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) {
1067 if (task_on_rq_queued(p))
1068 rq = __migrate_task(rq, p, arg->dest_cpu);
1070 p->wake_cpu = arg->dest_cpu;
1072 raw_spin_unlock(&rq->lock);
1073 raw_spin_unlock(&p->pi_lock);
1080 * sched_class::set_cpus_allowed must do the below, but is not required to
1081 * actually call this function.
1083 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1085 cpumask_copy(&p->cpus_allowed, new_mask);
1086 p->nr_cpus_allowed = cpumask_weight(new_mask);
1089 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1091 struct rq *rq = task_rq(p);
1092 bool queued, running;
1094 lockdep_assert_held(&p->pi_lock);
1096 queued = task_on_rq_queued(p);
1097 running = task_current(rq, p);
1101 * Because __kthread_bind() calls this on blocked tasks without
1104 lockdep_assert_held(&rq->lock);
1105 dequeue_task(rq, p, DEQUEUE_SAVE);
1108 put_prev_task(rq, p);
1110 p->sched_class->set_cpus_allowed(p, new_mask);
1113 p->sched_class->set_curr_task(rq);
1115 enqueue_task(rq, p, ENQUEUE_RESTORE);
1119 * Change a given task's CPU affinity. Migrate the thread to a
1120 * proper CPU and schedule it away if the CPU it's executing on
1121 * is removed from the allowed bitmask.
1123 * NOTE: the caller must have a valid reference to the task, the
1124 * task must not exit() & deallocate itself prematurely. The
1125 * call is not atomic; no spinlocks may be held.
1127 static int __set_cpus_allowed_ptr(struct task_struct *p,
1128 const struct cpumask *new_mask, bool check)
1130 const struct cpumask *cpu_valid_mask = cpu_active_mask;
1131 unsigned int dest_cpu;
1136 rq = task_rq_lock(p, &rf);
1138 if (p->flags & PF_KTHREAD) {
1140 * Kernel threads are allowed on online && !active CPUs
1142 cpu_valid_mask = cpu_online_mask;
1146 * Must re-check here, to close a race against __kthread_bind(),
1147 * sched_setaffinity() is not guaranteed to observe the flag.
1149 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1154 if (cpumask_equal(&p->cpus_allowed, new_mask))
1157 if (!cpumask_intersects(new_mask, cpu_valid_mask)) {
1162 do_set_cpus_allowed(p, new_mask);
1164 if (p->flags & PF_KTHREAD) {
1166 * For kernel threads that do indeed end up on online &&
1167 * !active we want to ensure they are strict per-cpu threads.
1169 WARN_ON(cpumask_intersects(new_mask, cpu_online_mask) &&
1170 !cpumask_intersects(new_mask, cpu_active_mask) &&
1171 p->nr_cpus_allowed != 1);
1174 /* Can the task run on the task's current CPU? If so, we're done */
1175 if (cpumask_test_cpu(task_cpu(p), new_mask))
1178 dest_cpu = cpumask_any_and(cpu_valid_mask, new_mask);
1179 if (task_running(rq, p) || p->state == TASK_WAKING) {
1180 struct migration_arg arg = { p, dest_cpu };
1181 /* Need help from migration thread: drop lock and wait. */
1182 task_rq_unlock(rq, p, &rf);
1183 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1184 tlb_migrate_finish(p->mm);
1186 } else if (task_on_rq_queued(p)) {
1188 * OK, since we're going to drop the lock immediately
1189 * afterwards anyway.
1191 lockdep_unpin_lock(&rq->lock, rf.cookie);
1192 rq = move_queued_task(rq, p, dest_cpu);
1193 lockdep_repin_lock(&rq->lock, rf.cookie);
1196 task_rq_unlock(rq, p, &rf);
1201 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1203 return __set_cpus_allowed_ptr(p, new_mask, false);
1205 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1207 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1209 #ifdef CONFIG_SCHED_DEBUG
1211 * We should never call set_task_cpu() on a blocked task,
1212 * ttwu() will sort out the placement.
1214 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1218 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
1219 * because schedstat_wait_{start,end} rebase migrating task's wait_start
1220 * time relying on p->on_rq.
1222 WARN_ON_ONCE(p->state == TASK_RUNNING &&
1223 p->sched_class == &fair_sched_class &&
1224 (p->on_rq && !task_on_rq_migrating(p)));
1226 #ifdef CONFIG_LOCKDEP
1228 * The caller should hold either p->pi_lock or rq->lock, when changing
1229 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1231 * sched_move_task() holds both and thus holding either pins the cgroup,
1234 * Furthermore, all task_rq users should acquire both locks, see
1237 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1238 lockdep_is_held(&task_rq(p)->lock)));
1242 trace_sched_migrate_task(p, new_cpu);
1244 if (task_cpu(p) != new_cpu) {
1245 if (p->sched_class->migrate_task_rq)
1246 p->sched_class->migrate_task_rq(p);
1247 p->se.nr_migrations++;
1248 perf_event_task_migrate(p);
1251 __set_task_cpu(p, new_cpu);
1254 static void __migrate_swap_task(struct task_struct *p, int cpu)
1256 if (task_on_rq_queued(p)) {
1257 struct rq *src_rq, *dst_rq;
1259 src_rq = task_rq(p);
1260 dst_rq = cpu_rq(cpu);
1262 p->on_rq = TASK_ON_RQ_MIGRATING;
1263 deactivate_task(src_rq, p, 0);
1264 set_task_cpu(p, cpu);
1265 activate_task(dst_rq, p, 0);
1266 p->on_rq = TASK_ON_RQ_QUEUED;
1267 check_preempt_curr(dst_rq, p, 0);
1270 * Task isn't running anymore; make it appear like we migrated
1271 * it before it went to sleep. This means on wakeup we make the
1272 * previous cpu our target instead of where it really is.
1278 struct migration_swap_arg {
1279 struct task_struct *src_task, *dst_task;
1280 int src_cpu, dst_cpu;
1283 static int migrate_swap_stop(void *data)
1285 struct migration_swap_arg *arg = data;
1286 struct rq *src_rq, *dst_rq;
1289 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1292 src_rq = cpu_rq(arg->src_cpu);
1293 dst_rq = cpu_rq(arg->dst_cpu);
1295 double_raw_lock(&arg->src_task->pi_lock,
1296 &arg->dst_task->pi_lock);
1297 double_rq_lock(src_rq, dst_rq);
1299 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1302 if (task_cpu(arg->src_task) != arg->src_cpu)
1305 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1308 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1311 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1312 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1317 double_rq_unlock(src_rq, dst_rq);
1318 raw_spin_unlock(&arg->dst_task->pi_lock);
1319 raw_spin_unlock(&arg->src_task->pi_lock);
1325 * Cross migrate two tasks
1327 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1329 struct migration_swap_arg arg;
1332 arg = (struct migration_swap_arg){
1334 .src_cpu = task_cpu(cur),
1336 .dst_cpu = task_cpu(p),
1339 if (arg.src_cpu == arg.dst_cpu)
1343 * These three tests are all lockless; this is OK since all of them
1344 * will be re-checked with proper locks held further down the line.
1346 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1349 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1352 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1355 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1356 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1363 * wait_task_inactive - wait for a thread to unschedule.
1365 * If @match_state is nonzero, it's the @p->state value just checked and
1366 * not expected to change. If it changes, i.e. @p might have woken up,
1367 * then return zero. When we succeed in waiting for @p to be off its CPU,
1368 * we return a positive number (its total switch count). If a second call
1369 * a short while later returns the same number, the caller can be sure that
1370 * @p has remained unscheduled the whole time.
1372 * The caller must ensure that the task *will* unschedule sometime soon,
1373 * else this function might spin for a *long* time. This function can't
1374 * be called with interrupts off, or it may introduce deadlock with
1375 * smp_call_function() if an IPI is sent by the same process we are
1376 * waiting to become inactive.
1378 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1380 int running, queued;
1387 * We do the initial early heuristics without holding
1388 * any task-queue locks at all. We'll only try to get
1389 * the runqueue lock when things look like they will
1395 * If the task is actively running on another CPU
1396 * still, just relax and busy-wait without holding
1399 * NOTE! Since we don't hold any locks, it's not
1400 * even sure that "rq" stays as the right runqueue!
1401 * But we don't care, since "task_running()" will
1402 * return false if the runqueue has changed and p
1403 * is actually now running somewhere else!
1405 while (task_running(rq, p)) {
1406 if (match_state && unlikely(p->state != match_state))
1412 * Ok, time to look more closely! We need the rq
1413 * lock now, to be *sure*. If we're wrong, we'll
1414 * just go back and repeat.
1416 rq = task_rq_lock(p, &rf);
1417 trace_sched_wait_task(p);
1418 running = task_running(rq, p);
1419 queued = task_on_rq_queued(p);
1421 if (!match_state || p->state == match_state)
1422 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1423 task_rq_unlock(rq, p, &rf);
1426 * If it changed from the expected state, bail out now.
1428 if (unlikely(!ncsw))
1432 * Was it really running after all now that we
1433 * checked with the proper locks actually held?
1435 * Oops. Go back and try again..
1437 if (unlikely(running)) {
1443 * It's not enough that it's not actively running,
1444 * it must be off the runqueue _entirely_, and not
1447 * So if it was still runnable (but just not actively
1448 * running right now), it's preempted, and we should
1449 * yield - it could be a while.
1451 if (unlikely(queued)) {
1452 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1454 set_current_state(TASK_UNINTERRUPTIBLE);
1455 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1460 * Ahh, all good. It wasn't running, and it wasn't
1461 * runnable, which means that it will never become
1462 * running in the future either. We're all done!
1471 * kick_process - kick a running thread to enter/exit the kernel
1472 * @p: the to-be-kicked thread
1474 * Cause a process which is running on another CPU to enter
1475 * kernel-mode, without any delay. (to get signals handled.)
1477 * NOTE: this function doesn't have to take the runqueue lock,
1478 * because all it wants to ensure is that the remote task enters
1479 * the kernel. If the IPI races and the task has been migrated
1480 * to another CPU then no harm is done and the purpose has been
1483 void kick_process(struct task_struct *p)
1489 if ((cpu != smp_processor_id()) && task_curr(p))
1490 smp_send_reschedule(cpu);
1493 EXPORT_SYMBOL_GPL(kick_process);
1496 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1498 * A few notes on cpu_active vs cpu_online:
1500 * - cpu_active must be a subset of cpu_online
1502 * - on cpu-up we allow per-cpu kthreads on the online && !active cpu,
1503 * see __set_cpus_allowed_ptr(). At this point the newly online
1504 * cpu isn't yet part of the sched domains, and balancing will not
1507 * - on cpu-down we clear cpu_active() to mask the sched domains and
1508 * avoid the load balancer to place new tasks on the to be removed
1509 * cpu. Existing tasks will remain running there and will be taken
1512 * This means that fallback selection must not select !active CPUs.
1513 * And can assume that any active CPU must be online. Conversely
1514 * select_task_rq() below may allow selection of !active CPUs in order
1515 * to satisfy the above rules.
1517 static int select_fallback_rq(int cpu, struct task_struct *p)
1519 int nid = cpu_to_node(cpu);
1520 const struct cpumask *nodemask = NULL;
1521 enum { cpuset, possible, fail } state = cpuset;
1525 * If the node that the cpu is on has been offlined, cpu_to_node()
1526 * will return -1. There is no cpu on the node, and we should
1527 * select the cpu on the other node.
1530 nodemask = cpumask_of_node(nid);
1532 /* Look for allowed, online CPU in same node. */
1533 for_each_cpu(dest_cpu, nodemask) {
1534 if (!cpu_active(dest_cpu))
1536 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1542 /* Any allowed, online CPU? */
1543 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1544 if (!(p->flags & PF_KTHREAD) && !cpu_active(dest_cpu))
1546 if (!cpu_online(dest_cpu))
1551 /* No more Mr. Nice Guy. */
1554 if (IS_ENABLED(CONFIG_CPUSETS)) {
1555 cpuset_cpus_allowed_fallback(p);
1561 do_set_cpus_allowed(p, cpu_possible_mask);
1572 if (state != cpuset) {
1574 * Don't tell them about moving exiting tasks or
1575 * kernel threads (both mm NULL), since they never
1578 if (p->mm && printk_ratelimit()) {
1579 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1580 task_pid_nr(p), p->comm, cpu);
1588 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1591 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1593 lockdep_assert_held(&p->pi_lock);
1595 if (tsk_nr_cpus_allowed(p) > 1)
1596 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1598 cpu = cpumask_any(tsk_cpus_allowed(p));
1601 * In order not to call set_task_cpu() on a blocking task we need
1602 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1605 * Since this is common to all placement strategies, this lives here.
1607 * [ this allows ->select_task() to simply return task_cpu(p) and
1608 * not worry about this generic constraint ]
1610 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1612 cpu = select_fallback_rq(task_cpu(p), p);
1617 static void update_avg(u64 *avg, u64 sample)
1619 s64 diff = sample - *avg;
1625 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1626 const struct cpumask *new_mask, bool check)
1628 return set_cpus_allowed_ptr(p, new_mask);
1631 #endif /* CONFIG_SMP */
1634 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1638 if (!schedstat_enabled())
1644 if (cpu == rq->cpu) {
1645 schedstat_inc(rq->ttwu_local);
1646 schedstat_inc(p->se.statistics.nr_wakeups_local);
1648 struct sched_domain *sd;
1650 schedstat_inc(p->se.statistics.nr_wakeups_remote);
1652 for_each_domain(rq->cpu, sd) {
1653 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1654 schedstat_inc(sd->ttwu_wake_remote);
1661 if (wake_flags & WF_MIGRATED)
1662 schedstat_inc(p->se.statistics.nr_wakeups_migrate);
1663 #endif /* CONFIG_SMP */
1665 schedstat_inc(rq->ttwu_count);
1666 schedstat_inc(p->se.statistics.nr_wakeups);
1668 if (wake_flags & WF_SYNC)
1669 schedstat_inc(p->se.statistics.nr_wakeups_sync);
1672 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1674 activate_task(rq, p, en_flags);
1675 p->on_rq = TASK_ON_RQ_QUEUED;
1677 /* if a worker is waking up, notify workqueue */
1678 if (p->flags & PF_WQ_WORKER)
1679 wq_worker_waking_up(p, cpu_of(rq));
1683 * Mark the task runnable and perform wakeup-preemption.
1685 static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
1686 struct pin_cookie cookie)
1688 check_preempt_curr(rq, p, wake_flags);
1689 p->state = TASK_RUNNING;
1690 trace_sched_wakeup(p);
1693 if (p->sched_class->task_woken) {
1695 * Our task @p is fully woken up and running; so its safe to
1696 * drop the rq->lock, hereafter rq is only used for statistics.
1698 lockdep_unpin_lock(&rq->lock, cookie);
1699 p->sched_class->task_woken(rq, p);
1700 lockdep_repin_lock(&rq->lock, cookie);
1703 if (rq->idle_stamp) {
1704 u64 delta = rq_clock(rq) - rq->idle_stamp;
1705 u64 max = 2*rq->max_idle_balance_cost;
1707 update_avg(&rq->avg_idle, delta);
1709 if (rq->avg_idle > max)
1718 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
1719 struct pin_cookie cookie)
1721 int en_flags = ENQUEUE_WAKEUP;
1723 lockdep_assert_held(&rq->lock);
1726 if (p->sched_contributes_to_load)
1727 rq->nr_uninterruptible--;
1729 if (wake_flags & WF_MIGRATED)
1730 en_flags |= ENQUEUE_MIGRATED;
1733 ttwu_activate(rq, p, en_flags);
1734 ttwu_do_wakeup(rq, p, wake_flags, cookie);
1738 * Called in case the task @p isn't fully descheduled from its runqueue,
1739 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1740 * since all we need to do is flip p->state to TASK_RUNNING, since
1741 * the task is still ->on_rq.
1743 static int ttwu_remote(struct task_struct *p, int wake_flags)
1749 rq = __task_rq_lock(p, &rf);
1750 if (task_on_rq_queued(p)) {
1751 /* check_preempt_curr() may use rq clock */
1752 update_rq_clock(rq);
1753 ttwu_do_wakeup(rq, p, wake_flags, rf.cookie);
1756 __task_rq_unlock(rq, &rf);
1762 void sched_ttwu_pending(void)
1764 struct rq *rq = this_rq();
1765 struct llist_node *llist = llist_del_all(&rq->wake_list);
1766 struct pin_cookie cookie;
1767 struct task_struct *p;
1768 unsigned long flags;
1773 raw_spin_lock_irqsave(&rq->lock, flags);
1774 cookie = lockdep_pin_lock(&rq->lock);
1779 p = llist_entry(llist, struct task_struct, wake_entry);
1780 llist = llist_next(llist);
1782 if (p->sched_remote_wakeup)
1783 wake_flags = WF_MIGRATED;
1785 ttwu_do_activate(rq, p, wake_flags, cookie);
1788 lockdep_unpin_lock(&rq->lock, cookie);
1789 raw_spin_unlock_irqrestore(&rq->lock, flags);
1792 void scheduler_ipi(void)
1795 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1796 * TIF_NEED_RESCHED remotely (for the first time) will also send
1799 preempt_fold_need_resched();
1801 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1805 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1806 * traditionally all their work was done from the interrupt return
1807 * path. Now that we actually do some work, we need to make sure
1810 * Some archs already do call them, luckily irq_enter/exit nest
1813 * Arguably we should visit all archs and update all handlers,
1814 * however a fair share of IPIs are still resched only so this would
1815 * somewhat pessimize the simple resched case.
1818 sched_ttwu_pending();
1821 * Check if someone kicked us for doing the nohz idle load balance.
1823 if (unlikely(got_nohz_idle_kick())) {
1824 this_rq()->idle_balance = 1;
1825 raise_softirq_irqoff(SCHED_SOFTIRQ);
1830 static void ttwu_queue_remote(struct task_struct *p, int cpu, int wake_flags)
1832 struct rq *rq = cpu_rq(cpu);
1834 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
1836 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1837 if (!set_nr_if_polling(rq->idle))
1838 smp_send_reschedule(cpu);
1840 trace_sched_wake_idle_without_ipi(cpu);
1844 void wake_up_if_idle(int cpu)
1846 struct rq *rq = cpu_rq(cpu);
1847 unsigned long flags;
1851 if (!is_idle_task(rcu_dereference(rq->curr)))
1854 if (set_nr_if_polling(rq->idle)) {
1855 trace_sched_wake_idle_without_ipi(cpu);
1857 raw_spin_lock_irqsave(&rq->lock, flags);
1858 if (is_idle_task(rq->curr))
1859 smp_send_reschedule(cpu);
1860 /* Else cpu is not in idle, do nothing here */
1861 raw_spin_unlock_irqrestore(&rq->lock, flags);
1868 bool cpus_share_cache(int this_cpu, int that_cpu)
1870 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1872 #endif /* CONFIG_SMP */
1874 static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
1876 struct rq *rq = cpu_rq(cpu);
1877 struct pin_cookie cookie;
1879 #if defined(CONFIG_SMP)
1880 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1881 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1882 ttwu_queue_remote(p, cpu, wake_flags);
1887 raw_spin_lock(&rq->lock);
1888 cookie = lockdep_pin_lock(&rq->lock);
1889 ttwu_do_activate(rq, p, wake_flags, cookie);
1890 lockdep_unpin_lock(&rq->lock, cookie);
1891 raw_spin_unlock(&rq->lock);
1895 * Notes on Program-Order guarantees on SMP systems.
1899 * The basic program-order guarantee on SMP systems is that when a task [t]
1900 * migrates, all its activity on its old cpu [c0] happens-before any subsequent
1901 * execution on its new cpu [c1].
1903 * For migration (of runnable tasks) this is provided by the following means:
1905 * A) UNLOCK of the rq(c0)->lock scheduling out task t
1906 * B) migration for t is required to synchronize *both* rq(c0)->lock and
1907 * rq(c1)->lock (if not at the same time, then in that order).
1908 * C) LOCK of the rq(c1)->lock scheduling in task
1910 * Transitivity guarantees that B happens after A and C after B.
1911 * Note: we only require RCpc transitivity.
1912 * Note: the cpu doing B need not be c0 or c1
1921 * UNLOCK rq(0)->lock
1923 * LOCK rq(0)->lock // orders against CPU0
1925 * UNLOCK rq(0)->lock
1929 * UNLOCK rq(1)->lock
1931 * LOCK rq(1)->lock // orders against CPU2
1934 * UNLOCK rq(1)->lock
1937 * BLOCKING -- aka. SLEEP + WAKEUP
1939 * For blocking we (obviously) need to provide the same guarantee as for
1940 * migration. However the means are completely different as there is no lock
1941 * chain to provide order. Instead we do:
1943 * 1) smp_store_release(X->on_cpu, 0)
1944 * 2) smp_cond_load_acquire(!X->on_cpu)
1948 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
1950 * LOCK rq(0)->lock LOCK X->pi_lock
1953 * smp_store_release(X->on_cpu, 0);
1955 * smp_cond_load_acquire(&X->on_cpu, !VAL);
1961 * X->state = RUNNING
1962 * UNLOCK rq(2)->lock
1964 * LOCK rq(2)->lock // orders against CPU1
1967 * UNLOCK rq(2)->lock
1970 * UNLOCK rq(0)->lock
1973 * However; for wakeups there is a second guarantee we must provide, namely we
1974 * must observe the state that lead to our wakeup. That is, not only must our
1975 * task observe its own prior state, it must also observe the stores prior to
1978 * This means that any means of doing remote wakeups must order the CPU doing
1979 * the wakeup against the CPU the task is going to end up running on. This,
1980 * however, is already required for the regular Program-Order guarantee above,
1981 * since the waking CPU is the one issueing the ACQUIRE (smp_cond_load_acquire).
