053 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2054 p->state = TASK_WAKING;
2056 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2057 if (task_cpu(p) != cpu) {
2058 wake_flags |= WF_MIGRATED;
2059 set_task_cpu(p, cpu);
2061 #endif /* CONFIG_SMP */
2063 ttwu_queue(p, cpu, wake_flags);
2065 if (schedstat_enabled())
2066 ttwu_stat(p, cpu, wake_flags);
2068 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2074 * try_to_wake_up_local - try to wake up a local task with rq lock held
2075 * @p: the thread to be awakened
2077 * Put @p on the run-queue if it's not already there. The caller must
2078 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2081 static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie)
2083 struct rq *rq = task_rq(p);
2085 if (WARN_ON_ONCE(rq != this_rq()) ||
2086 WARN_ON_ONCE(p == current))
2089 lockdep_assert_held(&rq->lock);
2091 if (!raw_spin_trylock(&p->pi_lock)) {
2093 * This is OK, because current is on_cpu, which avoids it being
2094 * picked for load-balance and preemption/IRQs are still
2095 * disabled avoiding further scheduler activity on it and we've
2096 * not yet picked a replacement task.
2098 lockdep_unpin_lock(&rq->lock, cookie);
2099 raw_spin_unlock(&rq->lock);
2100 raw_spin_lock(&p->pi_lock);
2101 raw_spin_lock(&rq->lock);
2102 lockdep_repin_lock(&rq->lock, cookie);
2105 if (!(p->state & TASK_NORMAL))
2108 trace_sched_waking(p);
2110 if (!task_on_rq_queued(p))
2111 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2113 ttwu_do_wakeup(rq, p, 0, cookie);
2114 if (schedstat_enabled())
2115 ttwu_stat(p, smp_processor_id(), 0);
2117 raw_spin_unlock(&p->pi_lock);
2121 * wake_up_process - Wake up a specific process
2122 * @p: The process to be woken up.
2124 * Attempt to wake up the nominated process and move it to the set of runnable
2127 * Return: 1 if the process was woken up, 0 if it was already running.
2129 * It may be assumed that this function implies a write memory barrier before
2130 * changing the task state if and only if any tasks are woken up.
2132 int wake_up_process(struct task_struct *p)
2134 return try_to_wake_up(p, TASK_NORMAL, 0);
2136 EXPORT_SYMBOL(wake_up_process);
2138 int wake_up_state(struct task_struct *p, unsigned int state)
2140 return try_to_wake_up(p, state, 0);
2144 * This function clears the sched_dl_entity static params.
2146 void __dl_clear_params(struct task_struct *p)
2148 struct sched_dl_entity *dl_se = &p->dl;
2150 dl_se->dl_runtime = 0;
2151 dl_se->dl_deadline = 0;
2152 dl_se->dl_period = 0;
2156 dl_se->dl_throttled = 0;
2157 dl_se->dl_yielded = 0;
2161 * Perform scheduler related setup for a newly forked process p.
2162 * p is forked by current.
2164 * __sched_fork() is basic setup used by init_idle() too:
2166 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2171 p->se.exec_start = 0;
2172 p->se.sum_exec_runtime = 0;
2173 p->se.prev_sum_exec_runtime = 0;
2174 p->se.nr_migrations = 0;
2176 INIT_LIST_HEAD(&p->se.group_node);
2178 #ifdef CONFIG_FAIR_GROUP_SCHED
2179 p->se.cfs_rq = NULL;
2182 #ifdef CONFIG_SCHEDSTATS
2183 /* Even if schedstat is disabled, there should not be garbage */
2184 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2187 RB_CLEAR_NODE(&p->dl.rb_node);
2188 init_dl_task_timer(&p->dl);
2189 __dl_clear_params(p);
2191 INIT_LIST_HEAD(&p->rt.run_list);
2193 p->rt.time_slice = sched_rr_timeslice;
2197 #ifdef CONFIG_PREEMPT_NOTIFIERS
2198 INIT_HLIST_HEAD(&p->preempt_notifiers);
2201 #ifdef CONFIG_NUMA_BALANCING
2202 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2203 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2204 p->mm->numa_scan_seq = 0;
2207 if (clone_flags & CLONE_VM)
2208 p->numa_preferred_nid = current->numa_preferred_nid;
2210 p->numa_preferred_nid = -1;
2212 p->node_stamp = 0ULL;
2213 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2214 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2215 p->numa_work.next = &p->numa_work;
2216 p->numa_faults = NULL;
2217 p->last_task_numa_placement = 0;
2218 p->last_sum_exec_runtime = 0;
2220 p->numa_group = NULL;
2221 #endif /* CONFIG_NUMA_BALANCING */
2224 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2226 #ifdef CONFIG_NUMA_BALANCING
2228 void set_numabalancing_state(bool enabled)
2231 static_branch_enable(&sched_numa_balancing);
2233 static_branch_disable(&sched_numa_balancing);
2236 #ifdef CONFIG_PROC_SYSCTL
2237 int sysctl_numa_balancing(struct ctl_table *table, int write,
2238 void __user *buffer, size_t *lenp, loff_t *ppos)
2242 int state = static_branch_likely(&sched_numa_balancing);
2244 if (write && !capable(CAP_SYS_ADMIN))
2249 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2253 set_numabalancing_state(state);
2259 #ifdef CONFIG_SCHEDSTATS
2261 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2262 static bool __initdata __sched_schedstats = false;
2264 static void set_schedstats(bool enabled)
2267 static_branch_enable(&sched_schedstats);
2269 static_branch_disable(&sched_schedstats);
2272 void force_schedstat_enabled(void)
2274 if (!schedstat_enabled()) {
2275 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2276 static_branch_enable(&sched_schedstats);
2280 static int __init setup_schedstats(char *str)
2287 * This code is called before jump labels have been set up, so we can't
2288 * change the static branch directly just yet. Instead set a temporary
2289 * variable so init_schedstats() can do it later.
2291 if (!strcmp(str, "enable")) {
2292 __sched_schedstats = true;
2294 } else if (!strcmp(str, "disable")) {
2295 __sched_schedstats = false;
2300 pr_warn("Unable to parse schedstats=\n");
2304 __setup("schedstats=", setup_schedstats);
2306 static void __init init_schedstats(void)
2308 set_schedstats(__sched_schedstats);
2311 #ifdef CONFIG_PROC_SYSCTL
2312 int sysctl_schedstats(struct ctl_table *table, int write,
2313 void __user *buffer, size_t *lenp, loff_t *ppos)
2317 int state = static_branch_likely(&sched_schedstats);
2319 if (write && !capable(CAP_SYS_ADMIN))
2324 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2328 set_schedstats(state);
2331 #endif /* CONFIG_PROC_SYSCTL */
2332 #else /* !CONFIG_SCHEDSTATS */
2333 static inline void init_schedstats(void) {}
2334 #endif /* CONFIG_SCHEDSTATS */
2337 * fork()/clone()-time setup:
2339 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2341 unsigned long flags;
2342 int cpu = get_cpu();
2344 __sched_fork(clone_flags, p);
2346 * We mark the process as NEW here. This guarantees that
2347 * nobody will actually run it, and a signal or other external
2348 * event cannot wake it up and insert it on the runqueue either.
2350 p->state = TASK_NEW;
2353 * Make sure we do not leak PI boosting priority to the child.
2355 p->prio = current->normal_prio;
2358 * Revert to default priority/policy on fork if requested.
2360 if (unlikely(p->sched_reset_on_fork)) {
2361 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2362 p->policy = SCHED_NORMAL;
2363 p->static_prio = NICE_TO_PRIO(0);
2365 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2366 p->static_prio = NICE_TO_PRIO(0);
2368 p->prio = p->normal_prio = __normal_prio(p);
2372 * We don't need the reset flag anymore after the fork. It has
2373 * fulfilled its duty:
2375 p->sched_reset_on_fork = 0;
2378 if (dl_prio(p->prio)) {
2381 } else if (rt_prio(p->prio)) {
2382 p->sched_class = &rt_sched_class;
2384 p->sched_class = &fair_sched_class;
2387 init_entity_runnable_average(&p->se);
2390 * The child is not yet in the pid-hash so no cgroup attach races,
2391 * and the cgroup is pinned to this child due to cgroup_fork()
2392 * is ran before sched_fork().
2394 * Silence PROVE_RCU.
2396 raw_spin_lock_irqsave(&p->pi_lock, flags);
2398 * We're setting the cpu for the first time, we don't migrate,
2399 * so use __set_task_cpu().
2401 __set_task_cpu(p, cpu);
2402 if (p->sched_class->task_fork)
2403 p->sched_class->task_fork(p);
2404 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2406 #ifdef CONFIG_SCHED_INFO
2407 if (likely(sched_info_on()))
2408 memset(&p->sched_info, 0, sizeof(p->sched_info));
2410 #if defined(CONFIG_SMP)
2413 init_task_preempt_count(p);
2415 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2416 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2423 unsigned long to_ratio(u64 period, u64 runtime)
2425 if (runtime == RUNTIME_INF)
2429 * Doing this here saves a lot of checks in all
2430 * the calling paths, and returning zero seems
2431 * safe for them anyway.
2436 return div64_u64(runtime << 20, period);
2440 inline struct dl_bw *dl_bw_of(int i)
2442 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2443 "sched RCU must be held");
2444 return &cpu_rq(i)->rd->dl_bw;
2447 static inline int dl_bw_cpus(int i)
2449 struct root_domain *rd = cpu_rq(i)->rd;
2452 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2453 "sched RCU must be held");
2454 for_each_cpu_and(i, rd->span, cpu_active_mask)
2460 inline struct dl_bw *dl_bw_of(int i)
2462 return &cpu_rq(i)->dl.dl_bw;
2465 static inline int dl_bw_cpus(int i)
2472 * We must be sure that accepting a new task (or allowing changing the
2473 * parameters of an existing one) is consistent with the bandwidth
2474 * constraints. If yes, this function also accordingly updates the currently
2475 * allocated bandwidth to reflect the new situation.
2477 * This function is called while holding p's rq->lock.
2479 * XXX we should delay bw change until the task's 0-lag point, see
2482 static int dl_overflow(struct task_struct *p, int policy,
2483 const struct sched_attr *attr)
2486 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2487 u64 period = attr->sched_period ?: attr->sched_deadline;
2488 u64 runtime = attr->sched_runtime;
2489 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2492 /* !deadline task may carry old deadline bandwidth */
2493 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2497 * Either if a task, enters, leave, or stays -deadline but changes
2498 * its parameters, we may need to update accordingly the total
2499 * allocated bandwidth of the container.
2501 raw_spin_lock(&dl_b->lock);
2502 cpus = dl_bw_cpus(task_cpu(p));
2503 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2504 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2505 __dl_add(dl_b, new_bw);
2507 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2508 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2509 __dl_clear(dl_b, p->dl.dl_bw);
2510 __dl_add(dl_b, new_bw);
2512 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2513 __dl_clear(dl_b, p->dl.dl_bw);
2516 raw_spin_unlock(&dl_b->lock);
2521 extern void init_dl_bw(struct dl_bw *dl_b);
2524 * wake_up_new_task - wake up a newly created task for the first time.
2526 * This function will do some initial scheduler statistics housekeeping
2527 * that must be done for every newly created context, then puts the task
2528 * on the runqueue and wakes it.
2530 void wake_up_new_task(struct task_struct *p)
2535 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2536 p->state = TASK_RUNNING;
2539 * Fork balancing, do it here and not earlier because:
2540 * - cpus_allowed can change in the fork path
2541 * - any previously selected cpu might disappear through hotplug
2543 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2544 * as we're not fully set-up yet.
2546 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2548 rq = __task_rq_lock(p, &rf);
2549 post_init_entity_util_avg(&p->se);
2551 activate_task(rq, p, 0);
2552 p->on_rq = TASK_ON_RQ_QUEUED;
2553 trace_sched_wakeup_new(p);
2554 check_preempt_curr(rq, p, WF_FORK);
2556 if (p->sched_class->task_woken) {
2558 * Nothing relies on rq->lock after this, so its fine to
2561 lockdep_unpin_lock(&rq->lock, rf.cookie);
2562 p->sched_class->task_woken(rq, p);
2563 lockdep_repin_lock(&rq->lock, rf.cookie);
2566 task_rq_unlock(rq, p, &rf);
2569 #ifdef CONFIG_PREEMPT_NOTIFIERS
2571 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2573 void preempt_notifier_inc(void)
2575 static_key_slow_inc(&preempt_notifier_key);
2577 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2579 void preempt_notifier_dec(void)
2581 static_key_slow_dec(&preempt_notifier_key);
2583 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2586 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2587 * @notifier: notifier struct to register
2589 void preempt_notifier_register(struct preempt_notifier *notifier)
2591 if (!static_key_false(&preempt_notifier_key))
2592 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2594 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2596 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2599 * preempt_notifier_unregister - no longer interested in preemption notifications
2600 * @notifier: notifier struct to unregister
2602 * This is *not* safe to call from within a preemption notifier.
2604 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2606 hlist_del(¬ifier->link);
2608 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2610 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2612 struct preempt_notifier *notifier;
2614 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2615 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2618 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2620 if (static_key_false(&preempt_notifier_key))
2621 __fire_sched_in_preempt_notifiers(curr);
2625 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2626 struct task_struct *next)
2628 struct preempt_notifier *notifier;
2630 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2631 notifier->ops->sched_out(notifier, next);
2634 static __always_inline void
2635 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2636 struct task_struct *next)
2638 if (static_key_false(&preempt_notifier_key))
2639 __fire_sched_out_preempt_notifiers(curr, next);
2642 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2644 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2649 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2650 struct task_struct *next)
2654 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2657 * prepare_task_switch - prepare to switch tasks
2658 * @rq: the runqueue preparing to switch
2659 * @prev: the current task that is being switched out
2660 * @next: the task we are going to switch to.
2662 * This is called with the rq lock held and interrupts off. It must
2663 * be paired with a subsequent finish_task_switch after the context
2666 * prepare_task_switch sets up locking and calls architecture specific
2670 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2671 struct task_struct *next)
2673 sched_info_switch(rq, prev, next);
2674 perf_event_task_sched_out(prev, next);
2675 fire_sched_out_preempt_notifiers(prev, next);
2676 prepare_lock_switch(rq, next);
2677 prepare_arch_switch(next);
2681 * finish_task_switch - clean up after a task-switch
2682 * @prev: the thread we just switched away from.
2684 * finish_task_switch must be called after the context switch, paired
2685 * with a prepare_task_switch call before the context switch.
2686 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2687 * and do any other architecture-specific cleanup actions.
2689 * Note that we may have delayed dropping an mm in context_switch(). If
2690 * so, we finish that here outside of the runqueue lock. (Doing it
2691 * with the lock held can cause deadlocks; see schedule() for
2694 * The context switch have flipped the stack from under us and restored the
2695 * local variables which were saved when this task called schedule() in the
2696 * past. prev == current is still correct but we need to recalculate this_rq
2697 * because prev may have moved to another CPU.
2699 static struct rq *finish_task_switch(struct task_struct *prev)
2700 __releases(rq->lock)
2702 struct rq *rq = this_rq();
2703 struct mm_struct *mm = rq->prev_mm;
2707 * The previous task will have left us with a preempt_count of 2
2708 * because it left us after:
2711 * preempt_disable(); // 1
2713 * raw_spin_lock_irq(&rq->lock) // 2
2715 * Also, see FORK_PREEMPT_COUNT.
2717 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2718 "corrupted preempt_count: %s/%d/0x%x\n",
2719 current->comm, current->pid, preempt_count()))
2720 preempt_count_set(FORK_PREEMPT_COUNT);
2725 * A task struct has one reference for the use as "current".
2726 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2727 * schedule one last time. The schedule call will never return, and
2728 * the scheduled task must drop that reference.
2730 * We must observe prev->state before clearing prev->on_cpu (in
2731 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2732 * running on another CPU and we could rave with its RUNNING -> DEAD
2733 * transition, resulting in a double drop.
2735 prev_state = prev->state;
2736 vtime_task_switch(prev);
2737 perf_event_task_sched_in(prev, current);
2738 finish_lock_switch(rq, prev);
2739 finish_arch_post_lock_switch();
2741 fire_sched_in_preempt_notifiers(current);
2744 if (unlikely(prev_state == TASK_DEAD)) {
2745 if (prev->sched_class->task_dead)
2746 prev->sched_class->task_dead(prev);
2749 * Remove function-return probe instances associated with this
2750 * task and put them back on the free list.
2752 kprobe_flush_task(prev);
2753 put_task_struct(prev);
2756 tick_nohz_task_switch();
2762 /* rq->lock is NOT held, but preemption is disabled */
2763 static void __balance_callback(struct rq *rq)
2765 struct callback_head *head, *next;
2766 void (*func)(struct rq *rq);
2767 unsigned long flags;
2769 raw_spin_lock_irqsave(&rq->lock, flags);
2770 head = rq->balance_callback;
2771 rq->balance_callback = NULL;
2773 func = (void (*)(struct rq *))head->func;
2780 raw_spin_unlock_irqrestore(&rq->lock, flags);
2783 static inline void balance_callback(struct rq *rq)
2785 if (unlikely(rq->balance_callback))
2786 __balance_callback(rq);
2791 static inline void balance_callback(struct rq *rq)
2798 * schedule_tail - first thing a freshly forked thread must call.
2799 * @prev: the thread we just switched away from.
2801 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2802 __releases(rq->lock)
2807 * New tasks start with FORK_PREEMPT_COUNT, see there and
2808 * finish_task_switch() for details.
2810 * finish_task_switch() will drop rq->lock() and lower preempt_count
2811 * and the preempt_enable() will end up enabling preemption (on
2812 * PREEMPT_COUNT kernels).
2815 rq = finish_task_switch(prev);
2816 balance_callback(rq);
2819 if (current->set_child_tid)
2820 put_user(task_pid_vnr(current), current->set_child_tid);
2824 * context_switch - switch to the new MM and the new thread's register state.
