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 #ifdef smp_mb__before_atomic
94 void __smp_mb__before_atomic(void)
96 smp_mb__before_atomic();
98 EXPORT_SYMBOL(__smp_mb__before_atomic);
101 #ifdef smp_mb__after_atomic
102 void __smp_mb__after_atomic(void)
104 smp_mb__after_atomic();
106 EXPORT_SYMBOL(__smp_mb__after_atomic);
109 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
112 ktime_t soft, hard, now;
115 if (hrtimer_active(period_timer))
118 now = hrtimer_cb_get_time(period_timer);
119 hrtimer_forward(period_timer, now, period);
121 soft = hrtimer_get_softexpires(period_timer);
122 hard = hrtimer_get_expires(period_timer);
123 delta = ktime_to_ns(ktime_sub(hard, soft));
124 __hrtimer_start_range_ns(period_timer, soft, delta,
125 HRTIMER_MODE_ABS_PINNED, 0);
129 DEFINE_MUTEX(sched_domains_mutex);
130 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
132 static void update_rq_clock_task(struct rq *rq, s64 delta);
134 void update_rq_clock(struct rq *rq)
138 if (rq->skip_clock_update > 0)
141 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
145 update_rq_clock_task(rq, delta);
149 * Debugging: various feature bits
152 #define SCHED_FEAT(name, enabled) \
153 (1UL << __SCHED_FEAT_##name) * enabled |
155 const_debug unsigned int sysctl_sched_features =
156 #include "features.h"
161 #ifdef CONFIG_SCHED_DEBUG
162 #define SCHED_FEAT(name, enabled) \
165 static const char * const sched_feat_names[] = {
166 #include "features.h"
171 static int sched_feat_show(struct seq_file *m, void *v)
175 for (i = 0; i < __SCHED_FEAT_NR; i++) {
176 if (!(sysctl_sched_features & (1UL << i)))
178 seq_printf(m, "%s ", sched_feat_names[i]);
185 #ifdef HAVE_JUMP_LABEL
187 #define jump_label_key__true STATIC_KEY_INIT_TRUE
188 #define jump_label_key__false STATIC_KEY_INIT_FALSE
190 #define SCHED_FEAT(name, enabled) \
191 jump_label_key__##enabled ,
193 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
194 #include "features.h"
199 static void sched_feat_disable(int i)
201 if (static_key_enabled(&sched_feat_keys[i]))
202 static_key_slow_dec(&sched_feat_keys[i]);
205 static void sched_feat_enable(int i)
207 if (!static_key_enabled(&sched_feat_keys[i]))
208 static_key_slow_inc(&sched_feat_keys[i]);
211 static void sched_feat_disable(int i) { };
212 static void sched_feat_enable(int i) { };
213 #endif /* HAVE_JUMP_LABEL */
215 static int sched_feat_set(char *cmp)
220 if (strncmp(cmp, "NO_", 3) == 0) {
225 for (i = 0; i < __SCHED_FEAT_NR; i++) {
226 if (strcmp(cmp, sched_feat_names[i]) == 0) {
228 sysctl_sched_features &= ~(1UL << i);
229 sched_feat_disable(i);
231 sysctl_sched_features |= (1UL << i);
232 sched_feat_enable(i);
242 sched_feat_write(struct file *filp, const char __user *ubuf,
243 size_t cnt, loff_t *ppos)
253 if (copy_from_user(&buf, ubuf, cnt))
259 /* Ensure the static_key remains in a consistent state */
260 inode = file_inode(filp);
261 mutex_lock(&inode->i_mutex);
262 i = sched_feat_set(cmp);
263 mutex_unlock(&inode->i_mutex);
264 if (i == __SCHED_FEAT_NR)
272 static int sched_feat_open(struct inode *inode, struct file *filp)
274 return single_open(filp, sched_feat_show, NULL);
277 static const struct file_operations sched_feat_fops = {
278 .open = sched_feat_open,
279 .write = sched_feat_write,
282 .release = single_release,
285 static __init int sched_init_debug(void)
287 debugfs_create_file("sched_features", 0644, NULL, NULL,
292 late_initcall(sched_init_debug);
293 #endif /* CONFIG_SCHED_DEBUG */
296 * Number of tasks to iterate in a single balance run.
297 * Limited because this is done with IRQs disabled.
299 const_debug unsigned int sysctl_sched_nr_migrate = 32;
302 * period over which we average the RT time consumption, measured
307 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
310 * period over which we measure -rt task cpu usage in us.
313 unsigned int sysctl_sched_rt_period = 1000000;
315 __read_mostly int scheduler_running;
318 * part of the period that we allow rt tasks to run in us.
321 int sysctl_sched_rt_runtime = 950000;
324 * __task_rq_lock - lock the rq @p resides on.
326 static inline struct rq *__task_rq_lock(struct task_struct *p)
331 lockdep_assert_held(&p->pi_lock);
335 raw_spin_lock(&rq->lock);
336 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
338 raw_spin_unlock(&rq->lock);
340 while (unlikely(task_on_rq_migrating(p)))
346 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
348 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
349 __acquires(p->pi_lock)
355 raw_spin_lock_irqsave(&p->pi_lock, *flags);
357 raw_spin_lock(&rq->lock);
358 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
360 raw_spin_unlock(&rq->lock);
361 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
363 while (unlikely(task_on_rq_migrating(p)))
368 static void __task_rq_unlock(struct rq *rq)
371 raw_spin_unlock(&rq->lock);
375 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
377 __releases(p->pi_lock)
379 raw_spin_unlock(&rq->lock);
380 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
384 * this_rq_lock - lock this runqueue and disable interrupts.
386 static struct rq *this_rq_lock(void)
393 raw_spin_lock(&rq->lock);
398 #ifdef CONFIG_SCHED_HRTICK
400 * Use HR-timers to deliver accurate preemption points.
403 static void hrtick_clear(struct rq *rq)
405 if (hrtimer_active(&rq->hrtick_timer))
406 hrtimer_cancel(&rq->hrtick_timer);
410 * High-resolution timer tick.
411 * Runs from hardirq context with interrupts disabled.
413 static enum hrtimer_restart hrtick(struct hrtimer *timer)
415 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
417 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
419 raw_spin_lock(&rq->lock);
421 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
422 raw_spin_unlock(&rq->lock);
424 return HRTIMER_NORESTART;
429 static int __hrtick_restart(struct rq *rq)
431 struct hrtimer *timer = &rq->hrtick_timer;
432 ktime_t time = hrtimer_get_softexpires(timer);
434 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
438 * called from hardirq (IPI) context
440 static void __hrtick_start(void *arg)
444 raw_spin_lock(&rq->lock);
445 __hrtick_restart(rq);
446 rq->hrtick_csd_pending = 0;
447 raw_spin_unlock(&rq->lock);
451 * Called to set the hrtick timer state.
453 * called with rq->lock held and irqs disabled
455 void hrtick_start(struct rq *rq, u64 delay)
457 struct hrtimer *timer = &rq->hrtick_timer;
462 * Don't schedule slices shorter than 10000ns, that just
463 * doesn't make sense and can cause timer DoS.
465 delta = max_t(s64, delay, 10000LL);
466 time = ktime_add_ns(timer->base->get_time(), delta);
468 hrtimer_set_expires(timer, time);
470 if (rq == this_rq()) {
471 __hrtick_restart(rq);
472 } else if (!rq->hrtick_csd_pending) {
473 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
474 rq->hrtick_csd_pending = 1;
479 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
481 int cpu = (int)(long)hcpu;
484 case CPU_UP_CANCELED:
485 case CPU_UP_CANCELED_FROZEN:
486 case CPU_DOWN_PREPARE:
487 case CPU_DOWN_PREPARE_FROZEN:
489 case CPU_DEAD_FROZEN:
490 hrtick_clear(cpu_rq(cpu));
497 static __init void init_hrtick(void)
499 hotcpu_notifier(hotplug_hrtick, 0);
503 * Called to set the hrtick timer state.
505 * called with rq->lock held and irqs disabled
507 void hrtick_start(struct rq *rq, u64 delay)
509 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
510 HRTIMER_MODE_REL_PINNED, 0);
513 static inline void init_hrtick(void)
516 #endif /* CONFIG_SMP */
518 static void init_rq_hrtick(struct rq *rq)
521 rq->hrtick_csd_pending = 0;
523 rq->hrtick_csd.flags = 0;
524 rq->hrtick_csd.func = __hrtick_start;
525 rq->hrtick_csd.info = rq;
528 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
529 rq->hrtick_timer.function = hrtick;
531 #else /* CONFIG_SCHED_HRTICK */
532 static inline void hrtick_clear(struct rq *rq)
536 static inline void init_rq_hrtick(struct rq *rq)
540 static inline void init_hrtick(void)
543 #endif /* CONFIG_SCHED_HRTICK */
546 * cmpxchg based fetch_or, macro so it works for different integer types
548 #define fetch_or(ptr, val) \
549 ({ typeof(*(ptr)) __old, __val = *(ptr); \
551 __old = cmpxchg((ptr), __val, __val | (val)); \
552 if (__old == __val) \
559 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
561 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
562 * this avoids any races wrt polling state changes and thereby avoids
565 static bool set_nr_and_not_polling(struct task_struct *p)
567 struct thread_info *ti = task_thread_info(p);
568 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
572 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
574 * If this returns true, then the idle task promises to call
575 * sched_ttwu_pending() and reschedule soon.
577 static bool set_nr_if_polling(struct task_struct *p)
579 struct thread_info *ti = task_thread_info(p);
580 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
583 if (!(val & _TIF_POLLING_NRFLAG))
585 if (val & _TIF_NEED_RESCHED)
587 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
596 static bool set_nr_and_not_polling(struct task_struct *p)
598 set_tsk_need_resched(p);
603 static bool set_nr_if_polling(struct task_struct *p)
611 * resched_curr - mark rq's current task 'to be rescheduled now'.
613 * On UP this means the setting of the need_resched flag, on SMP it
614 * might also involve a cross-CPU call to trigger the scheduler on
617 void resched_curr(struct rq *rq)
619 struct task_struct *curr = rq->curr;
622 lockdep_assert_held(&rq->lock);
624 if (test_tsk_need_resched(curr))
629 if (cpu == smp_processor_id()) {
630 set_tsk_need_resched(curr);
631 set_preempt_need_resched();
635 if (set_nr_and_not_polling(curr))
636 smp_send_reschedule(cpu);
638 trace_sched_wake_idle_without_ipi(cpu);
641 void resched_cpu(int cpu)
643 struct rq *rq = cpu_rq(cpu);
646 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
649 raw_spin_unlock_irqrestore(&rq->lock, flags);
653 #ifdef CONFIG_NO_HZ_COMMON
655 * In the semi idle case, use the nearest busy cpu for migrating timers
656 * from an idle cpu. This is good for power-savings.
658 * We don't do similar optimization for completely idle system, as
659 * selecting an idle cpu will add more delays to the timers than intended
660 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
662 int get_nohz_timer_target(int pinned)
664 int cpu = smp_processor_id();
666 struct sched_domain *sd;
668 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
672 for_each_domain(cpu, sd) {
673 for_each_cpu(i, sched_domain_span(sd)) {
685 * When add_timer_on() enqueues a timer into the timer wheel of an
686 * idle CPU then this timer might expire before the next timer event
687 * which is scheduled to wake up that CPU. In case of a completely
688 * idle system the next event might even be infinite time into the
689 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
690 * leaves the inner idle loop so the newly added timer is taken into
691 * account when the CPU goes back to idle and evaluates the timer
692 * wheel for the next timer event.
694 static void wake_up_idle_cpu(int cpu)
696 struct rq *rq = cpu_rq(cpu);
698 if (cpu == smp_processor_id())
701 if (set_nr_and_not_polling(rq->idle))
702 smp_send_reschedule(cpu);
704 trace_sched_wake_idle_without_ipi(cpu);
707 static bool wake_up_full_nohz_cpu(int cpu)
710 * We just need the target to call irq_exit() and re-evaluate
711 * the next tick. The nohz full kick at least implies that.
712 * If needed we can still optimize that later with an
715 if (tick_nohz_full_cpu(cpu)) {
716 if (cpu != smp_processor_id() ||
717 tick_nohz_tick_stopped())
718 tick_nohz_full_kick_cpu(cpu);
725 void wake_up_nohz_cpu(int cpu)
727 if (!wake_up_full_nohz_cpu(cpu))
728 wake_up_idle_cpu(cpu);
731 static inline bool got_nohz_idle_kick(void)
733 int cpu = smp_processor_id();
735 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
738 if (idle_cpu(cpu) && !need_resched())
742 * We can't run Idle Load Balance on this CPU for this time so we
743 * cancel it and clear NOHZ_BALANCE_KICK
745 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
749 #else /* CONFIG_NO_HZ_COMMON */
751 static inline bool got_nohz_idle_kick(void)
756 #endif /* CONFIG_NO_HZ_COMMON */
758 #ifdef CONFIG_NO_HZ_FULL
759 bool sched_can_stop_tick(void)
762 * More than one running task need preemption.
763 * nr_running update is assumed to be visible
764 * after IPI is sent from wakers.
766 if (this_rq()->nr_running > 1)
771 #endif /* CONFIG_NO_HZ_FULL */
773 void sched_avg_update(struct rq *rq)
775 s64 period = sched_avg_period();
777 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
779 * Inline assembly required to prevent the compiler
780 * optimising this loop into a divmod call.
781 * See __iter_div_u64_rem() for another example of this.
783 asm("" : "+rm" (rq->age_stamp));
784 rq->age_stamp += period;
789 #endif /* CONFIG_SMP */
791 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
792 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
794 * Iterate task_group tree rooted at *from, calling @down when first entering a
795 * node and @up when leaving it for the final time.
797 * Caller must hold rcu_lock or sufficient equivalent.
799 int walk_tg_tree_from(struct task_group *from,
800 tg_visitor down, tg_visitor up, void *data)
802 struct task_group *parent, *child;
808 ret = (*down)(parent, data);
811 list_for_each_entry_rcu(child, &parent->children, siblings) {
818 ret = (*up)(parent, data);
819 if (ret || parent == from)
823 parent = parent->parent;
830 int tg_nop(struct task_group *tg, void *data)
836 static void set_load_weight(struct task_struct *p)
838 int prio = p->static_prio - MAX_RT_PRIO;
839 struct load_weight *load = &p->se.load;
842 * SCHED_IDLE tasks get minimal weight:
844 if (p->policy == SCHED_IDLE) {
845 load->weight = scale_load(WEIGHT_IDLEPRIO);
846 load->inv_weight = WMULT_IDLEPRIO;
850 load->weight = scale_load(prio_to_weight[prio]);
851 load->inv_weight = prio_to_wmult[prio];
854 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
857 sched_info_queued(rq, p);
858 p->sched_class->enqueue_task(rq, p, flags);
861 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
864 sched_info_dequeued(rq, p);
865 p->sched_class->dequeue_task(rq, p, flags);
868 void activate_task(struct rq *rq, struct task_struct *p, int flags)
870 if (task_contributes_to_load(p))
871 rq->nr_uninterruptible--;
873 enqueue_task(rq, p, flags);
876 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
878 if (task_contributes_to_load(p))
879 rq->nr_uninterruptible++;
881 dequeue_task(rq, p, flags);
884 static void update_rq_clock_task(struct rq *rq, s64 delta)
887 * In theory, the compile should just see 0 here, and optimize out the call
888 * to sched_rt_avg_update. But I don't trust it...
890 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
891 s64 steal = 0, irq_delta = 0;
893 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
894 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
897 * Since irq_time is only updated on {soft,}irq_exit, we might run into
898 * this case when a previous update_rq_clock() happened inside a
901 * When this happens, we stop ->clock_task and only update the
902 * prev_irq_time stamp to account for the part that fit, so that a next
903 * update will consume the rest. This ensures ->clock_task is
906 * It does however cause some slight miss-attribution of {soft,}irq
907 * time, a more accurate solution would be to update the irq_time using
908 * the current rq->clock timestamp, except that would require using
911 if (irq_delta > delta)
914 rq->prev_irq_time += irq_delta;
917 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
918 if (static_key_false((¶virt_steal_rq_enabled))) {
919 steal = paravirt_steal_clock(cpu_of(rq));
920 steal -= rq->prev_steal_time_rq;
922 if (unlikely(steal > delta))
925 rq->prev_steal_time_rq += steal;
930 rq->clock_task += delta;
932 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
933 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
934 sched_rt_avg_update(rq, irq_delta + steal);
938 void sched_set_stop_task(int cpu, struct task_struct *stop)
940 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
941 struct task_struct *old_stop = cpu_rq(cpu)->stop;
945 * Make it appear like a SCHED_FIFO task, its something
946 * userspace knows about and won't get confused about.
948 * Also, it will make PI more or less work without too
949 * much confusion -- but then, stop work should not
950 * rely on PI working anyway.
952 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
954 stop->sched_class = &stop_sched_class;
957 cpu_rq(cpu)->stop = stop;
961 * Reset it back to a normal scheduling class so that
962 * it can die in pieces.
964 old_stop->sched_class = &rt_sched_class;
969 * __normal_prio - return the priority that is based on the static prio
971 static inline int __normal_prio(struct task_struct *p)
973 return p->static_prio;
977 * Calculate the expected normal priority: i.e. priority
978 * without taking RT-inheritance into account. Might be
979 * boosted by interactivity modifiers. Changes upon fork,
980 * setprio syscalls, and whenever the interactivity
981 * estimator recalculates.
983 static inline int normal_prio(struct task_struct *p)
987 if (task_has_dl_policy(p))
988 prio = MAX_DL_PRIO-1;
989 else if (task_has_rt_policy(p))
990 prio = MAX_RT_PRIO-1 - p->rt_priority;
992 prio = __normal_prio(p);
997 * Calculate the current priority, i.e. the priority
998 * taken into account by the scheduler. This value might
999 * be boosted by RT tasks, or might be boosted by
1000 * interactivity modifiers. Will be RT if the task got
1001 * RT-boosted. If not then it returns p->normal_prio.
1003 static int effective_prio(struct task_struct *p)
1005 p->normal_prio = normal_prio(p);
1007 * If we are RT tasks or we were boosted to RT priority,
1008 * keep the priority unchanged. Otherwise, update priority
1009 * to the normal priority:
1011 if (!rt_prio(p->prio))
1012 return p->normal_prio;
1017 * task_curr - is this task currently executing on a CPU?
1018 * @p: the task in question.
1020 * Return: 1 if the task is currently executing. 0 otherwise.
1022 inline int task_curr(const struct task_struct *p)
1024 return cpu_curr(task_cpu(p)) == p;
1027 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1028 const struct sched_class *prev_class,
1031 if (prev_class != p->sched_class) {
1032 if (prev_class->switched_from)
1033 prev_class->switched_from(rq, p);
1034 p->sched_class->switched_to(rq, p);
1035 } else if (oldprio != p->prio || dl_task(p))
1036 p->sched_class->prio_changed(rq, p, oldprio);
1039 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1041 const struct sched_class *class;
1043 if (p->sched_class == rq->curr->sched_class) {
1044 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1046 for_each_class(class) {
1047 if (class == rq->curr->sched_class)
1049 if (class == p->sched_class) {
1057 * A queue event has occurred, and we're going to schedule. In
1058 * this case, we can save a useless back to back clock update.
1060 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1061 rq->skip_clock_update = 1;
1065 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1067 #ifdef CONFIG_SCHED_DEBUG
1069 * We should never call set_task_cpu() on a blocked task,
1070 * ttwu() will sort out the placement.
1072 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1073 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1075 #ifdef CONFIG_LOCKDEP
1077 * The caller should hold either p->pi_lock or rq->lock, when changing
1078 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1080 * sched_move_task() holds both and thus holding either pins the cgroup,
1083 * Furthermore, all task_rq users should acquire both locks, see
1086 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1087 lockdep_is_held(&task_rq(p)->lock)));
1091 trace_sched_migrate_task(p, new_cpu);
1093 if (task_cpu(p) != new_cpu) {
1094 if (p->sched_class->migrate_task_rq)
1095 p->sched_class->migrate_task_rq(p, new_cpu);
1096 p->se.nr_migrations++;
1097 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1100 __set_task_cpu(p, new_cpu);
1103 static void __migrate_swap_task(struct task_struct *p, int cpu)
1105 if (task_on_rq_queued(p)) {
1106 struct rq *src_rq, *dst_rq;
1108 src_rq = task_rq(p);
1109 dst_rq = cpu_rq(cpu);
1111 deactivate_task(src_rq, p, 0);
1112 set_task_cpu(p, cpu);
1113 activate_task(dst_rq, p, 0);
1114 check_preempt_curr(dst_rq, p, 0);
1117 * Task isn't running anymore; make it appear like we migrated
1118 * it before it went to sleep. This means on wakeup we make the
1119 * previous cpu our targer instead of where it really is.
1125 struct migration_swap_arg {
1126 struct task_struct *src_task, *dst_task;
1127 int src_cpu, dst_cpu;
1130 static int migrate_swap_stop(void *data)
1132 struct migration_swap_arg *arg = data;
1133 struct rq *src_rq, *dst_rq;
1136 src_rq = cpu_rq(arg->src_cpu);
1137 dst_rq = cpu_rq(arg->dst_cpu);
1139 double_raw_lock(&arg->src_task->pi_lock,
1140 &arg->dst_task->pi_lock);
1141 double_rq_lock(src_rq, dst_rq);
1142 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1145 if (task_cpu(arg->src_task) != arg->src_cpu)
1148 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1151 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1154 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1155 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1160 double_rq_unlock(src_rq, dst_rq);
1161 raw_spin_unlock(&arg->dst_task->pi_lock);
1162 raw_spin_unlock(&arg->src_task->pi_lock);
1168 * Cross migrate two tasks
1170 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1172 struct migration_swap_arg arg;
1175 arg = (struct migration_swap_arg){
1177 .src_cpu = task_cpu(cur),
1179 .dst_cpu = task_cpu(p),
1182 if (arg.src_cpu == arg.dst_cpu)
1186 * These three tests are all lockless; this is OK since all of them
1187 * will be re-checked with proper locks held further down the line.
