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 bool cpus_share_cache(int this_cpu, int that_cpu)
1639 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1641 #endif /* CONFIG_SMP */
1643 static void ttwu_queue(struct task_struct *p, int cpu)
1645 struct rq *rq = cpu_rq(cpu);
1647 #if defined(CONFIG_SMP)
1648 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1649 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1650 ttwu_queue_remote(p, cpu);
1655 raw_spin_lock(&rq->lock);
1656 ttwu_do_activate(rq, p, 0);
1657 raw_spin_unlock(&rq->lock);
1661 * try_to_wake_up - wake up a thread
1662 * @p: the thread to be awakened
1663 * @state: the mask of task states that can be woken
1664 * @wake_flags: wake modifier flags (WF_*)
1666 * Put it on the run-queue if it's not already there. The "current"
1667 * thread is always on the run-queue (except when the actual
1668 * re-schedule is in progress), and as such you're allowed to do
1669 * the simpler "current->state = TASK_RUNNING" to mark yourself
1670 * runnable without the overhead of this.
1672 * Return: %true if @p was woken up, %false if it was already running.
1673 * or @state didn't match @p's state.
1676 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1678 unsigned long flags;
1679 int cpu, success = 0;
1682 * If we are going to wake up a thread waiting for CONDITION we
1683 * need to ensure that CONDITION=1 done by the caller can not be
1684 * reordered with p->state check below. This pairs with mb() in
1685 * set_current_state() the waiting thread does.
1687 smp_mb__before_spinlock();
1688 raw_spin_lock_irqsave(&p->pi_lock, flags);
1689 if (!(p->state & state))
1692 success = 1; /* we're going to change ->state */
1695 if (p->on_rq && ttwu_remote(p, wake_flags))
1700 * If the owning (remote) cpu is still in the middle of schedule() with
1701 * this task as prev, wait until its done referencing the task.
1706 * Pairs with the smp_wmb() in finish_lock_switch().
1710 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1711 p->state = TASK_WAKING;
1713 if (p->sched_class->task_waking)
1714 p->sched_class->task_waking(p);
1716 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1717 if (task_cpu(p) != cpu) {
1718 wake_flags |= WF_MIGRATED;
1719 set_task_cpu(p, cpu);
1721 #endif /* CONFIG_SMP */
1725 ttwu_stat(p, cpu, wake_flags);
1727 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1733 * try_to_wake_up_local - try to wake up a local task with rq lock held
1734 * @p: the thread to be awakened
1736 * Put @p on the run-queue if it's not already there. The caller must
1737 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1740 static void try_to_wake_up_local(struct task_struct *p)
1742 struct rq *rq = task_rq(p);
1744 if (WARN_ON_ONCE(rq != this_rq()) ||
1745 WARN_ON_ONCE(p == current))
1748 lockdep_assert_held(&rq->lock);
1750 if (!raw_spin_trylock(&p->pi_lock)) {
1751 raw_spin_unlock(&rq->lock);
1752 raw_spin_lock(&p->pi_lock);
1753 raw_spin_lock(&rq->lock);
1756 if (!(p->state & TASK_NORMAL))
1759 if (!task_on_rq_queued(p))
1760 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1762 ttwu_do_wakeup(rq, p, 0);
1763 ttwu_stat(p, smp_processor_id(), 0);
1765 raw_spin_unlock(&p->pi_lock);
1769 * wake_up_process - Wake up a specific process
1770 * @p: The process to be woken up.
1772 * Attempt to wake up the nominated process and move it to the set of runnable
1775 * Return: 1 if the process was woken up, 0 if it was already running.
1777 * It may be assumed that this function implies a write memory barrier before
1778 * changing the task state if and only if any tasks are woken up.
1780 int wake_up_process(struct task_struct *p)
1782 WARN_ON(task_is_stopped_or_traced(p));
1783 return try_to_wake_up(p, TASK_NORMAL, 0);
1785 EXPORT_SYMBOL(wake_up_process);
1787 int wake_up_state(struct task_struct *p, unsigned int state)
1789 return try_to_wake_up(p, state, 0);
1793 * Perform scheduler related setup for a newly forked process p.
1794 * p is forked by current.
1796 * __sched_fork() is basic setup used by init_idle() too:
1798 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1803 p->se.exec_start = 0;
1804 p->se.sum_exec_runtime = 0;
1805 p->se.prev_sum_exec_runtime = 0;
1806 p->se.nr_migrations = 0;
1808 INIT_LIST_HEAD(&p->se.group_node);
1810 #ifdef CONFIG_SCHEDSTATS
1811 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1814 RB_CLEAR_NODE(&p->dl.rb_node);
1815 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1816 p->dl.dl_runtime = p->dl.runtime = 0;
1817 p->dl.dl_deadline = p->dl.deadline = 0;
1818 p->dl.dl_period = 0;
1821 INIT_LIST_HEAD(&p->rt.run_list);
1823 #ifdef CONFIG_PREEMPT_NOTIFIERS
1824 INIT_HLIST_HEAD(&p->preempt_notifiers);
1827 #ifdef CONFIG_NUMA_BALANCING
1828 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1829 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1830 p->mm->numa_scan_seq = 0;
1833 if (clone_flags & CLONE_VM)
1834 p->numa_preferred_nid = current->numa_preferred_nid;
1836 p->numa_preferred_nid = -1;
1838 p->node_stamp = 0ULL;
1839 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1840 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1841 p->numa_work.next = &p->numa_work;
1842 p->numa_faults_memory = NULL;
1843 p->numa_faults_buffer_memory = NULL;
1844 p->last_task_numa_placement = 0;
1845 p->last_sum_exec_runtime = 0;
1847 INIT_LIST_HEAD(&p->numa_entry);
1848 p->numa_group = NULL;
1849 #endif /* CONFIG_NUMA_BALANCING */
1852 #ifdef CONFIG_NUMA_BALANCING
1853 #ifdef CONFIG_SCHED_DEBUG
1854 void set_numabalancing_state(bool enabled)
1857 sched_feat_set("NUMA");
1859 sched_feat_set("NO_NUMA");
1862 __read_mostly bool numabalancing_enabled;
1864 void set_numabalancing_state(bool enabled)
1866 numabalancing_enabled = enabled;
1868 #endif /* CONFIG_SCHED_DEBUG */
1870 #ifdef CONFIG_PROC_SYSCTL
1871 int sysctl_numa_balancing(struct ctl_table *table, int write,
1872 void __user *buffer, size_t *lenp, loff_t *ppos)
1876 int state = numabalancing_enabled;
1878 if (write && !capable(CAP_SYS_ADMIN))
1883 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1887 set_numabalancing_state(state);
1894 * fork()/clone()-time setup:
1896 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1898 unsigned long flags;
1899 int cpu = get_cpu();
1901 __sched_fork(clone_flags, p);
1903 * We mark the process as running here. This guarantees that
1904 * nobody will actually run it, and a signal or other external
1905 * event cannot wake it up and insert it on the runqueue either.
1907 p->state = TASK_RUNNING;
1910 * Make sure we do not leak PI boosting priority to the child.
1912 p->prio = current->normal_prio;
1915 * Revert to default priority/policy on fork if requested.
1917 if (unlikely(p->sched_reset_on_fork)) {
1918 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1919 p->policy = SCHED_NORMAL;
1920 p->static_prio = NICE_TO_PRIO(0);
1922 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1923 p->static_prio = NICE_TO_PRIO(0);
1925 p->prio = p->normal_prio = __normal_prio(p);
1929 * We don't need the reset flag anymore after the fork. It has
1930 * fulfilled its duty:
1932 p->sched_reset_on_fork = 0;
1935 if (dl_prio(p->prio)) {
1938 } else if (rt_prio(p->prio)) {
1939 p->sched_class = &rt_sched_class;
1941 p->sched_class = &fair_sched_class;
1944 if (p->sched_class->task_fork)
1945 p->sched_class->task_fork(p);
1948 * The child is not yet in the pid-hash so no cgroup attach races,
1949 * and the cgroup is pinned to this child due to cgroup_fork()
1950 * is ran before sched_fork().
1952 * Silence PROVE_RCU.
1954 raw_spin_lock_irqsave(&p->pi_lock, flags);
1955 set_task_cpu(p, cpu);
1956 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1958 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1959 if (likely(sched_info_on()))
1960 memset(&p->sched_info, 0, sizeof(p->sched_info));
1962 #if defined(CONFIG_SMP)
1965 init_task_preempt_count(p);
1967 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1968 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1975 unsigned long to_ratio(u64 period, u64 runtime)
1977 if (runtime == RUNTIME_INF)
1981 * Doing this here saves a lot of checks in all
1982 * the calling paths, and returning zero seems
1983 * safe for them anyway.
1988 return div64_u64(runtime << 20, period);
1992 inline struct dl_bw *dl_bw_of(int i)
1994 return &cpu_rq(i)->rd->dl_bw;
1997 static inline int dl_bw_cpus(int i)
1999 struct root_domain *rd = cpu_rq(i)->rd;
2002 for_each_cpu_and(i, rd->span, cpu_active_mask)
2008 inline struct dl_bw *dl_bw_of(int i)
2010 return &cpu_rq(i)->dl.dl_bw;
2013 static inline int dl_bw_cpus(int i)
2020 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2022 dl_b->total_bw -= tsk_bw;
2026 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2028 dl_b->total_bw += tsk_bw;
2032 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2034 return dl_b->bw != -1 &&
2035 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2039 * We must be sure that accepting a new task (or allowing changing the
2040 * parameters of an existing one) is consistent with the bandwidth
2041 * constraints. If yes, this function also accordingly updates the currently
2042 * allocated bandwidth to reflect the new situation.
2044 * This function is called while holding p's rq->lock.
2046 static int dl_overflow(struct task_struct *p, int policy,
2047 const struct sched_attr *attr)
2050 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2051 u64 period = attr->sched_period ?: attr->sched_deadline;
2052 u64 runtime = attr->sched_runtime;
2053 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2056 if (new_bw == p->dl.dl_bw)
2060 * Either if a task, enters, leave, or stays -deadline but changes
2061 * its parameters, we may need to update accordingly the total
2062 * allocated bandwidth of the container.
2064 raw_spin_lock(&dl_b->lock);
2065 cpus = dl_bw_cpus(task_cpu(p));
2066 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2067 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2068 __dl_add(dl_b, new_bw);
2070 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2071 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2072 __dl_clear(dl_b, p->dl.dl_bw);
2073 __dl_add(dl_b, new_bw);
2075 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2076 __dl_clear(dl_b, p->dl.dl_bw);
2079 raw_spin_unlock(&dl_b->lock);
2084 extern void init_dl_bw(struct dl_bw *dl_b);
2087 * wake_up_new_task - wake up a newly created task for the first time.
2089 * This function will do some initial scheduler statistics housekeeping
2090 * that must be done for every newly created context, then puts the task
2091 * on the runqueue and wakes it.
2093 void wake_up_new_task(struct task_struct *p)
2095 unsigned long flags;
2098 raw_spin_lock_irqsave(&p->pi_lock, flags);
2101 * Fork balancing, do it here and not earlier because:
2102 * - cpus_allowed can change in the fork path
2103 * - any previously selected cpu might disappear through hotplug
2105 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2108 /* Initialize new task's runnable average */
2109 init_task_runnable_average(p);
2110 rq = __task_rq_lock(p);
2111 activate_task(rq, p, 0);
2112 p->on_rq = TASK_ON_RQ_QUEUED;
2113 trace_sched_wakeup_new(p, true);
2114 check_preempt_curr(rq, p, WF_FORK);
2116 if (p->sched_class->task_woken)
2117 p->sched_class->task_woken(rq, p);
2119 task_rq_unlock(rq, p, &flags);
2122 #ifdef CONFIG_PREEMPT_NOTIFIERS
2125 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2126 * @notifier: notifier struct to register
2128 void preempt_notifier_register(struct preempt_notifier *notifier)
2130 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2132 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2135 * preempt_notifier_unregister - no longer interested in preemption notifications
2136 * @notifier: notifier struct to unregister
2138 * This is safe to call from within a preemption notifier.
2140 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2142 hlist_del(¬ifier->link);
2144 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2146 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2148 struct preempt_notifier *notifier;
2150 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2151 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2155 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2156 struct task_struct *next)
2158 struct preempt_notifier *notifier;
2160 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2161 notifier->ops->sched_out(notifier, next);
2164 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2166 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2171 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2172 struct task_struct *next)
2176 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2179 * prepare_task_switch - prepare to switch tasks
2180 * @rq: the runqueue preparing to switch
2181 * @prev: the current task that is being switched out
2182 * @next: the task we are going to switch to.
2184 * This is called with the rq lock held and interrupts off. It must
2185 * be paired with a subsequent finish_task_switch after the context
2188 * prepare_task_switch sets up locking and calls architecture specific
2192 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2193 struct task_struct *next)
2195 trace_sched_switch(prev, next);
2196 sched_info_switch(rq, prev, next);
2197 perf_event_task_sched_out(prev, next);
2198 fire_sched_out_preempt_notifiers(prev, next);
2199 prepare_lock_switch(rq, next);
2200 prepare_arch_switch(next);
2204 * finish_task_switch - clean up after a task-switch
2205 * @rq: runqueue associated with task-switch
2206 * @prev: the thread we just switched away from.
2208 * finish_task_switch must be called after the context switch, paired
2209 * with a prepare_task_switch call before the context switch.
2210 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2211 * and do any other architecture-specific cleanup actions.
2213 * Note that we may have delayed dropping an mm in context_switch(). If
2214 * so, we finish that here outside of the runqueue lock. (Doing it
2215 * with the lock held can cause deadlocks; see schedule() for
2218 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2219 __releases(rq->lock)
2221 struct mm_struct *mm = rq->prev_mm;
2227 * A task struct has one reference for the use as "current".
2228 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2229 * schedule one last time. The schedule call will never return, and
2230 * the scheduled task must drop that reference.
2231 * The test for TASK_DEAD must occur while the runqueue locks are
2232 * still held, otherwise prev could be scheduled on another cpu, die
2233 * there before we look at prev->state, and then the reference would
2235 * Manfred Spraul <manfred@colorfullife.com>
2237 prev_state = prev->state;
2238 vtime_task_switch(prev);
2239 finish_arch_switch(prev);
2240 perf_event_task_sched_in(prev, current);
2241 finish_lock_switch(rq, prev);
2242 finish_arch_post_lock_switch();
2244 fire_sched_in_preempt_notifiers(current);
2247 if (unlikely(prev_state == TASK_DEAD)) {
2248 if (prev->sched_class->task_dead)
2249 prev->sched_class->task_dead(prev);
2252 * Remove function-return probe instances associated with this
2253 * task and put them back on the free list.
2255 kprobe_flush_task(prev);
2256 put_task_struct(prev);
2259 tick_nohz_task_switch(current);
2264 /* rq->lock is NOT held, but preemption is disabled */
2265 static inline void post_schedule(struct rq *rq)
2267 if (rq->post_schedule) {
2268 unsigned long flags;
2270 raw_spin_lock_irqsave(&rq->lock, flags);
2271 if (rq->curr->sched_class->post_schedule)
2272 rq->curr->sched_class->post_schedule(rq);
2273 raw_spin_unlock_irqrestore(&rq->lock, flags);
2275 rq->post_schedule = 0;
2281 static inline void post_schedule(struct rq *rq)
2288 * schedule_tail - first thing a freshly forked thread must call.
2289 * @prev: the thread we just switched away from.
2291 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2292 __releases(rq->lock)
2294 struct rq *rq = this_rq();
2296 finish_task_switch(rq, prev);
2299 * FIXME: do we need to worry about rq being invalidated by the
2304 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2305 /* In this case, finish_task_switch does not reenable preemption */
2308 if (current->set_child_tid)
2309 put_user(task_pid_vnr(current), current->set_child_tid);
2313 * context_switch - switch to the new MM and the new
2314 * thread's register state.
2317 context_switch(struct rq *rq, struct task_struct *prev,
2318 struct task_struct *next)
2320 struct mm_struct *mm, *oldmm;
2322 prepare_task_switch(rq, prev, next);
2325 oldmm = prev->active_mm;
2327 * For paravirt, this is coupled with an exit in switch_to to
2328 * combine the page table reload and the switch backend into
2331 arch_start_context_switch(prev);
2334 next->active_mm = oldmm;
2335 atomic_inc(&oldmm->mm_count);
2336 enter_lazy_tlb(oldmm, next);
2338 switch_mm(oldmm, mm, next);
2341 prev->active_mm = NULL;
2342 rq->prev_mm = oldmm;
2345 * Since the runqueue lock will be released by the next
2346 * task (which is an invalid locking op but in the case
2347 * of the scheduler it's an obvious special-case), so we
2348 * do an early lockdep release here:
2350 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2351 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2354 context_tracking_task_switch(prev, next);
2355 /* Here we just switch the register state and the stack. */
2356 switch_to(prev, next, prev);
2360 * this_rq must be evaluated again because prev may have moved
2361 * CPUs since it called schedule(), thus the 'rq' on its stack
2362 * frame will be invalid.
2364 finish_task_switch(this_rq(), prev);
2368 * nr_running and nr_context_switches:
2370 * externally visible scheduler statistics: current number of runnable
2371 * threads, total number of context switches performed since bootup.