2826 static __always_inline struct rq *
2827 context_switch(struct rq *rq, struct task_struct *prev,
2828 struct task_struct *next, struct pin_cookie cookie)
2830 struct mm_struct *mm, *oldmm;
2832 prepare_task_switch(rq, prev, next);
2835 oldmm = prev->active_mm;
2837 * For paravirt, this is coupled with an exit in switch_to to
2838 * combine the page table reload and the switch backend into
2841 arch_start_context_switch(prev);
2844 next->active_mm = oldmm;
2845 atomic_inc(&oldmm->mm_count);
2846 enter_lazy_tlb(oldmm, next);
2848 switch_mm_irqs_off(oldmm, mm, next);
2851 prev->active_mm = NULL;
2852 rq->prev_mm = oldmm;
2855 * Since the runqueue lock will be released by the next
2856 * task (which is an invalid locking op but in the case
2857 * of the scheduler it's an obvious special-case), so we
2858 * do an early lockdep release here:
2860 lockdep_unpin_lock(&rq->lock, cookie);
2861 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2863 /* Here we just switch the register state and the stack. */
2864 switch_to(prev, next, prev);
2867 return finish_task_switch(prev);
2871 * nr_running and nr_context_switches:
2873 * externally visible scheduler statistics: current number of runnable
2874 * threads, total number of context switches performed since bootup.
2876 unsigned long nr_running(void)
2878 unsigned long i, sum = 0;
2880 for_each_online_cpu(i)
2881 sum += cpu_rq(i)->nr_running;
2887 * Check if only the current task is running on the cpu.
2889 * Caution: this function does not check that the caller has disabled
2890 * preemption, thus the result might have a time-of-check-to-time-of-use
2891 * race. The caller is responsible to use it correctly, for example:
2893 * - from a non-preemptable section (of course)
2895 * - from a thread that is bound to a single CPU
2897 * - in a loop with very short iterations (e.g. a polling loop)
2899 bool single_task_running(void)
2901 return raw_rq()->nr_running == 1;
2903 EXPORT_SYMBOL(single_task_running);
2905 unsigned long long nr_context_switches(void)
2908 unsigned long long sum = 0;
2910 for_each_possible_cpu(i)
2911 sum += cpu_rq(i)->nr_switches;
2916 unsigned long nr_iowait(void)
2918 unsigned long i, sum = 0;
2920 for_each_possible_cpu(i)
2921 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2926 unsigned long nr_iowait_cpu(int cpu)
2928 struct rq *this = cpu_rq(cpu);
2929 return atomic_read(&this->nr_iowait);
2932 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2934 struct rq *rq = this_rq();
2935 *nr_waiters = atomic_read(&rq->nr_iowait);
2936 *load = rq->load.weight;
2942 * sched_exec - execve() is a valuable balancing opportunity, because at
2943 * this point the task has the smallest effective memory and cache footprint.
2945 void sched_exec(void)
2947 struct task_struct *p = current;
2948 unsigned long flags;
2951 raw_spin_lock_irqsave(&p->pi_lock, flags);
2952 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2953 if (dest_cpu == smp_processor_id())
2956 if (likely(cpu_active(dest_cpu))) {
2957 struct migration_arg arg = { p, dest_cpu };
2959 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2960 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2964 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2969 DEFINE_PER_CPU(struct kernel_stat, kstat);
2970 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2972 EXPORT_PER_CPU_SYMBOL(kstat);
2973 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2976 * The function fair_sched_class.update_curr accesses the struct curr
2977 * and its field curr->exec_start; when called from task_sched_runtime(),
2978 * we observe a high rate of cache misses in practice.
2979 * Prefetching this data results in improved performance.
2981 static inline void prefetch_curr_exec_start(struct task_struct *p)
2983 #ifdef CONFIG_FAIR_GROUP_SCHED
2984 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
2986 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
2989 prefetch(&curr->exec_start);
2993 * Return accounted runtime for the task.
2994 * In case the task is currently running, return the runtime plus current's
2995 * pending runtime that have not been accounted yet.
2997 unsigned long long task_sched_runtime(struct task_struct *p)
3003 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3005 * 64-bit doesn't need locks to atomically read a 64bit value.
3006 * So we have a optimization chance when the task's delta_exec is 0.
3007 * Reading ->on_cpu is racy, but this is ok.
3009 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3010 * If we race with it entering cpu, unaccounted time is 0. This is
3011 * indistinguishable from the read occurring a few cycles earlier.
3012 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3013 * been accounted, so we're correct here as well.
3015 if (!p->on_cpu || !task_on_rq_queued(p))
3016 return p->se.sum_exec_runtime;
3019 rq = task_rq_lock(p, &rf);
3021 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3022 * project cycles that may never be accounted to this
3023 * thread, breaking clock_gettime().
3025 if (task_current(rq, p) && task_on_rq_queued(p)) {
3026 prefetch_curr_exec_start(p);
3027 update_rq_clock(rq);
3028 p->sched_class->update_curr(rq);
3030 ns = p->se.sum_exec_runtime;
3031 task_rq_unlock(rq, p, &rf);
3037 * This function gets called by the timer code, with HZ frequency.
3038 * We call it with interrupts disabled.
3040 void scheduler_tick(void)
3042 int cpu = smp_processor_id();
3043 struct rq *rq = cpu_rq(cpu);
3044 struct task_struct *curr = rq->curr;
3048 raw_spin_lock(&rq->lock);
3049 update_rq_clock(rq);
3050 curr->sched_class->task_tick(rq, curr, 0);
3051 cpu_load_update_active(rq);
3052 calc_global_load_tick(rq);
3053 raw_spin_unlock(&rq->lock);
3055 perf_event_task_tick();
3058 rq->idle_balance = idle_cpu(cpu);
3059 trigger_load_balance(rq);
3061 rq_last_tick_reset(rq);
3064 #ifdef CONFIG_NO_HZ_FULL
3066 * scheduler_tick_max_deferment
3068 * Keep at least one tick per second when a single
3069 * active task is running because the scheduler doesn't
3070 * yet completely support full dynticks environment.
3072 * This makes sure that uptime, CFS vruntime, load
3073 * balancing, etc... continue to move forward, even
3074 * with a very low granularity.
3076 * Return: Maximum deferment in nanoseconds.
3078 u64 scheduler_tick_max_deferment(void)
3080 struct rq *rq = this_rq();
3081 unsigned long next, now = READ_ONCE(jiffies);
3083 next = rq->last_sched_tick + HZ;
3085 if (time_before_eq(next, now))
3088 return jiffies_to_nsecs(next - now);
3092 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3093 defined(CONFIG_PREEMPT_TRACER))
3095 * If the value passed in is equal to the current preempt count
3096 * then we just disabled preemption. Start timing the latency.
3098 static inline void preempt_latency_start(int val)
3100 if (preempt_count() == val) {
3101 unsigned long ip = get_lock_parent_ip();
3102 #ifdef CONFIG_DEBUG_PREEMPT
3103 current->preempt_disable_ip = ip;
3105 trace_preempt_off(CALLER_ADDR0, ip);
3109 void preempt_count_add(int val)
3111 #ifdef CONFIG_DEBUG_PREEMPT
3115 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3118 __preempt_count_add(val);
3119 #ifdef CONFIG_DEBUG_PREEMPT
3121 * Spinlock count overflowing soon?
3123 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3126 preempt_latency_start(val);
3128 EXPORT_SYMBOL(preempt_count_add);
3129 NOKPROBE_SYMBOL(preempt_count_add);
3132 * If the value passed in equals to the current preempt count
3133 * then we just enabled preemption. Stop timing the latency.
3135 static inline void preempt_latency_stop(int val)
3137 if (preempt_count() == val)
3138 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3141 void preempt_count_sub(int val)
3143 #ifdef CONFIG_DEBUG_PREEMPT
3147 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3150 * Is the spinlock portion underflowing?
3152 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3153 !(preempt_count() & PREEMPT_MASK)))
3157 preempt_latency_stop(val);
3158 __preempt_count_sub(val);
3160 EXPORT_SYMBOL(preempt_count_sub);
3161 NOKPROBE_SYMBOL(preempt_count_sub);
3164 static inline void preempt_latency_start(int val) { }
3165 static inline void preempt_latency_stop(int val) { }
3169 * Print scheduling while atomic bug:
3171 static noinline void __schedule_bug(struct task_struct *prev)
3173 if (oops_in_progress)
3176 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3177 prev->comm, prev->pid, preempt_count());
3179 debug_show_held_locks(prev);
3181 if (irqs_disabled())
3182 print_irqtrace_events(prev);
3183 #ifdef CONFIG_DEBUG_PREEMPT
3184 if (in_atomic_preempt_off()) {
3185 pr_err("Preemption disabled at:");
3186 print_ip_sym(current->preempt_disable_ip);
3191 panic("scheduling while atomic\n");
3194 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3198 * Various schedule()-time debugging checks and statistics:
3200 static inline void schedule_debug(struct task_struct *prev)
3202 #ifdef CONFIG_SCHED_STACK_END_CHECK
3203 if (task_stack_end_corrupted(prev))
3204 panic("corrupted stack end detected inside scheduler\n");
3207 if (unlikely(in_atomic_preempt_off())) {
3208 __schedule_bug(prev);
3209 preempt_count_set(PREEMPT_DISABLED);
3213 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3215 schedstat_inc(this_rq(), sched_count);
3219 * Pick up the highest-prio task:
3221 static inline struct task_struct *
3222 pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
3224 const struct sched_class *class = &fair_sched_class;
3225 struct task_struct *p;
3228 * Optimization: we know that if all tasks are in
3229 * the fair class we can call that function directly:
3231 if (likely(prev->sched_class == class &&
3232 rq->nr_running == rq->cfs.h_nr_running)) {
3233 p = fair_sched_class.pick_next_task(rq, prev, cookie);
3234 if (unlikely(p == RETRY_TASK))
3237 /* assumes fair_sched_class->next == idle_sched_class */
3239 p = idle_sched_class.pick_next_task(rq, prev, cookie);
3245 for_each_class(class) {
3246 p = class->pick_next_task(rq, prev, cookie);
3248 if (unlikely(p == RETRY_TASK))
3254 BUG(); /* the idle class will always have a runnable task */
3258 * __schedule() is the main scheduler function.
3260 * The main means of driving the scheduler and thus entering this function are:
3262 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3264 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3265 * paths. For example, see arch/x86/entry_64.S.
3267 * To drive preemption between tasks, the scheduler sets the flag in timer
3268 * interrupt handler scheduler_tick().
3270 * 3. Wakeups don't really cause entry into schedule(). They add a
3271 * task to the run-queue and that's it.
3273 * Now, if the new task added to the run-queue preempts the current
3274 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3275 * called on the nearest possible occasion:
3277 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3279 * - in syscall or exception context, at the next outmost
3280 * preempt_enable(). (this might be as soon as the wake_up()'s
3283 * - in IRQ context, return from interrupt-handler to
3284 * preemptible context
3286 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3289 * - cond_resched() call
3290 * - explicit schedule() call
3291 * - return from syscall or exception to user-space
3292 * - return from interrupt-handler to user-space
3294 * WARNING: must be called with preemption disabled!
3296 static void __sched notrace __schedule(bool preempt)
3298 struct task_struct *prev, *next;
3299 unsigned long *switch_count;
3300 struct pin_cookie cookie;
3304 cpu = smp_processor_id();
3309 * do_exit() calls schedule() with preemption disabled as an exception;
3310 * however we must fix that up, otherwise the next task will see an
3311 * inconsistent (higher) preempt count.
3313 * It also avoids the below schedule_debug() test from complaining
3316 if (unlikely(prev->state == TASK_DEAD))
3317 preempt_enable_no_resched_notrace();
3319 schedule_debug(prev);
3321 if (sched_feat(HRTICK))
3324 local_irq_disable();
3325 rcu_note_context_switch();
3328 * Make sure that signal_pending_state()->signal_pending() below
3329 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3330 * done by the caller to avoid the race with signal_wake_up().
3332 smp_mb__before_spinlock();
3333 raw_spin_lock(&rq->lock);
3334 cookie = lockdep_pin_lock(&rq->lock);
3336 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3338 switch_count = &prev->nivcsw;
3339 if (!preempt && prev->state) {
3340 if (unlikely(signal_pending_state(prev->state, prev))) {
3341 prev->state = TASK_RUNNING;
3343 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3347 * If a worker went to sleep, notify and ask workqueue
3348 * whether it wants to wake up a task to maintain
3351 if (prev->flags & PF_WQ_WORKER) {
3352 struct task_struct *to_wakeup;
3354 to_wakeup = wq_worker_sleeping(prev);
3356 try_to_wake_up_local(to_wakeup, cookie);
3359 switch_count = &prev->nvcsw;
3362 if (task_on_rq_queued(prev))
3363 update_rq_clock(rq);
3365 next = pick_next_task(rq, prev, cookie);
3366 clear_tsk_need_resched(prev);
3367 clear_preempt_need_resched();
3368 rq->clock_skip_update = 0;
3370 if (likely(prev != next)) {
3375 trace_sched_switch(preempt, prev, next);
3376 rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */
3378 lockdep_unpin_lock(&rq->lock, cookie);
3379 raw_spin_unlock_irq(&rq->lock);
3382 balance_callback(rq);
3384 STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3386 static inline void sched_submit_work(struct task_struct *tsk)
3388 if (!tsk->state || tsk_is_pi_blocked(tsk))
3391 * If we are going to sleep and we have plugged IO queued,
3392 * make sure to submit it to avoid deadlocks.
3394 if (blk_needs_flush_plug(tsk))
3395 blk_schedule_flush_plug(tsk);
3398 asmlinkage __visible void __sched schedule(void)
3400 struct task_struct *tsk = current;
3402 sched_submit_work(tsk);
3406 sched_preempt_enable_no_resched();
3407 } while (need_resched());
3409 EXPORT_SYMBOL(schedule);
3411 #ifdef CONFIG_CONTEXT_TRACKING
3412 asmlinkage __visible void __sched schedule_user(void)
3415 * If we come here after a random call to set_need_resched(),
3416 * or we have been woken up remotely but the IPI has not yet arrived,
3417 * we haven't yet exited the RCU idle mode. Do it here manually until
3418 * we find a better solution.
3420 * NB: There are buggy callers of this function. Ideally we
3421 * should warn if prev_state != CONTEXT_USER, but that will trigger
3422 * too frequently to make sense yet.
3424 enum ctx_state prev_state = exception_enter();
3426 exception_exit(prev_state);
3431 * schedule_preempt_disabled - called with preemption disabled
3433 * Returns with preemption disabled. Note: preempt_count must be 1
3435 void __sched schedule_preempt_disabled(void)
3437 sched_preempt_enable_no_resched();
3442 static void __sched notrace preempt_schedule_common(void)
3446 * Because the function tracer can trace preempt_count_sub()
3447 * and it also uses preempt_enable/disable_notrace(), if
3448 * NEED_RESCHED is set, the preempt_enable_notrace() called
3449 * by the function tracer will call this function again and
3450 * cause infinite recursion.
3452 * Preemption must be disabled here before the function
3453 * tracer can trace. Break up preempt_disable() into two
3454 * calls. One to disable preemption without fear of being
3455 * traced. The other to still record the preemption latency,
3456 * which can also be traced by the function tracer.
3458 preempt_disable_notrace();
3459 preempt_latency_start(1);
3461 preempt_latency_stop(1);
3462 preempt_enable_no_resched_notrace();
3465 * Check again in case we missed a preemption opportunity
3466 * between schedule and now.
3468 } while (need_resched());
3471 #ifdef CONFIG_PREEMPT
3473 * this is the entry point to schedule() from in-kernel preemption
3474 * off of preempt_enable. Kernel preemptions off return from interrupt
3475 * occur there and call schedule directly.
3477 asmlinkage __visible void __sched notrace preempt_schedule(void)
3480 * If there is a non-zero preempt_count or interrupts are disabled,
3481 * we do not want to preempt the current task. Just return..
3483 if (likely(!preemptible()))
3486 preempt_schedule_common();
3488 NOKPROBE_SYMBOL(preempt_schedule);
3489 EXPORT_SYMBOL(preempt_schedule);
3492 * preempt_schedule_notrace - preempt_schedule called by tracing
3494 * The tracing infrastructure uses preempt_enable_notrace to prevent
3495 * recursion and tracing preempt enabling caused by the tracing
3496 * infrastructure itself. But as tracing can happen in areas coming
3497 * from userspace or just about to enter userspace, a preempt enable
3498 * can occur before user_exit() is called. This will cause the scheduler
3499 * to be called when the system is still in usermode.
3501 * To prevent this, the preempt_enable_notrace will use this function
3502 * instead of preempt_schedule() to exit user context if needed before
3503 * calling the scheduler.
3505 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3507 enum ctx_state prev_ctx;
3509 if (likely(!preemptible()))
3514 * Because the function tracer can trace preempt_count_sub()
3515 * and it also uses preempt_enable/disable_notrace(), if
3516 * NEED_RESCHED is set, the preempt_enable_notrace() called
3517 * by the function tracer will call this function again and
3518 * cause infinite recursion.
3520 * Preemption must be disabled here before the function
3521 * tracer can trace. Break up preempt_disable() into two
3522 * calls. One to disable preemption without fear of being
3523 * traced. The other to still record the preemption latency,
3524 * which can also be traced by the function tracer.
3526 preempt_disable_notrace();
3527 preempt_latency_start(1);
3529 * Needs preempt disabled in case user_exit() is traced
3530 * and the tracer calls preempt_enable_notrace() causing
3531 * an infinite recursion.
3533 prev_ctx = exception_enter();
3535 exception_exit(prev_ctx);
3537 preempt_latency_stop(1);
3538 preempt_enable_no_resched_notrace();
3539 } while (need_resched());
3541 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3543 #endif /* CONFIG_PREEMPT */
3546 * this is the entry point to schedule() from kernel preemption
3547 * off of irq context.
3548 * Note, that this is called and return with irqs disabled. This will
3549 * protect us against recursive calling from irq.
3551 asmlinkage __visible void __sched preempt_schedule_irq(void)
3553 enum ctx_state prev_state;
3555 /* Catch callers which need to be fixed */
3556 BUG_ON(preempt_count() || !irqs_disabled());
3558 prev_state = exception_enter();
3564 local_irq_disable();
3565 sched_preempt_enable_no_resched();
3566 } while (need_resched());
3568 exception_exit(prev_state);
3571 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3574 return try_to_wake_up(curr->private, mode, wake_flags);
3576 EXPORT_SYMBOL(default_wake_function);
3578 #ifdef CONFIG_RT_MUTEXES
3581 * rt_mutex_setprio - set the current priority of a task
3583 * @prio: prio value (kernel-internal form)
3585 * This function changes the 'effective' priority of a task. It does
3586 * not touch ->normal_prio like __setscheduler().