1986 * try_to_wake_up - wake up a thread
1987 * @p: the thread to be awakened
1988 * @state: the mask of task states that can be woken
1989 * @wake_flags: wake modifier flags (WF_*)
1991 * Put it on the run-queue if it's not already there. The "current"
1992 * thread is always on the run-queue (except when the actual
1993 * re-schedule is in progress), and as such you're allowed to do
1994 * the simpler "current->state = TASK_RUNNING" to mark yourself
1995 * runnable without the overhead of this.
1997 * Return: %true if @p was woken up, %false if it was already running.
1998 * or @state didn't match @p's state.
2001 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2003 unsigned long flags;
2004 int cpu, success = 0;
2007 * If we are going to wake up a thread waiting for CONDITION we
2008 * need to ensure that CONDITION=1 done by the caller can not be
2009 * reordered with p->state check below. This pairs with mb() in
2010 * set_current_state() the waiting thread does.
2012 smp_mb__before_spinlock();
2013 raw_spin_lock_irqsave(&p->pi_lock, flags);
2014 if (!(p->state & state))
2017 trace_sched_waking(p);
2019 success = 1; /* we're going to change ->state */
2023 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2024 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2025 * in smp_cond_load_acquire() below.
2027 * sched_ttwu_pending() try_to_wake_up()
2028 * [S] p->on_rq = 1; [L] P->state
2029 * UNLOCK rq->lock -----.
2033 * LOCK rq->lock -----'
2037 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2039 * Pairs with the UNLOCK+LOCK on rq->lock from the
2040 * last wakeup of our task and the schedule that got our task
2044 if (p->on_rq && ttwu_remote(p, wake_flags))
2049 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2050 * possible to, falsely, observe p->on_cpu == 0.
2052 * One must be running (->on_cpu == 1) in order to remove oneself
2053 * from the runqueue.
2055 * [S] ->on_cpu = 1; [L] ->on_rq
2059 * [S] ->on_rq = 0; [L] ->on_cpu
2061 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2062 * from the consecutive calls to schedule(); the first switching to our
2063 * task, the second putting it to sleep.
2068 * If the owning (remote) cpu is still in the middle of schedule() with
2069 * this task as prev, wait until its done referencing the task.
2071 * Pairs with the smp_store_release() in finish_lock_switch().
2073 * This ensures that tasks getting woken will be fully ordered against
2074 * their previous state and preserve Program Order.
2076 smp_cond_load_acquire(&p->on_cpu, !VAL);
2078 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2079 p->state = TASK_WAKING;
2081 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2082 if (task_cpu(p) != cpu) {
2083 wake_flags |= WF_MIGRATED;
2084 set_task_cpu(p, cpu);
2086 #endif /* CONFIG_SMP */
2088 ttwu_queue(p, cpu, wake_flags);
2090 ttwu_stat(p, cpu, wake_flags);
2092 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2098 * try_to_wake_up_local - try to wake up a local task with rq lock held
2099 * @p: the thread to be awakened
2100 * @cookie: context's cookie for pinning
2102 * Put @p on the run-queue if it's not already there. The caller must
2103 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2106 static void try_to_wake_up_local(struct task_struct *p, struct pin_cookie cookie)
2108 struct rq *rq = task_rq(p);
2110 if (WARN_ON_ONCE(rq != this_rq()) ||
2111 WARN_ON_ONCE(p == current))
2114 lockdep_assert_held(&rq->lock);
2116 if (!raw_spin_trylock(&p->pi_lock)) {
2118 * This is OK, because current is on_cpu, which avoids it being
2119 * picked for load-balance and preemption/IRQs are still
2120 * disabled avoiding further scheduler activity on it and we've
2121 * not yet picked a replacement task.
2123 lockdep_unpin_lock(&rq->lock, cookie);
2124 raw_spin_unlock(&rq->lock);
2125 raw_spin_lock(&p->pi_lock);
2126 raw_spin_lock(&rq->lock);
2127 lockdep_repin_lock(&rq->lock, cookie);
2130 if (!(p->state & TASK_NORMAL))
2133 trace_sched_waking(p);
2135 if (!task_on_rq_queued(p))
2136 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2138 ttwu_do_wakeup(rq, p, 0, cookie);
2139 ttwu_stat(p, smp_processor_id(), 0);
2141 raw_spin_unlock(&p->pi_lock);
2145 * wake_up_process - Wake up a specific process
2146 * @p: The process to be woken up.
2148 * Attempt to wake up the nominated process and move it to the set of runnable
2151 * Return: 1 if the process was woken up, 0 if it was already running.
2153 * It may be assumed that this function implies a write memory barrier before
2154 * changing the task state if and only if any tasks are woken up.
2156 int wake_up_process(struct task_struct *p)
2158 return try_to_wake_up(p, TASK_NORMAL, 0);
2160 EXPORT_SYMBOL(wake_up_process);
2162 int wake_up_state(struct task_struct *p, unsigned int state)
2164 return try_to_wake_up(p, state, 0);
2168 * This function clears the sched_dl_entity static params.
2170 void __dl_clear_params(struct task_struct *p)
2172 struct sched_dl_entity *dl_se = &p->dl;
2174 dl_se->dl_runtime = 0;
2175 dl_se->dl_deadline = 0;
2176 dl_se->dl_period = 0;
2180 dl_se->dl_throttled = 0;
2181 dl_se->dl_yielded = 0;
2185 * Perform scheduler related setup for a newly forked process p.
2186 * p is forked by current.
2188 * __sched_fork() is basic setup used by init_idle() too:
2190 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2195 p->se.exec_start = 0;
2196 p->se.sum_exec_runtime = 0;
2197 p->se.prev_sum_exec_runtime = 0;
2198 p->se.nr_migrations = 0;
2200 INIT_LIST_HEAD(&p->se.group_node);
2202 #ifdef CONFIG_FAIR_GROUP_SCHED
2203 p->se.cfs_rq = NULL;
2206 #ifdef CONFIG_SCHEDSTATS
2207 /* Even if schedstat is disabled, there should not be garbage */
2208 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2211 RB_CLEAR_NODE(&p->dl.rb_node);
2212 init_dl_task_timer(&p->dl);
2213 __dl_clear_params(p);
2215 INIT_LIST_HEAD(&p->rt.run_list);
2217 p->rt.time_slice = sched_rr_timeslice;
2221 #ifdef CONFIG_PREEMPT_NOTIFIERS
2222 INIT_HLIST_HEAD(&p->preempt_notifiers);
2225 #ifdef CONFIG_NUMA_BALANCING
2226 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2227 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2228 p->mm->numa_scan_seq = 0;
2231 if (clone_flags & CLONE_VM)
2232 p->numa_preferred_nid = current->numa_preferred_nid;
2234 p->numa_preferred_nid = -1;
2236 p->node_stamp = 0ULL;
2237 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2238 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2239 p->numa_work.next = &p->numa_work;
2240 p->numa_faults = NULL;
2241 p->last_task_numa_placement = 0;
2242 p->last_sum_exec_runtime = 0;
2244 p->numa_group = NULL;
2245 #endif /* CONFIG_NUMA_BALANCING */
2248 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2250 #ifdef CONFIG_NUMA_BALANCING
2252 void set_numabalancing_state(bool enabled)
2255 static_branch_enable(&sched_numa_balancing);
2257 static_branch_disable(&sched_numa_balancing);
2260 #ifdef CONFIG_PROC_SYSCTL
2261 int sysctl_numa_balancing(struct ctl_table *table, int write,
2262 void __user *buffer, size_t *lenp, loff_t *ppos)
2266 int state = static_branch_likely(&sched_numa_balancing);
2268 if (write && !capable(CAP_SYS_ADMIN))
2273 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2277 set_numabalancing_state(state);
2283 #ifdef CONFIG_SCHEDSTATS
2285 DEFINE_STATIC_KEY_FALSE(sched_schedstats);
2286 static bool __initdata __sched_schedstats = false;
2288 static void set_schedstats(bool enabled)
2291 static_branch_enable(&sched_schedstats);
2293 static_branch_disable(&sched_schedstats);
2296 void force_schedstat_enabled(void)
2298 if (!schedstat_enabled()) {
2299 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
2300 static_branch_enable(&sched_schedstats);
2304 static int __init setup_schedstats(char *str)
2311 * This code is called before jump labels have been set up, so we can't
2312 * change the static branch directly just yet. Instead set a temporary
2313 * variable so init_schedstats() can do it later.
2315 if (!strcmp(str, "enable")) {
2316 __sched_schedstats = true;
2318 } else if (!strcmp(str, "disable")) {
2319 __sched_schedstats = false;
2324 pr_warn("Unable to parse schedstats=\n");
2328 __setup("schedstats=", setup_schedstats);
2330 static void __init init_schedstats(void)
2332 set_schedstats(__sched_schedstats);
2335 #ifdef CONFIG_PROC_SYSCTL
2336 int sysctl_schedstats(struct ctl_table *table, int write,
2337 void __user *buffer, size_t *lenp, loff_t *ppos)
2341 int state = static_branch_likely(&sched_schedstats);
2343 if (write && !capable(CAP_SYS_ADMIN))
2348 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2352 set_schedstats(state);
2355 #endif /* CONFIG_PROC_SYSCTL */
2356 #else /* !CONFIG_SCHEDSTATS */
2357 static inline void init_schedstats(void) {}
2358 #endif /* CONFIG_SCHEDSTATS */
2361 * fork()/clone()-time setup:
2363 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2365 unsigned long flags;
2366 int cpu = get_cpu();
2368 __sched_fork(clone_flags, p);
2370 * We mark the process as NEW here. This guarantees that
2371 * nobody will actually run it, and a signal or other external
2372 * event cannot wake it up and insert it on the runqueue either.
2374 p->state = TASK_NEW;
2377 * Make sure we do not leak PI boosting priority to the child.
2379 p->prio = current->normal_prio;
2382 * Revert to default priority/policy on fork if requested.
2384 if (unlikely(p->sched_reset_on_fork)) {
2385 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2386 p->policy = SCHED_NORMAL;
2387 p->static_prio = NICE_TO_PRIO(0);
2389 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2390 p->static_prio = NICE_TO_PRIO(0);
2392 p->prio = p->normal_prio = __normal_prio(p);
2396 * We don't need the reset flag anymore after the fork. It has
2397 * fulfilled its duty:
2399 p->sched_reset_on_fork = 0;
2402 if (dl_prio(p->prio)) {
2405 } else if (rt_prio(p->prio)) {
2406 p->sched_class = &rt_sched_class;
2408 p->sched_class = &fair_sched_class;
2411 init_entity_runnable_average(&p->se);
2414 * The child is not yet in the pid-hash so no cgroup attach races,
2415 * and the cgroup is pinned to this child due to cgroup_fork()
2416 * is ran before sched_fork().
2418 * Silence PROVE_RCU.
2420 raw_spin_lock_irqsave(&p->pi_lock, flags);
2422 * We're setting the cpu for the first time, we don't migrate,
2423 * so use __set_task_cpu().
2425 __set_task_cpu(p, cpu);
2426 if (p->sched_class->task_fork)
2427 p->sched_class->task_fork(p);
2428 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2430 #ifdef CONFIG_SCHED_INFO
2431 if (likely(sched_info_on()))
2432 memset(&p->sched_info, 0, sizeof(p->sched_info));
2434 #if defined(CONFIG_SMP)
2437 init_task_preempt_count(p);
2439 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2440 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2447 unsigned long to_ratio(u64 period, u64 runtime)
2449 if (runtime == RUNTIME_INF)
2453 * Doing this here saves a lot of checks in all
2454 * the calling paths, and returning zero seems
2455 * safe for them anyway.
2460 return div64_u64(runtime << 20, period);
2464 inline struct dl_bw *dl_bw_of(int i)
2466 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2467 "sched RCU must be held");
2468 return &cpu_rq(i)->rd->dl_bw;
2471 static inline int dl_bw_cpus(int i)
2473 struct root_domain *rd = cpu_rq(i)->rd;
2476 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2477 "sched RCU must be held");
2478 for_each_cpu_and(i, rd->span, cpu_active_mask)
2484 inline struct dl_bw *dl_bw_of(int i)
2486 return &cpu_rq(i)->dl.dl_bw;
2489 static inline int dl_bw_cpus(int i)
2496 * We must be sure that accepting a new task (or allowing changing the
2497 * parameters of an existing one) is consistent with the bandwidth
2498 * constraints. If yes, this function also accordingly updates the currently
2499 * allocated bandwidth to reflect the new situation.
2501 * This function is called while holding p's rq->lock.
2503 * XXX we should delay bw change until the task's 0-lag point, see
2506 static int dl_overflow(struct task_struct *p, int policy,
2507 const struct sched_attr *attr)
2510 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2511 u64 period = attr->sched_period ?: attr->sched_deadline;
2512 u64 runtime = attr->sched_runtime;
2513 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2516 /* !deadline task may carry old deadline bandwidth */
2517 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
2521 * Either if a task, enters, leave, or stays -deadline but changes
2522 * its parameters, we may need to update accordingly the total
2523 * allocated bandwidth of the container.
2525 raw_spin_lock(&dl_b->lock);
2526 cpus = dl_bw_cpus(task_cpu(p));
2527 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2528 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2529 __dl_add(dl_b, new_bw);
2531 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2532 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2533 __dl_clear(dl_b, p->dl.dl_bw);
2534 __dl_add(dl_b, new_bw);
2536 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2537 __dl_clear(dl_b, p->dl.dl_bw);
2540 raw_spin_unlock(&dl_b->lock);
2545 extern void init_dl_bw(struct dl_bw *dl_b);
2548 * wake_up_new_task - wake up a newly created task for the first time.
2550 * This function will do some initial scheduler statistics housekeeping
2551 * that must be done for every newly created context, then puts the task
2552 * on the runqueue and wakes it.
2554 void wake_up_new_task(struct task_struct *p)
2559 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
2560 p->state = TASK_RUNNING;
2563 * Fork balancing, do it here and not earlier because:
2564 * - cpus_allowed can change in the fork path
2565 * - any previously selected cpu might disappear through hotplug
2567 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2568 * as we're not fully set-up yet.
2570 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2572 rq = __task_rq_lock(p, &rf);
2573 post_init_entity_util_avg(&p->se);
2575 activate_task(rq, p, 0);
2576 p->on_rq = TASK_ON_RQ_QUEUED;
2577 trace_sched_wakeup_new(p);
2578 check_preempt_curr(rq, p, WF_FORK);
2580 if (p->sched_class->task_woken) {
2582 * Nothing relies on rq->lock after this, so its fine to
2585 lockdep_unpin_lock(&rq->lock, rf.cookie);
2586 p->sched_class->task_woken(rq, p);
2587 lockdep_repin_lock(&rq->lock, rf.cookie);
2590 task_rq_unlock(rq, p, &rf);
2593 #ifdef CONFIG_PREEMPT_NOTIFIERS
2595 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2597 void preempt_notifier_inc(void)
2599 static_key_slow_inc(&preempt_notifier_key);
2601 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2603 void preempt_notifier_dec(void)
2605 static_key_slow_dec(&preempt_notifier_key);
2607 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2610 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2611 * @notifier: notifier struct to register
2613 void preempt_notifier_register(struct preempt_notifier *notifier)
2615 if (!static_key_false(&preempt_notifier_key))
2616 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2618 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2620 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2623 * preempt_notifier_unregister - no longer interested in preemption notifications
2624 * @notifier: notifier struct to unregister
2626 * This is *not* safe to call from within a preemption notifier.
2628 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2630 hlist_del(¬ifier->link);
2632 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2634 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2636 struct preempt_notifier *notifier;
2638 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2639 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2642 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2644 if (static_key_false(&preempt_notifier_key))
2645 __fire_sched_in_preempt_notifiers(curr);
2649 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2650 struct task_struct *next)
2652 struct preempt_notifier *notifier;
2654 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2655 notifier->ops->sched_out(notifier, next);
2658 static __always_inline void
2659 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2660 struct task_struct *next)
2662 if (static_key_false(&preempt_notifier_key))
2663 __fire_sched_out_preempt_notifiers(curr, next);
2666 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2668 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2673 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2674 struct task_struct *next)
2678 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2681 * prepare_task_switch - prepare to switch tasks
2682 * @rq: the runqueue preparing to switch
2683 * @prev: the current task that is being switched out
2684 * @next: the task we are going to switch to.
2686 * This is called with the rq lock held and interrupts off. It must
2687 * be paired with a subsequent finish_task_switch after the context
2690 * prepare_task_switch sets up locking and calls architecture specific
2694 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2695 struct task_struct *next)
2697 sched_info_switch(rq, prev, next);
2698 perf_event_task_sched_out(prev, next);
2699 fire_sched_out_preempt_notifiers(prev, next);
2700 prepare_lock_switch(rq, next);
2701 prepare_arch_switch(next);
2705 * finish_task_switch - clean up after a task-switch
2706 * @prev: the thread we just switched away from.
2708 * finish_task_switch must be called after the context switch, paired
2709 * with a prepare_task_switch call before the context switch.
2710 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2711 * and do any other architecture-specific cleanup actions.
2713 * Note that we may have delayed dropping an mm in context_switch(). If
2714 * so, we finish that here outside of the runqueue lock. (Doing it
2715 * with the lock held can cause deadlocks; see schedule() for
2718 * The context switch have flipped the stack from under us and restored the
2719 * local variables which were saved when this task called schedule() in the
2720 * past. prev == current is still correct but we need to recalculate this_rq
2721 * because prev may have moved to another CPU.
2723 static struct rq *finish_task_switch(struct task_struct *prev)
2724 __releases(rq->lock)
2726 struct rq *rq = this_rq();
2727 struct mm_struct *mm = rq->prev_mm;
2731 * The previous task will have left us with a preempt_count of 2
2732 * because it left us after:
2735 * preempt_disable(); // 1
2737 * raw_spin_lock_irq(&rq->lock) // 2
2739 * Also, see FORK_PREEMPT_COUNT.
2741 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2742 "corrupted preempt_count: %s/%d/0x%x\n",
2743 current->comm, current->pid, preempt_count()))
2744 preempt_count_set(FORK_PREEMPT_COUNT);
2749 * A task struct has one reference for the use as "current".
2750 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2751 * schedule one last time. The schedule call will never return, and
2752 * the scheduled task must drop that reference.
2754 * We must observe prev->state before clearing prev->on_cpu (in
2755 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2756 * running on another CPU and we could rave with its RUNNING -> DEAD
2757 * transition, resulting in a double drop.
2759 prev_state = prev->state;
2760 vtime_task_switch(prev);
2761 perf_event_task_sched_in(prev, current);
2762 finish_lock_switch(rq, prev);
2763 finish_arch_post_lock_switch();
2765 fire_sched_in_preempt_notifiers(current);
2768 if (unlikely(prev_state == TASK_DEAD)) {
2769 if (prev->sched_class->task_dead)
2770 prev->sched_class->task_dead(prev);
2773 * Remove function-return probe instances associated with this
2774 * task and put them back on the free list.
2776 kprobe_flush_task(prev);
2777 put_task_struct(prev);
2780 tick_nohz_task_switch();
2786 /* rq->lock is NOT held, but preemption is disabled */
2787 static void __balance_callback(struct rq *rq)
2789 struct callback_head *head, *next;
2790 void (*func)(struct rq *rq);
2791 unsigned long flags;
2793 raw_spin_lock_irqsave(&rq->lock, flags);
2794 head = rq->balance_callback;
2795 rq->balance_callback = NULL;
2797 func = (void (*)(struct rq *))head->func;
2804 raw_spin_unlock_irqrestore(&rq->lock, flags);
2807 static inline void balance_callback(struct rq *rq)
2809 if (unlikely(rq->balance_callback))
2810 __balance_callback(rq);
2815 static inline void balance_callback(struct rq *rq)
2822 * schedule_tail - first thing a freshly forked thread must call.
2823 * @prev: the thread we just switched away from.
2825 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2826 __releases(rq->lock)
2831 * New tasks start with FORK_PREEMPT_COUNT, see there and
2832 * finish_task_switch() for details.
2834 * finish_task_switch() will drop rq->lock() and lower preempt_count
2835 * and the preempt_enable() will end up enabling preemption (on
2836 * PREEMPT_COUNT kernels).
2839 rq = finish_task_switch(prev);
2840 balance_callback(rq);
2843 if (current->set_child_tid)
2844 put_user(task_pid_vnr(current), current->set_child_tid);
2848 * context_switch - switch to the new MM and the new thread's register state.
2850 static __always_inline struct rq *
2851 context_switch(struct rq *rq, struct task_struct *prev,
2852 struct task_struct *next, struct pin_cookie cookie)
2854 struct mm_struct *mm, *oldmm;
2856 prepare_task_switch(rq, prev, next);
2859 oldmm = prev->active_mm;
2861 * For paravirt, this is coupled with an exit in switch_to to
2862 * combine the page table reload and the switch backend into
2865 arch_start_context_switch(prev);
2868 next->active_mm = oldmm;
2869 atomic_inc(&oldmm->mm_count);
2870 enter_lazy_tlb(oldmm, next);
2872 switch_mm_irqs_off(oldmm, mm, next);
2875 prev->active_mm = NULL;
2876 rq->prev_mm = oldmm;
2879 * Since the runqueue lock will be released by the next
2880 * task (which is an invalid locking op but in the case
2881 * of the scheduler it's an obvious special-case), so we
2882 * do an early lockdep release here:
2884 lockdep_unpin_lock(&rq->lock, cookie);
2885 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2887 /* Here we just switch the register state and the stack. */
2888 switch_to(prev, next, prev);
2891 return finish_task_switch(prev);
2895 * nr_running and nr_context_switches:
2897 * externally visible scheduler statistics: current number of runnable
2898 * threads, total number of context switches performed since bootup.