2373 unsigned long nr_running(void)
2375 unsigned long i, sum = 0;
2377 for_each_online_cpu(i)
2378 sum += cpu_rq(i)->nr_running;
2383 unsigned long long nr_context_switches(void)
2386 unsigned long long sum = 0;
2388 for_each_possible_cpu(i)
2389 sum += cpu_rq(i)->nr_switches;
2394 unsigned long nr_iowait(void)
2396 unsigned long i, sum = 0;
2398 for_each_possible_cpu(i)
2399 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2404 unsigned long nr_iowait_cpu(int cpu)
2406 struct rq *this = cpu_rq(cpu);
2407 return atomic_read(&this->nr_iowait);
2410 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2412 struct rq *this = this_rq();
2413 *nr_waiters = atomic_read(&this->nr_iowait);
2414 *load = this->cpu_load[0];
2420 * sched_exec - execve() is a valuable balancing opportunity, because at
2421 * this point the task has the smallest effective memory and cache footprint.
2423 void sched_exec(void)
2425 struct task_struct *p = current;
2426 unsigned long flags;
2429 raw_spin_lock_irqsave(&p->pi_lock, flags);
2430 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2431 if (dest_cpu == smp_processor_id())
2434 if (likely(cpu_active(dest_cpu))) {
2435 struct migration_arg arg = { p, dest_cpu };
2437 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2438 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2442 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2447 DEFINE_PER_CPU(struct kernel_stat, kstat);
2448 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2450 EXPORT_PER_CPU_SYMBOL(kstat);
2451 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2454 * Return any ns on the sched_clock that have not yet been accounted in
2455 * @p in case that task is currently running.
2457 * Called with task_rq_lock() held on @rq.
2459 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2464 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2465 * project cycles that may never be accounted to this
2466 * thread, breaking clock_gettime().
2468 if (task_current(rq, p) && task_on_rq_queued(p)) {
2469 update_rq_clock(rq);
2470 ns = rq_clock_task(rq) - p->se.exec_start;
2478 unsigned long long task_delta_exec(struct task_struct *p)
2480 unsigned long flags;
2484 rq = task_rq_lock(p, &flags);
2485 ns = do_task_delta_exec(p, rq);
2486 task_rq_unlock(rq, p, &flags);
2492 * Return accounted runtime for the task.
2493 * In case the task is currently running, return the runtime plus current's
2494 * pending runtime that have not been accounted yet.
2496 unsigned long long task_sched_runtime(struct task_struct *p)
2498 unsigned long flags;
2502 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2504 * 64-bit doesn't need locks to atomically read a 64bit value.
2505 * So we have a optimization chance when the task's delta_exec is 0.
2506 * Reading ->on_cpu is racy, but this is ok.
2508 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2509 * If we race with it entering cpu, unaccounted time is 0. This is
2510 * indistinguishable from the read occurring a few cycles earlier.
2511 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2512 * been accounted, so we're correct here as well.
2514 if (!p->on_cpu || !task_on_rq_queued(p))
2515 return p->se.sum_exec_runtime;
2518 rq = task_rq_lock(p, &flags);
2519 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2520 task_rq_unlock(rq, p, &flags);
2526 * This function gets called by the timer code, with HZ frequency.
2527 * We call it with interrupts disabled.
2529 void scheduler_tick(void)
2531 int cpu = smp_processor_id();
2532 struct rq *rq = cpu_rq(cpu);
2533 struct task_struct *curr = rq->curr;
2537 raw_spin_lock(&rq->lock);
2538 update_rq_clock(rq);
2539 curr->sched_class->task_tick(rq, curr, 0);
2540 update_cpu_load_active(rq);
2541 raw_spin_unlock(&rq->lock);
2543 perf_event_task_tick();
2546 rq->idle_balance = idle_cpu(cpu);
2547 trigger_load_balance(rq);
2549 rq_last_tick_reset(rq);
2552 #ifdef CONFIG_NO_HZ_FULL
2554 * scheduler_tick_max_deferment
2556 * Keep at least one tick per second when a single
2557 * active task is running because the scheduler doesn't
2558 * yet completely support full dynticks environment.
2560 * This makes sure that uptime, CFS vruntime, load
2561 * balancing, etc... continue to move forward, even
2562 * with a very low granularity.
2564 * Return: Maximum deferment in nanoseconds.
2566 u64 scheduler_tick_max_deferment(void)
2568 struct rq *rq = this_rq();
2569 unsigned long next, now = ACCESS_ONCE(jiffies);
2571 next = rq->last_sched_tick + HZ;
2573 if (time_before_eq(next, now))
2576 return jiffies_to_nsecs(next - now);
2580 notrace unsigned long get_parent_ip(unsigned long addr)
2582 if (in_lock_functions(addr)) {
2583 addr = CALLER_ADDR2;
2584 if (in_lock_functions(addr))
2585 addr = CALLER_ADDR3;
2590 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2591 defined(CONFIG_PREEMPT_TRACER))
2593 void preempt_count_add(int val)
2595 #ifdef CONFIG_DEBUG_PREEMPT
2599 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2602 __preempt_count_add(val);
2603 #ifdef CONFIG_DEBUG_PREEMPT
2605 * Spinlock count overflowing soon?
2607 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2610 if (preempt_count() == val) {
2611 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2612 #ifdef CONFIG_DEBUG_PREEMPT
2613 current->preempt_disable_ip = ip;
2615 trace_preempt_off(CALLER_ADDR0, ip);
2618 EXPORT_SYMBOL(preempt_count_add);
2619 NOKPROBE_SYMBOL(preempt_count_add);
2621 void preempt_count_sub(int val)
2623 #ifdef CONFIG_DEBUG_PREEMPT
2627 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2630 * Is the spinlock portion underflowing?
2632 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2633 !(preempt_count() & PREEMPT_MASK)))
2637 if (preempt_count() == val)
2638 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2639 __preempt_count_sub(val);
2641 EXPORT_SYMBOL(preempt_count_sub);
2642 NOKPROBE_SYMBOL(preempt_count_sub);
2647 * Print scheduling while atomic bug:
2649 static noinline void __schedule_bug(struct task_struct *prev)
2651 if (oops_in_progress)
2654 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2655 prev->comm, prev->pid, preempt_count());
2657 debug_show_held_locks(prev);
2659 if (irqs_disabled())
2660 print_irqtrace_events(prev);
2661 #ifdef CONFIG_DEBUG_PREEMPT
2662 if (in_atomic_preempt_off()) {
2663 pr_err("Preemption disabled at:");
2664 print_ip_sym(current->preempt_disable_ip);
2669 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2673 * Various schedule()-time debugging checks and statistics:
2675 static inline void schedule_debug(struct task_struct *prev)
2678 * Test if we are atomic. Since do_exit() needs to call into
2679 * schedule() atomically, we ignore that path. Otherwise whine
2680 * if we are scheduling when we should not.
2682 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2683 __schedule_bug(prev);
2686 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2688 schedstat_inc(this_rq(), sched_count);
2692 * Pick up the highest-prio task:
2694 static inline struct task_struct *
2695 pick_next_task(struct rq *rq, struct task_struct *prev)
2697 const struct sched_class *class = &fair_sched_class;
2698 struct task_struct *p;
2701 * Optimization: we know that if all tasks are in
2702 * the fair class we can call that function directly:
2704 if (likely(prev->sched_class == class &&
2705 rq->nr_running == rq->cfs.h_nr_running)) {
2706 p = fair_sched_class.pick_next_task(rq, prev);
2707 if (unlikely(p == RETRY_TASK))
2710 /* assumes fair_sched_class->next == idle_sched_class */
2712 p = idle_sched_class.pick_next_task(rq, prev);
2718 for_each_class(class) {
2719 p = class->pick_next_task(rq, prev);
2721 if (unlikely(p == RETRY_TASK))
2727 BUG(); /* the idle class will always have a runnable task */
2731 * __schedule() is the main scheduler function.
2733 * The main means of driving the scheduler and thus entering this function are:
2735 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2737 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2738 * paths. For example, see arch/x86/entry_64.S.
2740 * To drive preemption between tasks, the scheduler sets the flag in timer
2741 * interrupt handler scheduler_tick().
2743 * 3. Wakeups don't really cause entry into schedule(). They add a
2744 * task to the run-queue and that's it.
2746 * Now, if the new task added to the run-queue preempts the current
2747 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2748 * called on the nearest possible occasion:
2750 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2752 * - in syscall or exception context, at the next outmost
2753 * preempt_enable(). (this might be as soon as the wake_up()'s
2756 * - in IRQ context, return from interrupt-handler to
2757 * preemptible context
2759 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2762 * - cond_resched() call
2763 * - explicit schedule() call
2764 * - return from syscall or exception to user-space
2765 * - return from interrupt-handler to user-space
2767 static void __sched __schedule(void)
2769 struct task_struct *prev, *next;
2770 unsigned long *switch_count;
2776 cpu = smp_processor_id();
2778 rcu_note_context_switch(cpu);
2781 schedule_debug(prev);
2783 if (sched_feat(HRTICK))
2787 * Make sure that signal_pending_state()->signal_pending() below
2788 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2789 * done by the caller to avoid the race with signal_wake_up().
2791 smp_mb__before_spinlock();
2792 raw_spin_lock_irq(&rq->lock);
2794 switch_count = &prev->nivcsw;
2795 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2796 if (unlikely(signal_pending_state(prev->state, prev))) {
2797 prev->state = TASK_RUNNING;
2799 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2803 * If a worker went to sleep, notify and ask workqueue
2804 * whether it wants to wake up a task to maintain
2807 if (prev->flags & PF_WQ_WORKER) {
2808 struct task_struct *to_wakeup;
2810 to_wakeup = wq_worker_sleeping(prev, cpu);
2812 try_to_wake_up_local(to_wakeup);
2815 switch_count = &prev->nvcsw;
2818 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0)
2819 update_rq_clock(rq);
2821 next = pick_next_task(rq, prev);
2822 clear_tsk_need_resched(prev);
2823 clear_preempt_need_resched();
2824 rq->skip_clock_update = 0;
2826 if (likely(prev != next)) {
2831 context_switch(rq, prev, next); /* unlocks the rq */
2833 * The context switch have flipped the stack from under us
2834 * and restored the local variables which were saved when
2835 * this task called schedule() in the past. prev == current
2836 * is still correct, but it can be moved to another cpu/rq.
2838 cpu = smp_processor_id();
2841 raw_spin_unlock_irq(&rq->lock);
2845 sched_preempt_enable_no_resched();
2850 static inline void sched_submit_work(struct task_struct *tsk)
2852 if (!tsk->state || tsk_is_pi_blocked(tsk))
2855 * If we are going to sleep and we have plugged IO queued,
2856 * make sure to submit it to avoid deadlocks.
2858 if (blk_needs_flush_plug(tsk))
2859 blk_schedule_flush_plug(tsk);
2862 asmlinkage __visible void __sched schedule(void)
2864 struct task_struct *tsk = current;
2866 sched_submit_work(tsk);
2869 EXPORT_SYMBOL(schedule);
2871 #ifdef CONFIG_CONTEXT_TRACKING
2872 asmlinkage __visible void __sched schedule_user(void)
2875 * If we come here after a random call to set_need_resched(),
2876 * or we have been woken up remotely but the IPI has not yet arrived,
2877 * we haven't yet exited the RCU idle mode. Do it here manually until
2878 * we find a better solution.
2887 * schedule_preempt_disabled - called with preemption disabled
2889 * Returns with preemption disabled. Note: preempt_count must be 1
2891 void __sched schedule_preempt_disabled(void)
2893 sched_preempt_enable_no_resched();
2898 #ifdef CONFIG_PREEMPT
2900 * this is the entry point to schedule() from in-kernel preemption
2901 * off of preempt_enable. Kernel preemptions off return from interrupt
2902 * occur there and call schedule directly.
2904 asmlinkage __visible void __sched notrace preempt_schedule(void)
2907 * If there is a non-zero preempt_count or interrupts are disabled,
2908 * we do not want to preempt the current task. Just return..
2910 if (likely(!preemptible()))
2914 __preempt_count_add(PREEMPT_ACTIVE);
2916 __preempt_count_sub(PREEMPT_ACTIVE);
2919 * Check again in case we missed a preemption opportunity
2920 * between schedule and now.
2923 } while (need_resched());
2925 NOKPROBE_SYMBOL(preempt_schedule);
2926 EXPORT_SYMBOL(preempt_schedule);
2927 #endif /* CONFIG_PREEMPT */
2930 * this is the entry point to schedule() from kernel preemption
2931 * off of irq context.
2932 * Note, that this is called and return with irqs disabled. This will
2933 * protect us against recursive calling from irq.
2935 asmlinkage __visible void __sched preempt_schedule_irq(void)
2937 enum ctx_state prev_state;
2939 /* Catch callers which need to be fixed */
2940 BUG_ON(preempt_count() || !irqs_disabled());
2942 prev_state = exception_enter();
2945 __preempt_count_add(PREEMPT_ACTIVE);
2948 local_irq_disable();
2949 __preempt_count_sub(PREEMPT_ACTIVE);
2952 * Check again in case we missed a preemption opportunity
2953 * between schedule and now.
2956 } while (need_resched());
2958 exception_exit(prev_state);
2961 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2964 return try_to_wake_up(curr->private, mode, wake_flags);
2966 EXPORT_SYMBOL(default_wake_function);
2968 #ifdef CONFIG_RT_MUTEXES
2971 * rt_mutex_setprio - set the current priority of a task
2973 * @prio: prio value (kernel-internal form)
2975 * This function changes the 'effective' priority of a task. It does
2976 * not touch ->normal_prio like __setscheduler().
2978 * Used by the rt_mutex code to implement priority inheritance
2979 * logic. Call site only calls if the priority of the task changed.
2981 void rt_mutex_setprio(struct task_struct *p, int prio)
2983 int oldprio, queued, running, enqueue_flag = 0;
2985 const struct sched_class *prev_class;
2987 BUG_ON(prio > MAX_PRIO);
2989 rq = __task_rq_lock(p);
2992 * Idle task boosting is a nono in general. There is one
2993 * exception, when PREEMPT_RT and NOHZ is active:
2995 * The idle task calls get_next_timer_interrupt() and holds
2996 * the timer wheel base->lock on the CPU and another CPU wants
2997 * to access the timer (probably to cancel it). We can safely
2998 * ignore the boosting request, as the idle CPU runs this code
2999 * with interrupts disabled and will complete the lock
3000 * protected section without being interrupted. So there is no
3001 * real need to boost.
3003 if (unlikely(p == rq->idle)) {
3004 WARN_ON(p != rq->curr);
3005 WARN_ON(p->pi_blocked_on);
3009 trace_sched_pi_setprio(p, prio);
3011 prev_class = p->sched_class;
3012 queued = task_on_rq_queued(p);
3013 running = task_current(rq, p);
3015 dequeue_task(rq, p, 0);
3017 p->sched_class->put_prev_task(rq, p);
3020 * Boosting condition are:
3021 * 1. -rt task is running and holds mutex A
3022 * --> -dl task blocks on mutex A
3024 * 2. -dl task is running and holds mutex A
3025 * --> -dl task blocks on mutex A and could preempt the
3028 if (dl_prio(prio)) {
3029 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3030 if (!dl_prio(p->normal_prio) ||
3031 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3032 p->dl.dl_boosted = 1;
3033 p->dl.dl_throttled = 0;
3034 enqueue_flag = ENQUEUE_REPLENISH;
3036 p->dl.dl_boosted = 0;
3037 p->sched_class = &dl_sched_class;
3038 } else if (rt_prio(prio)) {
3039 if (dl_prio(oldprio))
3040 p->dl.dl_boosted = 0;
3042 enqueue_flag = ENQUEUE_HEAD;
3043 p->sched_class = &rt_sched_class;
3045 if (dl_prio(oldprio))
3046 p->dl.dl_boosted = 0;
3047 p->sched_class = &fair_sched_class;
3053 p->sched_class->set_curr_task(rq);
3055 enqueue_task(rq, p, enqueue_flag);
3057 check_class_changed(rq, p, prev_class, oldprio);
3059 __task_rq_unlock(rq);
3063 void set_user_nice(struct task_struct *p, long nice)
3065 int old_prio, delta, queued;
3066 unsigned long flags;
3069 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3072 * We have to be careful, if called from sys_setpriority(),
3073 * the task might be in the middle of scheduling on another CPU.
3075 rq = task_rq_lock(p, &flags);
3077 * The RT priorities are set via sched_setscheduler(), but we still
3078 * allow the 'normal' nice value to be set - but as expected
3079 * it wont have any effect on scheduling until the task is
3080 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3082 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3083 p->static_prio = NICE_TO_PRIO(nice);
3086 queued = task_on_rq_queued(p);
3088 dequeue_task(rq, p, 0);
3090 p->static_prio = NICE_TO_PRIO(nice);
3093 p->prio = effective_prio(p);
3094 delta = p->prio - old_prio;
3097 enqueue_task(rq, p, 0);
3099 * If the task increased its priority or is running and
3100 * lowered its priority, then reschedule its CPU:
3102 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3106 task_rq_unlock(rq, p, &flags);
3108 EXPORT_SYMBOL(set_user_nice);
3111 * can_nice - check if a task can reduce its nice value
3115 int can_nice(const struct task_struct *p, const int nice)
3117 /* convert nice value [19,-20] to rlimit style value [1,40] */
3118 int nice_rlim = nice_to_rlimit(nice);
3120 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3121 capable(CAP_SYS_NICE));
3124 #ifdef __ARCH_WANT_SYS_NICE
3127 * sys_nice - change the priority of the current process.
3128 * @increment: priority increment
3130 * sys_setpriority is a more generic, but much slower function that
3131 * does similar things.
3133 SYSCALL_DEFINE1(nice, int, increment)
3138 * Setpriority might change our priority at the same moment.
3139 * We don't have to worry. Conceptually one call occurs first
3140 * and we have a single winner.
3142 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3143 nice = task_nice(current) + increment;
3145 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3146 if (increment < 0 && !can_nice(current, nice))
3149 retval = security_task_setnice(current, nice);
3153 set_user_nice(current, nice);
3160 * task_prio - return the priority value of a given task.