1189 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1192 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1195 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1198 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1199 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1205 struct migration_arg {
1206 struct task_struct *task;
1210 static int migration_cpu_stop(void *data);
1213 * wait_task_inactive - wait for a thread to unschedule.
1215 * If @match_state is nonzero, it's the @p->state value just checked and
1216 * not expected to change. If it changes, i.e. @p might have woken up,
1217 * then return zero. When we succeed in waiting for @p to be off its CPU,
1218 * we return a positive number (its total switch count). If a second call
1219 * a short while later returns the same number, the caller can be sure that
1220 * @p has remained unscheduled the whole time.
1222 * The caller must ensure that the task *will* unschedule sometime soon,
1223 * else this function might spin for a *long* time. This function can't
1224 * be called with interrupts off, or it may introduce deadlock with
1225 * smp_call_function() if an IPI is sent by the same process we are
1226 * waiting to become inactive.
1228 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1230 unsigned long flags;
1231 int running, queued;
1237 * We do the initial early heuristics without holding
1238 * any task-queue locks at all. We'll only try to get
1239 * the runqueue lock when things look like they will
1245 * If the task is actively running on another CPU
1246 * still, just relax and busy-wait without holding
1249 * NOTE! Since we don't hold any locks, it's not
1250 * even sure that "rq" stays as the right runqueue!
1251 * But we don't care, since "task_running()" will
1252 * return false if the runqueue has changed and p
1253 * is actually now running somewhere else!
1255 while (task_running(rq, p)) {
1256 if (match_state && unlikely(p->state != match_state))
1262 * Ok, time to look more closely! We need the rq
1263 * lock now, to be *sure*. If we're wrong, we'll
1264 * just go back and repeat.
1266 rq = task_rq_lock(p, &flags);
1267 trace_sched_wait_task(p);
1268 running = task_running(rq, p);
1269 queued = task_on_rq_queued(p);
1271 if (!match_state || p->state == match_state)
1272 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1273 task_rq_unlock(rq, p, &flags);
1276 * If it changed from the expected state, bail out now.
1278 if (unlikely(!ncsw))
1282 * Was it really running after all now that we
1283 * checked with the proper locks actually held?
1285 * Oops. Go back and try again..
1287 if (unlikely(running)) {
1293 * It's not enough that it's not actively running,
1294 * it must be off the runqueue _entirely_, and not
1297 * So if it was still runnable (but just not actively
1298 * running right now), it's preempted, and we should
1299 * yield - it could be a while.
1301 if (unlikely(queued)) {
1302 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1304 set_current_state(TASK_UNINTERRUPTIBLE);
1305 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1310 * Ahh, all good. It wasn't running, and it wasn't
1311 * runnable, which means that it will never become
1312 * running in the future either. We're all done!
1321 * kick_process - kick a running thread to enter/exit the kernel
1322 * @p: the to-be-kicked thread
1324 * Cause a process which is running on another CPU to enter
1325 * kernel-mode, without any delay. (to get signals handled.)
1327 * NOTE: this function doesn't have to take the runqueue lock,
1328 * because all it wants to ensure is that the remote task enters
1329 * the kernel. If the IPI races and the task has been migrated
1330 * to another CPU then no harm is done and the purpose has been
1333 void kick_process(struct task_struct *p)
1339 if ((cpu != smp_processor_id()) && task_curr(p))
1340 smp_send_reschedule(cpu);
1343 EXPORT_SYMBOL_GPL(kick_process);
1344 #endif /* CONFIG_SMP */
1348 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1350 static int select_fallback_rq(int cpu, struct task_struct *p)
1352 int nid = cpu_to_node(cpu);
1353 const struct cpumask *nodemask = NULL;
1354 enum { cpuset, possible, fail } state = cpuset;
1358 * If the node that the cpu is on has been offlined, cpu_to_node()
1359 * will return -1. There is no cpu on the node, and we should
1360 * select the cpu on the other node.
1363 nodemask = cpumask_of_node(nid);
1365 /* Look for allowed, online CPU in same node. */
1366 for_each_cpu(dest_cpu, nodemask) {
1367 if (!cpu_online(dest_cpu))
1369 if (!cpu_active(dest_cpu))
1371 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1377 /* Any allowed, online CPU? */
1378 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1379 if (!cpu_online(dest_cpu))
1381 if (!cpu_active(dest_cpu))
1388 /* No more Mr. Nice Guy. */
1389 cpuset_cpus_allowed_fallback(p);
1394 do_set_cpus_allowed(p, cpu_possible_mask);
1405 if (state != cpuset) {
1407 * Don't tell them about moving exiting tasks or
1408 * kernel threads (both mm NULL), since they never
1411 if (p->mm && printk_ratelimit()) {
1412 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1413 task_pid_nr(p), p->comm, cpu);
1421 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1424 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1426 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1429 * In order not to call set_task_cpu() on a blocking task we need
1430 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1433 * Since this is common to all placement strategies, this lives here.
1435 * [ this allows ->select_task() to simply return task_cpu(p) and
1436 * not worry about this generic constraint ]
1438 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1440 cpu = select_fallback_rq(task_cpu(p), p);
1445 static void update_avg(u64 *avg, u64 sample)
1447 s64 diff = sample - *avg;
1453 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1455 #ifdef CONFIG_SCHEDSTATS
1456 struct rq *rq = this_rq();
1459 int this_cpu = smp_processor_id();
1461 if (cpu == this_cpu) {
1462 schedstat_inc(rq, ttwu_local);
1463 schedstat_inc(p, se.statistics.nr_wakeups_local);
1465 struct sched_domain *sd;
1467 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1469 for_each_domain(this_cpu, sd) {
1470 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1471 schedstat_inc(sd, ttwu_wake_remote);
1478 if (wake_flags & WF_MIGRATED)
1479 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1481 #endif /* CONFIG_SMP */
1483 schedstat_inc(rq, ttwu_count);
1484 schedstat_inc(p, se.statistics.nr_wakeups);
1486 if (wake_flags & WF_SYNC)
1487 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1489 #endif /* CONFIG_SCHEDSTATS */
1492 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1494 activate_task(rq, p, en_flags);
1495 p->on_rq = TASK_ON_RQ_QUEUED;
1497 /* if a worker is waking up, notify workqueue */
1498 if (p->flags & PF_WQ_WORKER)
1499 wq_worker_waking_up(p, cpu_of(rq));
1503 * Mark the task runnable and perform wakeup-preemption.
1506 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1508 check_preempt_curr(rq, p, wake_flags);
1509 trace_sched_wakeup(p, true);
1511 p->state = TASK_RUNNING;
1513 if (p->sched_class->task_woken)
1514 p->sched_class->task_woken(rq, p);
1516 if (rq->idle_stamp) {
1517 u64 delta = rq_clock(rq) - rq->idle_stamp;
1518 u64 max = 2*rq->max_idle_balance_cost;
1520 update_avg(&rq->avg_idle, delta);
1522 if (rq->avg_idle > max)
1531 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1534 if (p->sched_contributes_to_load)
1535 rq->nr_uninterruptible--;
1538 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1539 ttwu_do_wakeup(rq, p, wake_flags);
1543 * Called in case the task @p isn't fully descheduled from its runqueue,
1544 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1545 * since all we need to do is flip p->state to TASK_RUNNING, since
1546 * the task is still ->on_rq.
1548 static int ttwu_remote(struct task_struct *p, int wake_flags)
1553 rq = __task_rq_lock(p);
1554 if (task_on_rq_queued(p)) {
1555 /* check_preempt_curr() may use rq clock */
1556 update_rq_clock(rq);
1557 ttwu_do_wakeup(rq, p, wake_flags);
1560 __task_rq_unlock(rq);
1566 void sched_ttwu_pending(void)
1568 struct rq *rq = this_rq();
1569 struct llist_node *llist = llist_del_all(&rq->wake_list);
1570 struct task_struct *p;
1571 unsigned long flags;
1576 raw_spin_lock_irqsave(&rq->lock, flags);
1579 p = llist_entry(llist, struct task_struct, wake_entry);
1580 llist = llist_next(llist);
1581 ttwu_do_activate(rq, p, 0);
1584 raw_spin_unlock_irqrestore(&rq->lock, flags);
1587 void scheduler_ipi(void)
1590 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1591 * TIF_NEED_RESCHED remotely (for the first time) will also send
1594 preempt_fold_need_resched();
1596 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1600 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1601 * traditionally all their work was done from the interrupt return
1602 * path. Now that we actually do some work, we need to make sure
1605 * Some archs already do call them, luckily irq_enter/exit nest
1608 * Arguably we should visit all archs and update all handlers,
1609 * however a fair share of IPIs are still resched only so this would
1610 * somewhat pessimize the simple resched case.
1613 sched_ttwu_pending();
1616 * Check if someone kicked us for doing the nohz idle load balance.
1618 if (unlikely(got_nohz_idle_kick())) {
1619 this_rq()->idle_balance = 1;
1620 raise_softirq_irqoff(SCHED_SOFTIRQ);
1625 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1627 struct rq *rq = cpu_rq(cpu);
1629 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1630 if (!set_nr_if_polling(rq->idle))
1631 smp_send_reschedule(cpu);
1633 trace_sched_wake_idle_without_ipi(cpu);
1637 void wake_up_if_idle(int cpu)
1639 struct rq *rq = cpu_rq(cpu);
1640 unsigned long flags;
1642 if (!is_idle_task(rq->curr))
1645 if (set_nr_if_polling(rq->idle)) {
1646 trace_sched_wake_idle_without_ipi(cpu);
1648 raw_spin_lock_irqsave(&rq->lock, flags);
1649 if (is_idle_task(rq->curr))
1650 smp_send_reschedule(cpu);
1651 /* Else cpu is not in idle, do nothing here */
1652 raw_spin_unlock_irqrestore(&rq->lock, flags);
1656 bool cpus_share_cache(int this_cpu, int that_cpu)
1658 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1660 #endif /* CONFIG_SMP */
1662 static void ttwu_queue(struct task_struct *p, int cpu)
1664 struct rq *rq = cpu_rq(cpu);
1666 #if defined(CONFIG_SMP)
1667 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1668 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1669 ttwu_queue_remote(p, cpu);
1674 raw_spin_lock(&rq->lock);
1675 ttwu_do_activate(rq, p, 0);
1676 raw_spin_unlock(&rq->lock);
1680 * try_to_wake_up - wake up a thread
1681 * @p: the thread to be awakened
1682 * @state: the mask of task states that can be woken
1683 * @wake_flags: wake modifier flags (WF_*)
1685 * Put it on the run-queue if it's not already there. The "current"
1686 * thread is always on the run-queue (except when the actual
1687 * re-schedule is in progress), and as such you're allowed to do
1688 * the simpler "current->state = TASK_RUNNING" to mark yourself
1689 * runnable without the overhead of this.
1691 * Return: %true if @p was woken up, %false if it was already running.
1692 * or @state didn't match @p's state.
1695 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1697 unsigned long flags;
1698 int cpu, success = 0;
1701 * If we are going to wake up a thread waiting for CONDITION we
1702 * need to ensure that CONDITION=1 done by the caller can not be
1703 * reordered with p->state check below. This pairs with mb() in
1704 * set_current_state() the waiting thread does.
1706 smp_mb__before_spinlock();
1707 raw_spin_lock_irqsave(&p->pi_lock, flags);
1708 if (!(p->state & state))
1711 success = 1; /* we're going to change ->state */
1714 if (p->on_rq && ttwu_remote(p, wake_flags))
1719 * If the owning (remote) cpu is still in the middle of schedule() with
1720 * this task as prev, wait until its done referencing the task.
1725 * Pairs with the smp_wmb() in finish_lock_switch().
1729 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1730 p->state = TASK_WAKING;
1732 if (p->sched_class->task_waking)
1733 p->sched_class->task_waking(p);
1735 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1736 if (task_cpu(p) != cpu) {
1737 wake_flags |= WF_MIGRATED;
1738 set_task_cpu(p, cpu);
1740 #endif /* CONFIG_SMP */
1744 ttwu_stat(p, cpu, wake_flags);
1746 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1752 * try_to_wake_up_local - try to wake up a local task with rq lock held
1753 * @p: the thread to be awakened
1755 * Put @p on the run-queue if it's not already there. The caller must
1756 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1759 static void try_to_wake_up_local(struct task_struct *p)
1761 struct rq *rq = task_rq(p);
1763 if (WARN_ON_ONCE(rq != this_rq()) ||
1764 WARN_ON_ONCE(p == current))
1767 lockdep_assert_held(&rq->lock);
1769 if (!raw_spin_trylock(&p->pi_lock)) {
1770 raw_spin_unlock(&rq->lock);
1771 raw_spin_lock(&p->pi_lock);
1772 raw_spin_lock(&rq->lock);
1775 if (!(p->state & TASK_NORMAL))
1778 if (!task_on_rq_queued(p))
1779 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1781 ttwu_do_wakeup(rq, p, 0);
1782 ttwu_stat(p, smp_processor_id(), 0);
1784 raw_spin_unlock(&p->pi_lock);
1788 * wake_up_process - Wake up a specific process
1789 * @p: The process to be woken up.
1791 * Attempt to wake up the nominated process and move it to the set of runnable
1794 * Return: 1 if the process was woken up, 0 if it was already running.
1796 * It may be assumed that this function implies a write memory barrier before
1797 * changing the task state if and only if any tasks are woken up.
1799 int wake_up_process(struct task_struct *p)
1801 WARN_ON(task_is_stopped_or_traced(p));
1802 return try_to_wake_up(p, TASK_NORMAL, 0);
1804 EXPORT_SYMBOL(wake_up_process);
1806 int wake_up_state(struct task_struct *p, unsigned int state)
1808 return try_to_wake_up(p, state, 0);
1812 * This function clears the sched_dl_entity static params.
1814 void __dl_clear_params(struct task_struct *p)
1816 struct sched_dl_entity *dl_se = &p->dl;
1818 dl_se->dl_runtime = 0;
1819 dl_se->dl_deadline = 0;
1820 dl_se->dl_period = 0;
1826 * Perform scheduler related setup for a newly forked process p.
1827 * p is forked by current.
1829 * __sched_fork() is basic setup used by init_idle() too:
1831 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1836 p->se.exec_start = 0;
1837 p->se.sum_exec_runtime = 0;
1838 p->se.prev_sum_exec_runtime = 0;
1839 p->se.nr_migrations = 0;
1841 INIT_LIST_HEAD(&p->se.group_node);
1843 #ifdef CONFIG_SCHEDSTATS
1844 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1847 RB_CLEAR_NODE(&p->dl.rb_node);
1848 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1849 __dl_clear_params(p);
1851 INIT_LIST_HEAD(&p->rt.run_list);
1853 #ifdef CONFIG_PREEMPT_NOTIFIERS
1854 INIT_HLIST_HEAD(&p->preempt_notifiers);
1857 #ifdef CONFIG_NUMA_BALANCING
1858 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1859 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1860 p->mm->numa_scan_seq = 0;
1863 if (clone_flags & CLONE_VM)
1864 p->numa_preferred_nid = current->numa_preferred_nid;
1866 p->numa_preferred_nid = -1;
1868 p->node_stamp = 0ULL;
1869 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1870 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1871 p->numa_work.next = &p->numa_work;
1872 p->numa_faults_memory = NULL;
1873 p->numa_faults_buffer_memory = NULL;
1874 p->last_task_numa_placement = 0;
1875 p->last_sum_exec_runtime = 0;
1877 INIT_LIST_HEAD(&p->numa_entry);
1878 p->numa_group = NULL;
1879 #endif /* CONFIG_NUMA_BALANCING */
1882 #ifdef CONFIG_NUMA_BALANCING
1883 #ifdef CONFIG_SCHED_DEBUG
1884 void set_numabalancing_state(bool enabled)
1887 sched_feat_set("NUMA");
1889 sched_feat_set("NO_NUMA");
1892 __read_mostly bool numabalancing_enabled;
1894 void set_numabalancing_state(bool enabled)
1896 numabalancing_enabled = enabled;
1898 #endif /* CONFIG_SCHED_DEBUG */
1900 #ifdef CONFIG_PROC_SYSCTL
1901 int sysctl_numa_balancing(struct ctl_table *table, int write,
1902 void __user *buffer, size_t *lenp, loff_t *ppos)
1906 int state = numabalancing_enabled;
1908 if (write && !capable(CAP_SYS_ADMIN))
1913 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1917 set_numabalancing_state(state);
1924 * fork()/clone()-time setup:
1926 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1928 unsigned long flags;
1929 int cpu = get_cpu();
1931 __sched_fork(clone_flags, p);
1933 * We mark the process as running here. This guarantees that
1934 * nobody will actually run it, and a signal or other external
1935 * event cannot wake it up and insert it on the runqueue either.
1937 p->state = TASK_RUNNING;
1940 * Make sure we do not leak PI boosting priority to the child.
1942 p->prio = current->normal_prio;
1945 * Revert to default priority/policy on fork if requested.
1947 if (unlikely(p->sched_reset_on_fork)) {
1948 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1949 p->policy = SCHED_NORMAL;
1950 p->static_prio = NICE_TO_PRIO(0);
1952 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1953 p->static_prio = NICE_TO_PRIO(0);
1955 p->prio = p->normal_prio = __normal_prio(p);
1959 * We don't need the reset flag anymore after the fork. It has
1960 * fulfilled its duty:
1962 p->sched_reset_on_fork = 0;
1965 if (dl_prio(p->prio)) {
1968 } else if (rt_prio(p->prio)) {
1969 p->sched_class = &rt_sched_class;
1971 p->sched_class = &fair_sched_class;
1974 if (p->sched_class->task_fork)
1975 p->sched_class->task_fork(p);
1978 * The child is not yet in the pid-hash so no cgroup attach races,
1979 * and the cgroup is pinned to this child due to cgroup_fork()
1980 * is ran before sched_fork().
1982 * Silence PROVE_RCU.
1984 raw_spin_lock_irqsave(&p->pi_lock, flags);
1985 set_task_cpu(p, cpu);
1986 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1988 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1989 if (likely(sched_info_on()))
1990 memset(&p->sched_info, 0, sizeof(p->sched_info));
1992 #if defined(CONFIG_SMP)
1995 init_task_preempt_count(p);
1997 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1998 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2005 unsigned long to_ratio(u64 period, u64 runtime)
2007 if (runtime == RUNTIME_INF)
2011 * Doing this here saves a lot of checks in all
2012 * the calling paths, and returning zero seems
2013 * safe for them anyway.
2018 return div64_u64(runtime << 20, period);
2022 inline struct dl_bw *dl_bw_of(int i)
2024 return &cpu_rq(i)->rd->dl_bw;
2027 static inline int dl_bw_cpus(int i)
2029 struct root_domain *rd = cpu_rq(i)->rd;
2032 for_each_cpu_and(i, rd->span, cpu_active_mask)
2038 inline struct dl_bw *dl_bw_of(int i)
2040 return &cpu_rq(i)->dl.dl_bw;
2043 static inline int dl_bw_cpus(int i)
2050 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2052 dl_b->total_bw -= tsk_bw;
2056 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2058 dl_b->total_bw += tsk_bw;
2062 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2064 return dl_b->bw != -1 &&
2065 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2069 * We must be sure that accepting a new task (or allowing changing the
2070 * parameters of an existing one) is consistent with the bandwidth
2071 * constraints. If yes, this function also accordingly updates the currently
2072 * allocated bandwidth to reflect the new situation.
2074 * This function is called while holding p's rq->lock.
2076 static int dl_overflow(struct task_struct *p, int policy,
2077 const struct sched_attr *attr)
2080 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2081 u64 period = attr->sched_period ?: attr->sched_deadline;
2082 u64 runtime = attr->sched_runtime;
2083 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2086 if (new_bw == p->dl.dl_bw)
2090 * Either if a task, enters, leave, or stays -deadline but changes
2091 * its parameters, we may need to update accordingly the total
2092 * allocated bandwidth of the container.
2094 raw_spin_lock(&dl_b->lock);
2095 cpus = dl_bw_cpus(task_cpu(p));
2096 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2097 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2098 __dl_add(dl_b, new_bw);
2100 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2101 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2102 __dl_clear(dl_b, p->dl.dl_bw);
2103 __dl_add(dl_b, new_bw);
2105 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2106 __dl_clear(dl_b, p->dl.dl_bw);
2109 raw_spin_unlock(&dl_b->lock);
2114 extern void init_dl_bw(struct dl_bw *dl_b);
2117 * wake_up_new_task - wake up a newly created task for the first time.
2119 * This function will do some initial scheduler statistics housekeeping
2120 * that must be done for every newly created context, then puts the task
2121 * on the runqueue and wakes it.
2123 void wake_up_new_task(struct task_struct *p)
2125 unsigned long flags;
2128 raw_spin_lock_irqsave(&p->pi_lock, flags);
2131 * Fork balancing, do it here and not earlier because:
2132 * - cpus_allowed can change in the fork path
2133 * - any previously selected cpu might disappear through hotplug
2135 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2138 /* Initialize new task's runnable average */
2139 init_task_runnable_average(p);
2140 rq = __task_rq_lock(p);
2141 activate_task(rq, p, 0);
2142 p->on_rq = TASK_ON_RQ_QUEUED;
2143 trace_sched_wakeup_new(p, true);
2144 check_preempt_curr(rq, p, WF_FORK);
2146 if (p->sched_class->task_woken)
2147 p->sched_class->task_woken(rq, p);
2149 task_rq_unlock(rq, p, &flags);
2152 #ifdef CONFIG_PREEMPT_NOTIFIERS
2155 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2156 * @notifier: notifier struct to register
2158 void preempt_notifier_register(struct preempt_notifier *notifier)
2160 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2162 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2165 * preempt_notifier_unregister - no longer interested in preemption notifications
2166 * @notifier: notifier struct to unregister
2168 * This is safe to call from within a preemption notifier.
