4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.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/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
96 static void update_rq_clock_task(struct rq *rq, s64 delta);
98 void update_rq_clock(struct rq *rq)
102 lockdep_assert_held(&rq->lock);
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
111 update_rq_clock_task(rq, delta);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug unsigned int sysctl_sched_features =
122 #include "features.h"
127 #ifdef CONFIG_SCHED_DEBUG
128 #define SCHED_FEAT(name, enabled) \
131 static const char * const sched_feat_names[] = {
132 #include "features.h"
137 static int sched_feat_show(struct seq_file *m, void *v)
141 for (i = 0; i < __SCHED_FEAT_NR; i++) {
142 if (!(sysctl_sched_features & (1UL << i)))
144 seq_printf(m, "%s ", sched_feat_names[i]);
151 #ifdef HAVE_JUMP_LABEL
153 #define jump_label_key__true STATIC_KEY_INIT_TRUE
154 #define jump_label_key__false STATIC_KEY_INIT_FALSE
156 #define SCHED_FEAT(name, enabled) \
157 jump_label_key__##enabled ,
159 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
160 #include "features.h"
165 static void sched_feat_disable(int i)
167 if (static_key_enabled(&sched_feat_keys[i]))
168 static_key_slow_dec(&sched_feat_keys[i]);
171 static void sched_feat_enable(int i)
173 if (!static_key_enabled(&sched_feat_keys[i]))
174 static_key_slow_inc(&sched_feat_keys[i]);
177 static void sched_feat_disable(int i) { };
178 static void sched_feat_enable(int i) { };
179 #endif /* HAVE_JUMP_LABEL */
181 static int sched_feat_set(char *cmp)
186 if (strncmp(cmp, "NO_", 3) == 0) {
191 for (i = 0; i < __SCHED_FEAT_NR; i++) {
192 if (strcmp(cmp, sched_feat_names[i]) == 0) {
194 sysctl_sched_features &= ~(1UL << i);
195 sched_feat_disable(i);
197 sysctl_sched_features |= (1UL << i);
198 sched_feat_enable(i);
208 sched_feat_write(struct file *filp, const char __user *ubuf,
209 size_t cnt, loff_t *ppos)
219 if (copy_from_user(&buf, ubuf, cnt))
225 /* Ensure the static_key remains in a consistent state */
226 inode = file_inode(filp);
227 mutex_lock(&inode->i_mutex);
228 i = sched_feat_set(cmp);
229 mutex_unlock(&inode->i_mutex);
230 if (i == __SCHED_FEAT_NR)
238 static int sched_feat_open(struct inode *inode, struct file *filp)
240 return single_open(filp, sched_feat_show, NULL);
243 static const struct file_operations sched_feat_fops = {
244 .open = sched_feat_open,
245 .write = sched_feat_write,
248 .release = single_release,
251 static __init int sched_init_debug(void)
253 debugfs_create_file("sched_features", 0644, NULL, NULL,
258 late_initcall(sched_init_debug);
259 #endif /* CONFIG_SCHED_DEBUG */
262 * Number of tasks to iterate in a single balance run.
263 * Limited because this is done with IRQs disabled.
265 const_debug unsigned int sysctl_sched_nr_migrate = 32;
268 * period over which we average the RT time consumption, measured
273 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
276 * period over which we measure -rt task cpu usage in us.
279 unsigned int sysctl_sched_rt_period = 1000000;
281 __read_mostly int scheduler_running;
284 * part of the period that we allow rt tasks to run in us.
287 int sysctl_sched_rt_runtime = 950000;
289 /* cpus with isolated domains */
290 cpumask_var_t cpu_isolated_map;
293 * this_rq_lock - lock this runqueue and disable interrupts.
295 static struct rq *this_rq_lock(void)
302 raw_spin_lock(&rq->lock);
307 #ifdef CONFIG_SCHED_HRTICK
309 * Use HR-timers to deliver accurate preemption points.
312 static void hrtick_clear(struct rq *rq)
314 if (hrtimer_active(&rq->hrtick_timer))
315 hrtimer_cancel(&rq->hrtick_timer);
319 * High-resolution timer tick.
320 * Runs from hardirq context with interrupts disabled.
322 static enum hrtimer_restart hrtick(struct hrtimer *timer)
324 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
326 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
328 raw_spin_lock(&rq->lock);
330 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
331 raw_spin_unlock(&rq->lock);
333 return HRTIMER_NORESTART;
338 static void __hrtick_restart(struct rq *rq)
340 struct hrtimer *timer = &rq->hrtick_timer;
342 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
346 * called from hardirq (IPI) context
348 static void __hrtick_start(void *arg)
352 raw_spin_lock(&rq->lock);
353 __hrtick_restart(rq);
354 rq->hrtick_csd_pending = 0;
355 raw_spin_unlock(&rq->lock);
359 * Called to set the hrtick timer state.
361 * called with rq->lock held and irqs disabled
363 void hrtick_start(struct rq *rq, u64 delay)
365 struct hrtimer *timer = &rq->hrtick_timer;
370 * Don't schedule slices shorter than 10000ns, that just
371 * doesn't make sense and can cause timer DoS.
373 delta = max_t(s64, delay, 10000LL);
374 time = ktime_add_ns(timer->base->get_time(), delta);
376 hrtimer_set_expires(timer, time);
378 if (rq == this_rq()) {
379 __hrtick_restart(rq);
380 } else if (!rq->hrtick_csd_pending) {
381 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
382 rq->hrtick_csd_pending = 1;
387 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
389 int cpu = (int)(long)hcpu;
392 case CPU_UP_CANCELED:
393 case CPU_UP_CANCELED_FROZEN:
394 case CPU_DOWN_PREPARE:
395 case CPU_DOWN_PREPARE_FROZEN:
397 case CPU_DEAD_FROZEN:
398 hrtick_clear(cpu_rq(cpu));
405 static __init void init_hrtick(void)
407 hotcpu_notifier(hotplug_hrtick, 0);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq *rq, u64 delay)
418 * Don't schedule slices shorter than 10000ns, that just
419 * doesn't make sense. Rely on vruntime for fairness.
421 delay = max_t(u64, delay, 10000LL);
422 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
423 HRTIMER_MODE_REL_PINNED);
426 static inline void init_hrtick(void)
429 #endif /* CONFIG_SMP */
431 static void init_rq_hrtick(struct rq *rq)
434 rq->hrtick_csd_pending = 0;
436 rq->hrtick_csd.flags = 0;
437 rq->hrtick_csd.func = __hrtick_start;
438 rq->hrtick_csd.info = rq;
441 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
442 rq->hrtick_timer.function = hrtick;
444 #else /* CONFIG_SCHED_HRTICK */
445 static inline void hrtick_clear(struct rq *rq)
449 static inline void init_rq_hrtick(struct rq *rq)
453 static inline void init_hrtick(void)
456 #endif /* CONFIG_SCHED_HRTICK */
459 * cmpxchg based fetch_or, macro so it works for different integer types
461 #define fetch_or(ptr, val) \
462 ({ typeof(*(ptr)) __old, __val = *(ptr); \
464 __old = cmpxchg((ptr), __val, __val | (val)); \
465 if (__old == __val) \
472 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
474 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
475 * this avoids any races wrt polling state changes and thereby avoids
478 static bool set_nr_and_not_polling(struct task_struct *p)
480 struct thread_info *ti = task_thread_info(p);
481 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
485 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
487 * If this returns true, then the idle task promises to call
488 * sched_ttwu_pending() and reschedule soon.
490 static bool set_nr_if_polling(struct task_struct *p)
492 struct thread_info *ti = task_thread_info(p);
493 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
496 if (!(val & _TIF_POLLING_NRFLAG))
498 if (val & _TIF_NEED_RESCHED)
500 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
509 static bool set_nr_and_not_polling(struct task_struct *p)
511 set_tsk_need_resched(p);
516 static bool set_nr_if_polling(struct task_struct *p)
523 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
525 struct wake_q_node *node = &task->wake_q;
528 * Atomically grab the task, if ->wake_q is !nil already it means
529 * its already queued (either by us or someone else) and will get the
530 * wakeup due to that.
532 * This cmpxchg() implies a full barrier, which pairs with the write
533 * barrier implied by the wakeup in wake_up_list().
535 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
538 get_task_struct(task);
541 * The head is context local, there can be no concurrency.
544 head->lastp = &node->next;
547 void wake_up_q(struct wake_q_head *head)
549 struct wake_q_node *node = head->first;
551 while (node != WAKE_Q_TAIL) {
552 struct task_struct *task;
554 task = container_of(node, struct task_struct, wake_q);
556 /* task can safely be re-inserted now */
558 task->wake_q.next = NULL;
561 * wake_up_process() implies a wmb() to pair with the queueing
562 * in wake_q_add() so as not to miss wakeups.
564 wake_up_process(task);
565 put_task_struct(task);
570 * resched_curr - mark rq's current task 'to be rescheduled now'.
572 * On UP this means the setting of the need_resched flag, on SMP it
573 * might also involve a cross-CPU call to trigger the scheduler on
576 void resched_curr(struct rq *rq)
578 struct task_struct *curr = rq->curr;
581 lockdep_assert_held(&rq->lock);
583 if (test_tsk_need_resched(curr))
588 if (cpu == smp_processor_id()) {
589 set_tsk_need_resched(curr);
590 set_preempt_need_resched();
594 if (set_nr_and_not_polling(curr))
595 smp_send_reschedule(cpu);
597 trace_sched_wake_idle_without_ipi(cpu);
600 void resched_cpu(int cpu)
602 struct rq *rq = cpu_rq(cpu);
605 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
608 raw_spin_unlock_irqrestore(&rq->lock, flags);
612 #ifdef CONFIG_NO_HZ_COMMON
614 * In the semi idle case, use the nearest busy cpu for migrating timers
615 * from an idle cpu. This is good for power-savings.
617 * We don't do similar optimization for completely idle system, as
618 * selecting an idle cpu will add more delays to the timers than intended
619 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
621 int get_nohz_timer_target(void)
623 int i, cpu = smp_processor_id();
624 struct sched_domain *sd;
626 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
630 for_each_domain(cpu, sd) {
631 for_each_cpu(i, sched_domain_span(sd)) {
632 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
639 if (!is_housekeeping_cpu(cpu))
640 cpu = housekeeping_any_cpu();
646 * When add_timer_on() enqueues a timer into the timer wheel of an
647 * idle CPU then this timer might expire before the next timer event
648 * which is scheduled to wake up that CPU. In case of a completely
649 * idle system the next event might even be infinite time into the
650 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
651 * leaves the inner idle loop so the newly added timer is taken into
652 * account when the CPU goes back to idle and evaluates the timer
653 * wheel for the next timer event.
655 static void wake_up_idle_cpu(int cpu)
657 struct rq *rq = cpu_rq(cpu);
659 if (cpu == smp_processor_id())
662 if (set_nr_and_not_polling(rq->idle))
663 smp_send_reschedule(cpu);
665 trace_sched_wake_idle_without_ipi(cpu);
668 static bool wake_up_full_nohz_cpu(int cpu)
671 * We just need the target to call irq_exit() and re-evaluate
672 * the next tick. The nohz full kick at least implies that.
673 * If needed we can still optimize that later with an
676 if (tick_nohz_full_cpu(cpu)) {
677 if (cpu != smp_processor_id() ||
678 tick_nohz_tick_stopped())
679 tick_nohz_full_kick_cpu(cpu);
686 void wake_up_nohz_cpu(int cpu)
688 if (!wake_up_full_nohz_cpu(cpu))
689 wake_up_idle_cpu(cpu);
692 static inline bool got_nohz_idle_kick(void)
694 int cpu = smp_processor_id();
696 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
699 if (idle_cpu(cpu) && !need_resched())
703 * We can't run Idle Load Balance on this CPU for this time so we
704 * cancel it and clear NOHZ_BALANCE_KICK
706 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
710 #else /* CONFIG_NO_HZ_COMMON */
712 static inline bool got_nohz_idle_kick(void)
717 #endif /* CONFIG_NO_HZ_COMMON */
719 #ifdef CONFIG_NO_HZ_FULL
720 bool sched_can_stop_tick(void)
723 * FIFO realtime policy runs the highest priority task. Other runnable
724 * tasks are of a lower priority. The scheduler tick does nothing.
726 if (current->policy == SCHED_FIFO)
730 * Round-robin realtime tasks time slice with other tasks at the same
731 * realtime priority. Is this task the only one at this priority?
733 if (current->policy == SCHED_RR) {
734 struct sched_rt_entity *rt_se = ¤t->rt;
736 return rt_se->run_list.prev == rt_se->run_list.next;
740 * More than one running task need preemption.
741 * nr_running update is assumed to be visible
742 * after IPI is sent from wakers.
744 if (this_rq()->nr_running > 1)
749 #endif /* CONFIG_NO_HZ_FULL */
751 void sched_avg_update(struct rq *rq)
753 s64 period = sched_avg_period();
755 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
757 * Inline assembly required to prevent the compiler
758 * optimising this loop into a divmod call.
759 * See __iter_div_u64_rem() for another example of this.
761 asm("" : "+rm" (rq->age_stamp));
762 rq->age_stamp += period;
767 #endif /* CONFIG_SMP */
769 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
770 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
772 * Iterate task_group tree rooted at *from, calling @down when first entering a
773 * node and @up when leaving it for the final time.
775 * Caller must hold rcu_lock or sufficient equivalent.
777 int walk_tg_tree_from(struct task_group *from,
778 tg_visitor down, tg_visitor up, void *data)
780 struct task_group *parent, *child;
786 ret = (*down)(parent, data);
789 list_for_each_entry_rcu(child, &parent->children, siblings) {
796 ret = (*up)(parent, data);
797 if (ret || parent == from)
801 parent = parent->parent;
808 int tg_nop(struct task_group *tg, void *data)
814 static void set_load_weight(struct task_struct *p)
816 int prio = p->static_prio - MAX_RT_PRIO;
817 struct load_weight *load = &p->se.load;
820 * SCHED_IDLE tasks get minimal weight:
822 if (p->policy == SCHED_IDLE) {
823 load->weight = scale_load(WEIGHT_IDLEPRIO);
824 load->inv_weight = WMULT_IDLEPRIO;
828 load->weight = scale_load(prio_to_weight[prio]);
829 load->inv_weight = prio_to_wmult[prio];
832 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
835 sched_info_queued(rq, p);
836 p->sched_class->enqueue_task(rq, p, flags);
839 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
842 sched_info_dequeued(rq, p);
843 p->sched_class->dequeue_task(rq, p, flags);
846 void activate_task(struct rq *rq, struct task_struct *p, int flags)
848 if (task_contributes_to_load(p))
849 rq->nr_uninterruptible--;
851 enqueue_task(rq, p, flags);
854 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
856 if (task_contributes_to_load(p))
857 rq->nr_uninterruptible++;
859 dequeue_task(rq, p, flags);
862 static void update_rq_clock_task(struct rq *rq, s64 delta)
865 * In theory, the compile should just see 0 here, and optimize out the call
866 * to sched_rt_avg_update. But I don't trust it...
868 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
869 s64 steal = 0, irq_delta = 0;
871 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
872 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
875 * Since irq_time is only updated on {soft,}irq_exit, we might run into
876 * this case when a previous update_rq_clock() happened inside a
879 * When this happens, we stop ->clock_task and only update the
880 * prev_irq_time stamp to account for the part that fit, so that a next
881 * update will consume the rest. This ensures ->clock_task is
884 * It does however cause some slight miss-attribution of {soft,}irq
885 * time, a more accurate solution would be to update the irq_time using
886 * the current rq->clock timestamp, except that would require using
889 if (irq_delta > delta)
892 rq->prev_irq_time += irq_delta;
895 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
896 if (static_key_false((¶virt_steal_rq_enabled))) {
897 steal = paravirt_steal_clock(cpu_of(rq));
898 steal -= rq->prev_steal_time_rq;
900 if (unlikely(steal > delta))
903 rq->prev_steal_time_rq += steal;
908 rq->clock_task += delta;
910 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
911 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
912 sched_rt_avg_update(rq, irq_delta + steal);
916 void sched_set_stop_task(int cpu, struct task_struct *stop)
918 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
919 struct task_struct *old_stop = cpu_rq(cpu)->stop;
923 * Make it appear like a SCHED_FIFO task, its something
924 * userspace knows about and won't get confused about.
926 * Also, it will make PI more or less work without too
927 * much confusion -- but then, stop work should not
928 * rely on PI working anyway.
930 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
932 stop->sched_class = &stop_sched_class;
935 cpu_rq(cpu)->stop = stop;
939 * Reset it back to a normal scheduling class so that
940 * it can die in pieces.
942 old_stop->sched_class = &rt_sched_class;
947 * __normal_prio - return the priority that is based on the static prio
949 static inline int __normal_prio(struct task_struct *p)
951 return p->static_prio;
955 * Calculate the expected normal priority: i.e. priority
956 * without taking RT-inheritance into account. Might be
957 * boosted by interactivity modifiers. Changes upon fork,
958 * setprio syscalls, and whenever the interactivity
959 * estimator recalculates.
961 static inline int normal_prio(struct task_struct *p)
965 if (task_has_dl_policy(p))
966 prio = MAX_DL_PRIO-1;
967 else if (task_has_rt_policy(p))
968 prio = MAX_RT_PRIO-1 - p->rt_priority;
970 prio = __normal_prio(p);
975 * Calculate the current priority, i.e. the priority
976 * taken into account by the scheduler. This value might
977 * be boosted by RT tasks, or might be boosted by
978 * interactivity modifiers. Will be RT if the task got
979 * RT-boosted. If not then it returns p->normal_prio.
981 static int effective_prio(struct task_struct *p)
983 p->normal_prio = normal_prio(p);
985 * If we are RT tasks or we were boosted to RT priority,
986 * keep the priority unchanged. Otherwise, update priority
987 * to the normal priority:
989 if (!rt_prio(p->prio))
990 return p->normal_prio;
995 * task_curr - is this task currently executing on a CPU?
996 * @p: the task in question.
998 * Return: 1 if the task is currently executing. 0 otherwise.
1000 inline int task_curr(const struct task_struct *p)
1002 return cpu_curr(task_cpu(p)) == p;
1006 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1007 * use the balance_callback list if you want balancing.
1009 * this means any call to check_class_changed() must be followed by a call to
1010 * balance_callback().
1012 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1013 const struct sched_class *prev_class,
1016 if (prev_class != p->sched_class) {
1017 if (prev_class->switched_from)
1018 prev_class->switched_from(rq, p);
1020 p->sched_class->switched_to(rq, p);
1021 } else if (oldprio != p->prio || dl_task(p))
1022 p->sched_class->prio_changed(rq, p, oldprio);
1025 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1027 const struct sched_class *class;
1029 if (p->sched_class == rq->curr->sched_class) {
1030 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1032 for_each_class(class) {
1033 if (class == rq->curr->sched_class)
1035 if (class == p->sched_class) {
1043 * A queue event has occurred, and we're going to schedule. In
1044 * this case, we can save a useless back to back clock update.
1046 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1047 rq_clock_skip_update(rq, true);
1052 * This is how migration works:
1054 * 1) we invoke migration_cpu_stop() on the target CPU using
1056 * 2) stopper starts to run (implicitly forcing the migrated thread
1058 * 3) it checks whether the migrated task is still in the wrong runqueue.
1059 * 4) if it's in the wrong runqueue then the migration thread removes
1060 * it and puts it into the right queue.
1061 * 5) stopper completes and stop_one_cpu() returns and the migration
1066 * move_queued_task - move a queued task to new rq.
1068 * Returns (locked) new rq. Old rq's lock is released.
1070 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1072 lockdep_assert_held(&rq->lock);
1074 dequeue_task(rq, p, 0);
1075 p->on_rq = TASK_ON_RQ_MIGRATING;
1076 set_task_cpu(p, new_cpu);
1077 raw_spin_unlock(&rq->lock);
1079 rq = cpu_rq(new_cpu);
1081 raw_spin_lock(&rq->lock);
1082 BUG_ON(task_cpu(p) != new_cpu);
1083 p->on_rq = TASK_ON_RQ_QUEUED;
1084 enqueue_task(rq, p, 0);
1085 check_preempt_curr(rq, p, 0);
1090 struct migration_arg {
1091 struct task_struct *task;
1096 * Move (not current) task off this cpu, onto dest cpu. We're doing
1097 * this because either it can't run here any more (set_cpus_allowed()
1098 * away from this CPU, or CPU going down), or because we're
1099 * attempting to rebalance this task on exec (sched_exec).
1101 * So we race with normal scheduler movements, but that's OK, as long
1102 * as the task is no longer on this CPU.
1104 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1106 if (unlikely(!cpu_active(dest_cpu)))
1109 /* Affinity changed (again). */
1110 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1113 rq = move_queued_task(rq, p, dest_cpu);
1119 * migration_cpu_stop - this will be executed by a highprio stopper thread
1120 * and performs thread migration by bumping thread off CPU then
1121 * 'pushing' onto another runqueue.
1123 static int migration_cpu_stop(void *data)
1125 struct migration_arg *arg = data;
1126 struct task_struct *p = arg->task;
1127 struct rq *rq = this_rq();
1130 * The original target cpu might have gone down and we might
1131 * be on another cpu but it doesn't matter.
1133 local_irq_disable();
1135 * We need to explicitly wake pending tasks before running
1136 * __migrate_task() such that we will not miss enforcing cpus_allowed
1137 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1139 sched_ttwu_pending();
1141 raw_spin_lock(&p->pi_lock);
1142 raw_spin_lock(&rq->lock);
1144 * If task_rq(p) != rq, it cannot be migrated here, because we're
1145 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1146 * we're holding p->pi_lock.
1148 if (task_rq(p) == rq && task_on_rq_queued(p))
1149 rq = __migrate_task(rq, p, arg->dest_cpu);
1150 raw_spin_unlock(&rq->lock);
1151 raw_spin_unlock(&p->pi_lock);
1158 * sched_class::set_cpus_allowed must do the below, but is not required to
1159 * actually call this function.
1161 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1163 cpumask_copy(&p->cpus_allowed, new_mask);
1164 p->nr_cpus_allowed = cpumask_weight(new_mask);
1167 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1169 struct rq *rq = task_rq(p);
1170 bool queued, running;
1172 lockdep_assert_held(&p->pi_lock);
1174 queued = task_on_rq_queued(p);
1175 running = task_current(rq, p);
1179 * Because __kthread_bind() calls this on blocked tasks without
1182 lockdep_assert_held(&rq->lock);
1183 dequeue_task(rq, p, 0);
1186 put_prev_task(rq, p);
1188 p->sched_class->set_cpus_allowed(p, new_mask);
1191 p->sched_class->set_curr_task(rq);
1193 enqueue_task(rq, p, 0);
1197 * Change a given task's CPU affinity. Migrate the thread to a
1198 * proper CPU and schedule it away if the CPU it's executing on
1199 * is removed from the allowed bitmask.