3161 * @p: the task in question.
3163 * Return: The priority value as seen by users in /proc.
3164 * RT tasks are offset by -200. Normal tasks are centered
3165 * around 0, value goes from -16 to +15.
3167 int task_prio(const struct task_struct *p)
3169 return p->prio - MAX_RT_PRIO;
3173 * idle_cpu - is a given cpu idle currently?
3174 * @cpu: the processor in question.
3176 * Return: 1 if the CPU is currently idle. 0 otherwise.
3178 int idle_cpu(int cpu)
3180 struct rq *rq = cpu_rq(cpu);
3182 if (rq->curr != rq->idle)
3189 if (!llist_empty(&rq->wake_list))
3197 * idle_task - return the idle task for a given cpu.
3198 * @cpu: the processor in question.
3200 * Return: The idle task for the cpu @cpu.
3202 struct task_struct *idle_task(int cpu)
3204 return cpu_rq(cpu)->idle;
3208 * find_process_by_pid - find a process with a matching PID value.
3209 * @pid: the pid in question.
3211 * The task of @pid, if found. %NULL otherwise.
3213 static struct task_struct *find_process_by_pid(pid_t pid)
3215 return pid ? find_task_by_vpid(pid) : current;
3219 * This function initializes the sched_dl_entity of a newly becoming
3220 * SCHED_DEADLINE task.
3222 * Only the static values are considered here, the actual runtime and the
3223 * absolute deadline will be properly calculated when the task is enqueued
3224 * for the first time with its new policy.
3227 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3229 struct sched_dl_entity *dl_se = &p->dl;
3231 init_dl_task_timer(dl_se);
3232 dl_se->dl_runtime = attr->sched_runtime;
3233 dl_se->dl_deadline = attr->sched_deadline;
3234 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3235 dl_se->flags = attr->sched_flags;
3236 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3237 dl_se->dl_throttled = 0;
3239 dl_se->dl_yielded = 0;
3243 * sched_setparam() passes in -1 for its policy, to let the functions
3244 * it calls know not to change it.
3246 #define SETPARAM_POLICY -1
3248 static void __setscheduler_params(struct task_struct *p,
3249 const struct sched_attr *attr)
3251 int policy = attr->sched_policy;
3253 if (policy == SETPARAM_POLICY)
3258 if (dl_policy(policy))
3259 __setparam_dl(p, attr);
3260 else if (fair_policy(policy))
3261 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3264 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3265 * !rt_policy. Always setting this ensures that things like
3266 * getparam()/getattr() don't report silly values for !rt tasks.
3268 p->rt_priority = attr->sched_priority;
3269 p->normal_prio = normal_prio(p);
3273 /* Actually do priority change: must hold pi & rq lock. */
3274 static void __setscheduler(struct rq *rq, struct task_struct *p,
3275 const struct sched_attr *attr)
3277 __setscheduler_params(p, attr);
3280 * If we get here, there was no pi waiters boosting the
3281 * task. It is safe to use the normal prio.
3283 p->prio = normal_prio(p);
3285 if (dl_prio(p->prio))
3286 p->sched_class = &dl_sched_class;
3287 else if (rt_prio(p->prio))
3288 p->sched_class = &rt_sched_class;
3290 p->sched_class = &fair_sched_class;
3294 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3296 struct sched_dl_entity *dl_se = &p->dl;
3298 attr->sched_priority = p->rt_priority;
3299 attr->sched_runtime = dl_se->dl_runtime;
3300 attr->sched_deadline = dl_se->dl_deadline;
3301 attr->sched_period = dl_se->dl_period;
3302 attr->sched_flags = dl_se->flags;
3306 * This function validates the new parameters of a -deadline task.
3307 * We ask for the deadline not being zero, and greater or equal
3308 * than the runtime, as well as the period of being zero or
3309 * greater than deadline. Furthermore, we have to be sure that
3310 * user parameters are above the internal resolution of 1us (we
3311 * check sched_runtime only since it is always the smaller one) and
3312 * below 2^63 ns (we have to check both sched_deadline and
3313 * sched_period, as the latter can be zero).
3316 __checkparam_dl(const struct sched_attr *attr)
3319 if (attr->sched_deadline == 0)
3323 * Since we truncate DL_SCALE bits, make sure we're at least
3326 if (attr->sched_runtime < (1ULL << DL_SCALE))
3330 * Since we use the MSB for wrap-around and sign issues, make
3331 * sure it's not set (mind that period can be equal to zero).
3333 if (attr->sched_deadline & (1ULL << 63) ||
3334 attr->sched_period & (1ULL << 63))
3337 /* runtime <= deadline <= period (if period != 0) */
3338 if ((attr->sched_period != 0 &&
3339 attr->sched_period < attr->sched_deadline) ||
3340 attr->sched_deadline < attr->sched_runtime)
3347 * check the target process has a UID that matches the current process's
3349 static bool check_same_owner(struct task_struct *p)
3351 const struct cred *cred = current_cred(), *pcred;
3355 pcred = __task_cred(p);
3356 match = (uid_eq(cred->euid, pcred->euid) ||
3357 uid_eq(cred->euid, pcred->uid));
3362 static int __sched_setscheduler(struct task_struct *p,
3363 const struct sched_attr *attr,
3366 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3367 MAX_RT_PRIO - 1 - attr->sched_priority;
3368 int retval, oldprio, oldpolicy = -1, queued, running;
3369 int policy = attr->sched_policy;
3370 unsigned long flags;
3371 const struct sched_class *prev_class;
3375 /* may grab non-irq protected spin_locks */
3376 BUG_ON(in_interrupt());
3378 /* double check policy once rq lock held */
3380 reset_on_fork = p->sched_reset_on_fork;
3381 policy = oldpolicy = p->policy;
3383 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3385 if (policy != SCHED_DEADLINE &&
3386 policy != SCHED_FIFO && policy != SCHED_RR &&
3387 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3388 policy != SCHED_IDLE)
3392 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3396 * Valid priorities for SCHED_FIFO and SCHED_RR are
3397 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3398 * SCHED_BATCH and SCHED_IDLE is 0.
3400 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3401 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3403 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3404 (rt_policy(policy) != (attr->sched_priority != 0)))
3408 * Allow unprivileged RT tasks to decrease priority:
3410 if (user && !capable(CAP_SYS_NICE)) {
3411 if (fair_policy(policy)) {
3412 if (attr->sched_nice < task_nice(p) &&
3413 !can_nice(p, attr->sched_nice))
3417 if (rt_policy(policy)) {
3418 unsigned long rlim_rtprio =
3419 task_rlimit(p, RLIMIT_RTPRIO);
3421 /* can't set/change the rt policy */
3422 if (policy != p->policy && !rlim_rtprio)
3425 /* can't increase priority */
3426 if (attr->sched_priority > p->rt_priority &&
3427 attr->sched_priority > rlim_rtprio)
3432 * Can't set/change SCHED_DEADLINE policy at all for now
3433 * (safest behavior); in the future we would like to allow
3434 * unprivileged DL tasks to increase their relative deadline
3435 * or reduce their runtime (both ways reducing utilization)
3437 if (dl_policy(policy))
3441 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3442 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3444 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3445 if (!can_nice(p, task_nice(p)))
3449 /* can't change other user's priorities */
3450 if (!check_same_owner(p))
3453 /* Normal users shall not reset the sched_reset_on_fork flag */
3454 if (p->sched_reset_on_fork && !reset_on_fork)
3459 retval = security_task_setscheduler(p);
3465 * make sure no PI-waiters arrive (or leave) while we are
3466 * changing the priority of the task:
3468 * To be able to change p->policy safely, the appropriate
3469 * runqueue lock must be held.
3471 rq = task_rq_lock(p, &flags);
3474 * Changing the policy of the stop threads its a very bad idea
3476 if (p == rq->stop) {
3477 task_rq_unlock(rq, p, &flags);
3482 * If not changing anything there's no need to proceed further,
3483 * but store a possible modification of reset_on_fork.
3485 if (unlikely(policy == p->policy)) {
3486 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3488 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3490 if (dl_policy(policy))
3493 p->sched_reset_on_fork = reset_on_fork;
3494 task_rq_unlock(rq, p, &flags);
3500 #ifdef CONFIG_RT_GROUP_SCHED
3502 * Do not allow realtime tasks into groups that have no runtime
3505 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3506 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3507 !task_group_is_autogroup(task_group(p))) {
3508 task_rq_unlock(rq, p, &flags);
3513 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3514 cpumask_t *span = rq->rd->span;
3517 * Don't allow tasks with an affinity mask smaller than
3518 * the entire root_domain to become SCHED_DEADLINE. We
3519 * will also fail if there's no bandwidth available.
3521 if (!cpumask_subset(span, &p->cpus_allowed) ||
3522 rq->rd->dl_bw.bw == 0) {
3523 task_rq_unlock(rq, p, &flags);
3530 /* recheck policy now with rq lock held */
3531 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3532 policy = oldpolicy = -1;
3533 task_rq_unlock(rq, p, &flags);
3538 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3539 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3542 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3543 task_rq_unlock(rq, p, &flags);
3547 p->sched_reset_on_fork = reset_on_fork;
3551 * Special case for priority boosted tasks.
3553 * If the new priority is lower or equal (user space view)
3554 * than the current (boosted) priority, we just store the new
3555 * normal parameters and do not touch the scheduler class and
3556 * the runqueue. This will be done when the task deboost
3559 if (rt_mutex_check_prio(p, newprio)) {
3560 __setscheduler_params(p, attr);
3561 task_rq_unlock(rq, p, &flags);
3565 queued = task_on_rq_queued(p);
3566 running = task_current(rq, p);
3568 dequeue_task(rq, p, 0);
3570 p->sched_class->put_prev_task(rq, p);
3572 prev_class = p->sched_class;
3573 __setscheduler(rq, p, attr);
3576 p->sched_class->set_curr_task(rq);
3579 * We enqueue to tail when the priority of a task is
3580 * increased (user space view).
3582 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3585 check_class_changed(rq, p, prev_class, oldprio);
3586 task_rq_unlock(rq, p, &flags);
3588 rt_mutex_adjust_pi(p);
3593 static int _sched_setscheduler(struct task_struct *p, int policy,
3594 const struct sched_param *param, bool check)
3596 struct sched_attr attr = {
3597 .sched_policy = policy,
3598 .sched_priority = param->sched_priority,
3599 .sched_nice = PRIO_TO_NICE(p->static_prio),
3602 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3603 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3604 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3605 policy &= ~SCHED_RESET_ON_FORK;
3606 attr.sched_policy = policy;
3609 return __sched_setscheduler(p, &attr, check);
3612 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3613 * @p: the task in question.
3614 * @policy: new policy.
3615 * @param: structure containing the new RT priority.
3617 * Return: 0 on success. An error code otherwise.
3619 * NOTE that the task may be already dead.
3621 int sched_setscheduler(struct task_struct *p, int policy,
3622 const struct sched_param *param)
3624 return _sched_setscheduler(p, policy, param, true);
3626 EXPORT_SYMBOL_GPL(sched_setscheduler);
3628 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3630 return __sched_setscheduler(p, attr, true);
3632 EXPORT_SYMBOL_GPL(sched_setattr);
3635 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3636 * @p: the task in question.
3637 * @policy: new policy.
3638 * @param: structure containing the new RT priority.
3640 * Just like sched_setscheduler, only don't bother checking if the
3641 * current context has permission. For example, this is needed in
3642 * stop_machine(): we create temporary high priority worker threads,
3643 * but our caller might not have that capability.
3645 * Return: 0 on success. An error code otherwise.
3647 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3648 const struct sched_param *param)
3650 return _sched_setscheduler(p, policy, param, false);
3654 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3656 struct sched_param lparam;
3657 struct task_struct *p;
3660 if (!param || pid < 0)
3662 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3667 p = find_process_by_pid(pid);
3669 retval = sched_setscheduler(p, policy, &lparam);
3676 * Mimics kernel/events/core.c perf_copy_attr().
3678 static int sched_copy_attr(struct sched_attr __user *uattr,
3679 struct sched_attr *attr)
3684 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3688 * zero the full structure, so that a short copy will be nice.
3690 memset(attr, 0, sizeof(*attr));
3692 ret = get_user(size, &uattr->size);
3696 if (size > PAGE_SIZE) /* silly large */
3699 if (!size) /* abi compat */
3700 size = SCHED_ATTR_SIZE_VER0;
3702 if (size < SCHED_ATTR_SIZE_VER0)
3706 * If we're handed a bigger struct than we know of,
3707 * ensure all the unknown bits are 0 - i.e. new
3708 * user-space does not rely on any kernel feature
3709 * extensions we dont know about yet.
3711 if (size > sizeof(*attr)) {
3712 unsigned char __user *addr;
3713 unsigned char __user *end;
3716 addr = (void __user *)uattr + sizeof(*attr);
3717 end = (void __user *)uattr + size;
3719 for (; addr < end; addr++) {
3720 ret = get_user(val, addr);
3726 size = sizeof(*attr);
3729 ret = copy_from_user(attr, uattr, size);
3734 * XXX: do we want to be lenient like existing syscalls; or do we want
3735 * to be strict and return an error on out-of-bounds values?
3737 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3742 put_user(sizeof(*attr), &uattr->size);
3747 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3748 * @pid: the pid in question.
3749 * @policy: new policy.
3750 * @param: structure containing the new RT priority.
3752 * Return: 0 on success. An error code otherwise.
3754 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3755 struct sched_param __user *, param)
3757 /* negative values for policy are not valid */
3761 return do_sched_setscheduler(pid, policy, param);
3765 * sys_sched_setparam - set/change the RT priority of a thread
3766 * @pid: the pid in question.
3767 * @param: structure containing the new RT priority.
3769 * Return: 0 on success. An error code otherwise.
3771 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3773 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3777 * sys_sched_setattr - same as above, but with extended sched_attr
3778 * @pid: the pid in question.
3779 * @uattr: structure containing the extended parameters.
3780 * @flags: for future extension.
3782 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3783 unsigned int, flags)
3785 struct sched_attr attr;
3786 struct task_struct *p;
3789 if (!uattr || pid < 0 || flags)
3792 retval = sched_copy_attr(uattr, &attr);
3796 if ((int)attr.sched_policy < 0)
3801 p = find_process_by_pid(pid);
3803 retval = sched_setattr(p, &attr);
3810 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3811 * @pid: the pid in question.
3813 * Return: On success, the policy of the thread. Otherwise, a negative error
3816 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3818 struct task_struct *p;
3826 p = find_process_by_pid(pid);
3828 retval = security_task_getscheduler(p);
3831 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3838 * sys_sched_getparam - get the RT priority of a thread
3839 * @pid: the pid in question.
3840 * @param: structure containing the RT priority.
3842 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3845 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3847 struct sched_param lp = { .sched_priority = 0 };
3848 struct task_struct *p;
3851 if (!param || pid < 0)
3855 p = find_process_by_pid(pid);
3860 retval = security_task_getscheduler(p);
3864 if (task_has_rt_policy(p))
3865 lp.sched_priority = p->rt_priority;
3869 * This one might sleep, we cannot do it with a spinlock held ...
3871 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3880 static int sched_read_attr(struct sched_attr __user *uattr,
3881 struct sched_attr *attr,
3886 if (!access_ok(VERIFY_WRITE, uattr, usize))
3890 * If we're handed a smaller struct than we know of,
3891 * ensure all the unknown bits are 0 - i.e. old
3892 * user-space does not get uncomplete information.
3894 if (usize < sizeof(*attr)) {
3895 unsigned char *addr;
3898 addr = (void *)attr + usize;
3899 end = (void *)attr + sizeof(*attr);
3901 for (; addr < end; addr++) {
3909 ret = copy_to_user(uattr, attr, attr->size);
3917 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3918 * @pid: the pid in question.
3919 * @uattr: structure containing the extended parameters.
3920 * @size: sizeof(attr) for fwd/bwd comp.
3921 * @flags: for future extension.
3923 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3924 unsigned int, size, unsigned int, flags)
3926 struct sched_attr attr = {
3927 .size = sizeof(struct sched_attr),
3929 struct task_struct *p;
3932 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3933 size < SCHED_ATTR_SIZE_VER0 || flags)
3937 p = find_process_by_pid(pid);
3942 retval = security_task_getscheduler(p);
3946 attr.sched_policy = p->policy;
3947 if (p->sched_reset_on_fork)
3948 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3949 if (task_has_dl_policy(p))
3950 __getparam_dl(p, &attr);
3951 else if (task_has_rt_policy(p))
3952 attr.sched_priority = p->rt_priority;
3954 attr.sched_nice = task_nice(p);
3958 retval = sched_read_attr(uattr, &attr, size);
3966 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3968 cpumask_var_t cpus_allowed, new_mask;
3969 struct task_struct *p;
3974 p = find_process_by_pid(pid);
3980 /* Prevent p going away */
3984 if (p->flags & PF_NO_SETAFFINITY) {
3988 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3992 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3994 goto out_free_cpus_allowed;
3997 if (!check_same_owner(p)) {
3999 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4006 retval = security_task_setscheduler(p);
4011 cpuset_cpus_allowed(p, cpus_allowed);
4012 cpumask_and(new_mask, in_mask, cpus_allowed);
4015 * Since bandwidth control happens on root_domain basis,
4016 * if admission test is enabled, we only admit -deadline
4017 * tasks allowed to run on all the CPUs in the task's
4021 if (task_has_dl_policy(p)) {
4022 const struct cpumask *span = task_rq(p)->rd->span;
4024 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
4031 retval = set_cpus_allowed_ptr(p, new_mask);
4034 cpuset_cpus_allowed(p, cpus_allowed);
4035 if (!cpumask_subset(new_mask, cpus_allowed)) {
4037 * We must have raced with a concurrent cpuset
4038 * update. Just reset the cpus_allowed to the
4039 * cpuset's cpus_allowed
4041 cpumask_copy(new_mask, cpus_allowed);
4046 free_cpumask_var(new_mask);
4047 out_free_cpus_allowed:
4048 free_cpumask_var(cpus_allowed);
4054 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4055 struct cpumask *new_mask)
4057 if (len < cpumask_size())
4058 cpumask_clear(new_mask);
4059 else if (len > cpumask_size())
4060 len = cpumask_size();
4062 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4066 * sys_sched_setaffinity - set the cpu affinity of a process
4067 * @pid: pid of the process
4068 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4069 * @user_mask_ptr: user-space pointer to the new cpu mask
4071 * Return: 0 on success. An error code otherwise.