2170 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2172 hlist_del(¬ifier->link);
2174 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2176 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2178 struct preempt_notifier *notifier;
2180 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2181 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2185 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2186 struct task_struct *next)
2188 struct preempt_notifier *notifier;
2190 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2191 notifier->ops->sched_out(notifier, next);
2194 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2196 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2201 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2202 struct task_struct *next)
2206 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2209 * prepare_task_switch - prepare to switch tasks
2210 * @rq: the runqueue preparing to switch
2211 * @prev: the current task that is being switched out
2212 * @next: the task we are going to switch to.
2214 * This is called with the rq lock held and interrupts off. It must
2215 * be paired with a subsequent finish_task_switch after the context
2218 * prepare_task_switch sets up locking and calls architecture specific
2222 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2223 struct task_struct *next)
2225 trace_sched_switch(prev, next);
2226 sched_info_switch(rq, prev, next);
2227 perf_event_task_sched_out(prev, next);
2228 fire_sched_out_preempt_notifiers(prev, next);
2229 prepare_lock_switch(rq, next);
2230 prepare_arch_switch(next);
2234 * finish_task_switch - clean up after a task-switch
2235 * @rq: runqueue associated with task-switch
2236 * @prev: the thread we just switched away from.
2238 * finish_task_switch must be called after the context switch, paired
2239 * with a prepare_task_switch call before the context switch.
2240 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2241 * and do any other architecture-specific cleanup actions.
2243 * Note that we may have delayed dropping an mm in context_switch(). If
2244 * so, we finish that here outside of the runqueue lock. (Doing it
2245 * with the lock held can cause deadlocks; see schedule() for
2248 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2249 __releases(rq->lock)
2251 struct mm_struct *mm = rq->prev_mm;
2257 * A task struct has one reference for the use as "current".
2258 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2259 * schedule one last time. The schedule call will never return, and
2260 * the scheduled task must drop that reference.
2261 * The test for TASK_DEAD must occur while the runqueue locks are
2262 * still held, otherwise prev could be scheduled on another cpu, die
2263 * there before we look at prev->state, and then the reference would
2265 * Manfred Spraul <manfred@colorfullife.com>
2267 prev_state = prev->state;
2268 vtime_task_switch(prev);
2269 finish_arch_switch(prev);
2270 perf_event_task_sched_in(prev, current);
2271 finish_lock_switch(rq, prev);
2272 finish_arch_post_lock_switch();
2274 fire_sched_in_preempt_notifiers(current);
2277 if (unlikely(prev_state == TASK_DEAD)) {
2278 if (prev->sched_class->task_dead)
2279 prev->sched_class->task_dead(prev);
2282 * Remove function-return probe instances associated with this
2283 * task and put them back on the free list.
2285 kprobe_flush_task(prev);
2286 put_task_struct(prev);
2289 tick_nohz_task_switch(current);
2294 /* rq->lock is NOT held, but preemption is disabled */
2295 static inline void post_schedule(struct rq *rq)
2297 if (rq->post_schedule) {
2298 unsigned long flags;
2300 raw_spin_lock_irqsave(&rq->lock, flags);
2301 if (rq->curr->sched_class->post_schedule)
2302 rq->curr->sched_class->post_schedule(rq);
2303 raw_spin_unlock_irqrestore(&rq->lock, flags);
2305 rq->post_schedule = 0;
2311 static inline void post_schedule(struct rq *rq)
2318 * schedule_tail - first thing a freshly forked thread must call.
2319 * @prev: the thread we just switched away from.
2321 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2322 __releases(rq->lock)
2324 struct rq *rq = this_rq();
2326 finish_task_switch(rq, prev);
2329 * FIXME: do we need to worry about rq being invalidated by the
2334 if (current->set_child_tid)
2335 put_user(task_pid_vnr(current), current->set_child_tid);
2339 * context_switch - switch to the new MM and the new
2340 * thread's register state.
2343 context_switch(struct rq *rq, struct task_struct *prev,
2344 struct task_struct *next)
2346 struct mm_struct *mm, *oldmm;
2348 prepare_task_switch(rq, prev, next);
2351 oldmm = prev->active_mm;
2353 * For paravirt, this is coupled with an exit in switch_to to
2354 * combine the page table reload and the switch backend into
2357 arch_start_context_switch(prev);
2360 next->active_mm = oldmm;
2361 atomic_inc(&oldmm->mm_count);
2362 enter_lazy_tlb(oldmm, next);
2364 switch_mm(oldmm, mm, next);
2367 prev->active_mm = NULL;
2368 rq->prev_mm = oldmm;
2371 * Since the runqueue lock will be released by the next
2372 * task (which is an invalid locking op but in the case
2373 * of the scheduler it's an obvious special-case), so we
2374 * do an early lockdep release here:
2376 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2378 context_tracking_task_switch(prev, next);
2379 /* Here we just switch the register state and the stack. */
2380 switch_to(prev, next, prev);
2384 * this_rq must be evaluated again because prev may have moved
2385 * CPUs since it called schedule(), thus the 'rq' on its stack
2386 * frame will be invalid.
2388 finish_task_switch(this_rq(), prev);
2392 * nr_running and nr_context_switches:
2394 * externally visible scheduler statistics: current number of runnable
2395 * threads, total number of context switches performed since bootup.
2397 unsigned long nr_running(void)
2399 unsigned long i, sum = 0;
2401 for_each_online_cpu(i)
2402 sum += cpu_rq(i)->nr_running;
2407 unsigned long long nr_context_switches(void)
2410 unsigned long long sum = 0;
2412 for_each_possible_cpu(i)
2413 sum += cpu_rq(i)->nr_switches;
2418 unsigned long nr_iowait(void)
2420 unsigned long i, sum = 0;
2422 for_each_possible_cpu(i)
2423 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2428 unsigned long nr_iowait_cpu(int cpu)
2430 struct rq *this = cpu_rq(cpu);
2431 return atomic_read(&this->nr_iowait);
2434 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2436 struct rq *this = this_rq();
2437 *nr_waiters = atomic_read(&this->nr_iowait);
2438 *load = this->cpu_load[0];
2444 * sched_exec - execve() is a valuable balancing opportunity, because at
2445 * this point the task has the smallest effective memory and cache footprint.
2447 void sched_exec(void)
2449 struct task_struct *p = current;
2450 unsigned long flags;
2453 raw_spin_lock_irqsave(&p->pi_lock, flags);
2454 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2455 if (dest_cpu == smp_processor_id())
2458 if (likely(cpu_active(dest_cpu))) {
2459 struct migration_arg arg = { p, dest_cpu };
2461 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2462 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2466 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2471 DEFINE_PER_CPU(struct kernel_stat, kstat);
2472 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2474 EXPORT_PER_CPU_SYMBOL(kstat);
2475 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2478 * Return any ns on the sched_clock that have not yet been accounted in
2479 * @p in case that task is currently running.
2481 * Called with task_rq_lock() held on @rq.
2483 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2488 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2489 * project cycles that may never be accounted to this
2490 * thread, breaking clock_gettime().
2492 if (task_current(rq, p) && task_on_rq_queued(p)) {
2493 update_rq_clock(rq);
2494 ns = rq_clock_task(rq) - p->se.exec_start;
2502 unsigned long long task_delta_exec(struct task_struct *p)
2504 unsigned long flags;
2508 rq = task_rq_lock(p, &flags);
2509 ns = do_task_delta_exec(p, rq);
2510 task_rq_unlock(rq, p, &flags);
2516 * Return accounted runtime for the task.
2517 * In case the task is currently running, return the runtime plus current's
2518 * pending runtime that have not been accounted yet.
2520 unsigned long long task_sched_runtime(struct task_struct *p)
2522 unsigned long flags;
2526 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2528 * 64-bit doesn't need locks to atomically read a 64bit value.
2529 * So we have a optimization chance when the task's delta_exec is 0.
2530 * Reading ->on_cpu is racy, but this is ok.
2532 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2533 * If we race with it entering cpu, unaccounted time is 0. This is
2534 * indistinguishable from the read occurring a few cycles earlier.
2535 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2536 * been accounted, so we're correct here as well.
2538 if (!p->on_cpu || !task_on_rq_queued(p))
2539 return p->se.sum_exec_runtime;
2542 rq = task_rq_lock(p, &flags);
2543 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2544 task_rq_unlock(rq, p, &flags);
2550 * This function gets called by the timer code, with HZ frequency.
2551 * We call it with interrupts disabled.
2553 void scheduler_tick(void)
2555 int cpu = smp_processor_id();
2556 struct rq *rq = cpu_rq(cpu);
2557 struct task_struct *curr = rq->curr;
2561 raw_spin_lock(&rq->lock);
2562 update_rq_clock(rq);
2563 curr->sched_class->task_tick(rq, curr, 0);
2564 update_cpu_load_active(rq);
2565 raw_spin_unlock(&rq->lock);
2567 perf_event_task_tick();
2570 rq->idle_balance = idle_cpu(cpu);
2571 trigger_load_balance(rq);
2573 rq_last_tick_reset(rq);
2576 #ifdef CONFIG_NO_HZ_FULL
2578 * scheduler_tick_max_deferment
2580 * Keep at least one tick per second when a single
2581 * active task is running because the scheduler doesn't
2582 * yet completely support full dynticks environment.
2584 * This makes sure that uptime, CFS vruntime, load
2585 * balancing, etc... continue to move forward, even
2586 * with a very low granularity.
2588 * Return: Maximum deferment in nanoseconds.
2590 u64 scheduler_tick_max_deferment(void)
2592 struct rq *rq = this_rq();
2593 unsigned long next, now = ACCESS_ONCE(jiffies);
2595 next = rq->last_sched_tick + HZ;
2597 if (time_before_eq(next, now))
2600 return jiffies_to_nsecs(next - now);
2604 notrace unsigned long get_parent_ip(unsigned long addr)
2606 if (in_lock_functions(addr)) {
2607 addr = CALLER_ADDR2;
2608 if (in_lock_functions(addr))
2609 addr = CALLER_ADDR3;
2614 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2615 defined(CONFIG_PREEMPT_TRACER))
2617 void preempt_count_add(int val)
2619 #ifdef CONFIG_DEBUG_PREEMPT
2623 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2626 __preempt_count_add(val);
2627 #ifdef CONFIG_DEBUG_PREEMPT
2629 * Spinlock count overflowing soon?
2631 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2634 if (preempt_count() == val) {
2635 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2636 #ifdef CONFIG_DEBUG_PREEMPT
2637 current->preempt_disable_ip = ip;
2639 trace_preempt_off(CALLER_ADDR0, ip);
2642 EXPORT_SYMBOL(preempt_count_add);
2643 NOKPROBE_SYMBOL(preempt_count_add);
2645 void preempt_count_sub(int val)
2647 #ifdef CONFIG_DEBUG_PREEMPT
2651 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2654 * Is the spinlock portion underflowing?
2656 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2657 !(preempt_count() & PREEMPT_MASK)))
2661 if (preempt_count() == val)
2662 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2663 __preempt_count_sub(val);
2665 EXPORT_SYMBOL(preempt_count_sub);
2666 NOKPROBE_SYMBOL(preempt_count_sub);
2671 * Print scheduling while atomic bug:
2673 static noinline void __schedule_bug(struct task_struct *prev)
2675 if (oops_in_progress)
2678 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2679 prev->comm, prev->pid, preempt_count());
2681 debug_show_held_locks(prev);
2683 if (irqs_disabled())
2684 print_irqtrace_events(prev);
2685 #ifdef CONFIG_DEBUG_PREEMPT
2686 if (in_atomic_preempt_off()) {
2687 pr_err("Preemption disabled at:");
2688 print_ip_sym(current->preempt_disable_ip);
2693 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2697 * Various schedule()-time debugging checks and statistics:
2699 static inline void schedule_debug(struct task_struct *prev)
2701 #ifdef CONFIG_SCHED_STACK_END_CHECK
2702 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2705 * Test if we are atomic. Since do_exit() needs to call into
2706 * schedule() atomically, we ignore that path. Otherwise whine
2707 * if we are scheduling when we should not.
2709 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2710 __schedule_bug(prev);
2713 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2715 schedstat_inc(this_rq(), sched_count);
2719 * Pick up the highest-prio task:
2721 static inline struct task_struct *
2722 pick_next_task(struct rq *rq, struct task_struct *prev)
2724 const struct sched_class *class = &fair_sched_class;
2725 struct task_struct *p;
2728 * Optimization: we know that if all tasks are in
2729 * the fair class we can call that function directly:
2731 if (likely(prev->sched_class == class &&
2732 rq->nr_running == rq->cfs.h_nr_running)) {
2733 p = fair_sched_class.pick_next_task(rq, prev);
2734 if (unlikely(p == RETRY_TASK))
2737 /* assumes fair_sched_class->next == idle_sched_class */
2739 p = idle_sched_class.pick_next_task(rq, prev);
2745 for_each_class(class) {
2746 p = class->pick_next_task(rq, prev);
2748 if (unlikely(p == RETRY_TASK))
2754 BUG(); /* the idle class will always have a runnable task */
2758 * __schedule() is the main scheduler function.
2760 * The main means of driving the scheduler and thus entering this function are:
2762 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2764 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2765 * paths. For example, see arch/x86/entry_64.S.
2767 * To drive preemption between tasks, the scheduler sets the flag in timer
2768 * interrupt handler scheduler_tick().
2770 * 3. Wakeups don't really cause entry into schedule(). They add a
2771 * task to the run-queue and that's it.
2773 * Now, if the new task added to the run-queue preempts the current
2774 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2775 * called on the nearest possible occasion:
2777 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2779 * - in syscall or exception context, at the next outmost
2780 * preempt_enable(). (this might be as soon as the wake_up()'s
2783 * - in IRQ context, return from interrupt-handler to
2784 * preemptible context
2786 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2789 * - cond_resched() call
2790 * - explicit schedule() call
2791 * - return from syscall or exception to user-space
2792 * - return from interrupt-handler to user-space
2794 static void __sched __schedule(void)
2796 struct task_struct *prev, *next;
2797 unsigned long *switch_count;
2803 cpu = smp_processor_id();
2805 rcu_note_context_switch(cpu);
2808 schedule_debug(prev);
2810 if (sched_feat(HRTICK))
2814 * Make sure that signal_pending_state()->signal_pending() below
2815 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2816 * done by the caller to avoid the race with signal_wake_up().
2818 smp_mb__before_spinlock();
2819 raw_spin_lock_irq(&rq->lock);
2821 switch_count = &prev->nivcsw;
2822 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2823 if (unlikely(signal_pending_state(prev->state, prev))) {
2824 prev->state = TASK_RUNNING;
2826 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2830 * If a worker went to sleep, notify and ask workqueue
2831 * whether it wants to wake up a task to maintain
2834 if (prev->flags & PF_WQ_WORKER) {
2835 struct task_struct *to_wakeup;
2837 to_wakeup = wq_worker_sleeping(prev, cpu);
2839 try_to_wake_up_local(to_wakeup);
2842 switch_count = &prev->nvcsw;
2845 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0)
2846 update_rq_clock(rq);
2848 next = pick_next_task(rq, prev);
2849 clear_tsk_need_resched(prev);
2850 clear_preempt_need_resched();
2851 rq->skip_clock_update = 0;
2853 if (likely(prev != next)) {
2858 context_switch(rq, prev, next); /* unlocks the rq */
2860 * The context switch have flipped the stack from under us
2861 * and restored the local variables which were saved when
2862 * this task called schedule() in the past. prev == current
2863 * is still correct, but it can be moved to another cpu/rq.
2865 cpu = smp_processor_id();
2868 raw_spin_unlock_irq(&rq->lock);
2872 sched_preempt_enable_no_resched();
2877 static inline void sched_submit_work(struct task_struct *tsk)
2879 if (!tsk->state || tsk_is_pi_blocked(tsk))
2882 * If we are going to sleep and we have plugged IO queued,
2883 * make sure to submit it to avoid deadlocks.
2885 if (blk_needs_flush_plug(tsk))
2886 blk_schedule_flush_plug(tsk);
2889 asmlinkage __visible void __sched schedule(void)
2891 struct task_struct *tsk = current;
2893 sched_submit_work(tsk);
2896 EXPORT_SYMBOL(schedule);
2898 #ifdef CONFIG_CONTEXT_TRACKING
2899 asmlinkage __visible void __sched schedule_user(void)
2902 * If we come here after a random call to set_need_resched(),
2903 * or we have been woken up remotely but the IPI has not yet arrived,
2904 * we haven't yet exited the RCU idle mode. Do it here manually until
2905 * we find a better solution.
2914 * schedule_preempt_disabled - called with preemption disabled
2916 * Returns with preemption disabled. Note: preempt_count must be 1
2918 void __sched schedule_preempt_disabled(void)
2920 sched_preempt_enable_no_resched();
2925 #ifdef CONFIG_PREEMPT
2927 * this is the entry point to schedule() from in-kernel preemption
2928 * off of preempt_enable. Kernel preemptions off return from interrupt
2929 * occur there and call schedule directly.
2931 asmlinkage __visible void __sched notrace preempt_schedule(void)
2934 * If there is a non-zero preempt_count or interrupts are disabled,
2935 * we do not want to preempt the current task. Just return..
2937 if (likely(!preemptible()))
2941 __preempt_count_add(PREEMPT_ACTIVE);
2943 __preempt_count_sub(PREEMPT_ACTIVE);
2946 * Check again in case we missed a preemption opportunity
2947 * between schedule and now.
2950 } while (need_resched());
2952 NOKPROBE_SYMBOL(preempt_schedule);
2953 EXPORT_SYMBOL(preempt_schedule);
2954 #endif /* CONFIG_PREEMPT */
2957 * this is the entry point to schedule() from kernel preemption
2958 * off of irq context.
2959 * Note, that this is called and return with irqs disabled. This will
2960 * protect us against recursive calling from irq.
2962 asmlinkage __visible void __sched preempt_schedule_irq(void)
2964 enum ctx_state prev_state;
2966 /* Catch callers which need to be fixed */
2967 BUG_ON(preempt_count() || !irqs_disabled());
2969 prev_state = exception_enter();
2972 __preempt_count_add(PREEMPT_ACTIVE);
2975 local_irq_disable();
2976 __preempt_count_sub(PREEMPT_ACTIVE);
2979 * Check again in case we missed a preemption opportunity
2980 * between schedule and now.
2983 } while (need_resched());
2985 exception_exit(prev_state);
2988 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2991 return try_to_wake_up(curr->private, mode, wake_flags);
2993 EXPORT_SYMBOL(default_wake_function);
2995 #ifdef CONFIG_RT_MUTEXES
2998 * rt_mutex_setprio - set the current priority of a task
3000 * @prio: prio value (kernel-internal form)
3002 * This function changes the 'effective' priority of a task. It does
3003 * not touch ->normal_prio like __setscheduler().
3005 * Used by the rt_mutex code to implement priority inheritance
3006 * logic. Call site only calls if the priority of the task changed.
3008 void rt_mutex_setprio(struct task_struct *p, int prio)
3010 int oldprio, queued, running, enqueue_flag = 0;
3012 const struct sched_class *prev_class;
3014 BUG_ON(prio > MAX_PRIO);
3016 rq = __task_rq_lock(p);
3019 * Idle task boosting is a nono in general. There is one
3020 * exception, when PREEMPT_RT and NOHZ is active:
3022 * The idle task calls get_next_timer_interrupt() and holds
3023 * the timer wheel base->lock on the CPU and another CPU wants
3024 * to access the timer (probably to cancel it). We can safely
3025 * ignore the boosting request, as the idle CPU runs this code
3026 * with interrupts disabled and will complete the lock
3027 * protected section without being interrupted. So there is no
3028 * real need to boost.
3030 if (unlikely(p == rq->idle)) {
3031 WARN_ON(p != rq->curr);
3032 WARN_ON(p->pi_blocked_on);
3036 trace_sched_pi_setprio(p, prio);
3038 prev_class = p->sched_class;
3039 queued = task_on_rq_queued(p);
3040 running = task_current(rq, p);
3042 dequeue_task(rq, p, 0);
3044 put_prev_task(rq, p);
3047 * Boosting condition are:
3048 * 1. -rt task is running and holds mutex A
3049 * --> -dl task blocks on mutex A
3051 * 2. -dl task is running and holds mutex A
3052 * --> -dl task blocks on mutex A and could preempt the
3055 if (dl_prio(prio)) {
3056 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3057 if (!dl_prio(p->normal_prio) ||
3058 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3059 p->dl.dl_boosted = 1;
3060 p->dl.dl_throttled = 0;
3061 enqueue_flag = ENQUEUE_REPLENISH;
3063 p->dl.dl_boosted = 0;
3064 p->sched_class = &dl_sched_class;
3065 } else if (rt_prio(prio)) {
3066 if (dl_prio(oldprio))
3067 p->dl.dl_boosted = 0;
3069 enqueue_flag = ENQUEUE_HEAD;
3070 p->sched_class = &rt_sched_class;
3072 if (dl_prio(oldprio))
3073 p->dl.dl_boosted = 0;
3074 p->sched_class = &fair_sched_class;
3080 p->sched_class->set_curr_task(rq);
3082 enqueue_task(rq, p, enqueue_flag);
3084 check_class_changed(rq, p, prev_class, oldprio);
3086 __task_rq_unlock(rq);
3090 void set_user_nice(struct task_struct *p, long nice)
3092 int old_prio, delta, queued;
3093 unsigned long flags;
3096 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3099 * We have to be careful, if called from sys_setpriority(),
3100 * the task might be in the middle of scheduling on another CPU.