1201 * NOTE: the caller must have a valid reference to the task, the
1202 * task must not exit() & deallocate itself prematurely. The
1203 * call is not atomic; no spinlocks may be held.
1205 static int __set_cpus_allowed_ptr(struct task_struct *p,
1206 const struct cpumask *new_mask, bool check)
1208 unsigned long flags;
1210 unsigned int dest_cpu;
1213 rq = task_rq_lock(p, &flags);
1216 * Must re-check here, to close a race against __kthread_bind(),
1217 * sched_setaffinity() is not guaranteed to observe the flag.
1219 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1224 if (cpumask_equal(&p->cpus_allowed, new_mask))
1227 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1232 do_set_cpus_allowed(p, new_mask);
1234 /* Can the task run on the task's current CPU? If so, we're done */
1235 if (cpumask_test_cpu(task_cpu(p), new_mask))
1238 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1239 if (task_running(rq, p) || p->state == TASK_WAKING) {
1240 struct migration_arg arg = { p, dest_cpu };
1241 /* Need help from migration thread: drop lock and wait. */
1242 task_rq_unlock(rq, p, &flags);
1243 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1244 tlb_migrate_finish(p->mm);
1246 } else if (task_on_rq_queued(p)) {
1248 * OK, since we're going to drop the lock immediately
1249 * afterwards anyway.
1251 lockdep_unpin_lock(&rq->lock);
1252 rq = move_queued_task(rq, p, dest_cpu);
1253 lockdep_pin_lock(&rq->lock);
1256 task_rq_unlock(rq, p, &flags);
1261 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1263 return __set_cpus_allowed_ptr(p, new_mask, false);
1265 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1267 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1269 #ifdef CONFIG_SCHED_DEBUG
1271 * We should never call set_task_cpu() on a blocked task,
1272 * ttwu() will sort out the placement.
1274 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1277 #ifdef CONFIG_LOCKDEP
1279 * The caller should hold either p->pi_lock or rq->lock, when changing
1280 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1282 * sched_move_task() holds both and thus holding either pins the cgroup,
1285 * Furthermore, all task_rq users should acquire both locks, see
1288 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1289 lockdep_is_held(&task_rq(p)->lock)));
1293 trace_sched_migrate_task(p, new_cpu);
1295 if (task_cpu(p) != new_cpu) {
1296 if (p->sched_class->migrate_task_rq)
1297 p->sched_class->migrate_task_rq(p, new_cpu);
1298 p->se.nr_migrations++;
1299 perf_event_task_migrate(p);
1302 __set_task_cpu(p, new_cpu);
1305 static void __migrate_swap_task(struct task_struct *p, int cpu)
1307 if (task_on_rq_queued(p)) {
1308 struct rq *src_rq, *dst_rq;
1310 src_rq = task_rq(p);
1311 dst_rq = cpu_rq(cpu);
1313 deactivate_task(src_rq, p, 0);
1314 set_task_cpu(p, cpu);
1315 activate_task(dst_rq, p, 0);
1316 check_preempt_curr(dst_rq, p, 0);
1319 * Task isn't running anymore; make it appear like we migrated
1320 * it before it went to sleep. This means on wakeup we make the
1321 * previous cpu our targer instead of where it really is.
1327 struct migration_swap_arg {
1328 struct task_struct *src_task, *dst_task;
1329 int src_cpu, dst_cpu;
1332 static int migrate_swap_stop(void *data)
1334 struct migration_swap_arg *arg = data;
1335 struct rq *src_rq, *dst_rq;
1338 src_rq = cpu_rq(arg->src_cpu);
1339 dst_rq = cpu_rq(arg->dst_cpu);
1341 double_raw_lock(&arg->src_task->pi_lock,
1342 &arg->dst_task->pi_lock);
1343 double_rq_lock(src_rq, dst_rq);
1344 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1347 if (task_cpu(arg->src_task) != arg->src_cpu)
1350 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1353 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1356 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1357 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1362 double_rq_unlock(src_rq, dst_rq);
1363 raw_spin_unlock(&arg->dst_task->pi_lock);
1364 raw_spin_unlock(&arg->src_task->pi_lock);
1370 * Cross migrate two tasks
1372 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1374 struct migration_swap_arg arg;
1377 arg = (struct migration_swap_arg){
1379 .src_cpu = task_cpu(cur),
1381 .dst_cpu = task_cpu(p),
1384 if (arg.src_cpu == arg.dst_cpu)
1388 * These three tests are all lockless; this is OK since all of them
1389 * will be re-checked with proper locks held further down the line.
1391 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1394 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1397 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1400 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1401 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1408 * wait_task_inactive - wait for a thread to unschedule.
1410 * If @match_state is nonzero, it's the @p->state value just checked and
1411 * not expected to change. If it changes, i.e. @p might have woken up,
1412 * then return zero. When we succeed in waiting for @p to be off its CPU,
1413 * we return a positive number (its total switch count). If a second call
1414 * a short while later returns the same number, the caller can be sure that
1415 * @p has remained unscheduled the whole time.
1417 * The caller must ensure that the task *will* unschedule sometime soon,
1418 * else this function might spin for a *long* time. This function can't
1419 * be called with interrupts off, or it may introduce deadlock with
1420 * smp_call_function() if an IPI is sent by the same process we are
1421 * waiting to become inactive.
1423 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1425 unsigned long flags;
1426 int running, queued;
1432 * We do the initial early heuristics without holding
1433 * any task-queue locks at all. We'll only try to get
1434 * the runqueue lock when things look like they will
1440 * If the task is actively running on another CPU
1441 * still, just relax and busy-wait without holding
1444 * NOTE! Since we don't hold any locks, it's not
1445 * even sure that "rq" stays as the right runqueue!
1446 * But we don't care, since "task_running()" will
1447 * return false if the runqueue has changed and p
1448 * is actually now running somewhere else!
1450 while (task_running(rq, p)) {
1451 if (match_state && unlikely(p->state != match_state))
1457 * Ok, time to look more closely! We need the rq
1458 * lock now, to be *sure*. If we're wrong, we'll
1459 * just go back and repeat.
1461 rq = task_rq_lock(p, &flags);
1462 trace_sched_wait_task(p);
1463 running = task_running(rq, p);
1464 queued = task_on_rq_queued(p);
1466 if (!match_state || p->state == match_state)
1467 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1468 task_rq_unlock(rq, p, &flags);
1471 * If it changed from the expected state, bail out now.
1473 if (unlikely(!ncsw))
1477 * Was it really running after all now that we
1478 * checked with the proper locks actually held?
1480 * Oops. Go back and try again..
1482 if (unlikely(running)) {
1488 * It's not enough that it's not actively running,
1489 * it must be off the runqueue _entirely_, and not
1492 * So if it was still runnable (but just not actively
1493 * running right now), it's preempted, and we should
1494 * yield - it could be a while.
1496 if (unlikely(queued)) {
1497 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1499 set_current_state(TASK_UNINTERRUPTIBLE);
1500 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1505 * Ahh, all good. It wasn't running, and it wasn't
1506 * runnable, which means that it will never become
1507 * running in the future either. We're all done!
1516 * kick_process - kick a running thread to enter/exit the kernel
1517 * @p: the to-be-kicked thread
1519 * Cause a process which is running on another CPU to enter
1520 * kernel-mode, without any delay. (to get signals handled.)
1522 * NOTE: this function doesn't have to take the runqueue lock,
1523 * because all it wants to ensure is that the remote task enters
1524 * the kernel. If the IPI races and the task has been migrated
1525 * to another CPU then no harm is done and the purpose has been
1528 void kick_process(struct task_struct *p)
1534 if ((cpu != smp_processor_id()) && task_curr(p))
1535 smp_send_reschedule(cpu);
1538 EXPORT_SYMBOL_GPL(kick_process);
1541 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1543 static int select_fallback_rq(int cpu, struct task_struct *p)
1545 int nid = cpu_to_node(cpu);
1546 const struct cpumask *nodemask = NULL;
1547 enum { cpuset, possible, fail } state = cpuset;
1551 * If the node that the cpu is on has been offlined, cpu_to_node()
1552 * will return -1. There is no cpu on the node, and we should
1553 * select the cpu on the other node.
1556 nodemask = cpumask_of_node(nid);
1558 /* Look for allowed, online CPU in same node. */
1559 for_each_cpu(dest_cpu, nodemask) {
1560 if (!cpu_online(dest_cpu))
1562 if (!cpu_active(dest_cpu))
1564 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1570 /* Any allowed, online CPU? */
1571 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1572 if (!cpu_online(dest_cpu))
1574 if (!cpu_active(dest_cpu))
1581 /* No more Mr. Nice Guy. */
1582 cpuset_cpus_allowed_fallback(p);
1587 do_set_cpus_allowed(p, cpu_possible_mask);
1598 if (state != cpuset) {
1600 * Don't tell them about moving exiting tasks or
1601 * kernel threads (both mm NULL), since they never
1604 if (p->mm && printk_ratelimit()) {
1605 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1606 task_pid_nr(p), p->comm, cpu);
1614 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1617 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1619 lockdep_assert_held(&p->pi_lock);
1621 if (p->nr_cpus_allowed > 1)
1622 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1625 * In order not to call set_task_cpu() on a blocking task we need
1626 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1629 * Since this is common to all placement strategies, this lives here.
1631 * [ this allows ->select_task() to simply return task_cpu(p) and
1632 * not worry about this generic constraint ]
1634 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1636 cpu = select_fallback_rq(task_cpu(p), p);
1641 static void update_avg(u64 *avg, u64 sample)
1643 s64 diff = sample - *avg;
1649 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1650 const struct cpumask *new_mask, bool check)
1652 return set_cpus_allowed_ptr(p, new_mask);
1655 #endif /* CONFIG_SMP */
1658 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1660 #ifdef CONFIG_SCHEDSTATS
1661 struct rq *rq = this_rq();
1664 int this_cpu = smp_processor_id();
1666 if (cpu == this_cpu) {
1667 schedstat_inc(rq, ttwu_local);
1668 schedstat_inc(p, se.statistics.nr_wakeups_local);
1670 struct sched_domain *sd;
1672 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1674 for_each_domain(this_cpu, sd) {
1675 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1676 schedstat_inc(sd, ttwu_wake_remote);
1683 if (wake_flags & WF_MIGRATED)
1684 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1686 #endif /* CONFIG_SMP */
1688 schedstat_inc(rq, ttwu_count);
1689 schedstat_inc(p, se.statistics.nr_wakeups);
1691 if (wake_flags & WF_SYNC)
1692 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1694 #endif /* CONFIG_SCHEDSTATS */
1697 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1699 activate_task(rq, p, en_flags);
1700 p->on_rq = TASK_ON_RQ_QUEUED;
1702 /* if a worker is waking up, notify workqueue */
1703 if (p->flags & PF_WQ_WORKER)
1704 wq_worker_waking_up(p, cpu_of(rq));
1708 * Mark the task runnable and perform wakeup-preemption.
1711 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1713 check_preempt_curr(rq, p, wake_flags);
1714 p->state = TASK_RUNNING;
1715 trace_sched_wakeup(p);
1718 if (p->sched_class->task_woken) {
1720 * Our task @p is fully woken up and running; so its safe to
1721 * drop the rq->lock, hereafter rq is only used for statistics.
1723 lockdep_unpin_lock(&rq->lock);
1724 p->sched_class->task_woken(rq, p);
1725 lockdep_pin_lock(&rq->lock);
1728 if (rq->idle_stamp) {
1729 u64 delta = rq_clock(rq) - rq->idle_stamp;
1730 u64 max = 2*rq->max_idle_balance_cost;
1732 update_avg(&rq->avg_idle, delta);
1734 if (rq->avg_idle > max)
1743 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1745 lockdep_assert_held(&rq->lock);
1748 if (p->sched_contributes_to_load)
1749 rq->nr_uninterruptible--;
1752 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1753 ttwu_do_wakeup(rq, p, wake_flags);
1757 * Called in case the task @p isn't fully descheduled from its runqueue,
1758 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1759 * since all we need to do is flip p->state to TASK_RUNNING, since
1760 * the task is still ->on_rq.
1762 static int ttwu_remote(struct task_struct *p, int wake_flags)
1767 rq = __task_rq_lock(p);
1768 if (task_on_rq_queued(p)) {
1769 /* check_preempt_curr() may use rq clock */
1770 update_rq_clock(rq);
1771 ttwu_do_wakeup(rq, p, wake_flags);
1774 __task_rq_unlock(rq);
1780 void sched_ttwu_pending(void)
1782 struct rq *rq = this_rq();
1783 struct llist_node *llist = llist_del_all(&rq->wake_list);
1784 struct task_struct *p;
1785 unsigned long flags;
1790 raw_spin_lock_irqsave(&rq->lock, flags);
1791 lockdep_pin_lock(&rq->lock);
1794 p = llist_entry(llist, struct task_struct, wake_entry);
1795 llist = llist_next(llist);
1796 ttwu_do_activate(rq, p, 0);
1799 lockdep_unpin_lock(&rq->lock);
1800 raw_spin_unlock_irqrestore(&rq->lock, flags);
1803 void scheduler_ipi(void)
1806 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1807 * TIF_NEED_RESCHED remotely (for the first time) will also send
1810 preempt_fold_need_resched();
1812 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1816 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1817 * traditionally all their work was done from the interrupt return
1818 * path. Now that we actually do some work, we need to make sure
1821 * Some archs already do call them, luckily irq_enter/exit nest
1824 * Arguably we should visit all archs and update all handlers,
1825 * however a fair share of IPIs are still resched only so this would
1826 * somewhat pessimize the simple resched case.
1829 sched_ttwu_pending();
1832 * Check if someone kicked us for doing the nohz idle load balance.
1834 if (unlikely(got_nohz_idle_kick())) {
1835 this_rq()->idle_balance = 1;
1836 raise_softirq_irqoff(SCHED_SOFTIRQ);
1841 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1843 struct rq *rq = cpu_rq(cpu);
1845 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1846 if (!set_nr_if_polling(rq->idle))
1847 smp_send_reschedule(cpu);
1849 trace_sched_wake_idle_without_ipi(cpu);
1853 void wake_up_if_idle(int cpu)
1855 struct rq *rq = cpu_rq(cpu);
1856 unsigned long flags;
1860 if (!is_idle_task(rcu_dereference(rq->curr)))
1863 if (set_nr_if_polling(rq->idle)) {
1864 trace_sched_wake_idle_without_ipi(cpu);
1866 raw_spin_lock_irqsave(&rq->lock, flags);
1867 if (is_idle_task(rq->curr))
1868 smp_send_reschedule(cpu);
1869 /* Else cpu is not in idle, do nothing here */
1870 raw_spin_unlock_irqrestore(&rq->lock, flags);
1877 bool cpus_share_cache(int this_cpu, int that_cpu)
1879 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1881 #endif /* CONFIG_SMP */
1883 static void ttwu_queue(struct task_struct *p, int cpu)
1885 struct rq *rq = cpu_rq(cpu);
1887 #if defined(CONFIG_SMP)
1888 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1889 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1890 ttwu_queue_remote(p, cpu);
1895 raw_spin_lock(&rq->lock);
1896 lockdep_pin_lock(&rq->lock);
1897 ttwu_do_activate(rq, p, 0);
1898 lockdep_unpin_lock(&rq->lock);
1899 raw_spin_unlock(&rq->lock);
1903 * try_to_wake_up - wake up a thread
1904 * @p: the thread to be awakened
1905 * @state: the mask of task states that can be woken
1906 * @wake_flags: wake modifier flags (WF_*)
1908 * Put it on the run-queue if it's not already there. The "current"
1909 * thread is always on the run-queue (except when the actual
1910 * re-schedule is in progress), and as such you're allowed to do
1911 * the simpler "current->state = TASK_RUNNING" to mark yourself
1912 * runnable without the overhead of this.
1914 * Return: %true if @p was woken up, %false if it was already running.
1915 * or @state didn't match @p's state.
1918 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1920 unsigned long flags;
1921 int cpu, success = 0;
1924 * If we are going to wake up a thread waiting for CONDITION we
1925 * need to ensure that CONDITION=1 done by the caller can not be
1926 * reordered with p->state check below. This pairs with mb() in
1927 * set_current_state() the waiting thread does.
1929 smp_mb__before_spinlock();
1930 raw_spin_lock_irqsave(&p->pi_lock, flags);
1931 if (!(p->state & state))
1934 trace_sched_waking(p);
1936 success = 1; /* we're going to change ->state */
1939 if (p->on_rq && ttwu_remote(p, wake_flags))
1944 * If the owning (remote) cpu is still in the middle of schedule() with
1945 * this task as prev, wait until its done referencing the task.
1950 * Pairs with the smp_wmb() in finish_lock_switch().
1954 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1955 p->state = TASK_WAKING;
1957 if (p->sched_class->task_waking)
1958 p->sched_class->task_waking(p);
1960 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1961 if (task_cpu(p) != cpu) {
1962 wake_flags |= WF_MIGRATED;
1963 set_task_cpu(p, cpu);
1965 #endif /* CONFIG_SMP */
1969 ttwu_stat(p, cpu, wake_flags);
1971 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1977 * try_to_wake_up_local - try to wake up a local task with rq lock held
1978 * @p: the thread to be awakened
1980 * Put @p on the run-queue if it's not already there. The caller must
1981 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1984 static void try_to_wake_up_local(struct task_struct *p)
1986 struct rq *rq = task_rq(p);
1988 if (WARN_ON_ONCE(rq != this_rq()) ||
1989 WARN_ON_ONCE(p == current))
1992 lockdep_assert_held(&rq->lock);
1994 if (!raw_spin_trylock(&p->pi_lock)) {
1996 * This is OK, because current is on_cpu, which avoids it being
1997 * picked for load-balance and preemption/IRQs are still
1998 * disabled avoiding further scheduler activity on it and we've
1999 * not yet picked a replacement task.
2001 lockdep_unpin_lock(&rq->lock);
2002 raw_spin_unlock(&rq->lock);
2003 raw_spin_lock(&p->pi_lock);
2004 raw_spin_lock(&rq->lock);
2005 lockdep_pin_lock(&rq->lock);
2008 if (!(p->state & TASK_NORMAL))
2011 trace_sched_waking(p);
2013 if (!task_on_rq_queued(p))
2014 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2016 ttwu_do_wakeup(rq, p, 0);
2017 ttwu_stat(p, smp_processor_id(), 0);
2019 raw_spin_unlock(&p->pi_lock);
2023 * wake_up_process - Wake up a specific process
2024 * @p: The process to be woken up.
2026 * Attempt to wake up the nominated process and move it to the set of runnable
2029 * Return: 1 if the process was woken up, 0 if it was already running.
2031 * It may be assumed that this function implies a write memory barrier before
2032 * changing the task state if and only if any tasks are woken up.
2034 int wake_up_process(struct task_struct *p)
2036 WARN_ON(task_is_stopped_or_traced(p));
2037 return try_to_wake_up(p, TASK_NORMAL, 0);
2039 EXPORT_SYMBOL(wake_up_process);
2041 int wake_up_state(struct task_struct *p, unsigned int state)
2043 return try_to_wake_up(p, state, 0);
2047 * This function clears the sched_dl_entity static params.
2049 void __dl_clear_params(struct task_struct *p)
2051 struct sched_dl_entity *dl_se = &p->dl;
2053 dl_se->dl_runtime = 0;
2054 dl_se->dl_deadline = 0;
2055 dl_se->dl_period = 0;
2059 dl_se->dl_throttled = 0;
2061 dl_se->dl_yielded = 0;
2065 * Perform scheduler related setup for a newly forked process p.
2066 * p is forked by current.
2068 * __sched_fork() is basic setup used by init_idle() too:
2070 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2075 p->se.exec_start = 0;
2076 p->se.sum_exec_runtime = 0;
2077 p->se.prev_sum_exec_runtime = 0;
2078 p->se.nr_migrations = 0;
2080 INIT_LIST_HEAD(&p->se.group_node);
2082 #ifdef CONFIG_SCHEDSTATS
2083 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2086 RB_CLEAR_NODE(&p->dl.rb_node);
2087 init_dl_task_timer(&p->dl);
2088 __dl_clear_params(p);
2090 INIT_LIST_HEAD(&p->rt.run_list);
2092 #ifdef CONFIG_PREEMPT_NOTIFIERS
2093 INIT_HLIST_HEAD(&p->preempt_notifiers);
2096 #ifdef CONFIG_NUMA_BALANCING
2097 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2098 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2099 p->mm->numa_scan_seq = 0;
2102 if (clone_flags & CLONE_VM)
2103 p->numa_preferred_nid = current->numa_preferred_nid;
2105 p->numa_preferred_nid = -1;
2107 p->node_stamp = 0ULL;
2108 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2109 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2110 p->numa_work.next = &p->numa_work;
2111 p->numa_faults = NULL;
2112 p->last_task_numa_placement = 0;
2113 p->last_sum_exec_runtime = 0;
2115 p->numa_group = NULL;
2116 #endif /* CONFIG_NUMA_BALANCING */
2119 #ifdef CONFIG_NUMA_BALANCING
2120 #ifdef CONFIG_SCHED_DEBUG
2121 void set_numabalancing_state(bool enabled)
2124 sched_feat_set("NUMA");
2126 sched_feat_set("NO_NUMA");
2129 __read_mostly bool numabalancing_enabled;
2131 void set_numabalancing_state(bool enabled)
2133 numabalancing_enabled = enabled;
2135 #endif /* CONFIG_SCHED_DEBUG */
2137 #ifdef CONFIG_PROC_SYSCTL
2138 int sysctl_numa_balancing(struct ctl_table *table, int write,
2139 void __user *buffer, size_t *lenp, loff_t *ppos)
2143 int state = numabalancing_enabled;
2145 if (write && !capable(CAP_SYS_ADMIN))
2150 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2154 set_numabalancing_state(state);
2161 * fork()/clone()-time setup:
2163 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2165 unsigned long flags;
2166 int cpu = get_cpu();
2168 __sched_fork(clone_flags, p);
2170 * We mark the process as running here. This guarantees that
2171 * nobody will actually run it, and a signal or other external
2172 * event cannot wake it up and insert it on the runqueue either.
2174 p->state = TASK_RUNNING;
2177 * Make sure we do not leak PI boosting priority to the child.
2179 p->prio = current->normal_prio;
2182 * Revert to default priority/policy on fork if requested.
