4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 #ifdef smp_mb__before_atomic
94 void __smp_mb__before_atomic(void)
96 smp_mb__before_atomic();
98 EXPORT_SYMBOL(__smp_mb__before_atomic);
101 #ifdef smp_mb__after_atomic
102 void __smp_mb__after_atomic(void)
104 smp_mb__after_atomic();
106 EXPORT_SYMBOL(__smp_mb__after_atomic);
109 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
112 ktime_t soft, hard, now;
115 if (hrtimer_active(period_timer))
118 now = hrtimer_cb_get_time(period_timer);
119 hrtimer_forward(period_timer, now, period);
121 soft = hrtimer_get_softexpires(period_timer);
122 hard = hrtimer_get_expires(period_timer);
123 delta = ktime_to_ns(ktime_sub(hard, soft));
124 __hrtimer_start_range_ns(period_timer, soft, delta,
125 HRTIMER_MODE_ABS_PINNED, 0);
129 DEFINE_MUTEX(sched_domains_mutex);
130 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
132 static void update_rq_clock_task(struct rq *rq, s64 delta);
134 void update_rq_clock(struct rq *rq)
138 if (rq->skip_clock_update > 0)
141 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
145 update_rq_clock_task(rq, delta);
149 * Debugging: various feature bits
152 #define SCHED_FEAT(name, enabled) \
153 (1UL << __SCHED_FEAT_##name) * enabled |
155 const_debug unsigned int sysctl_sched_features =
156 #include "features.h"
161 #ifdef CONFIG_SCHED_DEBUG
162 #define SCHED_FEAT(name, enabled) \
165 static const char * const sched_feat_names[] = {
166 #include "features.h"
171 static int sched_feat_show(struct seq_file *m, void *v)
175 for (i = 0; i < __SCHED_FEAT_NR; i++) {
176 if (!(sysctl_sched_features & (1UL << i)))
178 seq_printf(m, "%s ", sched_feat_names[i]);
185 #ifdef HAVE_JUMP_LABEL
187 #define jump_label_key__true STATIC_KEY_INIT_TRUE
188 #define jump_label_key__false STATIC_KEY_INIT_FALSE
190 #define SCHED_FEAT(name, enabled) \
191 jump_label_key__##enabled ,
193 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
194 #include "features.h"
199 static void sched_feat_disable(int i)
201 if (static_key_enabled(&sched_feat_keys[i]))
202 static_key_slow_dec(&sched_feat_keys[i]);
205 static void sched_feat_enable(int i)
207 if (!static_key_enabled(&sched_feat_keys[i]))
208 static_key_slow_inc(&sched_feat_keys[i]);
211 static void sched_feat_disable(int i) { };
212 static void sched_feat_enable(int i) { };
213 #endif /* HAVE_JUMP_LABEL */
215 static int sched_feat_set(char *cmp)
220 if (strncmp(cmp, "NO_", 3) == 0) {
225 for (i = 0; i < __SCHED_FEAT_NR; i++) {
226 if (strcmp(cmp, sched_feat_names[i]) == 0) {
228 sysctl_sched_features &= ~(1UL << i);
229 sched_feat_disable(i);
231 sysctl_sched_features |= (1UL << i);
232 sched_feat_enable(i);
242 sched_feat_write(struct file *filp, const char __user *ubuf,
243 size_t cnt, loff_t *ppos)
253 if (copy_from_user(&buf, ubuf, cnt))
259 /* Ensure the static_key remains in a consistent state */
260 inode = file_inode(filp);
261 mutex_lock(&inode->i_mutex);
262 i = sched_feat_set(cmp);
263 mutex_unlock(&inode->i_mutex);
264 if (i == __SCHED_FEAT_NR)
272 static int sched_feat_open(struct inode *inode, struct file *filp)
274 return single_open(filp, sched_feat_show, NULL);
277 static const struct file_operations sched_feat_fops = {
278 .open = sched_feat_open,
279 .write = sched_feat_write,
282 .release = single_release,
285 static __init int sched_init_debug(void)
287 debugfs_create_file("sched_features", 0644, NULL, NULL,
292 late_initcall(sched_init_debug);
293 #endif /* CONFIG_SCHED_DEBUG */
296 * Number of tasks to iterate in a single balance run.
297 * Limited because this is done with IRQs disabled.
299 const_debug unsigned int sysctl_sched_nr_migrate = 32;
302 * period over which we average the RT time consumption, measured
307 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
310 * period over which we measure -rt task cpu usage in us.
313 unsigned int sysctl_sched_rt_period = 1000000;
315 __read_mostly int scheduler_running;
318 * part of the period that we allow rt tasks to run in us.
321 int sysctl_sched_rt_runtime = 950000;
324 * __task_rq_lock - lock the rq @p resides on.
326 static inline struct rq *__task_rq_lock(struct task_struct *p)
331 lockdep_assert_held(&p->pi_lock);
335 raw_spin_lock(&rq->lock);
336 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
338 raw_spin_unlock(&rq->lock);
340 while (unlikely(task_on_rq_migrating(p)))
346 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
348 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
349 __acquires(p->pi_lock)
355 raw_spin_lock_irqsave(&p->pi_lock, *flags);
357 raw_spin_lock(&rq->lock);
358 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
360 raw_spin_unlock(&rq->lock);
361 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
363 while (unlikely(task_on_rq_migrating(p)))
368 static void __task_rq_unlock(struct rq *rq)
371 raw_spin_unlock(&rq->lock);
375 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
377 __releases(p->pi_lock)
379 raw_spin_unlock(&rq->lock);
380 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
384 * this_rq_lock - lock this runqueue and disable interrupts.
386 static struct rq *this_rq_lock(void)
393 raw_spin_lock(&rq->lock);
398 #ifdef CONFIG_SCHED_HRTICK
400 * Use HR-timers to deliver accurate preemption points.
403 static void hrtick_clear(struct rq *rq)
405 if (hrtimer_active(&rq->hrtick_timer))
406 hrtimer_cancel(&rq->hrtick_timer);
410 * High-resolution timer tick.
411 * Runs from hardirq context with interrupts disabled.
413 static enum hrtimer_restart hrtick(struct hrtimer *timer)
415 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
417 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
419 raw_spin_lock(&rq->lock);
421 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
422 raw_spin_unlock(&rq->lock);
424 return HRTIMER_NORESTART;
429 static int __hrtick_restart(struct rq *rq)
431 struct hrtimer *timer = &rq->hrtick_timer;
432 ktime_t time = hrtimer_get_softexpires(timer);
434 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
438 * called from hardirq (IPI) context
440 static void __hrtick_start(void *arg)
444 raw_spin_lock(&rq->lock);
445 __hrtick_restart(rq);
446 rq->hrtick_csd_pending = 0;
447 raw_spin_unlock(&rq->lock);
451 * Called to set the hrtick timer state.
453 * called with rq->lock held and irqs disabled
455 void hrtick_start(struct rq *rq, u64 delay)
457 struct hrtimer *timer = &rq->hrtick_timer;
462 * Don't schedule slices shorter than 10000ns, that just
463 * doesn't make sense and can cause timer DoS.
465 delta = max_t(s64, delay, 10000LL);
466 time = ktime_add_ns(timer->base->get_time(), delta);
468 hrtimer_set_expires(timer, time);
470 if (rq == this_rq()) {
471 __hrtick_restart(rq);
472 } else if (!rq->hrtick_csd_pending) {
473 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
474 rq->hrtick_csd_pending = 1;
479 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
481 int cpu = (int)(long)hcpu;
484 case CPU_UP_CANCELED:
485 case CPU_UP_CANCELED_FROZEN:
486 case CPU_DOWN_PREPARE:
487 case CPU_DOWN_PREPARE_FROZEN:
489 case CPU_DEAD_FROZEN:
490 hrtick_clear(cpu_rq(cpu));
497 static __init void init_hrtick(void)
499 hotcpu_notifier(hotplug_hrtick, 0);
503 * Called to set the hrtick timer state.
505 * called with rq->lock held and irqs disabled
507 void hrtick_start(struct rq *rq, u64 delay)
509 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
510 HRTIMER_MODE_REL_PINNED, 0);
513 static inline void init_hrtick(void)
516 #endif /* CONFIG_SMP */
518 static void init_rq_hrtick(struct rq *rq)
521 rq->hrtick_csd_pending = 0;
523 rq->hrtick_csd.flags = 0;
524 rq->hrtick_csd.func = __hrtick_start;
525 rq->hrtick_csd.info = rq;
528 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
529 rq->hrtick_timer.function = hrtick;
531 #else /* CONFIG_SCHED_HRTICK */
532 static inline void hrtick_clear(struct rq *rq)
536 static inline void init_rq_hrtick(struct rq *rq)
540 static inline void init_hrtick(void)
543 #endif /* CONFIG_SCHED_HRTICK */
546 * cmpxchg based fetch_or, macro so it works for different integer types
548 #define fetch_or(ptr, val) \
549 ({ typeof(*(ptr)) __old, __val = *(ptr); \
551 __old = cmpxchg((ptr), __val, __val | (val)); \
552 if (__old == __val) \
559 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
561 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
562 * this avoids any races wrt polling state changes and thereby avoids
565 static bool set_nr_and_not_polling(struct task_struct *p)
567 struct thread_info *ti = task_thread_info(p);
568 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
572 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
574 * If this returns true, then the idle task promises to call
575 * sched_ttwu_pending() and reschedule soon.
577 static bool set_nr_if_polling(struct task_struct *p)
579 struct thread_info *ti = task_thread_info(p);
580 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
583 if (!(val & _TIF_POLLING_NRFLAG))
585 if (val & _TIF_NEED_RESCHED)
587 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
596 static bool set_nr_and_not_polling(struct task_struct *p)
598 set_tsk_need_resched(p);
603 static bool set_nr_if_polling(struct task_struct *p)
611 * resched_curr - mark rq's current task 'to be rescheduled now'.
613 * On UP this means the setting of the need_resched flag, on SMP it
614 * might also involve a cross-CPU call to trigger the scheduler on
617 void resched_curr(struct rq *rq)
619 struct task_struct *curr = rq->curr;
622 lockdep_assert_held(&rq->lock);
624 if (test_tsk_need_resched(curr))
629 if (cpu == smp_processor_id()) {
630 set_tsk_need_resched(curr);
631 set_preempt_need_resched();
635 if (set_nr_and_not_polling(curr))
636 smp_send_reschedule(cpu);
638 trace_sched_wake_idle_without_ipi(cpu);
641 void resched_cpu(int cpu)
643 struct rq *rq = cpu_rq(cpu);
646 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
649 raw_spin_unlock_irqrestore(&rq->lock, flags);
653 #ifdef CONFIG_NO_HZ_COMMON
655 * In the semi idle case, use the nearest busy cpu for migrating timers
656 * from an idle cpu. This is good for power-savings.