4073 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4074 unsigned long __user *, user_mask_ptr)
4076 cpumask_var_t new_mask;
4079 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4082 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4084 retval = sched_setaffinity(pid, new_mask);
4085 free_cpumask_var(new_mask);
4089 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4091 struct task_struct *p;
4092 unsigned long flags;
4098 p = find_process_by_pid(pid);
4102 retval = security_task_getscheduler(p);
4106 raw_spin_lock_irqsave(&p->pi_lock, flags);
4107 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4108 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4117 * sys_sched_getaffinity - get the cpu affinity of a process
4118 * @pid: pid of the process
4119 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4120 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4122 * Return: 0 on success. An error code otherwise.
4124 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4125 unsigned long __user *, user_mask_ptr)
4130 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4132 if (len & (sizeof(unsigned long)-1))
4135 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4138 ret = sched_getaffinity(pid, mask);
4140 size_t retlen = min_t(size_t, len, cpumask_size());
4142 if (copy_to_user(user_mask_ptr, mask, retlen))
4147 free_cpumask_var(mask);
4153 * sys_sched_yield - yield the current processor to other threads.
4155 * This function yields the current CPU to other tasks. If there are no
4156 * other threads running on this CPU then this function will return.
4160 SYSCALL_DEFINE0(sched_yield)
4162 struct rq *rq = this_rq_lock();
4164 schedstat_inc(rq, yld_count);
4165 current->sched_class->yield_task(rq);
4168 * Since we are going to call schedule() anyway, there's
4169 * no need to preempt or enable interrupts:
4171 __release(rq->lock);
4172 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4173 do_raw_spin_unlock(&rq->lock);
4174 sched_preempt_enable_no_resched();
4181 static void __cond_resched(void)
4183 __preempt_count_add(PREEMPT_ACTIVE);
4185 __preempt_count_sub(PREEMPT_ACTIVE);
4188 int __sched _cond_resched(void)
4190 if (should_resched()) {
4196 EXPORT_SYMBOL(_cond_resched);
4199 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4200 * call schedule, and on return reacquire the lock.
4202 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4203 * operations here to prevent schedule() from being called twice (once via
4204 * spin_unlock(), once by hand).
4206 int __cond_resched_lock(spinlock_t *lock)
4208 int resched = should_resched();
4211 lockdep_assert_held(lock);
4213 if (spin_needbreak(lock) || resched) {
4224 EXPORT_SYMBOL(__cond_resched_lock);
4226 int __sched __cond_resched_softirq(void)
4228 BUG_ON(!in_softirq());
4230 if (should_resched()) {
4238 EXPORT_SYMBOL(__cond_resched_softirq);
4241 * yield - yield the current processor to other threads.
4243 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4245 * The scheduler is at all times free to pick the calling task as the most
4246 * eligible task to run, if removing the yield() call from your code breaks
4247 * it, its already broken.
4249 * Typical broken usage is:
4254 * where one assumes that yield() will let 'the other' process run that will
4255 * make event true. If the current task is a SCHED_FIFO task that will never
4256 * happen. Never use yield() as a progress guarantee!!
4258 * If you want to use yield() to wait for something, use wait_event().
4259 * If you want to use yield() to be 'nice' for others, use cond_resched().
4260 * If you still want to use yield(), do not!
4262 void __sched yield(void)
4264 set_current_state(TASK_RUNNING);
4267 EXPORT_SYMBOL(yield);
4270 * yield_to - yield the current processor to another thread in
4271 * your thread group, or accelerate that thread toward the
4272 * processor it's on.
4274 * @preempt: whether task preemption is allowed or not
4276 * It's the caller's job to ensure that the target task struct
4277 * can't go away on us before we can do any checks.
4280 * true (>0) if we indeed boosted the target task.
4281 * false (0) if we failed to boost the target.
4282 * -ESRCH if there's no task to yield to.
4284 int __sched yield_to(struct task_struct *p, bool preempt)
4286 struct task_struct *curr = current;
4287 struct rq *rq, *p_rq;
4288 unsigned long flags;
4291 local_irq_save(flags);
4297 * If we're the only runnable task on the rq and target rq also
4298 * has only one task, there's absolutely no point in yielding.
4300 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4305 double_rq_lock(rq, p_rq);
4306 if (task_rq(p) != p_rq) {
4307 double_rq_unlock(rq, p_rq);
4311 if (!curr->sched_class->yield_to_task)
4314 if (curr->sched_class != p->sched_class)
4317 if (task_running(p_rq, p) || p->state)
4320 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4322 schedstat_inc(rq, yld_count);
4324 * Make p's CPU reschedule; pick_next_entity takes care of
4327 if (preempt && rq != p_rq)
4332 double_rq_unlock(rq, p_rq);
4334 local_irq_restore(flags);
4341 EXPORT_SYMBOL_GPL(yield_to);
4344 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4345 * that process accounting knows that this is a task in IO wait state.
4347 void __sched io_schedule(void)
4349 struct rq *rq = raw_rq();
4351 delayacct_blkio_start();
4352 atomic_inc(&rq->nr_iowait);
4353 blk_flush_plug(current);
4354 current->in_iowait = 1;
4356 current->in_iowait = 0;
4357 atomic_dec(&rq->nr_iowait);
4358 delayacct_blkio_end();
4360 EXPORT_SYMBOL(io_schedule);
4362 long __sched io_schedule_timeout(long timeout)
4364 struct rq *rq = raw_rq();
4367 delayacct_blkio_start();
4368 atomic_inc(&rq->nr_iowait);
4369 blk_flush_plug(current);
4370 current->in_iowait = 1;
4371 ret = schedule_timeout(timeout);
4372 current->in_iowait = 0;
4373 atomic_dec(&rq->nr_iowait);
4374 delayacct_blkio_end();
4379 * sys_sched_get_priority_max - return maximum RT priority.
4380 * @policy: scheduling class.
4382 * Return: On success, this syscall returns the maximum
4383 * rt_priority that can be used by a given scheduling class.
4384 * On failure, a negative error code is returned.
4386 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4393 ret = MAX_USER_RT_PRIO-1;
4395 case SCHED_DEADLINE:
4406 * sys_sched_get_priority_min - return minimum RT priority.
4407 * @policy: scheduling class.
4409 * Return: On success, this syscall returns the minimum
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_min, int, policy)
4422 case SCHED_DEADLINE:
4432 * sys_sched_rr_get_interval - return the default timeslice of a process.
4433 * @pid: pid of the process.
4434 * @interval: userspace pointer to the timeslice value.
4436 * this syscall writes the default timeslice value of a given process
4437 * into the user-space timespec buffer. A value of '0' means infinity.
4439 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4442 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4443 struct timespec __user *, interval)
4445 struct task_struct *p;
4446 unsigned int time_slice;
4447 unsigned long flags;
4457 p = find_process_by_pid(pid);
4461 retval = security_task_getscheduler(p);
4465 rq = task_rq_lock(p, &flags);
4467 if (p->sched_class->get_rr_interval)
4468 time_slice = p->sched_class->get_rr_interval(rq, p);
4469 task_rq_unlock(rq, p, &flags);
4472 jiffies_to_timespec(time_slice, &t);
4473 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4481 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4483 void sched_show_task(struct task_struct *p)
4485 unsigned long free = 0;
4489 state = p->state ? __ffs(p->state) + 1 : 0;
4490 printk(KERN_INFO "%-15.15s %c", p->comm,
4491 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4492 #if BITS_PER_LONG == 32
4493 if (state == TASK_RUNNING)
4494 printk(KERN_CONT " running ");
4496 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4498 if (state == TASK_RUNNING)
4499 printk(KERN_CONT " running task ");
4501 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4503 #ifdef CONFIG_DEBUG_STACK_USAGE
4504 free = stack_not_used(p);
4507 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4509 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4510 task_pid_nr(p), ppid,
4511 (unsigned long)task_thread_info(p)->flags);
4513 print_worker_info(KERN_INFO, p);
4514 show_stack(p, NULL);
4517 void show_state_filter(unsigned long state_filter)
4519 struct task_struct *g, *p;
4521 #if BITS_PER_LONG == 32
4523 " task PC stack pid father\n");
4526 " task PC stack pid father\n");
4529 for_each_process_thread(g, p) {
4531 * reset the NMI-timeout, listing all files on a slow
4532 * console might take a lot of time:
4534 touch_nmi_watchdog();
4535 if (!state_filter || (p->state & state_filter))
4539 touch_all_softlockup_watchdogs();
4541 #ifdef CONFIG_SCHED_DEBUG
4542 sysrq_sched_debug_show();
4546 * Only show locks if all tasks are dumped:
4549 debug_show_all_locks();
4552 void init_idle_bootup_task(struct task_struct *idle)
4554 idle->sched_class = &idle_sched_class;
4558 * init_idle - set up an idle thread for a given CPU
4559 * @idle: task in question
4560 * @cpu: cpu the idle task belongs to
4562 * NOTE: this function does not set the idle thread's NEED_RESCHED
4563 * flag, to make booting more robust.
4565 void init_idle(struct task_struct *idle, int cpu)
4567 struct rq *rq = cpu_rq(cpu);
4568 unsigned long flags;
4570 raw_spin_lock_irqsave(&rq->lock, flags);
4572 __sched_fork(0, idle);
4573 idle->state = TASK_RUNNING;
4574 idle->se.exec_start = sched_clock();
4576 do_set_cpus_allowed(idle, cpumask_of(cpu));
4578 * We're having a chicken and egg problem, even though we are
4579 * holding rq->lock, the cpu isn't yet set to this cpu so the
4580 * lockdep check in task_group() will fail.
4582 * Similar case to sched_fork(). / Alternatively we could
4583 * use task_rq_lock() here and obtain the other rq->lock.
4588 __set_task_cpu(idle, cpu);
4591 rq->curr = rq->idle = idle;
4592 idle->on_rq = TASK_ON_RQ_QUEUED;
4593 #if defined(CONFIG_SMP)
4596 raw_spin_unlock_irqrestore(&rq->lock, flags);
4598 /* Set the preempt count _outside_ the spinlocks! */
4599 init_idle_preempt_count(idle, cpu);
4602 * The idle tasks have their own, simple scheduling class:
4604 idle->sched_class = &idle_sched_class;
4605 ftrace_graph_init_idle_task(idle, cpu);
4606 vtime_init_idle(idle, cpu);
4607 #if defined(CONFIG_SMP)
4608 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4613 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4615 if (p->sched_class && p->sched_class->set_cpus_allowed)
4616 p->sched_class->set_cpus_allowed(p, new_mask);
4618 cpumask_copy(&p->cpus_allowed, new_mask);
4619 p->nr_cpus_allowed = cpumask_weight(new_mask);
4623 * This is how migration works:
4625 * 1) we invoke migration_cpu_stop() on the target CPU using
4627 * 2) stopper starts to run (implicitly forcing the migrated thread
4629 * 3) it checks whether the migrated task is still in the wrong runqueue.
4630 * 4) if it's in the wrong runqueue then the migration thread removes
4631 * it and puts it into the right queue.
4632 * 5) stopper completes and stop_one_cpu() returns and the migration
4637 * Change a given task's CPU affinity. Migrate the thread to a
4638 * proper CPU and schedule it away if the CPU it's executing on
4639 * is removed from the allowed bitmask.
4641 * NOTE: the caller must have a valid reference to the task, the
4642 * task must not exit() & deallocate itself prematurely. The
4643 * call is not atomic; no spinlocks may be held.
4645 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4647 unsigned long flags;
4649 unsigned int dest_cpu;
4652 rq = task_rq_lock(p, &flags);
4654 if (cpumask_equal(&p->cpus_allowed, new_mask))
4657 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4662 do_set_cpus_allowed(p, new_mask);
4664 /* Can the task run on the task's current CPU? If so, we're done */
4665 if (cpumask_test_cpu(task_cpu(p), new_mask))
4668 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4669 if (task_on_rq_queued(p) || p->state == TASK_WAKING) {
4670 struct migration_arg arg = { p, dest_cpu };
4671 /* Need help from migration thread: drop lock and wait. */
4672 task_rq_unlock(rq, p, &flags);
4673 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4674 tlb_migrate_finish(p->mm);
4678 task_rq_unlock(rq, p, &flags);
4682 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4685 * Move (not current) task off this cpu, onto dest cpu. We're doing
4686 * this because either it can't run here any more (set_cpus_allowed()
4687 * away from this CPU, or CPU going down), or because we're
4688 * attempting to rebalance this task on exec (sched_exec).
4690 * So we race with normal scheduler movements, but that's OK, as long
4691 * as the task is no longer on this CPU.
4693 * Returns non-zero if task was successfully migrated.
4695 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4700 if (unlikely(!cpu_active(dest_cpu)))
4703 rq = cpu_rq(src_cpu);
4705 raw_spin_lock(&p->pi_lock);
4706 raw_spin_lock(&rq->lock);
4707 /* Already moved. */
4708 if (task_cpu(p) != src_cpu)
4711 /* Affinity changed (again). */
4712 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4716 * If we're not on a rq, the next wake-up will ensure we're
4719 if (task_on_rq_queued(p)) {
4720 dequeue_task(rq, p, 0);
4721 p->on_rq = TASK_ON_RQ_MIGRATING;
4722 set_task_cpu(p, dest_cpu);
4723 raw_spin_unlock(&rq->lock);
4725 rq = cpu_rq(dest_cpu);
4726 raw_spin_lock(&rq->lock);
4727 BUG_ON(task_rq(p) != rq);
4728 p->on_rq = TASK_ON_RQ_QUEUED;
4729 enqueue_task(rq, p, 0);
4730 check_preempt_curr(rq, p, 0);
4735 raw_spin_unlock(&rq->lock);
4736 raw_spin_unlock(&p->pi_lock);
4740 #ifdef CONFIG_NUMA_BALANCING
4741 /* Migrate current task p to target_cpu */
4742 int migrate_task_to(struct task_struct *p, int target_cpu)
4744 struct migration_arg arg = { p, target_cpu };
4745 int curr_cpu = task_cpu(p);
4747 if (curr_cpu == target_cpu)
4750 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4753 /* TODO: This is not properly updating schedstats */
4755 trace_sched_move_numa(p, curr_cpu, target_cpu);
4756 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4760 * Requeue a task on a given node and accurately track the number of NUMA
4761 * tasks on the runqueues
4763 void sched_setnuma(struct task_struct *p, int nid)
4766 unsigned long flags;
4767 bool queued, running;
4769 rq = task_rq_lock(p, &flags);
4770 queued = task_on_rq_queued(p);
4771 running = task_current(rq, p);
4774 dequeue_task(rq, p, 0);
4776 p->sched_class->put_prev_task(rq, p);
4778 p->numa_preferred_nid = nid;
4781 p->sched_class->set_curr_task(rq);
4783 enqueue_task(rq, p, 0);
4784 task_rq_unlock(rq, p, &flags);
4789 * migration_cpu_stop - this will be executed by a highprio stopper thread
4790 * and performs thread migration by bumping thread off CPU then
4791 * 'pushing' onto another runqueue.
4793 static int migration_cpu_stop(void *data)
4795 struct migration_arg *arg = data;
4798 * The original target cpu might have gone down and we might
4799 * be on another cpu but it doesn't matter.
4801 local_irq_disable();
4803 * We need to explicitly wake pending tasks before running
4804 * __migrate_task() such that we will not miss enforcing cpus_allowed
4805 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4807 sched_ttwu_pending();
4808 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4813 #ifdef CONFIG_HOTPLUG_CPU
4816 * Ensures that the idle task is using init_mm right before its cpu goes
4819 void idle_task_exit(void)
4821 struct mm_struct *mm = current->active_mm;
4823 BUG_ON(cpu_online(smp_processor_id()));
4825 if (mm != &init_mm) {
4826 switch_mm(mm, &init_mm, current);
4827 finish_arch_post_lock_switch();
4833 * Since this CPU is going 'away' for a while, fold any nr_active delta
4834 * we might have. Assumes we're called after migrate_tasks() so that the
4835 * nr_active count is stable.
4837 * Also see the comment "Global load-average calculations".
4839 static void calc_load_migrate(struct rq *rq)
4841 long delta = calc_load_fold_active(rq);
4843 atomic_long_add(delta, &calc_load_tasks);
4846 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4850 static const struct sched_class fake_sched_class = {
4851 .put_prev_task = put_prev_task_fake,
4854 static struct task_struct fake_task = {
4856 * Avoid pull_{rt,dl}_task()
4858 .prio = MAX_PRIO + 1,
4859 .sched_class = &fake_sched_class,
4863 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4864 * try_to_wake_up()->select_task_rq().
4866 * Called with rq->lock held even though we'er in stop_machine() and
4867 * there's no concurrency possible, we hold the required locks anyway
4868 * because of lock validation efforts.