2184 if (unlikely(p->sched_reset_on_fork)) {
2185 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2186 p->policy = SCHED_NORMAL;
2187 p->static_prio = NICE_TO_PRIO(0);
2189 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2190 p->static_prio = NICE_TO_PRIO(0);
2192 p->prio = p->normal_prio = __normal_prio(p);
2196 * We don't need the reset flag anymore after the fork. It has
2197 * fulfilled its duty:
2199 p->sched_reset_on_fork = 0;
2202 if (dl_prio(p->prio)) {
2205 } else if (rt_prio(p->prio)) {
2206 p->sched_class = &rt_sched_class;
2208 p->sched_class = &fair_sched_class;
2211 if (p->sched_class->task_fork)
2212 p->sched_class->task_fork(p);
2215 * The child is not yet in the pid-hash so no cgroup attach races,
2216 * and the cgroup is pinned to this child due to cgroup_fork()
2217 * is ran before sched_fork().
2219 * Silence PROVE_RCU.
2221 raw_spin_lock_irqsave(&p->pi_lock, flags);
2222 set_task_cpu(p, cpu);
2223 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2225 #ifdef CONFIG_SCHED_INFO
2226 if (likely(sched_info_on()))
2227 memset(&p->sched_info, 0, sizeof(p->sched_info));
2229 #if defined(CONFIG_SMP)
2232 init_task_preempt_count(p);
2234 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2235 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2242 unsigned long to_ratio(u64 period, u64 runtime)
2244 if (runtime == RUNTIME_INF)
2248 * Doing this here saves a lot of checks in all
2249 * the calling paths, and returning zero seems
2250 * safe for them anyway.
2255 return div64_u64(runtime << 20, period);
2259 inline struct dl_bw *dl_bw_of(int i)
2261 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2262 "sched RCU must be held");
2263 return &cpu_rq(i)->rd->dl_bw;
2266 static inline int dl_bw_cpus(int i)
2268 struct root_domain *rd = cpu_rq(i)->rd;
2271 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2272 "sched RCU must be held");
2273 for_each_cpu_and(i, rd->span, cpu_active_mask)
2279 inline struct dl_bw *dl_bw_of(int i)
2281 return &cpu_rq(i)->dl.dl_bw;
2284 static inline int dl_bw_cpus(int i)
2291 * We must be sure that accepting a new task (or allowing changing the
2292 * parameters of an existing one) is consistent with the bandwidth
2293 * constraints. If yes, this function also accordingly updates the currently
2294 * allocated bandwidth to reflect the new situation.
2296 * This function is called while holding p's rq->lock.
2298 * XXX we should delay bw change until the task's 0-lag point, see
2301 static int dl_overflow(struct task_struct *p, int policy,
2302 const struct sched_attr *attr)
2305 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2306 u64 period = attr->sched_period ?: attr->sched_deadline;
2307 u64 runtime = attr->sched_runtime;
2308 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2311 if (new_bw == p->dl.dl_bw)
2315 * Either if a task, enters, leave, or stays -deadline but changes
2316 * its parameters, we may need to update accordingly the total
2317 * allocated bandwidth of the container.
2319 raw_spin_lock(&dl_b->lock);
2320 cpus = dl_bw_cpus(task_cpu(p));
2321 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2322 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2323 __dl_add(dl_b, new_bw);
2325 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2326 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2327 __dl_clear(dl_b, p->dl.dl_bw);
2328 __dl_add(dl_b, new_bw);
2330 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2331 __dl_clear(dl_b, p->dl.dl_bw);
2334 raw_spin_unlock(&dl_b->lock);
2339 extern void init_dl_bw(struct dl_bw *dl_b);
2342 * wake_up_new_task - wake up a newly created task for the first time.
2344 * This function will do some initial scheduler statistics housekeeping
2345 * that must be done for every newly created context, then puts the task
2346 * on the runqueue and wakes it.
2348 void wake_up_new_task(struct task_struct *p)
2350 unsigned long flags;
2353 raw_spin_lock_irqsave(&p->pi_lock, flags);
2356 * Fork balancing, do it here and not earlier because:
2357 * - cpus_allowed can change in the fork path
2358 * - any previously selected cpu might disappear through hotplug
2360 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2363 /* Initialize new task's runnable average */
2364 init_entity_runnable_average(&p->se);
2365 rq = __task_rq_lock(p);
2366 activate_task(rq, p, 0);
2367 p->on_rq = TASK_ON_RQ_QUEUED;
2368 trace_sched_wakeup_new(p);
2369 check_preempt_curr(rq, p, WF_FORK);
2371 if (p->sched_class->task_woken)
2372 p->sched_class->task_woken(rq, p);
2374 task_rq_unlock(rq, p, &flags);
2377 #ifdef CONFIG_PREEMPT_NOTIFIERS
2379 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2381 void preempt_notifier_inc(void)
2383 static_key_slow_inc(&preempt_notifier_key);
2385 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2387 void preempt_notifier_dec(void)
2389 static_key_slow_dec(&preempt_notifier_key);
2391 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2394 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2395 * @notifier: notifier struct to register
2397 void preempt_notifier_register(struct preempt_notifier *notifier)
2399 if (!static_key_false(&preempt_notifier_key))
2400 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2402 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2404 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2407 * preempt_notifier_unregister - no longer interested in preemption notifications
2408 * @notifier: notifier struct to unregister
2410 * This is *not* safe to call from within a preemption notifier.
2412 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2414 hlist_del(¬ifier->link);
2416 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2418 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2420 struct preempt_notifier *notifier;
2422 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2423 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2426 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2428 if (static_key_false(&preempt_notifier_key))
2429 __fire_sched_in_preempt_notifiers(curr);
2433 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2434 struct task_struct *next)
2436 struct preempt_notifier *notifier;
2438 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2439 notifier->ops->sched_out(notifier, next);
2442 static __always_inline void
2443 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2444 struct task_struct *next)
2446 if (static_key_false(&preempt_notifier_key))
2447 __fire_sched_out_preempt_notifiers(curr, next);
2450 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2452 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2457 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2458 struct task_struct *next)
2462 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2465 * prepare_task_switch - prepare to switch tasks
2466 * @rq: the runqueue preparing to switch
2467 * @prev: the current task that is being switched out
2468 * @next: the task we are going to switch to.
2470 * This is called with the rq lock held and interrupts off. It must
2471 * be paired with a subsequent finish_task_switch after the context
2474 * prepare_task_switch sets up locking and calls architecture specific
2478 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2479 struct task_struct *next)
2481 trace_sched_switch(prev, next);
2482 sched_info_switch(rq, prev, next);
2483 perf_event_task_sched_out(prev, next);
2484 fire_sched_out_preempt_notifiers(prev, next);
2485 prepare_lock_switch(rq, next);
2486 prepare_arch_switch(next);
2490 * finish_task_switch - clean up after a task-switch
2491 * @prev: the thread we just switched away from.
2493 * finish_task_switch must be called after the context switch, paired
2494 * with a prepare_task_switch call before the context switch.
2495 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2496 * and do any other architecture-specific cleanup actions.
2498 * Note that we may have delayed dropping an mm in context_switch(). If
2499 * so, we finish that here outside of the runqueue lock. (Doing it
2500 * with the lock held can cause deadlocks; see schedule() for
2503 * The context switch have flipped the stack from under us and restored the
2504 * local variables which were saved when this task called schedule() in the
2505 * past. prev == current is still correct but we need to recalculate this_rq
2506 * because prev may have moved to another CPU.
2508 static struct rq *finish_task_switch(struct task_struct *prev)
2509 __releases(rq->lock)
2511 struct rq *rq = this_rq();
2512 struct mm_struct *mm = rq->prev_mm;
2518 * A task struct has one reference for the use as "current".
2519 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2520 * schedule one last time. The schedule call will never return, and
2521 * the scheduled task must drop that reference.
2522 * The test for TASK_DEAD must occur while the runqueue locks are
2523 * still held, otherwise prev could be scheduled on another cpu, die
2524 * there before we look at prev->state, and then the reference would
2526 * Manfred Spraul <manfred@colorfullife.com>
2528 prev_state = prev->state;
2529 vtime_task_switch(prev);
2530 perf_event_task_sched_in(prev, current);
2531 finish_lock_switch(rq, prev);
2532 finish_arch_post_lock_switch();
2534 fire_sched_in_preempt_notifiers(current);
2537 if (unlikely(prev_state == TASK_DEAD)) {
2538 if (prev->sched_class->task_dead)
2539 prev->sched_class->task_dead(prev);
2542 * Remove function-return probe instances associated with this
2543 * task and put them back on the free list.
2545 kprobe_flush_task(prev);
2546 put_task_struct(prev);
2549 tick_nohz_task_switch();
2555 /* rq->lock is NOT held, but preemption is disabled */
2556 static void __balance_callback(struct rq *rq)
2558 struct callback_head *head, *next;
2559 void (*func)(struct rq *rq);
2560 unsigned long flags;
2562 raw_spin_lock_irqsave(&rq->lock, flags);
2563 head = rq->balance_callback;
2564 rq->balance_callback = NULL;
2566 func = (void (*)(struct rq *))head->func;
2573 raw_spin_unlock_irqrestore(&rq->lock, flags);
2576 static inline void balance_callback(struct rq *rq)
2578 if (unlikely(rq->balance_callback))
2579 __balance_callback(rq);
2584 static inline void balance_callback(struct rq *rq)
2591 * schedule_tail - first thing a freshly forked thread must call.
2592 * @prev: the thread we just switched away from.
2594 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2595 __releases(rq->lock)
2599 /* finish_task_switch() drops rq->lock and enables preemtion */
2601 rq = finish_task_switch(prev);
2602 balance_callback(rq);
2605 if (current->set_child_tid)
2606 put_user(task_pid_vnr(current), current->set_child_tid);
2610 * context_switch - switch to the new MM and the new thread's register state.
2612 static inline struct rq *
2613 context_switch(struct rq *rq, struct task_struct *prev,
2614 struct task_struct *next)
2616 struct mm_struct *mm, *oldmm;
2618 prepare_task_switch(rq, prev, next);
2621 oldmm = prev->active_mm;
2623 * For paravirt, this is coupled with an exit in switch_to to
2624 * combine the page table reload and the switch backend into
2627 arch_start_context_switch(prev);
2630 next->active_mm = oldmm;
2631 atomic_inc(&oldmm->mm_count);
2632 enter_lazy_tlb(oldmm, next);
2634 switch_mm(oldmm, mm, next);
2637 prev->active_mm = NULL;
2638 rq->prev_mm = oldmm;
2641 * Since the runqueue lock will be released by the next
2642 * task (which is an invalid locking op but in the case
2643 * of the scheduler it's an obvious special-case), so we
2644 * do an early lockdep release here:
2646 lockdep_unpin_lock(&rq->lock);
2647 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2649 /* Here we just switch the register state and the stack. */
2650 switch_to(prev, next, prev);
2653 return finish_task_switch(prev);
2657 * nr_running and nr_context_switches:
2659 * externally visible scheduler statistics: current number of runnable
2660 * threads, total number of context switches performed since bootup.
2662 unsigned long nr_running(void)
2664 unsigned long i, sum = 0;
2666 for_each_online_cpu(i)
2667 sum += cpu_rq(i)->nr_running;
2673 * Check if only the current task is running on the cpu.
2675 bool single_task_running(void)
2677 if (cpu_rq(smp_processor_id())->nr_running == 1)
2682 EXPORT_SYMBOL(single_task_running);
2684 unsigned long long nr_context_switches(void)
2687 unsigned long long sum = 0;
2689 for_each_possible_cpu(i)
2690 sum += cpu_rq(i)->nr_switches;
2695 unsigned long nr_iowait(void)
2697 unsigned long i, sum = 0;
2699 for_each_possible_cpu(i)
2700 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2705 unsigned long nr_iowait_cpu(int cpu)
2707 struct rq *this = cpu_rq(cpu);
2708 return atomic_read(&this->nr_iowait);
2711 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2713 struct rq *rq = this_rq();
2714 *nr_waiters = atomic_read(&rq->nr_iowait);
2715 *load = rq->load.weight;
2721 * sched_exec - execve() is a valuable balancing opportunity, because at
2722 * this point the task has the smallest effective memory and cache footprint.
2724 void sched_exec(void)
2726 struct task_struct *p = current;
2727 unsigned long flags;
2730 raw_spin_lock_irqsave(&p->pi_lock, flags);
2731 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2732 if (dest_cpu == smp_processor_id())
2735 if (likely(cpu_active(dest_cpu))) {
2736 struct migration_arg arg = { p, dest_cpu };
2738 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2739 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2743 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2748 DEFINE_PER_CPU(struct kernel_stat, kstat);
2749 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2751 EXPORT_PER_CPU_SYMBOL(kstat);
2752 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2755 * Return accounted runtime for the task.
2756 * In case the task is currently running, return the runtime plus current's
2757 * pending runtime that have not been accounted yet.
2759 unsigned long long task_sched_runtime(struct task_struct *p)
2761 unsigned long flags;
2765 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2767 * 64-bit doesn't need locks to atomically read a 64bit value.
2768 * So we have a optimization chance when the task's delta_exec is 0.
2769 * Reading ->on_cpu is racy, but this is ok.
2771 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2772 * If we race with it entering cpu, unaccounted time is 0. This is
2773 * indistinguishable from the read occurring a few cycles earlier.
2774 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2775 * been accounted, so we're correct here as well.
2777 if (!p->on_cpu || !task_on_rq_queued(p))
2778 return p->se.sum_exec_runtime;
2781 rq = task_rq_lock(p, &flags);
2783 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2784 * project cycles that may never be accounted to this
2785 * thread, breaking clock_gettime().
2787 if (task_current(rq, p) && task_on_rq_queued(p)) {
2788 update_rq_clock(rq);
2789 p->sched_class->update_curr(rq);
2791 ns = p->se.sum_exec_runtime;
2792 task_rq_unlock(rq, p, &flags);
2798 * This function gets called by the timer code, with HZ frequency.
2799 * We call it with interrupts disabled.
2801 void scheduler_tick(void)
2803 int cpu = smp_processor_id();
2804 struct rq *rq = cpu_rq(cpu);
2805 struct task_struct *curr = rq->curr;
2809 raw_spin_lock(&rq->lock);
2810 update_rq_clock(rq);
2811 curr->sched_class->task_tick(rq, curr, 0);
2812 update_cpu_load_active(rq);
2813 calc_global_load_tick(rq);
2814 raw_spin_unlock(&rq->lock);
2816 perf_event_task_tick();
2819 rq->idle_balance = idle_cpu(cpu);
2820 trigger_load_balance(rq);
2822 rq_last_tick_reset(rq);
2825 #ifdef CONFIG_NO_HZ_FULL
2827 * scheduler_tick_max_deferment
2829 * Keep at least one tick per second when a single
2830 * active task is running because the scheduler doesn't
2831 * yet completely support full dynticks environment.
2833 * This makes sure that uptime, CFS vruntime, load
2834 * balancing, etc... continue to move forward, even
2835 * with a very low granularity.
2837 * Return: Maximum deferment in nanoseconds.
2839 u64 scheduler_tick_max_deferment(void)
2841 struct rq *rq = this_rq();
2842 unsigned long next, now = READ_ONCE(jiffies);
2844 next = rq->last_sched_tick + HZ;
2846 if (time_before_eq(next, now))
2849 return jiffies_to_nsecs(next - now);
2853 notrace unsigned long get_parent_ip(unsigned long addr)
2855 if (in_lock_functions(addr)) {
2856 addr = CALLER_ADDR2;
2857 if (in_lock_functions(addr))
2858 addr = CALLER_ADDR3;
2863 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2864 defined(CONFIG_PREEMPT_TRACER))
2866 void preempt_count_add(int val)
2868 #ifdef CONFIG_DEBUG_PREEMPT
2872 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2875 __preempt_count_add(val);
2876 #ifdef CONFIG_DEBUG_PREEMPT
2878 * Spinlock count overflowing soon?
2880 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2883 if (preempt_count() == val) {
2884 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2885 #ifdef CONFIG_DEBUG_PREEMPT
2886 current->preempt_disable_ip = ip;
2888 trace_preempt_off(CALLER_ADDR0, ip);
2891 EXPORT_SYMBOL(preempt_count_add);
2892 NOKPROBE_SYMBOL(preempt_count_add);
2894 void preempt_count_sub(int val)
2896 #ifdef CONFIG_DEBUG_PREEMPT
2900 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2903 * Is the spinlock portion underflowing?
2905 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2906 !(preempt_count() & PREEMPT_MASK)))
2910 if (preempt_count() == val)
2911 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2912 __preempt_count_sub(val);
2914 EXPORT_SYMBOL(preempt_count_sub);
2915 NOKPROBE_SYMBOL(preempt_count_sub);
2920 * Print scheduling while atomic bug:
2922 static noinline void __schedule_bug(struct task_struct *prev)
2924 if (oops_in_progress)
2927 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2928 prev->comm, prev->pid, preempt_count());
2930 debug_show_held_locks(prev);
2932 if (irqs_disabled())
2933 print_irqtrace_events(prev);
2934 #ifdef CONFIG_DEBUG_PREEMPT
2935 if (in_atomic_preempt_off()) {
2936 pr_err("Preemption disabled at:");
2937 print_ip_sym(current->preempt_disable_ip);
2942 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2946 * Various schedule()-time debugging checks and statistics:
2948 static inline void schedule_debug(struct task_struct *prev)
2950 #ifdef CONFIG_SCHED_STACK_END_CHECK
2951 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2954 * Test if we are atomic. Since do_exit() needs to call into
2955 * schedule() atomically, we ignore that path. Otherwise whine
2956 * if we are scheduling when we should not.
2958 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2959 __schedule_bug(prev);
2962 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2964 schedstat_inc(this_rq(), sched_count);
2968 * Pick up the highest-prio task:
2970 static inline struct task_struct *
2971 pick_next_task(struct rq *rq, struct task_struct *prev)
2973 const struct sched_class *class = &fair_sched_class;
2974 struct task_struct *p;
2977 * Optimization: we know that if all tasks are in
2978 * the fair class we can call that function directly:
2980 if (likely(prev->sched_class == class &&
2981 rq->nr_running == rq->cfs.h_nr_running)) {
2982 p = fair_sched_class.pick_next_task(rq, prev);
2983 if (unlikely(p == RETRY_TASK))
2986 /* assumes fair_sched_class->next == idle_sched_class */
2988 p = idle_sched_class.pick_next_task(rq, prev);
2994 for_each_class(class) {
2995 p = class->pick_next_task(rq, prev);
2997 if (unlikely(p == RETRY_TASK))
3003 BUG(); /* the idle class will always have a runnable task */
3007 * __schedule() is the main scheduler function.
3009 * The main means of driving the scheduler and thus entering this function are:
3011 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3013 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3014 * paths. For example, see arch/x86/entry_64.S.
3016 * To drive preemption between tasks, the scheduler sets the flag in timer
3017 * interrupt handler scheduler_tick().
3019 * 3. Wakeups don't really cause entry into schedule(). They add a
3020 * task to the run-queue and that's it.
3022 * Now, if the new task added to the run-queue preempts the current
3023 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3024 * called on the nearest possible occasion:
3026 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3028 * - in syscall or exception context, at the next outmost
3029 * preempt_enable(). (this might be as soon as the wake_up()'s
3032 * - in IRQ context, return from interrupt-handler to
3033 * preemptible context
3035 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3038 * - cond_resched() call
3039 * - explicit schedule() call
3040 * - return from syscall or exception to user-space
3041 * - return from interrupt-handler to user-space
3043 * WARNING: must be called with preemption disabled!
3045 static void __sched __schedule(void)
3047 struct task_struct *prev, *next;
3048 unsigned long *switch_count;
3052 cpu = smp_processor_id();
3054 rcu_note_context_switch();
3057 schedule_debug(prev);
3059 if (sched_feat(HRTICK))
3063 * Make sure that signal_pending_state()->signal_pending() below
3064 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3065 * done by the caller to avoid the race with signal_wake_up().
3067 smp_mb__before_spinlock();
3068 raw_spin_lock_irq(&rq->lock);
3069 lockdep_pin_lock(&rq->lock);
3071 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3073 switch_count = &prev->nivcsw;
3074 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3075 if (unlikely(signal_pending_state(prev->state, prev))) {
3076 prev->state = TASK_RUNNING;
3078 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3082 * If a worker went to sleep, notify and ask workqueue
3083 * whether it wants to wake up a task to maintain
3086 if (prev->flags & PF_WQ_WORKER) {
3087 struct task_struct *to_wakeup;
3089 to_wakeup = wq_worker_sleeping(prev, cpu);
3091 try_to_wake_up_local(to_wakeup);
3094 switch_count = &prev->nvcsw;
3097 if (task_on_rq_queued(prev))
3098 update_rq_clock(rq);
3100 next = pick_next_task(rq, prev);
3101 clear_tsk_need_resched(prev);
3102 clear_preempt_need_resched();
3103 rq->clock_skip_update = 0;
3105 if (likely(prev != next)) {
3110 rq = context_switch(rq, prev, next); /* unlocks the rq */
3113 lockdep_unpin_lock(&rq->lock);
3114 raw_spin_unlock_irq(&rq->lock);
3117 balance_callback(rq);
3120 static inline void sched_submit_work(struct task_struct *tsk)
3122 if (!tsk->state || tsk_is_pi_blocked(tsk))
3125 * If we are going to sleep and we have plugged IO queued,
3126 * make sure to submit it to avoid deadlocks.
3128 if (blk_needs_flush_plug(tsk))
3129 blk_schedule_flush_plug(tsk);
3132 asmlinkage __visible void __sched schedule(void)
3134 struct task_struct *tsk = current;
3136 sched_submit_work(tsk);
3140 sched_preempt_enable_no_resched();
3141 } while (need_resched());
3143 EXPORT_SYMBOL(schedule);
3145 #ifdef CONFIG_CONTEXT_TRACKING
3146 asmlinkage __visible void __sched schedule_user(void)
3149 * If we come here after a random call to set_need_resched(),
3150 * or we have been woken up remotely but the IPI has not yet arrived,
3151 * we haven't yet exited the RCU idle mode. Do it here manually until
3152 * we find a better solution.
3154 * NB: There are buggy callers of this function. Ideally we
3155 * should warn if prev_state != CONTEXT_USER, but that will trigger
3156 * too frequently to make sense yet.
3158 enum ctx_state prev_state = exception_enter();
3160 exception_exit(prev_state);
3165 * schedule_preempt_disabled - called with preemption disabled
3167 * Returns with preemption disabled. Note: preempt_count must be 1
3169 void __sched schedule_preempt_disabled(void)
3171 sched_preempt_enable_no_resched();
3176 static void __sched notrace preempt_schedule_common(void)
3179 preempt_active_enter();
3181 preempt_active_exit();
3184 * Check again in case we missed a preemption opportunity
3185 * between schedule and now.