2900 unsigned long nr_running(void)
2902 unsigned long i, sum = 0;
2904 for_each_online_cpu(i)
2905 sum += cpu_rq(i)->nr_running;
2911 * Check if only the current task is running on the cpu.
2913 * Caution: this function does not check that the caller has disabled
2914 * preemption, thus the result might have a time-of-check-to-time-of-use
2915 * race. The caller is responsible to use it correctly, for example:
2917 * - from a non-preemptable section (of course)
2919 * - from a thread that is bound to a single CPU
2921 * - in a loop with very short iterations (e.g. a polling loop)
2923 bool single_task_running(void)
2925 return raw_rq()->nr_running == 1;
2927 EXPORT_SYMBOL(single_task_running);
2929 unsigned long long nr_context_switches(void)
2932 unsigned long long sum = 0;
2934 for_each_possible_cpu(i)
2935 sum += cpu_rq(i)->nr_switches;
2940 unsigned long nr_iowait(void)
2942 unsigned long i, sum = 0;
2944 for_each_possible_cpu(i)
2945 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2950 unsigned long nr_iowait_cpu(int cpu)
2952 struct rq *this = cpu_rq(cpu);
2953 return atomic_read(&this->nr_iowait);
2956 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2958 struct rq *rq = this_rq();
2959 *nr_waiters = atomic_read(&rq->nr_iowait);
2960 *load = rq->load.weight;
2966 * sched_exec - execve() is a valuable balancing opportunity, because at
2967 * this point the task has the smallest effective memory and cache footprint.
2969 void sched_exec(void)
2971 struct task_struct *p = current;
2972 unsigned long flags;
2975 raw_spin_lock_irqsave(&p->pi_lock, flags);
2976 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2977 if (dest_cpu == smp_processor_id())
2980 if (likely(cpu_active(dest_cpu))) {
2981 struct migration_arg arg = { p, dest_cpu };
2983 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2984 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2988 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2993 DEFINE_PER_CPU(struct kernel_stat, kstat);
2994 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2996 EXPORT_PER_CPU_SYMBOL(kstat);
2997 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
3000 * The function fair_sched_class.update_curr accesses the struct curr
3001 * and its field curr->exec_start; when called from task_sched_runtime(),
3002 * we observe a high rate of cache misses in practice.
3003 * Prefetching this data results in improved performance.
3005 static inline void prefetch_curr_exec_start(struct task_struct *p)
3007 #ifdef CONFIG_FAIR_GROUP_SCHED
3008 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
3010 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
3013 prefetch(&curr->exec_start);
3017 * Return accounted runtime for the task.
3018 * In case the task is currently running, return the runtime plus current's
3019 * pending runtime that have not been accounted yet.
3021 unsigned long long task_sched_runtime(struct task_struct *p)
3027 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
3029 * 64-bit doesn't need locks to atomically read a 64bit value.
3030 * So we have a optimization chance when the task's delta_exec is 0.
3031 * Reading ->on_cpu is racy, but this is ok.
3033 * If we race with it leaving cpu, we'll take a lock. So we're correct.
3034 * If we race with it entering cpu, unaccounted time is 0. This is
3035 * indistinguishable from the read occurring a few cycles earlier.
3036 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
3037 * been accounted, so we're correct here as well.
3039 if (!p->on_cpu || !task_on_rq_queued(p))
3040 return p->se.sum_exec_runtime;
3043 rq = task_rq_lock(p, &rf);
3045 * Must be ->curr _and_ ->on_rq. If dequeued, we would
3046 * project cycles that may never be accounted to this
3047 * thread, breaking clock_gettime().
3049 if (task_current(rq, p) && task_on_rq_queued(p)) {
3050 prefetch_curr_exec_start(p);
3051 update_rq_clock(rq);
3052 p->sched_class->update_curr(rq);
3054 ns = p->se.sum_exec_runtime;
3055 task_rq_unlock(rq, p, &rf);
3061 * This function gets called by the timer code, with HZ frequency.
3062 * We call it with interrupts disabled.
3064 void scheduler_tick(void)
3066 int cpu = smp_processor_id();
3067 struct rq *rq = cpu_rq(cpu);
3068 struct task_struct *curr = rq->curr;
3072 raw_spin_lock(&rq->lock);
3073 update_rq_clock(rq);
3074 curr->sched_class->task_tick(rq, curr, 0);
3075 cpu_load_update_active(rq);
3076 calc_global_load_tick(rq);
3077 raw_spin_unlock(&rq->lock);
3079 perf_event_task_tick();
3082 rq->idle_balance = idle_cpu(cpu);
3083 trigger_load_balance(rq);
3085 rq_last_tick_reset(rq);
3088 #ifdef CONFIG_NO_HZ_FULL
3090 * scheduler_tick_max_deferment
3092 * Keep at least one tick per second when a single
3093 * active task is running because the scheduler doesn't
3094 * yet completely support full dynticks environment.
3096 * This makes sure that uptime, CFS vruntime, load
3097 * balancing, etc... continue to move forward, even
3098 * with a very low granularity.
3100 * Return: Maximum deferment in nanoseconds.
3102 u64 scheduler_tick_max_deferment(void)
3104 struct rq *rq = this_rq();
3105 unsigned long next, now = READ_ONCE(jiffies);
3107 next = rq->last_sched_tick + HZ;
3109 if (time_before_eq(next, now))
3112 return jiffies_to_nsecs(next - now);
3116 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3117 defined(CONFIG_PREEMPT_TRACER))
3119 * If the value passed in is equal to the current preempt count
3120 * then we just disabled preemption. Start timing the latency.
3122 static inline void preempt_latency_start(int val)
3124 if (preempt_count() == val) {
3125 unsigned long ip = get_lock_parent_ip();
3126 #ifdef CONFIG_DEBUG_PREEMPT
3127 current->preempt_disable_ip = ip;
3129 trace_preempt_off(CALLER_ADDR0, ip);
3133 void preempt_count_add(int val)
3135 #ifdef CONFIG_DEBUG_PREEMPT
3139 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3142 __preempt_count_add(val);
3143 #ifdef CONFIG_DEBUG_PREEMPT
3145 * Spinlock count overflowing soon?
3147 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3150 preempt_latency_start(val);
3152 EXPORT_SYMBOL(preempt_count_add);
3153 NOKPROBE_SYMBOL(preempt_count_add);
3156 * If the value passed in equals to the current preempt count
3157 * then we just enabled preemption. Stop timing the latency.
3159 static inline void preempt_latency_stop(int val)
3161 if (preempt_count() == val)
3162 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3165 void preempt_count_sub(int val)
3167 #ifdef CONFIG_DEBUG_PREEMPT
3171 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3174 * Is the spinlock portion underflowing?
3176 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3177 !(preempt_count() & PREEMPT_MASK)))
3181 preempt_latency_stop(val);
3182 __preempt_count_sub(val);
3184 EXPORT_SYMBOL(preempt_count_sub);
3185 NOKPROBE_SYMBOL(preempt_count_sub);
3188 static inline void preempt_latency_start(int val) { }
3189 static inline void preempt_latency_stop(int val) { }
3193 * Print scheduling while atomic bug:
3195 static noinline void __schedule_bug(struct task_struct *prev)
3197 /* Save this before calling printk(), since that will clobber it */
3198 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
3200 if (oops_in_progress)
3203 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3204 prev->comm, prev->pid, preempt_count());
3206 debug_show_held_locks(prev);
3208 if (irqs_disabled())
3209 print_irqtrace_events(prev);
3210 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
3211 && in_atomic_preempt_off()) {
3212 pr_err("Preemption disabled at:");
3213 print_ip_sym(preempt_disable_ip);
3217 panic("scheduling while atomic\n");
3220 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3224 * Various schedule()-time debugging checks and statistics:
3226 static inline void schedule_debug(struct task_struct *prev)
3228 #ifdef CONFIG_SCHED_STACK_END_CHECK
3229 if (task_stack_end_corrupted(prev))
3230 panic("corrupted stack end detected inside scheduler\n");
3233 if (unlikely(in_atomic_preempt_off())) {
3234 __schedule_bug(prev);
3235 preempt_count_set(PREEMPT_DISABLED);
3239 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3241 schedstat_inc(this_rq()->sched_count);
3245 * Pick up the highest-prio task:
3247 static inline struct task_struct *
3248 pick_next_task(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
3250 const struct sched_class *class = &fair_sched_class;
3251 struct task_struct *p;
3254 * Optimization: we know that if all tasks are in
3255 * the fair class we can call that function directly:
3257 if (likely(prev->sched_class == class &&
3258 rq->nr_running == rq->cfs.h_nr_running)) {
3259 p = fair_sched_class.pick_next_task(rq, prev, cookie);
3260 if (unlikely(p == RETRY_TASK))
3263 /* assumes fair_sched_class->next == idle_sched_class */
3265 p = idle_sched_class.pick_next_task(rq, prev, cookie);
3271 for_each_class(class) {
3272 p = class->pick_next_task(rq, prev, cookie);
3274 if (unlikely(p == RETRY_TASK))
3280 BUG(); /* the idle class will always have a runnable task */
3284 * __schedule() is the main scheduler function.
3286 * The main means of driving the scheduler and thus entering this function are:
3288 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3290 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3291 * paths. For example, see arch/x86/entry_64.S.
3293 * To drive preemption between tasks, the scheduler sets the flag in timer
3294 * interrupt handler scheduler_tick().
3296 * 3. Wakeups don't really cause entry into schedule(). They add a
3297 * task to the run-queue and that's it.
3299 * Now, if the new task added to the run-queue preempts the current
3300 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3301 * called on the nearest possible occasion:
3303 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3305 * - in syscall or exception context, at the next outmost
3306 * preempt_enable(). (this might be as soon as the wake_up()'s
3309 * - in IRQ context, return from interrupt-handler to
3310 * preemptible context
3312 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3315 * - cond_resched() call
3316 * - explicit schedule() call
3317 * - return from syscall or exception to user-space
3318 * - return from interrupt-handler to user-space
3320 * WARNING: must be called with preemption disabled!
3322 static void __sched notrace __schedule(bool preempt)
3324 struct task_struct *prev, *next;
3325 unsigned long *switch_count;
3326 struct pin_cookie cookie;
3330 cpu = smp_processor_id();
3334 schedule_debug(prev);
3336 if (sched_feat(HRTICK))
3339 local_irq_disable();
3340 rcu_note_context_switch();
3343 * Make sure that signal_pending_state()->signal_pending() below
3344 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3345 * done by the caller to avoid the race with signal_wake_up().
3347 smp_mb__before_spinlock();
3348 raw_spin_lock(&rq->lock);
3349 cookie = lockdep_pin_lock(&rq->lock);
3351 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3353 switch_count = &prev->nivcsw;
3354 if (!preempt && prev->state) {
3355 if (unlikely(signal_pending_state(prev->state, prev))) {
3356 prev->state = TASK_RUNNING;
3358 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3362 * If a worker went to sleep, notify and ask workqueue
3363 * whether it wants to wake up a task to maintain
3366 if (prev->flags & PF_WQ_WORKER) {
3367 struct task_struct *to_wakeup;
3369 to_wakeup = wq_worker_sleeping(prev);
3371 try_to_wake_up_local(to_wakeup, cookie);
3374 switch_count = &prev->nvcsw;
3377 if (task_on_rq_queued(prev))
3378 update_rq_clock(rq);
3380 next = pick_next_task(rq, prev, cookie);
3381 clear_tsk_need_resched(prev);
3382 clear_preempt_need_resched();
3383 rq->clock_skip_update = 0;
3385 if (likely(prev != next)) {
3390 trace_sched_switch(preempt, prev, next);
3391 rq = context_switch(rq, prev, next, cookie); /* unlocks the rq */
3393 lockdep_unpin_lock(&rq->lock, cookie);
3394 raw_spin_unlock_irq(&rq->lock);
3397 balance_callback(rq);
3399 STACK_FRAME_NON_STANDARD(__schedule); /* switch_to() */
3401 void __noreturn do_task_dead(void)
3404 * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
3405 * when the following two conditions become true.
3406 * - There is race condition of mmap_sem (It is acquired by
3408 * - SMI occurs before setting TASK_RUNINNG.
3409 * (or hypervisor of virtual machine switches to other guest)
3410 * As a result, we may become TASK_RUNNING after becoming TASK_DEAD
3412 * To avoid it, we have to wait for releasing tsk->pi_lock which
3413 * is held by try_to_wake_up()
3416 raw_spin_unlock_wait(¤t->pi_lock);
3418 /* causes final put_task_struct in finish_task_switch(). */
3419 __set_current_state(TASK_DEAD);
3420 current->flags |= PF_NOFREEZE; /* tell freezer to ignore us */
3423 /* Avoid "noreturn function does return". */
3425 cpu_relax(); /* For when BUG is null */
3428 static inline void sched_submit_work(struct task_struct *tsk)
3430 if (!tsk->state || tsk_is_pi_blocked(tsk))
3433 * If we are going to sleep and we have plugged IO queued,
3434 * make sure to submit it to avoid deadlocks.
3436 if (blk_needs_flush_plug(tsk))
3437 blk_schedule_flush_plug(tsk);
3440 asmlinkage __visible void __sched schedule(void)
3442 struct task_struct *tsk = current;
3444 sched_submit_work(tsk);
3448 sched_preempt_enable_no_resched();
3449 } while (need_resched());
3451 EXPORT_SYMBOL(schedule);
3453 #ifdef CONFIG_CONTEXT_TRACKING
3454 asmlinkage __visible void __sched schedule_user(void)
3457 * If we come here after a random call to set_need_resched(),
3458 * or we have been woken up remotely but the IPI has not yet arrived,
3459 * we haven't yet exited the RCU idle mode. Do it here manually until
3460 * we find a better solution.
3462 * NB: There are buggy callers of this function. Ideally we
3463 * should warn if prev_state != CONTEXT_USER, but that will trigger
3464 * too frequently to make sense yet.
3466 enum ctx_state prev_state = exception_enter();
3468 exception_exit(prev_state);
3473 * schedule_preempt_disabled - called with preemption disabled
3475 * Returns with preemption disabled. Note: preempt_count must be 1
3477 void __sched schedule_preempt_disabled(void)
3479 sched_preempt_enable_no_resched();
3484 static void __sched notrace preempt_schedule_common(void)
3488 * Because the function tracer can trace preempt_count_sub()
3489 * and it also uses preempt_enable/disable_notrace(), if
3490 * NEED_RESCHED is set, the preempt_enable_notrace() called
3491 * by the function tracer will call this function again and
3492 * cause infinite recursion.
3494 * Preemption must be disabled here before the function
3495 * tracer can trace. Break up preempt_disable() into two
3496 * calls. One to disable preemption without fear of being
3497 * traced. The other to still record the preemption latency,
3498 * which can also be traced by the function tracer.
3500 preempt_disable_notrace();
3501 preempt_latency_start(1);
3503 preempt_latency_stop(1);
3504 preempt_enable_no_resched_notrace();
3507 * Check again in case we missed a preemption opportunity
3508 * between schedule and now.
3510 } while (need_resched());
3513 #ifdef CONFIG_PREEMPT
3515 * this is the entry point to schedule() from in-kernel preemption
3516 * off of preempt_enable. Kernel preemptions off return from interrupt
3517 * occur there and call schedule directly.
3519 asmlinkage __visible void __sched notrace preempt_schedule(void)
3522 * If there is a non-zero preempt_count or interrupts are disabled,
3523 * we do not want to preempt the current task. Just return..
3525 if (likely(!preemptible()))
3528 preempt_schedule_common();
3530 NOKPROBE_SYMBOL(preempt_schedule);
3531 EXPORT_SYMBOL(preempt_schedule);
3534 * preempt_schedule_notrace - preempt_schedule called by tracing
3536 * The tracing infrastructure uses preempt_enable_notrace to prevent
3537 * recursion and tracing preempt enabling caused by the tracing
3538 * infrastructure itself. But as tracing can happen in areas coming
3539 * from userspace or just about to enter userspace, a preempt enable
3540 * can occur before user_exit() is called. This will cause the scheduler
3541 * to be called when the system is still in usermode.
3543 * To prevent this, the preempt_enable_notrace will use this function
3544 * instead of preempt_schedule() to exit user context if needed before
3545 * calling the scheduler.
3547 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3549 enum ctx_state prev_ctx;
3551 if (likely(!preemptible()))
3556 * Because the function tracer can trace preempt_count_sub()
3557 * and it also uses preempt_enable/disable_notrace(), if
3558 * NEED_RESCHED is set, the preempt_enable_notrace() called
3559 * by the function tracer will call this function again and
3560 * cause infinite recursion.
3562 * Preemption must be disabled here before the function
3563 * tracer can trace. Break up preempt_disable() into two
3564 * calls. One to disable preemption without fear of being
3565 * traced. The other to still record the preemption latency,
3566 * which can also be traced by the function tracer.
3568 preempt_disable_notrace();
3569 preempt_latency_start(1);
3571 * Needs preempt disabled in case user_exit() is traced
3572 * and the tracer calls preempt_enable_notrace() causing
3573 * an infinite recursion.
3575 prev_ctx = exception_enter();
3577 exception_exit(prev_ctx);
3579 preempt_latency_stop(1);
3580 preempt_enable_no_resched_notrace();
3581 } while (need_resched());
3583 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3585 #endif /* CONFIG_PREEMPT */
3588 * this is the entry point to schedule() from kernel preemption
3589 * off of irq context.
3590 * Note, that this is called and return with irqs disabled. This will
3591 * protect us against recursive calling from irq.
3593 asmlinkage __visible void __sched preempt_schedule_irq(void)
3595 enum ctx_state prev_state;
3597 /* Catch callers which need to be fixed */
3598 BUG_ON(preempt_count() || !irqs_disabled());
3600 prev_state = exception_enter();
3606 local_irq_disable();
3607 sched_preempt_enable_no_resched();
3608 } while (need_resched());
3610 exception_exit(prev_state);
3613 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3616 return try_to_wake_up(curr->private, mode, wake_flags);
3618 EXPORT_SYMBOL(default_wake_function);
3620 #ifdef CONFIG_RT_MUTEXES
3623 * rt_mutex_setprio - set the current priority of a task
3625 * @prio: prio value (kernel-internal form)
3627 * This function changes the 'effective' priority of a task. It does
3628 * not touch ->normal_prio like __setscheduler().
3630 * Used by the rt_mutex code to implement priority inheritance
3631 * logic. Call site only calls if the priority of the task changed.
3633 void rt_mutex_setprio(struct task_struct *p, int prio)
3635 int oldprio, queued, running, queue_flag = DEQUEUE_SAVE | DEQUEUE_MOVE;
3636 const struct sched_class *prev_class;
3640 BUG_ON(prio > MAX_PRIO);
3642 rq = __task_rq_lock(p, &rf);
3645 * Idle task boosting is a nono in general. There is one
3646 * exception, when PREEMPT_RT and NOHZ is active:
3648 * The idle task calls get_next_timer_interrupt() and holds
3649 * the timer wheel base->lock on the CPU and another CPU wants
3650 * to access the timer (probably to cancel it). We can safely
3651 * ignore the boosting request, as the idle CPU runs this code
3652 * with interrupts disabled and will complete the lock
3653 * protected section without being interrupted. So there is no
3654 * real need to boost.
3656 if (unlikely(p == rq->idle)) {
3657 WARN_ON(p != rq->curr);
3658 WARN_ON(p->pi_blocked_on);
3662 trace_sched_pi_setprio(p, prio);
3665 if (oldprio == prio)
3666 queue_flag &= ~DEQUEUE_MOVE;
3668 prev_class = p->sched_class;
3669 queued = task_on_rq_queued(p);
3670 running = task_current(rq, p);
3672 dequeue_task(rq, p, queue_flag);
3674 put_prev_task(rq, p);
3677 * Boosting condition are:
3678 * 1. -rt task is running and holds mutex A
3679 * --> -dl task blocks on mutex A
3681 * 2. -dl task is running and holds mutex A
3682 * --> -dl task blocks on mutex A and could preempt the
3685 if (dl_prio(prio)) {
3686 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3687 if (!dl_prio(p->normal_prio) ||
3688 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3689 p->dl.dl_boosted = 1;
3690 queue_flag |= ENQUEUE_REPLENISH;
3692 p->dl.dl_boosted = 0;
3693 p->sched_class = &dl_sched_class;
3694 } else if (rt_prio(prio)) {
3695 if (dl_prio(oldprio))
3696 p->dl.dl_boosted = 0;
3698 queue_flag |= ENQUEUE_HEAD;
3699 p->sched_class = &rt_sched_class;
3701 if (dl_prio(oldprio))
3702 p->dl.dl_boosted = 0;
3703 if (rt_prio(oldprio))
3705 p->sched_class = &fair_sched_class;
3711 p->sched_class->set_curr_task(rq);
3713 enqueue_task(rq, p, queue_flag);
3715 check_class_changed(rq, p, prev_class, oldprio);
3717 preempt_disable(); /* avoid rq from going away on us */
3718 __task_rq_unlock(rq, &rf);
3720 balance_callback(rq);
3725 void set_user_nice(struct task_struct *p, long nice)
3727 int old_prio, delta, queued;
3731 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3734 * We have to be careful, if called from sys_setpriority(),
3735 * the task might be in the middle of scheduling on another CPU.