3588 * Used by the rt_mutex code to implement priority inheritance
3589 * logic. Call site only calls if the priority of the task changed.
3591 void rt_mutex_setprio(struct task_struct *p, int prio)
3593 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3594 const struct sched_class *prev_class;
3598 BUG_ON(prio > MAX_PRIO);
3600 rq = __task_rq_lock(p, &rf);
3603 * Idle task boosting is a nono in general. There is one
3604 * exception, when PREEMPT_RT and NOHZ is active:
3606 * The idle task calls get_next_timer_interrupt() and holds
3607 * the timer wheel base->lock on the CPU and another CPU wants
3608 * to access the timer (probably to cancel it). We can safely
3609 * ignore the boosting request, as the idle CPU runs this code
3610 * with interrupts disabled and will complete the lock
3611 * protected section without being interrupted. So there is no
3612 * real need to boost.
3614 if (unlikely(p == rq->idle)) {
3615 WARN_ON(p != rq->curr);
3616 WARN_ON(p->pi_blocked_on);
3620 trace_sched_pi_setprio(p, prio);
3623 if (oldprio == prio)
3624 queue_flag &= ~DEQUEUE_MOVE;
3626 prev_class = p->sched_class;
3627 queued = task_on_rq_queued(p);
3628 running = task_current(rq, p);
3630 dequeue_task(rq, p, queue_flag);
3632 put_prev_task(rq, p);
3635 * Boosting condition are:
3636 * 1. -rt task is running and holds mutex A
3637 * --> -dl task blocks on mutex A
3639 * 2. -dl task is running and holds mutex A
3640 * --> -dl task blocks on mutex A and could preempt the
3643 if (dl_prio(prio)) {
3644 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3645 if (!dl_prio(p->normal_prio) ||
3646 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3647 p->dl.dl_boosted = 1;
3648 queue_flag |= ENQUEUE_REPLENISH;
3650 p->dl.dl_boosted = 0;
3651 p->sched_class = &dl_sched_class;
3652 } else if (rt_prio(prio)) {
3653 if (dl_prio(oldprio))
3654 p->dl.dl_boosted = 0;
3656 queue_flag |= ENQUEUE_HEAD;
3657 p->sched_class = &rt_sched_class;
3659 if (dl_prio(oldprio))
3660 p->dl.dl_boosted = 0;
3661 if (rt_prio(oldprio))
3663 p->sched_class = &fair_sched_class;
3669 p->sched_class->set_curr_task(rq);
3671 enqueue_task(rq, p, queue_flag);
3673 check_class_changed(rq, p, prev_class, oldprio);
3675 preempt_disable(); /* avoid rq from going away on us */
3676 __task_rq_unlock(rq, &rf);
3678 balance_callback(rq);
3683 void set_user_nice(struct task_struct *p, long nice)
3685 int old_prio, delta, queued;
3689 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3692 * We have to be careful, if called from sys_setpriority(),
3693 * the task might be in the middle of scheduling on another CPU.
3695 rq = task_rq_lock(p, &rf);
3697 * The RT priorities are set via sched_setscheduler(), but we still
3698 * allow the 'normal' nice value to be set - but as expected
3699 * it wont have any effect on scheduling until the task is
3700 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3702 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3703 p->static_prio = NICE_TO_PRIO(nice);
3706 queued = task_on_rq_queued(p);
3708 dequeue_task(rq, p, DEQUEUE_SAVE);
3710 p->static_prio = NICE_TO_PRIO(nice);
3713 p->prio = effective_prio(p);
3714 delta = p->prio - old_prio;
3717 enqueue_task(rq, p, ENQUEUE_RESTORE);
3719 * If the task increased its priority or is running and
3720 * lowered its priority, then reschedule its CPU:
3722 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3726 task_rq_unlock(rq, p, &rf);
3728 EXPORT_SYMBOL(set_user_nice);
3731 * can_nice - check if a task can reduce its nice value
3735 int can_nice(const struct task_struct *p, const int nice)
3737 /* convert nice value [19,-20] to rlimit style value [1,40] */
3738 int nice_rlim = nice_to_rlimit(nice);
3740 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3741 capable(CAP_SYS_NICE));
3744 #ifdef __ARCH_WANT_SYS_NICE
3747 * sys_nice - change the priority of the current process.
3748 * @increment: priority increment
3750 * sys_setpriority is a more generic, but much slower function that
3751 * does similar things.
3753 SYSCALL_DEFINE1(nice, int, increment)
3758 * Setpriority might change our priority at the same moment.
3759 * We don't have to worry. Conceptually one call occurs first
3760 * and we have a single winner.
3762 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3763 nice = task_nice(current) + increment;
3765 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3766 if (increment < 0 && !can_nice(current, nice))
3769 retval = security_task_setnice(current, nice);
3773 set_user_nice(current, nice);
3780 * task_prio - return the priority value of a given task.
3781 * @p: the task in question.
3783 * Return: The priority value as seen by users in /proc.
3784 * RT tasks are offset by -200. Normal tasks are centered
3785 * around 0, value goes from -16 to +15.
3787 int task_prio(const struct task_struct *p)
3789 return p->prio - MAX_RT_PRIO;
3793 * idle_cpu - is a given cpu idle currently?
3794 * @cpu: the processor in question.
3796 * Return: 1 if the CPU is currently idle. 0 otherwise.
3798 int idle_cpu(int cpu)
3800 struct rq *rq = cpu_rq(cpu);
3802 if (rq->curr != rq->idle)
3809 if (!llist_empty(&rq->wake_list))
3817 * idle_task - return the idle task for a given cpu.
3818 * @cpu: the processor in question.
3820 * Return: The idle task for the cpu @cpu.
3822 struct task_struct *idle_task(int cpu)
3824 return cpu_rq(cpu)->idle;
3828 * find_process_by_pid - find a process with a matching PID value.
3829 * @pid: the pid in question.
3831 * The task of @pid, if found. %NULL otherwise.
3833 static struct task_struct *find_process_by_pid(pid_t pid)
3835 return pid ? find_task_by_vpid(pid) : current;
3839 * This function initializes the sched_dl_entity of a newly becoming
3840 * SCHED_DEADLINE task.
3842 * Only the static values are considered here, the actual runtime and the
3843 * absolute deadline will be properly calculated when the task is enqueued
3844 * for the first time with its new policy.
3847 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3849 struct sched_dl_entity *dl_se = &p->dl;
3851 dl_se->dl_runtime = attr->sched_runtime;
3852 dl_se->dl_deadline = attr->sched_deadline;
3853 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3854 dl_se->flags = attr->sched_flags;
3855 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3858 * Changing the parameters of a task is 'tricky' and we're not doing
3859 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3861 * What we SHOULD do is delay the bandwidth release until the 0-lag
3862 * point. This would include retaining the task_struct until that time
3863 * and change dl_overflow() to not immediately decrement the current
3866 * Instead we retain the current runtime/deadline and let the new
3867 * parameters take effect after the current reservation period lapses.
3868 * This is safe (albeit pessimistic) because the 0-lag point is always
3869 * before the current scheduling deadline.
3871 * We can still have temporary overloads because we do not delay the
3872 * change in bandwidth until that time; so admission control is
3873 * not on the safe side. It does however guarantee tasks will never
3874 * consume more than promised.
3879 * sched_setparam() passes in -1 for its policy, to let the functions
3880 * it calls know not to change it.
3882 #define SETPARAM_POLICY -1
3884 static void __setscheduler_params(struct task_struct *p,
3885 const struct sched_attr *attr)
3887 int policy = attr->sched_policy;
3889 if (policy == SETPARAM_POLICY)
3894 if (dl_policy(policy))
3895 __setparam_dl(p, attr);
3896 else if (fair_policy(policy))
3897 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3900 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3901 * !rt_policy. Always setting this ensures that things like
3902 * getparam()/getattr() don't report silly values for !rt tasks.
3904 p->rt_priority = attr->sched_priority;
3905 p->normal_prio = normal_prio(p);
3909 /* Actually do priority change: must hold pi & rq lock. */
3910 static void __setscheduler(struct rq *rq, struct task_struct *p,
3911 const struct sched_attr *attr, bool keep_boost)
3913 __setscheduler_params(p, attr);
3916 * Keep a potential priority boosting if called from
3917 * sched_setscheduler().
3920 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3922 p->prio = normal_prio(p);
3924 if (dl_prio(p->prio))
3925 p->sched_class = &dl_sched_class;
3926 else if (rt_prio(p->prio))
3927 p->sched_class = &rt_sched_class;
3929 p->sched_class = &fair_sched_class;
3933 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3935 struct sched_dl_entity *dl_se = &p->dl;
3937 attr->sched_priority = p->rt_priority;
3938 attr->sched_runtime = dl_se->dl_runtime;
3939 attr->sched_deadline = dl_se->dl_deadline;
3940 attr->sched_period = dl_se->dl_period;
3941 attr->sched_flags = dl_se->flags;
3945 * This function validates the new parameters of a -deadline task.
3946 * We ask for the deadline not being zero, and greater or equal
3947 * than the runtime, as well as the period of being zero or
3948 * greater than deadline. Furthermore, we have to be sure that
3949 * user parameters are above the internal resolution of 1us (we
3950 * check sched_runtime only since it is always the smaller one) and
3951 * below 2^63 ns (we have to check both sched_deadline and
3952 * sched_period, as the latter can be zero).
3955 __checkparam_dl(const struct sched_attr *attr)
3958 if (attr->sched_deadline == 0)
3962 * Since we truncate DL_SCALE bits, make sure we're at least
3965 if (attr->sched_runtime < (1ULL << DL_SCALE))
3969 * Since we use the MSB for wrap-around and sign issues, make
3970 * sure it's not set (mind that period can be equal to zero).
3972 if (attr->sched_deadline & (1ULL << 63) ||
3973 attr->sched_period & (1ULL << 63))
3976 /* runtime <= deadline <= period (if period != 0) */
3977 if ((attr->sched_period != 0 &&
3978 attr->sched_period < attr->sched_deadline) ||
3979 attr->sched_deadline < attr->sched_runtime)
3986 * check the target process has a UID that matches the current process's
3988 static bool check_same_owner(struct task_struct *p)
3990 const struct cred *cred = current_cred(), *pcred;
3994 pcred = __task_cred(p);
3995 match = (uid_eq(cred->euid, pcred->euid) ||
3996 uid_eq(cred->euid, pcred->uid));
4001 static bool dl_param_changed(struct task_struct *p,
4002 const struct sched_attr *attr)
4004 struct sched_dl_entity *dl_se = &p->dl;
4006 if (dl_se->dl_runtime != attr->sched_runtime ||
4007 dl_se->dl_deadline != attr->sched_deadline ||
4008 dl_se->dl_period != attr->sched_period ||
4009 dl_se->flags != attr->sched_flags)
4015 static int __sched_setscheduler(struct task_struct *p,
4016 const struct sched_attr *attr,
4019 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4020 MAX_RT_PRIO - 1 - attr->sched_priority;
4021 int retval, oldprio, oldpolicy = -1, queued, running;
4022 int new_effective_prio, policy = attr->sched_policy;
4023 const struct sched_class *prev_class;
4026 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4029 /* may grab non-irq protected spin_locks */
4030 BUG_ON(in_interrupt());
4032 /* double check policy once rq lock held */
4034 reset_on_fork = p->sched_reset_on_fork;
4035 policy = oldpolicy = p->policy;
4037 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4039 if (!valid_policy(policy))
4043 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4047 * Valid priorities for SCHED_FIFO and SCHED_RR are
4048 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4049 * SCHED_BATCH and SCHED_IDLE is 0.
4051 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4052 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4054 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4055 (rt_policy(policy) != (attr->sched_priority != 0)))
4059 * Allow unprivileged RT tasks to decrease priority:
4061 if (user && !capable(CAP_SYS_NICE)) {
4062 if (fair_policy(policy)) {
4063 if (attr->sched_nice < task_nice(p) &&
4064 !can_nice(p, attr->sched_nice))
4068 if (rt_policy(policy)) {
4069 unsigned long rlim_rtprio =
4070 task_rlimit(p, RLIMIT_RTPRIO);
4072 /* can't set/change the rt policy */
4073 if (policy != p->policy && !rlim_rtprio)
4076 /* can't increase priority */
4077 if (attr->sched_priority > p->rt_priority &&
4078 attr->sched_priority > rlim_rtprio)
4083 * Can't set/change SCHED_DEADLINE policy at all for now
4084 * (safest behavior); in the future we would like to allow
4085 * unprivileged DL tasks to increase their relative deadline
4086 * or reduce their runtime (both ways reducing utilization)
4088 if (dl_policy(policy))
4092 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4093 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4095 if (idle_policy(p->policy) && !idle_policy(policy)) {
4096 if (!can_nice(p, task_nice(p)))
4100 /* can't change other user's priorities */
4101 if (!check_same_owner(p))
4104 /* Normal users shall not reset the sched_reset_on_fork flag */
4105 if (p->sched_reset_on_fork && !reset_on_fork)
4110 retval = security_task_setscheduler(p);
4116 * make sure no PI-waiters arrive (or leave) while we are
4117 * changing the priority of the task:
4119 * To be able to change p->policy safely, the appropriate
4120 * runqueue lock must be held.
4122 rq = task_rq_lock(p, &rf);
4125 * Changing the policy of the stop threads its a very bad idea
4127 if (p == rq->stop) {
4128 task_rq_unlock(rq, p, &rf);
4133 * If not changing anything there's no need to proceed further,
4134 * but store a possible modification of reset_on_fork.
4136 if (unlikely(policy == p->policy)) {
4137 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4139 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4141 if (dl_policy(policy) && dl_param_changed(p, attr))
4144 p->sched_reset_on_fork = reset_on_fork;
4145 task_rq_unlock(rq, p, &rf);
4151 #ifdef CONFIG_RT_GROUP_SCHED
4153 * Do not allow realtime tasks into groups that have no runtime
4156 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4157 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4158 !task_group_is_autogroup(task_group(p))) {
4159 task_rq_unlock(rq, p, &rf);
4164 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4165 cpumask_t *span = rq->rd->span;
4168 * Don't allow tasks with an affinity mask smaller than
4169 * the entire root_domain to become SCHED_DEADLINE. We
4170 * will also fail if there's no bandwidth available.
4172 if (!cpumask_subset(span, &p->cpus_allowed) ||
4173 rq->rd->dl_bw.bw == 0) {
4174 task_rq_unlock(rq, p, &rf);
4181 /* recheck policy now with rq lock held */
4182 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4183 policy = oldpolicy = -1;
4184 task_rq_unlock(rq, p, &rf);
4189 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4190 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4193 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4194 task_rq_unlock(rq, p, &rf);
4198 p->sched_reset_on_fork = reset_on_fork;
4203 * Take priority boosted tasks into account. If the new
4204 * effective priority is unchanged, we just store the new
4205 * normal parameters and do not touch the scheduler class and
4206 * the runqueue. This will be done when the task deboost
4209 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4210 if (new_effective_prio == oldprio)
4211 queue_flags &= ~DEQUEUE_MOVE;
4214 queued = task_on_rq_queued(p);
4215 running = task_current(rq, p);
4217 dequeue_task(rq, p, queue_flags);
4219 put_prev_task(rq, p);
4221 prev_class = p->sched_class;
4222 __setscheduler(rq, p, attr, pi);
4225 p->sched_class->set_curr_task(rq);
4228 * We enqueue to tail when the priority of a task is
4229 * increased (user space view).
4231 if (oldprio < p->prio)
4232 queue_flags |= ENQUEUE_HEAD;
4234 enqueue_task(rq, p, queue_flags);
4237 check_class_changed(rq, p, prev_class, oldprio);
4238 preempt_disable(); /* avoid rq from going away on us */
4239 task_rq_unlock(rq, p, &rf);
4242 rt_mutex_adjust_pi(p);
4245 * Run balance callbacks after we've adjusted the PI chain.
4247 balance_callback(rq);
4253 static int _sched_setscheduler(struct task_struct *p, int policy,
4254 const struct sched_param *param, bool check)
4256 struct sched_attr attr = {
4257 .sched_policy = policy,
4258 .sched_priority = param->sched_priority,
4259 .sched_nice = PRIO_TO_NICE(p->static_prio),
4262 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4263 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4264 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4265 policy &= ~SCHED_RESET_ON_FORK;
4266 attr.sched_policy = policy;
4269 return __sched_setscheduler(p, &attr, check, true);
4272 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4273 * @p: the task in question.
4274 * @policy: new policy.
4275 * @param: structure containing the new RT priority.
4277 * Return: 0 on success. An error code otherwise.
4279 * NOTE that the task may be already dead.
4281 int sched_setscheduler(struct task_struct *p, int policy,
4282 const struct sched_param *param)
4284 return _sched_setscheduler(p, policy, param, true);
4286 EXPORT_SYMBOL_GPL(sched_setscheduler);
4288 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4290 return __sched_setscheduler(p, attr, true, true);
4292 EXPORT_SYMBOL_GPL(sched_setattr);
4295 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4296 * @p: the task in question.
4297 * @policy: new policy.
4298 * @param: structure containing the new RT priority.
4300 * Just like sched_setscheduler, only don't bother checking if the
4301 * current context has permission. For example, this is needed in
4302 * stop_machine(): we create temporary high priority worker threads,
4303 * but our caller might not have that capability.
4305 * Return: 0 on success. An error code otherwise.
4307 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4308 const struct sched_param *param)
4310 return _sched_setscheduler(p, policy, param, false);
4312 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4315 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4317 struct sched_param lparam;
4318 struct task_struct *p;
4321 if (!param || pid < 0)
4323 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4328 p = find_process_by_pid(pid);
4330 retval = sched_setscheduler(p, policy, &lparam);
4337 * Mimics kernel/events/core.c perf_copy_attr().
4339 static int sched_copy_attr(struct sched_attr __user *uattr,
4340 struct sched_attr *attr)
4345 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4349 * zero the full structure, so that a short copy will be nice.