3187 } while (need_resched());
3190 #ifdef CONFIG_PREEMPT
3192 * this is the entry point to schedule() from in-kernel preemption
3193 * off of preempt_enable. Kernel preemptions off return from interrupt
3194 * occur there and call schedule directly.
3196 asmlinkage __visible void __sched notrace preempt_schedule(void)
3199 * If there is a non-zero preempt_count or interrupts are disabled,
3200 * we do not want to preempt the current task. Just return..
3202 if (likely(!preemptible()))
3205 preempt_schedule_common();
3207 NOKPROBE_SYMBOL(preempt_schedule);
3208 EXPORT_SYMBOL(preempt_schedule);
3211 * preempt_schedule_notrace - preempt_schedule called by tracing
3213 * The tracing infrastructure uses preempt_enable_notrace to prevent
3214 * recursion and tracing preempt enabling caused by the tracing
3215 * infrastructure itself. But as tracing can happen in areas coming
3216 * from userspace or just about to enter userspace, a preempt enable
3217 * can occur before user_exit() is called. This will cause the scheduler
3218 * to be called when the system is still in usermode.
3220 * To prevent this, the preempt_enable_notrace will use this function
3221 * instead of preempt_schedule() to exit user context if needed before
3222 * calling the scheduler.
3224 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3226 enum ctx_state prev_ctx;
3228 if (likely(!preemptible()))
3233 * Use raw __prempt_count() ops that don't call function.
3234 * We can't call functions before disabling preemption which
3235 * disarm preemption tracing recursions.
3237 __preempt_count_add(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
3240 * Needs preempt disabled in case user_exit() is traced
3241 * and the tracer calls preempt_enable_notrace() causing
3242 * an infinite recursion.
3244 prev_ctx = exception_enter();
3246 exception_exit(prev_ctx);
3249 __preempt_count_sub(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
3250 } while (need_resched());
3252 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3254 #endif /* CONFIG_PREEMPT */
3257 * this is the entry point to schedule() from kernel preemption
3258 * off of irq context.
3259 * Note, that this is called and return with irqs disabled. This will
3260 * protect us against recursive calling from irq.
3262 asmlinkage __visible void __sched preempt_schedule_irq(void)
3264 enum ctx_state prev_state;
3266 /* Catch callers which need to be fixed */
3267 BUG_ON(preempt_count() || !irqs_disabled());
3269 prev_state = exception_enter();
3272 preempt_active_enter();
3275 local_irq_disable();
3276 preempt_active_exit();
3277 } while (need_resched());
3279 exception_exit(prev_state);
3282 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3285 return try_to_wake_up(curr->private, mode, wake_flags);
3287 EXPORT_SYMBOL(default_wake_function);
3289 #ifdef CONFIG_RT_MUTEXES
3292 * rt_mutex_setprio - set the current priority of a task
3294 * @prio: prio value (kernel-internal form)
3296 * This function changes the 'effective' priority of a task. It does
3297 * not touch ->normal_prio like __setscheduler().
3299 * Used by the rt_mutex code to implement priority inheritance
3300 * logic. Call site only calls if the priority of the task changed.
3302 void rt_mutex_setprio(struct task_struct *p, int prio)
3304 int oldprio, queued, running, enqueue_flag = 0;
3306 const struct sched_class *prev_class;
3308 BUG_ON(prio > MAX_PRIO);
3310 rq = __task_rq_lock(p);
3313 * Idle task boosting is a nono in general. There is one
3314 * exception, when PREEMPT_RT and NOHZ is active:
3316 * The idle task calls get_next_timer_interrupt() and holds
3317 * the timer wheel base->lock on the CPU and another CPU wants
3318 * to access the timer (probably to cancel it). We can safely
3319 * ignore the boosting request, as the idle CPU runs this code
3320 * with interrupts disabled and will complete the lock
3321 * protected section without being interrupted. So there is no
3322 * real need to boost.
3324 if (unlikely(p == rq->idle)) {
3325 WARN_ON(p != rq->curr);
3326 WARN_ON(p->pi_blocked_on);
3330 trace_sched_pi_setprio(p, prio);
3332 prev_class = p->sched_class;
3333 queued = task_on_rq_queued(p);
3334 running = task_current(rq, p);
3336 dequeue_task(rq, p, 0);
3338 put_prev_task(rq, p);
3341 * Boosting condition are:
3342 * 1. -rt task is running and holds mutex A
3343 * --> -dl task blocks on mutex A
3345 * 2. -dl task is running and holds mutex A
3346 * --> -dl task blocks on mutex A and could preempt the
3349 if (dl_prio(prio)) {
3350 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3351 if (!dl_prio(p->normal_prio) ||
3352 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3353 p->dl.dl_boosted = 1;
3354 enqueue_flag = ENQUEUE_REPLENISH;
3356 p->dl.dl_boosted = 0;
3357 p->sched_class = &dl_sched_class;
3358 } else if (rt_prio(prio)) {
3359 if (dl_prio(oldprio))
3360 p->dl.dl_boosted = 0;
3362 enqueue_flag = ENQUEUE_HEAD;
3363 p->sched_class = &rt_sched_class;
3365 if (dl_prio(oldprio))
3366 p->dl.dl_boosted = 0;
3367 if (rt_prio(oldprio))
3369 p->sched_class = &fair_sched_class;
3375 p->sched_class->set_curr_task(rq);
3377 enqueue_task(rq, p, enqueue_flag);
3379 check_class_changed(rq, p, prev_class, oldprio);
3381 preempt_disable(); /* avoid rq from going away on us */
3382 __task_rq_unlock(rq);
3384 balance_callback(rq);
3389 void set_user_nice(struct task_struct *p, long nice)
3391 int old_prio, delta, queued;
3392 unsigned long flags;
3395 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3398 * We have to be careful, if called from sys_setpriority(),
3399 * the task might be in the middle of scheduling on another CPU.
3401 rq = task_rq_lock(p, &flags);
3403 * The RT priorities are set via sched_setscheduler(), but we still
3404 * allow the 'normal' nice value to be set - but as expected
3405 * it wont have any effect on scheduling until the task is
3406 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3408 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3409 p->static_prio = NICE_TO_PRIO(nice);
3412 queued = task_on_rq_queued(p);
3414 dequeue_task(rq, p, 0);
3416 p->static_prio = NICE_TO_PRIO(nice);
3419 p->prio = effective_prio(p);
3420 delta = p->prio - old_prio;
3423 enqueue_task(rq, p, 0);
3425 * If the task increased its priority or is running and
3426 * lowered its priority, then reschedule its CPU:
3428 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3432 task_rq_unlock(rq, p, &flags);
3434 EXPORT_SYMBOL(set_user_nice);
3437 * can_nice - check if a task can reduce its nice value
3441 int can_nice(const struct task_struct *p, const int nice)
3443 /* convert nice value [19,-20] to rlimit style value [1,40] */
3444 int nice_rlim = nice_to_rlimit(nice);
3446 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3447 capable(CAP_SYS_NICE));
3450 #ifdef __ARCH_WANT_SYS_NICE
3453 * sys_nice - change the priority of the current process.
3454 * @increment: priority increment
3456 * sys_setpriority is a more generic, but much slower function that
3457 * does similar things.
3459 SYSCALL_DEFINE1(nice, int, increment)
3464 * Setpriority might change our priority at the same moment.
3465 * We don't have to worry. Conceptually one call occurs first
3466 * and we have a single winner.
3468 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3469 nice = task_nice(current) + increment;
3471 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3472 if (increment < 0 && !can_nice(current, nice))
3475 retval = security_task_setnice(current, nice);
3479 set_user_nice(current, nice);
3486 * task_prio - return the priority value of a given task.
3487 * @p: the task in question.
3489 * Return: The priority value as seen by users in /proc.
3490 * RT tasks are offset by -200. Normal tasks are centered
3491 * around 0, value goes from -16 to +15.
3493 int task_prio(const struct task_struct *p)
3495 return p->prio - MAX_RT_PRIO;
3499 * idle_cpu - is a given cpu idle currently?
3500 * @cpu: the processor in question.
3502 * Return: 1 if the CPU is currently idle. 0 otherwise.
3504 int idle_cpu(int cpu)
3506 struct rq *rq = cpu_rq(cpu);
3508 if (rq->curr != rq->idle)
3515 if (!llist_empty(&rq->wake_list))
3523 * idle_task - return the idle task for a given cpu.
3524 * @cpu: the processor in question.
3526 * Return: The idle task for the cpu @cpu.
3528 struct task_struct *idle_task(int cpu)
3530 return cpu_rq(cpu)->idle;
3534 * find_process_by_pid - find a process with a matching PID value.
3535 * @pid: the pid in question.
3537 * The task of @pid, if found. %NULL otherwise.
3539 static struct task_struct *find_process_by_pid(pid_t pid)
3541 return pid ? find_task_by_vpid(pid) : current;
3545 * This function initializes the sched_dl_entity of a newly becoming
3546 * SCHED_DEADLINE task.
3548 * Only the static values are considered here, the actual runtime and the
3549 * absolute deadline will be properly calculated when the task is enqueued
3550 * for the first time with its new policy.
3553 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3555 struct sched_dl_entity *dl_se = &p->dl;
3557 dl_se->dl_runtime = attr->sched_runtime;
3558 dl_se->dl_deadline = attr->sched_deadline;
3559 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3560 dl_se->flags = attr->sched_flags;
3561 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3564 * Changing the parameters of a task is 'tricky' and we're not doing
3565 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3567 * What we SHOULD do is delay the bandwidth release until the 0-lag
3568 * point. This would include retaining the task_struct until that time
3569 * and change dl_overflow() to not immediately decrement the current
3572 * Instead we retain the current runtime/deadline and let the new
3573 * parameters take effect after the current reservation period lapses.
3574 * This is safe (albeit pessimistic) because the 0-lag point is always
3575 * before the current scheduling deadline.
3577 * We can still have temporary overloads because we do not delay the
3578 * change in bandwidth until that time; so admission control is
3579 * not on the safe side. It does however guarantee tasks will never
3580 * consume more than promised.
3585 * sched_setparam() passes in -1 for its policy, to let the functions
3586 * it calls know not to change it.
3588 #define SETPARAM_POLICY -1
3590 static void __setscheduler_params(struct task_struct *p,
3591 const struct sched_attr *attr)
3593 int policy = attr->sched_policy;
3595 if (policy == SETPARAM_POLICY)
3600 if (dl_policy(policy))
3601 __setparam_dl(p, attr);
3602 else if (fair_policy(policy))
3603 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3606 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3607 * !rt_policy. Always setting this ensures that things like
3608 * getparam()/getattr() don't report silly values for !rt tasks.
3610 p->rt_priority = attr->sched_priority;
3611 p->normal_prio = normal_prio(p);
3615 /* Actually do priority change: must hold pi & rq lock. */
3616 static void __setscheduler(struct rq *rq, struct task_struct *p,
3617 const struct sched_attr *attr, bool keep_boost)
3619 __setscheduler_params(p, attr);
3622 * Keep a potential priority boosting if called from
3623 * sched_setscheduler().
3626 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3628 p->prio = normal_prio(p);
3630 if (dl_prio(p->prio))
3631 p->sched_class = &dl_sched_class;
3632 else if (rt_prio(p->prio))
3633 p->sched_class = &rt_sched_class;
3635 p->sched_class = &fair_sched_class;
3639 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3641 struct sched_dl_entity *dl_se = &p->dl;
3643 attr->sched_priority = p->rt_priority;
3644 attr->sched_runtime = dl_se->dl_runtime;
3645 attr->sched_deadline = dl_se->dl_deadline;
3646 attr->sched_period = dl_se->dl_period;
3647 attr->sched_flags = dl_se->flags;
3651 * This function validates the new parameters of a -deadline task.
3652 * We ask for the deadline not being zero, and greater or equal
3653 * than the runtime, as well as the period of being zero or
3654 * greater than deadline. Furthermore, we have to be sure that
3655 * user parameters are above the internal resolution of 1us (we
3656 * check sched_runtime only since it is always the smaller one) and
3657 * below 2^63 ns (we have to check both sched_deadline and
3658 * sched_period, as the latter can be zero).
3661 __checkparam_dl(const struct sched_attr *attr)
3664 if (attr->sched_deadline == 0)
3668 * Since we truncate DL_SCALE bits, make sure we're at least
3671 if (attr->sched_runtime < (1ULL << DL_SCALE))
3675 * Since we use the MSB for wrap-around and sign issues, make
3676 * sure it's not set (mind that period can be equal to zero).
3678 if (attr->sched_deadline & (1ULL << 63) ||
3679 attr->sched_period & (1ULL << 63))
3682 /* runtime <= deadline <= period (if period != 0) */
3683 if ((attr->sched_period != 0 &&
3684 attr->sched_period < attr->sched_deadline) ||
3685 attr->sched_deadline < attr->sched_runtime)
3692 * check the target process has a UID that matches the current process's
3694 static bool check_same_owner(struct task_struct *p)
3696 const struct cred *cred = current_cred(), *pcred;
3700 pcred = __task_cred(p);
3701 match = (uid_eq(cred->euid, pcred->euid) ||
3702 uid_eq(cred->euid, pcred->uid));
3707 static bool dl_param_changed(struct task_struct *p,
3708 const struct sched_attr *attr)
3710 struct sched_dl_entity *dl_se = &p->dl;
3712 if (dl_se->dl_runtime != attr->sched_runtime ||
3713 dl_se->dl_deadline != attr->sched_deadline ||
3714 dl_se->dl_period != attr->sched_period ||
3715 dl_se->flags != attr->sched_flags)
3721 static int __sched_setscheduler(struct task_struct *p,
3722 const struct sched_attr *attr,
3725 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3726 MAX_RT_PRIO - 1 - attr->sched_priority;
3727 int retval, oldprio, oldpolicy = -1, queued, running;
3728 int new_effective_prio, policy = attr->sched_policy;
3729 unsigned long flags;
3730 const struct sched_class *prev_class;
3734 /* may grab non-irq protected spin_locks */
3735 BUG_ON(in_interrupt());
3737 /* double check policy once rq lock held */
3739 reset_on_fork = p->sched_reset_on_fork;
3740 policy = oldpolicy = p->policy;
3742 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3744 if (policy != SCHED_DEADLINE &&
3745 policy != SCHED_FIFO && policy != SCHED_RR &&
3746 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3747 policy != SCHED_IDLE)
3751 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3755 * Valid priorities for SCHED_FIFO and SCHED_RR are
3756 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3757 * SCHED_BATCH and SCHED_IDLE is 0.
3759 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3760 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3762 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3763 (rt_policy(policy) != (attr->sched_priority != 0)))
3767 * Allow unprivileged RT tasks to decrease priority:
3769 if (user && !capable(CAP_SYS_NICE)) {
3770 if (fair_policy(policy)) {
3771 if (attr->sched_nice < task_nice(p) &&
3772 !can_nice(p, attr->sched_nice))
3776 if (rt_policy(policy)) {
3777 unsigned long rlim_rtprio =
3778 task_rlimit(p, RLIMIT_RTPRIO);
3780 /* can't set/change the rt policy */
3781 if (policy != p->policy && !rlim_rtprio)
3784 /* can't increase priority */
3785 if (attr->sched_priority > p->rt_priority &&
3786 attr->sched_priority > rlim_rtprio)
3791 * Can't set/change SCHED_DEADLINE policy at all for now
3792 * (safest behavior); in the future we would like to allow
3793 * unprivileged DL tasks to increase their relative deadline
3794 * or reduce their runtime (both ways reducing utilization)
3796 if (dl_policy(policy))
3800 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3801 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3803 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3804 if (!can_nice(p, task_nice(p)))
3808 /* can't change other user's priorities */
3809 if (!check_same_owner(p))
3812 /* Normal users shall not reset the sched_reset_on_fork flag */
3813 if (p->sched_reset_on_fork && !reset_on_fork)
3818 retval = security_task_setscheduler(p);
3824 * make sure no PI-waiters arrive (or leave) while we are
3825 * changing the priority of the task:
3827 * To be able to change p->policy safely, the appropriate
3828 * runqueue lock must be held.
3830 rq = task_rq_lock(p, &flags);
3833 * Changing the policy of the stop threads its a very bad idea
3835 if (p == rq->stop) {
3836 task_rq_unlock(rq, p, &flags);
3841 * If not changing anything there's no need to proceed further,
3842 * but store a possible modification of reset_on_fork.
3844 if (unlikely(policy == p->policy)) {
3845 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3847 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3849 if (dl_policy(policy) && dl_param_changed(p, attr))
3852 p->sched_reset_on_fork = reset_on_fork;
3853 task_rq_unlock(rq, p, &flags);
3859 #ifdef CONFIG_RT_GROUP_SCHED
3861 * Do not allow realtime tasks into groups that have no runtime
3864 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3865 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3866 !task_group_is_autogroup(task_group(p))) {
3867 task_rq_unlock(rq, p, &flags);
3872 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3873 cpumask_t *span = rq->rd->span;
3876 * Don't allow tasks with an affinity mask smaller than
3877 * the entire root_domain to become SCHED_DEADLINE. We
3878 * will also fail if there's no bandwidth available.
3880 if (!cpumask_subset(span, &p->cpus_allowed) ||
3881 rq->rd->dl_bw.bw == 0) {
3882 task_rq_unlock(rq, p, &flags);
3889 /* recheck policy now with rq lock held */
3890 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3891 policy = oldpolicy = -1;
3892 task_rq_unlock(rq, p, &flags);
3897 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3898 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3901 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3902 task_rq_unlock(rq, p, &flags);
3906 p->sched_reset_on_fork = reset_on_fork;
3911 * Take priority boosted tasks into account. If the new
3912 * effective priority is unchanged, we just store the new
3913 * normal parameters and do not touch the scheduler class and
3914 * the runqueue. This will be done when the task deboost
3917 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3918 if (new_effective_prio == oldprio) {
3919 __setscheduler_params(p, attr);
3920 task_rq_unlock(rq, p, &flags);
3925 queued = task_on_rq_queued(p);
3926 running = task_current(rq, p);
3928 dequeue_task(rq, p, 0);
3930 put_prev_task(rq, p);
3932 prev_class = p->sched_class;
3933 __setscheduler(rq, p, attr, pi);
3936 p->sched_class->set_curr_task(rq);
3939 * We enqueue to tail when the priority of a task is
3940 * increased (user space view).
3942 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3945 check_class_changed(rq, p, prev_class, oldprio);
3946 preempt_disable(); /* avoid rq from going away on us */
3947 task_rq_unlock(rq, p, &flags);
3950 rt_mutex_adjust_pi(p);
3953 * Run balance callbacks after we've adjusted the PI chain.
3955 balance_callback(rq);
3961 static int _sched_setscheduler(struct task_struct *p, int policy,
3962 const struct sched_param *param, bool check)
3964 struct sched_attr attr = {
3965 .sched_policy = policy,
3966 .sched_priority = param->sched_priority,
3967 .sched_nice = PRIO_TO_NICE(p->static_prio),
3970 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3971 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3972 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3973 policy &= ~SCHED_RESET_ON_FORK;
3974 attr.sched_policy = policy;
3977 return __sched_setscheduler(p, &attr, check, true);
3980 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3981 * @p: the task in question.
3982 * @policy: new policy.
3983 * @param: structure containing the new RT priority.
3985 * Return: 0 on success. An error code otherwise.
3987 * NOTE that the task may be already dead.
3989 int sched_setscheduler(struct task_struct *p, int policy,
3990 const struct sched_param *param)
3992 return _sched_setscheduler(p, policy, param, true);
3994 EXPORT_SYMBOL_GPL(sched_setscheduler);
3996 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3998 return __sched_setscheduler(p, attr, true, true);
4000 EXPORT_SYMBOL_GPL(sched_setattr);
4003 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4004 * @p: the task in question.
4005 * @policy: new policy.
4006 * @param: structure containing the new RT priority.
4008 * Just like sched_setscheduler, only don't bother checking if the
4009 * current context has permission. For example, this is needed in
4010 * stop_machine(): we create temporary high priority worker threads,
4011 * but our caller might not have that capability.
4013 * Return: 0 on success. An error code otherwise.
4015 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4016 const struct sched_param *param)
4018 return _sched_setscheduler(p, policy, param, false);
4022 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4024 struct sched_param lparam;
4025 struct task_struct *p;
4028 if (!param || pid < 0)
4030 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4035 p = find_process_by_pid(pid);
4037 retval = sched_setscheduler(p, policy, &lparam);
4044 * Mimics kernel/events/core.c perf_copy_attr().
4046 static int sched_copy_attr(struct sched_attr __user *uattr,
4047 struct sched_attr *attr)
4052 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4056 * zero the full structure, so that a short copy will be nice.
4058 memset(attr, 0, sizeof(*attr));
4060 ret = get_user(size, &uattr->size);
4064 if (size > PAGE_SIZE) /* silly large */
4067 if (!size) /* abi compat */
4068 size = SCHED_ATTR_SIZE_VER0;
4070 if (size < SCHED_ATTR_SIZE_VER0)
4074 * If we're handed a bigger struct than we know of,
4075 * ensure all the unknown bits are 0 - i.e. new
4076 * user-space does not rely on any kernel feature
4077 * extensions we dont know about yet.
4079 if (size > sizeof(*attr)) {
4080 unsigned char __user *addr;
4081 unsigned char __user *end;
4084 addr = (void __user *)uattr + sizeof(*attr);
4085 end = (void __user *)uattr + size;
4087 for (; addr < end; addr++) {
4088 ret = get_user(val, addr);
4094 size = sizeof(*attr);
4097 ret = copy_from_user(attr, uattr, size);
4102 * XXX: do we want to be lenient like existing syscalls; or do we want
4103 * to be strict and return an error on out-of-bounds values?
4105 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4110 put_user(sizeof(*attr), &uattr->size);
4115 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4116 * @pid: the pid in question.
4117 * @policy: new policy.
4118 * @param: structure containing the new RT priority.
4120 * Return: 0 on success. An error code otherwise.
4122 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4123 struct sched_param __user *, param)
4125 /* negative values for policy are not valid */
4129 return do_sched_setscheduler(pid, policy, param);
4133 * sys_sched_setparam - set/change the RT priority of a thread
4134 * @pid: the pid in question.
4135 * @param: structure containing the new RT priority.
4137 * Return: 0 on success. An error code otherwise.
4139 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4141 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4145 * sys_sched_setattr - same as above, but with extended sched_attr
4146 * @pid: the pid in question.
4147 * @uattr: structure containing the extended parameters.
4148 * @flags: for future extension.
4150 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4151 unsigned int, flags)
4153 struct sched_attr attr;
4154 struct task_struct *p;
4157 if (!uattr || pid < 0 || flags)
4160 retval = sched_copy_attr(uattr, &attr);
4164 if ((int)attr.sched_policy < 0)
4169 p = find_process_by_pid(pid);
4171 retval = sched_setattr(p, &attr);
4178 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4179 * @pid: the pid in question.