3737 rq = task_rq_lock(p, &rf);
3739 * The RT priorities are set via sched_setscheduler(), but we still
3740 * allow the 'normal' nice value to be set - but as expected
3741 * it wont have any effect on scheduling until the task is
3742 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3744 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3745 p->static_prio = NICE_TO_PRIO(nice);
3748 queued = task_on_rq_queued(p);
3750 dequeue_task(rq, p, DEQUEUE_SAVE);
3752 p->static_prio = NICE_TO_PRIO(nice);
3755 p->prio = effective_prio(p);
3756 delta = p->prio - old_prio;
3759 enqueue_task(rq, p, ENQUEUE_RESTORE);
3761 * If the task increased its priority or is running and
3762 * lowered its priority, then reschedule its CPU:
3764 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3768 task_rq_unlock(rq, p, &rf);
3770 EXPORT_SYMBOL(set_user_nice);
3773 * can_nice - check if a task can reduce its nice value
3777 int can_nice(const struct task_struct *p, const int nice)
3779 /* convert nice value [19,-20] to rlimit style value [1,40] */
3780 int nice_rlim = nice_to_rlimit(nice);
3782 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3783 capable(CAP_SYS_NICE));
3786 #ifdef __ARCH_WANT_SYS_NICE
3789 * sys_nice - change the priority of the current process.
3790 * @increment: priority increment
3792 * sys_setpriority is a more generic, but much slower function that
3793 * does similar things.
3795 SYSCALL_DEFINE1(nice, int, increment)
3800 * Setpriority might change our priority at the same moment.
3801 * We don't have to worry. Conceptually one call occurs first
3802 * and we have a single winner.
3804 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3805 nice = task_nice(current) + increment;
3807 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3808 if (increment < 0 && !can_nice(current, nice))
3811 retval = security_task_setnice(current, nice);
3815 set_user_nice(current, nice);
3822 * task_prio - return the priority value of a given task.
3823 * @p: the task in question.
3825 * Return: The priority value as seen by users in /proc.
3826 * RT tasks are offset by -200. Normal tasks are centered
3827 * around 0, value goes from -16 to +15.
3829 int task_prio(const struct task_struct *p)
3831 return p->prio - MAX_RT_PRIO;
3835 * idle_cpu - is a given cpu idle currently?
3836 * @cpu: the processor in question.
3838 * Return: 1 if the CPU is currently idle. 0 otherwise.
3840 int idle_cpu(int cpu)
3842 struct rq *rq = cpu_rq(cpu);
3844 if (rq->curr != rq->idle)
3851 if (!llist_empty(&rq->wake_list))
3859 * idle_task - return the idle task for a given cpu.
3860 * @cpu: the processor in question.
3862 * Return: The idle task for the cpu @cpu.
3864 struct task_struct *idle_task(int cpu)
3866 return cpu_rq(cpu)->idle;
3870 * find_process_by_pid - find a process with a matching PID value.
3871 * @pid: the pid in question.
3873 * The task of @pid, if found. %NULL otherwise.
3875 static struct task_struct *find_process_by_pid(pid_t pid)
3877 return pid ? find_task_by_vpid(pid) : current;
3881 * This function initializes the sched_dl_entity of a newly becoming
3882 * SCHED_DEADLINE task.
3884 * Only the static values are considered here, the actual runtime and the
3885 * absolute deadline will be properly calculated when the task is enqueued
3886 * for the first time with its new policy.
3889 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3891 struct sched_dl_entity *dl_se = &p->dl;
3893 dl_se->dl_runtime = attr->sched_runtime;
3894 dl_se->dl_deadline = attr->sched_deadline;
3895 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3896 dl_se->flags = attr->sched_flags;
3897 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3900 * Changing the parameters of a task is 'tricky' and we're not doing
3901 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3903 * What we SHOULD do is delay the bandwidth release until the 0-lag
3904 * point. This would include retaining the task_struct until that time
3905 * and change dl_overflow() to not immediately decrement the current
3908 * Instead we retain the current runtime/deadline and let the new
3909 * parameters take effect after the current reservation period lapses.
3910 * This is safe (albeit pessimistic) because the 0-lag point is always
3911 * before the current scheduling deadline.
3913 * We can still have temporary overloads because we do not delay the
3914 * change in bandwidth until that time; so admission control is
3915 * not on the safe side. It does however guarantee tasks will never
3916 * consume more than promised.
3921 * sched_setparam() passes in -1 for its policy, to let the functions
3922 * it calls know not to change it.
3924 #define SETPARAM_POLICY -1
3926 static void __setscheduler_params(struct task_struct *p,
3927 const struct sched_attr *attr)
3929 int policy = attr->sched_policy;
3931 if (policy == SETPARAM_POLICY)
3936 if (dl_policy(policy))
3937 __setparam_dl(p, attr);
3938 else if (fair_policy(policy))
3939 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3942 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3943 * !rt_policy. Always setting this ensures that things like
3944 * getparam()/getattr() don't report silly values for !rt tasks.
3946 p->rt_priority = attr->sched_priority;
3947 p->normal_prio = normal_prio(p);
3951 /* Actually do priority change: must hold pi & rq lock. */
3952 static void __setscheduler(struct rq *rq, struct task_struct *p,
3953 const struct sched_attr *attr, bool keep_boost)
3955 __setscheduler_params(p, attr);
3958 * Keep a potential priority boosting if called from
3959 * sched_setscheduler().
3962 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3964 p->prio = normal_prio(p);
3966 if (dl_prio(p->prio))
3967 p->sched_class = &dl_sched_class;
3968 else if (rt_prio(p->prio))
3969 p->sched_class = &rt_sched_class;
3971 p->sched_class = &fair_sched_class;
3975 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3977 struct sched_dl_entity *dl_se = &p->dl;
3979 attr->sched_priority = p->rt_priority;
3980 attr->sched_runtime = dl_se->dl_runtime;
3981 attr->sched_deadline = dl_se->dl_deadline;
3982 attr->sched_period = dl_se->dl_period;
3983 attr->sched_flags = dl_se->flags;
3987 * This function validates the new parameters of a -deadline task.
3988 * We ask for the deadline not being zero, and greater or equal
3989 * than the runtime, as well as the period of being zero or
3990 * greater than deadline. Furthermore, we have to be sure that
3991 * user parameters are above the internal resolution of 1us (we
3992 * check sched_runtime only since it is always the smaller one) and
3993 * below 2^63 ns (we have to check both sched_deadline and
3994 * sched_period, as the latter can be zero).
3997 __checkparam_dl(const struct sched_attr *attr)
4000 if (attr->sched_deadline == 0)
4004 * Since we truncate DL_SCALE bits, make sure we're at least
4007 if (attr->sched_runtime < (1ULL << DL_SCALE))
4011 * Since we use the MSB for wrap-around and sign issues, make
4012 * sure it's not set (mind that period can be equal to zero).
4014 if (attr->sched_deadline & (1ULL << 63) ||
4015 attr->sched_period & (1ULL << 63))
4018 /* runtime <= deadline <= period (if period != 0) */
4019 if ((attr->sched_period != 0 &&
4020 attr->sched_period < attr->sched_deadline) ||
4021 attr->sched_deadline < attr->sched_runtime)
4028 * check the target process has a UID that matches the current process's
4030 static bool check_same_owner(struct task_struct *p)
4032 const struct cred *cred = current_cred(), *pcred;
4036 pcred = __task_cred(p);
4037 match = (uid_eq(cred->euid, pcred->euid) ||
4038 uid_eq(cred->euid, pcred->uid));
4043 static bool dl_param_changed(struct task_struct *p,
4044 const struct sched_attr *attr)
4046 struct sched_dl_entity *dl_se = &p->dl;
4048 if (dl_se->dl_runtime != attr->sched_runtime ||
4049 dl_se->dl_deadline != attr->sched_deadline ||
4050 dl_se->dl_period != attr->sched_period ||
4051 dl_se->flags != attr->sched_flags)
4057 static int __sched_setscheduler(struct task_struct *p,
4058 const struct sched_attr *attr,
4061 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4062 MAX_RT_PRIO - 1 - attr->sched_priority;
4063 int retval, oldprio, oldpolicy = -1, queued, running;
4064 int new_effective_prio, policy = attr->sched_policy;
4065 const struct sched_class *prev_class;
4068 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE;
4071 /* may grab non-irq protected spin_locks */
4072 BUG_ON(in_interrupt());
4074 /* double check policy once rq lock held */
4076 reset_on_fork = p->sched_reset_on_fork;
4077 policy = oldpolicy = p->policy;
4079 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4081 if (!valid_policy(policy))
4085 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4089 * Valid priorities for SCHED_FIFO and SCHED_RR are
4090 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4091 * SCHED_BATCH and SCHED_IDLE is 0.
4093 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4094 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4096 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4097 (rt_policy(policy) != (attr->sched_priority != 0)))
4101 * Allow unprivileged RT tasks to decrease priority:
4103 if (user && !capable(CAP_SYS_NICE)) {
4104 if (fair_policy(policy)) {
4105 if (attr->sched_nice < task_nice(p) &&
4106 !can_nice(p, attr->sched_nice))
4110 if (rt_policy(policy)) {
4111 unsigned long rlim_rtprio =
4112 task_rlimit(p, RLIMIT_RTPRIO);
4114 /* can't set/change the rt policy */
4115 if (policy != p->policy && !rlim_rtprio)
4118 /* can't increase priority */
4119 if (attr->sched_priority > p->rt_priority &&
4120 attr->sched_priority > rlim_rtprio)
4125 * Can't set/change SCHED_DEADLINE policy at all for now
4126 * (safest behavior); in the future we would like to allow
4127 * unprivileged DL tasks to increase their relative deadline
4128 * or reduce their runtime (both ways reducing utilization)
4130 if (dl_policy(policy))
4134 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4135 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4137 if (idle_policy(p->policy) && !idle_policy(policy)) {
4138 if (!can_nice(p, task_nice(p)))
4142 /* can't change other user's priorities */
4143 if (!check_same_owner(p))
4146 /* Normal users shall not reset the sched_reset_on_fork flag */
4147 if (p->sched_reset_on_fork && !reset_on_fork)
4152 retval = security_task_setscheduler(p);
4158 * make sure no PI-waiters arrive (or leave) while we are
4159 * changing the priority of the task:
4161 * To be able to change p->policy safely, the appropriate
4162 * runqueue lock must be held.
4164 rq = task_rq_lock(p, &rf);
4167 * Changing the policy of the stop threads its a very bad idea
4169 if (p == rq->stop) {
4170 task_rq_unlock(rq, p, &rf);
4175 * If not changing anything there's no need to proceed further,
4176 * but store a possible modification of reset_on_fork.
4178 if (unlikely(policy == p->policy)) {
4179 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4181 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4183 if (dl_policy(policy) && dl_param_changed(p, attr))
4186 p->sched_reset_on_fork = reset_on_fork;
4187 task_rq_unlock(rq, p, &rf);
4193 #ifdef CONFIG_RT_GROUP_SCHED
4195 * Do not allow realtime tasks into groups that have no runtime
4198 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4199 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4200 !task_group_is_autogroup(task_group(p))) {
4201 task_rq_unlock(rq, p, &rf);
4206 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4207 cpumask_t *span = rq->rd->span;
4210 * Don't allow tasks with an affinity mask smaller than
4211 * the entire root_domain to become SCHED_DEADLINE. We
4212 * will also fail if there's no bandwidth available.
4214 if (!cpumask_subset(span, &p->cpus_allowed) ||
4215 rq->rd->dl_bw.bw == 0) {
4216 task_rq_unlock(rq, p, &rf);
4223 /* recheck policy now with rq lock held */
4224 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4225 policy = oldpolicy = -1;
4226 task_rq_unlock(rq, p, &rf);
4231 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4232 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4235 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4236 task_rq_unlock(rq, p, &rf);
4240 p->sched_reset_on_fork = reset_on_fork;
4245 * Take priority boosted tasks into account. If the new
4246 * effective priority is unchanged, we just store the new
4247 * normal parameters and do not touch the scheduler class and
4248 * the runqueue. This will be done when the task deboost
4251 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4252 if (new_effective_prio == oldprio)
4253 queue_flags &= ~DEQUEUE_MOVE;
4256 queued = task_on_rq_queued(p);
4257 running = task_current(rq, p);
4259 dequeue_task(rq, p, queue_flags);
4261 put_prev_task(rq, p);
4263 prev_class = p->sched_class;
4264 __setscheduler(rq, p, attr, pi);
4267 p->sched_class->set_curr_task(rq);
4270 * We enqueue to tail when the priority of a task is
4271 * increased (user space view).
4273 if (oldprio < p->prio)
4274 queue_flags |= ENQUEUE_HEAD;
4276 enqueue_task(rq, p, queue_flags);
4279 check_class_changed(rq, p, prev_class, oldprio);
4280 preempt_disable(); /* avoid rq from going away on us */
4281 task_rq_unlock(rq, p, &rf);
4284 rt_mutex_adjust_pi(p);
4287 * Run balance callbacks after we've adjusted the PI chain.
4289 balance_callback(rq);
4295 static int _sched_setscheduler(struct task_struct *p, int policy,
4296 const struct sched_param *param, bool check)
4298 struct sched_attr attr = {
4299 .sched_policy = policy,
4300 .sched_priority = param->sched_priority,
4301 .sched_nice = PRIO_TO_NICE(p->static_prio),
4304 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4305 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4306 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4307 policy &= ~SCHED_RESET_ON_FORK;
4308 attr.sched_policy = policy;
4311 return __sched_setscheduler(p, &attr, check, true);
4314 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4315 * @p: the task in question.
4316 * @policy: new policy.
4317 * @param: structure containing the new RT priority.
4319 * Return: 0 on success. An error code otherwise.
4321 * NOTE that the task may be already dead.
4323 int sched_setscheduler(struct task_struct *p, int policy,
4324 const struct sched_param *param)
4326 return _sched_setscheduler(p, policy, param, true);
4328 EXPORT_SYMBOL_GPL(sched_setscheduler);
4330 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4332 return __sched_setscheduler(p, attr, true, true);
4334 EXPORT_SYMBOL_GPL(sched_setattr);
4337 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4338 * @p: the task in question.
4339 * @policy: new policy.
4340 * @param: structure containing the new RT priority.
4342 * Just like sched_setscheduler, only don't bother checking if the
4343 * current context has permission. For example, this is needed in
4344 * stop_machine(): we create temporary high priority worker threads,
4345 * but our caller might not have that capability.
4347 * Return: 0 on success. An error code otherwise.
4349 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4350 const struct sched_param *param)
4352 return _sched_setscheduler(p, policy, param, false);
4354 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4357 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4359 struct sched_param lparam;
4360 struct task_struct *p;
4363 if (!param || pid < 0)
4365 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4370 p = find_process_by_pid(pid);
4372 retval = sched_setscheduler(p, policy, &lparam);
4379 * Mimics kernel/events/core.c perf_copy_attr().
4381 static int sched_copy_attr(struct sched_attr __user *uattr,
4382 struct sched_attr *attr)
4387 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4391 * zero the full structure, so that a short copy will be nice.
4393 memset(attr, 0, sizeof(*attr));
4395 ret = get_user(size, &uattr->size);
4399 if (size > PAGE_SIZE) /* silly large */
4402 if (!size) /* abi compat */
4403 size = SCHED_ATTR_SIZE_VER0;
4405 if (size < SCHED_ATTR_SIZE_VER0)
4409 * If we're handed a bigger struct than we know of,
4410 * ensure all the unknown bits are 0 - i.e. new
4411 * user-space does not rely on any kernel feature
4412 * extensions we dont know about yet.
4414 if (size > sizeof(*attr)) {
4415 unsigned char __user *addr;
4416 unsigned char __user *end;
4419 addr = (void __user *)uattr + sizeof(*attr);
4420 end = (void __user *)uattr + size;
4422 for (; addr < end; addr++) {
4423 ret = get_user(val, addr);
4429 size = sizeof(*attr);
4432 ret = copy_from_user(attr, uattr, size);
4437 * XXX: do we want to be lenient like existing syscalls; or do we want
4438 * to be strict and return an error on out-of-bounds values?
4440 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4445 put_user(sizeof(*attr), &uattr->size);
4450 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4451 * @pid: the pid in question.
4452 * @policy: new policy.
4453 * @param: structure containing the new RT priority.
4455 * Return: 0 on success. An error code otherwise.
4457 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4458 struct sched_param __user *, param)
4460 /* negative values for policy are not valid */
4464 return do_sched_setscheduler(pid, policy, param);
4468 * sys_sched_setparam - set/change the RT priority of a thread
4469 * @pid: the pid in question.
4470 * @param: structure containing the new RT priority.
4472 * Return: 0 on success. An error code otherwise.
4474 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4476 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4480 * sys_sched_setattr - same as above, but with extended sched_attr
4481 * @pid: the pid in question.
4482 * @uattr: structure containing the extended parameters.
4483 * @flags: for future extension.
4485 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4486 unsigned int, flags)
4488 struct sched_attr attr;
4489 struct task_struct *p;
4492 if (!uattr || pid < 0 || flags)
4495 retval = sched_copy_attr(uattr, &attr);
4499 if ((int)attr.sched_policy < 0)
4504 p = find_process_by_pid(pid);
4506 retval = sched_setattr(p, &attr);
4513 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4514 * @pid: the pid in question.
4516 * Return: On success, the policy of the thread. Otherwise, a negative error
4519 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4521 struct task_struct *p;
4529 p = find_process_by_pid(pid);
4531 retval = security_task_getscheduler(p);
4534 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4541 * sys_sched_getparam - get the RT priority of a thread
4542 * @pid: the pid in question.
4543 * @param: structure containing the RT priority.
4545 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4548 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4550 struct sched_param lp = { .sched_priority = 0 };
4551 struct task_struct *p;
4554 if (!param || pid < 0)
4558 p = find_process_by_pid(pid);
4563 retval = security_task_getscheduler(p);
4567 if (task_has_rt_policy(p))
4568 lp.sched_priority = p->rt_priority;
4572 * This one might sleep, we cannot do it with a spinlock held ...
4574 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4583 static int sched_read_attr(struct sched_attr __user *uattr,
4584 struct sched_attr *attr,
4589 if (!access_ok(VERIFY_WRITE, uattr, usize))
4593 * If we're handed a smaller struct than we know of,
4594 * ensure all the unknown bits are 0 - i.e. old
4595 * user-space does not get uncomplete information.
4597 if (usize < sizeof(*attr)) {
4598 unsigned char *addr;
4601 addr = (void *)attr + usize;
4602 end = (void *)attr + sizeof(*attr);
4604 for (; addr < end; addr++) {
4612 ret = copy_to_user(uattr, attr, attr->size);
4620 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4621 * @pid: the pid in question.
4622 * @uattr: structure containing the extended parameters.
4623 * @size: sizeof(attr) for fwd/bwd comp.
4624 * @flags: for future extension.
4626 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4627 unsigned int, size, unsigned int, flags)
4629 struct sched_attr attr = {
4630 .size = sizeof(struct sched_attr),
4632 struct task_struct *p;
4635 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4636 size < SCHED_ATTR_SIZE_VER0 || flags)
4640 p = find_process_by_pid(pid);
4645 retval = security_task_getscheduler(p);
4649 attr.sched_policy = p->policy;
4650 if (p->sched_reset_on_fork)
4651 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4652 if (task_has_dl_policy(p))
4653 __getparam_dl(p, &attr);
4654 else if (task_has_rt_policy(p))
4655 attr.sched_priority = p->rt_priority;
4657 attr.sched_nice = task_nice(p);
4661 retval = sched_read_attr(uattr, &attr, size);
4669 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4671 cpumask_var_t cpus_allowed, new_mask;
4672 struct task_struct *p;
4677 p = find_process_by_pid(pid);
4683 /* Prevent p going away */
4687 if (p->flags & PF_NO_SETAFFINITY) {
4691 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4695 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4697 goto out_free_cpus_allowed;
4700 if (!check_same_owner(p)) {
4702 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4704 goto out_free_new_mask;
4709 retval = security_task_setscheduler(p);
4711 goto out_free_new_mask;
4714 cpuset_cpus_allowed(p, cpus_allowed);
4715 cpumask_and(new_mask, in_mask, cpus_allowed);
4718 * Since bandwidth control happens on root_domain basis,
4719 * if admission test is enabled, we only admit -deadline
4720 * tasks allowed to run on all the CPUs in the task's
4724 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4726 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4729 goto out_free_new_mask;
4735 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4738 cpuset_cpus_allowed(p, cpus_allowed);
4739 if (!cpumask_subset(new_mask, cpus_allowed)) {
4741 * We must have raced with a concurrent cpuset
4742 * update. Just reset the cpus_allowed to the
4743 * cpuset's cpus_allowed
4745 cpumask_copy(new_mask, cpus_allowed);
4750 free_cpumask_var(new_mask);
4751 out_free_cpus_allowed:
4752 free_cpumask_var(cpus_allowed);
4758 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4759 struct cpumask *new_mask)
4761 if (len < cpumask_size())
4762 cpumask_clear(new_mask);
4763 else if (len > cpumask_size())
4764 len = cpumask_size();
4766 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4770 * sys_sched_setaffinity - set the cpu affinity of a process
4771 * @pid: pid of the process
4772 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4773 * @user_mask_ptr: user-space pointer to the new cpu mask
4775 * Return: 0 on success. An error code otherwise.
4777 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4778 unsigned long __user *, user_mask_ptr)
4780 cpumask_var_t new_mask;
4783 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4786 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4788 retval = sched_setaffinity(pid, new_mask);
4789 free_cpumask_var(new_mask);
4793 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4795 struct task_struct *p;
4796 unsigned long flags;
4802 p = find_process_by_pid(pid);
4806 retval = security_task_getscheduler(p);
4810 raw_spin_lock_irqsave(&p->pi_lock, flags);
4811 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4812 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4821 * sys_sched_getaffinity - get the cpu affinity of a process
4822 * @pid: pid of the process
4823 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4824 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4826 * Return: size of CPU mask copied to user_mask_ptr on success. An
4827 * error code otherwise.