4351 memset(attr, 0, sizeof(*attr));
4353 ret = get_user(size, &uattr->size);
4357 if (size > PAGE_SIZE) /* silly large */
4360 if (!size) /* abi compat */
4361 size = SCHED_ATTR_SIZE_VER0;
4363 if (size < SCHED_ATTR_SIZE_VER0)
4367 * If we're handed a bigger struct than we know of,
4368 * ensure all the unknown bits are 0 - i.e. new
4369 * user-space does not rely on any kernel feature
4370 * extensions we dont know about yet.
4372 if (size > sizeof(*attr)) {
4373 unsigned char __user *addr;
4374 unsigned char __user *end;
4377 addr = (void __user *)uattr + sizeof(*attr);
4378 end = (void __user *)uattr + size;
4380 for (; addr < end; addr++) {
4381 ret = get_user(val, addr);
4387 size = sizeof(*attr);
4390 ret = copy_from_user(attr, uattr, size);
4395 * XXX: do we want to be lenient like existing syscalls; or do we want
4396 * to be strict and return an error on out-of-bounds values?
4398 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4403 put_user(sizeof(*attr), &uattr->size);
4408 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4409 * @pid: the pid in question.
4410 * @policy: new policy.
4411 * @param: structure containing the new RT priority.
4413 * Return: 0 on success. An error code otherwise.
4415 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4416 struct sched_param __user *, param)
4418 /* negative values for policy are not valid */
4422 return do_sched_setscheduler(pid, policy, param);
4426 * sys_sched_setparam - set/change the RT priority of a thread
4427 * @pid: the pid in question.
4428 * @param: structure containing the new RT priority.
4430 * Return: 0 on success. An error code otherwise.
4432 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4434 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4438 * sys_sched_setattr - same as above, but with extended sched_attr
4439 * @pid: the pid in question.
4440 * @uattr: structure containing the extended parameters.
4441 * @flags: for future extension.
4443 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4444 unsigned int, flags)
4446 struct sched_attr attr;
4447 struct task_struct *p;
4450 if (!uattr || pid < 0 || flags)
4453 retval = sched_copy_attr(uattr, &attr);
4457 if ((int)attr.sched_policy < 0)
4462 p = find_process_by_pid(pid);
4464 retval = sched_setattr(p, &attr);
4471 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4472 * @pid: the pid in question.
4474 * Return: On success, the policy of the thread. Otherwise, a negative error
4477 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4479 struct task_struct *p;
4487 p = find_process_by_pid(pid);
4489 retval = security_task_getscheduler(p);
4492 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4499 * sys_sched_getparam - get the RT priority of a thread
4500 * @pid: the pid in question.
4501 * @param: structure containing the RT priority.
4503 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4506 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4508 struct sched_param lp = { .sched_priority = 0 };
4509 struct task_struct *p;
4512 if (!param || pid < 0)
4516 p = find_process_by_pid(pid);
4521 retval = security_task_getscheduler(p);
4525 if (task_has_rt_policy(p))
4526 lp.sched_priority = p->rt_priority;
4530 * This one might sleep, we cannot do it with a spinlock held ...
4532 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4541 static int sched_read_attr(struct sched_attr __user *uattr,
4542 struct sched_attr *attr,
4547 if (!access_ok(VERIFY_WRITE, uattr, usize))
4551 * If we're handed a smaller struct than we know of,
4552 * ensure all the unknown bits are 0 - i.e. old
4553 * user-space does not get uncomplete information.
4555 if (usize < sizeof(*attr)) {
4556 unsigned char *addr;
4559 addr = (void *)attr + usize;
4560 end = (void *)attr + sizeof(*attr);
4562 for (; addr < end; addr++) {
4570 ret = copy_to_user(uattr, attr, attr->size);
4578 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4579 * @pid: the pid in question.
4580 * @uattr: structure containing the extended parameters.
4581 * @size: sizeof(attr) for fwd/bwd comp.
4582 * @flags: for future extension.
4584 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4585 unsigned int, size, unsigned int, flags)
4587 struct sched_attr attr = {
4588 .size = sizeof(struct sched_attr),
4590 struct task_struct *p;
4593 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4594 size < SCHED_ATTR_SIZE_VER0 || flags)
4598 p = find_process_by_pid(pid);
4603 retval = security_task_getscheduler(p);
4607 attr.sched_policy = p->policy;
4608 if (p->sched_reset_on_fork)
4609 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4610 if (task_has_dl_policy(p))
4611 __getparam_dl(p, &attr);
4612 else if (task_has_rt_policy(p))
4613 attr.sched_priority = p->rt_priority;
4615 attr.sched_nice = task_nice(p);
4619 retval = sched_read_attr(uattr, &attr, size);
4627 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4629 cpumask_var_t cpus_allowed, new_mask;
4630 struct task_struct *p;
4635 p = find_process_by_pid(pid);
4641 /* Prevent p going away */
4645 if (p->flags & PF_NO_SETAFFINITY) {
4649 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4653 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4655 goto out_free_cpus_allowed;
4658 if (!check_same_owner(p)) {
4660 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4662 goto out_free_new_mask;
4667 retval = security_task_setscheduler(p);
4669 goto out_free_new_mask;
4672 cpuset_cpus_allowed(p, cpus_allowed);
4673 cpumask_and(new_mask, in_mask, cpus_allowed);
4676 * Since bandwidth control happens on root_domain basis,
4677 * if admission test is enabled, we only admit -deadline
4678 * tasks allowed to run on all the CPUs in the task's
4682 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4684 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4687 goto out_free_new_mask;
4693 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4696 cpuset_cpus_allowed(p, cpus_allowed);
4697 if (!cpumask_subset(new_mask, cpus_allowed)) {
4699 * We must have raced with a concurrent cpuset
4700 * update. Just reset the cpus_allowed to the
4701 * cpuset's cpus_allowed
4703 cpumask_copy(new_mask, cpus_allowed);
4708 free_cpumask_var(new_mask);
4709 out_free_cpus_allowed:
4710 free_cpumask_var(cpus_allowed);
4716 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4717 struct cpumask *new_mask)
4719 if (len < cpumask_size())
4720 cpumask_clear(new_mask);
4721 else if (len > cpumask_size())
4722 len = cpumask_size();
4724 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4728 * sys_sched_setaffinity - set the cpu affinity of a process
4729 * @pid: pid of the process
4730 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4731 * @user_mask_ptr: user-space pointer to the new cpu mask
4733 * Return: 0 on success. An error code otherwise.
4735 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4736 unsigned long __user *, user_mask_ptr)
4738 cpumask_var_t new_mask;
4741 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4744 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4746 retval = sched_setaffinity(pid, new_mask);
4747 free_cpumask_var(new_mask);
4751 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4753 struct task_struct *p;
4754 unsigned long flags;
4760 p = find_process_by_pid(pid);
4764 retval = security_task_getscheduler(p);
4768 raw_spin_lock_irqsave(&p->pi_lock, flags);
4769 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4770 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4779 * sys_sched_getaffinity - get the cpu affinity of a process
4780 * @pid: pid of the process
4781 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4782 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4784 * Return: size of CPU mask copied to user_mask_ptr on success. An
4785 * error code otherwise.
4787 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4788 unsigned long __user *, user_mask_ptr)
4793 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4795 if (len & (sizeof(unsigned long)-1))
4798 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4801 ret = sched_getaffinity(pid, mask);
4803 size_t retlen = min_t(size_t, len, cpumask_size());
4805 if (copy_to_user(user_mask_ptr, mask, retlen))
4810 free_cpumask_var(mask);
4816 * sys_sched_yield - yield the current processor to other threads.
4818 * This function yields the current CPU to other tasks. If there are no
4819 * other threads running on this CPU then this function will return.
4823 SYSCALL_DEFINE0(sched_yield)
4825 struct rq *rq = this_rq_lock();
4827 schedstat_inc(rq, yld_count);
4828 current->sched_class->yield_task(rq);
4831 * Since we are going to call schedule() anyway, there's
4832 * no need to preempt or enable interrupts:
4834 __release(rq->lock);
4835 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4836 do_raw_spin_unlock(&rq->lock);
4837 sched_preempt_enable_no_resched();
4844 int __sched _cond_resched(void)
4846 if (should_resched(0)) {
4847 preempt_schedule_common();
4852 EXPORT_SYMBOL(_cond_resched);
4855 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4856 * call schedule, and on return reacquire the lock.
4858 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4859 * operations here to prevent schedule() from being called twice (once via
4860 * spin_unlock(), once by hand).
4862 int __cond_resched_lock(spinlock_t *lock)
4864 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4867 lockdep_assert_held(lock);
4869 if (spin_needbreak(lock) || resched) {
4872 preempt_schedule_common();
4880 EXPORT_SYMBOL(__cond_resched_lock);
4882 int __sched __cond_resched_softirq(void)
4884 BUG_ON(!in_softirq());
4886 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4888 preempt_schedule_common();
4894 EXPORT_SYMBOL(__cond_resched_softirq);
4897 * yield - yield the current processor to other threads.
4899 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4901 * The scheduler is at all times free to pick the calling task as the most
4902 * eligible task to run, if removing the yield() call from your code breaks
4903 * it, its already broken.
4905 * Typical broken usage is:
4910 * where one assumes that yield() will let 'the other' process run that will
4911 * make event true. If the current task is a SCHED_FIFO task that will never
4912 * happen. Never use yield() as a progress guarantee!!
4914 * If you want to use yield() to wait for something, use wait_event().
4915 * If you want to use yield() to be 'nice' for others, use cond_resched().
4916 * If you still want to use yield(), do not!
4918 void __sched yield(void)
4920 set_current_state(TASK_RUNNING);
4923 EXPORT_SYMBOL(yield);
4926 * yield_to - yield the current processor to another thread in
4927 * your thread group, or accelerate that thread toward the
4928 * processor it's on.
4930 * @preempt: whether task preemption is allowed or not
4932 * It's the caller's job to ensure that the target task struct
4933 * can't go away on us before we can do any checks.
4936 * true (>0) if we indeed boosted the target task.
4937 * false (0) if we failed to boost the target.
4938 * -ESRCH if there's no task to yield to.
4940 int __sched yield_to(struct task_struct *p, bool preempt)
4942 struct task_struct *curr = current;
4943 struct rq *rq, *p_rq;
4944 unsigned long flags;
4947 local_irq_save(flags);
4953 * If we're the only runnable task on the rq and target rq also
4954 * has only one task, there's absolutely no point in yielding.
4956 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4961 double_rq_lock(rq, p_rq);
4962 if (task_rq(p) != p_rq) {
4963 double_rq_unlock(rq, p_rq);
4967 if (!curr->sched_class->yield_to_task)
4970 if (curr->sched_class != p->sched_class)
4973 if (task_running(p_rq, p) || p->state)
4976 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4978 schedstat_inc(rq, yld_count);
4980 * Make p's CPU reschedule; pick_next_entity takes care of
4983 if (preempt && rq != p_rq)
4988 double_rq_unlock(rq, p_rq);
4990 local_irq_restore(flags);
4997 EXPORT_SYMBOL_GPL(yield_to);
5000 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5001 * that process accounting knows that this is a task in IO wait state.
5003 long __sched io_schedule_timeout(long timeout)
5005 int old_iowait = current->in_iowait;
5009 current->in_iowait = 1;
5010 blk_schedule_flush_plug(current);
5012 delayacct_blkio_start();
5014 atomic_inc(&rq->nr_iowait);
5015 ret = schedule_timeout(timeout);
5016 current->in_iowait = old_iowait;
5017 atomic_dec(&rq->nr_iowait);
5018 delayacct_blkio_end();
5022 EXPORT_SYMBOL(io_schedule_timeout);
5025 * sys_sched_get_priority_max - return maximum RT priority.
5026 * @policy: scheduling class.
5028 * Return: On success, this syscall returns the maximum
5029 * rt_priority that can be used by a given scheduling class.
5030 * On failure, a negative error code is returned.
5032 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5039 ret = MAX_USER_RT_PRIO-1;
5041 case SCHED_DEADLINE:
5052 * sys_sched_get_priority_min - return minimum RT priority.
5053 * @policy: scheduling class.
5055 * Return: On success, this syscall returns the minimum
5056 * rt_priority that can be used by a given scheduling class.
5057 * On failure, a negative error code is returned.
5059 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5068 case SCHED_DEADLINE:
5078 * sys_sched_rr_get_interval - return the default timeslice of a process.
5079 * @pid: pid of the process.
5080 * @interval: userspace pointer to the timeslice value.
5082 * this syscall writes the default timeslice value of a given process
5083 * into the user-space timespec buffer. A value of '0' means infinity.
5085 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5088 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5089 struct timespec __user *, interval)
5091 struct task_struct *p;
5092 unsigned int time_slice;
5103 p = find_process_by_pid(pid);
5107 retval = security_task_getscheduler(p);
5111 rq = task_rq_lock(p, &rf);
5113 if (p->sched_class->get_rr_interval)
5114 time_slice = p->sched_class->get_rr_interval(rq, p);
5115 task_rq_unlock(rq, p, &rf);
5118 jiffies_to_timespec(time_slice, &t);
5119 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5127 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5129 void sched_show_task(struct task_struct *p)
5131 unsigned long free = 0;
5133 unsigned long state = p->state;
5136 state = __ffs(state) + 1;
5137 printk(KERN_INFO "%-15.15s %c", p->comm,
5138 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5139 #if BITS_PER_LONG == 32
5140 if (state == TASK_RUNNING)
5141 printk(KERN_CONT " running ");
5143 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5145 if (state == TASK_RUNNING)
5146 printk(KERN_CONT " running task ");
5148 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5150 #ifdef CONFIG_DEBUG_STACK_USAGE
5151 free = stack_not_used(p);
5156 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5158 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5159 task_pid_nr(p), ppid,
5160 (unsigned long)task_thread_info(p)->flags);
5162 print_worker_info(KERN_INFO, p);
5163 show_stack(p, NULL);
5166 void show_state_filter(unsigned long state_filter)
5168 struct task_struct *g, *p;
5170 #if BITS_PER_LONG == 32
5172 " task PC stack pid father\n");
5175 " task PC stack pid father\n");
5178 for_each_process_thread(g, p) {
5180 * reset the NMI-timeout, listing all files on a slow
5181 * console might take a lot of time:
5182 * Also, reset softlockup watchdogs on all CPUs, because
5183 * another CPU might be blocked waiting for us to process
5186 touch_nmi_watchdog();
5187 touch_all_softlockup_watchdogs();
5188 if (!state_filter || (p->state & state_filter))
5192 #ifdef CONFIG_SCHED_DEBUG
5194 sysrq_sched_debug_show();
5198 * Only show locks if all tasks are dumped:
5201 debug_show_all_locks();
5204 void init_idle_bootup_task(struct task_struct *idle)
5206 idle->sched_class = &idle_sched_class;
5210 * init_idle - set up an idle thread for a given CPU
5211 * @idle: task in question
5212 * @cpu: cpu the idle task belongs to
5214 * NOTE: this function does not set the idle thread's NEED_RESCHED
5215 * flag, to make booting more robust.
5217 void init_idle(struct task_struct *idle, int cpu)
5219 struct rq *rq = cpu_rq(cpu);
5220 unsigned long flags;
5222 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5223 raw_spin_lock(&rq->lock);
5225 __sched_fork(0, idle);
5226 idle->state = TASK_RUNNING;
5227 idle->se.exec_start = sched_clock();
5229 kasan_unpoison_task_stack(idle);
5233 * Its possible that init_idle() gets called multiple times on a task,
5234 * in that case do_set_cpus_allowed() will not do the right thing.
5236 * And since this is boot we can forgo the serialization.
5238 set_cpus_allowed_common(idle, cpumask_of(cpu));
5241 * We're having a chicken and egg problem, even though we are
5242 * holding rq->lock, the cpu isn't yet set to this cpu so the
5243 * lockdep check in task_group() will fail.
5245 * Similar case to sched_fork(). / Alternatively we could
5246 * use task_rq_lock() here and obtain the other rq->lock.
5251 __set_task_cpu(idle, cpu);
5254 rq->curr = rq->idle = idle;
5255 idle->on_rq = TASK_ON_RQ_QUEUED;
5259 raw_spin_unlock(&rq->lock);
5260 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5262 /* Set the preempt count _outside_ the spinlocks! */
5263 init_idle_preempt_count(idle, cpu);
5266 * The idle tasks have their own, simple scheduling class:
5268 idle->sched_class = &idle_sched_class;
5269 ftrace_graph_init_idle_task(idle, cpu);
5270 vtime_init_idle(idle, cpu);
5272 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5276 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5277 const struct cpumask *trial)
5279 int ret = 1, trial_cpus;
5280 struct dl_bw *cur_dl_b;
5281 unsigned long flags;
5283 if (!cpumask_weight(cur))
5286 rcu_read_lock_sched();
5287 cur_dl_b = dl_bw_of(cpumask_any(cur));
5288 trial_cpus = cpumask_weight(trial);
5290 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5291 if (cur_dl_b->bw != -1 &&
5292 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5294 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5295 rcu_read_unlock_sched();
5300 int task_can_attach(struct task_struct *p,
5301 const struct cpumask *cs_cpus_allowed)
5306 * Kthreads which disallow setaffinity shouldn't be moved
5307 * to a new cpuset; we don't want to change their cpu
5308 * affinity and isolating such threads by their set of
5309 * allowed nodes is unnecessary. Thus, cpusets are not
5310 * applicable for such threads. This prevents checking for
5311 * success of set_cpus_allowed_ptr() on all attached tasks
5312 * before cpus_allowed may be changed.
5314 if (p->flags & PF_NO_SETAFFINITY) {
5320 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5322 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5327 unsigned long flags;
5329 rcu_read_lock_sched();
5330 dl_b = dl_bw_of(dest_cpu);
5331 raw_spin_lock_irqsave(&dl_b->lock, flags);
5332 cpus = dl_bw_cpus(dest_cpu);
5333 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5338 * We reserve space for this task in the destination
5339 * root_domain, as we can't fail after this point.
5340 * We will free resources in the source root_domain
5341 * later on (see set_cpus_allowed_dl()).