4870 static void migrate_tasks(unsigned int dead_cpu)
4872 struct rq *rq = cpu_rq(dead_cpu);
4873 struct task_struct *next, *stop = rq->stop;
4877 * Fudge the rq selection such that the below task selection loop
4878 * doesn't get stuck on the currently eligible stop task.
4880 * We're currently inside stop_machine() and the rq is either stuck
4881 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4882 * either way we should never end up calling schedule() until we're
4888 * put_prev_task() and pick_next_task() sched
4889 * class method both need to have an up-to-date
4890 * value of rq->clock[_task]
4892 update_rq_clock(rq);
4896 * There's this thread running, bail when that's the only
4899 if (rq->nr_running == 1)
4902 next = pick_next_task(rq, &fake_task);
4904 next->sched_class->put_prev_task(rq, next);
4906 /* Find suitable destination for @next, with force if needed. */
4907 dest_cpu = select_fallback_rq(dead_cpu, next);
4908 raw_spin_unlock(&rq->lock);
4910 __migrate_task(next, dead_cpu, dest_cpu);
4912 raw_spin_lock(&rq->lock);
4918 #endif /* CONFIG_HOTPLUG_CPU */
4920 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4922 static struct ctl_table sd_ctl_dir[] = {
4924 .procname = "sched_domain",
4930 static struct ctl_table sd_ctl_root[] = {
4932 .procname = "kernel",
4934 .child = sd_ctl_dir,
4939 static struct ctl_table *sd_alloc_ctl_entry(int n)
4941 struct ctl_table *entry =
4942 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4947 static void sd_free_ctl_entry(struct ctl_table **tablep)
4949 struct ctl_table *entry;
4952 * In the intermediate directories, both the child directory and
4953 * procname are dynamically allocated and could fail but the mode
4954 * will always be set. In the lowest directory the names are
4955 * static strings and all have proc handlers.
4957 for (entry = *tablep; entry->mode; entry++) {
4959 sd_free_ctl_entry(&entry->child);
4960 if (entry->proc_handler == NULL)
4961 kfree(entry->procname);
4968 static int min_load_idx = 0;
4969 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4972 set_table_entry(struct ctl_table *entry,
4973 const char *procname, void *data, int maxlen,
4974 umode_t mode, proc_handler *proc_handler,
4977 entry->procname = procname;
4979 entry->maxlen = maxlen;
4981 entry->proc_handler = proc_handler;
4984 entry->extra1 = &min_load_idx;
4985 entry->extra2 = &max_load_idx;
4989 static struct ctl_table *
4990 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4992 struct ctl_table *table = sd_alloc_ctl_entry(14);
4997 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4998 sizeof(long), 0644, proc_doulongvec_minmax, false);
4999 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5000 sizeof(long), 0644, proc_doulongvec_minmax, false);
5001 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5002 sizeof(int), 0644, proc_dointvec_minmax, true);
5003 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5004 sizeof(int), 0644, proc_dointvec_minmax, true);
5005 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5006 sizeof(int), 0644, proc_dointvec_minmax, true);
5007 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5008 sizeof(int), 0644, proc_dointvec_minmax, true);
5009 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5010 sizeof(int), 0644, proc_dointvec_minmax, true);
5011 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5012 sizeof(int), 0644, proc_dointvec_minmax, false);
5013 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5014 sizeof(int), 0644, proc_dointvec_minmax, false);
5015 set_table_entry(&table[9], "cache_nice_tries",
5016 &sd->cache_nice_tries,
5017 sizeof(int), 0644, proc_dointvec_minmax, false);
5018 set_table_entry(&table[10], "flags", &sd->flags,
5019 sizeof(int), 0644, proc_dointvec_minmax, false);
5020 set_table_entry(&table[11], "max_newidle_lb_cost",
5021 &sd->max_newidle_lb_cost,
5022 sizeof(long), 0644, proc_doulongvec_minmax, false);
5023 set_table_entry(&table[12], "name", sd->name,
5024 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5025 /* &table[13] is terminator */
5030 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5032 struct ctl_table *entry, *table;
5033 struct sched_domain *sd;
5034 int domain_num = 0, i;
5037 for_each_domain(cpu, sd)
5039 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5044 for_each_domain(cpu, sd) {
5045 snprintf(buf, 32, "domain%d", i);
5046 entry->procname = kstrdup(buf, GFP_KERNEL);
5048 entry->child = sd_alloc_ctl_domain_table(sd);
5055 static struct ctl_table_header *sd_sysctl_header;
5056 static void register_sched_domain_sysctl(void)
5058 int i, cpu_num = num_possible_cpus();
5059 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5062 WARN_ON(sd_ctl_dir[0].child);
5063 sd_ctl_dir[0].child = entry;
5068 for_each_possible_cpu(i) {
5069 snprintf(buf, 32, "cpu%d", i);
5070 entry->procname = kstrdup(buf, GFP_KERNEL);
5072 entry->child = sd_alloc_ctl_cpu_table(i);
5076 WARN_ON(sd_sysctl_header);
5077 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5080 /* may be called multiple times per register */
5081 static void unregister_sched_domain_sysctl(void)
5083 if (sd_sysctl_header)
5084 unregister_sysctl_table(sd_sysctl_header);
5085 sd_sysctl_header = NULL;
5086 if (sd_ctl_dir[0].child)
5087 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5090 static void register_sched_domain_sysctl(void)
5093 static void unregister_sched_domain_sysctl(void)
5098 static void set_rq_online(struct rq *rq)
5101 const struct sched_class *class;
5103 cpumask_set_cpu(rq->cpu, rq->rd->online);
5106 for_each_class(class) {
5107 if (class->rq_online)
5108 class->rq_online(rq);
5113 static void set_rq_offline(struct rq *rq)
5116 const struct sched_class *class;
5118 for_each_class(class) {
5119 if (class->rq_offline)
5120 class->rq_offline(rq);
5123 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5129 * migration_call - callback that gets triggered when a CPU is added.
5130 * Here we can start up the necessary migration thread for the new CPU.
5133 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5135 int cpu = (long)hcpu;
5136 unsigned long flags;
5137 struct rq *rq = cpu_rq(cpu);
5139 switch (action & ~CPU_TASKS_FROZEN) {
5141 case CPU_UP_PREPARE:
5142 rq->calc_load_update = calc_load_update;
5146 /* Update our root-domain */
5147 raw_spin_lock_irqsave(&rq->lock, flags);
5149 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5153 raw_spin_unlock_irqrestore(&rq->lock, flags);
5156 #ifdef CONFIG_HOTPLUG_CPU
5158 sched_ttwu_pending();
5159 /* Update our root-domain */
5160 raw_spin_lock_irqsave(&rq->lock, flags);
5162 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5166 BUG_ON(rq->nr_running != 1); /* the migration thread */
5167 raw_spin_unlock_irqrestore(&rq->lock, flags);
5171 calc_load_migrate(rq);
5176 update_max_interval();
5182 * Register at high priority so that task migration (migrate_all_tasks)
5183 * happens before everything else. This has to be lower priority than
5184 * the notifier in the perf_event subsystem, though.
5186 static struct notifier_block migration_notifier = {
5187 .notifier_call = migration_call,
5188 .priority = CPU_PRI_MIGRATION,
5191 static void __cpuinit set_cpu_rq_start_time(void)
5193 int cpu = smp_processor_id();
5194 struct rq *rq = cpu_rq(cpu);
5195 rq->age_stamp = sched_clock_cpu(cpu);
5198 static int sched_cpu_active(struct notifier_block *nfb,
5199 unsigned long action, void *hcpu)
5201 switch (action & ~CPU_TASKS_FROZEN) {
5203 set_cpu_rq_start_time();
5205 case CPU_DOWN_FAILED:
5206 set_cpu_active((long)hcpu, true);
5213 static int sched_cpu_inactive(struct notifier_block *nfb,
5214 unsigned long action, void *hcpu)
5216 unsigned long flags;
5217 long cpu = (long)hcpu;
5219 switch (action & ~CPU_TASKS_FROZEN) {
5220 case CPU_DOWN_PREPARE:
5221 set_cpu_active(cpu, false);
5223 /* explicitly allow suspend */
5224 if (!(action & CPU_TASKS_FROZEN)) {
5225 struct dl_bw *dl_b = dl_bw_of(cpu);
5229 raw_spin_lock_irqsave(&dl_b->lock, flags);
5230 cpus = dl_bw_cpus(cpu);
5231 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5232 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5235 return notifier_from_errno(-EBUSY);
5243 static int __init migration_init(void)
5245 void *cpu = (void *)(long)smp_processor_id();
5248 /* Initialize migration for the boot CPU */
5249 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5250 BUG_ON(err == NOTIFY_BAD);
5251 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5252 register_cpu_notifier(&migration_notifier);
5254 /* Register cpu active notifiers */
5255 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5256 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5260 early_initcall(migration_init);
5265 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5267 #ifdef CONFIG_SCHED_DEBUG
5269 static __read_mostly int sched_debug_enabled;
5271 static int __init sched_debug_setup(char *str)
5273 sched_debug_enabled = 1;
5277 early_param("sched_debug", sched_debug_setup);
5279 static inline bool sched_debug(void)
5281 return sched_debug_enabled;
5284 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5285 struct cpumask *groupmask)
5287 struct sched_group *group = sd->groups;
5290 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5291 cpumask_clear(groupmask);
5293 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5295 if (!(sd->flags & SD_LOAD_BALANCE)) {
5296 printk("does not load-balance\n");
5298 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5303 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5305 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5306 printk(KERN_ERR "ERROR: domain->span does not contain "
5309 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5310 printk(KERN_ERR "ERROR: domain->groups does not contain"
5314 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5318 printk(KERN_ERR "ERROR: group is NULL\n");
5323 * Even though we initialize ->capacity to something semi-sane,
5324 * we leave capacity_orig unset. This allows us to detect if
5325 * domain iteration is still funny without causing /0 traps.
5327 if (!group->sgc->capacity_orig) {
5328 printk(KERN_CONT "\n");
5329 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5333 if (!cpumask_weight(sched_group_cpus(group))) {
5334 printk(KERN_CONT "\n");
5335 printk(KERN_ERR "ERROR: empty group\n");
5339 if (!(sd->flags & SD_OVERLAP) &&
5340 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5341 printk(KERN_CONT "\n");
5342 printk(KERN_ERR "ERROR: repeated CPUs\n");
5346 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5348 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5350 printk(KERN_CONT " %s", str);
5351 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5352 printk(KERN_CONT " (cpu_capacity = %d)",
5353 group->sgc->capacity);
5356 group = group->next;
5357 } while (group != sd->groups);
5358 printk(KERN_CONT "\n");
5360 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5361 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5364 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5365 printk(KERN_ERR "ERROR: parent span is not a superset "
5366 "of domain->span\n");
5370 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5374 if (!sched_debug_enabled)
5378 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5382 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5385 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5393 #else /* !CONFIG_SCHED_DEBUG */
5394 # define sched_domain_debug(sd, cpu) do { } while (0)
5395 static inline bool sched_debug(void)
5399 #endif /* CONFIG_SCHED_DEBUG */
5401 static int sd_degenerate(struct sched_domain *sd)
5403 if (cpumask_weight(sched_domain_span(sd)) == 1)
5406 /* Following flags need at least 2 groups */
5407 if (sd->flags & (SD_LOAD_BALANCE |
5408 SD_BALANCE_NEWIDLE |
5411 SD_SHARE_CPUCAPACITY |
5412 SD_SHARE_PKG_RESOURCES |
5413 SD_SHARE_POWERDOMAIN)) {
5414 if (sd->groups != sd->groups->next)
5418 /* Following flags don't use groups */
5419 if (sd->flags & (SD_WAKE_AFFINE))
5426 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5428 unsigned long cflags = sd->flags, pflags = parent->flags;
5430 if (sd_degenerate(parent))
5433 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5436 /* Flags needing groups don't count if only 1 group in parent */
5437 if (parent->groups == parent->groups->next) {
5438 pflags &= ~(SD_LOAD_BALANCE |
5439 SD_BALANCE_NEWIDLE |
5442 SD_SHARE_CPUCAPACITY |
5443 SD_SHARE_PKG_RESOURCES |
5445 SD_SHARE_POWERDOMAIN);
5446 if (nr_node_ids == 1)
5447 pflags &= ~SD_SERIALIZE;
5449 if (~cflags & pflags)
5455 static void free_rootdomain(struct rcu_head *rcu)
5457 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5459 cpupri_cleanup(&rd->cpupri);
5460 cpudl_cleanup(&rd->cpudl);
5461 free_cpumask_var(rd->dlo_mask);
5462 free_cpumask_var(rd->rto_mask);
5463 free_cpumask_var(rd->online);
5464 free_cpumask_var(rd->span);
5468 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5470 struct root_domain *old_rd = NULL;
5471 unsigned long flags;
5473 raw_spin_lock_irqsave(&rq->lock, flags);
5478 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5481 cpumask_clear_cpu(rq->cpu, old_rd->span);
5484 * If we dont want to free the old_rd yet then
5485 * set old_rd to NULL to skip the freeing later
5488 if (!atomic_dec_and_test(&old_rd->refcount))
5492 atomic_inc(&rd->refcount);
5495 cpumask_set_cpu(rq->cpu, rd->span);
5496 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5499 raw_spin_unlock_irqrestore(&rq->lock, flags);
5502 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5505 static int init_rootdomain(struct root_domain *rd)
5507 memset(rd, 0, sizeof(*rd));
5509 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5511 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5513 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5515 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5518 init_dl_bw(&rd->dl_bw);
5519 if (cpudl_init(&rd->cpudl) != 0)
5522 if (cpupri_init(&rd->cpupri) != 0)
5527 free_cpumask_var(rd->rto_mask);
5529 free_cpumask_var(rd->dlo_mask);
5531 free_cpumask_var(rd->online);
5533 free_cpumask_var(rd->span);
5539 * By default the system creates a single root-domain with all cpus as
5540 * members (mimicking the global state we have today).
5542 struct root_domain def_root_domain;
5544 static void init_defrootdomain(void)
5546 init_rootdomain(&def_root_domain);
5548 atomic_set(&def_root_domain.refcount, 1);
5551 static struct root_domain *alloc_rootdomain(void)
5553 struct root_domain *rd;
5555 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5559 if (init_rootdomain(rd) != 0) {
5567 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5569 struct sched_group *tmp, *first;
5578 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5583 } while (sg != first);
5586 static void free_sched_domain(struct rcu_head *rcu)
5588 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5591 * If its an overlapping domain it has private groups, iterate and
5594 if (sd->flags & SD_OVERLAP) {
5595 free_sched_groups(sd->groups, 1);
5596 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5597 kfree(sd->groups->sgc);
5603 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5605 call_rcu(&sd->rcu, free_sched_domain);
5608 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5610 for (; sd; sd = sd->parent)
5611 destroy_sched_domain(sd, cpu);
5615 * Keep a special pointer to the highest sched_domain that has
5616 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5617 * allows us to avoid some pointer chasing select_idle_sibling().
5619 * Also keep a unique ID per domain (we use the first cpu number in
5620 * the cpumask of the domain), this allows us to quickly tell if
5621 * two cpus are in the same cache domain, see cpus_share_cache().
5623 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5624 DEFINE_PER_CPU(int, sd_llc_size);
5625 DEFINE_PER_CPU(int, sd_llc_id);
5626 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5627 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5628 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5630 static void update_top_cache_domain(int cpu)
5632 struct sched_domain *sd;
5633 struct sched_domain *busy_sd = NULL;
5637 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5639 id = cpumask_first(sched_domain_span(sd));
5640 size = cpumask_weight(sched_domain_span(sd));
5641 busy_sd = sd->parent; /* sd_busy */
5643 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5645 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5646 per_cpu(sd_llc_size, cpu) = size;
5647 per_cpu(sd_llc_id, cpu) = id;
5649 sd = lowest_flag_domain(cpu, SD_NUMA);
5650 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5652 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5653 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5657 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5658 * hold the hotplug lock.
5661 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5663 struct rq *rq = cpu_rq(cpu);
5664 struct sched_domain *tmp;
5666 /* Remove the sched domains which do not contribute to scheduling. */
5667 for (tmp = sd; tmp; ) {
5668 struct sched_domain *parent = tmp->parent;
5672 if (sd_parent_degenerate(tmp, parent)) {
5673 tmp->parent = parent->parent;
5675 parent->parent->child = tmp;
5677 * Transfer SD_PREFER_SIBLING down in case of a
5678 * degenerate parent; the spans match for this
5679 * so the property transfers.
5681 if (parent->flags & SD_PREFER_SIBLING)
5682 tmp->flags |= SD_PREFER_SIBLING;
5683 destroy_sched_domain(parent, cpu);
5688 if (sd && sd_degenerate(sd)) {
5691 destroy_sched_domain(tmp, cpu);
5696 sched_domain_debug(sd, cpu);
5698 rq_attach_root(rq, rd);
5700 rcu_assign_pointer(rq->sd, sd);
5701 destroy_sched_domains(tmp, cpu);
5703 update_top_cache_domain(cpu);
5706 /* cpus with isolated domains */
5707 static cpumask_var_t cpu_isolated_map;
5709 /* Setup the mask of cpus configured for isolated domains */
5710 static int __init isolated_cpu_setup(char *str)
5712 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5713 cpulist_parse(str, cpu_isolated_map);
5717 __setup("isolcpus=", isolated_cpu_setup);
5720 struct sched_domain ** __percpu sd;
5721 struct root_domain *rd;
5732 * Build an iteration mask that can exclude certain CPUs from the upwards
5735 * Asymmetric node setups can result in situations where the domain tree is of
5736 * unequal depth, make sure to skip domains that already cover the entire
5739 * In that case build_sched_domains() will have terminated the iteration early
5740 * and our sibling sd spans will be empty. Domains should always include the
5741 * cpu they're built on, so check that.