658 * We don't do similar optimization for completely idle system, as
659 * selecting an idle cpu will add more delays to the timers than intended
660 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
662 int get_nohz_timer_target(int pinned)
664 int cpu = smp_processor_id();
666 struct sched_domain *sd;
668 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
672 for_each_domain(cpu, sd) {
673 for_each_cpu(i, sched_domain_span(sd)) {
685 * When add_timer_on() enqueues a timer into the timer wheel of an
686 * idle CPU then this timer might expire before the next timer event
687 * which is scheduled to wake up that CPU. In case of a completely
688 * idle system the next event might even be infinite time into the
689 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
690 * leaves the inner idle loop so the newly added timer is taken into
691 * account when the CPU goes back to idle and evaluates the timer
692 * wheel for the next timer event.
694 static void wake_up_idle_cpu(int cpu)
696 struct rq *rq = cpu_rq(cpu);
698 if (cpu == smp_processor_id())
701 if (set_nr_and_not_polling(rq->idle))
702 smp_send_reschedule(cpu);
704 trace_sched_wake_idle_without_ipi(cpu);
707 static bool wake_up_full_nohz_cpu(int cpu)
710 * We just need the target to call irq_exit() and re-evaluate
711 * the next tick. The nohz full kick at least implies that.
712 * If needed we can still optimize that later with an
715 if (tick_nohz_full_cpu(cpu)) {
716 if (cpu != smp_processor_id() ||
717 tick_nohz_tick_stopped())
718 tick_nohz_full_kick_cpu(cpu);
725 void wake_up_nohz_cpu(int cpu)
727 if (!wake_up_full_nohz_cpu(cpu))
728 wake_up_idle_cpu(cpu);
731 static inline bool got_nohz_idle_kick(void)
733 int cpu = smp_processor_id();
735 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
738 if (idle_cpu(cpu) && !need_resched())
742 * We can't run Idle Load Balance on this CPU for this time so we
743 * cancel it and clear NOHZ_BALANCE_KICK
745 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
749 #else /* CONFIG_NO_HZ_COMMON */
751 static inline bool got_nohz_idle_kick(void)
756 #endif /* CONFIG_NO_HZ_COMMON */
758 #ifdef CONFIG_NO_HZ_FULL
759 bool sched_can_stop_tick(void)
762 * More than one running task need preemption.
763 * nr_running update is assumed to be visible
764 * after IPI is sent from wakers.
766 if (this_rq()->nr_running > 1)
771 #endif /* CONFIG_NO_HZ_FULL */
773 void sched_avg_update(struct rq *rq)
775 s64 period = sched_avg_period();
777 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
779 * Inline assembly required to prevent the compiler
780 * optimising this loop into a divmod call.
781 * See __iter_div_u64_rem() for another example of this.
783 asm("" : "+rm" (rq->age_stamp));
784 rq->age_stamp += period;
789 #endif /* CONFIG_SMP */
791 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
792 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
794 * Iterate task_group tree rooted at *from, calling @down when first entering a
795 * node and @up when leaving it for the final time.
797 * Caller must hold rcu_lock or sufficient equivalent.
799 int walk_tg_tree_from(struct task_group *from,
800 tg_visitor down, tg_visitor up, void *data)
802 struct task_group *parent, *child;
808 ret = (*down)(parent, data);
811 list_for_each_entry_rcu(child, &parent->children, siblings) {
818 ret = (*up)(parent, data);
819 if (ret || parent == from)
823 parent = parent->parent;
830 int tg_nop(struct task_group *tg, void *data)
836 static void set_load_weight(struct task_struct *p)
838 int prio = p->static_prio - MAX_RT_PRIO;
839 struct load_weight *load = &p->se.load;
842 * SCHED_IDLE tasks get minimal weight:
844 if (p->policy == SCHED_IDLE) {
845 load->weight = scale_load(WEIGHT_IDLEPRIO);
846 load->inv_weight = WMULT_IDLEPRIO;
850 load->weight = scale_load(prio_to_weight[prio]);
851 load->inv_weight = prio_to_wmult[prio];
854 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
857 sched_info_queued(rq, p);
858 p->sched_class->enqueue_task(rq, p, flags);
861 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
864 sched_info_dequeued(rq, p);
865 p->sched_class->dequeue_task(rq, p, flags);
868 void activate_task(struct rq *rq, struct task_struct *p, int flags)
870 if (task_contributes_to_load(p))
871 rq->nr_uninterruptible--;
873 enqueue_task(rq, p, flags);
876 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
878 if (task_contributes_to_load(p))
879 rq->nr_uninterruptible++;
881 dequeue_task(rq, p, flags);
884 static void update_rq_clock_task(struct rq *rq, s64 delta)
887 * In theory, the compile should just see 0 here, and optimize out the call
888 * to sched_rt_avg_update. But I don't trust it...
890 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
891 s64 steal = 0, irq_delta = 0;
893 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
894 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
897 * Since irq_time is only updated on {soft,}irq_exit, we might run into
898 * this case when a previous update_rq_clock() happened inside a
901 * When this happens, we stop ->clock_task and only update the
902 * prev_irq_time stamp to account for the part that fit, so that a next
903 * update will consume the rest. This ensures ->clock_task is
906 * It does however cause some slight miss-attribution of {soft,}irq
907 * time, a more accurate solution would be to update the irq_time using
908 * the current rq->clock timestamp, except that would require using
911 if (irq_delta > delta)
914 rq->prev_irq_time += irq_delta;
917 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
918 if (static_key_false((¶virt_steal_rq_enabled))) {
919 steal = paravirt_steal_clock(cpu_of(rq));
920 steal -= rq->prev_steal_time_rq;
922 if (unlikely(steal > delta))
925 rq->prev_steal_time_rq += steal;
930 rq->clock_task += delta;
932 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
933 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
934 sched_rt_avg_update(rq, irq_delta + steal);
938 void sched_set_stop_task(int cpu, struct task_struct *stop)
940 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
941 struct task_struct *old_stop = cpu_rq(cpu)->stop;
945 * Make it appear like a SCHED_FIFO task, its something
946 * userspace knows about and won't get confused about.
948 * Also, it will make PI more or less work without too
949 * much confusion -- but then, stop work should not
950 * rely on PI working anyway.
952 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
954 stop->sched_class = &stop_sched_class;
957 cpu_rq(cpu)->stop = stop;
961 * Reset it back to a normal scheduling class so that
962 * it can die in pieces.
964 old_stop->sched_class = &rt_sched_class;
969 * __normal_prio - return the priority that is based on the static prio
971 static inline int __normal_prio(struct task_struct *p)
973 return p->static_prio;
977 * Calculate the expected normal priority: i.e. priority
978 * without taking RT-inheritance into account. Might be
979 * boosted by interactivity modifiers. Changes upon fork,
980 * setprio syscalls, and whenever the interactivity
981 * estimator recalculates.
983 static inline int normal_prio(struct task_struct *p)
987 if (task_has_dl_policy(p))
988 prio = MAX_DL_PRIO-1;
989 else if (task_has_rt_policy(p))
990 prio = MAX_RT_PRIO-1 - p->rt_priority;
992 prio = __normal_prio(p);
997 * Calculate the current priority, i.e. the priority
998 * taken into account by the scheduler. This value might
999 * be boosted by RT tasks, or might be boosted by
1000 * interactivity modifiers. Will be RT if the task got
1001 * RT-boosted. If not then it returns p->normal_prio.
1003 static int effective_prio(struct task_struct *p)
1005 p->normal_prio = normal_prio(p);
1007 * If we are RT tasks or we were boosted to RT priority,
1008 * keep the priority unchanged. Otherwise, update priority
1009 * to the normal priority:
1011 if (!rt_prio(p->prio))
1012 return p->normal_prio;
1017 * task_curr - is this task currently executing on a CPU?
1018 * @p: the task in question.
1020 * Return: 1 if the task is currently executing. 0 otherwise.
1022 inline int task_curr(const struct task_struct *p)
1024 return cpu_curr(task_cpu(p)) == p;
1027 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1028 const struct sched_class *prev_class,
1031 if (prev_class != p->sched_class) {
1032 if (prev_class->switched_from)
1033 prev_class->switched_from(rq, p);
1034 p->sched_class->switched_to(rq, p);
1035 } else if (oldprio != p->prio || dl_task(p))
1036 p->sched_class->prio_changed(rq, p, oldprio);
1039 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1041 const struct sched_class *class;
1043 if (p->sched_class == rq->curr->sched_class) {
1044 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1046 for_each_class(class) {
1047 if (class == rq->curr->sched_class)
1049 if (class == p->sched_class) {
1057 * A queue event has occurred, and we're going to schedule. In
1058 * this case, we can save a useless back to back clock update.
1060 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1061 rq->skip_clock_update = 1;
1065 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1067 #ifdef CONFIG_SCHED_DEBUG
1069 * We should never call set_task_cpu() on a blocked task,
1070 * ttwu() will sort out the placement.
1072 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1073 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1075 #ifdef CONFIG_LOCKDEP
1077 * The caller should hold either p->pi_lock or rq->lock, when changing
1078 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1080 * sched_move_task() holds both and thus holding either pins the cgroup,
1083 * Furthermore, all task_rq users should acquire both locks, see
1086 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1087 lockdep_is_held(&task_rq(p)->lock)));
1091 trace_sched_migrate_task(p, new_cpu);
1093 if (task_cpu(p) != new_cpu) {
1094 if (p->sched_class->migrate_task_rq)
1095 p->sched_class->migrate_task_rq(p, new_cpu);
1096 p->se.nr_migrations++;
1097 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1100 __set_task_cpu(p, new_cpu);
1103 static void __migrate_swap_task(struct task_struct *p, int cpu)
1105 if (task_on_rq_queued(p)) {
1106 struct rq *src_rq, *dst_rq;
1108 src_rq = task_rq(p);
1109 dst_rq = cpu_rq(cpu);
1111 deactivate_task(src_rq, p, 0);
1112 set_task_cpu(p, cpu);
1113 activate_task(dst_rq, p, 0);
1114 check_preempt_curr(dst_rq, p, 0);
1117 * Task isn't running anymore; make it appear like we migrated
1118 * it before it went to sleep. This means on wakeup we make the
1119 * previous cpu our targer instead of where it really is.