5343 __dl_add(dl_b, p->dl.dl_bw);
5345 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5346 rcu_read_unlock_sched();
5356 static bool sched_smp_initialized __read_mostly;
5358 #ifdef CONFIG_NUMA_BALANCING
5359 /* Migrate current task p to target_cpu */
5360 int migrate_task_to(struct task_struct *p, int target_cpu)
5362 struct migration_arg arg = { p, target_cpu };
5363 int curr_cpu = task_cpu(p);
5365 if (curr_cpu == target_cpu)
5368 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5371 /* TODO: This is not properly updating schedstats */
5373 trace_sched_move_numa(p, curr_cpu, target_cpu);
5374 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5378 * Requeue a task on a given node and accurately track the number of NUMA
5379 * tasks on the runqueues
5381 void sched_setnuma(struct task_struct *p, int nid)
5383 bool queued, running;
5387 rq = task_rq_lock(p, &rf);
5388 queued = task_on_rq_queued(p);
5389 running = task_current(rq, p);
5392 dequeue_task(rq, p, DEQUEUE_SAVE);
5394 put_prev_task(rq, p);
5396 p->numa_preferred_nid = nid;
5399 p->sched_class->set_curr_task(rq);
5401 enqueue_task(rq, p, ENQUEUE_RESTORE);
5402 task_rq_unlock(rq, p, &rf);
5404 #endif /* CONFIG_NUMA_BALANCING */
5406 #ifdef CONFIG_HOTPLUG_CPU
5408 * Ensures that the idle task is using init_mm right before its cpu goes
5411 void idle_task_exit(void)
5413 struct mm_struct *mm = current->active_mm;
5415 BUG_ON(cpu_online(smp_processor_id()));
5417 if (mm != &init_mm) {
5418 switch_mm_irqs_off(mm, &init_mm, current);
5419 finish_arch_post_lock_switch();
5425 * Since this CPU is going 'away' for a while, fold any nr_active delta
5426 * we might have. Assumes we're called after migrate_tasks() so that the
5427 * nr_active count is stable. We need to take the teardown thread which
5428 * is calling this into account, so we hand in adjust = 1 to the load
5431 * Also see the comment "Global load-average calculations".
5433 static void calc_load_migrate(struct rq *rq)
5435 long delta = calc_load_fold_active(rq, 1);
5437 atomic_long_add(delta, &calc_load_tasks);
5440 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5444 static const struct sched_class fake_sched_class = {
5445 .put_prev_task = put_prev_task_fake,
5448 static struct task_struct fake_task = {
5450 * Avoid pull_{rt,dl}_task()
5452 .prio = MAX_PRIO + 1,
5453 .sched_class = &fake_sched_class,
5457 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5458 * try_to_wake_up()->select_task_rq().
5460 * Called with rq->lock held even though we'er in stop_machine() and
5461 * there's no concurrency possible, we hold the required locks anyway
5462 * because of lock validation efforts.
5464 static void migrate_tasks(struct rq *dead_rq)
5466 struct rq *rq = dead_rq;
5467 struct task_struct *next, *stop = rq->stop;
5468 struct pin_cookie cookie;
5472 * Fudge the rq selection such that the below task selection loop
5473 * doesn't get stuck on the currently eligible stop task.
5475 * We're currently inside stop_machine() and the rq is either stuck
5476 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5477 * either way we should never end up calling schedule() until we're
5483 * put_prev_task() and pick_next_task() sched
5484 * class method both need to have an up-to-date
5485 * value of rq->clock[_task]
5487 update_rq_clock(rq);
5491 * There's this thread running, bail when that's the only
5494 if (rq->nr_running == 1)
5498 * pick_next_task assumes pinned rq->lock.
5500 cookie = lockdep_pin_lock(&rq->lock);
5501 next = pick_next_task(rq, &fake_task, cookie);
5503 next->sched_class->put_prev_task(rq, next);
5506 * Rules for changing task_struct::cpus_allowed are holding
5507 * both pi_lock and rq->lock, such that holding either
5508 * stabilizes the mask.
5510 * Drop rq->lock is not quite as disastrous as it usually is
5511 * because !cpu_active at this point, which means load-balance
5512 * will not interfere. Also, stop-machine.
5514 lockdep_unpin_lock(&rq->lock, cookie);
5515 raw_spin_unlock(&rq->lock);
5516 raw_spin_lock(&next->pi_lock);
5517 raw_spin_lock(&rq->lock);
5520 * Since we're inside stop-machine, _nothing_ should have
5521 * changed the task, WARN if weird stuff happened, because in
5522 * that case the above rq->lock drop is a fail too.
5524 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5525 raw_spin_unlock(&next->pi_lock);
5529 /* Find suitable destination for @next, with force if needed. */
5530 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5532 rq = __migrate_task(rq, next, dest_cpu);
5533 if (rq != dead_rq) {
5534 raw_spin_unlock(&rq->lock);
5536 raw_spin_lock(&rq->lock);
5538 raw_spin_unlock(&next->pi_lock);
5543 #endif /* CONFIG_HOTPLUG_CPU */
5545 static void set_rq_online(struct rq *rq)
5548 const struct sched_class *class;
5550 cpumask_set_cpu(rq->cpu, rq->rd->online);
5553 for_each_class(class) {
5554 if (class->rq_online)
5555 class->rq_online(rq);
5560 static void set_rq_offline(struct rq *rq)
5563 const struct sched_class *class;
5565 for_each_class(class) {
5566 if (class->rq_offline)
5567 class->rq_offline(rq);
5570 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5575 static void set_cpu_rq_start_time(unsigned int cpu)
5577 struct rq *rq = cpu_rq(cpu);
5579 rq->age_stamp = sched_clock_cpu(cpu);
5582 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5584 #ifdef CONFIG_SCHED_DEBUG
5586 static __read_mostly int sched_debug_enabled;
5588 static int __init sched_debug_setup(char *str)
5590 sched_debug_enabled = 1;
5594 early_param("sched_debug", sched_debug_setup);
5596 static inline bool sched_debug(void)
5598 return sched_debug_enabled;
5601 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5602 struct cpumask *groupmask)
5604 struct sched_group *group = sd->groups;
5606 cpumask_clear(groupmask);
5608 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5610 if (!(sd->flags & SD_LOAD_BALANCE)) {
5611 printk("does not load-balance\n");
5613 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5618 printk(KERN_CONT "span %*pbl level %s\n",
5619 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5621 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5622 printk(KERN_ERR "ERROR: domain->span does not contain "
5625 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5626 printk(KERN_ERR "ERROR: domain->groups does not contain"
5630 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5634 printk(KERN_ERR "ERROR: group is NULL\n");
5638 if (!cpumask_weight(sched_group_cpus(group))) {
5639 printk(KERN_CONT "\n");
5640 printk(KERN_ERR "ERROR: empty group\n");
5644 if (!(sd->flags & SD_OVERLAP) &&
5645 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5646 printk(KERN_CONT "\n");
5647 printk(KERN_ERR "ERROR: repeated CPUs\n");
5651 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5653 printk(KERN_CONT " %*pbl",
5654 cpumask_pr_args(sched_group_cpus(group)));
5655 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5656 printk(KERN_CONT " (cpu_capacity = %d)",
5657 group->sgc->capacity);
5660 group = group->next;
5661 } while (group != sd->groups);
5662 printk(KERN_CONT "\n");
5664 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5665 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5668 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5669 printk(KERN_ERR "ERROR: parent span is not a superset "
5670 "of domain->span\n");
5674 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5678 if (!sched_debug_enabled)
5682 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5686 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5689 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5697 #else /* !CONFIG_SCHED_DEBUG */
5698 # define sched_domain_debug(sd, cpu) do { } while (0)
5699 static inline bool sched_debug(void)
5703 #endif /* CONFIG_SCHED_DEBUG */
5705 static int sd_degenerate(struct sched_domain *sd)
5707 if (cpumask_weight(sched_domain_span(sd)) == 1)
5710 /* Following flags need at least 2 groups */
5711 if (sd->flags & (SD_LOAD_BALANCE |
5712 SD_BALANCE_NEWIDLE |
5715 SD_SHARE_CPUCAPACITY |
5716 SD_SHARE_PKG_RESOURCES |
5717 SD_SHARE_POWERDOMAIN)) {
5718 if (sd->groups != sd->groups->next)
5722 /* Following flags don't use groups */
5723 if (sd->flags & (SD_WAKE_AFFINE))
5730 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5732 unsigned long cflags = sd->flags, pflags = parent->flags;
5734 if (sd_degenerate(parent))
5737 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5740 /* Flags needing groups don't count if only 1 group in parent */
5741 if (parent->groups == parent->groups->next) {
5742 pflags &= ~(SD_LOAD_BALANCE |
5743 SD_BALANCE_NEWIDLE |
5746 SD_SHARE_CPUCAPACITY |
5747 SD_SHARE_PKG_RESOURCES |
5749 SD_SHARE_POWERDOMAIN);
5750 if (nr_node_ids == 1)
5751 pflags &= ~SD_SERIALIZE;
5753 if (~cflags & pflags)
5759 static void free_rootdomain(struct rcu_head *rcu)
5761 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5763 cpupri_cleanup(&rd->cpupri);
5764 cpudl_cleanup(&rd->cpudl);
5765 free_cpumask_var(rd->dlo_mask);
5766 free_cpumask_var(rd->rto_mask);
5767 free_cpumask_var(rd->online);
5768 free_cpumask_var(rd->span);
5772 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5774 struct root_domain *old_rd = NULL;
5775 unsigned long flags;
5777 raw_spin_lock_irqsave(&rq->lock, flags);
5782 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5785 cpumask_clear_cpu(rq->cpu, old_rd->span);
5788 * If we dont want to free the old_rd yet then
5789 * set old_rd to NULL to skip the freeing later
5792 if (!atomic_dec_and_test(&old_rd->refcount))
5796 atomic_inc(&rd->refcount);
5799 cpumask_set_cpu(rq->cpu, rd->span);
5800 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5803 raw_spin_unlock_irqrestore(&rq->lock, flags);
5806 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5809 static int init_rootdomain(struct root_domain *rd)
5811 memset(rd, 0, sizeof(*rd));
5813 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5815 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5817 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5819 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5822 init_dl_bw(&rd->dl_bw);
5823 if (cpudl_init(&rd->cpudl) != 0)
5826 if (cpupri_init(&rd->cpupri) != 0)
5831 free_cpumask_var(rd->rto_mask);
5833 free_cpumask_var(rd->dlo_mask);
5835 free_cpumask_var(rd->online);
5837 free_cpumask_var(rd->span);
5843 * By default the system creates a single root-domain with all cpus as
5844 * members (mimicking the global state we have today).
5846 struct root_domain def_root_domain;
5848 static void init_defrootdomain(void)
5850 init_rootdomain(&def_root_domain);
5852 atomic_set(&def_root_domain.refcount, 1);
5855 static struct root_domain *alloc_rootdomain(void)
5857 struct root_domain *rd;
5859 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5863 if (init_rootdomain(rd) != 0) {
5871 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5873 struct sched_group *tmp, *first;
5882 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5887 } while (sg != first);
5890 static void free_sched_domain(struct rcu_head *rcu)
5892 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5895 * If its an overlapping domain it has private groups, iterate and
5898 if (sd->flags & SD_OVERLAP) {
5899 free_sched_groups(sd->groups, 1);
5900 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5901 kfree(sd->groups->sgc);
5907 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5909 call_rcu(&sd->rcu, free_sched_domain);
5912 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5914 for (; sd; sd = sd->parent)
5915 destroy_sched_domain(sd, cpu);
5919 * Keep a special pointer to the highest sched_domain that has
5920 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5921 * allows us to avoid some pointer chasing select_idle_sibling().
5923 * Also keep a unique ID per domain (we use the first cpu number in
5924 * the cpumask of the domain), this allows us to quickly tell if
5925 * two cpus are in the same cache domain, see cpus_share_cache().
5927 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5928 DEFINE_PER_CPU(int, sd_llc_size);
5929 DEFINE_PER_CPU(int, sd_llc_id);
5930 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5931 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5932 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5934 static void update_top_cache_domain(int cpu)
5936 struct sched_domain *sd;
5937 struct sched_domain *busy_sd = NULL;
5941 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5943 id = cpumask_first(sched_domain_span(sd));
5944 size = cpumask_weight(sched_domain_span(sd));
5945 busy_sd = sd->parent; /* sd_busy */
5947 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5949 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5950 per_cpu(sd_llc_size, cpu) = size;
5951 per_cpu(sd_llc_id, cpu) = id;
5953 sd = lowest_flag_domain(cpu, SD_NUMA);
5954 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5956 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5957 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5961 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5962 * hold the hotplug lock.
5965 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5967 struct rq *rq = cpu_rq(cpu);
5968 struct sched_domain *tmp;
5970 /* Remove the sched domains which do not contribute to scheduling. */
5971 for (tmp = sd; tmp; ) {
5972 struct sched_domain *parent = tmp->parent;
5976 if (sd_parent_degenerate(tmp, parent)) {
5977 tmp->parent = parent->parent;
5979 parent->parent->child = tmp;
5981 * Transfer SD_PREFER_SIBLING down in case of a
5982 * degenerate parent; the spans match for this
5983 * so the property transfers.
5985 if (parent->flags & SD_PREFER_SIBLING)
5986 tmp->flags |= SD_PREFER_SIBLING;
5987 destroy_sched_domain(parent, cpu);
5992 if (sd && sd_degenerate(sd)) {
5995 destroy_sched_domain(tmp, cpu);
6000 sched_domain_debug(sd, cpu);
6002 rq_attach_root(rq, rd);
6004 rcu_assign_pointer(rq->sd, sd);
6005 destroy_sched_domains(tmp, cpu);
6007 update_top_cache_domain(cpu);
6010 /* Setup the mask of cpus configured for isolated domains */
6011 static int __init isolated_cpu_setup(char *str)
6015 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6016 ret = cpulist_parse(str, cpu_isolated_map);
6018 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6023 __setup("isolcpus=", isolated_cpu_setup);
6026 struct sched_domain ** __percpu sd;
6027 struct root_domain *rd;
6038 * Build an iteration mask that can exclude certain CPUs from the upwards
6041 * Asymmetric node setups can result in situations where the domain tree is of
6042 * unequal depth, make sure to skip domains that already cover the entire
6045 * In that case build_sched_domains() will have terminated the iteration early
6046 * and our sibling sd spans will be empty. Domains should always include the
6047 * cpu they're built on, so check that.
6050 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6052 const struct cpumask *span = sched_domain_span(sd);
6053 struct sd_data *sdd = sd->private;
6054 struct sched_domain *sibling;
6057 for_each_cpu(i, span) {
6058 sibling = *per_cpu_ptr(sdd->sd, i);
6059 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6062 cpumask_set_cpu(i, sched_group_mask(sg));
6067 * Return the canonical balance cpu for this group, this is the first cpu
6068 * of this group that's also in the iteration mask.
6070 int group_balance_cpu(struct sched_group *sg)
6072 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6076 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6078 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6079 const struct cpumask *span = sched_domain_span(sd);
6080 struct cpumask *covered = sched_domains_tmpmask;
6081 struct sd_data *sdd = sd->private;
6082 struct sched_domain *sibling;
6085 cpumask_clear(covered);
6087 for_each_cpu(i, span) {
6088 struct cpumask *sg_span;
6090 if (cpumask_test_cpu(i, covered))
6093 sibling = *per_cpu_ptr(sdd->sd, i);
6095 /* See the comment near build_group_mask(). */
6096 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6099 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6100 GFP_KERNEL, cpu_to_node(cpu));
6105 sg_span = sched_group_cpus(sg);
6107 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6109 cpumask_set_cpu(i, sg_span);
6111 cpumask_or(covered, covered, sg_span);
6113 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6114 if (atomic_inc_return(&sg->sgc->ref) == 1)
6115 build_group_mask(sd, sg);
6118 * Initialize sgc->capacity such that even if we mess up the
6119 * domains and no possible iteration will get us here, we won't
6122 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6125 * Make sure the first group of this domain contains the
6126 * canonical balance cpu. Otherwise the sched_domain iteration
6127 * breaks. See update_sg_lb_stats().
6129 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6130 group_balance_cpu(sg) == cpu)
6140 sd->groups = groups;
6145 free_sched_groups(first, 0);
6150 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6152 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6153 struct sched_domain *child = sd->child;
6156 cpu = cpumask_first(sched_domain_span(child));
6159 *sg = *per_cpu_ptr(sdd->sg, cpu);
6160 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6161 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6168 * build_sched_groups will build a circular linked list of the groups
6169 * covered by the given span, and will set each group's ->cpumask correctly,
6170 * and ->cpu_capacity to 0.
6172 * Assumes the sched_domain tree is fully constructed
6175 build_sched_groups(struct sched_domain *sd, int cpu)
6177 struct sched_group *first = NULL, *last = NULL;
6178 struct sd_data *sdd = sd->private;
6179 const struct cpumask *span = sched_domain_span(sd);
6180 struct cpumask *covered;
6183 get_group(cpu, sdd, &sd->groups);
6184 atomic_inc(&sd->groups->ref);
6186 if (cpu != cpumask_first(span))
6189 lockdep_assert_held(&sched_domains_mutex);
6190 covered = sched_domains_tmpmask;
6192 cpumask_clear(covered);
6194 for_each_cpu(i, span) {
6195 struct sched_group *sg;
6198 if (cpumask_test_cpu(i, covered))
6201 group = get_group(i, sdd, &sg);
6202 cpumask_setall(sched_group_mask(sg));
6204 for_each_cpu(j, span) {
6205 if (get_group(j, sdd, NULL) != group)
6208 cpumask_set_cpu(j, covered);
6209 cpumask_set_cpu(j, sched_group_cpus(sg));
6224 * Initialize sched groups cpu_capacity.
6226 * cpu_capacity indicates the capacity of sched group, which is used while
6227 * distributing the load between different sched groups in a sched domain.
6228 * Typically cpu_capacity for all the groups in a sched domain will be same
6229 * unless there are asymmetries in the topology. If there are asymmetries,
6230 * group having more cpu_capacity will pickup more load compared to the
6231 * group having less cpu_capacity.