4181 * Return: On success, the policy of the thread. Otherwise, a negative error
4184 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4186 struct task_struct *p;
4194 p = find_process_by_pid(pid);
4196 retval = security_task_getscheduler(p);
4199 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4206 * sys_sched_getparam - get the RT priority of a thread
4207 * @pid: the pid in question.
4208 * @param: structure containing the RT priority.
4210 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4213 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4215 struct sched_param lp = { .sched_priority = 0 };
4216 struct task_struct *p;
4219 if (!param || pid < 0)
4223 p = find_process_by_pid(pid);
4228 retval = security_task_getscheduler(p);
4232 if (task_has_rt_policy(p))
4233 lp.sched_priority = p->rt_priority;
4237 * This one might sleep, we cannot do it with a spinlock held ...
4239 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4248 static int sched_read_attr(struct sched_attr __user *uattr,
4249 struct sched_attr *attr,
4254 if (!access_ok(VERIFY_WRITE, uattr, usize))
4258 * If we're handed a smaller struct than we know of,
4259 * ensure all the unknown bits are 0 - i.e. old
4260 * user-space does not get uncomplete information.
4262 if (usize < sizeof(*attr)) {
4263 unsigned char *addr;
4266 addr = (void *)attr + usize;
4267 end = (void *)attr + sizeof(*attr);
4269 for (; addr < end; addr++) {
4277 ret = copy_to_user(uattr, attr, attr->size);
4285 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4286 * @pid: the pid in question.
4287 * @uattr: structure containing the extended parameters.
4288 * @size: sizeof(attr) for fwd/bwd comp.
4289 * @flags: for future extension.
4291 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4292 unsigned int, size, unsigned int, flags)
4294 struct sched_attr attr = {
4295 .size = sizeof(struct sched_attr),
4297 struct task_struct *p;
4300 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4301 size < SCHED_ATTR_SIZE_VER0 || flags)
4305 p = find_process_by_pid(pid);
4310 retval = security_task_getscheduler(p);
4314 attr.sched_policy = p->policy;
4315 if (p->sched_reset_on_fork)
4316 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4317 if (task_has_dl_policy(p))
4318 __getparam_dl(p, &attr);
4319 else if (task_has_rt_policy(p))
4320 attr.sched_priority = p->rt_priority;
4322 attr.sched_nice = task_nice(p);
4326 retval = sched_read_attr(uattr, &attr, size);
4334 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4336 cpumask_var_t cpus_allowed, new_mask;
4337 struct task_struct *p;
4342 p = find_process_by_pid(pid);
4348 /* Prevent p going away */
4352 if (p->flags & PF_NO_SETAFFINITY) {
4356 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4360 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4362 goto out_free_cpus_allowed;
4365 if (!check_same_owner(p)) {
4367 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4369 goto out_free_new_mask;
4374 retval = security_task_setscheduler(p);
4376 goto out_free_new_mask;
4379 cpuset_cpus_allowed(p, cpus_allowed);
4380 cpumask_and(new_mask, in_mask, cpus_allowed);
4383 * Since bandwidth control happens on root_domain basis,
4384 * if admission test is enabled, we only admit -deadline
4385 * tasks allowed to run on all the CPUs in the task's
4389 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4391 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4394 goto out_free_new_mask;
4400 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4403 cpuset_cpus_allowed(p, cpus_allowed);
4404 if (!cpumask_subset(new_mask, cpus_allowed)) {
4406 * We must have raced with a concurrent cpuset
4407 * update. Just reset the cpus_allowed to the
4408 * cpuset's cpus_allowed
4410 cpumask_copy(new_mask, cpus_allowed);
4415 free_cpumask_var(new_mask);
4416 out_free_cpus_allowed:
4417 free_cpumask_var(cpus_allowed);
4423 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4424 struct cpumask *new_mask)
4426 if (len < cpumask_size())
4427 cpumask_clear(new_mask);
4428 else if (len > cpumask_size())
4429 len = cpumask_size();
4431 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4435 * sys_sched_setaffinity - set the cpu affinity of a process
4436 * @pid: pid of the process
4437 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4438 * @user_mask_ptr: user-space pointer to the new cpu mask
4440 * Return: 0 on success. An error code otherwise.
4442 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4443 unsigned long __user *, user_mask_ptr)
4445 cpumask_var_t new_mask;
4448 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4451 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4453 retval = sched_setaffinity(pid, new_mask);
4454 free_cpumask_var(new_mask);
4458 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4460 struct task_struct *p;
4461 unsigned long flags;
4467 p = find_process_by_pid(pid);
4471 retval = security_task_getscheduler(p);
4475 raw_spin_lock_irqsave(&p->pi_lock, flags);
4476 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4477 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4486 * sys_sched_getaffinity - get the cpu affinity of a process
4487 * @pid: pid of the process
4488 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4489 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4491 * Return: 0 on success. An error code otherwise.
4493 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4494 unsigned long __user *, user_mask_ptr)
4499 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4501 if (len & (sizeof(unsigned long)-1))
4504 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4507 ret = sched_getaffinity(pid, mask);
4509 size_t retlen = min_t(size_t, len, cpumask_size());
4511 if (copy_to_user(user_mask_ptr, mask, retlen))
4516 free_cpumask_var(mask);
4522 * sys_sched_yield - yield the current processor to other threads.
4524 * This function yields the current CPU to other tasks. If there are no
4525 * other threads running on this CPU then this function will return.
4529 SYSCALL_DEFINE0(sched_yield)
4531 struct rq *rq = this_rq_lock();
4533 schedstat_inc(rq, yld_count);
4534 current->sched_class->yield_task(rq);
4537 * Since we are going to call schedule() anyway, there's
4538 * no need to preempt or enable interrupts:
4540 __release(rq->lock);
4541 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4542 do_raw_spin_unlock(&rq->lock);
4543 sched_preempt_enable_no_resched();
4550 int __sched _cond_resched(void)
4552 if (should_resched(0)) {
4553 preempt_schedule_common();
4558 EXPORT_SYMBOL(_cond_resched);
4561 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4562 * call schedule, and on return reacquire the lock.
4564 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4565 * operations here to prevent schedule() from being called twice (once via
4566 * spin_unlock(), once by hand).
4568 int __cond_resched_lock(spinlock_t *lock)
4570 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4573 lockdep_assert_held(lock);
4575 if (spin_needbreak(lock) || resched) {
4578 preempt_schedule_common();
4586 EXPORT_SYMBOL(__cond_resched_lock);
4588 int __sched __cond_resched_softirq(void)
4590 BUG_ON(!in_softirq());
4592 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4594 preempt_schedule_common();
4600 EXPORT_SYMBOL(__cond_resched_softirq);
4603 * yield - yield the current processor to other threads.
4605 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4607 * The scheduler is at all times free to pick the calling task as the most
4608 * eligible task to run, if removing the yield() call from your code breaks
4609 * it, its already broken.
4611 * Typical broken usage is:
4616 * where one assumes that yield() will let 'the other' process run that will
4617 * make event true. If the current task is a SCHED_FIFO task that will never
4618 * happen. Never use yield() as a progress guarantee!!
4620 * If you want to use yield() to wait for something, use wait_event().
4621 * If you want to use yield() to be 'nice' for others, use cond_resched().
4622 * If you still want to use yield(), do not!
4624 void __sched yield(void)
4626 set_current_state(TASK_RUNNING);
4629 EXPORT_SYMBOL(yield);
4632 * yield_to - yield the current processor to another thread in
4633 * your thread group, or accelerate that thread toward the
4634 * processor it's on.
4636 * @preempt: whether task preemption is allowed or not
4638 * It's the caller's job to ensure that the target task struct
4639 * can't go away on us before we can do any checks.
4642 * true (>0) if we indeed boosted the target task.
4643 * false (0) if we failed to boost the target.
4644 * -ESRCH if there's no task to yield to.
4646 int __sched yield_to(struct task_struct *p, bool preempt)
4648 struct task_struct *curr = current;
4649 struct rq *rq, *p_rq;
4650 unsigned long flags;
4653 local_irq_save(flags);
4659 * If we're the only runnable task on the rq and target rq also
4660 * has only one task, there's absolutely no point in yielding.
4662 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4667 double_rq_lock(rq, p_rq);
4668 if (task_rq(p) != p_rq) {
4669 double_rq_unlock(rq, p_rq);
4673 if (!curr->sched_class->yield_to_task)
4676 if (curr->sched_class != p->sched_class)
4679 if (task_running(p_rq, p) || p->state)
4682 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4684 schedstat_inc(rq, yld_count);
4686 * Make p's CPU reschedule; pick_next_entity takes care of
4689 if (preempt && rq != p_rq)
4694 double_rq_unlock(rq, p_rq);
4696 local_irq_restore(flags);
4703 EXPORT_SYMBOL_GPL(yield_to);
4706 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4707 * that process accounting knows that this is a task in IO wait state.
4709 long __sched io_schedule_timeout(long timeout)
4711 int old_iowait = current->in_iowait;
4715 current->in_iowait = 1;
4716 blk_schedule_flush_plug(current);
4718 delayacct_blkio_start();
4720 atomic_inc(&rq->nr_iowait);
4721 ret = schedule_timeout(timeout);
4722 current->in_iowait = old_iowait;
4723 atomic_dec(&rq->nr_iowait);
4724 delayacct_blkio_end();
4728 EXPORT_SYMBOL(io_schedule_timeout);
4731 * sys_sched_get_priority_max - return maximum RT priority.
4732 * @policy: scheduling class.
4734 * Return: On success, this syscall returns the maximum
4735 * rt_priority that can be used by a given scheduling class.
4736 * On failure, a negative error code is returned.
4738 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4745 ret = MAX_USER_RT_PRIO-1;
4747 case SCHED_DEADLINE:
4758 * sys_sched_get_priority_min - return minimum RT priority.
4759 * @policy: scheduling class.
4761 * Return: On success, this syscall returns the minimum
4762 * rt_priority that can be used by a given scheduling class.
4763 * On failure, a negative error code is returned.
4765 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4774 case SCHED_DEADLINE:
4784 * sys_sched_rr_get_interval - return the default timeslice of a process.
4785 * @pid: pid of the process.
4786 * @interval: userspace pointer to the timeslice value.
4788 * this syscall writes the default timeslice value of a given process
4789 * into the user-space timespec buffer. A value of '0' means infinity.
4791 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4794 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4795 struct timespec __user *, interval)
4797 struct task_struct *p;
4798 unsigned int time_slice;
4799 unsigned long flags;
4809 p = find_process_by_pid(pid);
4813 retval = security_task_getscheduler(p);
4817 rq = task_rq_lock(p, &flags);
4819 if (p->sched_class->get_rr_interval)
4820 time_slice = p->sched_class->get_rr_interval(rq, p);
4821 task_rq_unlock(rq, p, &flags);
4824 jiffies_to_timespec(time_slice, &t);
4825 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4833 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4835 void sched_show_task(struct task_struct *p)
4837 unsigned long free = 0;
4839 unsigned long state = p->state;
4842 state = __ffs(state) + 1;
4843 printk(KERN_INFO "%-15.15s %c", p->comm,
4844 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4845 #if BITS_PER_LONG == 32
4846 if (state == TASK_RUNNING)
4847 printk(KERN_CONT " running ");
4849 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4851 if (state == TASK_RUNNING)
4852 printk(KERN_CONT " running task ");
4854 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4856 #ifdef CONFIG_DEBUG_STACK_USAGE
4857 free = stack_not_used(p);
4862 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4864 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4865 task_pid_nr(p), ppid,
4866 (unsigned long)task_thread_info(p)->flags);
4868 print_worker_info(KERN_INFO, p);
4869 show_stack(p, NULL);
4872 void show_state_filter(unsigned long state_filter)
4874 struct task_struct *g, *p;
4876 #if BITS_PER_LONG == 32
4878 " task PC stack pid father\n");
4881 " task PC stack pid father\n");
4884 for_each_process_thread(g, p) {
4886 * reset the NMI-timeout, listing all files on a slow
4887 * console might take a lot of time:
4889 touch_nmi_watchdog();
4890 if (!state_filter || (p->state & state_filter))
4894 touch_all_softlockup_watchdogs();
4896 #ifdef CONFIG_SCHED_DEBUG
4897 sysrq_sched_debug_show();
4901 * Only show locks if all tasks are dumped:
4904 debug_show_all_locks();
4907 void init_idle_bootup_task(struct task_struct *idle)
4909 idle->sched_class = &idle_sched_class;
4913 * init_idle - set up an idle thread for a given CPU
4914 * @idle: task in question
4915 * @cpu: cpu the idle task belongs to
4917 * NOTE: this function does not set the idle thread's NEED_RESCHED
4918 * flag, to make booting more robust.
4920 void init_idle(struct task_struct *idle, int cpu)
4922 struct rq *rq = cpu_rq(cpu);
4923 unsigned long flags;
4925 raw_spin_lock_irqsave(&idle->pi_lock, flags);
4926 raw_spin_lock(&rq->lock);
4928 __sched_fork(0, idle);
4929 idle->state = TASK_RUNNING;
4930 idle->se.exec_start = sched_clock();
4932 do_set_cpus_allowed(idle, cpumask_of(cpu));
4934 * We're having a chicken and egg problem, even though we are
4935 * holding rq->lock, the cpu isn't yet set to this cpu so the
4936 * lockdep check in task_group() will fail.
4938 * Similar case to sched_fork(). / Alternatively we could
4939 * use task_rq_lock() here and obtain the other rq->lock.
4944 __set_task_cpu(idle, cpu);
4947 rq->curr = rq->idle = idle;
4948 idle->on_rq = TASK_ON_RQ_QUEUED;
4949 #if defined(CONFIG_SMP)
4952 raw_spin_unlock(&rq->lock);
4953 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
4955 /* Set the preempt count _outside_ the spinlocks! */
4956 init_idle_preempt_count(idle, cpu);
4959 * The idle tasks have their own, simple scheduling class:
4961 idle->sched_class = &idle_sched_class;
4962 ftrace_graph_init_idle_task(idle, cpu);
4963 vtime_init_idle(idle, cpu);
4964 #if defined(CONFIG_SMP)
4965 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4969 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4970 const struct cpumask *trial)
4972 int ret = 1, trial_cpus;
4973 struct dl_bw *cur_dl_b;
4974 unsigned long flags;
4976 if (!cpumask_weight(cur))
4979 rcu_read_lock_sched();
4980 cur_dl_b = dl_bw_of(cpumask_any(cur));
4981 trial_cpus = cpumask_weight(trial);
4983 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4984 if (cur_dl_b->bw != -1 &&
4985 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4987 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4988 rcu_read_unlock_sched();
4993 int task_can_attach(struct task_struct *p,
4994 const struct cpumask *cs_cpus_allowed)
4999 * Kthreads which disallow setaffinity shouldn't be moved
5000 * to a new cpuset; we don't want to change their cpu
5001 * affinity and isolating such threads by their set of
5002 * allowed nodes is unnecessary. Thus, cpusets are not
5003 * applicable for such threads. This prevents checking for
5004 * success of set_cpus_allowed_ptr() on all attached tasks
5005 * before cpus_allowed may be changed.
5007 if (p->flags & PF_NO_SETAFFINITY) {
5013 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5015 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5020 unsigned long flags;
5022 rcu_read_lock_sched();
5023 dl_b = dl_bw_of(dest_cpu);
5024 raw_spin_lock_irqsave(&dl_b->lock, flags);
5025 cpus = dl_bw_cpus(dest_cpu);
5026 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5031 * We reserve space for this task in the destination
5032 * root_domain, as we can't fail after this point.
5033 * We will free resources in the source root_domain
5034 * later on (see set_cpus_allowed_dl()).
5036 __dl_add(dl_b, p->dl.dl_bw);
5038 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5039 rcu_read_unlock_sched();
5049 #ifdef CONFIG_NUMA_BALANCING
5050 /* Migrate current task p to target_cpu */
5051 int migrate_task_to(struct task_struct *p, int target_cpu)
5053 struct migration_arg arg = { p, target_cpu };
5054 int curr_cpu = task_cpu(p);
5056 if (curr_cpu == target_cpu)
5059 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5062 /* TODO: This is not properly updating schedstats */
5064 trace_sched_move_numa(p, curr_cpu, target_cpu);
5065 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5069 * Requeue a task on a given node and accurately track the number of NUMA
5070 * tasks on the runqueues
5072 void sched_setnuma(struct task_struct *p, int nid)
5075 unsigned long flags;
5076 bool queued, running;
5078 rq = task_rq_lock(p, &flags);
5079 queued = task_on_rq_queued(p);
5080 running = task_current(rq, p);
5083 dequeue_task(rq, p, 0);
5085 put_prev_task(rq, p);
5087 p->numa_preferred_nid = nid;
5090 p->sched_class->set_curr_task(rq);
5092 enqueue_task(rq, p, 0);
5093 task_rq_unlock(rq, p, &flags);
5095 #endif /* CONFIG_NUMA_BALANCING */
5097 #ifdef CONFIG_HOTPLUG_CPU
5099 * Ensures that the idle task is using init_mm right before its cpu goes
5102 void idle_task_exit(void)
5104 struct mm_struct *mm = current->active_mm;
5106 BUG_ON(cpu_online(smp_processor_id()));
5108 if (mm != &init_mm) {
5109 switch_mm(mm, &init_mm, current);
5110 finish_arch_post_lock_switch();
5116 * Since this CPU is going 'away' for a while, fold any nr_active delta
5117 * we might have. Assumes we're called after migrate_tasks() so that the
5118 * nr_active count is stable.
5120 * Also see the comment "Global load-average calculations".
5122 static void calc_load_migrate(struct rq *rq)
5124 long delta = calc_load_fold_active(rq);
5126 atomic_long_add(delta, &calc_load_tasks);
5129 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5133 static const struct sched_class fake_sched_class = {
5134 .put_prev_task = put_prev_task_fake,
5137 static struct task_struct fake_task = {
5139 * Avoid pull_{rt,dl}_task()
5141 .prio = MAX_PRIO + 1,
5142 .sched_class = &fake_sched_class,
5146 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5147 * try_to_wake_up()->select_task_rq().
5149 * Called with rq->lock held even though we'er in stop_machine() and
5150 * there's no concurrency possible, we hold the required locks anyway
5151 * because of lock validation efforts.
5153 static void migrate_tasks(struct rq *dead_rq)
5155 struct rq *rq = dead_rq;
5156 struct task_struct *next, *stop = rq->stop;
5160 * Fudge the rq selection such that the below task selection loop
5161 * doesn't get stuck on the currently eligible stop task.
5163 * We're currently inside stop_machine() and the rq is either stuck
5164 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5165 * either way we should never end up calling schedule() until we're
5171 * put_prev_task() and pick_next_task() sched
5172 * class method both need to have an up-to-date
5173 * value of rq->clock[_task]
5175 update_rq_clock(rq);
5179 * There's this thread running, bail when that's the only
5182 if (rq->nr_running == 1)
5186 * Ensure rq->lock covers the entire task selection
5187 * until the migration.
5189 lockdep_pin_lock(&rq->lock);
5190 next = pick_next_task(rq, &fake_task);
5192 next->sched_class->put_prev_task(rq, next);
5194 /* Find suitable destination for @next, with force if needed. */
5195 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5197 lockdep_unpin_lock(&rq->lock);
5198 rq = __migrate_task(rq, next, dest_cpu);
5199 if (rq != dead_rq) {
5200 raw_spin_unlock(&rq->lock);
5202 raw_spin_lock(&rq->lock);
5208 #endif /* CONFIG_HOTPLUG_CPU */
5210 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5212 static struct ctl_table sd_ctl_dir[] = {
5214 .procname = "sched_domain",
5220 static struct ctl_table sd_ctl_root[] = {
5222 .procname = "kernel",
5224 .child = sd_ctl_dir,
5229 static struct ctl_table *sd_alloc_ctl_entry(int n)
5231 struct ctl_table *entry =
5232 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5237 static void sd_free_ctl_entry(struct ctl_table **tablep)
5239 struct ctl_table *entry;
5242 * In the intermediate directories, both the child directory and
5243 * procname are dynamically allocated and could fail but the mode
5244 * will always be set. In the lowest directory the names are
5245 * static strings and all have proc handlers.
5247 for (entry = *tablep; entry->mode; entry++) {
5249 sd_free_ctl_entry(&entry->child);
5250 if (entry->proc_handler == NULL)
5251 kfree(entry->procname);
5258 static int min_load_idx = 0;
5259 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5262 set_table_entry(struct ctl_table *entry,
5263 const char *procname, void *data, int maxlen,
5264 umode_t mode, proc_handler *proc_handler,
5267 entry->procname = procname;
5269 entry->maxlen = maxlen;
5271 entry->proc_handler = proc_handler;
5274 entry->extra1 = &min_load_idx;
5275 entry->extra2 = &max_load_idx;
5279 static struct ctl_table *
5280 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5282 struct ctl_table *table = sd_alloc_ctl_entry(14);
5287 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5288 sizeof(long), 0644, proc_doulongvec_minmax, false);
5289 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5290 sizeof(long), 0644, proc_doulongvec_minmax, false);
5291 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5292 sizeof(int), 0644, proc_dointvec_minmax, true);
5293 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5294 sizeof(int), 0644, proc_dointvec_minmax, true);
5295 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5296 sizeof(int), 0644, proc_dointvec_minmax, true);
5297 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5298 sizeof(int), 0644, proc_dointvec_minmax, true);
5299 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5300 sizeof(int), 0644, proc_dointvec_minmax, true);
5301 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5302 sizeof(int), 0644, proc_dointvec_minmax, false);
5303 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5304 sizeof(int), 0644, proc_dointvec_minmax, false);
5305 set_table_entry(&table[9], "cache_nice_tries",
5306 &sd->cache_nice_tries,
5307 sizeof(int), 0644, proc_dointvec_minmax, false);
5308 set_table_entry(&table[10], "flags", &sd->flags,
5309 sizeof(int), 0644, proc_dointvec_minmax, false);
5310 set_table_entry(&table[11], "max_newidle_lb_cost",
5311 &sd->max_newidle_lb_cost,
5312 sizeof(long), 0644, proc_doulongvec_minmax, false);
5313 set_table_entry(&table[12], "name", sd->name,
5314 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5315 /* &table[13] is terminator */
5320 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5322 struct ctl_table *entry, *table;
5323 struct sched_domain *sd;
5324 int domain_num = 0, i;
5327 for_each_domain(cpu, sd)
5329 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5334 for_each_domain(cpu, sd) {
5335 snprintf(buf, 32, "domain%d", i);
5336 entry->procname = kstrdup(buf, GFP_KERNEL);
5338 entry->child = sd_alloc_ctl_domain_table(sd);
5345 static struct ctl_table_header *sd_sysctl_header;
5346 static void register_sched_domain_sysctl(void)
5348 int i, cpu_num = num_possible_cpus();
5349 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5352 WARN_ON(sd_ctl_dir[0].child);
5353 sd_ctl_dir[0].child = entry;
5358 for_each_possible_cpu(i) {
5359 snprintf(buf, 32, "cpu%d", i);
5360 entry->procname = kstrdup(buf, GFP_KERNEL);
5362 entry->child = sd_alloc_ctl_cpu_table(i);
5366 WARN_ON(sd_sysctl_header);
5367 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5370 /* may be called multiple times per register */
5371 static void unregister_sched_domain_sysctl(void)
5373 unregister_sysctl_table(sd_sysctl_header);
5374 sd_sysctl_header = NULL;
5375 if (sd_ctl_dir[0].child)
5376 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5379 static void register_sched_domain_sysctl(void)
5382 static void unregister_sched_domain_sysctl(void)
5385 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5387 static void set_rq_online(struct rq *rq)
5390 const struct sched_class *class;
5392 cpumask_set_cpu(rq->cpu, rq->rd->online);
5395 for_each_class(class) {
5396 if (class->rq_online)
5397 class->rq_online(rq);
5402 static void set_rq_offline(struct rq *rq)
5405 const struct sched_class *class;
5407 for_each_class(class) {
5408 if (class->rq_offline)
5409 class->rq_offline(rq);
5412 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5418 * migration_call - callback that gets triggered when a CPU is added.