3102 rq = task_rq_lock(p, &flags);
3104 * The RT priorities are set via sched_setscheduler(), but we still
3105 * allow the 'normal' nice value to be set - but as expected
3106 * it wont have any effect on scheduling until the task is
3107 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3109 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3110 p->static_prio = NICE_TO_PRIO(nice);
3113 queued = task_on_rq_queued(p);
3115 dequeue_task(rq, p, 0);
3117 p->static_prio = NICE_TO_PRIO(nice);
3120 p->prio = effective_prio(p);
3121 delta = p->prio - old_prio;
3124 enqueue_task(rq, p, 0);
3126 * If the task increased its priority or is running and
3127 * lowered its priority, then reschedule its CPU:
3129 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3133 task_rq_unlock(rq, p, &flags);
3135 EXPORT_SYMBOL(set_user_nice);
3138 * can_nice - check if a task can reduce its nice value
3142 int can_nice(const struct task_struct *p, const int nice)
3144 /* convert nice value [19,-20] to rlimit style value [1,40] */
3145 int nice_rlim = nice_to_rlimit(nice);
3147 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3148 capable(CAP_SYS_NICE));
3151 #ifdef __ARCH_WANT_SYS_NICE
3154 * sys_nice - change the priority of the current process.
3155 * @increment: priority increment
3157 * sys_setpriority is a more generic, but much slower function that
3158 * does similar things.
3160 SYSCALL_DEFINE1(nice, int, increment)
3165 * Setpriority might change our priority at the same moment.
3166 * We don't have to worry. Conceptually one call occurs first
3167 * and we have a single winner.
3169 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3170 nice = task_nice(current) + increment;
3172 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3173 if (increment < 0 && !can_nice(current, nice))
3176 retval = security_task_setnice(current, nice);
3180 set_user_nice(current, nice);
3187 * task_prio - return the priority value of a given task.
3188 * @p: the task in question.
3190 * Return: The priority value as seen by users in /proc.
3191 * RT tasks are offset by -200. Normal tasks are centered
3192 * around 0, value goes from -16 to +15.
3194 int task_prio(const struct task_struct *p)
3196 return p->prio - MAX_RT_PRIO;
3200 * idle_cpu - is a given cpu idle currently?
3201 * @cpu: the processor in question.
3203 * Return: 1 if the CPU is currently idle. 0 otherwise.
3205 int idle_cpu(int cpu)
3207 struct rq *rq = cpu_rq(cpu);
3209 if (rq->curr != rq->idle)
3216 if (!llist_empty(&rq->wake_list))
3224 * idle_task - return the idle task for a given cpu.
3225 * @cpu: the processor in question.
3227 * Return: The idle task for the cpu @cpu.
3229 struct task_struct *idle_task(int cpu)
3231 return cpu_rq(cpu)->idle;
3235 * find_process_by_pid - find a process with a matching PID value.
3236 * @pid: the pid in question.
3238 * The task of @pid, if found. %NULL otherwise.
3240 static struct task_struct *find_process_by_pid(pid_t pid)
3242 return pid ? find_task_by_vpid(pid) : current;
3246 * This function initializes the sched_dl_entity of a newly becoming
3247 * SCHED_DEADLINE task.
3249 * Only the static values are considered here, the actual runtime and the
3250 * absolute deadline will be properly calculated when the task is enqueued
3251 * for the first time with its new policy.
3254 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3256 struct sched_dl_entity *dl_se = &p->dl;
3258 init_dl_task_timer(dl_se);
3259 dl_se->dl_runtime = attr->sched_runtime;
3260 dl_se->dl_deadline = attr->sched_deadline;
3261 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3262 dl_se->flags = attr->sched_flags;
3263 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3264 dl_se->dl_throttled = 0;
3266 dl_se->dl_yielded = 0;
3270 * sched_setparam() passes in -1 for its policy, to let the functions
3271 * it calls know not to change it.
3273 #define SETPARAM_POLICY -1
3275 static void __setscheduler_params(struct task_struct *p,
3276 const struct sched_attr *attr)
3278 int policy = attr->sched_policy;
3280 if (policy == SETPARAM_POLICY)
3285 if (dl_policy(policy))
3286 __setparam_dl(p, attr);
3287 else if (fair_policy(policy))
3288 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3291 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3292 * !rt_policy. Always setting this ensures that things like
3293 * getparam()/getattr() don't report silly values for !rt tasks.
3295 p->rt_priority = attr->sched_priority;
3296 p->normal_prio = normal_prio(p);
3300 /* Actually do priority change: must hold pi & rq lock. */
3301 static void __setscheduler(struct rq *rq, struct task_struct *p,
3302 const struct sched_attr *attr)
3304 __setscheduler_params(p, attr);
3307 * If we get here, there was no pi waiters boosting the
3308 * task. It is safe to use the normal prio.
3310 p->prio = normal_prio(p);
3312 if (dl_prio(p->prio))
3313 p->sched_class = &dl_sched_class;
3314 else if (rt_prio(p->prio))
3315 p->sched_class = &rt_sched_class;
3317 p->sched_class = &fair_sched_class;
3321 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3323 struct sched_dl_entity *dl_se = &p->dl;
3325 attr->sched_priority = p->rt_priority;
3326 attr->sched_runtime = dl_se->dl_runtime;
3327 attr->sched_deadline = dl_se->dl_deadline;
3328 attr->sched_period = dl_se->dl_period;
3329 attr->sched_flags = dl_se->flags;
3333 * This function validates the new parameters of a -deadline task.
3334 * We ask for the deadline not being zero, and greater or equal
3335 * than the runtime, as well as the period of being zero or
3336 * greater than deadline. Furthermore, we have to be sure that
3337 * user parameters are above the internal resolution of 1us (we
3338 * check sched_runtime only since it is always the smaller one) and
3339 * below 2^63 ns (we have to check both sched_deadline and
3340 * sched_period, as the latter can be zero).
3343 __checkparam_dl(const struct sched_attr *attr)
3346 if (attr->sched_deadline == 0)
3350 * Since we truncate DL_SCALE bits, make sure we're at least
3353 if (attr->sched_runtime < (1ULL << DL_SCALE))
3357 * Since we use the MSB for wrap-around and sign issues, make
3358 * sure it's not set (mind that period can be equal to zero).
3360 if (attr->sched_deadline & (1ULL << 63) ||
3361 attr->sched_period & (1ULL << 63))
3364 /* runtime <= deadline <= period (if period != 0) */
3365 if ((attr->sched_period != 0 &&
3366 attr->sched_period < attr->sched_deadline) ||
3367 attr->sched_deadline < attr->sched_runtime)
3374 * check the target process has a UID that matches the current process's
3376 static bool check_same_owner(struct task_struct *p)
3378 const struct cred *cred = current_cred(), *pcred;
3382 pcred = __task_cred(p);
3383 match = (uid_eq(cred->euid, pcred->euid) ||
3384 uid_eq(cred->euid, pcred->uid));
3389 static int __sched_setscheduler(struct task_struct *p,
3390 const struct sched_attr *attr,
3393 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3394 MAX_RT_PRIO - 1 - attr->sched_priority;
3395 int retval, oldprio, oldpolicy = -1, queued, running;
3396 int policy = attr->sched_policy;
3397 unsigned long flags;
3398 const struct sched_class *prev_class;
3402 /* may grab non-irq protected spin_locks */
3403 BUG_ON(in_interrupt());
3405 /* double check policy once rq lock held */
3407 reset_on_fork = p->sched_reset_on_fork;
3408 policy = oldpolicy = p->policy;
3410 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3412 if (policy != SCHED_DEADLINE &&
3413 policy != SCHED_FIFO && policy != SCHED_RR &&
3414 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3415 policy != SCHED_IDLE)
3419 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3423 * Valid priorities for SCHED_FIFO and SCHED_RR are
3424 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3425 * SCHED_BATCH and SCHED_IDLE is 0.
3427 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3428 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3430 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3431 (rt_policy(policy) != (attr->sched_priority != 0)))
3435 * Allow unprivileged RT tasks to decrease priority:
3437 if (user && !capable(CAP_SYS_NICE)) {
3438 if (fair_policy(policy)) {
3439 if (attr->sched_nice < task_nice(p) &&
3440 !can_nice(p, attr->sched_nice))
3444 if (rt_policy(policy)) {
3445 unsigned long rlim_rtprio =
3446 task_rlimit(p, RLIMIT_RTPRIO);
3448 /* can't set/change the rt policy */
3449 if (policy != p->policy && !rlim_rtprio)
3452 /* can't increase priority */
3453 if (attr->sched_priority > p->rt_priority &&
3454 attr->sched_priority > rlim_rtprio)
3459 * Can't set/change SCHED_DEADLINE policy at all for now
3460 * (safest behavior); in the future we would like to allow
3461 * unprivileged DL tasks to increase their relative deadline
3462 * or reduce their runtime (both ways reducing utilization)
3464 if (dl_policy(policy))
3468 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3469 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3471 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3472 if (!can_nice(p, task_nice(p)))
3476 /* can't change other user's priorities */
3477 if (!check_same_owner(p))
3480 /* Normal users shall not reset the sched_reset_on_fork flag */
3481 if (p->sched_reset_on_fork && !reset_on_fork)
3486 retval = security_task_setscheduler(p);
3492 * make sure no PI-waiters arrive (or leave) while we are
3493 * changing the priority of the task:
3495 * To be able to change p->policy safely, the appropriate
3496 * runqueue lock must be held.
3498 rq = task_rq_lock(p, &flags);
3501 * Changing the policy of the stop threads its a very bad idea
3503 if (p == rq->stop) {
3504 task_rq_unlock(rq, p, &flags);
3509 * If not changing anything there's no need to proceed further,
3510 * but store a possible modification of reset_on_fork.
3512 if (unlikely(policy == p->policy)) {
3513 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3515 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3517 if (dl_policy(policy))
3520 p->sched_reset_on_fork = reset_on_fork;
3521 task_rq_unlock(rq, p, &flags);
3527 #ifdef CONFIG_RT_GROUP_SCHED
3529 * Do not allow realtime tasks into groups that have no runtime
3532 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3533 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3534 !task_group_is_autogroup(task_group(p))) {
3535 task_rq_unlock(rq, p, &flags);
3540 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3541 cpumask_t *span = rq->rd->span;
3544 * Don't allow tasks with an affinity mask smaller than
3545 * the entire root_domain to become SCHED_DEADLINE. We
3546 * will also fail if there's no bandwidth available.
3548 if (!cpumask_subset(span, &p->cpus_allowed) ||
3549 rq->rd->dl_bw.bw == 0) {
3550 task_rq_unlock(rq, p, &flags);
3557 /* recheck policy now with rq lock held */
3558 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3559 policy = oldpolicy = -1;
3560 task_rq_unlock(rq, p, &flags);
3565 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3566 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3569 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3570 task_rq_unlock(rq, p, &flags);
3574 p->sched_reset_on_fork = reset_on_fork;
3578 * Special case for priority boosted tasks.
3580 * If the new priority is lower or equal (user space view)
3581 * than the current (boosted) priority, we just store the new
3582 * normal parameters and do not touch the scheduler class and
3583 * the runqueue. This will be done when the task deboost
3586 if (rt_mutex_check_prio(p, newprio)) {
3587 __setscheduler_params(p, attr);
3588 task_rq_unlock(rq, p, &flags);
3592 queued = task_on_rq_queued(p);
3593 running = task_current(rq, p);
3595 dequeue_task(rq, p, 0);
3597 put_prev_task(rq, p);
3599 prev_class = p->sched_class;
3600 __setscheduler(rq, p, attr);
3603 p->sched_class->set_curr_task(rq);
3606 * We enqueue to tail when the priority of a task is
3607 * increased (user space view).
3609 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3612 check_class_changed(rq, p, prev_class, oldprio);
3613 task_rq_unlock(rq, p, &flags);
3615 rt_mutex_adjust_pi(p);
3620 static int _sched_setscheduler(struct task_struct *p, int policy,
3621 const struct sched_param *param, bool check)
3623 struct sched_attr attr = {
3624 .sched_policy = policy,
3625 .sched_priority = param->sched_priority,
3626 .sched_nice = PRIO_TO_NICE(p->static_prio),
3629 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3630 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3631 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3632 policy &= ~SCHED_RESET_ON_FORK;
3633 attr.sched_policy = policy;
3636 return __sched_setscheduler(p, &attr, check);
3639 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3640 * @p: the task in question.
3641 * @policy: new policy.
3642 * @param: structure containing the new RT priority.
3644 * Return: 0 on success. An error code otherwise.
3646 * NOTE that the task may be already dead.
3648 int sched_setscheduler(struct task_struct *p, int policy,
3649 const struct sched_param *param)
3651 return _sched_setscheduler(p, policy, param, true);
3653 EXPORT_SYMBOL_GPL(sched_setscheduler);
3655 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3657 return __sched_setscheduler(p, attr, true);
3659 EXPORT_SYMBOL_GPL(sched_setattr);
3662 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3663 * @p: the task in question.
3664 * @policy: new policy.
3665 * @param: structure containing the new RT priority.
3667 * Just like sched_setscheduler, only don't bother checking if the
3668 * current context has permission. For example, this is needed in
3669 * stop_machine(): we create temporary high priority worker threads,
3670 * but our caller might not have that capability.
3672 * Return: 0 on success. An error code otherwise.
3674 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3675 const struct sched_param *param)
3677 return _sched_setscheduler(p, policy, param, false);
3681 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3683 struct sched_param lparam;
3684 struct task_struct *p;
3687 if (!param || pid < 0)
3689 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3694 p = find_process_by_pid(pid);
3696 retval = sched_setscheduler(p, policy, &lparam);
3703 * Mimics kernel/events/core.c perf_copy_attr().
3705 static int sched_copy_attr(struct sched_attr __user *uattr,
3706 struct sched_attr *attr)
3711 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3715 * zero the full structure, so that a short copy will be nice.
3717 memset(attr, 0, sizeof(*attr));
3719 ret = get_user(size, &uattr->size);
3723 if (size > PAGE_SIZE) /* silly large */
3726 if (!size) /* abi compat */
3727 size = SCHED_ATTR_SIZE_VER0;
3729 if (size < SCHED_ATTR_SIZE_VER0)
3733 * If we're handed a bigger struct than we know of,
3734 * ensure all the unknown bits are 0 - i.e. new
3735 * user-space does not rely on any kernel feature
3736 * extensions we dont know about yet.
3738 if (size > sizeof(*attr)) {
3739 unsigned char __user *addr;
3740 unsigned char __user *end;
3743 addr = (void __user *)uattr + sizeof(*attr);
3744 end = (void __user *)uattr + size;
3746 for (; addr < end; addr++) {
3747 ret = get_user(val, addr);
3753 size = sizeof(*attr);
3756 ret = copy_from_user(attr, uattr, size);
3761 * XXX: do we want to be lenient like existing syscalls; or do we want
3762 * to be strict and return an error on out-of-bounds values?
3764 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3769 put_user(sizeof(*attr), &uattr->size);
3774 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3775 * @pid: the pid in question.
3776 * @policy: new policy.
3777 * @param: structure containing the new RT priority.
3779 * Return: 0 on success. An error code otherwise.
3781 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3782 struct sched_param __user *, param)
3784 /* negative values for policy are not valid */
3788 return do_sched_setscheduler(pid, policy, param);
3792 * sys_sched_setparam - set/change the RT priority of a thread
3793 * @pid: the pid in question.
3794 * @param: structure containing the new RT priority.
3796 * Return: 0 on success. An error code otherwise.
3798 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3800 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3804 * sys_sched_setattr - same as above, but with extended sched_attr
3805 * @pid: the pid in question.
3806 * @uattr: structure containing the extended parameters.
3807 * @flags: for future extension.
3809 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3810 unsigned int, flags)
3812 struct sched_attr attr;
3813 struct task_struct *p;
3816 if (!uattr || pid < 0 || flags)
3819 retval = sched_copy_attr(uattr, &attr);
3823 if ((int)attr.sched_policy < 0)
3828 p = find_process_by_pid(pid);
3830 retval = sched_setattr(p, &attr);
3837 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3838 * @pid: the pid in question.
3840 * Return: On success, the policy of the thread. Otherwise, a negative error
3843 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3845 struct task_struct *p;
3853 p = find_process_by_pid(pid);
3855 retval = security_task_getscheduler(p);
3858 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3865 * sys_sched_getparam - get the RT priority of a thread
3866 * @pid: the pid in question.
3867 * @param: structure containing the RT priority.
3869 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3872 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3874 struct sched_param lp = { .sched_priority = 0 };
3875 struct task_struct *p;
3878 if (!param || pid < 0)
3882 p = find_process_by_pid(pid);
3887 retval = security_task_getscheduler(p);
3891 if (task_has_rt_policy(p))
3892 lp.sched_priority = p->rt_priority;
3896 * This one might sleep, we cannot do it with a spinlock held ...
3898 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3907 static int sched_read_attr(struct sched_attr __user *uattr,
3908 struct sched_attr *attr,
3913 if (!access_ok(VERIFY_WRITE, uattr, usize))
3917 * If we're handed a smaller struct than we know of,
3918 * ensure all the unknown bits are 0 - i.e. old
3919 * user-space does not get uncomplete information.
3921 if (usize < sizeof(*attr)) {
3922 unsigned char *addr;
3925 addr = (void *)attr + usize;
3926 end = (void *)attr + sizeof(*attr);
3928 for (; addr < end; addr++) {
3936 ret = copy_to_user(uattr, attr, attr->size);
3944 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3945 * @pid: the pid in question.
3946 * @uattr: structure containing the extended parameters.
3947 * @size: sizeof(attr) for fwd/bwd comp.
3948 * @flags: for future extension.
3950 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3951 unsigned int, size, unsigned int, flags)
3953 struct sched_attr attr = {
3954 .size = sizeof(struct sched_attr),
3956 struct task_struct *p;
3959 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3960 size < SCHED_ATTR_SIZE_VER0 || flags)
3964 p = find_process_by_pid(pid);
3969 retval = security_task_getscheduler(p);
3973 attr.sched_policy = p->policy;
3974 if (p->sched_reset_on_fork)
3975 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3976 if (task_has_dl_policy(p))
3977 __getparam_dl(p, &attr);
3978 else if (task_has_rt_policy(p))
3979 attr.sched_priority = p->rt_priority;
3981 attr.sched_nice = task_nice(p);
3985 retval = sched_read_attr(uattr, &attr, size);
3993 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3995 cpumask_var_t cpus_allowed, new_mask;
3996 struct task_struct *p;
4001 p = find_process_by_pid(pid);
4007 /* Prevent p going away */
4011 if (p->flags & PF_NO_SETAFFINITY) {
4015 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4019 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4021 goto out_free_cpus_allowed;
4024 if (!check_same_owner(p)) {
4026 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4033 retval = security_task_setscheduler(p);
4038 cpuset_cpus_allowed(p, cpus_allowed);
4039 cpumask_and(new_mask, in_mask, cpus_allowed);
4042 * Since bandwidth control happens on root_domain basis,
4043 * if admission test is enabled, we only admit -deadline
4044 * tasks allowed to run on all the CPUs in the task's
4048 if (task_has_dl_policy(p)) {
4049 const struct cpumask *span = task_rq(p)->rd->span;
4051 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
4058 retval = set_cpus_allowed_ptr(p, new_mask);
4061 cpuset_cpus_allowed(p, cpus_allowed);
4062 if (!cpumask_subset(new_mask, cpus_allowed)) {
4064 * We must have raced with a concurrent cpuset
4065 * update. Just reset the cpus_allowed to the
4066 * cpuset's cpus_allowed
4068 cpumask_copy(new_mask, cpus_allowed);
4073 free_cpumask_var(new_mask);
4074 out_free_cpus_allowed:
4075 free_cpumask_var(cpus_allowed);
4081 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4082 struct cpumask *new_mask)
4084 if (len < cpumask_size())
4085 cpumask_clear(new_mask);
4086 else if (len > cpumask_size())
4087 len = cpumask_size();
4089 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4093 * sys_sched_setaffinity - set the cpu affinity of a process
4094 * @pid: pid of the process
4095 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4096 * @user_mask_ptr: user-space pointer to the new cpu mask
4098 * Return: 0 on success. An error code otherwise.
4100 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4101 unsigned long __user *, user_mask_ptr)
4103 cpumask_var_t new_mask;
4106 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4109 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4111 retval = sched_setaffinity(pid, new_mask);
4112 free_cpumask_var(new_mask);
4116 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4118 struct task_struct *p;
4119 unsigned long flags;
4125 p = find_process_by_pid(pid);
4129 retval = security_task_getscheduler(p);
4133 raw_spin_lock_irqsave(&p->pi_lock, flags);
4134 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4135 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4144 * sys_sched_getaffinity - get the cpu affinity of a process
4145 * @pid: pid of the process
4146 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4147 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4149 * Return: 0 on success. An error code otherwise.
4151 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4152 unsigned long __user *, user_mask_ptr)
4157 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4159 if (len & (sizeof(unsigned long)-1))
4162 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4165 ret = sched_getaffinity(pid, mask);
4167 size_t retlen = min_t(size_t, len, cpumask_size());
4169 if (copy_to_user(user_mask_ptr, mask, retlen))
4174 free_cpumask_var(mask);
4180 * sys_sched_yield - yield the current processor to other threads.