6233 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6235 struct sched_group *sg = sd->groups;
6240 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6242 } while (sg != sd->groups);
6244 if (cpu != group_balance_cpu(sg))
6247 update_group_capacity(sd, cpu);
6248 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6252 * Initializers for schedule domains
6253 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6256 static int default_relax_domain_level = -1;
6257 int sched_domain_level_max;
6259 static int __init setup_relax_domain_level(char *str)
6261 if (kstrtoint(str, 0, &default_relax_domain_level))
6262 pr_warn("Unable to set relax_domain_level\n");
6266 __setup("relax_domain_level=", setup_relax_domain_level);
6268 static void set_domain_attribute(struct sched_domain *sd,
6269 struct sched_domain_attr *attr)
6273 if (!attr || attr->relax_domain_level < 0) {
6274 if (default_relax_domain_level < 0)
6277 request = default_relax_domain_level;
6279 request = attr->relax_domain_level;
6280 if (request < sd->level) {
6281 /* turn off idle balance on this domain */
6282 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6284 /* turn on idle balance on this domain */
6285 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6289 static void __sdt_free(const struct cpumask *cpu_map);
6290 static int __sdt_alloc(const struct cpumask *cpu_map);
6292 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6293 const struct cpumask *cpu_map)
6297 if (!atomic_read(&d->rd->refcount))
6298 free_rootdomain(&d->rd->rcu); /* fall through */
6300 free_percpu(d->sd); /* fall through */
6302 __sdt_free(cpu_map); /* fall through */
6308 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6309 const struct cpumask *cpu_map)
6311 memset(d, 0, sizeof(*d));
6313 if (__sdt_alloc(cpu_map))
6314 return sa_sd_storage;
6315 d->sd = alloc_percpu(struct sched_domain *);
6317 return sa_sd_storage;
6318 d->rd = alloc_rootdomain();
6321 return sa_rootdomain;
6325 * NULL the sd_data elements we've used to build the sched_domain and
6326 * sched_group structure so that the subsequent __free_domain_allocs()
6327 * will not free the data we're using.
6329 static void claim_allocations(int cpu, struct sched_domain *sd)
6331 struct sd_data *sdd = sd->private;
6333 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6334 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6336 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6337 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6339 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6340 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6344 static int sched_domains_numa_levels;
6345 enum numa_topology_type sched_numa_topology_type;
6346 static int *sched_domains_numa_distance;
6347 int sched_max_numa_distance;
6348 static struct cpumask ***sched_domains_numa_masks;
6349 static int sched_domains_curr_level;
6353 * SD_flags allowed in topology descriptions.
6355 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6356 * SD_SHARE_PKG_RESOURCES - describes shared caches
6357 * SD_NUMA - describes NUMA topologies
6358 * SD_SHARE_POWERDOMAIN - describes shared power domain
6361 * SD_ASYM_PACKING - describes SMT quirks
6363 #define TOPOLOGY_SD_FLAGS \
6364 (SD_SHARE_CPUCAPACITY | \
6365 SD_SHARE_PKG_RESOURCES | \
6368 SD_SHARE_POWERDOMAIN)
6370 static struct sched_domain *
6371 sd_init(struct sched_domain_topology_level *tl, int cpu)
6373 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6374 int sd_weight, sd_flags = 0;
6378 * Ugly hack to pass state to sd_numa_mask()...
6380 sched_domains_curr_level = tl->numa_level;
6383 sd_weight = cpumask_weight(tl->mask(cpu));
6386 sd_flags = (*tl->sd_flags)();
6387 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6388 "wrong sd_flags in topology description\n"))
6389 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6391 *sd = (struct sched_domain){
6392 .min_interval = sd_weight,
6393 .max_interval = 2*sd_weight,
6395 .imbalance_pct = 125,
6397 .cache_nice_tries = 0,
6404 .flags = 1*SD_LOAD_BALANCE
6405 | 1*SD_BALANCE_NEWIDLE
6410 | 0*SD_SHARE_CPUCAPACITY
6411 | 0*SD_SHARE_PKG_RESOURCES
6413 | 0*SD_PREFER_SIBLING
6418 .last_balance = jiffies,
6419 .balance_interval = sd_weight,
6421 .max_newidle_lb_cost = 0,
6422 .next_decay_max_lb_cost = jiffies,
6423 #ifdef CONFIG_SCHED_DEBUG
6429 * Convert topological properties into behaviour.
6432 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6433 sd->flags |= SD_PREFER_SIBLING;
6434 sd->imbalance_pct = 110;
6435 sd->smt_gain = 1178; /* ~15% */
6437 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6438 sd->imbalance_pct = 117;
6439 sd->cache_nice_tries = 1;
6443 } else if (sd->flags & SD_NUMA) {
6444 sd->cache_nice_tries = 2;
6448 sd->flags |= SD_SERIALIZE;
6449 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6450 sd->flags &= ~(SD_BALANCE_EXEC |
6457 sd->flags |= SD_PREFER_SIBLING;
6458 sd->cache_nice_tries = 1;
6463 sd->private = &tl->data;
6469 * Topology list, bottom-up.
6471 static struct sched_domain_topology_level default_topology[] = {
6472 #ifdef CONFIG_SCHED_SMT
6473 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6475 #ifdef CONFIG_SCHED_MC
6476 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6478 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6482 static struct sched_domain_topology_level *sched_domain_topology =
6485 #define for_each_sd_topology(tl) \
6486 for (tl = sched_domain_topology; tl->mask; tl++)
6488 void set_sched_topology(struct sched_domain_topology_level *tl)
6490 sched_domain_topology = tl;
6495 static const struct cpumask *sd_numa_mask(int cpu)
6497 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6500 static void sched_numa_warn(const char *str)
6502 static int done = false;
6510 printk(KERN_WARNING "ERROR: %s\n\n", str);
6512 for (i = 0; i < nr_node_ids; i++) {
6513 printk(KERN_WARNING " ");
6514 for (j = 0; j < nr_node_ids; j++)
6515 printk(KERN_CONT "%02d ", node_distance(i,j));
6516 printk(KERN_CONT "\n");
6518 printk(KERN_WARNING "\n");
6521 bool find_numa_distance(int distance)
6525 if (distance == node_distance(0, 0))
6528 for (i = 0; i < sched_domains_numa_levels; i++) {
6529 if (sched_domains_numa_distance[i] == distance)
6537 * A system can have three types of NUMA topology:
6538 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6539 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6540 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6542 * The difference between a glueless mesh topology and a backplane
6543 * topology lies in whether communication between not directly
6544 * connected nodes goes through intermediary nodes (where programs
6545 * could run), or through backplane controllers. This affects
6546 * placement of programs.
6548 * The type of topology can be discerned with the following tests:
6549 * - If the maximum distance between any nodes is 1 hop, the system
6550 * is directly connected.
6551 * - If for two nodes A and B, located N > 1 hops away from each other,
6552 * there is an intermediary node C, which is < N hops away from both
6553 * nodes A and B, the system is a glueless mesh.
6555 static void init_numa_topology_type(void)
6559 n = sched_max_numa_distance;
6561 if (sched_domains_numa_levels <= 1) {
6562 sched_numa_topology_type = NUMA_DIRECT;
6566 for_each_online_node(a) {
6567 for_each_online_node(b) {
6568 /* Find two nodes furthest removed from each other. */
6569 if (node_distance(a, b) < n)
6572 /* Is there an intermediary node between a and b? */
6573 for_each_online_node(c) {
6574 if (node_distance(a, c) < n &&
6575 node_distance(b, c) < n) {
6576 sched_numa_topology_type =
6582 sched_numa_topology_type = NUMA_BACKPLANE;
6588 static void sched_init_numa(void)
6590 int next_distance, curr_distance = node_distance(0, 0);
6591 struct sched_domain_topology_level *tl;
6595 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6596 if (!sched_domains_numa_distance)
6600 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6601 * unique distances in the node_distance() table.
6603 * Assumes node_distance(0,j) includes all distances in
6604 * node_distance(i,j) in order to avoid cubic time.
6606 next_distance = curr_distance;
6607 for (i = 0; i < nr_node_ids; i++) {
6608 for (j = 0; j < nr_node_ids; j++) {
6609 for (k = 0; k < nr_node_ids; k++) {
6610 int distance = node_distance(i, k);
6612 if (distance > curr_distance &&
6613 (distance < next_distance ||
6614 next_distance == curr_distance))
6615 next_distance = distance;
6618 * While not a strong assumption it would be nice to know
6619 * about cases where if node A is connected to B, B is not
6620 * equally connected to A.
6622 if (sched_debug() && node_distance(k, i) != distance)
6623 sched_numa_warn("Node-distance not symmetric");
6625 if (sched_debug() && i && !find_numa_distance(distance))
6626 sched_numa_warn("Node-0 not representative");
6628 if (next_distance != curr_distance) {
6629 sched_domains_numa_distance[level++] = next_distance;
6630 sched_domains_numa_levels = level;
6631 curr_distance = next_distance;
6636 * In case of sched_debug() we verify the above assumption.
6646 * 'level' contains the number of unique distances, excluding the
6647 * identity distance node_distance(i,i).
6649 * The sched_domains_numa_distance[] array includes the actual distance
6654 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6655 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6656 * the array will contain less then 'level' members. This could be
6657 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6658 * in other functions.
6660 * We reset it to 'level' at the end of this function.
6662 sched_domains_numa_levels = 0;
6664 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6665 if (!sched_domains_numa_masks)
6669 * Now for each level, construct a mask per node which contains all
6670 * cpus of nodes that are that many hops away from us.
6672 for (i = 0; i < level; i++) {
6673 sched_domains_numa_masks[i] =
6674 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6675 if (!sched_domains_numa_masks[i])
6678 for (j = 0; j < nr_node_ids; j++) {
6679 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6683 sched_domains_numa_masks[i][j] = mask;
6686 if (node_distance(j, k) > sched_domains_numa_distance[i])
6689 cpumask_or(mask, mask, cpumask_of_node(k));
6694 /* Compute default topology size */
6695 for (i = 0; sched_domain_topology[i].mask; i++);
6697 tl = kzalloc((i + level + 1) *
6698 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6703 * Copy the default topology bits..
6705 for (i = 0; sched_domain_topology[i].mask; i++)
6706 tl[i] = sched_domain_topology[i];
6709 * .. and append 'j' levels of NUMA goodness.
6711 for (j = 0; j < level; i++, j++) {
6712 tl[i] = (struct sched_domain_topology_level){
6713 .mask = sd_numa_mask,
6714 .sd_flags = cpu_numa_flags,
6715 .flags = SDTL_OVERLAP,
6721 sched_domain_topology = tl;
6723 sched_domains_numa_levels = level;
6724 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6726 init_numa_topology_type();
6729 static void sched_domains_numa_masks_set(unsigned int cpu)
6731 int node = cpu_to_node(cpu);
6734 for (i = 0; i < sched_domains_numa_levels; i++) {
6735 for (j = 0; j < nr_node_ids; j++) {
6736 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6737 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6742 static void sched_domains_numa_masks_clear(unsigned int cpu)
6746 for (i = 0; i < sched_domains_numa_levels; i++) {
6747 for (j = 0; j < nr_node_ids; j++)
6748 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6753 static inline void sched_init_numa(void) { }
6754 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6755 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6756 #endif /* CONFIG_NUMA */
6758 static int __sdt_alloc(const struct cpumask *cpu_map)
6760 struct sched_domain_topology_level *tl;
6763 for_each_sd_topology(tl) {
6764 struct sd_data *sdd = &tl->data;
6766 sdd->sd = alloc_percpu(struct sched_domain *);
6770 sdd->sg = alloc_percpu(struct sched_group *);
6774 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6778 for_each_cpu(j, cpu_map) {
6779 struct sched_domain *sd;
6780 struct sched_group *sg;
6781 struct sched_group_capacity *sgc;
6783 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6784 GFP_KERNEL, cpu_to_node(j));
6788 *per_cpu_ptr(sdd->sd, j) = sd;
6790 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6791 GFP_KERNEL, cpu_to_node(j));
6797 *per_cpu_ptr(sdd->sg, j) = sg;
6799 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6800 GFP_KERNEL, cpu_to_node(j));
6804 *per_cpu_ptr(sdd->sgc, j) = sgc;
6811 static void __sdt_free(const struct cpumask *cpu_map)
6813 struct sched_domain_topology_level *tl;
6816 for_each_sd_topology(tl) {
6817 struct sd_data *sdd = &tl->data;
6819 for_each_cpu(j, cpu_map) {
6820 struct sched_domain *sd;
6823 sd = *per_cpu_ptr(sdd->sd, j);
6824 if (sd && (sd->flags & SD_OVERLAP))
6825 free_sched_groups(sd->groups, 0);
6826 kfree(*per_cpu_ptr(sdd->sd, j));
6830 kfree(*per_cpu_ptr(sdd->sg, j));
6832 kfree(*per_cpu_ptr(sdd->sgc, j));
6834 free_percpu(sdd->sd);
6836 free_percpu(sdd->sg);
6838 free_percpu(sdd->sgc);
6843 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6844 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6845 struct sched_domain *child, int cpu)
6847 struct sched_domain *sd = sd_init(tl, cpu);
6851 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6853 sd->level = child->level + 1;
6854 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6858 if (!cpumask_subset(sched_domain_span(child),
6859 sched_domain_span(sd))) {
6860 pr_err("BUG: arch topology borken\n");
6861 #ifdef CONFIG_SCHED_DEBUG
6862 pr_err(" the %s domain not a subset of the %s domain\n",
6863 child->name, sd->name);
6865 /* Fixup, ensure @sd has at least @child cpus. */
6866 cpumask_or(sched_domain_span(sd),
6867 sched_domain_span(sd),
6868 sched_domain_span(child));
6872 set_domain_attribute(sd, attr);
6878 * Build sched domains for a given set of cpus and attach the sched domains
6879 * to the individual cpus
6881 static int build_sched_domains(const struct cpumask *cpu_map,
6882 struct sched_domain_attr *attr)
6884 enum s_alloc alloc_state;
6885 struct sched_domain *sd;
6887 int i, ret = -ENOMEM;
6889 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6890 if (alloc_state != sa_rootdomain)
6893 /* Set up domains for cpus specified by the cpu_map. */
6894 for_each_cpu(i, cpu_map) {
6895 struct sched_domain_topology_level *tl;
6898 for_each_sd_topology(tl) {
6899 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6900 if (tl == sched_domain_topology)
6901 *per_cpu_ptr(d.sd, i) = sd;
6902 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6903 sd->flags |= SD_OVERLAP;
6904 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6909 /* Build the groups for the domains */
6910 for_each_cpu(i, cpu_map) {
6911 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6912 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6913 if (sd->flags & SD_OVERLAP) {
6914 if (build_overlap_sched_groups(sd, i))
6917 if (build_sched_groups(sd, i))
6923 /* Calculate CPU capacity for physical packages and nodes */
6924 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6925 if (!cpumask_test_cpu(i, cpu_map))
6928 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6929 claim_allocations(i, sd);
6930 init_sched_groups_capacity(i, sd);
6934 /* Attach the domains */
6936 for_each_cpu(i, cpu_map) {
6937 sd = *per_cpu_ptr(d.sd, i);
6938 cpu_attach_domain(sd, d.rd, i);
6944 __free_domain_allocs(&d, alloc_state, cpu_map);
6948 static cpumask_var_t *doms_cur; /* current sched domains */
6949 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6950 static struct sched_domain_attr *dattr_cur;
6951 /* attribues of custom domains in 'doms_cur' */
6954 * Special case: If a kmalloc of a doms_cur partition (array of
6955 * cpumask) fails, then fallback to a single sched domain,
6956 * as determined by the single cpumask fallback_doms.
6958 static cpumask_var_t fallback_doms;
6961 * arch_update_cpu_topology lets virtualized architectures update the
6962 * cpu core maps. It is supposed to return 1 if the topology changed
6963 * or 0 if it stayed the same.
6965 int __weak arch_update_cpu_topology(void)
6970 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6973 cpumask_var_t *doms;
6975 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6978 for (i = 0; i < ndoms; i++) {
6979 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6980 free_sched_domains(doms, i);
6987 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6990 for (i = 0; i < ndoms; i++)
6991 free_cpumask_var(doms[i]);
6996 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6997 * For now this just excludes isolated cpus, but could be used to
6998 * exclude other special cases in the future.
7000 static int init_sched_domains(const struct cpumask *cpu_map)
7004 arch_update_cpu_topology();
7006 doms_cur = alloc_sched_domains(ndoms_cur);
7008 doms_cur = &fallback_doms;
7009 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7010 err = build_sched_domains(doms_cur[0], NULL);
7011 register_sched_domain_sysctl();
7017 * Detach sched domains from a group of cpus specified in cpu_map
7018 * These cpus will now be attached to the NULL domain
7020 static void detach_destroy_domains(const struct cpumask *cpu_map)
7025 for_each_cpu(i, cpu_map)
7026 cpu_attach_domain(NULL, &def_root_domain, i);
7030 /* handle null as "default" */
7031 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7032 struct sched_domain_attr *new, int idx_new)
7034 struct sched_domain_attr tmp;
7041 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7042 new ? (new + idx_new) : &tmp,
7043 sizeof(struct sched_domain_attr));
7047 * Partition sched domains as specified by the 'ndoms_new'
7048 * cpumasks in the array doms_new[] of cpumasks. This compares
7049 * doms_new[] to the current sched domain partitioning, doms_cur[].
7050 * It destroys each deleted domain and builds each new domain.
7052 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7053 * The masks don't intersect (don't overlap.) We should setup one
7054 * sched domain for each mask. CPUs not in any of the cpumasks will
7055 * not be load balanced. If the same cpumask appears both in the
7056 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7059 * The passed in 'doms_new' should be allocated using
7060 * alloc_sched_domains. This routine takes ownership of it and will
7061 * free_sched_domains it when done with it. If the caller failed the
7062 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7063 * and partition_sched_domains() will fallback to the single partition
7064 * 'fallback_doms', it also forces the domains to be rebuilt.
7066 * If doms_new == NULL it will be replaced with cpu_online_mask.
7067 * ndoms_new == 0 is a special case for destroying existing domains,
7068 * and it will not create the default domain.