1125 struct migration_swap_arg {
1126 struct task_struct *src_task, *dst_task;
1127 int src_cpu, dst_cpu;
1130 static int migrate_swap_stop(void *data)
1132 struct migration_swap_arg *arg = data;
1133 struct rq *src_rq, *dst_rq;
1136 src_rq = cpu_rq(arg->src_cpu);
1137 dst_rq = cpu_rq(arg->dst_cpu);
1139 double_raw_lock(&arg->src_task->pi_lock,
1140 &arg->dst_task->pi_lock);
1141 double_rq_lock(src_rq, dst_rq);
1142 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1145 if (task_cpu(arg->src_task) != arg->src_cpu)
1148 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1151 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1154 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1155 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1160 double_rq_unlock(src_rq, dst_rq);
1161 raw_spin_unlock(&arg->dst_task->pi_lock);
1162 raw_spin_unlock(&arg->src_task->pi_lock);
1168 * Cross migrate two tasks
1170 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1172 struct migration_swap_arg arg;
1175 arg = (struct migration_swap_arg){
1177 .src_cpu = task_cpu(cur),
1179 .dst_cpu = task_cpu(p),
1182 if (arg.src_cpu == arg.dst_cpu)
1186 * These three tests are all lockless; this is OK since all of them
1187 * will be re-checked with proper locks held further down the line.
1189 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1192 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1195 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1198 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1199 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1205 struct migration_arg {
1206 struct task_struct *task;
1210 static int migration_cpu_stop(void *data);
1213 * wait_task_inactive - wait for a thread to unschedule.
1215 * If @match_state is nonzero, it's the @p->state value just checked and
1216 * not expected to change. If it changes, i.e. @p might have woken up,
1217 * then return zero. When we succeed in waiting for @p to be off its CPU,
1218 * we return a positive number (its total switch count). If a second call
1219 * a short while later returns the same number, the caller can be sure that
1220 * @p has remained unscheduled the whole time.
1222 * The caller must ensure that the task *will* unschedule sometime soon,
1223 * else this function might spin for a *long* time. This function can't
1224 * be called with interrupts off, or it may introduce deadlock with
1225 * smp_call_function() if an IPI is sent by the same process we are
1226 * waiting to become inactive.
1228 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1230 unsigned long flags;
1231 int running, queued;
1237 * We do the initial early heuristics without holding
1238 * any task-queue locks at all. We'll only try to get
1239 * the runqueue lock when things look like they will
1245 * If the task is actively running on another CPU
1246 * still, just relax and busy-wait without holding
1249 * NOTE! Since we don't hold any locks, it's not
1250 * even sure that "rq" stays as the right runqueue!
1251 * But we don't care, since "task_running()" will
1252 * return false if the runqueue has changed and p
1253 * is actually now running somewhere else!
1255 while (task_running(rq, p)) {
1256 if (match_state && unlikely(p->state != match_state))
1262 * Ok, time to look more closely! We need the rq
1263 * lock now, to be *sure*. If we're wrong, we'll
1264 * just go back and repeat.
1266 rq = task_rq_lock(p, &flags);
1267 trace_sched_wait_task(p);
1268 running = task_running(rq, p);
1269 queued = task_on_rq_queued(p);
1271 if (!match_state || p->state == match_state)
1272 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1273 task_rq_unlock(rq, p, &flags);
1276 * If it changed from the expected state, bail out now.
1278 if (unlikely(!ncsw))
1282 * Was it really running after all now that we
1283 * checked with the proper locks actually held?
1285 * Oops. Go back and try again..
1287 if (unlikely(running)) {
1293 * It's not enough that it's not actively running,
1294 * it must be off the runqueue _entirely_, and not
1297 * So if it was still runnable (but just not actively
1298 * running right now), it's preempted, and we should
1299 * yield - it could be a while.
1301 if (unlikely(queued)) {
1302 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1304 set_current_state(TASK_UNINTERRUPTIBLE);
1305 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1310 * Ahh, all good. It wasn't running, and it wasn't
1311 * runnable, which means that it will never become
1312 * running in the future either. We're all done!
1321 * kick_process - kick a running thread to enter/exit the kernel
1322 * @p: the to-be-kicked thread
1324 * Cause a process which is running on another CPU to enter
1325 * kernel-mode, without any delay. (to get signals handled.)
1327 * NOTE: this function doesn't have to take the runqueue lock,
1328 * because all it wants to ensure is that the remote task enters
1329 * the kernel. If the IPI races and the task has been migrated
1330 * to another CPU then no harm is done and the purpose has been
1333 void kick_process(struct task_struct *p)
1339 if ((cpu != smp_processor_id()) && task_curr(p))
1340 smp_send_reschedule(cpu);
1343 EXPORT_SYMBOL_GPL(kick_process);
1344 #endif /* CONFIG_SMP */
1348 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1350 static int select_fallback_rq(int cpu, struct task_struct *p)
1352 int nid = cpu_to_node(cpu);
1353 const struct cpumask *nodemask = NULL;
1354 enum { cpuset, possible, fail } state = cpuset;
1358 * If the node that the cpu is on has been offlined, cpu_to_node()
1359 * will return -1. There is no cpu on the node, and we should
1360 * select the cpu on the other node.
1363 nodemask = cpumask_of_node(nid);
1365 /* Look for allowed, online CPU in same node. */
1366 for_each_cpu(dest_cpu, nodemask) {
1367 if (!cpu_online(dest_cpu))
1369 if (!cpu_active(dest_cpu))
1371 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1377 /* Any allowed, online CPU? */
1378 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1379 if (!cpu_online(dest_cpu))
1381 if (!cpu_active(dest_cpu))
1388 /* No more Mr. Nice Guy. */
1389 cpuset_cpus_allowed_fallback(p);
1394 do_set_cpus_allowed(p, cpu_possible_mask);
1405 if (state != cpuset) {
1407 * Don't tell them about moving exiting tasks or
1408 * kernel threads (both mm NULL), since they never
1411 if (p->mm && printk_ratelimit()) {
1412 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1413 task_pid_nr(p), p->comm, cpu);
1421 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1424 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1426 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1429 * In order not to call set_task_cpu() on a blocking task we need
1430 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1433 * Since this is common to all placement strategies, this lives here.
1435 * [ this allows ->select_task() to simply return task_cpu(p) and
1436 * not worry about this generic constraint ]
1438 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1440 cpu = select_fallback_rq(task_cpu(p), p);
1445 static void update_avg(u64 *avg, u64 sample)
1447 s64 diff = sample - *avg;
1453 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1455 #ifdef CONFIG_SCHEDSTATS
1456 struct rq *rq = this_rq();
1459 int this_cpu = smp_processor_id();
1461 if (cpu == this_cpu) {
1462 schedstat_inc(rq, ttwu_local);
1463 schedstat_inc(p, se.statistics.nr_wakeups_local);
1465 struct sched_domain *sd;
1467 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1469 for_each_domain(this_cpu, sd) {
1470 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1471 schedstat_inc(sd, ttwu_wake_remote);
1478 if (wake_flags & WF_MIGRATED)
1479 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1481 #endif /* CONFIG_SMP */
1483 schedstat_inc(rq, ttwu_count);
1484 schedstat_inc(p, se.statistics.nr_wakeups);
1486 if (wake_flags & WF_SYNC)
1487 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1489 #endif /* CONFIG_SCHEDSTATS */
1492 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1494 activate_task(rq, p, en_flags);
1495 p->on_rq = TASK_ON_RQ_QUEUED;
1497 /* if a worker is waking up, notify workqueue */
1498 if (p->flags & PF_WQ_WORKER)
1499 wq_worker_waking_up(p, cpu_of(rq));
1503 * Mark the task runnable and perform wakeup-preemption.
1506 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1508 check_preempt_curr(rq, p, wake_flags);
1509 trace_sched_wakeup(p, true);
1511 p->state = TASK_RUNNING;
1513 if (p->sched_class->task_woken)
1514 p->sched_class->task_woken(rq, p);
1516 if (rq->idle_stamp) {
1517 u64 delta = rq_clock(rq) - rq->idle_stamp;
1518 u64 max = 2*rq->max_idle_balance_cost;
1520 update_avg(&rq->avg_idle, delta);
1522 if (rq->avg_idle > max)
1531 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1534 if (p->sched_contributes_to_load)
1535 rq->nr_uninterruptible--;
1538 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1539 ttwu_do_wakeup(rq, p, wake_flags);
1543 * Called in case the task @p isn't fully descheduled from its runqueue,
1544 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1545 * since all we need to do is flip p->state to TASK_RUNNING, since
1546 * the task is still ->on_rq.
1548 static int ttwu_remote(struct task_struct *p, int wake_flags)
1553 rq = __task_rq_lock(p);
1554 if (task_on_rq_queued(p)) {
1555 /* check_preempt_curr() may use rq clock */
1556 update_rq_clock(rq);
1557 ttwu_do_wakeup(rq, p, wake_flags);
1560 __task_rq_unlock(rq);
1566 void sched_ttwu_pending(void)
1568 struct rq *rq = this_rq();
1569 struct llist_node *llist = llist_del_all(&rq->wake_list);
1570 struct task_struct *p;
1571 unsigned long flags;
1576 raw_spin_lock_irqsave(&rq->lock, flags);
1579 p = llist_entry(llist, struct task_struct, wake_entry);
1580 llist = llist_next(llist);
1581 ttwu_do_activate(rq, p, 0);
1584 raw_spin_unlock_irqrestore(&rq->lock, flags);
1587 void scheduler_ipi(void)
1590 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1591 * TIF_NEED_RESCHED remotely (for the first time) will also send
1594 preempt_fold_need_resched();
1596 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1600 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1601 * traditionally all their work was done from the interrupt return
1602 * path. Now that we actually do some work, we need to make sure
1605 * Some archs already do call them, luckily irq_enter/exit nest
1608 * Arguably we should visit all archs and update all handlers,
1609 * however a fair share of IPIs are still resched only so this would
1610 * somewhat pessimize the simple resched case.
1613 sched_ttwu_pending();
1616 * Check if someone kicked us for doing the nohz idle load balance.