5744 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5746 const struct cpumask *span = sched_domain_span(sd);
5747 struct sd_data *sdd = sd->private;
5748 struct sched_domain *sibling;
5751 for_each_cpu(i, span) {
5752 sibling = *per_cpu_ptr(sdd->sd, i);
5753 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5756 cpumask_set_cpu(i, sched_group_mask(sg));
5761 * Return the canonical balance cpu for this group, this is the first cpu
5762 * of this group that's also in the iteration mask.
5764 int group_balance_cpu(struct sched_group *sg)
5766 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5770 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5772 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5773 const struct cpumask *span = sched_domain_span(sd);
5774 struct cpumask *covered = sched_domains_tmpmask;
5775 struct sd_data *sdd = sd->private;
5776 struct sched_domain *sibling;
5779 cpumask_clear(covered);
5781 for_each_cpu(i, span) {
5782 struct cpumask *sg_span;
5784 if (cpumask_test_cpu(i, covered))
5787 sibling = *per_cpu_ptr(sdd->sd, i);
5789 /* See the comment near build_group_mask(). */
5790 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5793 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5794 GFP_KERNEL, cpu_to_node(cpu));
5799 sg_span = sched_group_cpus(sg);
5801 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5803 cpumask_set_cpu(i, sg_span);
5805 cpumask_or(covered, covered, sg_span);
5807 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5808 if (atomic_inc_return(&sg->sgc->ref) == 1)
5809 build_group_mask(sd, sg);
5812 * Initialize sgc->capacity such that even if we mess up the
5813 * domains and no possible iteration will get us here, we won't
5816 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5817 sg->sgc->capacity_orig = sg->sgc->capacity;
5820 * Make sure the first group of this domain contains the
5821 * canonical balance cpu. Otherwise the sched_domain iteration
5822 * breaks. See update_sg_lb_stats().
5824 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5825 group_balance_cpu(sg) == cpu)
5835 sd->groups = groups;
5840 free_sched_groups(first, 0);
5845 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5847 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5848 struct sched_domain *child = sd->child;
5851 cpu = cpumask_first(sched_domain_span(child));
5854 *sg = *per_cpu_ptr(sdd->sg, cpu);
5855 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5856 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5863 * build_sched_groups will build a circular linked list of the groups
5864 * covered by the given span, and will set each group's ->cpumask correctly,
5865 * and ->cpu_capacity to 0.
5867 * Assumes the sched_domain tree is fully constructed
5870 build_sched_groups(struct sched_domain *sd, int cpu)
5872 struct sched_group *first = NULL, *last = NULL;
5873 struct sd_data *sdd = sd->private;
5874 const struct cpumask *span = sched_domain_span(sd);
5875 struct cpumask *covered;
5878 get_group(cpu, sdd, &sd->groups);
5879 atomic_inc(&sd->groups->ref);
5881 if (cpu != cpumask_first(span))
5884 lockdep_assert_held(&sched_domains_mutex);
5885 covered = sched_domains_tmpmask;
5887 cpumask_clear(covered);
5889 for_each_cpu(i, span) {
5890 struct sched_group *sg;
5893 if (cpumask_test_cpu(i, covered))
5896 group = get_group(i, sdd, &sg);
5897 cpumask_setall(sched_group_mask(sg));
5899 for_each_cpu(j, span) {
5900 if (get_group(j, sdd, NULL) != group)
5903 cpumask_set_cpu(j, covered);
5904 cpumask_set_cpu(j, sched_group_cpus(sg));
5919 * Initialize sched groups cpu_capacity.
5921 * cpu_capacity indicates the capacity of sched group, which is used while
5922 * distributing the load between different sched groups in a sched domain.
5923 * Typically cpu_capacity for all the groups in a sched domain will be same
5924 * unless there are asymmetries in the topology. If there are asymmetries,
5925 * group having more cpu_capacity will pickup more load compared to the
5926 * group having less cpu_capacity.
5928 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5930 struct sched_group *sg = sd->groups;
5935 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5937 } while (sg != sd->groups);
5939 if (cpu != group_balance_cpu(sg))
5942 update_group_capacity(sd, cpu);
5943 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
5947 * Initializers for schedule domains
5948 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5951 static int default_relax_domain_level = -1;
5952 int sched_domain_level_max;
5954 static int __init setup_relax_domain_level(char *str)
5956 if (kstrtoint(str, 0, &default_relax_domain_level))
5957 pr_warn("Unable to set relax_domain_level\n");
5961 __setup("relax_domain_level=", setup_relax_domain_level);
5963 static void set_domain_attribute(struct sched_domain *sd,
5964 struct sched_domain_attr *attr)
5968 if (!attr || attr->relax_domain_level < 0) {
5969 if (default_relax_domain_level < 0)
5972 request = default_relax_domain_level;
5974 request = attr->relax_domain_level;
5975 if (request < sd->level) {
5976 /* turn off idle balance on this domain */
5977 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5979 /* turn on idle balance on this domain */
5980 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5984 static void __sdt_free(const struct cpumask *cpu_map);
5985 static int __sdt_alloc(const struct cpumask *cpu_map);
5987 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5988 const struct cpumask *cpu_map)
5992 if (!atomic_read(&d->rd->refcount))
5993 free_rootdomain(&d->rd->rcu); /* fall through */
5995 free_percpu(d->sd); /* fall through */
5997 __sdt_free(cpu_map); /* fall through */
6003 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6004 const struct cpumask *cpu_map)
6006 memset(d, 0, sizeof(*d));
6008 if (__sdt_alloc(cpu_map))
6009 return sa_sd_storage;
6010 d->sd = alloc_percpu(struct sched_domain *);
6012 return sa_sd_storage;
6013 d->rd = alloc_rootdomain();
6016 return sa_rootdomain;
6020 * NULL the sd_data elements we've used to build the sched_domain and
6021 * sched_group structure so that the subsequent __free_domain_allocs()
6022 * will not free the data we're using.
6024 static void claim_allocations(int cpu, struct sched_domain *sd)
6026 struct sd_data *sdd = sd->private;
6028 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6029 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6031 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6032 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6034 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6035 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6039 static int sched_domains_numa_levels;
6040 static int *sched_domains_numa_distance;
6041 static struct cpumask ***sched_domains_numa_masks;
6042 static int sched_domains_curr_level;
6046 * SD_flags allowed in topology descriptions.
6048 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6049 * SD_SHARE_PKG_RESOURCES - describes shared caches
6050 * SD_NUMA - describes NUMA topologies
6051 * SD_SHARE_POWERDOMAIN - describes shared power domain
6054 * SD_ASYM_PACKING - describes SMT quirks
6056 #define TOPOLOGY_SD_FLAGS \
6057 (SD_SHARE_CPUCAPACITY | \
6058 SD_SHARE_PKG_RESOURCES | \
6061 SD_SHARE_POWERDOMAIN)
6063 static struct sched_domain *
6064 sd_init(struct sched_domain_topology_level *tl, int cpu)
6066 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6067 int sd_weight, sd_flags = 0;
6071 * Ugly hack to pass state to sd_numa_mask()...
6073 sched_domains_curr_level = tl->numa_level;
6076 sd_weight = cpumask_weight(tl->mask(cpu));
6079 sd_flags = (*tl->sd_flags)();
6080 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6081 "wrong sd_flags in topology description\n"))
6082 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6084 *sd = (struct sched_domain){
6085 .min_interval = sd_weight,
6086 .max_interval = 2*sd_weight,
6088 .imbalance_pct = 125,
6090 .cache_nice_tries = 0,
6097 .flags = 1*SD_LOAD_BALANCE
6098 | 1*SD_BALANCE_NEWIDLE
6103 | 0*SD_SHARE_CPUCAPACITY
6104 | 0*SD_SHARE_PKG_RESOURCES
6106 | 0*SD_PREFER_SIBLING
6111 .last_balance = jiffies,
6112 .balance_interval = sd_weight,
6114 .max_newidle_lb_cost = 0,
6115 .next_decay_max_lb_cost = jiffies,
6116 #ifdef CONFIG_SCHED_DEBUG
6122 * Convert topological properties into behaviour.
6125 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6126 sd->imbalance_pct = 110;
6127 sd->smt_gain = 1178; /* ~15% */
6129 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6130 sd->imbalance_pct = 117;
6131 sd->cache_nice_tries = 1;
6135 } else if (sd->flags & SD_NUMA) {
6136 sd->cache_nice_tries = 2;
6140 sd->flags |= SD_SERIALIZE;
6141 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6142 sd->flags &= ~(SD_BALANCE_EXEC |
6149 sd->flags |= SD_PREFER_SIBLING;
6150 sd->cache_nice_tries = 1;
6155 sd->private = &tl->data;
6161 * Topology list, bottom-up.
6163 static struct sched_domain_topology_level default_topology[] = {
6164 #ifdef CONFIG_SCHED_SMT
6165 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6167 #ifdef CONFIG_SCHED_MC
6168 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6170 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6174 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6176 #define for_each_sd_topology(tl) \
6177 for (tl = sched_domain_topology; tl->mask; tl++)
6179 void set_sched_topology(struct sched_domain_topology_level *tl)
6181 sched_domain_topology = tl;
6186 static const struct cpumask *sd_numa_mask(int cpu)
6188 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6191 static void sched_numa_warn(const char *str)
6193 static int done = false;
6201 printk(KERN_WARNING "ERROR: %s\n\n", str);
6203 for (i = 0; i < nr_node_ids; i++) {
6204 printk(KERN_WARNING " ");
6205 for (j = 0; j < nr_node_ids; j++)
6206 printk(KERN_CONT "%02d ", node_distance(i,j));
6207 printk(KERN_CONT "\n");
6209 printk(KERN_WARNING "\n");
6212 static bool find_numa_distance(int distance)
6216 if (distance == node_distance(0, 0))
6219 for (i = 0; i < sched_domains_numa_levels; i++) {
6220 if (sched_domains_numa_distance[i] == distance)
6227 static void sched_init_numa(void)
6229 int next_distance, curr_distance = node_distance(0, 0);
6230 struct sched_domain_topology_level *tl;
6234 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6235 if (!sched_domains_numa_distance)
6239 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6240 * unique distances in the node_distance() table.
6242 * Assumes node_distance(0,j) includes all distances in
6243 * node_distance(i,j) in order to avoid cubic time.
6245 next_distance = curr_distance;
6246 for (i = 0; i < nr_node_ids; i++) {
6247 for (j = 0; j < nr_node_ids; j++) {
6248 for (k = 0; k < nr_node_ids; k++) {
6249 int distance = node_distance(i, k);
6251 if (distance > curr_distance &&
6252 (distance < next_distance ||
6253 next_distance == curr_distance))
6254 next_distance = distance;
6257 * While not a strong assumption it would be nice to know
6258 * about cases where if node A is connected to B, B is not
6259 * equally connected to A.
6261 if (sched_debug() && node_distance(k, i) != distance)
6262 sched_numa_warn("Node-distance not symmetric");
6264 if (sched_debug() && i && !find_numa_distance(distance))
6265 sched_numa_warn("Node-0 not representative");
6267 if (next_distance != curr_distance) {
6268 sched_domains_numa_distance[level++] = next_distance;
6269 sched_domains_numa_levels = level;
6270 curr_distance = next_distance;
6275 * In case of sched_debug() we verify the above assumption.
6281 * 'level' contains the number of unique distances, excluding the
6282 * identity distance node_distance(i,i).
6284 * The sched_domains_numa_distance[] array includes the actual distance
6289 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6290 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6291 * the array will contain less then 'level' members. This could be
6292 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6293 * in other functions.
6295 * We reset it to 'level' at the end of this function.
6297 sched_domains_numa_levels = 0;
6299 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6300 if (!sched_domains_numa_masks)
6304 * Now for each level, construct a mask per node which contains all
6305 * cpus of nodes that are that many hops away from us.
6307 for (i = 0; i < level; i++) {
6308 sched_domains_numa_masks[i] =
6309 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6310 if (!sched_domains_numa_masks[i])
6313 for (j = 0; j < nr_node_ids; j++) {
6314 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6318 sched_domains_numa_masks[i][j] = mask;
6320 for (k = 0; k < nr_node_ids; k++) {
6321 if (node_distance(j, k) > sched_domains_numa_distance[i])
6324 cpumask_or(mask, mask, cpumask_of_node(k));
6329 /* Compute default topology size */
6330 for (i = 0; sched_domain_topology[i].mask; i++);
6332 tl = kzalloc((i + level + 1) *
6333 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6338 * Copy the default topology bits..
6340 for (i = 0; sched_domain_topology[i].mask; i++)
6341 tl[i] = sched_domain_topology[i];
6344 * .. and append 'j' levels of NUMA goodness.
6346 for (j = 0; j < level; i++, j++) {
6347 tl[i] = (struct sched_domain_topology_level){
6348 .mask = sd_numa_mask,
6349 .sd_flags = cpu_numa_flags,
6350 .flags = SDTL_OVERLAP,
6356 sched_domain_topology = tl;
6358 sched_domains_numa_levels = level;
6361 static void sched_domains_numa_masks_set(int cpu)
6364 int node = cpu_to_node(cpu);
6366 for (i = 0; i < sched_domains_numa_levels; i++) {
6367 for (j = 0; j < nr_node_ids; j++) {
6368 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6369 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6374 static void sched_domains_numa_masks_clear(int cpu)
6377 for (i = 0; i < sched_domains_numa_levels; i++) {
6378 for (j = 0; j < nr_node_ids; j++)
6379 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6384 * Update sched_domains_numa_masks[level][node] array when new cpus
6387 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6388 unsigned long action,
6391 int cpu = (long)hcpu;
6393 switch (action & ~CPU_TASKS_FROZEN) {
6395 sched_domains_numa_masks_set(cpu);
6399 sched_domains_numa_masks_clear(cpu);
6409 static inline void sched_init_numa(void)
6413 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6414 unsigned long action,
6419 #endif /* CONFIG_NUMA */
6421 static int __sdt_alloc(const struct cpumask *cpu_map)
6423 struct sched_domain_topology_level *tl;
6426 for_each_sd_topology(tl) {
6427 struct sd_data *sdd = &tl->data;
6429 sdd->sd = alloc_percpu(struct sched_domain *);
6433 sdd->sg = alloc_percpu(struct sched_group *);
6437 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6441 for_each_cpu(j, cpu_map) {
6442 struct sched_domain *sd;
6443 struct sched_group *sg;
6444 struct sched_group_capacity *sgc;
6446 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6447 GFP_KERNEL, cpu_to_node(j));
6451 *per_cpu_ptr(sdd->sd, j) = sd;
6453 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6454 GFP_KERNEL, cpu_to_node(j));
6460 *per_cpu_ptr(sdd->sg, j) = sg;
6462 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6463 GFP_KERNEL, cpu_to_node(j));
6467 *per_cpu_ptr(sdd->sgc, j) = sgc;
6474 static void __sdt_free(const struct cpumask *cpu_map)
6476 struct sched_domain_topology_level *tl;
6479 for_each_sd_topology(tl) {
6480 struct sd_data *sdd = &tl->data;
6482 for_each_cpu(j, cpu_map) {
6483 struct sched_domain *sd;
6486 sd = *per_cpu_ptr(sdd->sd, j);
6487 if (sd && (sd->flags & SD_OVERLAP))
6488 free_sched_groups(sd->groups, 0);
6489 kfree(*per_cpu_ptr(sdd->sd, j));
6493 kfree(*per_cpu_ptr(sdd->sg, j));
6495 kfree(*per_cpu_ptr(sdd->sgc, j));
6497 free_percpu(sdd->sd);
6499 free_percpu(sdd->sg);
6501 free_percpu(sdd->sgc);
6506 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6507 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6508 struct sched_domain *child, int cpu)
6510 struct sched_domain *sd = sd_init(tl, cpu);
6514 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6516 sd->level = child->level + 1;
6517 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6521 if (!cpumask_subset(sched_domain_span(child),
6522 sched_domain_span(sd))) {
6523 pr_err("BUG: arch topology borken\n");
6524 #ifdef CONFIG_SCHED_DEBUG
6525 pr_err(" the %s domain not a subset of the %s domain\n",
6526 child->name, sd->name);
6528 /* Fixup, ensure @sd has at least @child cpus. */
6529 cpumask_or(sched_domain_span(sd),
6530 sched_domain_span(sd),
6531 sched_domain_span(child));
6535 set_domain_attribute(sd, attr);
6541 * Build sched domains for a given set of cpus and attach the sched domains
6542 * to the individual cpus
6544 static int build_sched_domains(const struct cpumask *cpu_map,
6545 struct sched_domain_attr *attr)
6547 enum s_alloc alloc_state;
6548 struct sched_domain *sd;
6550 int i, ret = -ENOMEM;
6552 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6553 if (alloc_state != sa_rootdomain)
6556 /* Set up domains for cpus specified by the cpu_map. */
6557 for_each_cpu(i, cpu_map) {
6558 struct sched_domain_topology_level *tl;
6561 for_each_sd_topology(tl) {
6562 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6563 if (tl == sched_domain_topology)
6564 *per_cpu_ptr(d.sd, i) = sd;
6565 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6566 sd->flags |= SD_OVERLAP;
6567 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6572 /* Build the groups for the domains */
6573 for_each_cpu(i, cpu_map) {
6574 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6575 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6576 if (sd->flags & SD_OVERLAP) {
6577 if (build_overlap_sched_groups(sd, i))
6580 if (build_sched_groups(sd, i))
6586 /* Calculate CPU capacity for physical packages and nodes */
6587 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6588 if (!cpumask_test_cpu(i, cpu_map))
6591 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6592 claim_allocations(i, sd);
6593 init_sched_groups_capacity(i, sd);
6597 /* Attach the domains */
6599 for_each_cpu(i, cpu_map) {
6600 sd = *per_cpu_ptr(d.sd, i);
6601 cpu_attach_domain(sd, d.rd, i);
6607 __free_domain_allocs(&d, alloc_state, cpu_map);
6611 static cpumask_var_t *doms_cur; /* current sched domains */
6612 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6613 static struct sched_domain_attr *dattr_cur;
6614 /* attribues of custom domains in 'doms_cur' */
6617 * Special case: If a kmalloc of a doms_cur partition (array of
6618 * cpumask) fails, then fallback to a single sched domain,
6619 * as determined by the single cpumask fallback_doms.