4182 * This function yields the current CPU to other tasks. If there are no
4183 * other threads running on this CPU then this function will return.
4187 SYSCALL_DEFINE0(sched_yield)
4189 struct rq *rq = this_rq_lock();
4191 schedstat_inc(rq, yld_count);
4192 current->sched_class->yield_task(rq);
4195 * Since we are going to call schedule() anyway, there's
4196 * no need to preempt or enable interrupts:
4198 __release(rq->lock);
4199 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4200 do_raw_spin_unlock(&rq->lock);
4201 sched_preempt_enable_no_resched();
4208 static void __cond_resched(void)
4210 __preempt_count_add(PREEMPT_ACTIVE);
4212 __preempt_count_sub(PREEMPT_ACTIVE);
4215 int __sched _cond_resched(void)
4217 if (should_resched()) {
4223 EXPORT_SYMBOL(_cond_resched);
4226 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4227 * call schedule, and on return reacquire the lock.
4229 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4230 * operations here to prevent schedule() from being called twice (once via
4231 * spin_unlock(), once by hand).
4233 int __cond_resched_lock(spinlock_t *lock)
4235 int resched = should_resched();
4238 lockdep_assert_held(lock);
4240 if (spin_needbreak(lock) || resched) {
4251 EXPORT_SYMBOL(__cond_resched_lock);
4253 int __sched __cond_resched_softirq(void)
4255 BUG_ON(!in_softirq());
4257 if (should_resched()) {
4265 EXPORT_SYMBOL(__cond_resched_softirq);
4268 * yield - yield the current processor to other threads.
4270 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4272 * The scheduler is at all times free to pick the calling task as the most
4273 * eligible task to run, if removing the yield() call from your code breaks
4274 * it, its already broken.
4276 * Typical broken usage is:
4281 * where one assumes that yield() will let 'the other' process run that will
4282 * make event true. If the current task is a SCHED_FIFO task that will never
4283 * happen. Never use yield() as a progress guarantee!!
4285 * If you want to use yield() to wait for something, use wait_event().
4286 * If you want to use yield() to be 'nice' for others, use cond_resched().
4287 * If you still want to use yield(), do not!
4289 void __sched yield(void)
4291 set_current_state(TASK_RUNNING);
4294 EXPORT_SYMBOL(yield);
4297 * yield_to - yield the current processor to another thread in
4298 * your thread group, or accelerate that thread toward the
4299 * processor it's on.
4301 * @preempt: whether task preemption is allowed or not
4303 * It's the caller's job to ensure that the target task struct
4304 * can't go away on us before we can do any checks.
4307 * true (>0) if we indeed boosted the target task.
4308 * false (0) if we failed to boost the target.
4309 * -ESRCH if there's no task to yield to.
4311 int __sched yield_to(struct task_struct *p, bool preempt)
4313 struct task_struct *curr = current;
4314 struct rq *rq, *p_rq;
4315 unsigned long flags;
4318 local_irq_save(flags);
4324 * If we're the only runnable task on the rq and target rq also
4325 * has only one task, there's absolutely no point in yielding.
4327 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4332 double_rq_lock(rq, p_rq);
4333 if (task_rq(p) != p_rq) {
4334 double_rq_unlock(rq, p_rq);
4338 if (!curr->sched_class->yield_to_task)
4341 if (curr->sched_class != p->sched_class)
4344 if (task_running(p_rq, p) || p->state)
4347 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4349 schedstat_inc(rq, yld_count);
4351 * Make p's CPU reschedule; pick_next_entity takes care of
4354 if (preempt && rq != p_rq)
4359 double_rq_unlock(rq, p_rq);
4361 local_irq_restore(flags);
4368 EXPORT_SYMBOL_GPL(yield_to);
4371 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4372 * that process accounting knows that this is a task in IO wait state.
4374 void __sched io_schedule(void)
4376 struct rq *rq = raw_rq();
4378 delayacct_blkio_start();
4379 atomic_inc(&rq->nr_iowait);
4380 blk_flush_plug(current);
4381 current->in_iowait = 1;
4383 current->in_iowait = 0;
4384 atomic_dec(&rq->nr_iowait);
4385 delayacct_blkio_end();
4387 EXPORT_SYMBOL(io_schedule);
4389 long __sched io_schedule_timeout(long timeout)
4391 struct rq *rq = raw_rq();
4394 delayacct_blkio_start();
4395 atomic_inc(&rq->nr_iowait);
4396 blk_flush_plug(current);
4397 current->in_iowait = 1;
4398 ret = schedule_timeout(timeout);
4399 current->in_iowait = 0;
4400 atomic_dec(&rq->nr_iowait);
4401 delayacct_blkio_end();
4406 * sys_sched_get_priority_max - return maximum RT priority.
4407 * @policy: scheduling class.
4409 * Return: On success, this syscall returns the maximum
4410 * rt_priority that can be used by a given scheduling class.
4411 * On failure, a negative error code is returned.
4413 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4420 ret = MAX_USER_RT_PRIO-1;
4422 case SCHED_DEADLINE:
4433 * sys_sched_get_priority_min - return minimum RT priority.
4434 * @policy: scheduling class.
4436 * Return: On success, this syscall returns the minimum
4437 * rt_priority that can be used by a given scheduling class.
4438 * On failure, a negative error code is returned.
4440 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4449 case SCHED_DEADLINE:
4459 * sys_sched_rr_get_interval - return the default timeslice of a process.
4460 * @pid: pid of the process.
4461 * @interval: userspace pointer to the timeslice value.
4463 * this syscall writes the default timeslice value of a given process
4464 * into the user-space timespec buffer. A value of '0' means infinity.
4466 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4469 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4470 struct timespec __user *, interval)
4472 struct task_struct *p;
4473 unsigned int time_slice;
4474 unsigned long flags;
4484 p = find_process_by_pid(pid);
4488 retval = security_task_getscheduler(p);
4492 rq = task_rq_lock(p, &flags);
4494 if (p->sched_class->get_rr_interval)
4495 time_slice = p->sched_class->get_rr_interval(rq, p);
4496 task_rq_unlock(rq, p, &flags);
4499 jiffies_to_timespec(time_slice, &t);
4500 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4508 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4510 void sched_show_task(struct task_struct *p)
4512 unsigned long free = 0;
4516 state = p->state ? __ffs(p->state) + 1 : 0;
4517 printk(KERN_INFO "%-15.15s %c", p->comm,
4518 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4519 #if BITS_PER_LONG == 32
4520 if (state == TASK_RUNNING)
4521 printk(KERN_CONT " running ");
4523 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4525 if (state == TASK_RUNNING)
4526 printk(KERN_CONT " running task ");
4528 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4530 #ifdef CONFIG_DEBUG_STACK_USAGE
4531 free = stack_not_used(p);
4534 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4536 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4537 task_pid_nr(p), ppid,
4538 (unsigned long)task_thread_info(p)->flags);
4540 print_worker_info(KERN_INFO, p);
4541 show_stack(p, NULL);
4544 void show_state_filter(unsigned long state_filter)
4546 struct task_struct *g, *p;
4548 #if BITS_PER_LONG == 32
4550 " task PC stack pid father\n");
4553 " task PC stack pid father\n");
4556 for_each_process_thread(g, p) {
4558 * reset the NMI-timeout, listing all files on a slow
4559 * console might take a lot of time:
4561 touch_nmi_watchdog();
4562 if (!state_filter || (p->state & state_filter))
4566 touch_all_softlockup_watchdogs();
4568 #ifdef CONFIG_SCHED_DEBUG
4569 sysrq_sched_debug_show();
4573 * Only show locks if all tasks are dumped:
4576 debug_show_all_locks();
4579 void init_idle_bootup_task(struct task_struct *idle)
4581 idle->sched_class = &idle_sched_class;
4585 * init_idle - set up an idle thread for a given CPU
4586 * @idle: task in question
4587 * @cpu: cpu the idle task belongs to
4589 * NOTE: this function does not set the idle thread's NEED_RESCHED
4590 * flag, to make booting more robust.
4592 void init_idle(struct task_struct *idle, int cpu)
4594 struct rq *rq = cpu_rq(cpu);
4595 unsigned long flags;
4597 raw_spin_lock_irqsave(&rq->lock, flags);
4599 __sched_fork(0, idle);
4600 idle->state = TASK_RUNNING;
4601 idle->se.exec_start = sched_clock();
4603 do_set_cpus_allowed(idle, cpumask_of(cpu));
4605 * We're having a chicken and egg problem, even though we are
4606 * holding rq->lock, the cpu isn't yet set to this cpu so the
4607 * lockdep check in task_group() will fail.
4609 * Similar case to sched_fork(). / Alternatively we could
4610 * use task_rq_lock() here and obtain the other rq->lock.
4615 __set_task_cpu(idle, cpu);
4618 rq->curr = rq->idle = idle;
4619 idle->on_rq = TASK_ON_RQ_QUEUED;
4620 #if defined(CONFIG_SMP)
4623 raw_spin_unlock_irqrestore(&rq->lock, flags);
4625 /* Set the preempt count _outside_ the spinlocks! */
4626 init_idle_preempt_count(idle, cpu);
4629 * The idle tasks have their own, simple scheduling class:
4631 idle->sched_class = &idle_sched_class;
4632 ftrace_graph_init_idle_task(idle, cpu);
4633 vtime_init_idle(idle, cpu);
4634 #if defined(CONFIG_SMP)
4635 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4641 * move_queued_task - move a queued task to new rq.
4643 * Returns (locked) new rq. Old rq's lock is released.
4645 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4647 struct rq *rq = task_rq(p);
4649 lockdep_assert_held(&rq->lock);
4651 dequeue_task(rq, p, 0);
4652 p->on_rq = TASK_ON_RQ_MIGRATING;
4653 set_task_cpu(p, new_cpu);
4654 raw_spin_unlock(&rq->lock);
4656 rq = cpu_rq(new_cpu);
4658 raw_spin_lock(&rq->lock);
4659 BUG_ON(task_cpu(p) != new_cpu);
4660 p->on_rq = TASK_ON_RQ_QUEUED;
4661 enqueue_task(rq, p, 0);
4662 check_preempt_curr(rq, p, 0);
4667 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4669 if (p->sched_class && p->sched_class->set_cpus_allowed)
4670 p->sched_class->set_cpus_allowed(p, new_mask);
4672 cpumask_copy(&p->cpus_allowed, new_mask);
4673 p->nr_cpus_allowed = cpumask_weight(new_mask);
4677 * This is how migration works:
4679 * 1) we invoke migration_cpu_stop() on the target CPU using
4681 * 2) stopper starts to run (implicitly forcing the migrated thread
4683 * 3) it checks whether the migrated task is still in the wrong runqueue.
4684 * 4) if it's in the wrong runqueue then the migration thread removes
4685 * it and puts it into the right queue.
4686 * 5) stopper completes and stop_one_cpu() returns and the migration
4691 * Change a given task's CPU affinity. Migrate the thread to a
4692 * proper CPU and schedule it away if the CPU it's executing on
4693 * is removed from the allowed bitmask.
4695 * NOTE: the caller must have a valid reference to the task, the
4696 * task must not exit() & deallocate itself prematurely. The
4697 * call is not atomic; no spinlocks may be held.
4699 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4701 unsigned long flags;
4703 unsigned int dest_cpu;
4706 rq = task_rq_lock(p, &flags);
4708 if (cpumask_equal(&p->cpus_allowed, new_mask))
4711 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4716 do_set_cpus_allowed(p, new_mask);
4718 /* Can the task run on the task's current CPU? If so, we're done */
4719 if (cpumask_test_cpu(task_cpu(p), new_mask))
4722 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4723 if (task_running(rq, p) || p->state == TASK_WAKING) {
4724 struct migration_arg arg = { p, dest_cpu };
4725 /* Need help from migration thread: drop lock and wait. */
4726 task_rq_unlock(rq, p, &flags);
4727 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4728 tlb_migrate_finish(p->mm);
4730 } else if (task_on_rq_queued(p))
4731 rq = move_queued_task(p, dest_cpu);
4733 task_rq_unlock(rq, p, &flags);
4737 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4740 * Move (not current) task off this cpu, onto dest cpu. We're doing
4741 * this because either it can't run here any more (set_cpus_allowed()
4742 * away from this CPU, or CPU going down), or because we're
4743 * attempting to rebalance this task on exec (sched_exec).
4745 * So we race with normal scheduler movements, but that's OK, as long
4746 * as the task is no longer on this CPU.
4748 * Returns non-zero if task was successfully migrated.
4750 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4755 if (unlikely(!cpu_active(dest_cpu)))
4758 rq = cpu_rq(src_cpu);
4760 raw_spin_lock(&p->pi_lock);
4761 raw_spin_lock(&rq->lock);
4762 /* Already moved. */
4763 if (task_cpu(p) != src_cpu)
4766 /* Affinity changed (again). */
4767 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4771 * If we're not on a rq, the next wake-up will ensure we're
4774 if (task_on_rq_queued(p))
4775 rq = move_queued_task(p, dest_cpu);
4779 raw_spin_unlock(&rq->lock);
4780 raw_spin_unlock(&p->pi_lock);
4784 #ifdef CONFIG_NUMA_BALANCING
4785 /* Migrate current task p to target_cpu */
4786 int migrate_task_to(struct task_struct *p, int target_cpu)
4788 struct migration_arg arg = { p, target_cpu };
4789 int curr_cpu = task_cpu(p);
4791 if (curr_cpu == target_cpu)
4794 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4797 /* TODO: This is not properly updating schedstats */
4799 trace_sched_move_numa(p, curr_cpu, target_cpu);
4800 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4804 * Requeue a task on a given node and accurately track the number of NUMA
4805 * tasks on the runqueues
4807 void sched_setnuma(struct task_struct *p, int nid)
4810 unsigned long flags;
4811 bool queued, running;
4813 rq = task_rq_lock(p, &flags);
4814 queued = task_on_rq_queued(p);
4815 running = task_current(rq, p);
4818 dequeue_task(rq, p, 0);
4820 put_prev_task(rq, p);
4822 p->numa_preferred_nid = nid;
4825 p->sched_class->set_curr_task(rq);
4827 enqueue_task(rq, p, 0);
4828 task_rq_unlock(rq, p, &flags);
4833 * migration_cpu_stop - this will be executed by a highprio stopper thread
4834 * and performs thread migration by bumping thread off CPU then
4835 * 'pushing' onto another runqueue.
4837 static int migration_cpu_stop(void *data)
4839 struct migration_arg *arg = data;
4842 * The original target cpu might have gone down and we might
4843 * be on another cpu but it doesn't matter.
4845 local_irq_disable();
4847 * We need to explicitly wake pending tasks before running
4848 * __migrate_task() such that we will not miss enforcing cpus_allowed
4849 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4851 sched_ttwu_pending();
4852 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4857 #ifdef CONFIG_HOTPLUG_CPU
4860 * Ensures that the idle task is using init_mm right before its cpu goes
4863 void idle_task_exit(void)
4865 struct mm_struct *mm = current->active_mm;
4867 BUG_ON(cpu_online(smp_processor_id()));
4869 if (mm != &init_mm) {
4870 switch_mm(mm, &init_mm, current);
4871 finish_arch_post_lock_switch();
4877 * Since this CPU is going 'away' for a while, fold any nr_active delta
4878 * we might have. Assumes we're called after migrate_tasks() so that the
4879 * nr_active count is stable.
4881 * Also see the comment "Global load-average calculations".
4883 static void calc_load_migrate(struct rq *rq)
4885 long delta = calc_load_fold_active(rq);
4887 atomic_long_add(delta, &calc_load_tasks);
4890 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4894 static const struct sched_class fake_sched_class = {
4895 .put_prev_task = put_prev_task_fake,
4898 static struct task_struct fake_task = {
4900 * Avoid pull_{rt,dl}_task()
4902 .prio = MAX_PRIO + 1,
4903 .sched_class = &fake_sched_class,
4907 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4908 * try_to_wake_up()->select_task_rq().
4910 * Called with rq->lock held even though we'er in stop_machine() and
4911 * there's no concurrency possible, we hold the required locks anyway
4912 * because of lock validation efforts.
4914 static void migrate_tasks(unsigned int dead_cpu)
4916 struct rq *rq = cpu_rq(dead_cpu);
4917 struct task_struct *next, *stop = rq->stop;
4921 * Fudge the rq selection such that the below task selection loop
4922 * doesn't get stuck on the currently eligible stop task.
4924 * We're currently inside stop_machine() and the rq is either stuck
4925 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4926 * either way we should never end up calling schedule() until we're
4932 * put_prev_task() and pick_next_task() sched
4933 * class method both need to have an up-to-date
4934 * value of rq->clock[_task]
4936 update_rq_clock(rq);
4940 * There's this thread running, bail when that's the only
4943 if (rq->nr_running == 1)
4946 next = pick_next_task(rq, &fake_task);
4948 next->sched_class->put_prev_task(rq, next);
4950 /* Find suitable destination for @next, with force if needed. */
4951 dest_cpu = select_fallback_rq(dead_cpu, next);
4952 raw_spin_unlock(&rq->lock);
4954 __migrate_task(next, dead_cpu, dest_cpu);
4956 raw_spin_lock(&rq->lock);
4962 #endif /* CONFIG_HOTPLUG_CPU */
4964 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4966 static struct ctl_table sd_ctl_dir[] = {
4968 .procname = "sched_domain",
4974 static struct ctl_table sd_ctl_root[] = {
4976 .procname = "kernel",
4978 .child = sd_ctl_dir,
4983 static struct ctl_table *sd_alloc_ctl_entry(int n)
4985 struct ctl_table *entry =
4986 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4991 static void sd_free_ctl_entry(struct ctl_table **tablep)
4993 struct ctl_table *entry;
4996 * In the intermediate directories, both the child directory and
4997 * procname are dynamically allocated and could fail but the mode
4998 * will always be set. In the lowest directory the names are
4999 * static strings and all have proc handlers.
5001 for (entry = *tablep; entry->mode; entry++) {
5003 sd_free_ctl_entry(&entry->child);
5004 if (entry->proc_handler == NULL)
5005 kfree(entry->procname);
5012 static int min_load_idx = 0;
5013 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5016 set_table_entry(struct ctl_table *entry,
5017 const char *procname, void *data, int maxlen,
5018 umode_t mode, proc_handler *proc_handler,
5021 entry->procname = procname;
5023 entry->maxlen = maxlen;
5025 entry->proc_handler = proc_handler;
5028 entry->extra1 = &min_load_idx;
5029 entry->extra2 = &max_load_idx;
5033 static struct ctl_table *
5034 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5036 struct ctl_table *table = sd_alloc_ctl_entry(14);
5041 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5042 sizeof(long), 0644, proc_doulongvec_minmax, false);
5043 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5044 sizeof(long), 0644, proc_doulongvec_minmax, false);
5045 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5046 sizeof(int), 0644, proc_dointvec_minmax, true);
5047 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5048 sizeof(int), 0644, proc_dointvec_minmax, true);
5049 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5050 sizeof(int), 0644, proc_dointvec_minmax, true);
5051 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5052 sizeof(int), 0644, proc_dointvec_minmax, true);
5053 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5054 sizeof(int), 0644, proc_dointvec_minmax, true);
5055 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5056 sizeof(int), 0644, proc_dointvec_minmax, false);
5057 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5058 sizeof(int), 0644, proc_dointvec_minmax, false);
5059 set_table_entry(&table[9], "cache_nice_tries",
5060 &sd->cache_nice_tries,
5061 sizeof(int), 0644, proc_dointvec_minmax, false);
5062 set_table_entry(&table[10], "flags", &sd->flags,
5063 sizeof(int), 0644, proc_dointvec_minmax, false);
5064 set_table_entry(&table[11], "max_newidle_lb_cost",
5065 &sd->max_newidle_lb_cost,
5066 sizeof(long), 0644, proc_doulongvec_minmax, false);
5067 set_table_entry(&table[12], "name", sd->name,
5068 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5069 /* &table[13] is terminator */
5074 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5076 struct ctl_table *entry, *table;
5077 struct sched_domain *sd;
5078 int domain_num = 0, i;
5081 for_each_domain(cpu, sd)
5083 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5088 for_each_domain(cpu, sd) {
5089 snprintf(buf, 32, "domain%d", i);
5090 entry->procname = kstrdup(buf, GFP_KERNEL);
5092 entry->child = sd_alloc_ctl_domain_table(sd);
5099 static struct ctl_table_header *sd_sysctl_header;
5100 static void register_sched_domain_sysctl(void)
5102 int i, cpu_num = num_possible_cpus();
5103 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5106 WARN_ON(sd_ctl_dir[0].child);
5107 sd_ctl_dir[0].child = entry;
5112 for_each_possible_cpu(i) {
5113 snprintf(buf, 32, "cpu%d", i);
5114 entry->procname = kstrdup(buf, GFP_KERNEL);
5116 entry->child = sd_alloc_ctl_cpu_table(i);
5120 WARN_ON(sd_sysctl_header);
5121 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5124 /* may be called multiple times per register */
5125 static void unregister_sched_domain_sysctl(void)
5127 if (sd_sysctl_header)
5128 unregister_sysctl_table(sd_sysctl_header);
5129 sd_sysctl_header = NULL;
5130 if (sd_ctl_dir[0].child)
5131 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5134 static void register_sched_domain_sysctl(void)
5137 static void unregister_sched_domain_sysctl(void)
5142 static void set_rq_online(struct rq *rq)
5145 const struct sched_class *class;
5147 cpumask_set_cpu(rq->cpu, rq->rd->online);
5150 for_each_class(class) {
5151 if (class->rq_online)
5152 class->rq_online(rq);
5157 static void set_rq_offline(struct rq *rq)
5160 const struct sched_class *class;
5162 for_each_class(class) {
5163 if (class->rq_offline)
5164 class->rq_offline(rq);
5167 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5173 * migration_call - callback that gets triggered when a CPU is added.