1618 if (unlikely(got_nohz_idle_kick())) {
1619 this_rq()->idle_balance = 1;
1620 raise_softirq_irqoff(SCHED_SOFTIRQ);
1625 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1627 struct rq *rq = cpu_rq(cpu);
1629 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1630 if (!set_nr_if_polling(rq->idle))
1631 smp_send_reschedule(cpu);
1633 trace_sched_wake_idle_without_ipi(cpu);
1637 void wake_up_if_idle(int cpu)
1639 struct rq *rq = cpu_rq(cpu);
1640 unsigned long flags;
1642 if (!is_idle_task(rq->curr))
1645 if (set_nr_if_polling(rq->idle)) {
1646 trace_sched_wake_idle_without_ipi(cpu);
1648 raw_spin_lock_irqsave(&rq->lock, flags);
1649 if (is_idle_task(rq->curr))
1650 smp_send_reschedule(cpu);
1651 /* Else cpu is not in idle, do nothing here */
1652 raw_spin_unlock_irqrestore(&rq->lock, flags);
1656 bool cpus_share_cache(int this_cpu, int that_cpu)
1658 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1660 #endif /* CONFIG_SMP */
1662 static void ttwu_queue(struct task_struct *p, int cpu)
1664 struct rq *rq = cpu_rq(cpu);
1666 #if defined(CONFIG_SMP)
1667 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1668 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1669 ttwu_queue_remote(p, cpu);
1674 raw_spin_lock(&rq->lock);
1675 ttwu_do_activate(rq, p, 0);
1676 raw_spin_unlock(&rq->lock);
1680 * try_to_wake_up - wake up a thread
1681 * @p: the thread to be awakened
1682 * @state: the mask of task states that can be woken
1683 * @wake_flags: wake modifier flags (WF_*)
1685 * Put it on the run-queue if it's not already there. The "current"
1686 * thread is always on the run-queue (except when the actual
1687 * re-schedule is in progress), and as such you're allowed to do
1688 * the simpler "current->state = TASK_RUNNING" to mark yourself
1689 * runnable without the overhead of this.
1691 * Return: %true if @p was woken up, %false if it was already running.
1692 * or @state didn't match @p's state.
1695 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1697 unsigned long flags;
1698 int cpu, success = 0;
1701 * If we are going to wake up a thread waiting for CONDITION we
1702 * need to ensure that CONDITION=1 done by the caller can not be
1703 * reordered with p->state check below. This pairs with mb() in
1704 * set_current_state() the waiting thread does.
1706 smp_mb__before_spinlock();
1707 raw_spin_lock_irqsave(&p->pi_lock, flags);
1708 if (!(p->state & state))
1711 success = 1; /* we're going to change ->state */
1714 if (p->on_rq && ttwu_remote(p, wake_flags))
1719 * If the owning (remote) cpu is still in the middle of schedule() with
1720 * this task as prev, wait until its done referencing the task.
1725 * Pairs with the smp_wmb() in finish_lock_switch().
1729 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1730 p->state = TASK_WAKING;
1732 if (p->sched_class->task_waking)
1733 p->sched_class->task_waking(p);
1735 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1736 if (task_cpu(p) != cpu) {
1737 wake_flags |= WF_MIGRATED;
1738 set_task_cpu(p, cpu);
1740 #endif /* CONFIG_SMP */
1744 ttwu_stat(p, cpu, wake_flags);
1746 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1752 * try_to_wake_up_local - try to wake up a local task with rq lock held
1753 * @p: the thread to be awakened
1755 * Put @p on the run-queue if it's not already there. The caller must
1756 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1759 static void try_to_wake_up_local(struct task_struct *p)
1761 struct rq *rq = task_rq(p);
1763 if (WARN_ON_ONCE(rq != this_rq()) ||
1764 WARN_ON_ONCE(p == current))
1767 lockdep_assert_held(&rq->lock);
1769 if (!raw_spin_trylock(&p->pi_lock)) {
1770 raw_spin_unlock(&rq->lock);
1771 raw_spin_lock(&p->pi_lock);
1772 raw_spin_lock(&rq->lock);
1775 if (!(p->state & TASK_NORMAL))
1778 if (!task_on_rq_queued(p))
1779 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1781 ttwu_do_wakeup(rq, p, 0);
1782 ttwu_stat(p, smp_processor_id(), 0);
1784 raw_spin_unlock(&p->pi_lock);
1788 * wake_up_process - Wake up a specific process
1789 * @p: The process to be woken up.
1791 * Attempt to wake up the nominated process and move it to the set of runnable
1794 * Return: 1 if the process was woken up, 0 if it was already running.
1796 * It may be assumed that this function implies a write memory barrier before
1797 * changing the task state if and only if any tasks are woken up.
1799 int wake_up_process(struct task_struct *p)
1801 WARN_ON(task_is_stopped_or_traced(p));
1802 return try_to_wake_up(p, TASK_NORMAL, 0);
1804 EXPORT_SYMBOL(wake_up_process);
1806 int wake_up_state(struct task_struct *p, unsigned int state)
1808 return try_to_wake_up(p, state, 0);
1812 * Perform scheduler related setup for a newly forked process p.
1813 * p is forked by current.
1815 * __sched_fork() is basic setup used by init_idle() too:
1817 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1822 p->se.exec_start = 0;
1823 p->se.sum_exec_runtime = 0;
1824 p->se.prev_sum_exec_runtime = 0;
1825 p->se.nr_migrations = 0;
1827 INIT_LIST_HEAD(&p->se.group_node);
1829 #ifdef CONFIG_SCHEDSTATS
1830 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1833 RB_CLEAR_NODE(&p->dl.rb_node);
1834 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1835 p->dl.dl_runtime = p->dl.runtime = 0;
1836 p->dl.dl_deadline = p->dl.deadline = 0;
1837 p->dl.dl_period = 0;
1840 INIT_LIST_HEAD(&p->rt.run_list);
1842 #ifdef CONFIG_PREEMPT_NOTIFIERS
1843 INIT_HLIST_HEAD(&p->preempt_notifiers);
1846 #ifdef CONFIG_NUMA_BALANCING
1847 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1848 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1849 p->mm->numa_scan_seq = 0;
1852 if (clone_flags & CLONE_VM)
1853 p->numa_preferred_nid = current->numa_preferred_nid;
1855 p->numa_preferred_nid = -1;
1857 p->node_stamp = 0ULL;
1858 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1859 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1860 p->numa_work.next = &p->numa_work;
1861 p->numa_faults_memory = NULL;
1862 p->numa_faults_buffer_memory = NULL;
1863 p->last_task_numa_placement = 0;
1864 p->last_sum_exec_runtime = 0;
1866 INIT_LIST_HEAD(&p->numa_entry);
1867 p->numa_group = NULL;
1868 #endif /* CONFIG_NUMA_BALANCING */
1871 #ifdef CONFIG_NUMA_BALANCING
1872 #ifdef CONFIG_SCHED_DEBUG
1873 void set_numabalancing_state(bool enabled)
1876 sched_feat_set("NUMA");
1878 sched_feat_set("NO_NUMA");
1881 __read_mostly bool numabalancing_enabled;
1883 void set_numabalancing_state(bool enabled)
1885 numabalancing_enabled = enabled;
1887 #endif /* CONFIG_SCHED_DEBUG */
1889 #ifdef CONFIG_PROC_SYSCTL
1890 int sysctl_numa_balancing(struct ctl_table *table, int write,
1891 void __user *buffer, size_t *lenp, loff_t *ppos)
1895 int state = numabalancing_enabled;
1897 if (write && !capable(CAP_SYS_ADMIN))
1902 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1906 set_numabalancing_state(state);
1913 * fork()/clone()-time setup:
1915 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1917 unsigned long flags;
1918 int cpu = get_cpu();
1920 __sched_fork(clone_flags, p);
1922 * We mark the process as running here. This guarantees that
1923 * nobody will actually run it, and a signal or other external
1924 * event cannot wake it up and insert it on the runqueue either.
1926 p->state = TASK_RUNNING;
1929 * Make sure we do not leak PI boosting priority to the child.
1931 p->prio = current->normal_prio;
1934 * Revert to default priority/policy on fork if requested.
1936 if (unlikely(p->sched_reset_on_fork)) {
1937 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1938 p->policy = SCHED_NORMAL;
1939 p->static_prio = NICE_TO_PRIO(0);
1941 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1942 p->static_prio = NICE_TO_PRIO(0);
1944 p->prio = p->normal_prio = __normal_prio(p);
1948 * We don't need the reset flag anymore after the fork. It has
1949 * fulfilled its duty:
1951 p->sched_reset_on_fork = 0;
1954 if (dl_prio(p->prio)) {
1957 } else if (rt_prio(p->prio)) {
1958 p->sched_class = &rt_sched_class;
1960 p->sched_class = &fair_sched_class;
1963 if (p->sched_class->task_fork)
1964 p->sched_class->task_fork(p);
1967 * The child is not yet in the pid-hash so no cgroup attach races,
1968 * and the cgroup is pinned to this child due to cgroup_fork()
1969 * is ran before sched_fork().
1971 * Silence PROVE_RCU.
1973 raw_spin_lock_irqsave(&p->pi_lock, flags);
1974 set_task_cpu(p, cpu);
1975 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1977 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1978 if (likely(sched_info_on()))
1979 memset(&p->sched_info, 0, sizeof(p->sched_info));
1981 #if defined(CONFIG_SMP)
1984 init_task_preempt_count(p);
1986 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1987 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1994 unsigned long to_ratio(u64 period, u64 runtime)
1996 if (runtime == RUNTIME_INF)
2000 * Doing this here saves a lot of checks in all
2001 * the calling paths, and returning zero seems
2002 * safe for them anyway.
2007 return div64_u64(runtime << 20, period);
2011 inline struct dl_bw *dl_bw_of(int i)
2013 return &cpu_rq(i)->rd->dl_bw;
2016 static inline int dl_bw_cpus(int i)
2018 struct root_domain *rd = cpu_rq(i)->rd;
2021 for_each_cpu_and(i, rd->span, cpu_active_mask)
2027 inline struct dl_bw *dl_bw_of(int i)
2029 return &cpu_rq(i)->dl.dl_bw;
2032 static inline int dl_bw_cpus(int i)
2039 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2041 dl_b->total_bw -= tsk_bw;
2045 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2047 dl_b->total_bw += tsk_bw;
2051 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2053 return dl_b->bw != -1 &&
2054 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2058 * We must be sure that accepting a new task (or allowing changing the
2059 * parameters of an existing one) is consistent with the bandwidth
2060 * constraints. If yes, this function also accordingly updates the currently
2061 * allocated bandwidth to reflect the new situation.
2063 * This function is called while holding p's rq->lock.
2065 static int dl_overflow(struct task_struct *p, int policy,
2066 const struct sched_attr *attr)
2069 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2070 u64 period = attr->sched_period ?: attr->sched_deadline;
2071 u64 runtime = attr->sched_runtime;
2072 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2075 if (new_bw == p->dl.dl_bw)
2079 * Either if a task, enters, leave, or stays -deadline but changes
2080 * its parameters, we may need to update accordingly the total
2081 * allocated bandwidth of the container.