6621 static cpumask_var_t fallback_doms;
6624 * arch_update_cpu_topology lets virtualized architectures update the
6625 * cpu core maps. It is supposed to return 1 if the topology changed
6626 * or 0 if it stayed the same.
6628 int __weak arch_update_cpu_topology(void)
6633 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6636 cpumask_var_t *doms;
6638 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6641 for (i = 0; i < ndoms; i++) {
6642 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6643 free_sched_domains(doms, i);
6650 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6653 for (i = 0; i < ndoms; i++)
6654 free_cpumask_var(doms[i]);
6659 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6660 * For now this just excludes isolated cpus, but could be used to
6661 * exclude other special cases in the future.
6663 static int init_sched_domains(const struct cpumask *cpu_map)
6667 arch_update_cpu_topology();
6669 doms_cur = alloc_sched_domains(ndoms_cur);
6671 doms_cur = &fallback_doms;
6672 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6673 err = build_sched_domains(doms_cur[0], NULL);
6674 register_sched_domain_sysctl();
6680 * Detach sched domains from a group of cpus specified in cpu_map
6681 * These cpus will now be attached to the NULL domain
6683 static void detach_destroy_domains(const struct cpumask *cpu_map)
6688 for_each_cpu(i, cpu_map)
6689 cpu_attach_domain(NULL, &def_root_domain, i);
6693 /* handle null as "default" */
6694 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6695 struct sched_domain_attr *new, int idx_new)
6697 struct sched_domain_attr tmp;
6704 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6705 new ? (new + idx_new) : &tmp,
6706 sizeof(struct sched_domain_attr));
6710 * Partition sched domains as specified by the 'ndoms_new'
6711 * cpumasks in the array doms_new[] of cpumasks. This compares
6712 * doms_new[] to the current sched domain partitioning, doms_cur[].
6713 * It destroys each deleted domain and builds each new domain.
6715 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6716 * The masks don't intersect (don't overlap.) We should setup one
6717 * sched domain for each mask. CPUs not in any of the cpumasks will
6718 * not be load balanced. If the same cpumask appears both in the
6719 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6722 * The passed in 'doms_new' should be allocated using
6723 * alloc_sched_domains. This routine takes ownership of it and will
6724 * free_sched_domains it when done with it. If the caller failed the
6725 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6726 * and partition_sched_domains() will fallback to the single partition
6727 * 'fallback_doms', it also forces the domains to be rebuilt.
6729 * If doms_new == NULL it will be replaced with cpu_online_mask.
6730 * ndoms_new == 0 is a special case for destroying existing domains,
6731 * and it will not create the default domain.
6733 * Call with hotplug lock held
6735 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6736 struct sched_domain_attr *dattr_new)
6741 mutex_lock(&sched_domains_mutex);
6743 /* always unregister in case we don't destroy any domains */
6744 unregister_sched_domain_sysctl();
6746 /* Let architecture update cpu core mappings. */
6747 new_topology = arch_update_cpu_topology();
6749 n = doms_new ? ndoms_new : 0;
6751 /* Destroy deleted domains */
6752 for (i = 0; i < ndoms_cur; i++) {
6753 for (j = 0; j < n && !new_topology; j++) {
6754 if (cpumask_equal(doms_cur[i], doms_new[j])
6755 && dattrs_equal(dattr_cur, i, dattr_new, j))
6758 /* no match - a current sched domain not in new doms_new[] */
6759 detach_destroy_domains(doms_cur[i]);
6765 if (doms_new == NULL) {
6767 doms_new = &fallback_doms;
6768 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6769 WARN_ON_ONCE(dattr_new);
6772 /* Build new domains */
6773 for (i = 0; i < ndoms_new; i++) {
6774 for (j = 0; j < n && !new_topology; j++) {
6775 if (cpumask_equal(doms_new[i], doms_cur[j])
6776 && dattrs_equal(dattr_new, i, dattr_cur, j))
6779 /* no match - add a new doms_new */
6780 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6785 /* Remember the new sched domains */
6786 if (doms_cur != &fallback_doms)
6787 free_sched_domains(doms_cur, ndoms_cur);
6788 kfree(dattr_cur); /* kfree(NULL) is safe */
6789 doms_cur = doms_new;
6790 dattr_cur = dattr_new;
6791 ndoms_cur = ndoms_new;
6793 register_sched_domain_sysctl();
6795 mutex_unlock(&sched_domains_mutex);
6798 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6801 * Update cpusets according to cpu_active mask. If cpusets are
6802 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6803 * around partition_sched_domains().
6805 * If we come here as part of a suspend/resume, don't touch cpusets because we
6806 * want to restore it back to its original state upon resume anyway.
6808 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6812 case CPU_ONLINE_FROZEN:
6813 case CPU_DOWN_FAILED_FROZEN:
6816 * num_cpus_frozen tracks how many CPUs are involved in suspend
6817 * resume sequence. As long as this is not the last online
6818 * operation in the resume sequence, just build a single sched
6819 * domain, ignoring cpusets.
6822 if (likely(num_cpus_frozen)) {
6823 partition_sched_domains(1, NULL, NULL);
6828 * This is the last CPU online operation. So fall through and
6829 * restore the original sched domains by considering the
6830 * cpuset configurations.
6834 case CPU_DOWN_FAILED:
6835 cpuset_update_active_cpus(true);
6843 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6847 case CPU_DOWN_PREPARE:
6848 cpuset_update_active_cpus(false);
6850 case CPU_DOWN_PREPARE_FROZEN:
6852 partition_sched_domains(1, NULL, NULL);
6860 void __init sched_init_smp(void)
6862 cpumask_var_t non_isolated_cpus;
6864 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6865 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6870 * There's no userspace yet to cause hotplug operations; hence all the
6871 * cpu masks are stable and all blatant races in the below code cannot
6874 mutex_lock(&sched_domains_mutex);
6875 init_sched_domains(cpu_active_mask);
6876 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6877 if (cpumask_empty(non_isolated_cpus))
6878 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6879 mutex_unlock(&sched_domains_mutex);
6881 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6882 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6883 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6887 /* Move init over to a non-isolated CPU */
6888 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6890 sched_init_granularity();
6891 free_cpumask_var(non_isolated_cpus);
6893 init_sched_rt_class();
6894 init_sched_dl_class();
6897 void __init sched_init_smp(void)
6899 sched_init_granularity();
6901 #endif /* CONFIG_SMP */
6903 const_debug unsigned int sysctl_timer_migration = 1;
6905 int in_sched_functions(unsigned long addr)
6907 return in_lock_functions(addr) ||
6908 (addr >= (unsigned long)__sched_text_start
6909 && addr < (unsigned long)__sched_text_end);
6912 #ifdef CONFIG_CGROUP_SCHED
6914 * Default task group.
6915 * Every task in system belongs to this group at bootup.
6917 struct task_group root_task_group;
6918 LIST_HEAD(task_groups);
6921 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6923 void __init sched_init(void)
6926 unsigned long alloc_size = 0, ptr;
6928 #ifdef CONFIG_FAIR_GROUP_SCHED
6929 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6931 #ifdef CONFIG_RT_GROUP_SCHED
6932 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6934 #ifdef CONFIG_CPUMASK_OFFSTACK
6935 alloc_size += num_possible_cpus() * cpumask_size();
6938 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6940 #ifdef CONFIG_FAIR_GROUP_SCHED
6941 root_task_group.se = (struct sched_entity **)ptr;
6942 ptr += nr_cpu_ids * sizeof(void **);
6944 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6945 ptr += nr_cpu_ids * sizeof(void **);
6947 #endif /* CONFIG_FAIR_GROUP_SCHED */
6948 #ifdef CONFIG_RT_GROUP_SCHED
6949 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6950 ptr += nr_cpu_ids * sizeof(void **);
6952 root_task_group.rt_rq = (struct rt_rq **)ptr;
6953 ptr += nr_cpu_ids * sizeof(void **);
6955 #endif /* CONFIG_RT_GROUP_SCHED */
6956 #ifdef CONFIG_CPUMASK_OFFSTACK
6957 for_each_possible_cpu(i) {
6958 per_cpu(load_balance_mask, i) = (void *)ptr;
6959 ptr += cpumask_size();
6961 #endif /* CONFIG_CPUMASK_OFFSTACK */
6964 init_rt_bandwidth(&def_rt_bandwidth,
6965 global_rt_period(), global_rt_runtime());
6966 init_dl_bandwidth(&def_dl_bandwidth,
6967 global_rt_period(), global_rt_runtime());
6970 init_defrootdomain();
6973 #ifdef CONFIG_RT_GROUP_SCHED
6974 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6975 global_rt_period(), global_rt_runtime());
6976 #endif /* CONFIG_RT_GROUP_SCHED */
6978 #ifdef CONFIG_CGROUP_SCHED
6979 list_add(&root_task_group.list, &task_groups);
6980 INIT_LIST_HEAD(&root_task_group.children);
6981 INIT_LIST_HEAD(&root_task_group.siblings);
6982 autogroup_init(&init_task);
6984 #endif /* CONFIG_CGROUP_SCHED */
6986 for_each_possible_cpu(i) {
6990 raw_spin_lock_init(&rq->lock);
6992 rq->calc_load_active = 0;
6993 rq->calc_load_update = jiffies + LOAD_FREQ;
6994 init_cfs_rq(&rq->cfs);
6995 init_rt_rq(&rq->rt, rq);
6996 init_dl_rq(&rq->dl, rq);
6997 #ifdef CONFIG_FAIR_GROUP_SCHED
6998 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6999 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7001 * How much cpu bandwidth does root_task_group get?
7003 * In case of task-groups formed thr' the cgroup filesystem, it
7004 * gets 100% of the cpu resources in the system. This overall
7005 * system cpu resource is divided among the tasks of
7006 * root_task_group and its child task-groups in a fair manner,
7007 * based on each entity's (task or task-group's) weight
7008 * (se->load.weight).
7010 * In other words, if root_task_group has 10 tasks of weight
7011 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7012 * then A0's share of the cpu resource is:
7014 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7016 * We achieve this by letting root_task_group's tasks sit
7017 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7019 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7020 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7021 #endif /* CONFIG_FAIR_GROUP_SCHED */
7023 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7024 #ifdef CONFIG_RT_GROUP_SCHED
7025 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7028 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7029 rq->cpu_load[j] = 0;
7031 rq->last_load_update_tick = jiffies;
7036 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7037 rq->post_schedule = 0;
7038 rq->active_balance = 0;
7039 rq->next_balance = jiffies;
7044 rq->avg_idle = 2*sysctl_sched_migration_cost;
7045 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7047 INIT_LIST_HEAD(&rq->cfs_tasks);
7049 rq_attach_root(rq, &def_root_domain);
7050 #ifdef CONFIG_NO_HZ_COMMON
7053 #ifdef CONFIG_NO_HZ_FULL
7054 rq->last_sched_tick = 0;
7058 atomic_set(&rq->nr_iowait, 0);
7061 set_load_weight(&init_task);
7063 #ifdef CONFIG_PREEMPT_NOTIFIERS
7064 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7068 * The boot idle thread does lazy MMU switching as well:
7070 atomic_inc(&init_mm.mm_count);
7071 enter_lazy_tlb(&init_mm, current);
7074 * Make us the idle thread. Technically, schedule() should not be
7075 * called from this thread, however somewhere below it might be,
7076 * but because we are the idle thread, we just pick up running again
7077 * when this runqueue becomes "idle".
7079 init_idle(current, smp_processor_id());
7081 calc_load_update = jiffies + LOAD_FREQ;
7084 * During early bootup we pretend to be a normal task:
7086 current->sched_class = &fair_sched_class;
7089 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7090 /* May be allocated at isolcpus cmdline parse time */
7091 if (cpu_isolated_map == NULL)
7092 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7093 idle_thread_set_boot_cpu();
7094 set_cpu_rq_start_time();
7096 init_sched_fair_class();
7098 scheduler_running = 1;
7101 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7102 static inline int preempt_count_equals(int preempt_offset)
7104 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7106 return (nested == preempt_offset);
7109 void __might_sleep(const char *file, int line, int preempt_offset)
7111 static unsigned long prev_jiffy; /* ratelimiting */
7113 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7114 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7115 !is_idle_task(current)) ||
7116 system_state != SYSTEM_RUNNING || oops_in_progress)
7118 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7120 prev_jiffy = jiffies;
7123 "BUG: sleeping function called from invalid context at %s:%d\n",
7126 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7127 in_atomic(), irqs_disabled(),
7128 current->pid, current->comm);
7130 debug_show_held_locks(current);
7131 if (irqs_disabled())
7132 print_irqtrace_events(current);
7133 #ifdef CONFIG_DEBUG_PREEMPT
7134 if (!preempt_count_equals(preempt_offset)) {
7135 pr_err("Preemption disabled at:");
7136 print_ip_sym(current->preempt_disable_ip);
7142 EXPORT_SYMBOL(__might_sleep);
7145 #ifdef CONFIG_MAGIC_SYSRQ
7146 static void normalize_task(struct rq *rq, struct task_struct *p)
7148 const struct sched_class *prev_class = p->sched_class;
7149 struct sched_attr attr = {
7150 .sched_policy = SCHED_NORMAL,
7152 int old_prio = p->prio;
7155 queued = task_on_rq_queued(p);
7157 dequeue_task(rq, p, 0);
7158 __setscheduler(rq, p, &attr);
7160 enqueue_task(rq, p, 0);
7164 check_class_changed(rq, p, prev_class, old_prio);
7167 void normalize_rt_tasks(void)
7169 struct task_struct *g, *p;
7170 unsigned long flags;
7173 read_lock_irqsave(&tasklist_lock, flags);
7174 for_each_process_thread(g, p) {
7176 * Only normalize user tasks:
7181 p->se.exec_start = 0;
7182 #ifdef CONFIG_SCHEDSTATS
7183 p->se.statistics.wait_start = 0;
7184 p->se.statistics.sleep_start = 0;
7185 p->se.statistics.block_start = 0;
7188 if (!dl_task(p) && !rt_task(p)) {
7190 * Renice negative nice level userspace
7193 if (task_nice(p) < 0 && p->mm)
7194 set_user_nice(p, 0);
7198 raw_spin_lock(&p->pi_lock);
7199 rq = __task_rq_lock(p);
7201 normalize_task(rq, p);
7203 __task_rq_unlock(rq);
7204 raw_spin_unlock(&p->pi_lock);
7206 read_unlock_irqrestore(&tasklist_lock, flags);
7209 #endif /* CONFIG_MAGIC_SYSRQ */
7211 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7213 * These functions are only useful for the IA64 MCA handling, or kdb.
7215 * They can only be called when the whole system has been
7216 * stopped - every CPU needs to be quiescent, and no scheduling
7217 * activity can take place. Using them for anything else would
7218 * be a serious bug, and as a result, they aren't even visible
7219 * under any other configuration.
7223 * curr_task - return the current task for a given cpu.
7224 * @cpu: the processor in question.
7226 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7228 * Return: The current task for @cpu.
7230 struct task_struct *curr_task(int cpu)
7232 return cpu_curr(cpu);
7235 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7239 * set_curr_task - set the current task for a given cpu.
7240 * @cpu: the processor in question.
7241 * @p: the task pointer to set.
7243 * Description: This function must only be used when non-maskable interrupts
7244 * are serviced on a separate stack. It allows the architecture to switch the
7245 * notion of the current task on a cpu in a non-blocking manner. This function
7246 * must be called with all CPU's synchronized, and interrupts disabled, the
7247 * and caller must save the original value of the current task (see
7248 * curr_task() above) and restore that value before reenabling interrupts and
7249 * re-starting the system.
7251 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7253 void set_curr_task(int cpu, struct task_struct *p)
7260 #ifdef CONFIG_CGROUP_SCHED
7261 /* task_group_lock serializes the addition/removal of task groups */
7262 static DEFINE_SPINLOCK(task_group_lock);
7264 static void free_sched_group(struct task_group *tg)
7266 free_fair_sched_group(tg);
7267 free_rt_sched_group(tg);
7272 /* allocate runqueue etc for a new task group */
7273 struct task_group *sched_create_group(struct task_group *parent)
7275 struct task_group *tg;
7277 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7279 return ERR_PTR(-ENOMEM);
7281 if (!alloc_fair_sched_group(tg, parent))
7284 if (!alloc_rt_sched_group(tg, parent))
7290 free_sched_group(tg);
7291 return ERR_PTR(-ENOMEM);
7294 void sched_online_group(struct task_group *tg, struct task_group *parent)
7296 unsigned long flags;
7298 spin_lock_irqsave(&task_group_lock, flags);
7299 list_add_rcu(&tg->list, &task_groups);
7301 WARN_ON(!parent); /* root should already exist */
7303 tg->parent = parent;
7304 INIT_LIST_HEAD(&tg->children);
7305 list_add_rcu(&tg->siblings, &parent->children);
7306 spin_unlock_irqrestore(&task_group_lock, flags);
7309 /* rcu callback to free various structures associated with a task group */
7310 static void free_sched_group_rcu(struct rcu_head *rhp)
7312 /* now it should be safe to free those cfs_rqs */
7313 free_sched_group(container_of(rhp, struct task_group, rcu));
7316 /* Destroy runqueue etc associated with a task group */
7317 void sched_destroy_group(struct task_group *tg)
7319 /* wait for possible concurrent references to cfs_rqs complete */
7320 call_rcu(&tg->rcu, free_sched_group_rcu);
7323 void sched_offline_group(struct task_group *tg)
7325 unsigned long flags;
7328 /* end participation in shares distribution */
7329 for_each_possible_cpu(i)
7330 unregister_fair_sched_group(tg, i);
7332 spin_lock_irqsave(&task_group_lock, flags);
7333 list_del_rcu(&tg->list);
7334 list_del_rcu(&tg->siblings);
7335 spin_unlock_irqrestore(&task_group_lock, flags);
7338 /* change task's runqueue when it moves between groups.