5174 * Here we can start up the necessary migration thread for the new CPU.
5177 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5179 int cpu = (long)hcpu;
5180 unsigned long flags;
5181 struct rq *rq = cpu_rq(cpu);
5183 switch (action & ~CPU_TASKS_FROZEN) {
5185 case CPU_UP_PREPARE:
5186 rq->calc_load_update = calc_load_update;
5190 /* Update our root-domain */
5191 raw_spin_lock_irqsave(&rq->lock, flags);
5193 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5197 raw_spin_unlock_irqrestore(&rq->lock, flags);
5200 #ifdef CONFIG_HOTPLUG_CPU
5202 sched_ttwu_pending();
5203 /* Update our root-domain */
5204 raw_spin_lock_irqsave(&rq->lock, flags);
5206 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5210 BUG_ON(rq->nr_running != 1); /* the migration thread */
5211 raw_spin_unlock_irqrestore(&rq->lock, flags);
5215 calc_load_migrate(rq);
5220 update_max_interval();
5226 * Register at high priority so that task migration (migrate_all_tasks)
5227 * happens before everything else. This has to be lower priority than
5228 * the notifier in the perf_event subsystem, though.
5230 static struct notifier_block migration_notifier = {
5231 .notifier_call = migration_call,
5232 .priority = CPU_PRI_MIGRATION,
5235 static void __cpuinit set_cpu_rq_start_time(void)
5237 int cpu = smp_processor_id();
5238 struct rq *rq = cpu_rq(cpu);
5239 rq->age_stamp = sched_clock_cpu(cpu);
5242 static int sched_cpu_active(struct notifier_block *nfb,
5243 unsigned long action, void *hcpu)
5245 switch (action & ~CPU_TASKS_FROZEN) {
5247 set_cpu_rq_start_time();
5249 case CPU_DOWN_FAILED:
5250 set_cpu_active((long)hcpu, true);
5257 static int sched_cpu_inactive(struct notifier_block *nfb,
5258 unsigned long action, void *hcpu)
5260 unsigned long flags;
5261 long cpu = (long)hcpu;
5263 switch (action & ~CPU_TASKS_FROZEN) {
5264 case CPU_DOWN_PREPARE:
5265 set_cpu_active(cpu, false);
5267 /* explicitly allow suspend */
5268 if (!(action & CPU_TASKS_FROZEN)) {
5269 struct dl_bw *dl_b = dl_bw_of(cpu);
5273 raw_spin_lock_irqsave(&dl_b->lock, flags);
5274 cpus = dl_bw_cpus(cpu);
5275 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5276 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5279 return notifier_from_errno(-EBUSY);
5287 static int __init migration_init(void)
5289 void *cpu = (void *)(long)smp_processor_id();
5292 /* Initialize migration for the boot CPU */
5293 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5294 BUG_ON(err == NOTIFY_BAD);
5295 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5296 register_cpu_notifier(&migration_notifier);
5298 /* Register cpu active notifiers */
5299 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5300 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5304 early_initcall(migration_init);
5309 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5311 #ifdef CONFIG_SCHED_DEBUG
5313 static __read_mostly int sched_debug_enabled;
5315 static int __init sched_debug_setup(char *str)
5317 sched_debug_enabled = 1;
5321 early_param("sched_debug", sched_debug_setup);
5323 static inline bool sched_debug(void)
5325 return sched_debug_enabled;
5328 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5329 struct cpumask *groupmask)
5331 struct sched_group *group = sd->groups;
5334 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5335 cpumask_clear(groupmask);
5337 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5339 if (!(sd->flags & SD_LOAD_BALANCE)) {
5340 printk("does not load-balance\n");
5342 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5347 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5349 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5350 printk(KERN_ERR "ERROR: domain->span does not contain "
5353 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5354 printk(KERN_ERR "ERROR: domain->groups does not contain"
5358 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5362 printk(KERN_ERR "ERROR: group is NULL\n");
5367 * Even though we initialize ->capacity to something semi-sane,
5368 * we leave capacity_orig unset. This allows us to detect if
5369 * domain iteration is still funny without causing /0 traps.
5371 if (!group->sgc->capacity_orig) {
5372 printk(KERN_CONT "\n");
5373 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5377 if (!cpumask_weight(sched_group_cpus(group))) {
5378 printk(KERN_CONT "\n");
5379 printk(KERN_ERR "ERROR: empty group\n");
5383 if (!(sd->flags & SD_OVERLAP) &&
5384 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5385 printk(KERN_CONT "\n");
5386 printk(KERN_ERR "ERROR: repeated CPUs\n");
5390 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5392 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5394 printk(KERN_CONT " %s", str);
5395 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5396 printk(KERN_CONT " (cpu_capacity = %d)",
5397 group->sgc->capacity);
5400 group = group->next;
5401 } while (group != sd->groups);
5402 printk(KERN_CONT "\n");
5404 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5405 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5408 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5409 printk(KERN_ERR "ERROR: parent span is not a superset "
5410 "of domain->span\n");
5414 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5418 if (!sched_debug_enabled)
5422 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5426 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5429 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5437 #else /* !CONFIG_SCHED_DEBUG */
5438 # define sched_domain_debug(sd, cpu) do { } while (0)
5439 static inline bool sched_debug(void)
5443 #endif /* CONFIG_SCHED_DEBUG */
5445 static int sd_degenerate(struct sched_domain *sd)
5447 if (cpumask_weight(sched_domain_span(sd)) == 1)
5450 /* Following flags need at least 2 groups */
5451 if (sd->flags & (SD_LOAD_BALANCE |
5452 SD_BALANCE_NEWIDLE |
5455 SD_SHARE_CPUCAPACITY |
5456 SD_SHARE_PKG_RESOURCES |
5457 SD_SHARE_POWERDOMAIN)) {
5458 if (sd->groups != sd->groups->next)
5462 /* Following flags don't use groups */
5463 if (sd->flags & (SD_WAKE_AFFINE))
5470 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5472 unsigned long cflags = sd->flags, pflags = parent->flags;
5474 if (sd_degenerate(parent))
5477 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5480 /* Flags needing groups don't count if only 1 group in parent */
5481 if (parent->groups == parent->groups->next) {
5482 pflags &= ~(SD_LOAD_BALANCE |
5483 SD_BALANCE_NEWIDLE |
5486 SD_SHARE_CPUCAPACITY |
5487 SD_SHARE_PKG_RESOURCES |
5489 SD_SHARE_POWERDOMAIN);
5490 if (nr_node_ids == 1)
5491 pflags &= ~SD_SERIALIZE;
5493 if (~cflags & pflags)
5499 static void free_rootdomain(struct rcu_head *rcu)
5501 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5503 cpupri_cleanup(&rd->cpupri);
5504 cpudl_cleanup(&rd->cpudl);
5505 free_cpumask_var(rd->dlo_mask);
5506 free_cpumask_var(rd->rto_mask);
5507 free_cpumask_var(rd->online);
5508 free_cpumask_var(rd->span);
5512 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5514 struct root_domain *old_rd = NULL;
5515 unsigned long flags;
5517 raw_spin_lock_irqsave(&rq->lock, flags);
5522 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5525 cpumask_clear_cpu(rq->cpu, old_rd->span);
5528 * If we dont want to free the old_rd yet then
5529 * set old_rd to NULL to skip the freeing later
5532 if (!atomic_dec_and_test(&old_rd->refcount))
5536 atomic_inc(&rd->refcount);
5539 cpumask_set_cpu(rq->cpu, rd->span);
5540 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5543 raw_spin_unlock_irqrestore(&rq->lock, flags);
5546 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5549 static int init_rootdomain(struct root_domain *rd)
5551 memset(rd, 0, sizeof(*rd));
5553 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5555 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5557 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5559 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5562 init_dl_bw(&rd->dl_bw);
5563 if (cpudl_init(&rd->cpudl) != 0)
5566 if (cpupri_init(&rd->cpupri) != 0)
5571 free_cpumask_var(rd->rto_mask);
5573 free_cpumask_var(rd->dlo_mask);
5575 free_cpumask_var(rd->online);
5577 free_cpumask_var(rd->span);
5583 * By default the system creates a single root-domain with all cpus as
5584 * members (mimicking the global state we have today).
5586 struct root_domain def_root_domain;
5588 static void init_defrootdomain(void)
5590 init_rootdomain(&def_root_domain);
5592 atomic_set(&def_root_domain.refcount, 1);
5595 static struct root_domain *alloc_rootdomain(void)
5597 struct root_domain *rd;
5599 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5603 if (init_rootdomain(rd) != 0) {
5611 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5613 struct sched_group *tmp, *first;
5622 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5627 } while (sg != first);
5630 static void free_sched_domain(struct rcu_head *rcu)
5632 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5635 * If its an overlapping domain it has private groups, iterate and
5638 if (sd->flags & SD_OVERLAP) {
5639 free_sched_groups(sd->groups, 1);
5640 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5641 kfree(sd->groups->sgc);
5647 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5649 call_rcu(&sd->rcu, free_sched_domain);
5652 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5654 for (; sd; sd = sd->parent)
5655 destroy_sched_domain(sd, cpu);
5659 * Keep a special pointer to the highest sched_domain that has
5660 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5661 * allows us to avoid some pointer chasing select_idle_sibling().
5663 * Also keep a unique ID per domain (we use the first cpu number in
5664 * the cpumask of the domain), this allows us to quickly tell if
5665 * two cpus are in the same cache domain, see cpus_share_cache().
5667 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5668 DEFINE_PER_CPU(int, sd_llc_size);
5669 DEFINE_PER_CPU(int, sd_llc_id);
5670 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5671 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5672 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5674 static void update_top_cache_domain(int cpu)
5676 struct sched_domain *sd;
5677 struct sched_domain *busy_sd = NULL;
5681 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5683 id = cpumask_first(sched_domain_span(sd));
5684 size = cpumask_weight(sched_domain_span(sd));
5685 busy_sd = sd->parent; /* sd_busy */
5687 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5689 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5690 per_cpu(sd_llc_size, cpu) = size;
5691 per_cpu(sd_llc_id, cpu) = id;
5693 sd = lowest_flag_domain(cpu, SD_NUMA);
5694 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5696 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5697 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5701 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5702 * hold the hotplug lock.
5705 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5707 struct rq *rq = cpu_rq(cpu);
5708 struct sched_domain *tmp;
5710 /* Remove the sched domains which do not contribute to scheduling. */
5711 for (tmp = sd; tmp; ) {
5712 struct sched_domain *parent = tmp->parent;
5716 if (sd_parent_degenerate(tmp, parent)) {
5717 tmp->parent = parent->parent;
5719 parent->parent->child = tmp;
5721 * Transfer SD_PREFER_SIBLING down in case of a
5722 * degenerate parent; the spans match for this
5723 * so the property transfers.
5725 if (parent->flags & SD_PREFER_SIBLING)
5726 tmp->flags |= SD_PREFER_SIBLING;
5727 destroy_sched_domain(parent, cpu);
5732 if (sd && sd_degenerate(sd)) {
5735 destroy_sched_domain(tmp, cpu);
5740 sched_domain_debug(sd, cpu);
5742 rq_attach_root(rq, rd);
5744 rcu_assign_pointer(rq->sd, sd);
5745 destroy_sched_domains(tmp, cpu);
5747 update_top_cache_domain(cpu);
5750 /* cpus with isolated domains */
5751 static cpumask_var_t cpu_isolated_map;
5753 /* Setup the mask of cpus configured for isolated domains */
5754 static int __init isolated_cpu_setup(char *str)
5756 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5757 cpulist_parse(str, cpu_isolated_map);
5761 __setup("isolcpus=", isolated_cpu_setup);
5764 struct sched_domain ** __percpu sd;
5765 struct root_domain *rd;
5776 * Build an iteration mask that can exclude certain CPUs from the upwards
5779 * Asymmetric node setups can result in situations where the domain tree is of
5780 * unequal depth, make sure to skip domains that already cover the entire
5783 * In that case build_sched_domains() will have terminated the iteration early
5784 * and our sibling sd spans will be empty. Domains should always include the
5785 * cpu they're built on, so check that.
5788 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5790 const struct cpumask *span = sched_domain_span(sd);
5791 struct sd_data *sdd = sd->private;
5792 struct sched_domain *sibling;
5795 for_each_cpu(i, span) {
5796 sibling = *per_cpu_ptr(sdd->sd, i);
5797 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5800 cpumask_set_cpu(i, sched_group_mask(sg));
5805 * Return the canonical balance cpu for this group, this is the first cpu
5806 * of this group that's also in the iteration mask.
5808 int group_balance_cpu(struct sched_group *sg)
5810 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5814 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5816 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5817 const struct cpumask *span = sched_domain_span(sd);
5818 struct cpumask *covered = sched_domains_tmpmask;
5819 struct sd_data *sdd = sd->private;
5820 struct sched_domain *sibling;
5823 cpumask_clear(covered);
5825 for_each_cpu(i, span) {
5826 struct cpumask *sg_span;
5828 if (cpumask_test_cpu(i, covered))
5831 sibling = *per_cpu_ptr(sdd->sd, i);
5833 /* See the comment near build_group_mask(). */
5834 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5837 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5838 GFP_KERNEL, cpu_to_node(cpu));
5843 sg_span = sched_group_cpus(sg);
5845 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5847 cpumask_set_cpu(i, sg_span);
5849 cpumask_or(covered, covered, sg_span);
5851 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5852 if (atomic_inc_return(&sg->sgc->ref) == 1)
5853 build_group_mask(sd, sg);
5856 * Initialize sgc->capacity such that even if we mess up the
5857 * domains and no possible iteration will get us here, we won't
5860 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5861 sg->sgc->capacity_orig = sg->sgc->capacity;
5864 * Make sure the first group of this domain contains the
5865 * canonical balance cpu. Otherwise the sched_domain iteration
5866 * breaks. See update_sg_lb_stats().
5868 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5869 group_balance_cpu(sg) == cpu)
5879 sd->groups = groups;
5884 free_sched_groups(first, 0);
5889 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5891 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5892 struct sched_domain *child = sd->child;
5895 cpu = cpumask_first(sched_domain_span(child));
5898 *sg = *per_cpu_ptr(sdd->sg, cpu);
5899 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5900 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5907 * build_sched_groups will build a circular linked list of the groups
5908 * covered by the given span, and will set each group's ->cpumask correctly,
5909 * and ->cpu_capacity to 0.
5911 * Assumes the sched_domain tree is fully constructed
5914 build_sched_groups(struct sched_domain *sd, int cpu)
5916 struct sched_group *first = NULL, *last = NULL;
5917 struct sd_data *sdd = sd->private;
5918 const struct cpumask *span = sched_domain_span(sd);
5919 struct cpumask *covered;
5922 get_group(cpu, sdd, &sd->groups);
5923 atomic_inc(&sd->groups->ref);
5925 if (cpu != cpumask_first(span))
5928 lockdep_assert_held(&sched_domains_mutex);
5929 covered = sched_domains_tmpmask;
5931 cpumask_clear(covered);
5933 for_each_cpu(i, span) {
5934 struct sched_group *sg;
5937 if (cpumask_test_cpu(i, covered))
5940 group = get_group(i, sdd, &sg);
5941 cpumask_setall(sched_group_mask(sg));
5943 for_each_cpu(j, span) {
5944 if (get_group(j, sdd, NULL) != group)
5947 cpumask_set_cpu(j, covered);
5948 cpumask_set_cpu(j, sched_group_cpus(sg));
5963 * Initialize sched groups cpu_capacity.
5965 * cpu_capacity indicates the capacity of sched group, which is used while
5966 * distributing the load between different sched groups in a sched domain.
5967 * Typically cpu_capacity for all the groups in a sched domain will be same
5968 * unless there are asymmetries in the topology. If there are asymmetries,
5969 * group having more cpu_capacity will pickup more load compared to the
5970 * group having less cpu_capacity.
5972 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5974 struct sched_group *sg = sd->groups;
5979 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5981 } while (sg != sd->groups);
5983 if (cpu != group_balance_cpu(sg))
5986 update_group_capacity(sd, cpu);
5987 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
5991 * Initializers for schedule domains
5992 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5995 static int default_relax_domain_level = -1;
5996 int sched_domain_level_max;
5998 static int __init setup_relax_domain_level(char *str)
6000 if (kstrtoint(str, 0, &default_relax_domain_level))
6001 pr_warn("Unable to set relax_domain_level\n");
6005 __setup("relax_domain_level=", setup_relax_domain_level);
6007 static void set_domain_attribute(struct sched_domain *sd,
6008 struct sched_domain_attr *attr)
6012 if (!attr || attr->relax_domain_level < 0) {
6013 if (default_relax_domain_level < 0)
6016 request = default_relax_domain_level;
6018 request = attr->relax_domain_level;
6019 if (request < sd->level) {
6020 /* turn off idle balance on this domain */
6021 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6023 /* turn on idle balance on this domain */
6024 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6028 static void __sdt_free(const struct cpumask *cpu_map);
6029 static int __sdt_alloc(const struct cpumask *cpu_map);
6031 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6032 const struct cpumask *cpu_map)
6036 if (!atomic_read(&d->rd->refcount))
6037 free_rootdomain(&d->rd->rcu); /* fall through */
6039 free_percpu(d->sd); /* fall through */
6041 __sdt_free(cpu_map); /* fall through */
6047 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6048 const struct cpumask *cpu_map)
6050 memset(d, 0, sizeof(*d));
6052 if (__sdt_alloc(cpu_map))
6053 return sa_sd_storage;
6054 d->sd = alloc_percpu(struct sched_domain *);
6056 return sa_sd_storage;
6057 d->rd = alloc_rootdomain();
6060 return sa_rootdomain;
6064 * NULL the sd_data elements we've used to build the sched_domain and
6065 * sched_group structure so that the subsequent __free_domain_allocs()
6066 * will not free the data we're using.
6068 static void claim_allocations(int cpu, struct sched_domain *sd)
6070 struct sd_data *sdd = sd->private;
6072 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6073 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6075 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6076 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6078 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6079 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6083 static int sched_domains_numa_levels;
6084 static int *sched_domains_numa_distance;
6085 static struct cpumask ***sched_domains_numa_masks;
6086 static int sched_domains_curr_level;
6090 * SD_flags allowed in topology descriptions.
6092 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6093 * SD_SHARE_PKG_RESOURCES - describes shared caches
6094 * SD_NUMA - describes NUMA topologies
6095 * SD_SHARE_POWERDOMAIN - describes shared power domain
6098 * SD_ASYM_PACKING - describes SMT quirks
6100 #define TOPOLOGY_SD_FLAGS \
6101 (SD_SHARE_CPUCAPACITY | \
6102 SD_SHARE_PKG_RESOURCES | \
6105 SD_SHARE_POWERDOMAIN)
6107 static struct sched_domain *
6108 sd_init(struct sched_domain_topology_level *tl, int cpu)
6110 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6111 int sd_weight, sd_flags = 0;
6115 * Ugly hack to pass state to sd_numa_mask()...
6117 sched_domains_curr_level = tl->numa_level;
6120 sd_weight = cpumask_weight(tl->mask(cpu));
6123 sd_flags = (*tl->sd_flags)();
6124 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6125 "wrong sd_flags in topology description\n"))
6126 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6128 *sd = (struct sched_domain){
6129 .min_interval = sd_weight,
6130 .max_interval = 2*sd_weight,
6132 .imbalance_pct = 125,
6134 .cache_nice_tries = 0,
6141 .flags = 1*SD_LOAD_BALANCE
6142 | 1*SD_BALANCE_NEWIDLE
6147 | 0*SD_SHARE_CPUCAPACITY
6148 | 0*SD_SHARE_PKG_RESOURCES
6150 | 0*SD_PREFER_SIBLING
6155 .last_balance = jiffies,
6156 .balance_interval = sd_weight,
6158 .max_newidle_lb_cost = 0,
6159 .next_decay_max_lb_cost = jiffies,
6160 #ifdef CONFIG_SCHED_DEBUG
6166 * Convert topological properties into behaviour.
6169 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6170 sd->imbalance_pct = 110;
6171 sd->smt_gain = 1178; /* ~15% */
6173 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6174 sd->imbalance_pct = 117;
6175 sd->cache_nice_tries = 1;
6179 } else if (sd->flags & SD_NUMA) {
6180 sd->cache_nice_tries = 2;
6184 sd->flags |= SD_SERIALIZE;
6185 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6186 sd->flags &= ~(SD_BALANCE_EXEC |
6193 sd->flags |= SD_PREFER_SIBLING;
6194 sd->cache_nice_tries = 1;
6199 sd->private = &tl->data;
6205 * Topology list, bottom-up.
6207 static struct sched_domain_topology_level default_topology[] = {
6208 #ifdef CONFIG_SCHED_SMT
6209 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6211 #ifdef CONFIG_SCHED_MC
6212 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6214 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6218 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6220 #define for_each_sd_topology(tl) \
6221 for (tl = sched_domain_topology; tl->mask; tl++)
6223 void set_sched_topology(struct sched_domain_topology_level *tl)
6225 sched_domain_topology = tl;
6230 static const struct cpumask *sd_numa_mask(int cpu)
6232 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6235 static void sched_numa_warn(const char *str)
6237 static int done = false;
6245 printk(KERN_WARNING "ERROR: %s\n\n", str);
6247 for (i = 0; i < nr_node_ids; i++) {
6248 printk(KERN_WARNING " ");
6249 for (j = 0; j < nr_node_ids; j++)
6250 printk(KERN_CONT "%02d ", node_distance(i,j));
6251 printk(KERN_CONT "\n");
6253 printk(KERN_WARNING "\n");
6256 static bool find_numa_distance(int distance)
6260 if (distance == node_distance(0, 0))
6263 for (i = 0; i < sched_domains_numa_levels; i++) {
6264 if (sched_domains_numa_distance[i] == distance)
6271 static void sched_init_numa(void)
6273 int next_distance, curr_distance = node_distance(0, 0);
6274 struct sched_domain_topology_level *tl;
6278 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6279 if (!sched_domains_numa_distance)
6283 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6284 * unique distances in the node_distance() table.