2083 raw_spin_lock(&dl_b->lock);
2084 cpus = dl_bw_cpus(task_cpu(p));
2085 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2086 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2087 __dl_add(dl_b, new_bw);
2089 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2090 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2091 __dl_clear(dl_b, p->dl.dl_bw);
2092 __dl_add(dl_b, new_bw);
2094 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2095 __dl_clear(dl_b, p->dl.dl_bw);
2098 raw_spin_unlock(&dl_b->lock);
2103 extern void init_dl_bw(struct dl_bw *dl_b);
2106 * wake_up_new_task - wake up a newly created task for the first time.
2108 * This function will do some initial scheduler statistics housekeeping
2109 * that must be done for every newly created context, then puts the task
2110 * on the runqueue and wakes it.
2112 void wake_up_new_task(struct task_struct *p)
2114 unsigned long flags;
2117 raw_spin_lock_irqsave(&p->pi_lock, flags);
2120 * Fork balancing, do it here and not earlier because:
2121 * - cpus_allowed can change in the fork path
2122 * - any previously selected cpu might disappear through hotplug
2124 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2127 /* Initialize new task's runnable average */
2128 init_task_runnable_average(p);
2129 rq = __task_rq_lock(p);
2130 activate_task(rq, p, 0);
2131 p->on_rq = TASK_ON_RQ_QUEUED;
2132 trace_sched_wakeup_new(p, true);
2133 check_preempt_curr(rq, p, WF_FORK);
2135 if (p->sched_class->task_woken)
2136 p->sched_class->task_woken(rq, p);
2138 task_rq_unlock(rq, p, &flags);
2141 #ifdef CONFIG_PREEMPT_NOTIFIERS
2144 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2145 * @notifier: notifier struct to register
2147 void preempt_notifier_register(struct preempt_notifier *notifier)
2149 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2151 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2154 * preempt_notifier_unregister - no longer interested in preemption notifications
2155 * @notifier: notifier struct to unregister
2157 * This is safe to call from within a preemption notifier.
2159 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2161 hlist_del(¬ifier->link);
2163 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2165 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2167 struct preempt_notifier *notifier;
2169 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2170 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2174 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2175 struct task_struct *next)
2177 struct preempt_notifier *notifier;
2179 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2180 notifier->ops->sched_out(notifier, next);
2183 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2185 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2190 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2191 struct task_struct *next)
2195 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2198 * prepare_task_switch - prepare to switch tasks
2199 * @rq: the runqueue preparing to switch
2200 * @prev: the current task that is being switched out
2201 * @next: the task we are going to switch to.
2203 * This is called with the rq lock held and interrupts off. It must
2204 * be paired with a subsequent finish_task_switch after the context
2207 * prepare_task_switch sets up locking and calls architecture specific
2211 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2212 struct task_struct *next)
2214 trace_sched_switch(prev, next);
2215 sched_info_switch(rq, prev, next);
2216 perf_event_task_sched_out(prev, next);
2217 fire_sched_out_preempt_notifiers(prev, next);
2218 prepare_lock_switch(rq, next);
2219 prepare_arch_switch(next);
2223 * finish_task_switch - clean up after a task-switch
2224 * @rq: runqueue associated with task-switch
2225 * @prev: the thread we just switched away from.
2227 * finish_task_switch must be called after the context switch, paired
2228 * with a prepare_task_switch call before the context switch.
2229 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2230 * and do any other architecture-specific cleanup actions.
2232 * Note that we may have delayed dropping an mm in context_switch(). If
2233 * so, we finish that here outside of the runqueue lock. (Doing it
2234 * with the lock held can cause deadlocks; see schedule() for
2237 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2238 __releases(rq->lock)
2240 struct mm_struct *mm = rq->prev_mm;
2246 * A task struct has one reference for the use as "current".
2247 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2248 * schedule one last time. The schedule call will never return, and
2249 * the scheduled task must drop that reference.
2250 * The test for TASK_DEAD must occur while the runqueue locks are
2251 * still held, otherwise prev could be scheduled on another cpu, die
2252 * there before we look at prev->state, and then the reference would
2254 * Manfred Spraul <manfred@colorfullife.com>
2256 prev_state = prev->state;
2257 vtime_task_switch(prev);
2258 finish_arch_switch(prev);
2259 perf_event_task_sched_in(prev, current);
2260 finish_lock_switch(rq, prev);
2261 finish_arch_post_lock_switch();
2263 fire_sched_in_preempt_notifiers(current);
2266 if (unlikely(prev_state == TASK_DEAD)) {
2267 if (prev->sched_class->task_dead)
2268 prev->sched_class->task_dead(prev);
2271 * Remove function-return probe instances associated with this
2272 * task and put them back on the free list.
2274 kprobe_flush_task(prev);
2275 put_task_struct(prev);
2278 tick_nohz_task_switch(current);
2283 /* rq->lock is NOT held, but preemption is disabled */
2284 static inline void post_schedule(struct rq *rq)
2286 if (rq->post_schedule) {
2287 unsigned long flags;
2289 raw_spin_lock_irqsave(&rq->lock, flags);
2290 if (rq->curr->sched_class->post_schedule)
2291 rq->curr->sched_class->post_schedule(rq);
2292 raw_spin_unlock_irqrestore(&rq->lock, flags);
2294 rq->post_schedule = 0;
2300 static inline void post_schedule(struct rq *rq)
2307 * schedule_tail - first thing a freshly forked thread must call.
2308 * @prev: the thread we just switched away from.
2310 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2311 __releases(rq->lock)
2313 struct rq *rq = this_rq();
2315 finish_task_switch(rq, prev);
2318 * FIXME: do we need to worry about rq being invalidated by the
2323 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2324 /* In this case, finish_task_switch does not reenable preemption */
2327 if (current->set_child_tid)
2328 put_user(task_pid_vnr(current), current->set_child_tid);
2332 * context_switch - switch to the new MM and the new
2333 * thread's register state.
2336 context_switch(struct rq *rq, struct task_struct *prev,
2337 struct task_struct *next)
2339 struct mm_struct *mm, *oldmm;
2341 prepare_task_switch(rq, prev, next);
2344 oldmm = prev->active_mm;
2346 * For paravirt, this is coupled with an exit in switch_to to
2347 * combine the page table reload and the switch backend into
2350 arch_start_context_switch(prev);
2353 next->active_mm = oldmm;
2354 atomic_inc(&oldmm->mm_count);
2355 enter_lazy_tlb(oldmm, next);
2357 switch_mm(oldmm, mm, next);
2360 prev->active_mm = NULL;
2361 rq->prev_mm = oldmm;
2364 * Since the runqueue lock will be released by the next
2365 * task (which is an invalid locking op but in the case
2366 * of the scheduler it's an obvious special-case), so we
2367 * do an early lockdep release here:
2369 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2370 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2373 context_tracking_task_switch(prev, next);
2374 /* Here we just switch the register state and the stack. */
2375 switch_to(prev, next, prev);
2379 * this_rq must be evaluated again because prev may have moved
2380 * CPUs since it called schedule(), thus the 'rq' on its stack
2381 * frame will be invalid.
2383 finish_task_switch(this_rq(), prev);
2387 * nr_running and nr_context_switches:
2389 * externally visible scheduler statistics: current number of runnable
2390 * threads, total number of context switches performed since bootup.
2392 unsigned long nr_running(void)
2394 unsigned long i, sum = 0;
2396 for_each_online_cpu(i)
2397 sum += cpu_rq(i)->nr_running;
2402 unsigned long long nr_context_switches(void)
2405 unsigned long long sum = 0;
2407 for_each_possible_cpu(i)
2408 sum += cpu_rq(i)->nr_switches;
2413 unsigned long nr_iowait(void)
2415 unsigned long i, sum = 0;
2417 for_each_possible_cpu(i)
2418 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2423 unsigned long nr_iowait_cpu(int cpu)
2425 struct rq *this = cpu_rq(cpu);
2426 return atomic_read(&this->nr_iowait);
2429 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2431 struct rq *this = this_rq();
2432 *nr_waiters = atomic_read(&this->nr_iowait);
2433 *load = this->cpu_load[0];
2439 * sched_exec - execve() is a valuable balancing opportunity, because at
2440 * this point the task has the smallest effective memory and cache footprint.
2442 void sched_exec(void)
2444 struct task_struct *p = current;
2445 unsigned long flags;
2448 raw_spin_lock_irqsave(&p->pi_lock, flags);
2449 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2450 if (dest_cpu == smp_processor_id())
2453 if (likely(cpu_active(dest_cpu))) {
2454 struct migration_arg arg = { p, dest_cpu };
2456 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2457 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2461 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2466 DEFINE_PER_CPU(struct kernel_stat, kstat);
2467 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2469 EXPORT_PER_CPU_SYMBOL(kstat);
2470 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2473 * Return any ns on the sched_clock that have not yet been accounted in
2474 * @p in case that task is currently running.
2476 * Called with task_rq_lock() held on @rq.
2478 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2483 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2484 * project cycles that may never be accounted to this
2485 * thread, breaking clock_gettime().
2487 if (task_current(rq, p) && task_on_rq_queued(p)) {
2488 update_rq_clock(rq);
2489 ns = rq_clock_task(rq) - p->se.exec_start;
2497 unsigned long long task_delta_exec(struct task_struct *p)
2499 unsigned long flags;
2503 rq = task_rq_lock(p, &flags);
2504 ns = do_task_delta_exec(p, rq);
2505 task_rq_unlock(rq, p, &flags);
2511 * Return accounted runtime for the task.
2512 * In case the task is currently running, return the runtime plus current's
2513 * pending runtime that have not been accounted yet.
2515 unsigned long long task_sched_runtime(struct task_struct *p)
2517 unsigned long flags;
2521 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2523 * 64-bit doesn't need locks to atomically read a 64bit value.
2524 * So we have a optimization chance when the task's delta_exec is 0.
2525 * Reading ->on_cpu is racy, but this is ok.
2527 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2528 * If we race with it entering cpu, unaccounted time is 0. This is
2529 * indistinguishable from the read occurring a few cycles earlier.
2530 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2531 * been accounted, so we're correct here as well.
2533 if (!p->on_cpu || !task_on_rq_queued(p))
2534 return p->se.sum_exec_runtime;
2537 rq = task_rq_lock(p, &flags);
2538 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2539 task_rq_unlock(rq, p, &flags);
2545 * This function gets called by the timer code, with HZ frequency.