7070 * Call with hotplug lock held
7072 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7073 struct sched_domain_attr *dattr_new)
7078 mutex_lock(&sched_domains_mutex);
7080 /* always unregister in case we don't destroy any domains */
7081 unregister_sched_domain_sysctl();
7083 /* Let architecture update cpu core mappings. */
7084 new_topology = arch_update_cpu_topology();
7086 n = doms_new ? ndoms_new : 0;
7088 /* Destroy deleted domains */
7089 for (i = 0; i < ndoms_cur; i++) {
7090 for (j = 0; j < n && !new_topology; j++) {
7091 if (cpumask_equal(doms_cur[i], doms_new[j])
7092 && dattrs_equal(dattr_cur, i, dattr_new, j))
7095 /* no match - a current sched domain not in new doms_new[] */
7096 detach_destroy_domains(doms_cur[i]);
7102 if (doms_new == NULL) {
7104 doms_new = &fallback_doms;
7105 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7106 WARN_ON_ONCE(dattr_new);
7109 /* Build new domains */
7110 for (i = 0; i < ndoms_new; i++) {
7111 for (j = 0; j < n && !new_topology; j++) {
7112 if (cpumask_equal(doms_new[i], doms_cur[j])
7113 && dattrs_equal(dattr_new, i, dattr_cur, j))
7116 /* no match - add a new doms_new */
7117 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7122 /* Remember the new sched domains */
7123 if (doms_cur != &fallback_doms)
7124 free_sched_domains(doms_cur, ndoms_cur);
7125 kfree(dattr_cur); /* kfree(NULL) is safe */
7126 doms_cur = doms_new;
7127 dattr_cur = dattr_new;
7128 ndoms_cur = ndoms_new;
7130 register_sched_domain_sysctl();
7132 mutex_unlock(&sched_domains_mutex);
7135 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7138 * Update cpusets according to cpu_active mask. If cpusets are
7139 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7140 * around partition_sched_domains().
7142 * If we come here as part of a suspend/resume, don't touch cpusets because we
7143 * want to restore it back to its original state upon resume anyway.
7145 static void cpuset_cpu_active(void)
7147 if (cpuhp_tasks_frozen) {
7149 * num_cpus_frozen tracks how many CPUs are involved in suspend
7150 * resume sequence. As long as this is not the last online
7151 * operation in the resume sequence, just build a single sched
7152 * domain, ignoring cpusets.
7155 if (likely(num_cpus_frozen)) {
7156 partition_sched_domains(1, NULL, NULL);
7160 * This is the last CPU online operation. So fall through and
7161 * restore the original sched domains by considering the
7162 * cpuset configurations.
7165 cpuset_update_active_cpus(true);
7168 static int cpuset_cpu_inactive(unsigned int cpu)
7170 unsigned long flags;
7175 if (!cpuhp_tasks_frozen) {
7176 rcu_read_lock_sched();
7177 dl_b = dl_bw_of(cpu);
7179 raw_spin_lock_irqsave(&dl_b->lock, flags);
7180 cpus = dl_bw_cpus(cpu);
7181 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7182 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7184 rcu_read_unlock_sched();
7188 cpuset_update_active_cpus(false);
7191 partition_sched_domains(1, NULL, NULL);
7196 int sched_cpu_activate(unsigned int cpu)
7198 struct rq *rq = cpu_rq(cpu);
7199 unsigned long flags;
7201 set_cpu_active(cpu, true);
7203 if (sched_smp_initialized) {
7204 sched_domains_numa_masks_set(cpu);
7205 cpuset_cpu_active();
7209 * Put the rq online, if not already. This happens:
7211 * 1) In the early boot process, because we build the real domains
7212 * after all cpus have been brought up.
7214 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7217 raw_spin_lock_irqsave(&rq->lock, flags);
7219 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7222 raw_spin_unlock_irqrestore(&rq->lock, flags);
7224 update_max_interval();
7229 int sched_cpu_deactivate(unsigned int cpu)
7233 set_cpu_active(cpu, false);
7235 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7236 * users of this state to go away such that all new such users will
7239 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7240 * not imply sync_sched(), so wait for both.
7242 * Do sync before park smpboot threads to take care the rcu boost case.
7244 if (IS_ENABLED(CONFIG_PREEMPT))
7245 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7249 if (!sched_smp_initialized)
7252 ret = cpuset_cpu_inactive(cpu);
7254 set_cpu_active(cpu, true);
7257 sched_domains_numa_masks_clear(cpu);
7261 static void sched_rq_cpu_starting(unsigned int cpu)
7263 struct rq *rq = cpu_rq(cpu);
7265 rq->calc_load_update = calc_load_update;
7266 update_max_interval();
7269 int sched_cpu_starting(unsigned int cpu)
7271 set_cpu_rq_start_time(cpu);
7272 sched_rq_cpu_starting(cpu);
7276 #ifdef CONFIG_HOTPLUG_CPU
7277 int sched_cpu_dying(unsigned int cpu)
7279 struct rq *rq = cpu_rq(cpu);
7280 unsigned long flags;
7282 /* Handle pending wakeups and then migrate everything off */
7283 sched_ttwu_pending();
7284 raw_spin_lock_irqsave(&rq->lock, flags);
7286 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7290 BUG_ON(rq->nr_running != 1);
7291 raw_spin_unlock_irqrestore(&rq->lock, flags);
7292 calc_load_migrate(rq);
7293 update_max_interval();
7294 nohz_balance_exit_idle(cpu);
7300 void __init sched_init_smp(void)
7302 cpumask_var_t non_isolated_cpus;
7304 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7305 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7310 * There's no userspace yet to cause hotplug operations; hence all the
7311 * cpu masks are stable and all blatant races in the below code cannot
7314 mutex_lock(&sched_domains_mutex);
7315 init_sched_domains(cpu_active_mask);
7316 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7317 if (cpumask_empty(non_isolated_cpus))
7318 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7319 mutex_unlock(&sched_domains_mutex);
7321 /* Move init over to a non-isolated CPU */
7322 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7324 sched_init_granularity();
7325 free_cpumask_var(non_isolated_cpus);
7327 init_sched_rt_class();
7328 init_sched_dl_class();
7329 sched_smp_initialized = true;
7332 static int __init migration_init(void)
7334 sched_rq_cpu_starting(smp_processor_id());
7337 early_initcall(migration_init);
7340 void __init sched_init_smp(void)
7342 sched_init_granularity();
7344 #endif /* CONFIG_SMP */
7346 int in_sched_functions(unsigned long addr)
7348 return in_lock_functions(addr) ||
7349 (addr >= (unsigned long)__sched_text_start
7350 && addr < (unsigned long)__sched_text_end);
7353 #ifdef CONFIG_CGROUP_SCHED
7355 * Default task group.
7356 * Every task in system belongs to this group at bootup.
7358 struct task_group root_task_group;
7359 LIST_HEAD(task_groups);
7361 /* Cacheline aligned slab cache for task_group */
7362 static struct kmem_cache *task_group_cache __read_mostly;
7365 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7367 void __init sched_init(void)
7370 unsigned long alloc_size = 0, ptr;
7372 #ifdef CONFIG_FAIR_GROUP_SCHED
7373 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7375 #ifdef CONFIG_RT_GROUP_SCHED
7376 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7379 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7381 #ifdef CONFIG_FAIR_GROUP_SCHED
7382 root_task_group.se = (struct sched_entity **)ptr;
7383 ptr += nr_cpu_ids * sizeof(void **);
7385 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7386 ptr += nr_cpu_ids * sizeof(void **);
7388 #endif /* CONFIG_FAIR_GROUP_SCHED */
7389 #ifdef CONFIG_RT_GROUP_SCHED
7390 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7391 ptr += nr_cpu_ids * sizeof(void **);
7393 root_task_group.rt_rq = (struct rt_rq **)ptr;
7394 ptr += nr_cpu_ids * sizeof(void **);
7396 #endif /* CONFIG_RT_GROUP_SCHED */
7398 #ifdef CONFIG_CPUMASK_OFFSTACK
7399 for_each_possible_cpu(i) {
7400 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7401 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7403 #endif /* CONFIG_CPUMASK_OFFSTACK */
7405 init_rt_bandwidth(&def_rt_bandwidth,
7406 global_rt_period(), global_rt_runtime());
7407 init_dl_bandwidth(&def_dl_bandwidth,
7408 global_rt_period(), global_rt_runtime());
7411 init_defrootdomain();
7414 #ifdef CONFIG_RT_GROUP_SCHED
7415 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7416 global_rt_period(), global_rt_runtime());
7417 #endif /* CONFIG_RT_GROUP_SCHED */
7419 #ifdef CONFIG_CGROUP_SCHED
7420 task_group_cache = KMEM_CACHE(task_group, 0);
7422 list_add(&root_task_group.list, &task_groups);
7423 INIT_LIST_HEAD(&root_task_group.children);
7424 INIT_LIST_HEAD(&root_task_group.siblings);
7425 autogroup_init(&init_task);
7426 #endif /* CONFIG_CGROUP_SCHED */
7428 for_each_possible_cpu(i) {
7432 raw_spin_lock_init(&rq->lock);
7434 rq->calc_load_active = 0;
7435 rq->calc_load_update = jiffies + LOAD_FREQ;
7436 init_cfs_rq(&rq->cfs);
7437 init_rt_rq(&rq->rt);
7438 init_dl_rq(&rq->dl);
7439 #ifdef CONFIG_FAIR_GROUP_SCHED
7440 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7441 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7443 * How much cpu bandwidth does root_task_group get?
7445 * In case of task-groups formed thr' the cgroup filesystem, it
7446 * gets 100% of the cpu resources in the system. This overall
7447 * system cpu resource is divided among the tasks of
7448 * root_task_group and its child task-groups in a fair manner,
7449 * based on each entity's (task or task-group's) weight
7450 * (se->load.weight).
7452 * In other words, if root_task_group has 10 tasks of weight
7453 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7454 * then A0's share of the cpu resource is:
7456 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7458 * We achieve this by letting root_task_group's tasks sit
7459 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7461 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7462 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7463 #endif /* CONFIG_FAIR_GROUP_SCHED */
7465 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7466 #ifdef CONFIG_RT_GROUP_SCHED
7467 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7470 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7471 rq->cpu_load[j] = 0;
7476 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7477 rq->balance_callback = NULL;
7478 rq->active_balance = 0;
7479 rq->next_balance = jiffies;
7484 rq->avg_idle = 2*sysctl_sched_migration_cost;
7485 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7487 INIT_LIST_HEAD(&rq->cfs_tasks);
7489 rq_attach_root(rq, &def_root_domain);
7490 #ifdef CONFIG_NO_HZ_COMMON
7491 rq->last_load_update_tick = jiffies;
7494 #ifdef CONFIG_NO_HZ_FULL
7495 rq->last_sched_tick = 0;
7497 #endif /* CONFIG_SMP */
7499 atomic_set(&rq->nr_iowait, 0);
7502 set_load_weight(&init_task);
7504 #ifdef CONFIG_PREEMPT_NOTIFIERS
7505 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7509 * The boot idle thread does lazy MMU switching as well:
7511 atomic_inc(&init_mm.mm_count);
7512 enter_lazy_tlb(&init_mm, current);
7515 * During early bootup we pretend to be a normal task:
7517 current->sched_class = &fair_sched_class;
7520 * Make us the idle thread. Technically, schedule() should not be
7521 * called from this thread, however somewhere below it might be,
7522 * but because we are the idle thread, we just pick up running again
7523 * when this runqueue becomes "idle".
7525 init_idle(current, smp_processor_id());
7527 calc_load_update = jiffies + LOAD_FREQ;
7530 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7531 /* May be allocated at isolcpus cmdline parse time */
7532 if (cpu_isolated_map == NULL)
7533 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7534 idle_thread_set_boot_cpu();
7535 set_cpu_rq_start_time(smp_processor_id());
7537 init_sched_fair_class();
7541 scheduler_running = 1;
7544 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7545 static inline int preempt_count_equals(int preempt_offset)
7547 int nested = preempt_count() + rcu_preempt_depth();
7549 return (nested == preempt_offset);
7552 void __might_sleep(const char *file, int line, int preempt_offset)
7555 * Blocking primitives will set (and therefore destroy) current->state,
7556 * since we will exit with TASK_RUNNING make sure we enter with it,
7557 * otherwise we will destroy state.
7559 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7560 "do not call blocking ops when !TASK_RUNNING; "
7561 "state=%lx set at [<%p>] %pS\n",
7563 (void *)current->task_state_change,
7564 (void *)current->task_state_change);
7566 ___might_sleep(file, line, preempt_offset);
7568 EXPORT_SYMBOL(__might_sleep);
7570 void ___might_sleep(const char *file, int line, int preempt_offset)
7572 static unsigned long prev_jiffy; /* ratelimiting */
7574 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7575 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7576 !is_idle_task(current)) ||
7577 system_state != SYSTEM_RUNNING || oops_in_progress)
7579 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7581 prev_jiffy = jiffies;
7584 "BUG: sleeping function called from invalid context at %s:%d\n",
7587 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7588 in_atomic(), irqs_disabled(),
7589 current->pid, current->comm);
7591 if (task_stack_end_corrupted(current))
7592 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7594 debug_show_held_locks(current);
7595 if (irqs_disabled())
7596 print_irqtrace_events(current);
7597 #ifdef CONFIG_DEBUG_PREEMPT
7598 if (!preempt_count_equals(preempt_offset)) {
7599 pr_err("Preemption disabled at:");
7600 print_ip_sym(current->preempt_disable_ip);
7606 EXPORT_SYMBOL(___might_sleep);
7609 #ifdef CONFIG_MAGIC_SYSRQ
7610 void normalize_rt_tasks(void)
7612 struct task_struct *g, *p;
7613 struct sched_attr attr = {
7614 .sched_policy = SCHED_NORMAL,
7617 read_lock(&tasklist_lock);
7618 for_each_process_thread(g, p) {
7620 * Only normalize user tasks:
7622 if (p->flags & PF_KTHREAD)
7625 p->se.exec_start = 0;
7626 #ifdef CONFIG_SCHEDSTATS
7627 p->se.statistics.wait_start = 0;
7628 p->se.statistics.sleep_start = 0;
7629 p->se.statistics.block_start = 0;
7632 if (!dl_task(p) && !rt_task(p)) {
7634 * Renice negative nice level userspace
7637 if (task_nice(p) < 0)
7638 set_user_nice(p, 0);
7642 __sched_setscheduler(p, &attr, false, false);
7644 read_unlock(&tasklist_lock);
7647 #endif /* CONFIG_MAGIC_SYSRQ */
7649 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7651 * These functions are only useful for the IA64 MCA handling, or kdb.
7653 * They can only be called when the whole system has been
7654 * stopped - every CPU needs to be quiescent, and no scheduling
7655 * activity can take place. Using them for anything else would
7656 * be a serious bug, and as a result, they aren't even visible
7657 * under any other configuration.
7661 * curr_task - return the current task for a given cpu.
7662 * @cpu: the processor in question.
7664 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7666 * Return: The current task for @cpu.
7668 struct task_struct *curr_task(int cpu)
7670 return cpu_curr(cpu);
7673 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7677 * set_curr_task - set the current task for a given cpu.
7678 * @cpu: the processor in question.
7679 * @p: the task pointer to set.
7681 * Description: This function must only be used when non-maskable interrupts
7682 * are serviced on a separate stack. It allows the architecture to switch the
7683 * notion of the current task on a cpu in a non-blocking manner. This function
7684 * must be called with all CPU's synchronized, and interrupts disabled, the
7685 * and caller must save the original value of the current task (see
7686 * curr_task() above) and restore that value before reenabling interrupts and
7687 * re-starting the system.
7689 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7691 void set_curr_task(int cpu, struct task_struct *p)
7698 #ifdef CONFIG_CGROUP_SCHED
7699 /* task_group_lock serializes the addition/removal of task groups */
7700 static DEFINE_SPINLOCK(task_group_lock);
7702 static void sched_free_group(struct task_group *tg)
7704 free_fair_sched_group(tg);
7705 free_rt_sched_group(tg);
7707 kmem_cache_free(task_group_cache, tg);
7710 /* allocate runqueue etc for a new task group */
7711 struct task_group *sched_create_group(struct task_group *parent)
7713 struct task_group *tg;
7715 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7717 return ERR_PTR(-ENOMEM);
7719 if (!alloc_fair_sched_group(tg, parent))
7722 if (!alloc_rt_sched_group(tg, parent))
7728 sched_free_group(tg);
7729 return ERR_PTR(-ENOMEM);
7732 void sched_online_group(struct task_group *tg, struct task_group *parent)
7734 unsigned long flags;
7736 spin_lock_irqsave(&task_group_lock, flags);
7737 list_add_rcu(&tg->list, &task_groups);
7739 WARN_ON(!parent); /* root should already exist */
7741 tg->parent = parent;
7742 INIT_LIST_HEAD(&tg->children);
7743 list_add_rcu(&tg->siblings, &parent->children);
7744 spin_unlock_irqrestore(&task_group_lock, flags);
7746 online_fair_sched_group(tg);
7749 /* rcu callback to free various structures associated with a task group */
7750 static void sched_free_group_rcu(struct rcu_head *rhp)
7752 /* now it should be safe to free those cfs_rqs */
7753 sched_free_group(container_of(rhp, struct task_group, rcu));
7756 void sched_destroy_group(struct task_group *tg)
7758 /* wait for possible concurrent references to cfs_rqs complete */
7759 call_rcu(&tg->rcu, sched_free_group_rcu);
7762 void sched_offline_group(struct task_group *tg)
7764 unsigned long flags;
7766 /* end participation in shares distribution */
7767 unregister_fair_sched_group(tg);
7769 spin_lock_irqsave(&task_group_lock, flags);
7770 list_del_rcu(&tg->list);
7771 list_del_rcu(&tg->siblings);
7772 spin_unlock_irqrestore(&task_group_lock, flags);
7775 static void sched_change_group(struct task_struct *tsk, int type)
7777 struct task_group *tg;
7780 * All callers are synchronized by task_rq_lock(); we do not use RCU
7781 * which is pointless here. Thus, we pass "true" to task_css_check()
7782 * to prevent lockdep warnings.
7784 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7785 struct task_group, css);
7786 tg = autogroup_task_group(tsk, tg);
7787 tsk->sched_task_group = tg;
7789 #ifdef CONFIG_FAIR_GROUP_SCHED
7790 if (tsk->sched_class->task_change_group)
7791 tsk->sched_class->task_change_group(tsk, type);
7794 set_task_rq(tsk, task_cpu(tsk));
7798 * Change task's runqueue when it moves between groups.