4829 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4830 unsigned long __user *, user_mask_ptr)
4835 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4837 if (len & (sizeof(unsigned long)-1))
4840 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4843 ret = sched_getaffinity(pid, mask);
4845 size_t retlen = min_t(size_t, len, cpumask_size());
4847 if (copy_to_user(user_mask_ptr, mask, retlen))
4852 free_cpumask_var(mask);
4858 * sys_sched_yield - yield the current processor to other threads.
4860 * This function yields the current CPU to other tasks. If there are no
4861 * other threads running on this CPU then this function will return.
4865 SYSCALL_DEFINE0(sched_yield)
4867 struct rq *rq = this_rq_lock();
4869 schedstat_inc(rq->yld_count);
4870 current->sched_class->yield_task(rq);
4873 * Since we are going to call schedule() anyway, there's
4874 * no need to preempt or enable interrupts:
4876 __release(rq->lock);
4877 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4878 do_raw_spin_unlock(&rq->lock);
4879 sched_preempt_enable_no_resched();
4886 #ifndef CONFIG_PREEMPT
4887 int __sched _cond_resched(void)
4889 if (should_resched(0)) {
4890 preempt_schedule_common();
4895 EXPORT_SYMBOL(_cond_resched);
4899 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4900 * call schedule, and on return reacquire the lock.
4902 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4903 * operations here to prevent schedule() from being called twice (once via
4904 * spin_unlock(), once by hand).
4906 int __cond_resched_lock(spinlock_t *lock)
4908 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4911 lockdep_assert_held(lock);
4913 if (spin_needbreak(lock) || resched) {
4916 preempt_schedule_common();
4924 EXPORT_SYMBOL(__cond_resched_lock);
4926 int __sched __cond_resched_softirq(void)
4928 BUG_ON(!in_softirq());
4930 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4932 preempt_schedule_common();
4938 EXPORT_SYMBOL(__cond_resched_softirq);
4941 * yield - yield the current processor to other threads.
4943 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4945 * The scheduler is at all times free to pick the calling task as the most
4946 * eligible task to run, if removing the yield() call from your code breaks
4947 * it, its already broken.
4949 * Typical broken usage is:
4954 * where one assumes that yield() will let 'the other' process run that will
4955 * make event true. If the current task is a SCHED_FIFO task that will never
4956 * happen. Never use yield() as a progress guarantee!!
4958 * If you want to use yield() to wait for something, use wait_event().
4959 * If you want to use yield() to be 'nice' for others, use cond_resched().
4960 * If you still want to use yield(), do not!
4962 void __sched yield(void)
4964 set_current_state(TASK_RUNNING);
4967 EXPORT_SYMBOL(yield);
4970 * yield_to - yield the current processor to another thread in
4971 * your thread group, or accelerate that thread toward the
4972 * processor it's on.
4974 * @preempt: whether task preemption is allowed or not
4976 * It's the caller's job to ensure that the target task struct
4977 * can't go away on us before we can do any checks.
4980 * true (>0) if we indeed boosted the target task.
4981 * false (0) if we failed to boost the target.
4982 * -ESRCH if there's no task to yield to.
4984 int __sched yield_to(struct task_struct *p, bool preempt)
4986 struct task_struct *curr = current;
4987 struct rq *rq, *p_rq;
4988 unsigned long flags;
4991 local_irq_save(flags);
4997 * If we're the only runnable task on the rq and target rq also
4998 * has only one task, there's absolutely no point in yielding.
5000 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
5005 double_rq_lock(rq, p_rq);
5006 if (task_rq(p) != p_rq) {
5007 double_rq_unlock(rq, p_rq);
5011 if (!curr->sched_class->yield_to_task)
5014 if (curr->sched_class != p->sched_class)
5017 if (task_running(p_rq, p) || p->state)
5020 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5022 schedstat_inc(rq->yld_count);
5024 * Make p's CPU reschedule; pick_next_entity takes care of
5027 if (preempt && rq != p_rq)
5032 double_rq_unlock(rq, p_rq);
5034 local_irq_restore(flags);
5041 EXPORT_SYMBOL_GPL(yield_to);
5044 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5045 * that process accounting knows that this is a task in IO wait state.
5047 long __sched io_schedule_timeout(long timeout)
5049 int old_iowait = current->in_iowait;
5053 current->in_iowait = 1;
5054 blk_schedule_flush_plug(current);
5056 delayacct_blkio_start();
5058 atomic_inc(&rq->nr_iowait);
5059 ret = schedule_timeout(timeout);
5060 current->in_iowait = old_iowait;
5061 atomic_dec(&rq->nr_iowait);
5062 delayacct_blkio_end();
5066 EXPORT_SYMBOL(io_schedule_timeout);
5069 * sys_sched_get_priority_max - return maximum RT priority.
5070 * @policy: scheduling class.
5072 * Return: On success, this syscall returns the maximum
5073 * rt_priority that can be used by a given scheduling class.
5074 * On failure, a negative error code is returned.
5076 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5083 ret = MAX_USER_RT_PRIO-1;
5085 case SCHED_DEADLINE:
5096 * sys_sched_get_priority_min - return minimum RT priority.
5097 * @policy: scheduling class.
5099 * Return: On success, this syscall returns the minimum
5100 * rt_priority that can be used by a given scheduling class.
5101 * On failure, a negative error code is returned.
5103 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5112 case SCHED_DEADLINE:
5122 * sys_sched_rr_get_interval - return the default timeslice of a process.
5123 * @pid: pid of the process.
5124 * @interval: userspace pointer to the timeslice value.
5126 * this syscall writes the default timeslice value of a given process
5127 * into the user-space timespec buffer. A value of '0' means infinity.
5129 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5132 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5133 struct timespec __user *, interval)
5135 struct task_struct *p;
5136 unsigned int time_slice;
5147 p = find_process_by_pid(pid);
5151 retval = security_task_getscheduler(p);
5155 rq = task_rq_lock(p, &rf);
5157 if (p->sched_class->get_rr_interval)
5158 time_slice = p->sched_class->get_rr_interval(rq, p);
5159 task_rq_unlock(rq, p, &rf);
5162 jiffies_to_timespec(time_slice, &t);
5163 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5171 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5173 void sched_show_task(struct task_struct *p)
5175 unsigned long free = 0;
5177 unsigned long state = p->state;
5180 state = __ffs(state) + 1;
5181 printk(KERN_INFO "%-15.15s %c", p->comm,
5182 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5183 #if BITS_PER_LONG == 32
5184 if (state == TASK_RUNNING)
5185 printk(KERN_CONT " running ");
5187 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5189 if (state == TASK_RUNNING)
5190 printk(KERN_CONT " running task ");
5192 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5194 #ifdef CONFIG_DEBUG_STACK_USAGE
5195 free = stack_not_used(p);
5200 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5202 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5203 task_pid_nr(p), ppid,
5204 (unsigned long)task_thread_info(p)->flags);
5206 print_worker_info(KERN_INFO, p);
5207 show_stack(p, NULL);
5210 void show_state_filter(unsigned long state_filter)
5212 struct task_struct *g, *p;
5214 #if BITS_PER_LONG == 32
5216 " task PC stack pid father\n");
5219 " task PC stack pid father\n");
5222 for_each_process_thread(g, p) {
5224 * reset the NMI-timeout, listing all files on a slow
5225 * console might take a lot of time:
5226 * Also, reset softlockup watchdogs on all CPUs, because
5227 * another CPU might be blocked waiting for us to process
5230 touch_nmi_watchdog();
5231 touch_all_softlockup_watchdogs();
5232 if (!state_filter || (p->state & state_filter))
5236 #ifdef CONFIG_SCHED_DEBUG
5238 sysrq_sched_debug_show();
5242 * Only show locks if all tasks are dumped:
5245 debug_show_all_locks();
5248 void init_idle_bootup_task(struct task_struct *idle)
5250 idle->sched_class = &idle_sched_class;
5254 * init_idle - set up an idle thread for a given CPU
5255 * @idle: task in question
5256 * @cpu: cpu the idle task belongs to
5258 * NOTE: this function does not set the idle thread's NEED_RESCHED
5259 * flag, to make booting more robust.
5261 void init_idle(struct task_struct *idle, int cpu)
5263 struct rq *rq = cpu_rq(cpu);
5264 unsigned long flags;
5266 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5267 raw_spin_lock(&rq->lock);
5269 __sched_fork(0, idle);
5270 idle->state = TASK_RUNNING;
5271 idle->se.exec_start = sched_clock();
5273 kasan_unpoison_task_stack(idle);
5277 * Its possible that init_idle() gets called multiple times on a task,
5278 * in that case do_set_cpus_allowed() will not do the right thing.
5280 * And since this is boot we can forgo the serialization.
5282 set_cpus_allowed_common(idle, cpumask_of(cpu));
5285 * We're having a chicken and egg problem, even though we are
5286 * holding rq->lock, the cpu isn't yet set to this cpu so the
5287 * lockdep check in task_group() will fail.
5289 * Similar case to sched_fork(). / Alternatively we could
5290 * use task_rq_lock() here and obtain the other rq->lock.
5295 __set_task_cpu(idle, cpu);
5298 rq->curr = rq->idle = idle;
5299 idle->on_rq = TASK_ON_RQ_QUEUED;
5303 raw_spin_unlock(&rq->lock);
5304 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5306 /* Set the preempt count _outside_ the spinlocks! */
5307 init_idle_preempt_count(idle, cpu);
5310 * The idle tasks have their own, simple scheduling class:
5312 idle->sched_class = &idle_sched_class;
5313 ftrace_graph_init_idle_task(idle, cpu);
5314 vtime_init_idle(idle, cpu);
5316 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5320 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5321 const struct cpumask *trial)
5323 int ret = 1, trial_cpus;
5324 struct dl_bw *cur_dl_b;
5325 unsigned long flags;
5327 if (!cpumask_weight(cur))
5330 rcu_read_lock_sched();
5331 cur_dl_b = dl_bw_of(cpumask_any(cur));
5332 trial_cpus = cpumask_weight(trial);
5334 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5335 if (cur_dl_b->bw != -1 &&
5336 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5338 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5339 rcu_read_unlock_sched();
5344 int task_can_attach(struct task_struct *p,
5345 const struct cpumask *cs_cpus_allowed)
5350 * Kthreads which disallow setaffinity shouldn't be moved
5351 * to a new cpuset; we don't want to change their cpu
5352 * affinity and isolating such threads by their set of
5353 * allowed nodes is unnecessary. Thus, cpusets are not
5354 * applicable for such threads. This prevents checking for
5355 * success of set_cpus_allowed_ptr() on all attached tasks
5356 * before cpus_allowed may be changed.
5358 if (p->flags & PF_NO_SETAFFINITY) {
5364 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5366 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5371 unsigned long flags;
5373 rcu_read_lock_sched();
5374 dl_b = dl_bw_of(dest_cpu);
5375 raw_spin_lock_irqsave(&dl_b->lock, flags);
5376 cpus = dl_bw_cpus(dest_cpu);
5377 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5382 * We reserve space for this task in the destination
5383 * root_domain, as we can't fail after this point.
5384 * We will free resources in the source root_domain
5385 * later on (see set_cpus_allowed_dl()).
5387 __dl_add(dl_b, p->dl.dl_bw);
5389 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5390 rcu_read_unlock_sched();
5400 static bool sched_smp_initialized __read_mostly;
5402 #ifdef CONFIG_NUMA_BALANCING
5403 /* Migrate current task p to target_cpu */
5404 int migrate_task_to(struct task_struct *p, int target_cpu)
5406 struct migration_arg arg = { p, target_cpu };
5407 int curr_cpu = task_cpu(p);
5409 if (curr_cpu == target_cpu)
5412 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5415 /* TODO: This is not properly updating schedstats */
5417 trace_sched_move_numa(p, curr_cpu, target_cpu);
5418 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5422 * Requeue a task on a given node and accurately track the number of NUMA
5423 * tasks on the runqueues
5425 void sched_setnuma(struct task_struct *p, int nid)
5427 bool queued, running;
5431 rq = task_rq_lock(p, &rf);
5432 queued = task_on_rq_queued(p);
5433 running = task_current(rq, p);
5436 dequeue_task(rq, p, DEQUEUE_SAVE);
5438 put_prev_task(rq, p);
5440 p->numa_preferred_nid = nid;
5443 p->sched_class->set_curr_task(rq);
5445 enqueue_task(rq, p, ENQUEUE_RESTORE);
5446 task_rq_unlock(rq, p, &rf);
5448 #endif /* CONFIG_NUMA_BALANCING */
5450 #ifdef CONFIG_HOTPLUG_CPU
5452 * Ensures that the idle task is using init_mm right before its cpu goes
5455 void idle_task_exit(void)
5457 struct mm_struct *mm = current->active_mm;
5459 BUG_ON(cpu_online(smp_processor_id()));
5461 if (mm != &init_mm) {
5462 switch_mm_irqs_off(mm, &init_mm, current);
5463 finish_arch_post_lock_switch();
5469 * Since this CPU is going 'away' for a while, fold any nr_active delta
5470 * we might have. Assumes we're called after migrate_tasks() so that the
5471 * nr_active count is stable. We need to take the teardown thread which
5472 * is calling this into account, so we hand in adjust = 1 to the load
5475 * Also see the comment "Global load-average calculations".
5477 static void calc_load_migrate(struct rq *rq)
5479 long delta = calc_load_fold_active(rq, 1);
5481 atomic_long_add(delta, &calc_load_tasks);
5484 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5488 static const struct sched_class fake_sched_class = {
5489 .put_prev_task = put_prev_task_fake,
5492 static struct task_struct fake_task = {
5494 * Avoid pull_{rt,dl}_task()
5496 .prio = MAX_PRIO + 1,
5497 .sched_class = &fake_sched_class,
5501 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5502 * try_to_wake_up()->select_task_rq().
5504 * Called with rq->lock held even though we'er in stop_machine() and
5505 * there's no concurrency possible, we hold the required locks anyway
5506 * because of lock validation efforts.
5508 static void migrate_tasks(struct rq *dead_rq)
5510 struct rq *rq = dead_rq;
5511 struct task_struct *next, *stop = rq->stop;
5512 struct pin_cookie cookie;
5516 * Fudge the rq selection such that the below task selection loop
5517 * doesn't get stuck on the currently eligible stop task.
5519 * We're currently inside stop_machine() and the rq is either stuck
5520 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5521 * either way we should never end up calling schedule() until we're
5527 * put_prev_task() and pick_next_task() sched
5528 * class method both need to have an up-to-date
5529 * value of rq->clock[_task]
5531 update_rq_clock(rq);
5535 * There's this thread running, bail when that's the only
5538 if (rq->nr_running == 1)
5542 * pick_next_task assumes pinned rq->lock.
5544 cookie = lockdep_pin_lock(&rq->lock);
5545 next = pick_next_task(rq, &fake_task, cookie);
5547 next->sched_class->put_prev_task(rq, next);
5550 * Rules for changing task_struct::cpus_allowed are holding
5551 * both pi_lock and rq->lock, such that holding either
5552 * stabilizes the mask.
5554 * Drop rq->lock is not quite as disastrous as it usually is
5555 * because !cpu_active at this point, which means load-balance
5556 * will not interfere. Also, stop-machine.
5558 lockdep_unpin_lock(&rq->lock, cookie);
5559 raw_spin_unlock(&rq->lock);
5560 raw_spin_lock(&next->pi_lock);
5561 raw_spin_lock(&rq->lock);
5564 * Since we're inside stop-machine, _nothing_ should have
5565 * changed the task, WARN if weird stuff happened, because in
5566 * that case the above rq->lock drop is a fail too.
5568 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5569 raw_spin_unlock(&next->pi_lock);
5573 /* Find suitable destination for @next, with force if needed. */
5574 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5576 rq = __migrate_task(rq, next, dest_cpu);
5577 if (rq != dead_rq) {
5578 raw_spin_unlock(&rq->lock);
5580 raw_spin_lock(&rq->lock);
5582 raw_spin_unlock(&next->pi_lock);
5587 #endif /* CONFIG_HOTPLUG_CPU */
5589 static void set_rq_online(struct rq *rq)
5592 const struct sched_class *class;
5594 cpumask_set_cpu(rq->cpu, rq->rd->online);
5597 for_each_class(class) {
5598 if (class->rq_online)
5599 class->rq_online(rq);
5604 static void set_rq_offline(struct rq *rq)
5607 const struct sched_class *class;
5609 for_each_class(class) {
5610 if (class->rq_offline)
5611 class->rq_offline(rq);
5614 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5619 static void set_cpu_rq_start_time(unsigned int cpu)
5621 struct rq *rq = cpu_rq(cpu);
5623 rq->age_stamp = sched_clock_cpu(cpu);
5626 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5628 #ifdef CONFIG_SCHED_DEBUG
5630 static __read_mostly int sched_debug_enabled;
5632 static int __init sched_debug_setup(char *str)
5634 sched_debug_enabled = 1;
5638 early_param("sched_debug", sched_debug_setup);
5640 static inline bool sched_debug(void)
5642 return sched_debug_enabled;
5645 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5646 struct cpumask *groupmask)
5648 struct sched_group *group = sd->groups;
5650 cpumask_clear(groupmask);
5652 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5654 if (!(sd->flags & SD_LOAD_BALANCE)) {
5655 printk("does not load-balance\n");
5657 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5662 printk(KERN_CONT "span %*pbl level %s\n",
5663 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5665 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5666 printk(KERN_ERR "ERROR: domain->span does not contain "
5669 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5670 printk(KERN_ERR "ERROR: domain->groups does not contain"
5674 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5678 printk(KERN_ERR "ERROR: group is NULL\n");
5682 if (!cpumask_weight(sched_group_cpus(group))) {
5683 printk(KERN_CONT "\n");
5684 printk(KERN_ERR "ERROR: empty group\n");
5688 if (!(sd->flags & SD_OVERLAP) &&
5689 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5690 printk(KERN_CONT "\n");
5691 printk(KERN_ERR "ERROR: repeated CPUs\n");
5695 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5697 printk(KERN_CONT " %*pbl",
5698 cpumask_pr_args(sched_group_cpus(group)));
5699 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5700 printk(KERN_CONT " (cpu_capacity = %d)",
5701 group->sgc->capacity);
5704 group = group->next;
5705 } while (group != sd->groups);
5706 printk(KERN_CONT "\n");
5708 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5709 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5712 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5713 printk(KERN_ERR "ERROR: parent span is not a superset "
5714 "of domain->span\n");
5718 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5722 if (!sched_debug_enabled)
5726 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5730 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5733 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5741 #else /* !CONFIG_SCHED_DEBUG */
5743 # define sched_debug_enabled 0
5744 # define sched_domain_debug(sd, cpu) do { } while (0)
5745 static inline bool sched_debug(void)
5749 #endif /* CONFIG_SCHED_DEBUG */
5751 static int sd_degenerate(struct sched_domain *sd)
5753 if (cpumask_weight(sched_domain_span(sd)) == 1)
5756 /* Following flags need at least 2 groups */
5757 if (sd->flags & (SD_LOAD_BALANCE |
5758 SD_BALANCE_NEWIDLE |
5761 SD_SHARE_CPUCAPACITY |
5762 SD_ASYM_CPUCAPACITY |
5763 SD_SHARE_PKG_RESOURCES |
5764 SD_SHARE_POWERDOMAIN)) {
5765 if (sd->groups != sd->groups->next)
5769 /* Following flags don't use groups */
5770 if (sd->flags & (SD_WAKE_AFFINE))
5777 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5779 unsigned long cflags = sd->flags, pflags = parent->flags;
5781 if (sd_degenerate(parent))
5784 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5787 /* Flags needing groups don't count if only 1 group in parent */
5788 if (parent->groups == parent->groups->next) {
5789 pflags &= ~(SD_LOAD_BALANCE |
5790 SD_BALANCE_NEWIDLE |
5793 SD_ASYM_CPUCAPACITY |
5794 SD_SHARE_CPUCAPACITY |
5795 SD_SHARE_PKG_RESOURCES |
5797 SD_SHARE_POWERDOMAIN);
5798 if (nr_node_ids == 1)
5799 pflags &= ~SD_SERIALIZE;
5801 if (~cflags & pflags)
5807 static void free_rootdomain(struct rcu_head *rcu)
5809 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5811 cpupri_cleanup(&rd->cpupri);
5812 cpudl_cleanup(&rd->cpudl);
5813 free_cpumask_var(rd->dlo_mask);
5814 free_cpumask_var(rd->rto_mask);
5815 free_cpumask_var(rd->online);
5816 free_cpumask_var(rd->span);
5820 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5822 struct root_domain *old_rd = NULL;
5823 unsigned long flags;
5825 raw_spin_lock_irqsave(&rq->lock, flags);
5830 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5833 cpumask_clear_cpu(rq->cpu, old_rd->span);
5836 * If we dont want to free the old_rd yet then
5837 * set old_rd to NULL to skip the freeing later
5840 if (!atomic_dec_and_test(&old_rd->refcount))
5844 atomic_inc(&rd->refcount);
5847 cpumask_set_cpu(rq->cpu, rd->span);
5848 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5851 raw_spin_unlock_irqrestore(&rq->lock, flags);
5854 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5857 static int init_rootdomain(struct root_domain *rd)
5859 memset(rd, 0, sizeof(*rd));
5861 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5863 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5865 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5867 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5870 init_dl_bw(&rd->dl_bw);
5871 if (cpudl_init(&rd->cpudl) != 0)
5874 if (cpupri_init(&rd->cpupri) != 0)
5879 free_cpumask_var(rd->rto_mask);
5881 free_cpumask_var(rd->dlo_mask);
5883 free_cpumask_var(rd->online);
5885 free_cpumask_var(rd->span);
5891 * By default the system creates a single root-domain with all cpus as
5892 * members (mimicking the global state we have today).