7339 * The caller of this function should have put the task in its new group
7340 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7341 * reflect its new group.
7343 void sched_move_task(struct task_struct *tsk)
7345 struct task_group *tg;
7346 int queued, running;
7347 unsigned long flags;
7350 rq = task_rq_lock(tsk, &flags);
7352 running = task_current(rq, tsk);
7353 queued = task_on_rq_queued(tsk);
7356 dequeue_task(rq, tsk, 0);
7357 if (unlikely(running))
7358 tsk->sched_class->put_prev_task(rq, tsk);
7360 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7361 lockdep_is_held(&tsk->sighand->siglock)),
7362 struct task_group, css);
7363 tg = autogroup_task_group(tsk, tg);
7364 tsk->sched_task_group = tg;
7366 #ifdef CONFIG_FAIR_GROUP_SCHED
7367 if (tsk->sched_class->task_move_group)
7368 tsk->sched_class->task_move_group(tsk, queued);
7371 set_task_rq(tsk, task_cpu(tsk));
7373 if (unlikely(running))
7374 tsk->sched_class->set_curr_task(rq);
7376 enqueue_task(rq, tsk, 0);
7378 task_rq_unlock(rq, tsk, &flags);
7380 #endif /* CONFIG_CGROUP_SCHED */
7382 #ifdef CONFIG_RT_GROUP_SCHED
7384 * Ensure that the real time constraints are schedulable.
7386 static DEFINE_MUTEX(rt_constraints_mutex);
7388 /* Must be called with tasklist_lock held */
7389 static inline int tg_has_rt_tasks(struct task_group *tg)
7391 struct task_struct *g, *p;
7393 for_each_process_thread(g, p) {
7394 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7401 struct rt_schedulable_data {
7402 struct task_group *tg;
7407 static int tg_rt_schedulable(struct task_group *tg, void *data)
7409 struct rt_schedulable_data *d = data;
7410 struct task_group *child;
7411 unsigned long total, sum = 0;
7412 u64 period, runtime;
7414 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7415 runtime = tg->rt_bandwidth.rt_runtime;
7418 period = d->rt_period;
7419 runtime = d->rt_runtime;
7423 * Cannot have more runtime than the period.
7425 if (runtime > period && runtime != RUNTIME_INF)
7429 * Ensure we don't starve existing RT tasks.
7431 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7434 total = to_ratio(period, runtime);
7437 * Nobody can have more than the global setting allows.
7439 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7443 * The sum of our children's runtime should not exceed our own.
7445 list_for_each_entry_rcu(child, &tg->children, siblings) {
7446 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7447 runtime = child->rt_bandwidth.rt_runtime;
7449 if (child == d->tg) {
7450 period = d->rt_period;
7451 runtime = d->rt_runtime;
7454 sum += to_ratio(period, runtime);
7463 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7467 struct rt_schedulable_data data = {
7469 .rt_period = period,
7470 .rt_runtime = runtime,
7474 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7480 static int tg_set_rt_bandwidth(struct task_group *tg,
7481 u64 rt_period, u64 rt_runtime)
7485 mutex_lock(&rt_constraints_mutex);
7486 read_lock(&tasklist_lock);
7487 err = __rt_schedulable(tg, rt_period, rt_runtime);
7491 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7492 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7493 tg->rt_bandwidth.rt_runtime = rt_runtime;
7495 for_each_possible_cpu(i) {
7496 struct rt_rq *rt_rq = tg->rt_rq[i];
7498 raw_spin_lock(&rt_rq->rt_runtime_lock);
7499 rt_rq->rt_runtime = rt_runtime;
7500 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7502 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7504 read_unlock(&tasklist_lock);
7505 mutex_unlock(&rt_constraints_mutex);
7510 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7512 u64 rt_runtime, rt_period;
7514 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7515 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7516 if (rt_runtime_us < 0)
7517 rt_runtime = RUNTIME_INF;
7519 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7522 static long sched_group_rt_runtime(struct task_group *tg)
7526 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7529 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7530 do_div(rt_runtime_us, NSEC_PER_USEC);
7531 return rt_runtime_us;
7534 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7536 u64 rt_runtime, rt_period;
7538 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7539 rt_runtime = tg->rt_bandwidth.rt_runtime;
7544 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7547 static long sched_group_rt_period(struct task_group *tg)
7551 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7552 do_div(rt_period_us, NSEC_PER_USEC);
7553 return rt_period_us;
7555 #endif /* CONFIG_RT_GROUP_SCHED */
7557 #ifdef CONFIG_RT_GROUP_SCHED
7558 static int sched_rt_global_constraints(void)
7562 mutex_lock(&rt_constraints_mutex);
7563 read_lock(&tasklist_lock);
7564 ret = __rt_schedulable(NULL, 0, 0);
7565 read_unlock(&tasklist_lock);
7566 mutex_unlock(&rt_constraints_mutex);
7571 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7573 /* Don't accept realtime tasks when there is no way for them to run */
7574 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7580 #else /* !CONFIG_RT_GROUP_SCHED */
7581 static int sched_rt_global_constraints(void)
7583 unsigned long flags;
7586 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7587 for_each_possible_cpu(i) {
7588 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7590 raw_spin_lock(&rt_rq->rt_runtime_lock);
7591 rt_rq->rt_runtime = global_rt_runtime();
7592 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7594 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7598 #endif /* CONFIG_RT_GROUP_SCHED */
7600 static int sched_dl_global_constraints(void)
7602 u64 runtime = global_rt_runtime();
7603 u64 period = global_rt_period();
7604 u64 new_bw = to_ratio(period, runtime);
7606 unsigned long flags;
7609 * Here we want to check the bandwidth not being set to some
7610 * value smaller than the currently allocated bandwidth in
7611 * any of the root_domains.
7613 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7614 * cycling on root_domains... Discussion on different/better
7615 * solutions is welcome!
7617 for_each_possible_cpu(cpu) {
7618 struct dl_bw *dl_b = dl_bw_of(cpu);
7620 raw_spin_lock_irqsave(&dl_b->lock, flags);
7621 if (new_bw < dl_b->total_bw)
7623 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7632 static void sched_dl_do_global(void)
7636 unsigned long flags;
7638 def_dl_bandwidth.dl_period = global_rt_period();
7639 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7641 if (global_rt_runtime() != RUNTIME_INF)
7642 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7645 * FIXME: As above...
7647 for_each_possible_cpu(cpu) {
7648 struct dl_bw *dl_b = dl_bw_of(cpu);
7650 raw_spin_lock_irqsave(&dl_b->lock, flags);
7652 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7656 static int sched_rt_global_validate(void)
7658 if (sysctl_sched_rt_period <= 0)
7661 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7662 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7668 static void sched_rt_do_global(void)
7670 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7671 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7674 int sched_rt_handler(struct ctl_table *table, int write,
7675 void __user *buffer, size_t *lenp,
7678 int old_period, old_runtime;
7679 static DEFINE_MUTEX(mutex);
7683 old_period = sysctl_sched_rt_period;
7684 old_runtime = sysctl_sched_rt_runtime;
7686 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7688 if (!ret && write) {
7689 ret = sched_rt_global_validate();
7693 ret = sched_rt_global_constraints();
7697 ret = sched_dl_global_constraints();
7701 sched_rt_do_global();
7702 sched_dl_do_global();
7706 sysctl_sched_rt_period = old_period;
7707 sysctl_sched_rt_runtime = old_runtime;
7709 mutex_unlock(&mutex);
7714 int sched_rr_handler(struct ctl_table *table, int write,
7715 void __user *buffer, size_t *lenp,
7719 static DEFINE_MUTEX(mutex);
7722 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7723 /* make sure that internally we keep jiffies */
7724 /* also, writing zero resets timeslice to default */
7725 if (!ret && write) {
7726 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7727 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7729 mutex_unlock(&mutex);
7733 #ifdef CONFIG_CGROUP_SCHED
7735 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7737 return css ? container_of(css, struct task_group, css) : NULL;
7740 static struct cgroup_subsys_state *
7741 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7743 struct task_group *parent = css_tg(parent_css);
7744 struct task_group *tg;
7747 /* This is early initialization for the top cgroup */
7748 return &root_task_group.css;
7751 tg = sched_create_group(parent);
7753 return ERR_PTR(-ENOMEM);
7758 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7760 struct task_group *tg = css_tg(css);
7761 struct task_group *parent = css_tg(css->parent);
7764 sched_online_group(tg, parent);
7768 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7770 struct task_group *tg = css_tg(css);
7772 sched_destroy_group(tg);
7775 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7777 struct task_group *tg = css_tg(css);
7779 sched_offline_group(tg);
7782 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7783 struct cgroup_taskset *tset)
7785 struct task_struct *task;
7787 cgroup_taskset_for_each(task, tset) {
7788 #ifdef CONFIG_RT_GROUP_SCHED
7789 if (!sched_rt_can_attach(css_tg(css), task))
7792 /* We don't support RT-tasks being in separate groups */
7793 if (task->sched_class != &fair_sched_class)
7800 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7801 struct cgroup_taskset *tset)
7803 struct task_struct *task;
7805 cgroup_taskset_for_each(task, tset)
7806 sched_move_task(task);
7809 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7810 struct cgroup_subsys_state *old_css,
7811 struct task_struct *task)
7814 * cgroup_exit() is called in the copy_process() failure path.
7815 * Ignore this case since the task hasn't ran yet, this avoids
7816 * trying to poke a half freed task state from generic code.
7818 if (!(task->flags & PF_EXITING))
7821 sched_move_task(task);
7824 #ifdef CONFIG_FAIR_GROUP_SCHED
7825 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7826 struct cftype *cftype, u64 shareval)
7828 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7831 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7834 struct task_group *tg = css_tg(css);
7836 return (u64) scale_load_down(tg->shares);
7839 #ifdef CONFIG_CFS_BANDWIDTH
7840 static DEFINE_MUTEX(cfs_constraints_mutex);
7842 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7843 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7845 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7847 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7849 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7850 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7852 if (tg == &root_task_group)
7856 * Ensure we have at some amount of bandwidth every period. This is
7857 * to prevent reaching a state of large arrears when throttled via
7858 * entity_tick() resulting in prolonged exit starvation.
7860 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7864 * Likewise, bound things on the otherside by preventing insane quota
7865 * periods. This also allows us to normalize in computing quota
7868 if (period > max_cfs_quota_period)
7872 * Prevent race between setting of cfs_rq->runtime_enabled and
7873 * unthrottle_offline_cfs_rqs().
7876 mutex_lock(&cfs_constraints_mutex);
7877 ret = __cfs_schedulable(tg, period, quota);
7881 runtime_enabled = quota != RUNTIME_INF;
7882 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7884 * If we need to toggle cfs_bandwidth_used, off->on must occur
7885 * before making related changes, and on->off must occur afterwards
7887 if (runtime_enabled && !runtime_was_enabled)
7888 cfs_bandwidth_usage_inc();
7889 raw_spin_lock_irq(&cfs_b->lock);
7890 cfs_b->period = ns_to_ktime(period);
7891 cfs_b->quota = quota;
7893 __refill_cfs_bandwidth_runtime(cfs_b);
7894 /* restart the period timer (if active) to handle new period expiry */
7895 if (runtime_enabled && cfs_b->timer_active) {
7896 /* force a reprogram */
7897 __start_cfs_bandwidth(cfs_b, true);
7899 raw_spin_unlock_irq(&cfs_b->lock);
7901 for_each_online_cpu(i) {
7902 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7903 struct rq *rq = cfs_rq->rq;
7905 raw_spin_lock_irq(&rq->lock);
7906 cfs_rq->runtime_enabled = runtime_enabled;
7907 cfs_rq->runtime_remaining = 0;
7909 if (cfs_rq->throttled)
7910 unthrottle_cfs_rq(cfs_rq);
7911 raw_spin_unlock_irq(&rq->lock);
7913 if (runtime_was_enabled && !runtime_enabled)
7914 cfs_bandwidth_usage_dec();
7916 mutex_unlock(&cfs_constraints_mutex);
7922 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7926 period = ktime_to_ns(tg->cfs_bandwidth.period);
7927 if (cfs_quota_us < 0)
7928 quota = RUNTIME_INF;
7930 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7932 return tg_set_cfs_bandwidth(tg, period, quota);
7935 long tg_get_cfs_quota(struct task_group *tg)
7939 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7942 quota_us = tg->cfs_bandwidth.quota;
7943 do_div(quota_us, NSEC_PER_USEC);
7948 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7952 period = (u64)cfs_period_us * NSEC_PER_USEC;
7953 quota = tg->cfs_bandwidth.quota;
7955 return tg_set_cfs_bandwidth(tg, period, quota);
7958 long tg_get_cfs_period(struct task_group *tg)
7962 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7963 do_div(cfs_period_us, NSEC_PER_USEC);
7965 return cfs_period_us;
7968 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7971 return tg_get_cfs_quota(css_tg(css));
7974 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7975 struct cftype *cftype, s64 cfs_quota_us)
7977 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7980 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7983 return tg_get_cfs_period(css_tg(css));
7986 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7987 struct cftype *cftype, u64 cfs_period_us)
7989 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7992 struct cfs_schedulable_data {
7993 struct task_group *tg;
7998 * normalize group quota/period to be quota/max_period
7999 * note: units are usecs
8001 static u64 normalize_cfs_quota(struct task_group *tg,
8002 struct cfs_schedulable_data *d)
8010 period = tg_get_cfs_period(tg);
8011 quota = tg_get_cfs_quota(tg);
8014 /* note: these should typically be equivalent */
8015 if (quota == RUNTIME_INF || quota == -1)
8018 return to_ratio(period, quota);
8021 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8023 struct cfs_schedulable_data *d = data;
8024 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8025 s64 quota = 0, parent_quota = -1;
8028 quota = RUNTIME_INF;
8030 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8032 quota = normalize_cfs_quota(tg, d);
8033 parent_quota = parent_b->hierarchal_quota;
8036 * ensure max(child_quota) <= parent_quota, inherit when no
8039 if (quota == RUNTIME_INF)
8040 quota = parent_quota;
8041 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8044 cfs_b->hierarchal_quota = quota;
8049 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8052 struct cfs_schedulable_data data = {
8058 if (quota != RUNTIME_INF) {
8059 do_div(data.period, NSEC_PER_USEC);
8060 do_div(data.quota, NSEC_PER_USEC);
8064 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8070 static int cpu_stats_show(struct seq_file *sf, void *v)
8072 struct task_group *tg = css_tg(seq_css(sf));
8073 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8075 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8076 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8077 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8081 #endif /* CONFIG_CFS_BANDWIDTH */
8082 #endif /* CONFIG_FAIR_GROUP_SCHED */
8084 #ifdef CONFIG_RT_GROUP_SCHED
8085 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8086 struct cftype *cft, s64 val)
8088 return sched_group_set_rt_runtime(css_tg(css), val);
8091 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8094 return sched_group_rt_runtime(css_tg(css));
8097 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8098 struct cftype *cftype, u64 rt_period_us)
8100 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8103 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8106 return sched_group_rt_period(css_tg(css));
8108 #endif /* CONFIG_RT_GROUP_SCHED */
8110 static struct cftype cpu_files[] = {
8111 #ifdef CONFIG_FAIR_GROUP_SCHED
8114 .read_u64 = cpu_shares_read_u64,
8115 .write_u64 = cpu_shares_write_u64,
8118 #ifdef CONFIG_CFS_BANDWIDTH
8120 .name = "cfs_quota_us",
8121 .read_s64 = cpu_cfs_quota_read_s64,
8122 .write_s64 = cpu_cfs_quota_write_s64,
8125 .name = "cfs_period_us",
8126 .read_u64 = cpu_cfs_period_read_u64,
8127 .write_u64 = cpu_cfs_period_write_u64,
8131 .seq_show = cpu_stats_show,
8134 #ifdef CONFIG_RT_GROUP_SCHED
8136 .name = "rt_runtime_us",
8137 .read_s64 = cpu_rt_runtime_read,
8138 .write_s64 = cpu_rt_runtime_write,
8141 .name = "rt_period_us",
8142 .read_u64 = cpu_rt_period_read_uint,
8143 .write_u64 = cpu_rt_period_write_uint,
8149 struct cgroup_subsys cpu_cgrp_subsys = {
8150 .css_alloc = cpu_cgroup_css_alloc,
8151 .css_free = cpu_cgroup_css_free,
8152 .css_online = cpu_cgroup_css_online,
8153 .css_offline = cpu_cgroup_css_offline,
8154 .can_attach = cpu_cgroup_can_attach,
8155 .attach = cpu_cgroup_attach,
8156 .exit = cpu_cgroup_exit,
8157 .legacy_cftypes = cpu_files,
8161 #endif /* CONFIG_CGROUP_SCHED */
8163 void dump_cpu_task(int cpu)
8165 pr_info("Task dump for CPU %d:\n", cpu);
8166 sched_show_task(cpu_curr(cpu));