6286 * Assumes node_distance(0,j) includes all distances in
6287 * node_distance(i,j) in order to avoid cubic time.
6289 next_distance = curr_distance;
6290 for (i = 0; i < nr_node_ids; i++) {
6291 for (j = 0; j < nr_node_ids; j++) {
6292 for (k = 0; k < nr_node_ids; k++) {
6293 int distance = node_distance(i, k);
6295 if (distance > curr_distance &&
6296 (distance < next_distance ||
6297 next_distance == curr_distance))
6298 next_distance = distance;
6301 * While not a strong assumption it would be nice to know
6302 * about cases where if node A is connected to B, B is not
6303 * equally connected to A.
6305 if (sched_debug() && node_distance(k, i) != distance)
6306 sched_numa_warn("Node-distance not symmetric");
6308 if (sched_debug() && i && !find_numa_distance(distance))
6309 sched_numa_warn("Node-0 not representative");
6311 if (next_distance != curr_distance) {
6312 sched_domains_numa_distance[level++] = next_distance;
6313 sched_domains_numa_levels = level;
6314 curr_distance = next_distance;
6319 * In case of sched_debug() we verify the above assumption.
6325 * 'level' contains the number of unique distances, excluding the
6326 * identity distance node_distance(i,i).
6328 * The sched_domains_numa_distance[] array includes the actual distance
6333 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6334 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6335 * the array will contain less then 'level' members. This could be
6336 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6337 * in other functions.
6339 * We reset it to 'level' at the end of this function.
6341 sched_domains_numa_levels = 0;
6343 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6344 if (!sched_domains_numa_masks)
6348 * Now for each level, construct a mask per node which contains all
6349 * cpus of nodes that are that many hops away from us.
6351 for (i = 0; i < level; i++) {
6352 sched_domains_numa_masks[i] =
6353 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6354 if (!sched_domains_numa_masks[i])
6357 for (j = 0; j < nr_node_ids; j++) {
6358 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6362 sched_domains_numa_masks[i][j] = mask;
6364 for (k = 0; k < nr_node_ids; k++) {
6365 if (node_distance(j, k) > sched_domains_numa_distance[i])
6368 cpumask_or(mask, mask, cpumask_of_node(k));
6373 /* Compute default topology size */
6374 for (i = 0; sched_domain_topology[i].mask; i++);
6376 tl = kzalloc((i + level + 1) *
6377 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6382 * Copy the default topology bits..
6384 for (i = 0; sched_domain_topology[i].mask; i++)
6385 tl[i] = sched_domain_topology[i];
6388 * .. and append 'j' levels of NUMA goodness.
6390 for (j = 0; j < level; i++, j++) {
6391 tl[i] = (struct sched_domain_topology_level){
6392 .mask = sd_numa_mask,
6393 .sd_flags = cpu_numa_flags,
6394 .flags = SDTL_OVERLAP,
6400 sched_domain_topology = tl;
6402 sched_domains_numa_levels = level;
6405 static void sched_domains_numa_masks_set(int cpu)
6408 int node = cpu_to_node(cpu);
6410 for (i = 0; i < sched_domains_numa_levels; i++) {
6411 for (j = 0; j < nr_node_ids; j++) {
6412 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6413 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6418 static void sched_domains_numa_masks_clear(int cpu)
6421 for (i = 0; i < sched_domains_numa_levels; i++) {
6422 for (j = 0; j < nr_node_ids; j++)
6423 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6428 * Update sched_domains_numa_masks[level][node] array when new cpus
6431 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6432 unsigned long action,
6435 int cpu = (long)hcpu;
6437 switch (action & ~CPU_TASKS_FROZEN) {
6439 sched_domains_numa_masks_set(cpu);
6443 sched_domains_numa_masks_clear(cpu);
6453 static inline void sched_init_numa(void)
6457 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6458 unsigned long action,
6463 #endif /* CONFIG_NUMA */
6465 static int __sdt_alloc(const struct cpumask *cpu_map)
6467 struct sched_domain_topology_level *tl;
6470 for_each_sd_topology(tl) {
6471 struct sd_data *sdd = &tl->data;
6473 sdd->sd = alloc_percpu(struct sched_domain *);
6477 sdd->sg = alloc_percpu(struct sched_group *);
6481 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6485 for_each_cpu(j, cpu_map) {
6486 struct sched_domain *sd;
6487 struct sched_group *sg;
6488 struct sched_group_capacity *sgc;
6490 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6491 GFP_KERNEL, cpu_to_node(j));
6495 *per_cpu_ptr(sdd->sd, j) = sd;
6497 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6498 GFP_KERNEL, cpu_to_node(j));
6504 *per_cpu_ptr(sdd->sg, j) = sg;
6506 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6507 GFP_KERNEL, cpu_to_node(j));
6511 *per_cpu_ptr(sdd->sgc, j) = sgc;
6518 static void __sdt_free(const struct cpumask *cpu_map)
6520 struct sched_domain_topology_level *tl;
6523 for_each_sd_topology(tl) {
6524 struct sd_data *sdd = &tl->data;
6526 for_each_cpu(j, cpu_map) {
6527 struct sched_domain *sd;
6530 sd = *per_cpu_ptr(sdd->sd, j);
6531 if (sd && (sd->flags & SD_OVERLAP))
6532 free_sched_groups(sd->groups, 0);
6533 kfree(*per_cpu_ptr(sdd->sd, j));
6537 kfree(*per_cpu_ptr(sdd->sg, j));
6539 kfree(*per_cpu_ptr(sdd->sgc, j));
6541 free_percpu(sdd->sd);
6543 free_percpu(sdd->sg);
6545 free_percpu(sdd->sgc);
6550 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6551 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6552 struct sched_domain *child, int cpu)
6554 struct sched_domain *sd = sd_init(tl, cpu);
6558 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6560 sd->level = child->level + 1;
6561 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6565 if (!cpumask_subset(sched_domain_span(child),
6566 sched_domain_span(sd))) {
6567 pr_err("BUG: arch topology borken\n");
6568 #ifdef CONFIG_SCHED_DEBUG
6569 pr_err(" the %s domain not a subset of the %s domain\n",
6570 child->name, sd->name);
6572 /* Fixup, ensure @sd has at least @child cpus. */
6573 cpumask_or(sched_domain_span(sd),
6574 sched_domain_span(sd),
6575 sched_domain_span(child));
6579 set_domain_attribute(sd, attr);
6585 * Build sched domains for a given set of cpus and attach the sched domains
6586 * to the individual cpus
6588 static int build_sched_domains(const struct cpumask *cpu_map,
6589 struct sched_domain_attr *attr)
6591 enum s_alloc alloc_state;
6592 struct sched_domain *sd;
6594 int i, ret = -ENOMEM;
6596 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6597 if (alloc_state != sa_rootdomain)
6600 /* Set up domains for cpus specified by the cpu_map. */
6601 for_each_cpu(i, cpu_map) {
6602 struct sched_domain_topology_level *tl;
6605 for_each_sd_topology(tl) {
6606 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6607 if (tl == sched_domain_topology)
6608 *per_cpu_ptr(d.sd, i) = sd;
6609 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6610 sd->flags |= SD_OVERLAP;
6611 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6616 /* Build the groups for the domains */
6617 for_each_cpu(i, cpu_map) {
6618 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6619 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6620 if (sd->flags & SD_OVERLAP) {
6621 if (build_overlap_sched_groups(sd, i))
6624 if (build_sched_groups(sd, i))
6630 /* Calculate CPU capacity for physical packages and nodes */
6631 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6632 if (!cpumask_test_cpu(i, cpu_map))
6635 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6636 claim_allocations(i, sd);
6637 init_sched_groups_capacity(i, sd);
6641 /* Attach the domains */
6643 for_each_cpu(i, cpu_map) {
6644 sd = *per_cpu_ptr(d.sd, i);
6645 cpu_attach_domain(sd, d.rd, i);
6651 __free_domain_allocs(&d, alloc_state, cpu_map);
6655 static cpumask_var_t *doms_cur; /* current sched domains */
6656 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6657 static struct sched_domain_attr *dattr_cur;
6658 /* attribues of custom domains in 'doms_cur' */
6661 * Special case: If a kmalloc of a doms_cur partition (array of
6662 * cpumask) fails, then fallback to a single sched domain,
6663 * as determined by the single cpumask fallback_doms.
6665 static cpumask_var_t fallback_doms;
6668 * arch_update_cpu_topology lets virtualized architectures update the
6669 * cpu core maps. It is supposed to return 1 if the topology changed
6670 * or 0 if it stayed the same.
6672 int __weak arch_update_cpu_topology(void)
6677 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6680 cpumask_var_t *doms;
6682 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6685 for (i = 0; i < ndoms; i++) {
6686 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6687 free_sched_domains(doms, i);
6694 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6697 for (i = 0; i < ndoms; i++)
6698 free_cpumask_var(doms[i]);
6703 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6704 * For now this just excludes isolated cpus, but could be used to
6705 * exclude other special cases in the future.
6707 static int init_sched_domains(const struct cpumask *cpu_map)
6711 arch_update_cpu_topology();
6713 doms_cur = alloc_sched_domains(ndoms_cur);
6715 doms_cur = &fallback_doms;
6716 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6717 err = build_sched_domains(doms_cur[0], NULL);
6718 register_sched_domain_sysctl();
6724 * Detach sched domains from a group of cpus specified in cpu_map
6725 * These cpus will now be attached to the NULL domain
6727 static void detach_destroy_domains(const struct cpumask *cpu_map)
6732 for_each_cpu(i, cpu_map)
6733 cpu_attach_domain(NULL, &def_root_domain, i);
6737 /* handle null as "default" */
6738 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6739 struct sched_domain_attr *new, int idx_new)
6741 struct sched_domain_attr tmp;
6748 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6749 new ? (new + idx_new) : &tmp,
6750 sizeof(struct sched_domain_attr));
6754 * Partition sched domains as specified by the 'ndoms_new'
6755 * cpumasks in the array doms_new[] of cpumasks. This compares
6756 * doms_new[] to the current sched domain partitioning, doms_cur[].
6757 * It destroys each deleted domain and builds each new domain.
6759 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6760 * The masks don't intersect (don't overlap.) We should setup one
6761 * sched domain for each mask. CPUs not in any of the cpumasks will
6762 * not be load balanced. If the same cpumask appears both in the
6763 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6766 * The passed in 'doms_new' should be allocated using
6767 * alloc_sched_domains. This routine takes ownership of it and will
6768 * free_sched_domains it when done with it. If the caller failed the
6769 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6770 * and partition_sched_domains() will fallback to the single partition
6771 * 'fallback_doms', it also forces the domains to be rebuilt.
6773 * If doms_new == NULL it will be replaced with cpu_online_mask.
6774 * ndoms_new == 0 is a special case for destroying existing domains,
6775 * and it will not create the default domain.
6777 * Call with hotplug lock held
6779 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6780 struct sched_domain_attr *dattr_new)
6785 mutex_lock(&sched_domains_mutex);
6787 /* always unregister in case we don't destroy any domains */
6788 unregister_sched_domain_sysctl();
6790 /* Let architecture update cpu core mappings. */
6791 new_topology = arch_update_cpu_topology();
6793 n = doms_new ? ndoms_new : 0;
6795 /* Destroy deleted domains */
6796 for (i = 0; i < ndoms_cur; i++) {
6797 for (j = 0; j < n && !new_topology; j++) {
6798 if (cpumask_equal(doms_cur[i], doms_new[j])
6799 && dattrs_equal(dattr_cur, i, dattr_new, j))
6802 /* no match - a current sched domain not in new doms_new[] */
6803 detach_destroy_domains(doms_cur[i]);
6809 if (doms_new == NULL) {
6811 doms_new = &fallback_doms;
6812 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6813 WARN_ON_ONCE(dattr_new);
6816 /* Build new domains */
6817 for (i = 0; i < ndoms_new; i++) {
6818 for (j = 0; j < n && !new_topology; j++) {
6819 if (cpumask_equal(doms_new[i], doms_cur[j])
6820 && dattrs_equal(dattr_new, i, dattr_cur, j))
6823 /* no match - add a new doms_new */
6824 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6829 /* Remember the new sched domains */
6830 if (doms_cur != &fallback_doms)
6831 free_sched_domains(doms_cur, ndoms_cur);
6832 kfree(dattr_cur); /* kfree(NULL) is safe */
6833 doms_cur = doms_new;
6834 dattr_cur = dattr_new;
6835 ndoms_cur = ndoms_new;
6837 register_sched_domain_sysctl();
6839 mutex_unlock(&sched_domains_mutex);
6842 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6845 * Update cpusets according to cpu_active mask. If cpusets are
6846 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6847 * around partition_sched_domains().
6849 * If we come here as part of a suspend/resume, don't touch cpusets because we
6850 * want to restore it back to its original state upon resume anyway.
6852 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6856 case CPU_ONLINE_FROZEN:
6857 case CPU_DOWN_FAILED_FROZEN:
6860 * num_cpus_frozen tracks how many CPUs are involved in suspend
6861 * resume sequence. As long as this is not the last online
6862 * operation in the resume sequence, just build a single sched
6863 * domain, ignoring cpusets.
6866 if (likely(num_cpus_frozen)) {
6867 partition_sched_domains(1, NULL, NULL);
6872 * This is the last CPU online operation. So fall through and
6873 * restore the original sched domains by considering the
6874 * cpuset configurations.
6878 case CPU_DOWN_FAILED:
6879 cpuset_update_active_cpus(true);
6887 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6891 case CPU_DOWN_PREPARE:
6892 cpuset_update_active_cpus(false);
6894 case CPU_DOWN_PREPARE_FROZEN:
6896 partition_sched_domains(1, NULL, NULL);
6904 void __init sched_init_smp(void)
6906 cpumask_var_t non_isolated_cpus;
6908 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6909 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6914 * There's no userspace yet to cause hotplug operations; hence all the
6915 * cpu masks are stable and all blatant races in the below code cannot
6918 mutex_lock(&sched_domains_mutex);
6919 init_sched_domains(cpu_active_mask);
6920 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6921 if (cpumask_empty(non_isolated_cpus))
6922 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6923 mutex_unlock(&sched_domains_mutex);
6925 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6926 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6927 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6931 /* Move init over to a non-isolated CPU */
6932 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6934 sched_init_granularity();
6935 free_cpumask_var(non_isolated_cpus);
6937 init_sched_rt_class();
6938 init_sched_dl_class();
6941 void __init sched_init_smp(void)
6943 sched_init_granularity();
6945 #endif /* CONFIG_SMP */
6947 const_debug unsigned int sysctl_timer_migration = 1;
6949 int in_sched_functions(unsigned long addr)
6951 return in_lock_functions(addr) ||
6952 (addr >= (unsigned long)__sched_text_start
6953 && addr < (unsigned long)__sched_text_end);
6956 #ifdef CONFIG_CGROUP_SCHED
6958 * Default task group.
6959 * Every task in system belongs to this group at bootup.
6961 struct task_group root_task_group;
6962 LIST_HEAD(task_groups);
6965 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6967 void __init sched_init(void)
6970 unsigned long alloc_size = 0, ptr;
6972 #ifdef CONFIG_FAIR_GROUP_SCHED
6973 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6975 #ifdef CONFIG_RT_GROUP_SCHED
6976 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6978 #ifdef CONFIG_CPUMASK_OFFSTACK
6979 alloc_size += num_possible_cpus() * cpumask_size();
6982 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6984 #ifdef CONFIG_FAIR_GROUP_SCHED
6985 root_task_group.se = (struct sched_entity **)ptr;
6986 ptr += nr_cpu_ids * sizeof(void **);
6988 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6989 ptr += nr_cpu_ids * sizeof(void **);
6991 #endif /* CONFIG_FAIR_GROUP_SCHED */
6992 #ifdef CONFIG_RT_GROUP_SCHED
6993 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6994 ptr += nr_cpu_ids * sizeof(void **);
6996 root_task_group.rt_rq = (struct rt_rq **)ptr;
6997 ptr += nr_cpu_ids * sizeof(void **);
6999 #endif /* CONFIG_RT_GROUP_SCHED */
7000 #ifdef CONFIG_CPUMASK_OFFSTACK
7001 for_each_possible_cpu(i) {
7002 per_cpu(load_balance_mask, i) = (void *)ptr;
7003 ptr += cpumask_size();
7005 #endif /* CONFIG_CPUMASK_OFFSTACK */
7008 init_rt_bandwidth(&def_rt_bandwidth,
7009 global_rt_period(), global_rt_runtime());
7010 init_dl_bandwidth(&def_dl_bandwidth,
7011 global_rt_period(), global_rt_runtime());
7014 init_defrootdomain();
7017 #ifdef CONFIG_RT_GROUP_SCHED
7018 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7019 global_rt_period(), global_rt_runtime());
7020 #endif /* CONFIG_RT_GROUP_SCHED */
7022 #ifdef CONFIG_CGROUP_SCHED
7023 list_add(&root_task_group.list, &task_groups);
7024 INIT_LIST_HEAD(&root_task_group.children);
7025 INIT_LIST_HEAD(&root_task_group.siblings);
7026 autogroup_init(&init_task);
7028 #endif /* CONFIG_CGROUP_SCHED */
7030 for_each_possible_cpu(i) {
7034 raw_spin_lock_init(&rq->lock);
7036 rq->calc_load_active = 0;
7037 rq->calc_load_update = jiffies + LOAD_FREQ;
7038 init_cfs_rq(&rq->cfs);
7039 init_rt_rq(&rq->rt, rq);
7040 init_dl_rq(&rq->dl, rq);
7041 #ifdef CONFIG_FAIR_GROUP_SCHED
7042 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7043 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7045 * How much cpu bandwidth does root_task_group get?
7047 * In case of task-groups formed thr' the cgroup filesystem, it
7048 * gets 100% of the cpu resources in the system. This overall
7049 * system cpu resource is divided among the tasks of
7050 * root_task_group and its child task-groups in a fair manner,
7051 * based on each entity's (task or task-group's) weight
7052 * (se->load.weight).
7054 * In other words, if root_task_group has 10 tasks of weight
7055 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7056 * then A0's share of the cpu resource is:
7058 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7060 * We achieve this by letting root_task_group's tasks sit
7061 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7063 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7064 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7065 #endif /* CONFIG_FAIR_GROUP_SCHED */
7067 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7068 #ifdef CONFIG_RT_GROUP_SCHED
7069 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7072 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7073 rq->cpu_load[j] = 0;
7075 rq->last_load_update_tick = jiffies;
7080 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7081 rq->post_schedule = 0;
7082 rq->active_balance = 0;
7083 rq->next_balance = jiffies;
7088 rq->avg_idle = 2*sysctl_sched_migration_cost;
7089 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7091 INIT_LIST_HEAD(&rq->cfs_tasks);
7093 rq_attach_root(rq, &def_root_domain);
7094 #ifdef CONFIG_NO_HZ_COMMON
7097 #ifdef CONFIG_NO_HZ_FULL
7098 rq->last_sched_tick = 0;
7102 atomic_set(&rq->nr_iowait, 0);
7105 set_load_weight(&init_task);
7107 #ifdef CONFIG_PREEMPT_NOTIFIERS
7108 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7112 * The boot idle thread does lazy MMU switching as well:
7114 atomic_inc(&init_mm.mm_count);
7115 enter_lazy_tlb(&init_mm, current);
7118 * Make us the idle thread. Technically, schedule() should not be
7119 * called from this thread, however somewhere below it might be,
7120 * but because we are the idle thread, we just pick up running again
7121 * when this runqueue becomes "idle".
7123 init_idle(current, smp_processor_id());
7125 calc_load_update = jiffies + LOAD_FREQ;
7128 * During early bootup we pretend to be a normal task:
7130 current->sched_class = &fair_sched_class;
7133 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7134 /* May be allocated at isolcpus cmdline parse time */
7135 if (cpu_isolated_map == NULL)
7136 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7137 idle_thread_set_boot_cpu();
7138 set_cpu_rq_start_time();
7140 init_sched_fair_class();
7142 scheduler_running = 1;
7145 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7146 static inline int preempt_count_equals(int preempt_offset)
7148 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7150 return (nested == preempt_offset);
7153 void __might_sleep(const char *file, int line, int preempt_offset)
7155 static unsigned long prev_jiffy; /* ratelimiting */
7157 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7158 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7159 !is_idle_task(current)) ||
7160 system_state != SYSTEM_RUNNING || oops_in_progress)
7162 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7164 prev_jiffy = jiffies;
7167 "BUG: sleeping function called from invalid context at %s:%d\n",
7170 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7171 in_atomic(), irqs_disabled(),
7172 current->pid, current->comm);
7174 debug_show_held_locks(current);
7175 if (irqs_disabled())
7176 print_irqtrace_events(current);
7177 #ifdef CONFIG_DEBUG_PREEMPT
7178 if (!preempt_count_equals(preempt_offset)) {
7179 pr_err("Preemption disabled at:");
7180 print_ip_sym(current->preempt_disable_ip);
7186 EXPORT_SYMBOL(__might_sleep);
7189 #ifdef CONFIG_MAGIC_SYSRQ
7190 static void normalize_task(struct rq *rq, struct task_struct *p)
7192 const struct sched_class *prev_class = p->sched_class;
7193 struct sched_attr attr = {
7194 .sched_policy = SCHED_NORMAL,
7196 int old_prio = p->prio;
7199 queued = task_on_rq_queued(p);
7201 dequeue_task(rq, p, 0);
7202 __setscheduler(rq, p, &attr);
7204 enqueue_task(rq, p, 0);
7208 check_class_changed(rq, p, prev_class, old_prio);
7211 void normalize_rt_tasks(void)
7213 struct task_struct *g, *p;
7214 unsigned long flags;
7217 read_lock(&tasklist_lock);
7218 for_each_process_thread(g, p) {
7220 * Only normalize user tasks:
7222 if (p->flags & PF_KTHREAD)
7225 p->se.exec_start = 0;
7226 #ifdef CONFIG_SCHEDSTATS
7227 p->se.statistics.wait_start = 0;
7228 p->se.statistics.sleep_start = 0;
7229 p->se.statistics.block_start = 0;
7232 if (!dl_task(p) && !rt_task(p)) {
7234 * Renice negative nice level userspace
7237 if (task_nice(p) < 0)
7238 set_user_nice(p, 0);
7242 rq = task_rq_lock(p, &flags);
7243 normalize_task(rq, p);
7244 task_rq_unlock(rq, p, &flags);
7246 read_unlock(&tasklist_lock);
7249 #endif /* CONFIG_MAGIC_SYSRQ */
7251 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7253 * These functions are only useful for the IA64 MCA handling, or kdb.