2546 * We call it with interrupts disabled.
2548 void scheduler_tick(void)
2550 int cpu = smp_processor_id();
2551 struct rq *rq = cpu_rq(cpu);
2552 struct task_struct *curr = rq->curr;
2556 raw_spin_lock(&rq->lock);
2557 update_rq_clock(rq);
2558 curr->sched_class->task_tick(rq, curr, 0);
2559 update_cpu_load_active(rq);
2560 raw_spin_unlock(&rq->lock);
2562 perf_event_task_tick();
2565 rq->idle_balance = idle_cpu(cpu);
2566 trigger_load_balance(rq);
2568 rq_last_tick_reset(rq);
2571 #ifdef CONFIG_NO_HZ_FULL
2573 * scheduler_tick_max_deferment
2575 * Keep at least one tick per second when a single
2576 * active task is running because the scheduler doesn't
2577 * yet completely support full dynticks environment.
2579 * This makes sure that uptime, CFS vruntime, load
2580 * balancing, etc... continue to move forward, even
2581 * with a very low granularity.
2583 * Return: Maximum deferment in nanoseconds.
2585 u64 scheduler_tick_max_deferment(void)
2587 struct rq *rq = this_rq();
2588 unsigned long next, now = ACCESS_ONCE(jiffies);
2590 next = rq->last_sched_tick + HZ;
2592 if (time_before_eq(next, now))
2595 return jiffies_to_nsecs(next - now);
2599 notrace unsigned long get_parent_ip(unsigned long addr)
2601 if (in_lock_functions(addr)) {
2602 addr = CALLER_ADDR2;
2603 if (in_lock_functions(addr))
2604 addr = CALLER_ADDR3;
2609 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2610 defined(CONFIG_PREEMPT_TRACER))
2612 void preempt_count_add(int val)
2614 #ifdef CONFIG_DEBUG_PREEMPT
2618 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2621 __preempt_count_add(val);
2622 #ifdef CONFIG_DEBUG_PREEMPT
2624 * Spinlock count overflowing soon?
2626 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2629 if (preempt_count() == val) {
2630 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2631 #ifdef CONFIG_DEBUG_PREEMPT
2632 current->preempt_disable_ip = ip;
2634 trace_preempt_off(CALLER_ADDR0, ip);
2637 EXPORT_SYMBOL(preempt_count_add);
2638 NOKPROBE_SYMBOL(preempt_count_add);
2640 void preempt_count_sub(int val)
2642 #ifdef CONFIG_DEBUG_PREEMPT
2646 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2649 * Is the spinlock portion underflowing?
2651 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2652 !(preempt_count() & PREEMPT_MASK)))
2656 if (preempt_count() == val)
2657 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2658 __preempt_count_sub(val);
2660 EXPORT_SYMBOL(preempt_count_sub);
2661 NOKPROBE_SYMBOL(preempt_count_sub);
2666 * Print scheduling while atomic bug:
2668 static noinline void __schedule_bug(struct task_struct *prev)
2670 if (oops_in_progress)
2673 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2674 prev->comm, prev->pid, preempt_count());
2676 debug_show_held_locks(prev);
2678 if (irqs_disabled())
2679 print_irqtrace_events(prev);
2680 #ifdef CONFIG_DEBUG_PREEMPT
2681 if (in_atomic_preempt_off()) {
2682 pr_err("Preemption disabled at:");
2683 print_ip_sym(current->preempt_disable_ip);
2688 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2692 * Various schedule()-time debugging checks and statistics:
2694 static inline void schedule_debug(struct task_struct *prev)
2697 * Test if we are atomic. Since do_exit() needs to call into
2698 * schedule() atomically, we ignore that path. Otherwise whine
2699 * if we are scheduling when we should not.
2701 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2702 __schedule_bug(prev);
2705 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2707 schedstat_inc(this_rq(), sched_count);
2711 * Pick up the highest-prio task:
2713 static inline struct task_struct *
2714 pick_next_task(struct rq *rq, struct task_struct *prev)
2716 const struct sched_class *class = &fair_sched_class;
2717 struct task_struct *p;
2720 * Optimization: we know that if all tasks are in
2721 * the fair class we can call that function directly:
2723 if (likely(prev->sched_class == class &&
2724 rq->nr_running == rq->cfs.h_nr_running)) {
2725 p = fair_sched_class.pick_next_task(rq, prev);
2726 if (unlikely(p == RETRY_TASK))
2729 /* assumes fair_sched_class->next == idle_sched_class */
2731 p = idle_sched_class.pick_next_task(rq, prev);
2737 for_each_class(class) {
2738 p = class->pick_next_task(rq, prev);
2740 if (unlikely(p == RETRY_TASK))
2746 BUG(); /* the idle class will always have a runnable task */
2750 * __schedule() is the main scheduler function.
2752 * The main means of driving the scheduler and thus entering this function are:
2754 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2756 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2757 * paths. For example, see arch/x86/entry_64.S.
2759 * To drive preemption between tasks, the scheduler sets the flag in timer
2760 * interrupt handler scheduler_tick().
2762 * 3. Wakeups don't really cause entry into schedule(). They add a
2763 * task to the run-queue and that's it.
2765 * Now, if the new task added to the run-queue preempts the current
2766 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2767 * called on the nearest possible occasion:
2769 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2771 * - in syscall or exception context, at the next outmost
2772 * preempt_enable(). (this might be as soon as the wake_up()'s
2775 * - in IRQ context, return from interrupt-handler to
2776 * preemptible context
2778 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2781 * - cond_resched() call
2782 * - explicit schedule() call
2783 * - return from syscall or exception to user-space
2784 * - return from interrupt-handler to user-space
2786 static void __sched __schedule(void)
2788 struct task_struct *prev, *next;
2789 unsigned long *switch_count;
2795 cpu = smp_processor_id();
2797 rcu_note_context_switch(cpu);
2800 schedule_debug(prev);
2802 if (sched_feat(HRTICK))
2806 * Make sure that signal_pending_state()->signal_pending() below
2807 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2808 * done by the caller to avoid the race with signal_wake_up().
2810 smp_mb__before_spinlock();
2811 raw_spin_lock_irq(&rq->lock);
2813 switch_count = &prev->nivcsw;
2814 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2815 if (unlikely(signal_pending_state(prev->state, prev))) {
2816 prev->state = TASK_RUNNING;
2818 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2822 * If a worker went to sleep, notify and ask workqueue
2823 * whether it wants to wake up a task to maintain
2826 if (prev->flags & PF_WQ_WORKER) {
2827 struct task_struct *to_wakeup;
2829 to_wakeup = wq_worker_sleeping(prev, cpu);
2831 try_to_wake_up_local(to_wakeup);
2834 switch_count = &prev->nvcsw;
2837 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0)
2838 update_rq_clock(rq);
2840 next = pick_next_task(rq, prev);
2841 clear_tsk_need_resched(prev);
2842 clear_preempt_need_resched();
2843 rq->skip_clock_update = 0;
2845 if (likely(prev != next)) {
2850 context_switch(rq, prev, next); /* unlocks the rq */
2852 * The context switch have flipped the stack from under us
2853 * and restored the local variables which were saved when
2854 * this task called schedule() in the past. prev == current
2855 * is still correct, but it can be moved to another cpu/rq.
2857 cpu = smp_processor_id();
2860 raw_spin_unlock_irq(&rq->lock);
2864 sched_preempt_enable_no_resched();
2869 static inline void sched_submit_work(struct task_struct *tsk)
2871 if (!tsk->state || tsk_is_pi_blocked(tsk))
2874 * If we are going to sleep and we have plugged IO queued,
2875 * make sure to submit it to avoid deadlocks.
2877 if (blk_needs_flush_plug(tsk))
2878 blk_schedule_flush_plug(tsk);
2881 asmlinkage __visible void __sched schedule(void)
2883 struct task_struct *tsk = current;
2885 sched_submit_work(tsk);
2888 EXPORT_SYMBOL(schedule);
2890 #ifdef CONFIG_CONTEXT_TRACKING
2891 asmlinkage __visible void __sched schedule_user(void)
2894 * If we come here after a random call to set_need_resched(),
2895 * or we have been woken up remotely but the IPI has not yet arrived,
2896 * we haven't yet exited the RCU idle mode. Do it here manually until
2897 * we find a better solution.
2906 * schedule_preempt_disabled - called with preemption disabled
2908 * Returns with preemption disabled. Note: preempt_count must be 1
2910 void __sched schedule_preempt_disabled(void)
2912 sched_preempt_enable_no_resched();
2917 #ifdef CONFIG_PREEMPT
2919 * this is the entry point to schedule() from in-kernel preemption
2920 * off of preempt_enable. Kernel preemptions off return from interrupt
2921 * occur there and call schedule directly.
2923 asmlinkage __visible void __sched notrace preempt_schedule(void)
2926 * If there is a non-zero preempt_count or interrupts are disabled,
2927 * we do not want to preempt the current task. Just return..
2929 if (likely(!preemptible()))
2933 __preempt_count_add(PREEMPT_ACTIVE);
2935 __preempt_count_sub(PREEMPT_ACTIVE);
2938 * Check again in case we missed a preemption opportunity
2939 * between schedule and now.
2942 } while (need_resched());
2944 NOKPROBE_SYMBOL(preempt_schedule);
2945 EXPORT_SYMBOL(preempt_schedule);
2946 #endif /* CONFIG_PREEMPT */
2949 * this is the entry point to schedule() from kernel preemption
2950 * off of irq context.
2951 * Note, that this is called and return with irqs disabled. This will
2952 * protect us against recursive calling from irq.
2954 asmlinkage __visible void __sched preempt_schedule_irq(void)
2956 enum ctx_state prev_state;
2958 /* Catch callers which need to be fixed */
2959 BUG_ON(preempt_count() || !irqs_disabled());
2961 prev_state = exception_enter();
2964 __preempt_count_add(PREEMPT_ACTIVE);
2967 local_irq_disable();
2968 __preempt_count_sub(PREEMPT_ACTIVE);
2971 * Check again in case we missed a preemption opportunity
2972 * between schedule and now.
2975 } while (need_resched());
2977 exception_exit(prev_state);
2980 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2983 return try_to_wake_up(curr->private, mode, wake_flags);
2985 EXPORT_SYMBOL(default_wake_function);
2987 #ifdef CONFIG_RT_MUTEXES
2990 * rt_mutex_setprio - set the current priority of a task
2992 * @prio: prio value (kernel-internal form)
2994 * This function changes the 'effective' priority of a task. It does
2995 * not touch ->normal_prio like __setscheduler().