5419 * Here we can start up the necessary migration thread for the new CPU.
5422 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5424 int cpu = (long)hcpu;
5425 unsigned long flags;
5426 struct rq *rq = cpu_rq(cpu);
5428 switch (action & ~CPU_TASKS_FROZEN) {
5430 case CPU_UP_PREPARE:
5431 rq->calc_load_update = calc_load_update;
5435 /* Update our root-domain */
5436 raw_spin_lock_irqsave(&rq->lock, flags);
5438 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5442 raw_spin_unlock_irqrestore(&rq->lock, flags);
5445 #ifdef CONFIG_HOTPLUG_CPU
5447 sched_ttwu_pending();
5448 /* Update our root-domain */
5449 raw_spin_lock_irqsave(&rq->lock, flags);
5451 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5455 BUG_ON(rq->nr_running != 1); /* the migration thread */
5456 raw_spin_unlock_irqrestore(&rq->lock, flags);
5460 calc_load_migrate(rq);
5465 update_max_interval();
5471 * Register at high priority so that task migration (migrate_all_tasks)
5472 * happens before everything else. This has to be lower priority than
5473 * the notifier in the perf_event subsystem, though.
5475 static struct notifier_block migration_notifier = {
5476 .notifier_call = migration_call,
5477 .priority = CPU_PRI_MIGRATION,
5480 static void set_cpu_rq_start_time(void)
5482 int cpu = smp_processor_id();
5483 struct rq *rq = cpu_rq(cpu);
5484 rq->age_stamp = sched_clock_cpu(cpu);
5487 static int sched_cpu_active(struct notifier_block *nfb,
5488 unsigned long action, void *hcpu)
5490 switch (action & ~CPU_TASKS_FROZEN) {
5492 set_cpu_rq_start_time();
5496 * At this point a starting CPU has marked itself as online via
5497 * set_cpu_online(). But it might not yet have marked itself
5498 * as active, which is essential from here on.
5500 * Thus, fall-through and help the starting CPU along.
5502 case CPU_DOWN_FAILED:
5503 set_cpu_active((long)hcpu, true);
5510 static int sched_cpu_inactive(struct notifier_block *nfb,
5511 unsigned long action, void *hcpu)
5513 switch (action & ~CPU_TASKS_FROZEN) {
5514 case CPU_DOWN_PREPARE:
5515 set_cpu_active((long)hcpu, false);
5522 static int __init migration_init(void)
5524 void *cpu = (void *)(long)smp_processor_id();
5527 /* Initialize migration for the boot CPU */
5528 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5529 BUG_ON(err == NOTIFY_BAD);
5530 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5531 register_cpu_notifier(&migration_notifier);
5533 /* Register cpu active notifiers */
5534 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5535 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5539 early_initcall(migration_init);
5541 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5543 #ifdef CONFIG_SCHED_DEBUG
5545 static __read_mostly int sched_debug_enabled;
5547 static int __init sched_debug_setup(char *str)
5549 sched_debug_enabled = 1;
5553 early_param("sched_debug", sched_debug_setup);
5555 static inline bool sched_debug(void)
5557 return sched_debug_enabled;
5560 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5561 struct cpumask *groupmask)
5563 struct sched_group *group = sd->groups;
5565 cpumask_clear(groupmask);
5567 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5569 if (!(sd->flags & SD_LOAD_BALANCE)) {
5570 printk("does not load-balance\n");
5572 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5577 printk(KERN_CONT "span %*pbl level %s\n",
5578 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5580 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5581 printk(KERN_ERR "ERROR: domain->span does not contain "
5584 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5585 printk(KERN_ERR "ERROR: domain->groups does not contain"
5589 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5593 printk(KERN_ERR "ERROR: group is NULL\n");
5597 if (!cpumask_weight(sched_group_cpus(group))) {
5598 printk(KERN_CONT "\n");
5599 printk(KERN_ERR "ERROR: empty group\n");
5603 if (!(sd->flags & SD_OVERLAP) &&
5604 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5605 printk(KERN_CONT "\n");
5606 printk(KERN_ERR "ERROR: repeated CPUs\n");
5610 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5612 printk(KERN_CONT " %*pbl",
5613 cpumask_pr_args(sched_group_cpus(group)));
5614 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5615 printk(KERN_CONT " (cpu_capacity = %d)",
5616 group->sgc->capacity);
5619 group = group->next;
5620 } while (group != sd->groups);
5621 printk(KERN_CONT "\n");
5623 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5624 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5627 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5628 printk(KERN_ERR "ERROR: parent span is not a superset "
5629 "of domain->span\n");
5633 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5637 if (!sched_debug_enabled)
5641 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5645 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5648 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5656 #else /* !CONFIG_SCHED_DEBUG */
5657 # define sched_domain_debug(sd, cpu) do { } while (0)
5658 static inline bool sched_debug(void)
5662 #endif /* CONFIG_SCHED_DEBUG */
5664 static int sd_degenerate(struct sched_domain *sd)
5666 if (cpumask_weight(sched_domain_span(sd)) == 1)
5669 /* Following flags need at least 2 groups */
5670 if (sd->flags & (SD_LOAD_BALANCE |
5671 SD_BALANCE_NEWIDLE |
5674 SD_SHARE_CPUCAPACITY |
5675 SD_SHARE_PKG_RESOURCES |
5676 SD_SHARE_POWERDOMAIN)) {
5677 if (sd->groups != sd->groups->next)
5681 /* Following flags don't use groups */
5682 if (sd->flags & (SD_WAKE_AFFINE))
5689 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5691 unsigned long cflags = sd->flags, pflags = parent->flags;
5693 if (sd_degenerate(parent))
5696 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5699 /* Flags needing groups don't count if only 1 group in parent */
5700 if (parent->groups == parent->groups->next) {
5701 pflags &= ~(SD_LOAD_BALANCE |
5702 SD_BALANCE_NEWIDLE |
5705 SD_SHARE_CPUCAPACITY |
5706 SD_SHARE_PKG_RESOURCES |
5708 SD_SHARE_POWERDOMAIN);
5709 if (nr_node_ids == 1)
5710 pflags &= ~SD_SERIALIZE;
5712 if (~cflags & pflags)
5718 static void free_rootdomain(struct rcu_head *rcu)
5720 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5722 cpupri_cleanup(&rd->cpupri);
5723 cpudl_cleanup(&rd->cpudl);
5724 free_cpumask_var(rd->dlo_mask);
5725 free_cpumask_var(rd->rto_mask);
5726 free_cpumask_var(rd->online);
5727 free_cpumask_var(rd->span);
5731 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5733 struct root_domain *old_rd = NULL;
5734 unsigned long flags;
5736 raw_spin_lock_irqsave(&rq->lock, flags);
5741 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5744 cpumask_clear_cpu(rq->cpu, old_rd->span);
5747 * If we dont want to free the old_rd yet then
5748 * set old_rd to NULL to skip the freeing later
5751 if (!atomic_dec_and_test(&old_rd->refcount))
5755 atomic_inc(&rd->refcount);
5758 cpumask_set_cpu(rq->cpu, rd->span);
5759 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5762 raw_spin_unlock_irqrestore(&rq->lock, flags);
5765 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5768 static int init_rootdomain(struct root_domain *rd)
5770 memset(rd, 0, sizeof(*rd));
5772 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5774 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5776 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5778 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5781 init_dl_bw(&rd->dl_bw);
5782 if (cpudl_init(&rd->cpudl) != 0)
5785 if (cpupri_init(&rd->cpupri) != 0)
5790 free_cpumask_var(rd->rto_mask);
5792 free_cpumask_var(rd->dlo_mask);
5794 free_cpumask_var(rd->online);
5796 free_cpumask_var(rd->span);
5802 * By default the system creates a single root-domain with all cpus as
5803 * members (mimicking the global state we have today).
5805 struct root_domain def_root_domain;
5807 static void init_defrootdomain(void)
5809 init_rootdomain(&def_root_domain);
5811 atomic_set(&def_root_domain.refcount, 1);
5814 static struct root_domain *alloc_rootdomain(void)
5816 struct root_domain *rd;
5818 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5822 if (init_rootdomain(rd) != 0) {
5830 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5832 struct sched_group *tmp, *first;
5841 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5846 } while (sg != first);
5849 static void free_sched_domain(struct rcu_head *rcu)
5851 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5854 * If its an overlapping domain it has private groups, iterate and
5857 if (sd->flags & SD_OVERLAP) {
5858 free_sched_groups(sd->groups, 1);
5859 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5860 kfree(sd->groups->sgc);
5866 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5868 call_rcu(&sd->rcu, free_sched_domain);
5871 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5873 for (; sd; sd = sd->parent)
5874 destroy_sched_domain(sd, cpu);
5878 * Keep a special pointer to the highest sched_domain that has
5879 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5880 * allows us to avoid some pointer chasing select_idle_sibling().
5882 * Also keep a unique ID per domain (we use the first cpu number in
5883 * the cpumask of the domain), this allows us to quickly tell if
5884 * two cpus are in the same cache domain, see cpus_share_cache().
5886 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5887 DEFINE_PER_CPU(int, sd_llc_size);
5888 DEFINE_PER_CPU(int, sd_llc_id);
5889 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5890 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5891 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5893 static void update_top_cache_domain(int cpu)
5895 struct sched_domain *sd;
5896 struct sched_domain *busy_sd = NULL;
5900 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5902 id = cpumask_first(sched_domain_span(sd));
5903 size = cpumask_weight(sched_domain_span(sd));
5904 busy_sd = sd->parent; /* sd_busy */
5906 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5908 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5909 per_cpu(sd_llc_size, cpu) = size;
5910 per_cpu(sd_llc_id, cpu) = id;
5912 sd = lowest_flag_domain(cpu, SD_NUMA);
5913 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5915 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5916 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5920 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5921 * hold the hotplug lock.
5924 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5926 struct rq *rq = cpu_rq(cpu);
5927 struct sched_domain *tmp;
5929 /* Remove the sched domains which do not contribute to scheduling. */
5930 for (tmp = sd; tmp; ) {
5931 struct sched_domain *parent = tmp->parent;
5935 if (sd_parent_degenerate(tmp, parent)) {
5936 tmp->parent = parent->parent;
5938 parent->parent->child = tmp;
5940 * Transfer SD_PREFER_SIBLING down in case of a
5941 * degenerate parent; the spans match for this
5942 * so the property transfers.
5944 if (parent->flags & SD_PREFER_SIBLING)
5945 tmp->flags |= SD_PREFER_SIBLING;
5946 destroy_sched_domain(parent, cpu);
5951 if (sd && sd_degenerate(sd)) {
5954 destroy_sched_domain(tmp, cpu);
5959 sched_domain_debug(sd, cpu);
5961 rq_attach_root(rq, rd);
5963 rcu_assign_pointer(rq->sd, sd);
5964 destroy_sched_domains(tmp, cpu);
5966 update_top_cache_domain(cpu);
5969 /* Setup the mask of cpus configured for isolated domains */
5970 static int __init isolated_cpu_setup(char *str)
5972 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5973 cpulist_parse(str, cpu_isolated_map);
5977 __setup("isolcpus=", isolated_cpu_setup);
5980 struct sched_domain ** __percpu sd;
5981 struct root_domain *rd;
5992 * Build an iteration mask that can exclude certain CPUs from the upwards
5995 * Asymmetric node setups can result in situations where the domain tree is of
5996 * unequal depth, make sure to skip domains that already cover the entire
5999 * In that case build_sched_domains() will have terminated the iteration early
6000 * and our sibling sd spans will be empty. Domains should always include the
6001 * cpu they're built on, so check that.
6004 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6006 const struct cpumask *span = sched_domain_span(sd);
6007 struct sd_data *sdd = sd->private;
6008 struct sched_domain *sibling;
6011 for_each_cpu(i, span) {
6012 sibling = *per_cpu_ptr(sdd->sd, i);
6013 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6016 cpumask_set_cpu(i, sched_group_mask(sg));
6021 * Return the canonical balance cpu for this group, this is the first cpu
6022 * of this group that's also in the iteration mask.
6024 int group_balance_cpu(struct sched_group *sg)
6026 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6030 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6032 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6033 const struct cpumask *span = sched_domain_span(sd);
6034 struct cpumask *covered = sched_domains_tmpmask;
6035 struct sd_data *sdd = sd->private;
6036 struct sched_domain *sibling;
6039 cpumask_clear(covered);
6041 for_each_cpu(i, span) {
6042 struct cpumask *sg_span;
6044 if (cpumask_test_cpu(i, covered))
6047 sibling = *per_cpu_ptr(sdd->sd, i);
6049 /* See the comment near build_group_mask(). */
6050 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6053 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6054 GFP_KERNEL, cpu_to_node(cpu));
6059 sg_span = sched_group_cpus(sg);
6061 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6063 cpumask_set_cpu(i, sg_span);
6065 cpumask_or(covered, covered, sg_span);
6067 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6068 if (atomic_inc_return(&sg->sgc->ref) == 1)
6069 build_group_mask(sd, sg);
6072 * Initialize sgc->capacity such that even if we mess up the
6073 * domains and no possible iteration will get us here, we won't
6076 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6079 * Make sure the first group of this domain contains the
6080 * canonical balance cpu. Otherwise the sched_domain iteration
6081 * breaks. See update_sg_lb_stats().
6083 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6084 group_balance_cpu(sg) == cpu)
6094 sd->groups = groups;
6099 free_sched_groups(first, 0);
6104 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6106 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6107 struct sched_domain *child = sd->child;
6110 cpu = cpumask_first(sched_domain_span(child));
6113 *sg = *per_cpu_ptr(sdd->sg, cpu);
6114 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6115 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6122 * build_sched_groups will build a circular linked list of the groups
6123 * covered by the given span, and will set each group's ->cpumask correctly,
6124 * and ->cpu_capacity to 0.
6126 * Assumes the sched_domain tree is fully constructed
6129 build_sched_groups(struct sched_domain *sd, int cpu)
6131 struct sched_group *first = NULL, *last = NULL;
6132 struct sd_data *sdd = sd->private;
6133 const struct cpumask *span = sched_domain_span(sd);
6134 struct cpumask *covered;
6137 get_group(cpu, sdd, &sd->groups);
6138 atomic_inc(&sd->groups->ref);
6140 if (cpu != cpumask_first(span))
6143 lockdep_assert_held(&sched_domains_mutex);
6144 covered = sched_domains_tmpmask;
6146 cpumask_clear(covered);
6148 for_each_cpu(i, span) {
6149 struct sched_group *sg;
6152 if (cpumask_test_cpu(i, covered))
6155 group = get_group(i, sdd, &sg);
6156 cpumask_setall(sched_group_mask(sg));
6158 for_each_cpu(j, span) {
6159 if (get_group(j, sdd, NULL) != group)
6162 cpumask_set_cpu(j, covered);
6163 cpumask_set_cpu(j, sched_group_cpus(sg));
6178 * Initialize sched groups cpu_capacity.
6180 * cpu_capacity indicates the capacity of sched group, which is used while
6181 * distributing the load between different sched groups in a sched domain.
6182 * Typically cpu_capacity for all the groups in a sched domain will be same
6183 * unless there are asymmetries in the topology. If there are asymmetries,
6184 * group having more cpu_capacity will pickup more load compared to the
6185 * group having less cpu_capacity.
6187 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6189 struct sched_group *sg = sd->groups;
6194 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6196 } while (sg != sd->groups);
6198 if (cpu != group_balance_cpu(sg))
6201 update_group_capacity(sd, cpu);
6202 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6206 * Initializers for schedule domains
6207 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6210 static int default_relax_domain_level = -1;
6211 int sched_domain_level_max;
6213 static int __init setup_relax_domain_level(char *str)
6215 if (kstrtoint(str, 0, &default_relax_domain_level))
6216 pr_warn("Unable to set relax_domain_level\n");
6220 __setup("relax_domain_level=", setup_relax_domain_level);
6222 static void set_domain_attribute(struct sched_domain *sd,
6223 struct sched_domain_attr *attr)
6227 if (!attr || attr->relax_domain_level < 0) {
6228 if (default_relax_domain_level < 0)
6231 request = default_relax_domain_level;
6233 request = attr->relax_domain_level;
6234 if (request < sd->level) {
6235 /* turn off idle balance on this domain */
6236 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6238 /* turn on idle balance on this domain */
6239 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6243 static void __sdt_free(const struct cpumask *cpu_map);
6244 static int __sdt_alloc(const struct cpumask *cpu_map);
6246 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6247 const struct cpumask *cpu_map)
6251 if (!atomic_read(&d->rd->refcount))
6252 free_rootdomain(&d->rd->rcu); /* fall through */
6254 free_percpu(d->sd); /* fall through */
6256 __sdt_free(cpu_map); /* fall through */
6262 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6263 const struct cpumask *cpu_map)
6265 memset(d, 0, sizeof(*d));
6267 if (__sdt_alloc(cpu_map))
6268 return sa_sd_storage;
6269 d->sd = alloc_percpu(struct sched_domain *);
6271 return sa_sd_storage;
6272 d->rd = alloc_rootdomain();
6275 return sa_rootdomain;
6279 * NULL the sd_data elements we've used to build the sched_domain and
6280 * sched_group structure so that the subsequent __free_domain_allocs()
6281 * will not free the data we're using.
6283 static void claim_allocations(int cpu, struct sched_domain *sd)
6285 struct sd_data *sdd = sd->private;
6287 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6288 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6290 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6291 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6293 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6294 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6298 static int sched_domains_numa_levels;
6299 enum numa_topology_type sched_numa_topology_type;
6300 static int *sched_domains_numa_distance;
6301 int sched_max_numa_distance;
6302 static struct cpumask ***sched_domains_numa_masks;
6303 static int sched_domains_curr_level;
6307 * SD_flags allowed in topology descriptions.
6309 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6310 * SD_SHARE_PKG_RESOURCES - describes shared caches
6311 * SD_NUMA - describes NUMA topologies
6312 * SD_SHARE_POWERDOMAIN - describes shared power domain
6315 * SD_ASYM_PACKING - describes SMT quirks
6317 #define TOPOLOGY_SD_FLAGS \
6318 (SD_SHARE_CPUCAPACITY | \
6319 SD_SHARE_PKG_RESOURCES | \
6322 SD_SHARE_POWERDOMAIN)
6324 static struct sched_domain *
6325 sd_init(struct sched_domain_topology_level *tl, int cpu)
6327 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6328 int sd_weight, sd_flags = 0;
6332 * Ugly hack to pass state to sd_numa_mask()...
6334 sched_domains_curr_level = tl->numa_level;
6337 sd_weight = cpumask_weight(tl->mask(cpu));
6340 sd_flags = (*tl->sd_flags)();
6341 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6342 "wrong sd_flags in topology description\n"))
6343 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6345 *sd = (struct sched_domain){
6346 .min_interval = sd_weight,
6347 .max_interval = 2*sd_weight,
6349 .imbalance_pct = 125,
6351 .cache_nice_tries = 0,
6358 .flags = 1*SD_LOAD_BALANCE
6359 | 1*SD_BALANCE_NEWIDLE
6364 | 0*SD_SHARE_CPUCAPACITY
6365 | 0*SD_SHARE_PKG_RESOURCES
6367 | 0*SD_PREFER_SIBLING
6372 .last_balance = jiffies,
6373 .balance_interval = sd_weight,
6375 .max_newidle_lb_cost = 0,
6376 .next_decay_max_lb_cost = jiffies,
6377 #ifdef CONFIG_SCHED_DEBUG
6383 * Convert topological properties into behaviour.
6386 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6387 sd->flags |= SD_PREFER_SIBLING;
6388 sd->imbalance_pct = 110;
6389 sd->smt_gain = 1178; /* ~15% */
6391 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6392 sd->imbalance_pct = 117;
6393 sd->cache_nice_tries = 1;
6397 } else if (sd->flags & SD_NUMA) {
6398 sd->cache_nice_tries = 2;
6402 sd->flags |= SD_SERIALIZE;
6403 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6404 sd->flags &= ~(SD_BALANCE_EXEC |
6411 sd->flags |= SD_PREFER_SIBLING;
6412 sd->cache_nice_tries = 1;
6417 sd->private = &tl->data;
6423 * Topology list, bottom-up.
6425 static struct sched_domain_topology_level default_topology[] = {
6426 #ifdef CONFIG_SCHED_SMT
6427 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6429 #ifdef CONFIG_SCHED_MC
6430 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6432 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6436 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6438 #define for_each_sd_topology(tl) \
6439 for (tl = sched_domain_topology; tl->mask; tl++)
6441 void set_sched_topology(struct sched_domain_topology_level *tl)
6443 sched_domain_topology = tl;
6448 static const struct cpumask *sd_numa_mask(int cpu)
6450 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6453 static void sched_numa_warn(const char *str)
6455 static int done = false;
6463 printk(KERN_WARNING "ERROR: %s\n\n", str);
6465 for (i = 0; i < nr_node_ids; i++) {
6466 printk(KERN_WARNING " ");
6467 for (j = 0; j < nr_node_ids; j++)
6468 printk(KERN_CONT "%02d ", node_distance(i,j));
6469 printk(KERN_CONT "\n");
6471 printk(KERN_WARNING "\n");
6474 bool find_numa_distance(int distance)
6478 if (distance == node_distance(0, 0))
6481 for (i = 0; i < sched_domains_numa_levels; i++) {
6482 if (sched_domains_numa_distance[i] == distance)
6490 * A system can have three types of NUMA topology:
6491 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6492 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6493 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6495 * The difference between a glueless mesh topology and a backplane
6496 * topology lies in whether communication between not directly
6497 * connected nodes goes through intermediary nodes (where programs
6498 * could run), or through backplane controllers. This affects
6499 * placement of programs.