7255 * They can only be called when the whole system has been
7256 * stopped - every CPU needs to be quiescent, and no scheduling
7257 * activity can take place. Using them for anything else would
7258 * be a serious bug, and as a result, they aren't even visible
7259 * under any other configuration.
7263 * curr_task - return the current task for a given cpu.
7264 * @cpu: the processor in question.
7266 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7268 * Return: The current task for @cpu.
7270 struct task_struct *curr_task(int cpu)
7272 return cpu_curr(cpu);
7275 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7279 * set_curr_task - set the current task for a given cpu.
7280 * @cpu: the processor in question.
7281 * @p: the task pointer to set.
7283 * Description: This function must only be used when non-maskable interrupts
7284 * are serviced on a separate stack. It allows the architecture to switch the
7285 * notion of the current task on a cpu in a non-blocking manner. This function
7286 * must be called with all CPU's synchronized, and interrupts disabled, the
7287 * and caller must save the original value of the current task (see
7288 * curr_task() above) and restore that value before reenabling interrupts and
7289 * re-starting the system.
7291 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7293 void set_curr_task(int cpu, struct task_struct *p)
7300 #ifdef CONFIG_CGROUP_SCHED
7301 /* task_group_lock serializes the addition/removal of task groups */
7302 static DEFINE_SPINLOCK(task_group_lock);
7304 static void free_sched_group(struct task_group *tg)
7306 free_fair_sched_group(tg);
7307 free_rt_sched_group(tg);
7312 /* allocate runqueue etc for a new task group */
7313 struct task_group *sched_create_group(struct task_group *parent)
7315 struct task_group *tg;
7317 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7319 return ERR_PTR(-ENOMEM);
7321 if (!alloc_fair_sched_group(tg, parent))
7324 if (!alloc_rt_sched_group(tg, parent))
7330 free_sched_group(tg);
7331 return ERR_PTR(-ENOMEM);
7334 void sched_online_group(struct task_group *tg, struct task_group *parent)
7336 unsigned long flags;
7338 spin_lock_irqsave(&task_group_lock, flags);
7339 list_add_rcu(&tg->list, &task_groups);
7341 WARN_ON(!parent); /* root should already exist */
7343 tg->parent = parent;
7344 INIT_LIST_HEAD(&tg->children);
7345 list_add_rcu(&tg->siblings, &parent->children);
7346 spin_unlock_irqrestore(&task_group_lock, flags);
7349 /* rcu callback to free various structures associated with a task group */
7350 static void free_sched_group_rcu(struct rcu_head *rhp)
7352 /* now it should be safe to free those cfs_rqs */
7353 free_sched_group(container_of(rhp, struct task_group, rcu));
7356 /* Destroy runqueue etc associated with a task group */
7357 void sched_destroy_group(struct task_group *tg)
7359 /* wait for possible concurrent references to cfs_rqs complete */
7360 call_rcu(&tg->rcu, free_sched_group_rcu);
7363 void sched_offline_group(struct task_group *tg)
7365 unsigned long flags;
7368 /* end participation in shares distribution */
7369 for_each_possible_cpu(i)
7370 unregister_fair_sched_group(tg, i);
7372 spin_lock_irqsave(&task_group_lock, flags);
7373 list_del_rcu(&tg->list);
7374 list_del_rcu(&tg->siblings);
7375 spin_unlock_irqrestore(&task_group_lock, flags);
7378 /* change task's runqueue when it moves between groups.
7379 * The caller of this function should have put the task in its new group
7380 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7381 * reflect its new group.
7383 void sched_move_task(struct task_struct *tsk)
7385 struct task_group *tg;
7386 int queued, running;
7387 unsigned long flags;
7390 rq = task_rq_lock(tsk, &flags);
7392 running = task_current(rq, tsk);
7393 queued = task_on_rq_queued(tsk);
7396 dequeue_task(rq, tsk, 0);
7397 if (unlikely(running))
7398 put_prev_task(rq, tsk);
7400 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7401 lockdep_is_held(&tsk->sighand->siglock)),
7402 struct task_group, css);
7403 tg = autogroup_task_group(tsk, tg);
7404 tsk->sched_task_group = tg;
7406 #ifdef CONFIG_FAIR_GROUP_SCHED
7407 if (tsk->sched_class->task_move_group)
7408 tsk->sched_class->task_move_group(tsk, queued);
7411 set_task_rq(tsk, task_cpu(tsk));
7413 if (unlikely(running))
7414 tsk->sched_class->set_curr_task(rq);
7416 enqueue_task(rq, tsk, 0);
7418 task_rq_unlock(rq, tsk, &flags);
7420 #endif /* CONFIG_CGROUP_SCHED */
7422 #ifdef CONFIG_RT_GROUP_SCHED
7424 * Ensure that the real time constraints are schedulable.
7426 static DEFINE_MUTEX(rt_constraints_mutex);
7428 /* Must be called with tasklist_lock held */
7429 static inline int tg_has_rt_tasks(struct task_group *tg)
7431 struct task_struct *g, *p;
7433 for_each_process_thread(g, p) {
7434 if (rt_task(p) && task_group(p) == tg)
7441 struct rt_schedulable_data {
7442 struct task_group *tg;
7447 static int tg_rt_schedulable(struct task_group *tg, void *data)
7449 struct rt_schedulable_data *d = data;
7450 struct task_group *child;
7451 unsigned long total, sum = 0;
7452 u64 period, runtime;
7454 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7455 runtime = tg->rt_bandwidth.rt_runtime;
7458 period = d->rt_period;
7459 runtime = d->rt_runtime;
7463 * Cannot have more runtime than the period.
7465 if (runtime > period && runtime != RUNTIME_INF)
7469 * Ensure we don't starve existing RT tasks.
7471 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7474 total = to_ratio(period, runtime);
7477 * Nobody can have more than the global setting allows.
7479 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7483 * The sum of our children's runtime should not exceed our own.
7485 list_for_each_entry_rcu(child, &tg->children, siblings) {
7486 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7487 runtime = child->rt_bandwidth.rt_runtime;
7489 if (child == d->tg) {
7490 period = d->rt_period;
7491 runtime = d->rt_runtime;
7494 sum += to_ratio(period, runtime);
7503 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7507 struct rt_schedulable_data data = {
7509 .rt_period = period,
7510 .rt_runtime = runtime,
7514 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7520 static int tg_set_rt_bandwidth(struct task_group *tg,
7521 u64 rt_period, u64 rt_runtime)
7525 mutex_lock(&rt_constraints_mutex);
7526 read_lock(&tasklist_lock);
7527 err = __rt_schedulable(tg, rt_period, rt_runtime);
7531 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7532 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7533 tg->rt_bandwidth.rt_runtime = rt_runtime;
7535 for_each_possible_cpu(i) {
7536 struct rt_rq *rt_rq = tg->rt_rq[i];
7538 raw_spin_lock(&rt_rq->rt_runtime_lock);
7539 rt_rq->rt_runtime = rt_runtime;
7540 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7542 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7544 read_unlock(&tasklist_lock);
7545 mutex_unlock(&rt_constraints_mutex);
7550 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7552 u64 rt_runtime, rt_period;
7554 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7555 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7556 if (rt_runtime_us < 0)
7557 rt_runtime = RUNTIME_INF;
7559 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7562 static long sched_group_rt_runtime(struct task_group *tg)
7566 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7569 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7570 do_div(rt_runtime_us, NSEC_PER_USEC);
7571 return rt_runtime_us;
7574 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7576 u64 rt_runtime, rt_period;
7578 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7579 rt_runtime = tg->rt_bandwidth.rt_runtime;
7584 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7587 static long sched_group_rt_period(struct task_group *tg)
7591 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7592 do_div(rt_period_us, NSEC_PER_USEC);
7593 return rt_period_us;
7595 #endif /* CONFIG_RT_GROUP_SCHED */
7597 #ifdef CONFIG_RT_GROUP_SCHED
7598 static int sched_rt_global_constraints(void)
7602 mutex_lock(&rt_constraints_mutex);
7603 read_lock(&tasklist_lock);
7604 ret = __rt_schedulable(NULL, 0, 0);
7605 read_unlock(&tasklist_lock);
7606 mutex_unlock(&rt_constraints_mutex);
7611 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7613 /* Don't accept realtime tasks when there is no way for them to run */
7614 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7620 #else /* !CONFIG_RT_GROUP_SCHED */
7621 static int sched_rt_global_constraints(void)
7623 unsigned long flags;
7626 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7627 for_each_possible_cpu(i) {
7628 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7630 raw_spin_lock(&rt_rq->rt_runtime_lock);
7631 rt_rq->rt_runtime = global_rt_runtime();
7632 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7634 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7638 #endif /* CONFIG_RT_GROUP_SCHED */
7640 static int sched_dl_global_constraints(void)
7642 u64 runtime = global_rt_runtime();
7643 u64 period = global_rt_period();
7644 u64 new_bw = to_ratio(period, runtime);
7646 unsigned long flags;
7649 * Here we want to check the bandwidth not being set to some
7650 * value smaller than the currently allocated bandwidth in
7651 * any of the root_domains.
7653 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7654 * cycling on root_domains... Discussion on different/better
7655 * solutions is welcome!
7657 for_each_possible_cpu(cpu) {
7658 struct dl_bw *dl_b = dl_bw_of(cpu);
7660 raw_spin_lock_irqsave(&dl_b->lock, flags);
7661 if (new_bw < dl_b->total_bw)
7663 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7672 static void sched_dl_do_global(void)
7676 unsigned long flags;
7678 def_dl_bandwidth.dl_period = global_rt_period();
7679 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7681 if (global_rt_runtime() != RUNTIME_INF)
7682 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7685 * FIXME: As above...
7687 for_each_possible_cpu(cpu) {
7688 struct dl_bw *dl_b = dl_bw_of(cpu);
7690 raw_spin_lock_irqsave(&dl_b->lock, flags);
7692 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7696 static int sched_rt_global_validate(void)
7698 if (sysctl_sched_rt_period <= 0)
7701 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7702 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7708 static void sched_rt_do_global(void)
7710 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7711 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7714 int sched_rt_handler(struct ctl_table *table, int write,
7715 void __user *buffer, size_t *lenp,
7718 int old_period, old_runtime;
7719 static DEFINE_MUTEX(mutex);
7723 old_period = sysctl_sched_rt_period;
7724 old_runtime = sysctl_sched_rt_runtime;
7726 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7728 if (!ret && write) {
7729 ret = sched_rt_global_validate();
7733 ret = sched_rt_global_constraints();
7737 ret = sched_dl_global_constraints();
7741 sched_rt_do_global();
7742 sched_dl_do_global();
7746 sysctl_sched_rt_period = old_period;
7747 sysctl_sched_rt_runtime = old_runtime;
7749 mutex_unlock(&mutex);
7754 int sched_rr_handler(struct ctl_table *table, int write,
7755 void __user *buffer, size_t *lenp,
7759 static DEFINE_MUTEX(mutex);
7762 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7763 /* make sure that internally we keep jiffies */
7764 /* also, writing zero resets timeslice to default */
7765 if (!ret && write) {
7766 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7767 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7769 mutex_unlock(&mutex);
7773 #ifdef CONFIG_CGROUP_SCHED
7775 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7777 return css ? container_of(css, struct task_group, css) : NULL;
7780 static struct cgroup_subsys_state *
7781 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7783 struct task_group *parent = css_tg(parent_css);
7784 struct task_group *tg;
7787 /* This is early initialization for the top cgroup */
7788 return &root_task_group.css;
7791 tg = sched_create_group(parent);
7793 return ERR_PTR(-ENOMEM);
7798 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7800 struct task_group *tg = css_tg(css);
7801 struct task_group *parent = css_tg(css->parent);
7804 sched_online_group(tg, parent);
7808 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7810 struct task_group *tg = css_tg(css);
7812 sched_destroy_group(tg);
7815 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7817 struct task_group *tg = css_tg(css);
7819 sched_offline_group(tg);
7822 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7823 struct cgroup_taskset *tset)
7825 struct task_struct *task;
7827 cgroup_taskset_for_each(task, tset) {
7828 #ifdef CONFIG_RT_GROUP_SCHED
7829 if (!sched_rt_can_attach(css_tg(css), task))
7832 /* We don't support RT-tasks being in separate groups */
7833 if (task->sched_class != &fair_sched_class)
7840 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7841 struct cgroup_taskset *tset)
7843 struct task_struct *task;
7845 cgroup_taskset_for_each(task, tset)
7846 sched_move_task(task);
7849 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7850 struct cgroup_subsys_state *old_css,
7851 struct task_struct *task)
7854 * cgroup_exit() is called in the copy_process() failure path.
7855 * Ignore this case since the task hasn't ran yet, this avoids
7856 * trying to poke a half freed task state from generic code.
7858 if (!(task->flags & PF_EXITING))
7861 sched_move_task(task);
7864 #ifdef CONFIG_FAIR_GROUP_SCHED
7865 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7866 struct cftype *cftype, u64 shareval)
7868 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7871 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7874 struct task_group *tg = css_tg(css);
7876 return (u64) scale_load_down(tg->shares);
7879 #ifdef CONFIG_CFS_BANDWIDTH
7880 static DEFINE_MUTEX(cfs_constraints_mutex);
7882 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7883 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7885 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7887 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7889 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7890 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7892 if (tg == &root_task_group)
7896 * Ensure we have at some amount of bandwidth every period. This is
7897 * to prevent reaching a state of large arrears when throttled via
7898 * entity_tick() resulting in prolonged exit starvation.
7900 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7904 * Likewise, bound things on the otherside by preventing insane quota
7905 * periods. This also allows us to normalize in computing quota
7908 if (period > max_cfs_quota_period)
7912 * Prevent race between setting of cfs_rq->runtime_enabled and
7913 * unthrottle_offline_cfs_rqs().
7916 mutex_lock(&cfs_constraints_mutex);
7917 ret = __cfs_schedulable(tg, period, quota);
7921 runtime_enabled = quota != RUNTIME_INF;
7922 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7924 * If we need to toggle cfs_bandwidth_used, off->on must occur
7925 * before making related changes, and on->off must occur afterwards
7927 if (runtime_enabled && !runtime_was_enabled)
7928 cfs_bandwidth_usage_inc();
7929 raw_spin_lock_irq(&cfs_b->lock);
7930 cfs_b->period = ns_to_ktime(period);
7931 cfs_b->quota = quota;
7933 __refill_cfs_bandwidth_runtime(cfs_b);
7934 /* restart the period timer (if active) to handle new period expiry */
7935 if (runtime_enabled && cfs_b->timer_active) {
7936 /* force a reprogram */
7937 __start_cfs_bandwidth(cfs_b, true);
7939 raw_spin_unlock_irq(&cfs_b->lock);
7941 for_each_online_cpu(i) {
7942 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7943 struct rq *rq = cfs_rq->rq;
7945 raw_spin_lock_irq(&rq->lock);
7946 cfs_rq->runtime_enabled = runtime_enabled;
7947 cfs_rq->runtime_remaining = 0;
7949 if (cfs_rq->throttled)
7950 unthrottle_cfs_rq(cfs_rq);
7951 raw_spin_unlock_irq(&rq->lock);
7953 if (runtime_was_enabled && !runtime_enabled)
7954 cfs_bandwidth_usage_dec();
7956 mutex_unlock(&cfs_constraints_mutex);
7962 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7966 period = ktime_to_ns(tg->cfs_bandwidth.period);
7967 if (cfs_quota_us < 0)
7968 quota = RUNTIME_INF;
7970 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7972 return tg_set_cfs_bandwidth(tg, period, quota);
7975 long tg_get_cfs_quota(struct task_group *tg)
7979 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7982 quota_us = tg->cfs_bandwidth.quota;
7983 do_div(quota_us, NSEC_PER_USEC);
7988 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7992 period = (u64)cfs_period_us * NSEC_PER_USEC;
7993 quota = tg->cfs_bandwidth.quota;
7995 return tg_set_cfs_bandwidth(tg, period, quota);
7998 long tg_get_cfs_period(struct task_group *tg)
8002 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8003 do_div(cfs_period_us, NSEC_PER_USEC);
8005 return cfs_period_us;
8008 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8011 return tg_get_cfs_quota(css_tg(css));
8014 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8015 struct cftype *cftype, s64 cfs_quota_us)
8017 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8020 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8023 return tg_get_cfs_period(css_tg(css));
8026 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8027 struct cftype *cftype, u64 cfs_period_us)
8029 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8032 struct cfs_schedulable_data {
8033 struct task_group *tg;
8038 * normalize group quota/period to be quota/max_period
8039 * note: units are usecs
8041 static u64 normalize_cfs_quota(struct task_group *tg,
8042 struct cfs_schedulable_data *d)
8050 period = tg_get_cfs_period(tg);
8051 quota = tg_get_cfs_quota(tg);
8054 /* note: these should typically be equivalent */
8055 if (quota == RUNTIME_INF || quota == -1)
8058 return to_ratio(period, quota);
8061 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8063 struct cfs_schedulable_data *d = data;
8064 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8065 s64 quota = 0, parent_quota = -1;
8068 quota = RUNTIME_INF;
8070 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8072 quota = normalize_cfs_quota(tg, d);
8073 parent_quota = parent_b->hierarchical_quota;
8076 * ensure max(child_quota) <= parent_quota, inherit when no
8079 if (quota == RUNTIME_INF)
8080 quota = parent_quota;
8081 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8084 cfs_b->hierarchical_quota = quota;
8089 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8092 struct cfs_schedulable_data data = {
8098 if (quota != RUNTIME_INF) {
8099 do_div(data.period, NSEC_PER_USEC);
8100 do_div(data.quota, NSEC_PER_USEC);
8104 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8110 static int cpu_stats_show(struct seq_file *sf, void *v)
8112 struct task_group *tg = css_tg(seq_css(sf));
8113 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8115 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8116 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8117 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8121 #endif /* CONFIG_CFS_BANDWIDTH */
8122 #endif /* CONFIG_FAIR_GROUP_SCHED */
8124 #ifdef CONFIG_RT_GROUP_SCHED
8125 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8126 struct cftype *cft, s64 val)
8128 return sched_group_set_rt_runtime(css_tg(css), val);
8131 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8134 return sched_group_rt_runtime(css_tg(css));
8137 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8138 struct cftype *cftype, u64 rt_period_us)
8140 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8143 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8146 return sched_group_rt_period(css_tg(css));
8148 #endif /* CONFIG_RT_GROUP_SCHED */
8150 static struct cftype cpu_files[] = {
8151 #ifdef CONFIG_FAIR_GROUP_SCHED
8154 .read_u64 = cpu_shares_read_u64,
8155 .write_u64 = cpu_shares_write_u64,
8158 #ifdef CONFIG_CFS_BANDWIDTH
8160 .name = "cfs_quota_us",
8161 .read_s64 = cpu_cfs_quota_read_s64,
8162 .write_s64 = cpu_cfs_quota_write_s64,
8165 .name = "cfs_period_us",
8166 .read_u64 = cpu_cfs_period_read_u64,
8167 .write_u64 = cpu_cfs_period_write_u64,
8171 .seq_show = cpu_stats_show,
8174 #ifdef CONFIG_RT_GROUP_SCHED
8176 .name = "rt_runtime_us",
8177 .read_s64 = cpu_rt_runtime_read,
8178 .write_s64 = cpu_rt_runtime_write,
8181 .name = "rt_period_us",
8182 .read_u64 = cpu_rt_period_read_uint,
8183 .write_u64 = cpu_rt_period_write_uint,
8189 struct cgroup_subsys cpu_cgrp_subsys = {
8190 .css_alloc = cpu_cgroup_css_alloc,
8191 .css_free = cpu_cgroup_css_free,
8192 .css_online = cpu_cgroup_css_online,
8193 .css_offline = cpu_cgroup_css_offline,
8194 .can_attach = cpu_cgroup_can_attach,
8195 .attach = cpu_cgroup_attach,
8196 .exit = cpu_cgroup_exit,
8197 .legacy_cftypes = cpu_files,
8201 #endif /* CONFIG_CGROUP_SCHED */
8203 void dump_cpu_task(int cpu)
8205 pr_info("Task dump for CPU %d:\n", cpu);
8206 sched_show_task(cpu_curr(cpu));