2997 * Used by the rt_mutex code to implement priority inheritance
2998 * logic. Call site only calls if the priority of the task changed.
3000 void rt_mutex_setprio(struct task_struct *p, int prio)
3002 int oldprio, queued, running, enqueue_flag = 0;
3004 const struct sched_class *prev_class;
3006 BUG_ON(prio > MAX_PRIO);
3008 rq = __task_rq_lock(p);
3011 * Idle task boosting is a nono in general. There is one
3012 * exception, when PREEMPT_RT and NOHZ is active:
3014 * The idle task calls get_next_timer_interrupt() and holds
3015 * the timer wheel base->lock on the CPU and another CPU wants
3016 * to access the timer (probably to cancel it). We can safely
3017 * ignore the boosting request, as the idle CPU runs this code
3018 * with interrupts disabled and will complete the lock
3019 * protected section without being interrupted. So there is no
3020 * real need to boost.
3022 if (unlikely(p == rq->idle)) {
3023 WARN_ON(p != rq->curr);
3024 WARN_ON(p->pi_blocked_on);
3028 trace_sched_pi_setprio(p, prio);
3030 prev_class = p->sched_class;
3031 queued = task_on_rq_queued(p);
3032 running = task_current(rq, p);
3034 dequeue_task(rq, p, 0);
3036 put_prev_task(rq, p);
3039 * Boosting condition are:
3040 * 1. -rt task is running and holds mutex A
3041 * --> -dl task blocks on mutex A
3043 * 2. -dl task is running and holds mutex A
3044 * --> -dl task blocks on mutex A and could preempt the
3047 if (dl_prio(prio)) {
3048 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3049 if (!dl_prio(p->normal_prio) ||
3050 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3051 p->dl.dl_boosted = 1;
3052 p->dl.dl_throttled = 0;
3053 enqueue_flag = ENQUEUE_REPLENISH;
3055 p->dl.dl_boosted = 0;
3056 p->sched_class = &dl_sched_class;
3057 } else if (rt_prio(prio)) {
3058 if (dl_prio(oldprio))
3059 p->dl.dl_boosted = 0;
3061 enqueue_flag = ENQUEUE_HEAD;
3062 p->sched_class = &rt_sched_class;
3064 if (dl_prio(oldprio))
3065 p->dl.dl_boosted = 0;
3066 p->sched_class = &fair_sched_class;
3072 p->sched_class->set_curr_task(rq);
3074 enqueue_task(rq, p, enqueue_flag);
3076 check_class_changed(rq, p, prev_class, oldprio);
3078 __task_rq_unlock(rq);
3082 void set_user_nice(struct task_struct *p, long nice)
3084 int old_prio, delta, queued;
3085 unsigned long flags;
3088 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3091 * We have to be careful, if called from sys_setpriority(),
3092 * the task might be in the middle of scheduling on another CPU.
3094 rq = task_rq_lock(p, &flags);
3096 * The RT priorities are set via sched_setscheduler(), but we still
3097 * allow the 'normal' nice value to be set - but as expected
3098 * it wont have any effect on scheduling until the task is
3099 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3101 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3102 p->static_prio = NICE_TO_PRIO(nice);
3105 queued = task_on_rq_queued(p);
3107 dequeue_task(rq, p, 0);
3109 p->static_prio = NICE_TO_PRIO(nice);
3112 p->prio = effective_prio(p);
3113 delta = p->prio - old_prio;
3116 enqueue_task(rq, p, 0);
3118 * If the task increased its priority or is running and
3119 * lowered its priority, then reschedule its CPU:
3121 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3125 task_rq_unlock(rq, p, &flags);
3127 EXPORT_SYMBOL(set_user_nice);
3130 * can_nice - check if a task can reduce its nice value
3134 int can_nice(const struct task_struct *p, const int nice)
3136 /* convert nice value [19,-20] to rlimit style value [1,40] */
3137 int nice_rlim = nice_to_rlimit(nice);
3139 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3140 capable(CAP_SYS_NICE));
3143 #ifdef __ARCH_WANT_SYS_NICE
3146 * sys_nice - change the priority of the current process.
3147 * @increment: priority increment
3149 * sys_setpriority is a more generic, but much slower function that
3150 * does similar things.
3152 SYSCALL_DEFINE1(nice, int, increment)
3157 * Setpriority might change our priority at the same moment.
3158 * We don't have to worry. Conceptually one call occurs first
3159 * and we have a single winner.
3161 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3162 nice = task_nice(current) + increment;
3164 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3165 if (increment < 0 && !can_nice(current, nice))
3168 retval = security_task_setnice(current, nice);
3172 set_user_nice(current, nice);
3179 * task_prio - return the priority value of a given task.
3180 * @p: the task in question.
3182 * Return: The priority value as seen by users in /proc.
3183 * RT tasks are offset by -200. Normal tasks are centered
3184 * around 0, value goes from -16 to +15.
3186 int task_prio(const struct task_struct *p)
3188 return p->prio - MAX_RT_PRIO;
3192 * idle_cpu - is a given cpu idle currently?
3193 * @cpu: the processor in question.
3195 * Return: 1 if the CPU is currently idle. 0 otherwise.
3197 int idle_cpu(int cpu)
3199 struct rq *rq = cpu_rq(cpu);
3201 if (rq->curr != rq->idle)
3208 if (!llist_empty(&rq->wake_list))
3216 * idle_task - return the idle task for a given cpu.
3217 * @cpu: the processor in question.
3219 * Return: The idle task for the cpu @cpu.
3221 struct task_struct *idle_task(int cpu)
3223 return cpu_rq(cpu)->idle;
3227 * find_process_by_pid - find a process with a matching PID value.
3228 * @pid: the pid in question.
3230 * The task of @pid, if found. %NULL otherwise.
3232 static struct task_struct *find_process_by_pid(pid_t pid)
3234 return pid ? find_task_by_vpid(pid) : current;
3238 * This function initializes the sched_dl_entity of a newly becoming
3239 * SCHED_DEADLINE task.
3241 * Only the static values are considered here, the actual runtime and the
3242 * absolute deadline will be properly calculated when the task is enqueued
3243 * for the first time with its new policy.
3246 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3248 struct sched_dl_entity *dl_se = &p->dl;
3250 init_dl_task_timer(dl_se);
3251 dl_se->dl_runtime = attr->sched_runtime;
3252 dl_se->dl_deadline = attr->sched_deadline;
3253 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3254 dl_se->flags = attr->sched_flags;
3255 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3256 dl_se->dl_throttled = 0;
3258 dl_se->dl_yielded = 0;
3262 * sched_setparam() passes in -1 for its policy, to let the functions
3263 * it calls know not to change it.
3265 #define SETPARAM_POLICY -1
3267 static void __setscheduler_params(struct task_struct *p,
3268 const struct sched_attr *attr)
3270 int policy = attr->sched_policy;
3272 if (policy == SETPARAM_POLICY)
3277 if (dl_policy(policy))
3278 __setparam_dl(p, attr);
3279 else if (fair_policy(policy))
3280 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3283 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3284 * !rt_policy. Always setting this ensures that things like
3285 * getparam()/getattr() don't report silly values for !rt tasks.
3287 p->rt_priority = attr->sched_priority;
3288 p->normal_prio = normal_prio(p);
3292 /* Actually do priority change: must hold pi & rq lock. */
3293 static void __setscheduler(struct rq *rq, struct task_struct *p,
3294 const struct sched_attr *attr)
3296 __setscheduler_params(p, attr);
3299 * If we get here, there was no pi waiters boosting the
3300 * task. It is safe to use the normal prio.
3302 p->prio = normal_prio(p);
3304 if (dl_prio(p->prio))
3305 p->sched_class = &dl_sched_class;
3306 else if (rt_prio(p->prio))
3307 p->sched_class = &rt_sched_class;
3309 p->sched_class = &fair_sched_class;
3313 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3315 struct sched_dl_entity *dl_se = &p->dl;
3317 attr->sched_priority = p->rt_priority;
3318 attr->sched_runtime = dl_se->dl_runtime;
3319 attr->sched_deadline = dl_se->dl_deadline;
3320 attr->sched_period = dl_se->dl_period;
3321 attr->sched_flags = dl_se->flags;
3325 * This function validates the new parameters of a -deadline task.
3326 * We ask for the deadline not being zero, and greater or equal
3327 * than the runtime, as well as the period of being zero or
3328 * greater than deadline. Furthermore, we have to be sure that
3329 * user parameters are above the internal resolution of 1us (we
3330 * check sched_runtime only since it is always the smaller one) and
3331 * below 2^63 ns (we have to check both sched_deadline and
3332 * sched_period, as the latter can be zero).
3335 __checkparam_dl(const struct sched_attr *attr)
3338 if (attr->sched_deadline == 0)
3342 * Since we truncate DL_SCALE bits, make sure we're at least
3345 if (attr->sched_runtime < (1ULL << DL_SCALE))
3349 * Since we use the MSB for wrap-around and sign issues, make
3350 * sure it's not set (mind that period can be equal to zero).
3352 if (attr->sched_deadline & (1ULL << 63) ||
3353 attr->sched_period & (1ULL << 63))
3356 /* runtime <= deadline <= period (if period != 0) */
3357 if ((attr->sched_period != 0 &&
3358 attr->sched_period < attr->sched_deadline) ||
3359 attr->sched_deadline < attr->sched_runtime)
3366 * check the target process has a UID that matches the current process's
3368 static bool check_same_owner(struct task_struct *p)
3370 const struct cred *cred = current_cred(), *pcred;
3374 pcred = __task_cred(p);
3375 match = (uid_eq(cred->euid, pcred->euid) ||
3376 uid_eq(cred->euid, pcred->uid));
3381 static int __sched_setscheduler(struct task_struct *p,
3382 const struct sched_attr *attr,
3385 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3386 MAX_RT_PRIO - 1 - attr->sched_priority;
3387 int retval, oldprio, oldpolicy = -1, queued, running;
3388 int policy = attr->sched_policy;
3389 unsigned long flags;
3390 const struct sched_class *prev_class;
3394 /* may grab non-irq protected spin_locks */
3395 BUG_ON(in_interrupt());
3397 /* double check policy once rq lock held */
3399 reset_on_fork = p->sched_reset_on_fork;
3400 policy = oldpolicy = p->policy;
3402 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3404 if (policy != SCHED_DEADLINE &&
3405 policy != SCHED_FIFO && policy != SCHED_RR &&
3406 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3407 policy != SCHED_IDLE)
3411 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3415 * Valid priorities for SCHED_FIFO and SCHED_RR are
3416 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3417 * SCHED_BATCH and SCHED_IDLE is 0.