6501 * The type of topology can be discerned with the following tests:
6502 * - If the maximum distance between any nodes is 1 hop, the system
6503 * is directly connected.
6504 * - If for two nodes A and B, located N > 1 hops away from each other,
6505 * there is an intermediary node C, which is < N hops away from both
6506 * nodes A and B, the system is a glueless mesh.
6508 static void init_numa_topology_type(void)
6512 n = sched_max_numa_distance;
6514 if (sched_domains_numa_levels <= 1) {
6515 sched_numa_topology_type = NUMA_DIRECT;
6519 for_each_online_node(a) {
6520 for_each_online_node(b) {
6521 /* Find two nodes furthest removed from each other. */
6522 if (node_distance(a, b) < n)
6525 /* Is there an intermediary node between a and b? */
6526 for_each_online_node(c) {
6527 if (node_distance(a, c) < n &&
6528 node_distance(b, c) < n) {
6529 sched_numa_topology_type =
6535 sched_numa_topology_type = NUMA_BACKPLANE;
6541 static void sched_init_numa(void)
6543 int next_distance, curr_distance = node_distance(0, 0);
6544 struct sched_domain_topology_level *tl;
6548 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6549 if (!sched_domains_numa_distance)
6553 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6554 * unique distances in the node_distance() table.
6556 * Assumes node_distance(0,j) includes all distances in
6557 * node_distance(i,j) in order to avoid cubic time.
6559 next_distance = curr_distance;
6560 for (i = 0; i < nr_node_ids; i++) {
6561 for (j = 0; j < nr_node_ids; j++) {
6562 for (k = 0; k < nr_node_ids; k++) {
6563 int distance = node_distance(i, k);
6565 if (distance > curr_distance &&
6566 (distance < next_distance ||
6567 next_distance == curr_distance))
6568 next_distance = distance;
6571 * While not a strong assumption it would be nice to know
6572 * about cases where if node A is connected to B, B is not
6573 * equally connected to A.
6575 if (sched_debug() && node_distance(k, i) != distance)
6576 sched_numa_warn("Node-distance not symmetric");
6578 if (sched_debug() && i && !find_numa_distance(distance))
6579 sched_numa_warn("Node-0 not representative");
6581 if (next_distance != curr_distance) {
6582 sched_domains_numa_distance[level++] = next_distance;
6583 sched_domains_numa_levels = level;
6584 curr_distance = next_distance;
6589 * In case of sched_debug() we verify the above assumption.
6599 * 'level' contains the number of unique distances, excluding the
6600 * identity distance node_distance(i,i).
6602 * The sched_domains_numa_distance[] array includes the actual distance
6607 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6608 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6609 * the array will contain less then 'level' members. This could be
6610 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6611 * in other functions.
6613 * We reset it to 'level' at the end of this function.
6615 sched_domains_numa_levels = 0;
6617 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6618 if (!sched_domains_numa_masks)
6622 * Now for each level, construct a mask per node which contains all
6623 * cpus of nodes that are that many hops away from us.
6625 for (i = 0; i < level; i++) {
6626 sched_domains_numa_masks[i] =
6627 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6628 if (!sched_domains_numa_masks[i])
6631 for (j = 0; j < nr_node_ids; j++) {
6632 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6636 sched_domains_numa_masks[i][j] = mask;
6638 for (k = 0; k < nr_node_ids; k++) {
6639 if (node_distance(j, k) > sched_domains_numa_distance[i])
6642 cpumask_or(mask, mask, cpumask_of_node(k));
6647 /* Compute default topology size */
6648 for (i = 0; sched_domain_topology[i].mask; i++);
6650 tl = kzalloc((i + level + 1) *
6651 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6656 * Copy the default topology bits..
6658 for (i = 0; sched_domain_topology[i].mask; i++)
6659 tl[i] = sched_domain_topology[i];
6662 * .. and append 'j' levels of NUMA goodness.
6664 for (j = 0; j < level; i++, j++) {
6665 tl[i] = (struct sched_domain_topology_level){
6666 .mask = sd_numa_mask,
6667 .sd_flags = cpu_numa_flags,
6668 .flags = SDTL_OVERLAP,
6674 sched_domain_topology = tl;
6676 sched_domains_numa_levels = level;
6677 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6679 init_numa_topology_type();
6682 static void sched_domains_numa_masks_set(int cpu)
6685 int node = cpu_to_node(cpu);
6687 for (i = 0; i < sched_domains_numa_levels; i++) {
6688 for (j = 0; j < nr_node_ids; j++) {
6689 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6690 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6695 static void sched_domains_numa_masks_clear(int cpu)
6698 for (i = 0; i < sched_domains_numa_levels; i++) {
6699 for (j = 0; j < nr_node_ids; j++)
6700 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6705 * Update sched_domains_numa_masks[level][node] array when new cpus
6708 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6709 unsigned long action,
6712 int cpu = (long)hcpu;
6714 switch (action & ~CPU_TASKS_FROZEN) {
6716 sched_domains_numa_masks_set(cpu);
6720 sched_domains_numa_masks_clear(cpu);
6730 static inline void sched_init_numa(void)
6734 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6735 unsigned long action,
6740 #endif /* CONFIG_NUMA */
6742 static int __sdt_alloc(const struct cpumask *cpu_map)
6744 struct sched_domain_topology_level *tl;
6747 for_each_sd_topology(tl) {
6748 struct sd_data *sdd = &tl->data;
6750 sdd->sd = alloc_percpu(struct sched_domain *);
6754 sdd->sg = alloc_percpu(struct sched_group *);
6758 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6762 for_each_cpu(j, cpu_map) {
6763 struct sched_domain *sd;
6764 struct sched_group *sg;
6765 struct sched_group_capacity *sgc;
6767 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6768 GFP_KERNEL, cpu_to_node(j));
6772 *per_cpu_ptr(sdd->sd, j) = sd;
6774 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6775 GFP_KERNEL, cpu_to_node(j));
6781 *per_cpu_ptr(sdd->sg, j) = sg;
6783 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6784 GFP_KERNEL, cpu_to_node(j));
6788 *per_cpu_ptr(sdd->sgc, j) = sgc;
6795 static void __sdt_free(const struct cpumask *cpu_map)
6797 struct sched_domain_topology_level *tl;
6800 for_each_sd_topology(tl) {
6801 struct sd_data *sdd = &tl->data;
6803 for_each_cpu(j, cpu_map) {
6804 struct sched_domain *sd;
6807 sd = *per_cpu_ptr(sdd->sd, j);
6808 if (sd && (sd->flags & SD_OVERLAP))
6809 free_sched_groups(sd->groups, 0);
6810 kfree(*per_cpu_ptr(sdd->sd, j));
6814 kfree(*per_cpu_ptr(sdd->sg, j));
6816 kfree(*per_cpu_ptr(sdd->sgc, j));
6818 free_percpu(sdd->sd);
6820 free_percpu(sdd->sg);
6822 free_percpu(sdd->sgc);
6827 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6828 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6829 struct sched_domain *child, int cpu)
6831 struct sched_domain *sd = sd_init(tl, cpu);
6835 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6837 sd->level = child->level + 1;
6838 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6842 if (!cpumask_subset(sched_domain_span(child),
6843 sched_domain_span(sd))) {
6844 pr_err("BUG: arch topology borken\n");
6845 #ifdef CONFIG_SCHED_DEBUG
6846 pr_err(" the %s domain not a subset of the %s domain\n",
6847 child->name, sd->name);
6849 /* Fixup, ensure @sd has at least @child cpus. */
6850 cpumask_or(sched_domain_span(sd),
6851 sched_domain_span(sd),
6852 sched_domain_span(child));
6856 set_domain_attribute(sd, attr);
6862 * Build sched domains for a given set of cpus and attach the sched domains
6863 * to the individual cpus
6865 static int build_sched_domains(const struct cpumask *cpu_map,
6866 struct sched_domain_attr *attr)
6868 enum s_alloc alloc_state;
6869 struct sched_domain *sd;
6871 int i, ret = -ENOMEM;
6873 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6874 if (alloc_state != sa_rootdomain)
6877 /* Set up domains for cpus specified by the cpu_map. */
6878 for_each_cpu(i, cpu_map) {
6879 struct sched_domain_topology_level *tl;
6882 for_each_sd_topology(tl) {
6883 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6884 if (tl == sched_domain_topology)
6885 *per_cpu_ptr(d.sd, i) = sd;
6886 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6887 sd->flags |= SD_OVERLAP;
6888 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6893 /* Build the groups for the domains */
6894 for_each_cpu(i, cpu_map) {
6895 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6896 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6897 if (sd->flags & SD_OVERLAP) {
6898 if (build_overlap_sched_groups(sd, i))
6901 if (build_sched_groups(sd, i))
6907 /* Calculate CPU capacity for physical packages and nodes */
6908 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6909 if (!cpumask_test_cpu(i, cpu_map))
6912 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6913 claim_allocations(i, sd);
6914 init_sched_groups_capacity(i, sd);
6918 /* Attach the domains */
6920 for_each_cpu(i, cpu_map) {
6921 sd = *per_cpu_ptr(d.sd, i);
6922 cpu_attach_domain(sd, d.rd, i);
6928 __free_domain_allocs(&d, alloc_state, cpu_map);
6932 static cpumask_var_t *doms_cur; /* current sched domains */
6933 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6934 static struct sched_domain_attr *dattr_cur;
6935 /* attribues of custom domains in 'doms_cur' */
6938 * Special case: If a kmalloc of a doms_cur partition (array of
6939 * cpumask) fails, then fallback to a single sched domain,
6940 * as determined by the single cpumask fallback_doms.
6942 static cpumask_var_t fallback_doms;
6945 * arch_update_cpu_topology lets virtualized architectures update the
6946 * cpu core maps. It is supposed to return 1 if the topology changed
6947 * or 0 if it stayed the same.
6949 int __weak arch_update_cpu_topology(void)
6954 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6957 cpumask_var_t *doms;
6959 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6962 for (i = 0; i < ndoms; i++) {
6963 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6964 free_sched_domains(doms, i);
6971 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6974 for (i = 0; i < ndoms; i++)
6975 free_cpumask_var(doms[i]);
6980 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6981 * For now this just excludes isolated cpus, but could be used to
6982 * exclude other special cases in the future.
6984 static int init_sched_domains(const struct cpumask *cpu_map)
6988 arch_update_cpu_topology();
6990 doms_cur = alloc_sched_domains(ndoms_cur);
6992 doms_cur = &fallback_doms;
6993 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6994 err = build_sched_domains(doms_cur[0], NULL);
6995 register_sched_domain_sysctl();
7001 * Detach sched domains from a group of cpus specified in cpu_map
7002 * These cpus will now be attached to the NULL domain
7004 static void detach_destroy_domains(const struct cpumask *cpu_map)
7009 for_each_cpu(i, cpu_map)
7010 cpu_attach_domain(NULL, &def_root_domain, i);
7014 /* handle null as "default" */
7015 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7016 struct sched_domain_attr *new, int idx_new)
7018 struct sched_domain_attr tmp;
7025 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7026 new ? (new + idx_new) : &tmp,
7027 sizeof(struct sched_domain_attr));
7031 * Partition sched domains as specified by the 'ndoms_new'
7032 * cpumasks in the array doms_new[] of cpumasks. This compares
7033 * doms_new[] to the current sched domain partitioning, doms_cur[].
7034 * It destroys each deleted domain and builds each new domain.
7036 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7037 * The masks don't intersect (don't overlap.) We should setup one
7038 * sched domain for each mask. CPUs not in any of the cpumasks will
7039 * not be load balanced. If the same cpumask appears both in the
7040 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7043 * The passed in 'doms_new' should be allocated using
7044 * alloc_sched_domains. This routine takes ownership of it and will
7045 * free_sched_domains it when done with it. If the caller failed the
7046 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7047 * and partition_sched_domains() will fallback to the single partition
7048 * 'fallback_doms', it also forces the domains to be rebuilt.
7050 * If doms_new == NULL it will be replaced with cpu_online_mask.
7051 * ndoms_new == 0 is a special case for destroying existing domains,
7052 * and it will not create the default domain.
7054 * Call with hotplug lock held
7056 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7057 struct sched_domain_attr *dattr_new)
7062 mutex_lock(&sched_domains_mutex);
7064 /* always unregister in case we don't destroy any domains */
7065 unregister_sched_domain_sysctl();
7067 /* Let architecture update cpu core mappings. */
7068 new_topology = arch_update_cpu_topology();
7070 n = doms_new ? ndoms_new : 0;
7072 /* Destroy deleted domains */
7073 for (i = 0; i < ndoms_cur; i++) {
7074 for (j = 0; j < n && !new_topology; j++) {
7075 if (cpumask_equal(doms_cur[i], doms_new[j])
7076 && dattrs_equal(dattr_cur, i, dattr_new, j))
7079 /* no match - a current sched domain not in new doms_new[] */
7080 detach_destroy_domains(doms_cur[i]);
7086 if (doms_new == NULL) {
7088 doms_new = &fallback_doms;
7089 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7090 WARN_ON_ONCE(dattr_new);
7093 /* Build new domains */
7094 for (i = 0; i < ndoms_new; i++) {
7095 for (j = 0; j < n && !new_topology; j++) {
7096 if (cpumask_equal(doms_new[i], doms_cur[j])
7097 && dattrs_equal(dattr_new, i, dattr_cur, j))
7100 /* no match - add a new doms_new */
7101 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7106 /* Remember the new sched domains */
7107 if (doms_cur != &fallback_doms)
7108 free_sched_domains(doms_cur, ndoms_cur);
7109 kfree(dattr_cur); /* kfree(NULL) is safe */
7110 doms_cur = doms_new;
7111 dattr_cur = dattr_new;
7112 ndoms_cur = ndoms_new;
7114 register_sched_domain_sysctl();
7116 mutex_unlock(&sched_domains_mutex);
7119 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7122 * Update cpusets according to cpu_active mask. If cpusets are
7123 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7124 * around partition_sched_domains().
7126 * If we come here as part of a suspend/resume, don't touch cpusets because we
7127 * want to restore it back to its original state upon resume anyway.
7129 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7133 case CPU_ONLINE_FROZEN:
7134 case CPU_DOWN_FAILED_FROZEN:
7137 * num_cpus_frozen tracks how many CPUs are involved in suspend
7138 * resume sequence. As long as this is not the last online
7139 * operation in the resume sequence, just build a single sched
7140 * domain, ignoring cpusets.
7143 if (likely(num_cpus_frozen)) {
7144 partition_sched_domains(1, NULL, NULL);
7149 * This is the last CPU online operation. So fall through and
7150 * restore the original sched domains by considering the
7151 * cpuset configurations.
7155 cpuset_update_active_cpus(true);
7163 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7166 unsigned long flags;
7167 long cpu = (long)hcpu;
7173 case CPU_DOWN_PREPARE:
7174 rcu_read_lock_sched();
7175 dl_b = dl_bw_of(cpu);
7177 raw_spin_lock_irqsave(&dl_b->lock, flags);
7178 cpus = dl_bw_cpus(cpu);
7179 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7180 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7182 rcu_read_unlock_sched();
7185 return notifier_from_errno(-EBUSY);
7186 cpuset_update_active_cpus(false);
7188 case CPU_DOWN_PREPARE_FROZEN:
7190 partition_sched_domains(1, NULL, NULL);
7198 void __init sched_init_smp(void)
7200 cpumask_var_t non_isolated_cpus;
7202 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7203 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7205 /* nohz_full won't take effect without isolating the cpus. */
7206 tick_nohz_full_add_cpus_to(cpu_isolated_map);
7211 * There's no userspace yet to cause hotplug operations; hence all the
7212 * cpu masks are stable and all blatant races in the below code cannot
7215 mutex_lock(&sched_domains_mutex);
7216 init_sched_domains(cpu_active_mask);
7217 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7218 if (cpumask_empty(non_isolated_cpus))
7219 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7220 mutex_unlock(&sched_domains_mutex);
7222 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7223 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7224 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7228 /* Move init over to a non-isolated CPU */
7229 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7231 sched_init_granularity();
7232 free_cpumask_var(non_isolated_cpus);
7234 init_sched_rt_class();
7235 init_sched_dl_class();
7238 void __init sched_init_smp(void)
7240 sched_init_granularity();
7242 #endif /* CONFIG_SMP */
7244 int in_sched_functions(unsigned long addr)
7246 return in_lock_functions(addr) ||
7247 (addr >= (unsigned long)__sched_text_start
7248 && addr < (unsigned long)__sched_text_end);
7251 #ifdef CONFIG_CGROUP_SCHED
7253 * Default task group.
7254 * Every task in system belongs to this group at bootup.
7256 struct task_group root_task_group;
7257 LIST_HEAD(task_groups);
7260 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7262 void __init sched_init(void)
7265 unsigned long alloc_size = 0, ptr;
7267 #ifdef CONFIG_FAIR_GROUP_SCHED
7268 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7270 #ifdef CONFIG_RT_GROUP_SCHED
7271 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7274 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7276 #ifdef CONFIG_FAIR_GROUP_SCHED
7277 root_task_group.se = (struct sched_entity **)ptr;
7278 ptr += nr_cpu_ids * sizeof(void **);
7280 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7281 ptr += nr_cpu_ids * sizeof(void **);
7283 #endif /* CONFIG_FAIR_GROUP_SCHED */
7284 #ifdef CONFIG_RT_GROUP_SCHED
7285 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7286 ptr += nr_cpu_ids * sizeof(void **);
7288 root_task_group.rt_rq = (struct rt_rq **)ptr;
7289 ptr += nr_cpu_ids * sizeof(void **);
7291 #endif /* CONFIG_RT_GROUP_SCHED */
7293 #ifdef CONFIG_CPUMASK_OFFSTACK
7294 for_each_possible_cpu(i) {
7295 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7296 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7298 #endif /* CONFIG_CPUMASK_OFFSTACK */
7300 init_rt_bandwidth(&def_rt_bandwidth,
7301 global_rt_period(), global_rt_runtime());
7302 init_dl_bandwidth(&def_dl_bandwidth,
7303 global_rt_period(), global_rt_runtime());
7306 init_defrootdomain();
7309 #ifdef CONFIG_RT_GROUP_SCHED
7310 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7311 global_rt_period(), global_rt_runtime());
7312 #endif /* CONFIG_RT_GROUP_SCHED */
7314 #ifdef CONFIG_CGROUP_SCHED
7315 list_add(&root_task_group.list, &task_groups);
7316 INIT_LIST_HEAD(&root_task_group.children);
7317 INIT_LIST_HEAD(&root_task_group.siblings);
7318 autogroup_init(&init_task);
7320 #endif /* CONFIG_CGROUP_SCHED */
7322 for_each_possible_cpu(i) {
7326 raw_spin_lock_init(&rq->lock);
7328 rq->calc_load_active = 0;
7329 rq->calc_load_update = jiffies + LOAD_FREQ;
7330 init_cfs_rq(&rq->cfs);
7331 init_rt_rq(&rq->rt);
7332 init_dl_rq(&rq->dl);
7333 #ifdef CONFIG_FAIR_GROUP_SCHED
7334 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7335 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7337 * How much cpu bandwidth does root_task_group get?
7339 * In case of task-groups formed thr' the cgroup filesystem, it
7340 * gets 100% of the cpu resources in the system. This overall
7341 * system cpu resource is divided among the tasks of
7342 * root_task_group and its child task-groups in a fair manner,
7343 * based on each entity's (task or task-group's) weight
7344 * (se->load.weight).
7346 * In other words, if root_task_group has 10 tasks of weight
7347 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7348 * then A0's share of the cpu resource is:
7350 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7352 * We achieve this by letting root_task_group's tasks sit
7353 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7355 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7356 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7357 #endif /* CONFIG_FAIR_GROUP_SCHED */
7359 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7360 #ifdef CONFIG_RT_GROUP_SCHED
7361 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7364 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7365 rq->cpu_load[j] = 0;
7367 rq->last_load_update_tick = jiffies;
7372 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7373 rq->balance_callback = NULL;
7374 rq->active_balance = 0;
7375 rq->next_balance = jiffies;
7380 rq->avg_idle = 2*sysctl_sched_migration_cost;
7381 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7383 INIT_LIST_HEAD(&rq->cfs_tasks);
7385 rq_attach_root(rq, &def_root_domain);
7386 #ifdef CONFIG_NO_HZ_COMMON
7389 #ifdef CONFIG_NO_HZ_FULL
7390 rq->last_sched_tick = 0;
7394 atomic_set(&rq->nr_iowait, 0);
7397 set_load_weight(&init_task);
7399 #ifdef CONFIG_PREEMPT_NOTIFIERS
7400 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7404 * The boot idle thread does lazy MMU switching as well:
7406 atomic_inc(&init_mm.mm_count);
7407 enter_lazy_tlb(&init_mm, current);
7410 * During early bootup we pretend to be a normal task:
7412 current->sched_class = &fair_sched_class;
7415 * Make us the idle thread. Technically, schedule() should not be
7416 * called from this thread, however somewhere below it might be,
7417 * but because we are the idle thread, we just pick up running again
7418 * when this runqueue becomes "idle".
7420 init_idle(current, smp_processor_id());
7422 calc_load_update = jiffies + LOAD_FREQ;
7425 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7426 /* May be allocated at isolcpus cmdline parse time */
7427 if (cpu_isolated_map == NULL)
7428 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7429 idle_thread_set_boot_cpu();
7430 set_cpu_rq_start_time();
7432 init_sched_fair_class();
7434 scheduler_running = 1;
7437 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7438 static inline int preempt_count_equals(int preempt_offset)
7440 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7442 return (nested == preempt_offset);
7445 void __might_sleep(const char *file, int line, int preempt_offset)
7448 * Blocking primitives will set (and therefore destroy) current->state,
7449 * since we will exit with TASK_RUNNING make sure we enter with it,
7450 * otherwise we will destroy state.