7800 * The caller of this function should have put the task in its new group by
7801 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7804 void sched_move_task(struct task_struct *tsk)
7806 int queued, running;
7810 rq = task_rq_lock(tsk, &rf);
7812 running = task_current(rq, tsk);
7813 queued = task_on_rq_queued(tsk);
7816 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7817 if (unlikely(running))
7818 put_prev_task(rq, tsk);
7820 sched_change_group(tsk, TASK_MOVE_GROUP);
7822 if (unlikely(running))
7823 tsk->sched_class->set_curr_task(rq);
7825 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7827 task_rq_unlock(rq, tsk, &rf);
7829 #endif /* CONFIG_CGROUP_SCHED */
7831 #ifdef CONFIG_RT_GROUP_SCHED
7833 * Ensure that the real time constraints are schedulable.
7835 static DEFINE_MUTEX(rt_constraints_mutex);
7837 /* Must be called with tasklist_lock held */
7838 static inline int tg_has_rt_tasks(struct task_group *tg)
7840 struct task_struct *g, *p;
7843 * Autogroups do not have RT tasks; see autogroup_create().
7845 if (task_group_is_autogroup(tg))
7848 for_each_process_thread(g, p) {
7849 if (rt_task(p) && task_group(p) == tg)
7856 struct rt_schedulable_data {
7857 struct task_group *tg;
7862 static int tg_rt_schedulable(struct task_group *tg, void *data)
7864 struct rt_schedulable_data *d = data;
7865 struct task_group *child;
7866 unsigned long total, sum = 0;
7867 u64 period, runtime;
7869 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7870 runtime = tg->rt_bandwidth.rt_runtime;
7873 period = d->rt_period;
7874 runtime = d->rt_runtime;
7878 * Cannot have more runtime than the period.
7880 if (runtime > period && runtime != RUNTIME_INF)
7884 * Ensure we don't starve existing RT tasks.
7886 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7889 total = to_ratio(period, runtime);
7892 * Nobody can have more than the global setting allows.
7894 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7898 * The sum of our children's runtime should not exceed our own.
7900 list_for_each_entry_rcu(child, &tg->children, siblings) {
7901 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7902 runtime = child->rt_bandwidth.rt_runtime;
7904 if (child == d->tg) {
7905 period = d->rt_period;
7906 runtime = d->rt_runtime;
7909 sum += to_ratio(period, runtime);
7918 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7922 struct rt_schedulable_data data = {
7924 .rt_period = period,
7925 .rt_runtime = runtime,
7929 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7935 static int tg_set_rt_bandwidth(struct task_group *tg,
7936 u64 rt_period, u64 rt_runtime)
7941 * Disallowing the root group RT runtime is BAD, it would disallow the
7942 * kernel creating (and or operating) RT threads.
7944 if (tg == &root_task_group && rt_runtime == 0)
7947 /* No period doesn't make any sense. */
7951 mutex_lock(&rt_constraints_mutex);
7952 read_lock(&tasklist_lock);
7953 err = __rt_schedulable(tg, rt_period, rt_runtime);
7957 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7958 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7959 tg->rt_bandwidth.rt_runtime = rt_runtime;
7961 for_each_possible_cpu(i) {
7962 struct rt_rq *rt_rq = tg->rt_rq[i];
7964 raw_spin_lock(&rt_rq->rt_runtime_lock);
7965 rt_rq->rt_runtime = rt_runtime;
7966 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7968 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7970 read_unlock(&tasklist_lock);
7971 mutex_unlock(&rt_constraints_mutex);
7976 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7978 u64 rt_runtime, rt_period;
7980 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7981 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7982 if (rt_runtime_us < 0)
7983 rt_runtime = RUNTIME_INF;
7985 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7988 static long sched_group_rt_runtime(struct task_group *tg)
7992 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7995 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7996 do_div(rt_runtime_us, NSEC_PER_USEC);
7997 return rt_runtime_us;
8000 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8002 u64 rt_runtime, rt_period;
8004 rt_period = rt_period_us * NSEC_PER_USEC;
8005 rt_runtime = tg->rt_bandwidth.rt_runtime;
8007 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8010 static long sched_group_rt_period(struct task_group *tg)
8014 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8015 do_div(rt_period_us, NSEC_PER_USEC);
8016 return rt_period_us;
8018 #endif /* CONFIG_RT_GROUP_SCHED */
8020 #ifdef CONFIG_RT_GROUP_SCHED
8021 static int sched_rt_global_constraints(void)
8025 mutex_lock(&rt_constraints_mutex);
8026 read_lock(&tasklist_lock);
8027 ret = __rt_schedulable(NULL, 0, 0);
8028 read_unlock(&tasklist_lock);
8029 mutex_unlock(&rt_constraints_mutex);
8034 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8036 /* Don't accept realtime tasks when there is no way for them to run */
8037 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8043 #else /* !CONFIG_RT_GROUP_SCHED */
8044 static int sched_rt_global_constraints(void)
8046 unsigned long flags;
8049 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8050 for_each_possible_cpu(i) {
8051 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8053 raw_spin_lock(&rt_rq->rt_runtime_lock);
8054 rt_rq->rt_runtime = global_rt_runtime();
8055 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8057 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8061 #endif /* CONFIG_RT_GROUP_SCHED */
8063 static int sched_dl_global_validate(void)
8065 u64 runtime = global_rt_runtime();
8066 u64 period = global_rt_period();
8067 u64 new_bw = to_ratio(period, runtime);
8070 unsigned long flags;
8073 * Here we want to check the bandwidth not being set to some
8074 * value smaller than the currently allocated bandwidth in
8075 * any of the root_domains.
8077 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8078 * cycling on root_domains... Discussion on different/better
8079 * solutions is welcome!
8081 for_each_possible_cpu(cpu) {
8082 rcu_read_lock_sched();
8083 dl_b = dl_bw_of(cpu);
8085 raw_spin_lock_irqsave(&dl_b->lock, flags);
8086 if (new_bw < dl_b->total_bw)
8088 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8090 rcu_read_unlock_sched();
8099 static void sched_dl_do_global(void)
8104 unsigned long flags;
8106 def_dl_bandwidth.dl_period = global_rt_period();
8107 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8109 if (global_rt_runtime() != RUNTIME_INF)
8110 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8113 * FIXME: As above...
8115 for_each_possible_cpu(cpu) {
8116 rcu_read_lock_sched();
8117 dl_b = dl_bw_of(cpu);
8119 raw_spin_lock_irqsave(&dl_b->lock, flags);
8121 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8123 rcu_read_unlock_sched();
8127 static int sched_rt_global_validate(void)
8129 if (sysctl_sched_rt_period <= 0)
8132 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8133 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8139 static void sched_rt_do_global(void)
8141 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8142 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8145 int sched_rt_handler(struct ctl_table *table, int write,
8146 void __user *buffer, size_t *lenp,
8149 int old_period, old_runtime;
8150 static DEFINE_MUTEX(mutex);
8154 old_period = sysctl_sched_rt_period;
8155 old_runtime = sysctl_sched_rt_runtime;
8157 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8159 if (!ret && write) {
8160 ret = sched_rt_global_validate();
8164 ret = sched_dl_global_validate();
8168 ret = sched_rt_global_constraints();
8172 sched_rt_do_global();
8173 sched_dl_do_global();
8177 sysctl_sched_rt_period = old_period;
8178 sysctl_sched_rt_runtime = old_runtime;
8180 mutex_unlock(&mutex);
8185 int sched_rr_handler(struct ctl_table *table, int write,
8186 void __user *buffer, size_t *lenp,
8190 static DEFINE_MUTEX(mutex);
8193 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8194 /* make sure that internally we keep jiffies */
8195 /* also, writing zero resets timeslice to default */
8196 if (!ret && write) {
8197 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8198 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8200 mutex_unlock(&mutex);
8204 #ifdef CONFIG_CGROUP_SCHED
8206 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8208 return css ? container_of(css, struct task_group, css) : NULL;
8211 static struct cgroup_subsys_state *
8212 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8214 struct task_group *parent = css_tg(parent_css);
8215 struct task_group *tg;
8218 /* This is early initialization for the top cgroup */
8219 return &root_task_group.css;
8222 tg = sched_create_group(parent);
8224 return ERR_PTR(-ENOMEM);
8226 sched_online_group(tg, parent);
8231 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8233 struct task_group *tg = css_tg(css);
8235 sched_offline_group(tg);
8238 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8240 struct task_group *tg = css_tg(css);
8243 * Relies on the RCU grace period between css_released() and this.
8245 sched_free_group(tg);
8249 * This is called before wake_up_new_task(), therefore we really only
8250 * have to set its group bits, all the other stuff does not apply.
8252 static void cpu_cgroup_fork(struct task_struct *task)
8257 rq = task_rq_lock(task, &rf);
8259 sched_change_group(task, TASK_SET_GROUP);
8261 task_rq_unlock(rq, task, &rf);
8264 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8266 struct task_struct *task;
8267 struct cgroup_subsys_state *css;
8270 cgroup_taskset_for_each(task, css, tset) {
8271 #ifdef CONFIG_RT_GROUP_SCHED
8272 if (!sched_rt_can_attach(css_tg(css), task))
8275 /* We don't support RT-tasks being in separate groups */
8276 if (task->sched_class != &fair_sched_class)
8280 * Serialize against wake_up_new_task() such that if its
8281 * running, we're sure to observe its full state.
8283 raw_spin_lock_irq(&task->pi_lock);
8285 * Avoid calling sched_move_task() before wake_up_new_task()
8286 * has happened. This would lead to problems with PELT, due to
8287 * move wanting to detach+attach while we're not attached yet.
8289 if (task->state == TASK_NEW)
8291 raw_spin_unlock_irq(&task->pi_lock);
8299 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8301 struct task_struct *task;
8302 struct cgroup_subsys_state *css;
8304 cgroup_taskset_for_each(task, css, tset)
8305 sched_move_task(task);
8308 #ifdef CONFIG_FAIR_GROUP_SCHED
8309 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8310 struct cftype *cftype, u64 shareval)
8312 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8315 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8318 struct task_group *tg = css_tg(css);
8320 return (u64) scale_load_down(tg->shares);
8323 #ifdef CONFIG_CFS_BANDWIDTH
8324 static DEFINE_MUTEX(cfs_constraints_mutex);
8326 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8327 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8329 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8331 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8333 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8334 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8336 if (tg == &root_task_group)
8340 * Ensure we have at some amount of bandwidth every period. This is
8341 * to prevent reaching a state of large arrears when throttled via
8342 * entity_tick() resulting in prolonged exit starvation.
8344 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8348 * Likewise, bound things on the otherside by preventing insane quota
8349 * periods. This also allows us to normalize in computing quota
8352 if (period > max_cfs_quota_period)
8356 * Prevent race between setting of cfs_rq->runtime_enabled and
8357 * unthrottle_offline_cfs_rqs().
8360 mutex_lock(&cfs_constraints_mutex);
8361 ret = __cfs_schedulable(tg, period, quota);
8365 runtime_enabled = quota != RUNTIME_INF;
8366 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8368 * If we need to toggle cfs_bandwidth_used, off->on must occur
8369 * before making related changes, and on->off must occur afterwards
8371 if (runtime_enabled && !runtime_was_enabled)
8372 cfs_bandwidth_usage_inc();
8373 raw_spin_lock_irq(&cfs_b->lock);
8374 cfs_b->period = ns_to_ktime(period);
8375 cfs_b->quota = quota;
8377 __refill_cfs_bandwidth_runtime(cfs_b);
8378 /* restart the period timer (if active) to handle new period expiry */
8379 if (runtime_enabled)
8380 start_cfs_bandwidth(cfs_b);
8381 raw_spin_unlock_irq(&cfs_b->lock);
8383 for_each_online_cpu(i) {
8384 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8385 struct rq *rq = cfs_rq->rq;
8387 raw_spin_lock_irq(&rq->lock);
8388 cfs_rq->runtime_enabled = runtime_enabled;
8389 cfs_rq->runtime_remaining = 0;
8391 if (cfs_rq->throttled)
8392 unthrottle_cfs_rq(cfs_rq);
8393 raw_spin_unlock_irq(&rq->lock);
8395 if (runtime_was_enabled && !runtime_enabled)
8396 cfs_bandwidth_usage_dec();
8398 mutex_unlock(&cfs_constraints_mutex);
8404 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8408 period = ktime_to_ns(tg->cfs_bandwidth.period);
8409 if (cfs_quota_us < 0)
8410 quota = RUNTIME_INF;
8412 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8414 return tg_set_cfs_bandwidth(tg, period, quota);
8417 long tg_get_cfs_quota(struct task_group *tg)
8421 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8424 quota_us = tg->cfs_bandwidth.quota;
8425 do_div(quota_us, NSEC_PER_USEC);
8430 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8434 period = (u64)cfs_period_us * NSEC_PER_USEC;
8435 quota = tg->cfs_bandwidth.quota;
8437 return tg_set_cfs_bandwidth(tg, period, quota);
8440 long tg_get_cfs_period(struct task_group *tg)
8444 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8445 do_div(cfs_period_us, NSEC_PER_USEC);
8447 return cfs_period_us;
8450 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8453 return tg_get_cfs_quota(css_tg(css));
8456 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8457 struct cftype *cftype, s64 cfs_quota_us)
8459 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8462 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8465 return tg_get_cfs_period(css_tg(css));
8468 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8469 struct cftype *cftype, u64 cfs_period_us)
8471 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8474 struct cfs_schedulable_data {
8475 struct task_group *tg;
8480 * normalize group quota/period to be quota/max_period
8481 * note: units are usecs
8483 static u64 normalize_cfs_quota(struct task_group *tg,
8484 struct cfs_schedulable_data *d)
8492 period = tg_get_cfs_period(tg);
8493 quota = tg_get_cfs_quota(tg);
8496 /* note: these should typically be equivalent */
8497 if (quota == RUNTIME_INF || quota == -1)
8500 return to_ratio(period, quota);
8503 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8505 struct cfs_schedulable_data *d = data;
8506 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8507 s64 quota = 0, parent_quota = -1;
8510 quota = RUNTIME_INF;
8512 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8514 quota = normalize_cfs_quota(tg, d);
8515 parent_quota = parent_b->hierarchical_quota;
8518 * ensure max(child_quota) <= parent_quota, inherit when no
8521 if (quota == RUNTIME_INF)
8522 quota = parent_quota;
8523 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8526 cfs_b->hierarchical_quota = quota;
8531 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8534 struct cfs_schedulable_data data = {
8540 if (quota != RUNTIME_INF) {
8541 do_div(data.period, NSEC_PER_USEC);
8542 do_div(data.quota, NSEC_PER_USEC);
8546 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8552 static int cpu_stats_show(struct seq_file *sf, void *v)
8554 struct task_group *tg = css_tg(seq_css(sf));
8555 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8557 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8558 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8559 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8563 #endif /* CONFIG_CFS_BANDWIDTH */
8564 #endif /* CONFIG_FAIR_GROUP_SCHED */
8566 #ifdef CONFIG_RT_GROUP_SCHED
8567 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8568 struct cftype *cft, s64 val)
8570 return sched_group_set_rt_runtime(css_tg(css), val);
8573 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8576 return sched_group_rt_runtime(css_tg(css));
8579 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8580 struct cftype *cftype, u64 rt_period_us)
8582 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8585 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8588 return sched_group_rt_period(css_tg(css));
8590 #endif /* CONFIG_RT_GROUP_SCHED */
8592 static struct cftype cpu_files[] = {
8593 #ifdef CONFIG_FAIR_GROUP_SCHED
8596 .read_u64 = cpu_shares_read_u64,
8597 .write_u64 = cpu_shares_write_u64,
8600 #ifdef CONFIG_CFS_BANDWIDTH
8602 .name = "cfs_quota_us",
8603 .read_s64 = cpu_cfs_quota_read_s64,
8604 .write_s64 = cpu_cfs_quota_write_s64,
8607 .name = "cfs_period_us",
8608 .read_u64 = cpu_cfs_period_read_u64,
8609 .write_u64 = cpu_cfs_period_write_u64,
8613 .seq_show = cpu_stats_show,
8616 #ifdef CONFIG_RT_GROUP_SCHED
8618 .name = "rt_runtime_us",
8619 .read_s64 = cpu_rt_runtime_read,
8620 .write_s64 = cpu_rt_runtime_write,
8623 .name = "rt_period_us",
8624 .read_u64 = cpu_rt_period_read_uint,
8625 .write_u64 = cpu_rt_period_write_uint,
8631 struct cgroup_subsys cpu_cgrp_subsys = {
8632 .css_alloc = cpu_cgroup_css_alloc,
8633 .css_released = cpu_cgroup_css_released,
8634 .css_free = cpu_cgroup_css_free,
8635 .fork = cpu_cgroup_fork,
8636 .can_attach = cpu_cgroup_can_attach,
8637 .attach = cpu_cgroup_attach,
8638 .legacy_cftypes = cpu_files,
8642 #endif /* CONFIG_CGROUP_SCHED */
8644 void dump_cpu_task(int cpu)
8646 pr_info("Task dump for CPU %d:\n", cpu);
8647 sched_show_task(cpu_curr(cpu));
8651 * Nice levels are multiplicative, with a gentle 10% change for every
8652 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8653 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8654 * that remained on nice 0.
8656 * The "10% effect" is relative and cumulative: from _any_ nice level,
8657 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8658 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8659 * If a task goes up by ~10% and another task goes down by ~10% then
8660 * the relative distance between them is ~25%.)
8662 const int sched_prio_to_weight[40] = {
8663 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8664 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8665 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8666 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8667 /* 0 */ 1024, 820, 655, 526, 423,
8668 /* 5 */ 335, 272, 215, 172, 137,
8669 /* 10 */ 110, 87, 70, 56, 45,
8670 /* 15 */ 36, 29, 23, 18, 15,
8674 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8676 * In cases where the weight does not change often, we can use the
8677 * precalculated inverse to speed up arithmetics by turning divisions
8678 * into multiplications:
8680 const u32 sched_prio_to_wmult[40] = {
8681 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8682 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8683 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8684 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8685 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8686 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8687 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8688 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,