5894 struct root_domain def_root_domain;
5896 static void init_defrootdomain(void)
5898 init_rootdomain(&def_root_domain);
5900 atomic_set(&def_root_domain.refcount, 1);
5903 static struct root_domain *alloc_rootdomain(void)
5905 struct root_domain *rd;
5907 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5911 if (init_rootdomain(rd) != 0) {
5919 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5921 struct sched_group *tmp, *first;
5930 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5935 } while (sg != first);
5938 static void destroy_sched_domain(struct sched_domain *sd)
5941 * If its an overlapping domain it has private groups, iterate and
5944 if (sd->flags & SD_OVERLAP) {
5945 free_sched_groups(sd->groups, 1);
5946 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5947 kfree(sd->groups->sgc);
5950 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
5955 static void destroy_sched_domains_rcu(struct rcu_head *rcu)
5957 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5960 struct sched_domain *parent = sd->parent;
5961 destroy_sched_domain(sd);
5966 static void destroy_sched_domains(struct sched_domain *sd)
5969 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
5973 * Keep a special pointer to the highest sched_domain that has
5974 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5975 * allows us to avoid some pointer chasing select_idle_sibling().
5977 * Also keep a unique ID per domain (we use the first cpu number in
5978 * the cpumask of the domain), this allows us to quickly tell if
5979 * two cpus are in the same cache domain, see cpus_share_cache().
5981 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5982 DEFINE_PER_CPU(int, sd_llc_size);
5983 DEFINE_PER_CPU(int, sd_llc_id);
5984 DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
5985 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5986 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5988 static void update_top_cache_domain(int cpu)
5990 struct sched_domain_shared *sds = NULL;
5991 struct sched_domain *sd;
5995 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5997 id = cpumask_first(sched_domain_span(sd));
5998 size = cpumask_weight(sched_domain_span(sd));
6002 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6003 per_cpu(sd_llc_size, cpu) = size;
6004 per_cpu(sd_llc_id, cpu) = id;
6005 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
6007 sd = lowest_flag_domain(cpu, SD_NUMA);
6008 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6010 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6011 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6015 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6016 * hold the hotplug lock.
6019 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6021 struct rq *rq = cpu_rq(cpu);
6022 struct sched_domain *tmp;
6024 /* Remove the sched domains which do not contribute to scheduling. */
6025 for (tmp = sd; tmp; ) {
6026 struct sched_domain *parent = tmp->parent;
6030 if (sd_parent_degenerate(tmp, parent)) {
6031 tmp->parent = parent->parent;
6033 parent->parent->child = tmp;
6035 * Transfer SD_PREFER_SIBLING down in case of a
6036 * degenerate parent; the spans match for this
6037 * so the property transfers.
6039 if (parent->flags & SD_PREFER_SIBLING)
6040 tmp->flags |= SD_PREFER_SIBLING;
6041 destroy_sched_domain(parent);
6046 if (sd && sd_degenerate(sd)) {
6049 destroy_sched_domain(tmp);
6054 sched_domain_debug(sd, cpu);
6056 rq_attach_root(rq, rd);
6058 rcu_assign_pointer(rq->sd, sd);
6059 destroy_sched_domains(tmp);
6061 update_top_cache_domain(cpu);
6064 /* Setup the mask of cpus configured for isolated domains */
6065 static int __init isolated_cpu_setup(char *str)
6069 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6070 ret = cpulist_parse(str, cpu_isolated_map);
6072 pr_err("sched: Error, all isolcpus= values must be between 0 and %d\n", nr_cpu_ids);
6077 __setup("isolcpus=", isolated_cpu_setup);
6080 struct sched_domain ** __percpu sd;
6081 struct root_domain *rd;
6092 * Build an iteration mask that can exclude certain CPUs from the upwards
6095 * Asymmetric node setups can result in situations where the domain tree is of
6096 * unequal depth, make sure to skip domains that already cover the entire
6099 * In that case build_sched_domains() will have terminated the iteration early
6100 * and our sibling sd spans will be empty. Domains should always include the
6101 * cpu they're built on, so check that.
6104 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6106 const struct cpumask *span = sched_domain_span(sd);
6107 struct sd_data *sdd = sd->private;
6108 struct sched_domain *sibling;
6111 for_each_cpu(i, span) {
6112 sibling = *per_cpu_ptr(sdd->sd, i);
6113 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6116 cpumask_set_cpu(i, sched_group_mask(sg));
6121 * Return the canonical balance cpu for this group, this is the first cpu
6122 * of this group that's also in the iteration mask.
6124 int group_balance_cpu(struct sched_group *sg)
6126 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6130 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6132 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6133 const struct cpumask *span = sched_domain_span(sd);
6134 struct cpumask *covered = sched_domains_tmpmask;
6135 struct sd_data *sdd = sd->private;
6136 struct sched_domain *sibling;
6139 cpumask_clear(covered);
6141 for_each_cpu(i, span) {
6142 struct cpumask *sg_span;
6144 if (cpumask_test_cpu(i, covered))
6147 sibling = *per_cpu_ptr(sdd->sd, i);
6149 /* See the comment near build_group_mask(). */
6150 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6153 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6154 GFP_KERNEL, cpu_to_node(cpu));
6159 sg_span = sched_group_cpus(sg);
6161 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6163 cpumask_set_cpu(i, sg_span);
6165 cpumask_or(covered, covered, sg_span);
6167 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6168 if (atomic_inc_return(&sg->sgc->ref) == 1)
6169 build_group_mask(sd, sg);
6172 * Initialize sgc->capacity such that even if we mess up the
6173 * domains and no possible iteration will get us here, we won't
6176 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6179 * Make sure the first group of this domain contains the
6180 * canonical balance cpu. Otherwise the sched_domain iteration
6181 * breaks. See update_sg_lb_stats().
6183 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6184 group_balance_cpu(sg) == cpu)
6194 sd->groups = groups;
6199 free_sched_groups(first, 0);
6204 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6206 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6207 struct sched_domain *child = sd->child;
6210 cpu = cpumask_first(sched_domain_span(child));
6213 *sg = *per_cpu_ptr(sdd->sg, cpu);
6214 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6215 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6222 * build_sched_groups will build a circular linked list of the groups
6223 * covered by the given span, and will set each group's ->cpumask correctly,
6224 * and ->cpu_capacity to 0.
6226 * Assumes the sched_domain tree is fully constructed
6229 build_sched_groups(struct sched_domain *sd, int cpu)
6231 struct sched_group *first = NULL, *last = NULL;
6232 struct sd_data *sdd = sd->private;
6233 const struct cpumask *span = sched_domain_span(sd);
6234 struct cpumask *covered;
6237 get_group(cpu, sdd, &sd->groups);
6238 atomic_inc(&sd->groups->ref);
6240 if (cpu != cpumask_first(span))
6243 lockdep_assert_held(&sched_domains_mutex);
6244 covered = sched_domains_tmpmask;
6246 cpumask_clear(covered);
6248 for_each_cpu(i, span) {
6249 struct sched_group *sg;
6252 if (cpumask_test_cpu(i, covered))
6255 group = get_group(i, sdd, &sg);
6256 cpumask_setall(sched_group_mask(sg));
6258 for_each_cpu(j, span) {
6259 if (get_group(j, sdd, NULL) != group)
6262 cpumask_set_cpu(j, covered);
6263 cpumask_set_cpu(j, sched_group_cpus(sg));
6278 * Initialize sched groups cpu_capacity.
6280 * cpu_capacity indicates the capacity of sched group, which is used while
6281 * distributing the load between different sched groups in a sched domain.
6282 * Typically cpu_capacity for all the groups in a sched domain will be same
6283 * unless there are asymmetries in the topology. If there are asymmetries,
6284 * group having more cpu_capacity will pickup more load compared to the
6285 * group having less cpu_capacity.
6287 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6289 struct sched_group *sg = sd->groups;
6294 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6296 } while (sg != sd->groups);
6298 if (cpu != group_balance_cpu(sg))
6301 update_group_capacity(sd, cpu);
6305 * Initializers for schedule domains
6306 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6309 static int default_relax_domain_level = -1;
6310 int sched_domain_level_max;
6312 static int __init setup_relax_domain_level(char *str)
6314 if (kstrtoint(str, 0, &default_relax_domain_level))
6315 pr_warn("Unable to set relax_domain_level\n");
6319 __setup("relax_domain_level=", setup_relax_domain_level);
6321 static void set_domain_attribute(struct sched_domain *sd,
6322 struct sched_domain_attr *attr)
6326 if (!attr || attr->relax_domain_level < 0) {
6327 if (default_relax_domain_level < 0)
6330 request = default_relax_domain_level;
6332 request = attr->relax_domain_level;
6333 if (request < sd->level) {
6334 /* turn off idle balance on this domain */
6335 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6337 /* turn on idle balance on this domain */
6338 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6342 static void __sdt_free(const struct cpumask *cpu_map);
6343 static int __sdt_alloc(const struct cpumask *cpu_map);
6345 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6346 const struct cpumask *cpu_map)
6350 if (!atomic_read(&d->rd->refcount))
6351 free_rootdomain(&d->rd->rcu); /* fall through */
6353 free_percpu(d->sd); /* fall through */
6355 __sdt_free(cpu_map); /* fall through */
6361 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6362 const struct cpumask *cpu_map)
6364 memset(d, 0, sizeof(*d));
6366 if (__sdt_alloc(cpu_map))
6367 return sa_sd_storage;
6368 d->sd = alloc_percpu(struct sched_domain *);
6370 return sa_sd_storage;
6371 d->rd = alloc_rootdomain();
6374 return sa_rootdomain;
6378 * NULL the sd_data elements we've used to build the sched_domain and
6379 * sched_group structure so that the subsequent __free_domain_allocs()
6380 * will not free the data we're using.
6382 static void claim_allocations(int cpu, struct sched_domain *sd)
6384 struct sd_data *sdd = sd->private;
6386 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6387 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6389 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
6390 *per_cpu_ptr(sdd->sds, cpu) = NULL;
6392 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6393 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6395 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6396 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6400 static int sched_domains_numa_levels;
6401 enum numa_topology_type sched_numa_topology_type;
6402 static int *sched_domains_numa_distance;
6403 int sched_max_numa_distance;
6404 static struct cpumask ***sched_domains_numa_masks;
6405 static int sched_domains_curr_level;
6409 * SD_flags allowed in topology descriptions.
6411 * These flags are purely descriptive of the topology and do not prescribe
6412 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6415 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6416 * SD_SHARE_PKG_RESOURCES - describes shared caches
6417 * SD_NUMA - describes NUMA topologies
6418 * SD_SHARE_POWERDOMAIN - describes shared power domain
6419 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6421 * Odd one out, which beside describing the topology has a quirk also
6422 * prescribes the desired behaviour that goes along with it:
6424 * SD_ASYM_PACKING - describes SMT quirks
6426 #define TOPOLOGY_SD_FLAGS \
6427 (SD_SHARE_CPUCAPACITY | \
6428 SD_SHARE_PKG_RESOURCES | \
6431 SD_ASYM_CPUCAPACITY | \
6432 SD_SHARE_POWERDOMAIN)
6434 static struct sched_domain *
6435 sd_init(struct sched_domain_topology_level *tl,
6436 const struct cpumask *cpu_map,
6437 struct sched_domain *child, int cpu)
6439 struct sd_data *sdd = &tl->data;
6440 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6441 int sd_id, sd_weight, sd_flags = 0;
6445 * Ugly hack to pass state to sd_numa_mask()...
6447 sched_domains_curr_level = tl->numa_level;
6450 sd_weight = cpumask_weight(tl->mask(cpu));
6453 sd_flags = (*tl->sd_flags)();
6454 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6455 "wrong sd_flags in topology description\n"))
6456 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6458 *sd = (struct sched_domain){
6459 .min_interval = sd_weight,
6460 .max_interval = 2*sd_weight,
6462 .imbalance_pct = 125,
6464 .cache_nice_tries = 0,
6471 .flags = 1*SD_LOAD_BALANCE
6472 | 1*SD_BALANCE_NEWIDLE
6477 | 0*SD_SHARE_CPUCAPACITY
6478 | 0*SD_SHARE_PKG_RESOURCES
6480 | 0*SD_PREFER_SIBLING
6485 .last_balance = jiffies,
6486 .balance_interval = sd_weight,
6488 .max_newidle_lb_cost = 0,
6489 .next_decay_max_lb_cost = jiffies,
6491 #ifdef CONFIG_SCHED_DEBUG
6496 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6497 sd_id = cpumask_first(sched_domain_span(sd));
6500 * Convert topological properties into behaviour.
6503 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6504 struct sched_domain *t = sd;
6506 for_each_lower_domain(t)
6507 t->flags |= SD_BALANCE_WAKE;
6510 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6511 sd->flags |= SD_PREFER_SIBLING;
6512 sd->imbalance_pct = 110;
6513 sd->smt_gain = 1178; /* ~15% */
6515 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6516 sd->imbalance_pct = 117;
6517 sd->cache_nice_tries = 1;
6521 } else if (sd->flags & SD_NUMA) {
6522 sd->cache_nice_tries = 2;
6526 sd->flags |= SD_SERIALIZE;
6527 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6528 sd->flags &= ~(SD_BALANCE_EXEC |
6535 sd->flags |= SD_PREFER_SIBLING;
6536 sd->cache_nice_tries = 1;
6542 * For all levels sharing cache; connect a sched_domain_shared
6545 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6546 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
6547 atomic_inc(&sd->shared->ref);
6548 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
6557 * Topology list, bottom-up.
6559 static struct sched_domain_topology_level default_topology[] = {
6560 #ifdef CONFIG_SCHED_SMT
6561 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6563 #ifdef CONFIG_SCHED_MC
6564 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6566 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6570 static struct sched_domain_topology_level *sched_domain_topology =
6573 #define for_each_sd_topology(tl) \
6574 for (tl = sched_domain_topology; tl->mask; tl++)
6576 void set_sched_topology(struct sched_domain_topology_level *tl)
6578 if (WARN_ON_ONCE(sched_smp_initialized))
6581 sched_domain_topology = tl;
6586 static const struct cpumask *sd_numa_mask(int cpu)
6588 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6591 static void sched_numa_warn(const char *str)
6593 static int done = false;
6601 printk(KERN_WARNING "ERROR: %s\n\n", str);
6603 for (i = 0; i < nr_node_ids; i++) {
6604 printk(KERN_WARNING " ");
6605 for (j = 0; j < nr_node_ids; j++)
6606 printk(KERN_CONT "%02d ", node_distance(i,j));
6607 printk(KERN_CONT "\n");
6609 printk(KERN_WARNING "\n");
6612 bool find_numa_distance(int distance)
6616 if (distance == node_distance(0, 0))
6619 for (i = 0; i < sched_domains_numa_levels; i++) {
6620 if (sched_domains_numa_distance[i] == distance)
6628 * A system can have three types of NUMA topology:
6629 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6630 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6631 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6633 * The difference between a glueless mesh topology and a backplane
6634 * topology lies in whether communication between not directly
6635 * connected nodes goes through intermediary nodes (where programs
6636 * could run), or through backplane controllers. This affects
6637 * placement of programs.
6639 * The type of topology can be discerned with the following tests:
6640 * - If the maximum distance between any nodes is 1 hop, the system
6641 * is directly connected.
6642 * - If for two nodes A and B, located N > 1 hops away from each other,
6643 * there is an intermediary node C, which is < N hops away from both
6644 * nodes A and B, the system is a glueless mesh.
6646 static void init_numa_topology_type(void)
6650 n = sched_max_numa_distance;
6652 if (sched_domains_numa_levels <= 1) {
6653 sched_numa_topology_type = NUMA_DIRECT;
6657 for_each_online_node(a) {
6658 for_each_online_node(b) {
6659 /* Find two nodes furthest removed from each other. */
6660 if (node_distance(a, b) < n)
6663 /* Is there an intermediary node between a and b? */
6664 for_each_online_node(c) {
6665 if (node_distance(a, c) < n &&
6666 node_distance(b, c) < n) {
6667 sched_numa_topology_type =
6673 sched_numa_topology_type = NUMA_BACKPLANE;
6679 static void sched_init_numa(void)
6681 int next_distance, curr_distance = node_distance(0, 0);
6682 struct sched_domain_topology_level *tl;
6686 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6687 if (!sched_domains_numa_distance)
6691 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6692 * unique distances in the node_distance() table.
6694 * Assumes node_distance(0,j) includes all distances in
6695 * node_distance(i,j) in order to avoid cubic time.
6697 next_distance = curr_distance;
6698 for (i = 0; i < nr_node_ids; i++) {
6699 for (j = 0; j < nr_node_ids; j++) {
6700 for (k = 0; k < nr_node_ids; k++) {
6701 int distance = node_distance(i, k);
6703 if (distance > curr_distance &&
6704 (distance < next_distance ||
6705 next_distance == curr_distance))
6706 next_distance = distance;
6709 * While not a strong assumption it would be nice to know
6710 * about cases where if node A is connected to B, B is not
6711 * equally connected to A.
6713 if (sched_debug() && node_distance(k, i) != distance)
6714 sched_numa_warn("Node-distance not symmetric");
6716 if (sched_debug() && i && !find_numa_distance(distance))
6717 sched_numa_warn("Node-0 not representative");
6719 if (next_distance != curr_distance) {
6720 sched_domains_numa_distance[level++] = next_distance;
6721 sched_domains_numa_levels = level;
6722 curr_distance = next_distance;
6727 * In case of sched_debug() we verify the above assumption.
6737 * 'level' contains the number of unique distances, excluding the
6738 * identity distance node_distance(i,i).
6740 * The sched_domains_numa_distance[] array includes the actual distance
6745 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6746 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6747 * the array will contain less then 'level' members. This could be
6748 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6749 * in other functions.
6751 * We reset it to 'level' at the end of this function.
6753 sched_domains_numa_levels = 0;
6755 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6756 if (!sched_domains_numa_masks)
6760 * Now for each level, construct a mask per node which contains all
6761 * cpus of nodes that are that many hops away from us.
6763 for (i = 0; i < level; i++) {
6764 sched_domains_numa_masks[i] =
6765 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6766 if (!sched_domains_numa_masks[i])
6769 for (j = 0; j < nr_node_ids; j++) {
6770 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6774 sched_domains_numa_masks[i][j] = mask;
6777 if (node_distance(j, k) > sched_domains_numa_distance[i])
6780 cpumask_or(mask, mask, cpumask_of_node(k));
6785 /* Compute default topology size */
6786 for (i = 0; sched_domain_topology[i].mask; i++);
6788 tl = kzalloc((i + level + 1) *
6789 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6794 * Copy the default topology bits..
6796 for (i = 0; sched_domain_topology[i].mask; i++)
6797 tl[i] = sched_domain_topology[i];
6800 * .. and append 'j' levels of NUMA goodness.