7452 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7453 "do not call blocking ops when !TASK_RUNNING; "
7454 "state=%lx set at [<%p>] %pS\n",
7456 (void *)current->task_state_change,
7457 (void *)current->task_state_change);
7459 ___might_sleep(file, line, preempt_offset);
7461 EXPORT_SYMBOL(__might_sleep);
7463 void ___might_sleep(const char *file, int line, int preempt_offset)
7465 static unsigned long prev_jiffy; /* ratelimiting */
7467 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7468 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7469 !is_idle_task(current)) ||
7470 system_state != SYSTEM_RUNNING || oops_in_progress)
7472 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7474 prev_jiffy = jiffies;
7477 "BUG: sleeping function called from invalid context at %s:%d\n",
7480 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7481 in_atomic(), irqs_disabled(),
7482 current->pid, current->comm);
7484 if (task_stack_end_corrupted(current))
7485 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7487 debug_show_held_locks(current);
7488 if (irqs_disabled())
7489 print_irqtrace_events(current);
7490 #ifdef CONFIG_DEBUG_PREEMPT
7491 if (!preempt_count_equals(preempt_offset)) {
7492 pr_err("Preemption disabled at:");
7493 print_ip_sym(current->preempt_disable_ip);
7499 EXPORT_SYMBOL(___might_sleep);
7502 #ifdef CONFIG_MAGIC_SYSRQ
7503 void normalize_rt_tasks(void)
7505 struct task_struct *g, *p;
7506 struct sched_attr attr = {
7507 .sched_policy = SCHED_NORMAL,
7510 read_lock(&tasklist_lock);
7511 for_each_process_thread(g, p) {
7513 * Only normalize user tasks:
7515 if (p->flags & PF_KTHREAD)
7518 p->se.exec_start = 0;
7519 #ifdef CONFIG_SCHEDSTATS
7520 p->se.statistics.wait_start = 0;
7521 p->se.statistics.sleep_start = 0;
7522 p->se.statistics.block_start = 0;
7525 if (!dl_task(p) && !rt_task(p)) {
7527 * Renice negative nice level userspace
7530 if (task_nice(p) < 0)
7531 set_user_nice(p, 0);
7535 __sched_setscheduler(p, &attr, false, false);
7537 read_unlock(&tasklist_lock);
7540 #endif /* CONFIG_MAGIC_SYSRQ */
7542 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7544 * These functions are only useful for the IA64 MCA handling, or kdb.
7546 * They can only be called when the whole system has been
7547 * stopped - every CPU needs to be quiescent, and no scheduling
7548 * activity can take place. Using them for anything else would
7549 * be a serious bug, and as a result, they aren't even visible
7550 * under any other configuration.
7554 * curr_task - return the current task for a given cpu.
7555 * @cpu: the processor in question.
7557 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7559 * Return: The current task for @cpu.
7561 struct task_struct *curr_task(int cpu)
7563 return cpu_curr(cpu);
7566 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7570 * set_curr_task - set the current task for a given cpu.
7571 * @cpu: the processor in question.
7572 * @p: the task pointer to set.
7574 * Description: This function must only be used when non-maskable interrupts
7575 * are serviced on a separate stack. It allows the architecture to switch the
7576 * notion of the current task on a cpu in a non-blocking manner. This function
7577 * must be called with all CPU's synchronized, and interrupts disabled, the
7578 * and caller must save the original value of the current task (see
7579 * curr_task() above) and restore that value before reenabling interrupts and
7580 * re-starting the system.
7582 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7584 void set_curr_task(int cpu, struct task_struct *p)
7591 #ifdef CONFIG_CGROUP_SCHED
7592 /* task_group_lock serializes the addition/removal of task groups */
7593 static DEFINE_SPINLOCK(task_group_lock);
7595 static void free_sched_group(struct task_group *tg)
7597 free_fair_sched_group(tg);
7598 free_rt_sched_group(tg);
7603 /* allocate runqueue etc for a new task group */
7604 struct task_group *sched_create_group(struct task_group *parent)
7606 struct task_group *tg;
7608 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7610 return ERR_PTR(-ENOMEM);
7612 if (!alloc_fair_sched_group(tg, parent))
7615 if (!alloc_rt_sched_group(tg, parent))
7621 free_sched_group(tg);
7622 return ERR_PTR(-ENOMEM);
7625 void sched_online_group(struct task_group *tg, struct task_group *parent)
7627 unsigned long flags;
7629 spin_lock_irqsave(&task_group_lock, flags);
7630 list_add_rcu(&tg->list, &task_groups);
7632 WARN_ON(!parent); /* root should already exist */
7634 tg->parent = parent;
7635 INIT_LIST_HEAD(&tg->children);
7636 list_add_rcu(&tg->siblings, &parent->children);
7637 spin_unlock_irqrestore(&task_group_lock, flags);
7640 /* rcu callback to free various structures associated with a task group */
7641 static void free_sched_group_rcu(struct rcu_head *rhp)
7643 /* now it should be safe to free those cfs_rqs */
7644 free_sched_group(container_of(rhp, struct task_group, rcu));
7647 /* Destroy runqueue etc associated with a task group */
7648 void sched_destroy_group(struct task_group *tg)
7650 /* wait for possible concurrent references to cfs_rqs complete */
7651 call_rcu(&tg->rcu, free_sched_group_rcu);
7654 void sched_offline_group(struct task_group *tg)
7656 unsigned long flags;
7659 /* end participation in shares distribution */
7660 for_each_possible_cpu(i)
7661 unregister_fair_sched_group(tg, i);
7663 spin_lock_irqsave(&task_group_lock, flags);
7664 list_del_rcu(&tg->list);
7665 list_del_rcu(&tg->siblings);
7666 spin_unlock_irqrestore(&task_group_lock, flags);
7669 /* change task's runqueue when it moves between groups.
7670 * The caller of this function should have put the task in its new group
7671 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7672 * reflect its new group.
7674 void sched_move_task(struct task_struct *tsk)
7676 struct task_group *tg;
7677 int queued, running;
7678 unsigned long flags;
7681 rq = task_rq_lock(tsk, &flags);
7683 running = task_current(rq, tsk);
7684 queued = task_on_rq_queued(tsk);
7687 dequeue_task(rq, tsk, 0);
7688 if (unlikely(running))
7689 put_prev_task(rq, tsk);
7692 * All callers are synchronized by task_rq_lock(); we do not use RCU
7693 * which is pointless here. Thus, we pass "true" to task_css_check()
7694 * to prevent lockdep warnings.
7696 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7697 struct task_group, css);
7698 tg = autogroup_task_group(tsk, tg);
7699 tsk->sched_task_group = tg;
7701 #ifdef CONFIG_FAIR_GROUP_SCHED
7702 if (tsk->sched_class->task_move_group)
7703 tsk->sched_class->task_move_group(tsk, queued);
7706 set_task_rq(tsk, task_cpu(tsk));
7708 if (unlikely(running))
7709 tsk->sched_class->set_curr_task(rq);
7711 enqueue_task(rq, tsk, 0);
7713 task_rq_unlock(rq, tsk, &flags);
7715 #endif /* CONFIG_CGROUP_SCHED */
7717 #ifdef CONFIG_RT_GROUP_SCHED
7719 * Ensure that the real time constraints are schedulable.
7721 static DEFINE_MUTEX(rt_constraints_mutex);
7723 /* Must be called with tasklist_lock held */
7724 static inline int tg_has_rt_tasks(struct task_group *tg)
7726 struct task_struct *g, *p;
7729 * Autogroups do not have RT tasks; see autogroup_create().
7731 if (task_group_is_autogroup(tg))
7734 for_each_process_thread(g, p) {
7735 if (rt_task(p) && task_group(p) == tg)
7742 struct rt_schedulable_data {
7743 struct task_group *tg;
7748 static int tg_rt_schedulable(struct task_group *tg, void *data)
7750 struct rt_schedulable_data *d = data;
7751 struct task_group *child;
7752 unsigned long total, sum = 0;
7753 u64 period, runtime;
7755 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7756 runtime = tg->rt_bandwidth.rt_runtime;
7759 period = d->rt_period;
7760 runtime = d->rt_runtime;
7764 * Cannot have more runtime than the period.
7766 if (runtime > period && runtime != RUNTIME_INF)
7770 * Ensure we don't starve existing RT tasks.
7772 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7775 total = to_ratio(period, runtime);
7778 * Nobody can have more than the global setting allows.
7780 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7784 * The sum of our children's runtime should not exceed our own.
7786 list_for_each_entry_rcu(child, &tg->children, siblings) {
7787 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7788 runtime = child->rt_bandwidth.rt_runtime;
7790 if (child == d->tg) {
7791 period = d->rt_period;
7792 runtime = d->rt_runtime;
7795 sum += to_ratio(period, runtime);
7804 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7808 struct rt_schedulable_data data = {
7810 .rt_period = period,
7811 .rt_runtime = runtime,
7815 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7821 static int tg_set_rt_bandwidth(struct task_group *tg,
7822 u64 rt_period, u64 rt_runtime)
7827 * Disallowing the root group RT runtime is BAD, it would disallow the
7828 * kernel creating (and or operating) RT threads.
7830 if (tg == &root_task_group && rt_runtime == 0)
7833 /* No period doesn't make any sense. */
7837 mutex_lock(&rt_constraints_mutex);
7838 read_lock(&tasklist_lock);
7839 err = __rt_schedulable(tg, rt_period, rt_runtime);
7843 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7844 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7845 tg->rt_bandwidth.rt_runtime = rt_runtime;
7847 for_each_possible_cpu(i) {
7848 struct rt_rq *rt_rq = tg->rt_rq[i];
7850 raw_spin_lock(&rt_rq->rt_runtime_lock);
7851 rt_rq->rt_runtime = rt_runtime;
7852 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7854 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7856 read_unlock(&tasklist_lock);
7857 mutex_unlock(&rt_constraints_mutex);
7862 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7864 u64 rt_runtime, rt_period;
7866 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7867 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7868 if (rt_runtime_us < 0)
7869 rt_runtime = RUNTIME_INF;
7871 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7874 static long sched_group_rt_runtime(struct task_group *tg)
7878 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7881 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7882 do_div(rt_runtime_us, NSEC_PER_USEC);
7883 return rt_runtime_us;
7886 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7888 u64 rt_runtime, rt_period;
7890 rt_period = rt_period_us * NSEC_PER_USEC;
7891 rt_runtime = tg->rt_bandwidth.rt_runtime;
7893 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7896 static long sched_group_rt_period(struct task_group *tg)
7900 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7901 do_div(rt_period_us, NSEC_PER_USEC);
7902 return rt_period_us;
7904 #endif /* CONFIG_RT_GROUP_SCHED */
7906 #ifdef CONFIG_RT_GROUP_SCHED
7907 static int sched_rt_global_constraints(void)
7911 mutex_lock(&rt_constraints_mutex);
7912 read_lock(&tasklist_lock);
7913 ret = __rt_schedulable(NULL, 0, 0);
7914 read_unlock(&tasklist_lock);
7915 mutex_unlock(&rt_constraints_mutex);
7920 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7922 /* Don't accept realtime tasks when there is no way for them to run */
7923 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7929 #else /* !CONFIG_RT_GROUP_SCHED */
7930 static int sched_rt_global_constraints(void)
7932 unsigned long flags;
7935 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7936 for_each_possible_cpu(i) {
7937 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7939 raw_spin_lock(&rt_rq->rt_runtime_lock);
7940 rt_rq->rt_runtime = global_rt_runtime();
7941 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7943 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7947 #endif /* CONFIG_RT_GROUP_SCHED */
7949 static int sched_dl_global_validate(void)
7951 u64 runtime = global_rt_runtime();
7952 u64 period = global_rt_period();
7953 u64 new_bw = to_ratio(period, runtime);
7956 unsigned long flags;
7959 * Here we want to check the bandwidth not being set to some
7960 * value smaller than the currently allocated bandwidth in
7961 * any of the root_domains.
7963 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7964 * cycling on root_domains... Discussion on different/better
7965 * solutions is welcome!
7967 for_each_possible_cpu(cpu) {
7968 rcu_read_lock_sched();
7969 dl_b = dl_bw_of(cpu);
7971 raw_spin_lock_irqsave(&dl_b->lock, flags);
7972 if (new_bw < dl_b->total_bw)
7974 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7976 rcu_read_unlock_sched();
7985 static void sched_dl_do_global(void)
7990 unsigned long flags;
7992 def_dl_bandwidth.dl_period = global_rt_period();
7993 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7995 if (global_rt_runtime() != RUNTIME_INF)
7996 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7999 * FIXME: As above...
8001 for_each_possible_cpu(cpu) {
8002 rcu_read_lock_sched();
8003 dl_b = dl_bw_of(cpu);
8005 raw_spin_lock_irqsave(&dl_b->lock, flags);
8007 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8009 rcu_read_unlock_sched();
8013 static int sched_rt_global_validate(void)
8015 if (sysctl_sched_rt_period <= 0)
8018 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8019 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8025 static void sched_rt_do_global(void)
8027 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8028 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8031 int sched_rt_handler(struct ctl_table *table, int write,
8032 void __user *buffer, size_t *lenp,
8035 int old_period, old_runtime;
8036 static DEFINE_MUTEX(mutex);
8040 old_period = sysctl_sched_rt_period;
8041 old_runtime = sysctl_sched_rt_runtime;
8043 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8045 if (!ret && write) {
8046 ret = sched_rt_global_validate();
8050 ret = sched_dl_global_validate();
8054 ret = sched_rt_global_constraints();
8058 sched_rt_do_global();
8059 sched_dl_do_global();
8063 sysctl_sched_rt_period = old_period;
8064 sysctl_sched_rt_runtime = old_runtime;
8066 mutex_unlock(&mutex);
8071 int sched_rr_handler(struct ctl_table *table, int write,
8072 void __user *buffer, size_t *lenp,
8076 static DEFINE_MUTEX(mutex);
8079 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8080 /* make sure that internally we keep jiffies */
8081 /* also, writing zero resets timeslice to default */
8082 if (!ret && write) {
8083 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8084 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8086 mutex_unlock(&mutex);
8090 #ifdef CONFIG_CGROUP_SCHED
8092 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8094 return css ? container_of(css, struct task_group, css) : NULL;
8097 static struct cgroup_subsys_state *
8098 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8100 struct task_group *parent = css_tg(parent_css);
8101 struct task_group *tg;
8104 /* This is early initialization for the top cgroup */
8105 return &root_task_group.css;
8108 tg = sched_create_group(parent);
8110 return ERR_PTR(-ENOMEM);
8115 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8117 struct task_group *tg = css_tg(css);
8118 struct task_group *parent = css_tg(css->parent);
8121 sched_online_group(tg, parent);
8125 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8127 struct task_group *tg = css_tg(css);
8129 sched_destroy_group(tg);
8132 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8134 struct task_group *tg = css_tg(css);
8136 sched_offline_group(tg);
8139 static void cpu_cgroup_fork(struct task_struct *task)
8141 sched_move_task(task);
8144 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8145 struct cgroup_taskset *tset)
8147 struct task_struct *task;
8149 cgroup_taskset_for_each(task, tset) {
8150 #ifdef CONFIG_RT_GROUP_SCHED
8151 if (!sched_rt_can_attach(css_tg(css), task))
8154 /* We don't support RT-tasks being in separate groups */
8155 if (task->sched_class != &fair_sched_class)
8162 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8163 struct cgroup_taskset *tset)
8165 struct task_struct *task;
8167 cgroup_taskset_for_each(task, tset)
8168 sched_move_task(task);
8171 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8172 struct cgroup_subsys_state *old_css,
8173 struct task_struct *task)
8176 * cgroup_exit() is called in the copy_process() failure path.
8177 * Ignore this case since the task hasn't ran yet, this avoids
8178 * trying to poke a half freed task state from generic code.
8180 if (!(task->flags & PF_EXITING))
8183 sched_move_task(task);
8186 #ifdef CONFIG_FAIR_GROUP_SCHED
8187 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8188 struct cftype *cftype, u64 shareval)
8190 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8193 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8196 struct task_group *tg = css_tg(css);
8198 return (u64) scale_load_down(tg->shares);
8201 #ifdef CONFIG_CFS_BANDWIDTH
8202 static DEFINE_MUTEX(cfs_constraints_mutex);
8204 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8205 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8207 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8209 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8211 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8212 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8214 if (tg == &root_task_group)
8218 * Ensure we have at some amount of bandwidth every period. This is
8219 * to prevent reaching a state of large arrears when throttled via
8220 * entity_tick() resulting in prolonged exit starvation.
8222 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8226 * Likewise, bound things on the otherside by preventing insane quota
8227 * periods. This also allows us to normalize in computing quota
8230 if (period > max_cfs_quota_period)
8234 * Prevent race between setting of cfs_rq->runtime_enabled and
8235 * unthrottle_offline_cfs_rqs().
8238 mutex_lock(&cfs_constraints_mutex);
8239 ret = __cfs_schedulable(tg, period, quota);
8243 runtime_enabled = quota != RUNTIME_INF;
8244 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8246 * If we need to toggle cfs_bandwidth_used, off->on must occur
8247 * before making related changes, and on->off must occur afterwards
8249 if (runtime_enabled && !runtime_was_enabled)
8250 cfs_bandwidth_usage_inc();
8251 raw_spin_lock_irq(&cfs_b->lock);
8252 cfs_b->period = ns_to_ktime(period);
8253 cfs_b->quota = quota;
8255 __refill_cfs_bandwidth_runtime(cfs_b);
8256 /* restart the period timer (if active) to handle new period expiry */
8257 if (runtime_enabled)
8258 start_cfs_bandwidth(cfs_b);
8259 raw_spin_unlock_irq(&cfs_b->lock);
8261 for_each_online_cpu(i) {
8262 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8263 struct rq *rq = cfs_rq->rq;
8265 raw_spin_lock_irq(&rq->lock);
8266 cfs_rq->runtime_enabled = runtime_enabled;
8267 cfs_rq->runtime_remaining = 0;
8269 if (cfs_rq->throttled)
8270 unthrottle_cfs_rq(cfs_rq);
8271 raw_spin_unlock_irq(&rq->lock);
8273 if (runtime_was_enabled && !runtime_enabled)
8274 cfs_bandwidth_usage_dec();
8276 mutex_unlock(&cfs_constraints_mutex);
8282 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8286 period = ktime_to_ns(tg->cfs_bandwidth.period);
8287 if (cfs_quota_us < 0)
8288 quota = RUNTIME_INF;
8290 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8292 return tg_set_cfs_bandwidth(tg, period, quota);
8295 long tg_get_cfs_quota(struct task_group *tg)
8299 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8302 quota_us = tg->cfs_bandwidth.quota;
8303 do_div(quota_us, NSEC_PER_USEC);
8308 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8312 period = (u64)cfs_period_us * NSEC_PER_USEC;
8313 quota = tg->cfs_bandwidth.quota;
8315 return tg_set_cfs_bandwidth(tg, period, quota);
8318 long tg_get_cfs_period(struct task_group *tg)
8322 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8323 do_div(cfs_period_us, NSEC_PER_USEC);
8325 return cfs_period_us;
8328 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8331 return tg_get_cfs_quota(css_tg(css));
8334 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8335 struct cftype *cftype, s64 cfs_quota_us)
8337 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8340 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8343 return tg_get_cfs_period(css_tg(css));
8346 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8347 struct cftype *cftype, u64 cfs_period_us)
8349 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8352 struct cfs_schedulable_data {
8353 struct task_group *tg;
8358 * normalize group quota/period to be quota/max_period
8359 * note: units are usecs
8361 static u64 normalize_cfs_quota(struct task_group *tg,
8362 struct cfs_schedulable_data *d)
8370 period = tg_get_cfs_period(tg);
8371 quota = tg_get_cfs_quota(tg);
8374 /* note: these should typically be equivalent */
8375 if (quota == RUNTIME_INF || quota == -1)
8378 return to_ratio(period, quota);
8381 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8383 struct cfs_schedulable_data *d = data;
8384 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8385 s64 quota = 0, parent_quota = -1;
8388 quota = RUNTIME_INF;
8390 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8392 quota = normalize_cfs_quota(tg, d);
8393 parent_quota = parent_b->hierarchical_quota;
8396 * ensure max(child_quota) <= parent_quota, inherit when no
8399 if (quota == RUNTIME_INF)
8400 quota = parent_quota;
8401 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8404 cfs_b->hierarchical_quota = quota;
8409 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8412 struct cfs_schedulable_data data = {
8418 if (quota != RUNTIME_INF) {
8419 do_div(data.period, NSEC_PER_USEC);
8420 do_div(data.quota, NSEC_PER_USEC);
8424 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8430 static int cpu_stats_show(struct seq_file *sf, void *v)
8432 struct task_group *tg = css_tg(seq_css(sf));
8433 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8435 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8436 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8437 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8441 #endif /* CONFIG_CFS_BANDWIDTH */
8442 #endif /* CONFIG_FAIR_GROUP_SCHED */
8444 #ifdef CONFIG_RT_GROUP_SCHED
8445 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8446 struct cftype *cft, s64 val)
8448 return sched_group_set_rt_runtime(css_tg(css), val);
8451 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8454 return sched_group_rt_runtime(css_tg(css));
8457 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8458 struct cftype *cftype, u64 rt_period_us)
8460 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8463 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8466 return sched_group_rt_period(css_tg(css));
8468 #endif /* CONFIG_RT_GROUP_SCHED */
8470 static struct cftype cpu_files[] = {
8471 #ifdef CONFIG_FAIR_GROUP_SCHED
8474 .read_u64 = cpu_shares_read_u64,
8475 .write_u64 = cpu_shares_write_u64,
8478 #ifdef CONFIG_CFS_BANDWIDTH
8480 .name = "cfs_quota_us",
8481 .read_s64 = cpu_cfs_quota_read_s64,
8482 .write_s64 = cpu_cfs_quota_write_s64,
8485 .name = "cfs_period_us",
8486 .read_u64 = cpu_cfs_period_read_u64,
8487 .write_u64 = cpu_cfs_period_write_u64,
8491 .seq_show = cpu_stats_show,
8494 #ifdef CONFIG_RT_GROUP_SCHED
8496 .name = "rt_runtime_us",
8497 .read_s64 = cpu_rt_runtime_read,
8498 .write_s64 = cpu_rt_runtime_write,
8501 .name = "rt_period_us",
8502 .read_u64 = cpu_rt_period_read_uint,
8503 .write_u64 = cpu_rt_period_write_uint,
8509 struct cgroup_subsys cpu_cgrp_subsys = {
8510 .css_alloc = cpu_cgroup_css_alloc,
8511 .css_free = cpu_cgroup_css_free,
8512 .css_online = cpu_cgroup_css_online,
8513 .css_offline = cpu_cgroup_css_offline,
8514 .fork = cpu_cgroup_fork,
8515 .can_attach = cpu_cgroup_can_attach,
8516 .attach = cpu_cgroup_attach,
8517 .exit = cpu_cgroup_exit,
8518 .legacy_cftypes = cpu_files,
8522 #endif /* CONFIG_CGROUP_SCHED */
8524 void dump_cpu_task(int cpu)
8526 pr_info("Task dump for CPU %d:\n", cpu);
8527 sched_show_task(cpu_curr(cpu));