6802 for (j = 0; j < level; i++, j++) {
6803 tl[i] = (struct sched_domain_topology_level){
6804 .mask = sd_numa_mask,
6805 .sd_flags = cpu_numa_flags,
6806 .flags = SDTL_OVERLAP,
6812 sched_domain_topology = tl;
6814 sched_domains_numa_levels = level;
6815 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6817 init_numa_topology_type();
6820 static void sched_domains_numa_masks_set(unsigned int cpu)
6822 int node = cpu_to_node(cpu);
6825 for (i = 0; i < sched_domains_numa_levels; i++) {
6826 for (j = 0; j < nr_node_ids; j++) {
6827 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6828 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6833 static void sched_domains_numa_masks_clear(unsigned int cpu)
6837 for (i = 0; i < sched_domains_numa_levels; i++) {
6838 for (j = 0; j < nr_node_ids; j++)
6839 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6844 static inline void sched_init_numa(void) { }
6845 static void sched_domains_numa_masks_set(unsigned int cpu) { }
6846 static void sched_domains_numa_masks_clear(unsigned int cpu) { }
6847 #endif /* CONFIG_NUMA */
6849 static int __sdt_alloc(const struct cpumask *cpu_map)
6851 struct sched_domain_topology_level *tl;
6854 for_each_sd_topology(tl) {
6855 struct sd_data *sdd = &tl->data;
6857 sdd->sd = alloc_percpu(struct sched_domain *);
6861 sdd->sds = alloc_percpu(struct sched_domain_shared *);
6865 sdd->sg = alloc_percpu(struct sched_group *);
6869 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6873 for_each_cpu(j, cpu_map) {
6874 struct sched_domain *sd;
6875 struct sched_domain_shared *sds;
6876 struct sched_group *sg;
6877 struct sched_group_capacity *sgc;
6879 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6880 GFP_KERNEL, cpu_to_node(j));
6884 *per_cpu_ptr(sdd->sd, j) = sd;
6886 sds = kzalloc_node(sizeof(struct sched_domain_shared),
6887 GFP_KERNEL, cpu_to_node(j));
6891 *per_cpu_ptr(sdd->sds, j) = sds;
6893 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6894 GFP_KERNEL, cpu_to_node(j));
6900 *per_cpu_ptr(sdd->sg, j) = sg;
6902 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6903 GFP_KERNEL, cpu_to_node(j));
6907 *per_cpu_ptr(sdd->sgc, j) = sgc;
6914 static void __sdt_free(const struct cpumask *cpu_map)
6916 struct sched_domain_topology_level *tl;
6919 for_each_sd_topology(tl) {
6920 struct sd_data *sdd = &tl->data;
6922 for_each_cpu(j, cpu_map) {
6923 struct sched_domain *sd;
6926 sd = *per_cpu_ptr(sdd->sd, j);
6927 if (sd && (sd->flags & SD_OVERLAP))
6928 free_sched_groups(sd->groups, 0);
6929 kfree(*per_cpu_ptr(sdd->sd, j));
6933 kfree(*per_cpu_ptr(sdd->sds, j));
6935 kfree(*per_cpu_ptr(sdd->sg, j));
6937 kfree(*per_cpu_ptr(sdd->sgc, j));
6939 free_percpu(sdd->sd);
6941 free_percpu(sdd->sds);
6943 free_percpu(sdd->sg);
6945 free_percpu(sdd->sgc);
6950 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6951 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6952 struct sched_domain *child, int cpu)
6954 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
6957 sd->level = child->level + 1;
6958 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6961 if (!cpumask_subset(sched_domain_span(child),
6962 sched_domain_span(sd))) {
6963 pr_err("BUG: arch topology borken\n");
6964 #ifdef CONFIG_SCHED_DEBUG
6965 pr_err(" the %s domain not a subset of the %s domain\n",
6966 child->name, sd->name);
6968 /* Fixup, ensure @sd has at least @child cpus. */
6969 cpumask_or(sched_domain_span(sd),
6970 sched_domain_span(sd),
6971 sched_domain_span(child));
6975 set_domain_attribute(sd, attr);
6981 * Build sched domains for a given set of cpus and attach the sched domains
6982 * to the individual cpus
6984 static int build_sched_domains(const struct cpumask *cpu_map,
6985 struct sched_domain_attr *attr)
6987 enum s_alloc alloc_state;
6988 struct sched_domain *sd;
6990 struct rq *rq = NULL;
6991 int i, ret = -ENOMEM;
6993 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6994 if (alloc_state != sa_rootdomain)
6997 /* Set up domains for cpus specified by the cpu_map. */
6998 for_each_cpu(i, cpu_map) {
6999 struct sched_domain_topology_level *tl;
7002 for_each_sd_topology(tl) {
7003 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7004 if (tl == sched_domain_topology)
7005 *per_cpu_ptr(d.sd, i) = sd;
7006 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7007 sd->flags |= SD_OVERLAP;
7008 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7013 /* Build the groups for the domains */
7014 for_each_cpu(i, cpu_map) {
7015 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7016 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7017 if (sd->flags & SD_OVERLAP) {
7018 if (build_overlap_sched_groups(sd, i))
7021 if (build_sched_groups(sd, i))
7027 /* Calculate CPU capacity for physical packages and nodes */
7028 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7029 if (!cpumask_test_cpu(i, cpu_map))
7032 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7033 claim_allocations(i, sd);
7034 init_sched_groups_capacity(i, sd);
7038 /* Attach the domains */
7040 for_each_cpu(i, cpu_map) {
7042 sd = *per_cpu_ptr(d.sd, i);
7044 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
7045 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
7046 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
7048 cpu_attach_domain(sd, d.rd, i);
7052 if (rq && sched_debug_enabled) {
7053 pr_info("span: %*pbl (max cpu_capacity = %lu)\n",
7054 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
7059 __free_domain_allocs(&d, alloc_state, cpu_map);
7063 static cpumask_var_t *doms_cur; /* current sched domains */
7064 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7065 static struct sched_domain_attr *dattr_cur;
7066 /* attribues of custom domains in 'doms_cur' */
7069 * Special case: If a kmalloc of a doms_cur partition (array of
7070 * cpumask) fails, then fallback to a single sched domain,
7071 * as determined by the single cpumask fallback_doms.
7073 static cpumask_var_t fallback_doms;
7076 * arch_update_cpu_topology lets virtualized architectures update the
7077 * cpu core maps. It is supposed to return 1 if the topology changed
7078 * or 0 if it stayed the same.
7080 int __weak arch_update_cpu_topology(void)
7085 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7088 cpumask_var_t *doms;
7090 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7093 for (i = 0; i < ndoms; i++) {
7094 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7095 free_sched_domains(doms, i);
7102 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7105 for (i = 0; i < ndoms; i++)
7106 free_cpumask_var(doms[i]);
7111 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7112 * For now this just excludes isolated cpus, but could be used to
7113 * exclude other special cases in the future.
7115 static int init_sched_domains(const struct cpumask *cpu_map)
7119 arch_update_cpu_topology();
7121 doms_cur = alloc_sched_domains(ndoms_cur);
7123 doms_cur = &fallback_doms;
7124 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7125 err = build_sched_domains(doms_cur[0], NULL);
7126 register_sched_domain_sysctl();
7132 * Detach sched domains from a group of cpus specified in cpu_map
7133 * These cpus will now be attached to the NULL domain
7135 static void detach_destroy_domains(const struct cpumask *cpu_map)
7140 for_each_cpu(i, cpu_map)
7141 cpu_attach_domain(NULL, &def_root_domain, i);
7145 /* handle null as "default" */
7146 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7147 struct sched_domain_attr *new, int idx_new)
7149 struct sched_domain_attr tmp;
7156 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7157 new ? (new + idx_new) : &tmp,
7158 sizeof(struct sched_domain_attr));
7162 * Partition sched domains as specified by the 'ndoms_new'
7163 * cpumasks in the array doms_new[] of cpumasks. This compares
7164 * doms_new[] to the current sched domain partitioning, doms_cur[].
7165 * It destroys each deleted domain and builds each new domain.
7167 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7168 * The masks don't intersect (don't overlap.) We should setup one
7169 * sched domain for each mask. CPUs not in any of the cpumasks will
7170 * not be load balanced. If the same cpumask appears both in the
7171 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7174 * The passed in 'doms_new' should be allocated using
7175 * alloc_sched_domains. This routine takes ownership of it and will
7176 * free_sched_domains it when done with it. If the caller failed the
7177 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7178 * and partition_sched_domains() will fallback to the single partition
7179 * 'fallback_doms', it also forces the domains to be rebuilt.
7181 * If doms_new == NULL it will be replaced with cpu_online_mask.
7182 * ndoms_new == 0 is a special case for destroying existing domains,
7183 * and it will not create the default domain.
7185 * Call with hotplug lock held
7187 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7188 struct sched_domain_attr *dattr_new)
7193 mutex_lock(&sched_domains_mutex);
7195 /* always unregister in case we don't destroy any domains */
7196 unregister_sched_domain_sysctl();
7198 /* Let architecture update cpu core mappings. */
7199 new_topology = arch_update_cpu_topology();
7201 n = doms_new ? ndoms_new : 0;
7203 /* Destroy deleted domains */
7204 for (i = 0; i < ndoms_cur; i++) {
7205 for (j = 0; j < n && !new_topology; j++) {
7206 if (cpumask_equal(doms_cur[i], doms_new[j])
7207 && dattrs_equal(dattr_cur, i, dattr_new, j))
7210 /* no match - a current sched domain not in new doms_new[] */
7211 detach_destroy_domains(doms_cur[i]);
7217 if (doms_new == NULL) {
7219 doms_new = &fallback_doms;
7220 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7221 WARN_ON_ONCE(dattr_new);
7224 /* Build new domains */
7225 for (i = 0; i < ndoms_new; i++) {
7226 for (j = 0; j < n && !new_topology; j++) {
7227 if (cpumask_equal(doms_new[i], doms_cur[j])
7228 && dattrs_equal(dattr_new, i, dattr_cur, j))
7231 /* no match - add a new doms_new */
7232 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7237 /* Remember the new sched domains */
7238 if (doms_cur != &fallback_doms)
7239 free_sched_domains(doms_cur, ndoms_cur);
7240 kfree(dattr_cur); /* kfree(NULL) is safe */
7241 doms_cur = doms_new;
7242 dattr_cur = dattr_new;
7243 ndoms_cur = ndoms_new;
7245 register_sched_domain_sysctl();
7247 mutex_unlock(&sched_domains_mutex);
7250 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7253 * Update cpusets according to cpu_active mask. If cpusets are
7254 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7255 * around partition_sched_domains().
7257 * If we come here as part of a suspend/resume, don't touch cpusets because we
7258 * want to restore it back to its original state upon resume anyway.
7260 static void cpuset_cpu_active(void)
7262 if (cpuhp_tasks_frozen) {
7264 * num_cpus_frozen tracks how many CPUs are involved in suspend
7265 * resume sequence. As long as this is not the last online
7266 * operation in the resume sequence, just build a single sched
7267 * domain, ignoring cpusets.
7270 if (likely(num_cpus_frozen)) {
7271 partition_sched_domains(1, NULL, NULL);
7275 * This is the last CPU online operation. So fall through and
7276 * restore the original sched domains by considering the
7277 * cpuset configurations.
7280 cpuset_update_active_cpus(true);
7283 static int cpuset_cpu_inactive(unsigned int cpu)
7285 unsigned long flags;
7290 if (!cpuhp_tasks_frozen) {
7291 rcu_read_lock_sched();
7292 dl_b = dl_bw_of(cpu);
7294 raw_spin_lock_irqsave(&dl_b->lock, flags);
7295 cpus = dl_bw_cpus(cpu);
7296 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7297 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7299 rcu_read_unlock_sched();
7303 cpuset_update_active_cpus(false);
7306 partition_sched_domains(1, NULL, NULL);
7311 int sched_cpu_activate(unsigned int cpu)
7313 struct rq *rq = cpu_rq(cpu);
7314 unsigned long flags;
7316 set_cpu_active(cpu, true);
7318 if (sched_smp_initialized) {
7319 sched_domains_numa_masks_set(cpu);
7320 cpuset_cpu_active();
7324 * Put the rq online, if not already. This happens:
7326 * 1) In the early boot process, because we build the real domains
7327 * after all cpus have been brought up.
7329 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
7332 raw_spin_lock_irqsave(&rq->lock, flags);
7334 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7337 raw_spin_unlock_irqrestore(&rq->lock, flags);
7339 update_max_interval();
7344 int sched_cpu_deactivate(unsigned int cpu)
7348 set_cpu_active(cpu, false);
7350 * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU
7351 * users of this state to go away such that all new such users will
7354 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
7355 * not imply sync_sched(), so wait for both.
7357 * Do sync before park smpboot threads to take care the rcu boost case.
7359 if (IS_ENABLED(CONFIG_PREEMPT))
7360 synchronize_rcu_mult(call_rcu, call_rcu_sched);
7364 if (!sched_smp_initialized)
7367 ret = cpuset_cpu_inactive(cpu);
7369 set_cpu_active(cpu, true);
7372 sched_domains_numa_masks_clear(cpu);
7376 static void sched_rq_cpu_starting(unsigned int cpu)
7378 struct rq *rq = cpu_rq(cpu);
7380 rq->calc_load_update = calc_load_update;
7381 update_max_interval();
7384 int sched_cpu_starting(unsigned int cpu)
7386 set_cpu_rq_start_time(cpu);
7387 sched_rq_cpu_starting(cpu);
7391 #ifdef CONFIG_HOTPLUG_CPU
7392 int sched_cpu_dying(unsigned int cpu)
7394 struct rq *rq = cpu_rq(cpu);
7395 unsigned long flags;
7397 /* Handle pending wakeups and then migrate everything off */
7398 sched_ttwu_pending();
7399 raw_spin_lock_irqsave(&rq->lock, flags);
7401 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7405 BUG_ON(rq->nr_running != 1);
7406 raw_spin_unlock_irqrestore(&rq->lock, flags);
7407 calc_load_migrate(rq);
7408 update_max_interval();
7409 nohz_balance_exit_idle(cpu);
7415 void __init sched_init_smp(void)
7417 cpumask_var_t non_isolated_cpus;
7419 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7420 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7425 * There's no userspace yet to cause hotplug operations; hence all the
7426 * cpu masks are stable and all blatant races in the below code cannot
7429 mutex_lock(&sched_domains_mutex);
7430 init_sched_domains(cpu_active_mask);
7431 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7432 if (cpumask_empty(non_isolated_cpus))
7433 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7434 mutex_unlock(&sched_domains_mutex);
7436 /* Move init over to a non-isolated CPU */
7437 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7439 sched_init_granularity();
7440 free_cpumask_var(non_isolated_cpus);
7442 init_sched_rt_class();
7443 init_sched_dl_class();
7444 sched_smp_initialized = true;
7447 static int __init migration_init(void)
7449 sched_rq_cpu_starting(smp_processor_id());
7452 early_initcall(migration_init);
7455 void __init sched_init_smp(void)
7457 sched_init_granularity();
7459 #endif /* CONFIG_SMP */
7461 int in_sched_functions(unsigned long addr)
7463 return in_lock_functions(addr) ||
7464 (addr >= (unsigned long)__sched_text_start
7465 && addr < (unsigned long)__sched_text_end);
7468 #ifdef CONFIG_CGROUP_SCHED
7470 * Default task group.
7471 * Every task in system belongs to this group at bootup.
7473 struct task_group root_task_group;
7474 LIST_HEAD(task_groups);
7476 /* Cacheline aligned slab cache for task_group */
7477 static struct kmem_cache *task_group_cache __read_mostly;
7480 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7481 DECLARE_PER_CPU(cpumask_var_t, select_idle_mask);
7483 void __init sched_init(void)
7486 unsigned long alloc_size = 0, ptr;
7488 #ifdef CONFIG_FAIR_GROUP_SCHED
7489 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7491 #ifdef CONFIG_RT_GROUP_SCHED
7492 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7495 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7497 #ifdef CONFIG_FAIR_GROUP_SCHED
7498 root_task_group.se = (struct sched_entity **)ptr;
7499 ptr += nr_cpu_ids * sizeof(void **);
7501 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7502 ptr += nr_cpu_ids * sizeof(void **);
7504 #endif /* CONFIG_FAIR_GROUP_SCHED */
7505 #ifdef CONFIG_RT_GROUP_SCHED
7506 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7507 ptr += nr_cpu_ids * sizeof(void **);
7509 root_task_group.rt_rq = (struct rt_rq **)ptr;
7510 ptr += nr_cpu_ids * sizeof(void **);
7512 #endif /* CONFIG_RT_GROUP_SCHED */
7514 #ifdef CONFIG_CPUMASK_OFFSTACK
7515 for_each_possible_cpu(i) {
7516 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7517 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7518 per_cpu(select_idle_mask, i) = (cpumask_var_t)kzalloc_node(
7519 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7521 #endif /* CONFIG_CPUMASK_OFFSTACK */
7523 init_rt_bandwidth(&def_rt_bandwidth,
7524 global_rt_period(), global_rt_runtime());
7525 init_dl_bandwidth(&def_dl_bandwidth,
7526 global_rt_period(), global_rt_runtime());
7529 init_defrootdomain();
7532 #ifdef CONFIG_RT_GROUP_SCHED
7533 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7534 global_rt_period(), global_rt_runtime());
7535 #endif /* CONFIG_RT_GROUP_SCHED */
7537 #ifdef CONFIG_CGROUP_SCHED
7538 task_group_cache = KMEM_CACHE(task_group, 0);
7540 list_add(&root_task_group.list, &task_groups);
7541 INIT_LIST_HEAD(&root_task_group.children);
7542 INIT_LIST_HEAD(&root_task_group.siblings);
7543 autogroup_init(&init_task);
7544 #endif /* CONFIG_CGROUP_SCHED */
7546 for_each_possible_cpu(i) {
7550 raw_spin_lock_init(&rq->lock);
7552 rq->calc_load_active = 0;
7553 rq->calc_load_update = jiffies + LOAD_FREQ;
7554 init_cfs_rq(&rq->cfs);
7555 init_rt_rq(&rq->rt);
7556 init_dl_rq(&rq->dl);
7557 #ifdef CONFIG_FAIR_GROUP_SCHED
7558 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7559 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7561 * How much cpu bandwidth does root_task_group get?
7563 * In case of task-groups formed thr' the cgroup filesystem, it
7564 * gets 100% of the cpu resources in the system. This overall
7565 * system cpu resource is divided among the tasks of
7566 * root_task_group and its child task-groups in a fair manner,
7567 * based on each entity's (task or task-group's) weight
7568 * (se->load.weight).
7570 * In other words, if root_task_group has 10 tasks of weight
7571 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7572 * then A0's share of the cpu resource is:
7574 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7576 * We achieve this by letting root_task_group's tasks sit
7577 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7579 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7580 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7581 #endif /* CONFIG_FAIR_GROUP_SCHED */
7583 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7584 #ifdef CONFIG_RT_GROUP_SCHED
7585 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7588 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7589 rq->cpu_load[j] = 0;
7594 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7595 rq->balance_callback = NULL;
7596 rq->active_balance = 0;
7597 rq->next_balance = jiffies;
7602 rq->avg_idle = 2*sysctl_sched_migration_cost;
7603 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7605 INIT_LIST_HEAD(&rq->cfs_tasks);
7607 rq_attach_root(rq, &def_root_domain);
7608 #ifdef CONFIG_NO_HZ_COMMON
7609 rq->last_load_update_tick = jiffies;
7612 #ifdef CONFIG_NO_HZ_FULL
7613 rq->last_sched_tick = 0;
7615 #endif /* CONFIG_SMP */
7617 atomic_set(&rq->nr_iowait, 0);
7620 set_load_weight(&init_task);
7623 * The boot idle thread does lazy MMU switching as well:
7625 atomic_inc(&init_mm.mm_count);
7626 enter_lazy_tlb(&init_mm, current);
7629 * Make us the idle thread. Technically, schedule() should not be
7630 * called from this thread, however somewhere below it might be,
7631 * but because we are the idle thread, we just pick up running again
7632 * when this runqueue becomes "idle".
7634 init_idle(current, smp_processor_id());
7636 calc_load_update = jiffies + LOAD_FREQ;
7639 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7640 /* May be allocated at isolcpus cmdline parse time */
7641 if (cpu_isolated_map == NULL)
7642 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7643 idle_thread_set_boot_cpu();
7644 set_cpu_rq_start_time(smp_processor_id());
7646 init_sched_fair_class();
7650 scheduler_running = 1;
7653 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7654 static inline int preempt_count_equals(int preempt_offset)
7656 int nested = preempt_count() + rcu_preempt_depth();
7658 return (nested == preempt_offset);
7661 void __might_sleep(const char *file, int line, int preempt_offset)
7664 * Blocking primitives will set (and therefore destroy) current->state,
7665 * since we will exit with TASK_RUNNING make sure we enter with it,
7666 * otherwise we will destroy state.
7668 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7669 "do not call blocking ops when !TASK_RUNNING; "
7670 "state=%lx set at [<%p>] %pS\n",
7672 (void *)current->task_state_change,
7673 (void *)current->task_state_change);
7675 ___might_sleep(file, line, preempt_offset);
7677 EXPORT_SYMBOL(__might_sleep);
7679 void ___might_sleep(const char *file, int line, int preempt_offset)
7681 static unsigned long prev_jiffy; /* ratelimiting */
7682 unsigned long preempt_disable_ip;
7684 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7685 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7686 !is_idle_task(current)) ||
7687 system_state != SYSTEM_RUNNING || oops_in_progress)
7689 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7691 prev_jiffy = jiffies;
7693 /* Save this before calling printk(), since that will clobber it */
7694 preempt_disable_ip = get_preempt_disable_ip(current);
7697 "BUG: sleeping function called from invalid context at %s:%d\n",
7700 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7701 in_atomic(), irqs_disabled(),
7702 current->pid, current->comm);
7704 if (task_stack_end_corrupted(current))
7705 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7707 debug_show_held_locks(current);
7708 if (irqs_disabled())
7709 print_irqtrace_events(current);
7710 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
7711 && !preempt_count_equals(preempt_offset)) {
7712 pr_err("Preemption disabled at:");
7713 print_ip_sym(preempt_disable_ip);
7717 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
7719 EXPORT_SYMBOL(___might_sleep);
7722 #ifdef CONFIG_MAGIC_SYSRQ
7723 void normalize_rt_tasks(void)
7725 struct task_struct *g, *p;
7726 struct sched_attr attr = {
7727 .sched_policy = SCHED_NORMAL,
7730 read_lock(&tasklist_lock);
7731 for_each_process_thread(g, p) {
7733 * Only normalize user tasks:
7735 if (p->flags & PF_KTHREAD)
7738 p->se.exec_start = 0;
7739 schedstat_set(p->se.statistics.wait_start, 0);
7740 schedstat_set(p->se.statistics.sleep_start, 0);
7741 schedstat_set(p->se.statistics.block_start, 0);
7743 if (!dl_task(p) && !rt_task(p)) {
7745 * Renice negative nice level userspace
7748 if (task_nice(p) < 0)
7749 set_user_nice(p, 0);
7753 __sched_setscheduler(p, &attr, false, false);
7755 read_unlock(&tasklist_lock);
7758 #endif /* CONFIG_MAGIC_SYSRQ */
7760 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7762 * These functions are only useful for the IA64 MCA handling, or kdb.
7764 * They can only be called when the whole system has been
7765 * stopped - every CPU needs to be quiescent, and no scheduling
7766 * activity can take place. Using them for anything else would
7767 * be a serious bug, and as a result, they aren't even visible
7768 * under any other configuration.
7772 * curr_task - return the current task for a given cpu.
7773 * @cpu: the processor in question.
7775 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7777 * Return: The current task for @cpu.
7779 struct task_struct *curr_task(int cpu)
7781 return cpu_curr(cpu);
7784 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7788 * set_curr_task - set the current task for a given cpu.
7789 * @cpu: the processor in question.
7790 * @p: the task pointer to set.
7792 * Description: This function must only be used when non-maskable interrupts
7793 * are serviced on a separate stack. It allows the architecture to switch the
7794 * notion of the current task on a cpu in a non-blocking manner. This function
7795 * must be called with all CPU's synchronized, and interrupts disabled, the
7796 * and caller must save the original value of the current task (see
7797 * curr_task() above) and restore that value before reenabling interrupts and
7798 * re-starting the system.
7800 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7802 void set_curr_task(int cpu, struct task_struct *p)
7809 #ifdef CONFIG_CGROUP_SCHED
7810 /* task_group_lock serializes the addition/removal of task groups */
7811 static DEFINE_SPINLOCK(task_group_lock);
7813 static void sched_free_group(struct task_group *tg)
7815 free_fair_sched_group(tg);
7816 free_rt_sched_group(tg);
7818 kmem_cache_free(task_group_cache, tg);
7821 /* allocate runqueue etc for a new task group */
7822 struct task_group *sched_create_group(struct task_group *parent)
7824 struct task_group *tg;
7826 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
7828 return ERR_PTR(-ENOMEM);
7830 if (!alloc_fair_sched_group(tg, parent))
7833 if (!alloc_rt_sched_group(tg, parent))
7839 sched_free_group(tg);
7840 return ERR_PTR(-ENOMEM);
7843 void sched_online_group(struct task_group *tg, struct task_group *parent)
7845 unsigned long flags;
7847 spin_lock_irqsave(&task_group_lock, flags);
7848 list_add_rcu(&tg->list, &task_groups);
7850 WARN_ON(!parent); /* root should already exist */
7852 tg->parent = parent;
7853 INIT_LIST_HEAD(&tg->children);
7854 list_add_rcu(&tg->siblings, &parent->children);
7855 spin_unlock_irqrestore(&task_group_lock, flags);
7857 online_fair_sched_group(tg);
7860 /* rcu callback to free various structures associated with a task group */
7861 static void sched_free_group_rcu(struct rcu_head *rhp)
7863 /* now it should be safe to free those cfs_rqs */
7864 sched_free_group(container_of(rhp, struct task_group, rcu));
7867 void sched_destroy_group(struct task_group *tg)
7869 /* wait for possible concurrent references to cfs_rqs complete */
7870 call_rcu(&tg->rcu, sched_free_group_rcu);
7873 void sched_offline_group(struct task_group *tg)
7875 unsigned long flags;
7877 /* end participation in shares distribution */
7878 unregister_fair_sched_group(tg);
7880 spin_lock_irqsave(&task_group_lock, flags);
7881 list_del_rcu(&tg->list);
7882 list_del_rcu(&tg->siblings);
7883 spin_unlock_irqrestore(&task_group_lock, flags);
7886 static void sched_change_group(struct task_struct *tsk, int type)
7888 struct task_group *tg;
7891 * All callers are synchronized by task_rq_lock(); we do not use RCU
7892 * which is pointless here. Thus, we pass "true" to task_css_check()
7893 * to prevent lockdep warnings.
7895 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7896 struct task_group, css);
7897 tg = autogroup_task_group(tsk, tg);
7898 tsk->sched_task_group = tg;
7900 #ifdef CONFIG_FAIR_GROUP_SCHED
7901 if (tsk->sched_class->task_change_group)
7902 tsk->sched_class->task_change_group(tsk, type);
7905 set_task_rq(tsk, task_cpu(tsk));
7909 * Change task's runqueue when it moves between groups.
7911 * The caller of this function should have put the task in its new group by
7912 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
7915 void sched_move_task(struct task_struct *tsk)
7917 int queued, running;
7921 rq = task_rq_lock(tsk, &rf);
7923 running = task_current(rq, tsk);
7924 queued = task_on_rq_queued(tsk);
7927 dequeue_task(rq, tsk, DEQUEUE_SAVE | DEQUEUE_MOVE);
7928 if (unlikely(running))
7929 put_prev_task(rq, tsk);
7931 sched_change_group(tsk, TASK_MOVE_GROUP);
7933 if (unlikely(running))
7934 tsk->sched_class->set_curr_task(rq);
7936 enqueue_task(rq, tsk, ENQUEUE_RESTORE | ENQUEUE_MOVE);
7938 task_rq_unlock(rq, tsk, &rf);
7940 #endif /* CONFIG_CGROUP_SCHED */
7942 #ifdef CONFIG_RT_GROUP_SCHED
7944 * Ensure that the real time constraints are schedulable.
7946 static DEFINE_MUTEX(rt_constraints_mutex);
7948 /* Must be called with tasklist_lock held */
7949 static inline int tg_has_rt_tasks(struct task_group *tg)
7951 struct task_struct *g, *p;
7954 * Autogroups do not have RT tasks; see autogroup_create().
7956 if (task_group_is_autogroup(tg))
7959 for_each_process_thread(g, p) {
7960 if (rt_task(p) && task_group(p) == tg)
7967 struct rt_schedulable_data {
7968 struct task_group *tg;
7973 static int tg_rt_schedulable(struct task_group *tg, void *data)
7975 struct rt_schedulable_data *d = data;
7976 struct task_group *child;
7977 unsigned long total, sum = 0;
7978 u64 period, runtime;
7980 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7981 runtime = tg->rt_bandwidth.rt_runtime;
7984 period = d->rt_period;
7985 runtime = d->rt_runtime;
7989 * Cannot have more runtime than the period.
7991 if (runtime > period && runtime != RUNTIME_INF)
7995 * Ensure we don't starve existing RT tasks.
7997 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8000 total = to_ratio(period, runtime);
8003 * Nobody can have more than the global setting allows.
8005 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8009 * The sum of our children's runtime should not exceed our own.
8011 list_for_each_entry_rcu(child, &tg->children, siblings) {
8012 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8013 runtime = child->rt_bandwidth.rt_runtime;
8015 if (child == d->tg) {
8016 period = d->rt_period;
8017 runtime = d->rt_runtime;
8020 sum += to_ratio(period, runtime);
8029 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8033 struct rt_schedulable_data data = {
8035 .rt_period = period,
8036 .rt_runtime = runtime,
8040 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8046 static int tg_set_rt_bandwidth(struct task_group *tg,
8047 u64 rt_period, u64 rt_runtime)
8052 * Disallowing the root group RT runtime is BAD, it would disallow the
8053 * kernel creating (and or operating) RT threads.
8055 if (tg == &root_task_group && rt_runtime == 0)
8058 /* No period doesn't make any sense. */
8062 mutex_lock(&rt_constraints_mutex);
8063 read_lock(&tasklist_lock);
8064 err = __rt_schedulable(tg, rt_period, rt_runtime);
8068 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8069 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8070 tg->rt_bandwidth.rt_runtime = rt_runtime;
8072 for_each_possible_cpu(i) {
8073 struct rt_rq *rt_rq = tg->rt_rq[i];
8075 raw_spin_lock(&rt_rq->rt_runtime_lock);
8076 rt_rq->rt_runtime = rt_runtime;
8077 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8079 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8081 read_unlock(&tasklist_lock);
8082 mutex_unlock(&rt_constraints_mutex);
8087 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8089 u64 rt_runtime, rt_period;
8091 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8092 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8093 if (rt_runtime_us < 0)
8094 rt_runtime = RUNTIME_INF;
8096 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8099 static long sched_group_rt_runtime(struct task_group *tg)
8103 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8106 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8107 do_div(rt_runtime_us, NSEC_PER_USEC);
8108 return rt_runtime_us;
8111 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8113 u64 rt_runtime, rt_period;
8115 rt_period = rt_period_us * NSEC_PER_USEC;
8116 rt_runtime = tg->rt_bandwidth.rt_runtime;
8118 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8121 static long sched_group_rt_period(struct task_group *tg)
8125 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8126 do_div(rt_period_us, NSEC_PER_USEC);
8127 return rt_period_us;
8129 #endif /* CONFIG_RT_GROUP_SCHED */
8131 #ifdef CONFIG_RT_GROUP_SCHED
8132 static int sched_rt_global_constraints(void)
8136 mutex_lock(&rt_constraints_mutex);
8137 read_lock(&tasklist_lock);
8138 ret = __rt_schedulable(NULL, 0, 0);
8139 read_unlock(&tasklist_lock);
8140 mutex_unlock(&rt_constraints_mutex);
8145 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8147 /* Don't accept realtime tasks when there is no way for them to run */
8148 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8154 #else /* !CONFIG_RT_GROUP_SCHED */
8155 static int sched_rt_global_constraints(void)
8157 unsigned long flags;
8160 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8161 for_each_possible_cpu(i) {
8162 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8164 raw_spin_lock(&rt_rq->rt_runtime_lock);
8165 rt_rq->rt_runtime = global_rt_runtime();
8166 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8168 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8172 #endif /* CONFIG_RT_GROUP_SCHED */
8174 static int sched_dl_global_validate(void)
8176 u64 runtime = global_rt_runtime();
8177 u64 period = global_rt_period();
8178 u64 new_bw = to_ratio(period, runtime);
8181 unsigned long flags;
8184 * Here we want to check the bandwidth not being set to some
8185 * value smaller than the currently allocated bandwidth in
8186 * any of the root_domains.
8188 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8189 * cycling on root_domains... Discussion on different/better
8190 * solutions is welcome!
8192 for_each_possible_cpu(cpu) {
8193 rcu_read_lock_sched();
8194 dl_b = dl_bw_of(cpu);
8196 raw_spin_lock_irqsave(&dl_b->lock, flags);
8197 if (new_bw < dl_b->total_bw)
8199 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8201 rcu_read_unlock_sched();
8210 static void sched_dl_do_global(void)
8215 unsigned long flags;
8217 def_dl_bandwidth.dl_period = global_rt_period();
8218 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8220 if (global_rt_runtime() != RUNTIME_INF)
8221 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8224 * FIXME: As above...
8226 for_each_possible_cpu(cpu) {
8227 rcu_read_lock_sched();
8228 dl_b = dl_bw_of(cpu);
8230 raw_spin_lock_irqsave(&dl_b->lock, flags);
8232 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8234 rcu_read_unlock_sched();
8238 static int sched_rt_global_validate(void)
8240 if (sysctl_sched_rt_period <= 0)
8243 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8244 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8250 static void sched_rt_do_global(void)
8252 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8253 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8256 int sched_rt_handler(struct ctl_table *table, int write,
8257 void __user *buffer, size_t *lenp,
8260 int old_period, old_runtime;
8261 static DEFINE_MUTEX(mutex);
8265 old_period = sysctl_sched_rt_period;
8266 old_runtime = sysctl_sched_rt_runtime;
8268 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8270 if (!ret && write) {
8271 ret = sched_rt_global_validate();
8275 ret = sched_dl_global_validate();
8279 ret = sched_rt_global_constraints();
8283 sched_rt_do_global();
8284 sched_dl_do_global();
8288 sysctl_sched_rt_period = old_period;
8289 sysctl_sched_rt_runtime = old_runtime;
8291 mutex_unlock(&mutex);
8296 int sched_rr_handler(struct ctl_table *table, int write,
8297 void __user *buffer, size_t *lenp,
8301 static DEFINE_MUTEX(mutex);
8304 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8305 /* make sure that internally we keep jiffies */
8306 /* also, writing zero resets timeslice to default */
8307 if (!ret && write) {
8308 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8309 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8311 mutex_unlock(&mutex);
8315 #ifdef CONFIG_CGROUP_SCHED
8317 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8319 return css ? container_of(css, struct task_group, css) : NULL;
8322 static struct cgroup_subsys_state *
8323 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8325 struct task_group *parent = css_tg(parent_css);
8326 struct task_group *tg;
8329 /* This is early initialization for the top cgroup */
8330 return &root_task_group.css;
8333 tg = sched_create_group(parent);
8335 return ERR_PTR(-ENOMEM);
8337 sched_online_group(tg, parent);
8342 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8344 struct task_group *tg = css_tg(css);
8346 sched_offline_group(tg);
8349 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8351 struct task_group *tg = css_tg(css);
8354 * Relies on the RCU grace period between css_released() and this.
8356 sched_free_group(tg);
8360 * This is called before wake_up_new_task(), therefore we really only
8361 * have to set its group bits, all the other stuff does not apply.
8363 static void cpu_cgroup_fork(struct task_struct *task)
8368 rq = task_rq_lock(task, &rf);
8370 sched_change_group(task, TASK_SET_GROUP);
8372 task_rq_unlock(rq, task, &rf);
8375 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8377 struct task_struct *task;
8378 struct cgroup_subsys_state *css;
8381 cgroup_taskset_for_each(task, css, tset) {
8382 #ifdef CONFIG_RT_GROUP_SCHED
8383 if (!sched_rt_can_attach(css_tg(css), task))
8386 /* We don't support RT-tasks being in separate groups */
8387 if (task->sched_class != &fair_sched_class)
8391 * Serialize against wake_up_new_task() such that if its
8392 * running, we're sure to observe its full state.
8394 raw_spin_lock_irq(&task->pi_lock);
8396 * Avoid calling sched_move_task() before wake_up_new_task()
8397 * has happened. This would lead to problems with PELT, due to
8398 * move wanting to detach+attach while we're not attached yet.
8400 if (task->state == TASK_NEW)
8402 raw_spin_unlock_irq(&task->pi_lock);
8410 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8412 struct task_struct *task;
8413 struct cgroup_subsys_state *css;
8415 cgroup_taskset_for_each(task, css, tset)
8416 sched_move_task(task);
8419 #ifdef CONFIG_FAIR_GROUP_SCHED
8420 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8421 struct cftype *cftype, u64 shareval)
8423 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8426 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8429 struct task_group *tg = css_tg(css);
8431 return (u64) scale_load_down(tg->shares);
8434 #ifdef CONFIG_CFS_BANDWIDTH
8435 static DEFINE_MUTEX(cfs_constraints_mutex);
8437 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8438 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8440 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8442 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8444 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8445 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8447 if (tg == &root_task_group)
8451 * Ensure we have at some amount of bandwidth every period. This is
8452 * to prevent reaching a state of large arrears when throttled via
8453 * entity_tick() resulting in prolonged exit starvation.
8455 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8459 * Likewise, bound things on the otherside by preventing insane quota
8460 * periods. This also allows us to normalize in computing quota
8463 if (period > max_cfs_quota_period)
8467 * Prevent race between setting of cfs_rq->runtime_enabled and
8468 * unthrottle_offline_cfs_rqs().
8471 mutex_lock(&cfs_constraints_mutex);
8472 ret = __cfs_schedulable(tg, period, quota);
8476 runtime_enabled = quota != RUNTIME_INF;
8477 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8479 * If we need to toggle cfs_bandwidth_used, off->on must occur
8480 * before making related changes, and on->off must occur afterwards
8482 if (runtime_enabled && !runtime_was_enabled)
8483 cfs_bandwidth_usage_inc();
8484 raw_spin_lock_irq(&cfs_b->lock);
8485 cfs_b->period = ns_to_ktime(period);
8486 cfs_b->quota = quota;
8488 __refill_cfs_bandwidth_runtime(cfs_b);
8489 /* restart the period timer (if active) to handle new period expiry */
8490 if (runtime_enabled)
8491 start_cfs_bandwidth(cfs_b);
8492 raw_spin_unlock_irq(&cfs_b->lock);
8494 for_each_online_cpu(i) {
8495 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8496 struct rq *rq = cfs_rq->rq;
8498 raw_spin_lock_irq(&rq->lock);
8499 cfs_rq->runtime_enabled = runtime_enabled;
8500 cfs_rq->runtime_remaining = 0;
8502 if (cfs_rq->throttled)
8503 unthrottle_cfs_rq(cfs_rq);
8504 raw_spin_unlock_irq(&rq->lock);
8506 if (runtime_was_enabled && !runtime_enabled)
8507 cfs_bandwidth_usage_dec();
8509 mutex_unlock(&cfs_constraints_mutex);
8515 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8519 period = ktime_to_ns(tg->cfs_bandwidth.period);
8520 if (cfs_quota_us < 0)
8521 quota = RUNTIME_INF;
8523 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8525 return tg_set_cfs_bandwidth(tg, period, quota);
8528 long tg_get_cfs_quota(struct task_group *tg)
8532 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8535 quota_us = tg->cfs_bandwidth.quota;
8536 do_div(quota_us, NSEC_PER_USEC);
8541 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8545 period = (u64)cfs_period_us * NSEC_PER_USEC;
8546 quota = tg->cfs_bandwidth.quota;
8548 return tg_set_cfs_bandwidth(tg, period, quota);
8551 long tg_get_cfs_period(struct task_group *tg)
8555 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8556 do_div(cfs_period_us, NSEC_PER_USEC);
8558 return cfs_period_us;
8561 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8564 return tg_get_cfs_quota(css_tg(css));
8567 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8568 struct cftype *cftype, s64 cfs_quota_us)
8570 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8573 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8576 return tg_get_cfs_period(css_tg(css));
8579 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8580 struct cftype *cftype, u64 cfs_period_us)
8582 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8585 struct cfs_schedulable_data {
8586 struct task_group *tg;
8591 * normalize group quota/period to be quota/max_period
8592 * note: units are usecs
8594 static u64 normalize_cfs_quota(struct task_group *tg,
8595 struct cfs_schedulable_data *d)
8603 period = tg_get_cfs_period(tg);
8604 quota = tg_get_cfs_quota(tg);
8607 /* note: these should typically be equivalent */
8608 if (quota == RUNTIME_INF || quota == -1)
8611 return to_ratio(period, quota);
8614 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8616 struct cfs_schedulable_data *d = data;
8617 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8618 s64 quota = 0, parent_quota = -1;
8621 quota = RUNTIME_INF;
8623 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8625 quota = normalize_cfs_quota(tg, d);
8626 parent_quota = parent_b->hierarchical_quota;
8629 * ensure max(child_quota) <= parent_quota, inherit when no
8632 if (quota == RUNTIME_INF)
8633 quota = parent_quota;
8634 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8637 cfs_b->hierarchical_quota = quota;
8642 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8645 struct cfs_schedulable_data data = {
8651 if (quota != RUNTIME_INF) {
8652 do_div(data.period, NSEC_PER_USEC);
8653 do_div(data.quota, NSEC_PER_USEC);
8657 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8663 static int cpu_stats_show(struct seq_file *sf, void *v)
8665 struct task_group *tg = css_tg(seq_css(sf));
8666 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8668 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8669 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8670 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8674 #endif /* CONFIG_CFS_BANDWIDTH */
8675 #endif /* CONFIG_FAIR_GROUP_SCHED */
8677 #ifdef CONFIG_RT_GROUP_SCHED
8678 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8679 struct cftype *cft, s64 val)
8681 return sched_group_set_rt_runtime(css_tg(css), val);
8684 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8687 return sched_group_rt_runtime(css_tg(css));
8690 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8691 struct cftype *cftype, u64 rt_period_us)
8693 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8696 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8699 return sched_group_rt_period(css_tg(css));
8701 #endif /* CONFIG_RT_GROUP_SCHED */
8703 static struct cftype cpu_files[] = {
8704 #ifdef CONFIG_FAIR_GROUP_SCHED
8707 .read_u64 = cpu_shares_read_u64,
8708 .write_u64 = cpu_shares_write_u64,
8711 #ifdef CONFIG_CFS_BANDWIDTH
8713 .name = "cfs_quota_us",
8714 .read_s64 = cpu_cfs_quota_read_s64,
8715 .write_s64 = cpu_cfs_quota_write_s64,
8718 .name = "cfs_period_us",
8719 .read_u64 = cpu_cfs_period_read_u64,
8720 .write_u64 = cpu_cfs_period_write_u64,
8724 .seq_show = cpu_stats_show,
8727 #ifdef CONFIG_RT_GROUP_SCHED
8729 .name = "rt_runtime_us",
8730 .read_s64 = cpu_rt_runtime_read,
8731 .write_s64 = cpu_rt_runtime_write,
8734 .name = "rt_period_us",
8735 .read_u64 = cpu_rt_period_read_uint,
8736 .write_u64 = cpu_rt_period_write_uint,
8742 struct cgroup_subsys cpu_cgrp_subsys = {
8743 .css_alloc = cpu_cgroup_css_alloc,
8744 .css_released = cpu_cgroup_css_released,
8745 .css_free = cpu_cgroup_css_free,
8746 .fork = cpu_cgroup_fork,
8747 .can_attach = cpu_cgroup_can_attach,
8748 .attach = cpu_cgroup_attach,
8749 .legacy_cftypes = cpu_files,
8753 #endif /* CONFIG_CGROUP_SCHED */
8755 void dump_cpu_task(int cpu)
8757 pr_info("Task dump for CPU %d:\n", cpu);
8758 sched_show_task(cpu_curr(cpu));
8762 * Nice levels are multiplicative, with a gentle 10% change for every
8763 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
8764 * nice 1, it will get ~10% less CPU time than another CPU-bound task
8765 * that remained on nice 0.
8767 * The "10% effect" is relative and cumulative: from _any_ nice level,
8768 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
8769 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
8770 * If a task goes up by ~10% and another task goes down by ~10% then
8771 * the relative distance between them is ~25%.)
8773 const int sched_prio_to_weight[40] = {
8774 /* -20 */ 88761, 71755, 56483, 46273, 36291,
8775 /* -15 */ 29154, 23254, 18705, 14949, 11916,
8776 /* -10 */ 9548, 7620, 6100, 4904, 3906,
8777 /* -5 */ 3121, 2501, 1991, 1586, 1277,
8778 /* 0 */ 1024, 820, 655, 526, 423,
8779 /* 5 */ 335, 272, 215, 172, 137,
8780 /* 10 */ 110, 87, 70, 56, 45,
8781 /* 15 */ 36, 29, 23, 18, 15,
8785 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
8787 * In cases where the weight does not change often, we can use the
8788 * precalculated inverse to speed up arithmetics by turning divisions
8789 * into multiplications:
8791 const u32 sched_prio_to_wmult[40] = {
8792 /* -20 */ 48388, 59856, 76040, 92818, 118348,
8793 /* -15 */ 147320, 184698, 229616, 287308, 360437,
8794 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
8795 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
8796 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
8797 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
8798 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
8799 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,