3419 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3420 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3422 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3423 (rt_policy(policy) != (attr->sched_priority != 0)))
3427 * Allow unprivileged RT tasks to decrease priority:
3429 if (user && !capable(CAP_SYS_NICE)) {
3430 if (fair_policy(policy)) {
3431 if (attr->sched_nice < task_nice(p) &&
3432 !can_nice(p, attr->sched_nice))
3436 if (rt_policy(policy)) {
3437 unsigned long rlim_rtprio =
3438 task_rlimit(p, RLIMIT_RTPRIO);
3440 /* can't set/change the rt policy */
3441 if (policy != p->policy && !rlim_rtprio)
3444 /* can't increase priority */
3445 if (attr->sched_priority > p->rt_priority &&
3446 attr->sched_priority > rlim_rtprio)
3451 * Can't set/change SCHED_DEADLINE policy at all for now
3452 * (safest behavior); in the future we would like to allow
3453 * unprivileged DL tasks to increase their relative deadline
3454 * or reduce their runtime (both ways reducing utilization)
3456 if (dl_policy(policy))
3460 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3461 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3463 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3464 if (!can_nice(p, task_nice(p)))
3468 /* can't change other user's priorities */
3469 if (!check_same_owner(p))
3472 /* Normal users shall not reset the sched_reset_on_fork flag */
3473 if (p->sched_reset_on_fork && !reset_on_fork)
3478 retval = security_task_setscheduler(p);
3484 * make sure no PI-waiters arrive (or leave) while we are
3485 * changing the priority of the task:
3487 * To be able to change p->policy safely, the appropriate
3488 * runqueue lock must be held.
3490 rq = task_rq_lock(p, &flags);
3493 * Changing the policy of the stop threads its a very bad idea
3495 if (p == rq->stop) {
3496 task_rq_unlock(rq, p, &flags);
3501 * If not changing anything there's no need to proceed further,
3502 * but store a possible modification of reset_on_fork.
3504 if (unlikely(policy == p->policy)) {
3505 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3507 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3509 if (dl_policy(policy))
3512 p->sched_reset_on_fork = reset_on_fork;
3513 task_rq_unlock(rq, p, &flags);
3519 #ifdef CONFIG_RT_GROUP_SCHED
3521 * Do not allow realtime tasks into groups that have no runtime
3524 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3525 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3526 !task_group_is_autogroup(task_group(p))) {
3527 task_rq_unlock(rq, p, &flags);
3532 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3533 cpumask_t *span = rq->rd->span;
3536 * Don't allow tasks with an affinity mask smaller than
3537 * the entire root_domain to become SCHED_DEADLINE. We
3538 * will also fail if there's no bandwidth available.
3540 if (!cpumask_subset(span, &p->cpus_allowed) ||
3541 rq->rd->dl_bw.bw == 0) {
3542 task_rq_unlock(rq, p, &flags);
3549 /* recheck policy now with rq lock held */
3550 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3551 policy = oldpolicy = -1;
3552 task_rq_unlock(rq, p, &flags);
3557 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3558 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3561 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3562 task_rq_unlock(rq, p, &flags);
3566 p->sched_reset_on_fork = reset_on_fork;
3570 * Special case for priority boosted tasks.
3572 * If the new priority is lower or equal (user space view)
3573 * than the current (boosted) priority, we just store the new
3574 * normal parameters and do not touch the scheduler class and
3575 * the runqueue. This will be done when the task deboost
3578 if (rt_mutex_check_prio(p, newprio)) {
3579 __setscheduler_params(p, attr);
3580 task_rq_unlock(rq, p, &flags);
3584 queued = task_on_rq_queued(p);
3585 running = task_current(rq, p);
3587 dequeue_task(rq, p, 0);
3589 put_prev_task(rq, p);
3591 prev_class = p->sched_class;
3592 __setscheduler(rq, p, attr);
3595 p->sched_class->set_curr_task(rq);
3598 * We enqueue to tail when the priority of a task is
3599 * increased (user space view).
3601 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3604 check_class_changed(rq, p, prev_class, oldprio);
3605 task_rq_unlock(rq, p, &flags);
3607 rt_mutex_adjust_pi(p);
3612 static int _sched_setscheduler(struct task_struct *p, int policy,
3613 const struct sched_param *param, bool check)
3615 struct sched_attr attr = {
3616 .sched_policy = policy,
3617 .sched_priority = param->sched_priority,
3618 .sched_nice = PRIO_TO_NICE(p->static_prio),
3621 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3622 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3623 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3624 policy &= ~SCHED_RESET_ON_FORK;
3625 attr.sched_policy = policy;
3628 return __sched_setscheduler(p, &attr, check);
3631 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3632 * @p: the task in question.
3633 * @policy: new policy.
3634 * @param: structure containing the new RT priority.
3636 * Return: 0 on success. An error code otherwise.
3638 * NOTE that the task may be already dead.
3640 int sched_setscheduler(struct task_struct *p, int policy,
3641 const struct sched_param *param)
3643 return _sched_setscheduler(p, policy, param, true);
3645 EXPORT_SYMBOL_GPL(sched_setscheduler);
3647 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3649 return __sched_setscheduler(p, attr, true);
3651 EXPORT_SYMBOL_GPL(sched_setattr);
3654 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3655 * @p: the task in question.
3656 * @policy: new policy.
3657 * @param: structure containing the new RT priority.
3659 * Just like sched_setscheduler, only don't bother checking if the
3660 * current context has permission. For example, this is needed in
3661 * stop_machine(): we create temporary high priority worker threads,
3662 * but our caller might not have that capability.
3664 * Return: 0 on success. An error code otherwise.
3666 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3667 const struct sched_param *param)
3669 return _sched_setscheduler(p, policy, param, false);
3673 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3675 struct sched_param lparam;
3676 struct task_struct *p;
3679 if (!param || pid < 0)
3681 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3686 p = find_process_by_pid(pid);
3688 retval = sched_setscheduler(p, policy, &lparam);
3695 * Mimics kernel/events/core.c perf_copy_attr().
3697 static int sched_copy_attr(struct sched_attr __user *uattr,
3698 struct sched_attr *attr)
3703 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3707 * zero the full structure, so that a short copy will be nice.
3709 memset(attr, 0, sizeof(*attr));
3711 ret = get_user(size, &uattr->size);
3715 if (size > PAGE_SIZE) /* silly large */
3718 if (!size) /* abi compat */
3719 size = SCHED_ATTR_SIZE_VER0;
3721 if (size < SCHED_ATTR_SIZE_VER0)
3725 * If we're handed a bigger struct than we know of,
3726 * ensure all the unknown bits are 0 - i.e. new
3727 * user-space does not rely on any kernel feature
3728 * extensions we dont know about yet.
3730 if (size > sizeof(*attr)) {
3731 unsigned char __user *addr;
3732 unsigned char __user *end;
3735 addr = (void __user *)uattr + sizeof(*attr);
3736 end = (void __user *)uattr + size;
3738 for (; addr < end; addr++) {
3739 ret = get_user(val, addr);
3745 size = sizeof(*attr);
3748 ret = copy_from_user(attr, uattr, size);
3753 * XXX: do we want to be lenient like existing syscalls; or do we want
3754 * to be strict and return an error on out-of-bounds values?
3756 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3761 put_user(sizeof(*attr), &uattr->size);
3766 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3767 * @pid: the pid in question.
3768 * @policy: new policy.
3769 * @param: structure containing the new RT priority.
3771 * Return: 0 on success. An error code otherwise.
3773 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3774 struct sched_param __user *, param)
3776 /* negative values for policy are not valid */
3780 return do_sched_setscheduler(pid, policy, param);
3784 * sys_sched_setparam - set/change the RT priority of a thread
3785 * @pid: the pid in question.
3786 * @param: structure containing the new RT priority.
3788 * Return: 0 on success. An error code otherwise.
3790 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3792 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3796 * sys_sched_setattr - same as above, but with extended sched_attr
3797 * @pid: the pid in question.
3798 * @uattr: structure containing the extended parameters.
3799 * @flags: for future extension.
3801 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3802 unsigned int, flags)
3804 struct sched_attr attr;
3805 struct task_struct *p;
3808 if (!uattr || pid < 0 || flags)
3811 retval = sched_copy_attr(uattr, &attr);
3815 if ((int)attr.sched_policy < 0)
3820 p = find_process_by_pid(pid);
3822 retval = sched_setattr(p, &attr);
3829 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3830 * @pid: the pid in question.
3832 * Return: On success, the policy of the thread. Otherwise, a negative error
3835 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3837 struct task_struct *p;
3845 p = find_process_by_pid(pid);
3847 retval = security_task_getscheduler(p);
3850 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3857 * sys_sched_getparam - get the RT priority of a thread
3858 * @pid: the pid in question.
3859 * @param: structure containing the RT priority.
3861 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3864 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3866 struct sched_param lp = { .sched_priority = 0 };
3867 struct task_struct *p;
3870 if (!param || pid < 0)
3874 p = find_process_by_pid(pid);
3879 retval = security_task_getscheduler(p);
3883 if (task_has_rt_policy(p))
3884 lp.sched_priority = p->rt_priority;
3888 * This one might sleep, we cannot do it with a spinlock held ...
3890 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3899 static int sched_read_attr(struct sched_attr __user *uattr,
3900 struct sched_attr *attr,
3905 if (!access_ok(VERIFY_WRITE, uattr, usize))
3909 * If we're handed a smaller struct than we know of,
3910 * ensure all the unknown bits are 0 - i.e. old
3911 * user-space does not get uncomplete information.
3913 if (usize < sizeof(*attr)) {
3914 unsigned char *addr;
3917 addr = (void *)attr + usize;
3918 end = (void *)attr + sizeof(*attr);
3920 for (; addr < end; addr++) {
3928 ret = copy_to_user(uattr, attr, attr->size);
3936 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3937 * @pid: the pid in question.
3938 * @uattr: structure containing the extended parameters.
3939 * @size: sizeof(attr) for fwd/bwd comp.
3940 * @flags: for future extension.
3942 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3943 unsigned int, size, unsigned int, flags)
3945 struct sched_attr attr = {
3946 .size = sizeof(struct sched_attr),
3948 struct task_struct *p;
3951 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3952 size < SCHED_ATTR_SIZE_VER0 || flags)
3956 p = find_process_by_pid(pid);
3961 retval = security_task_getscheduler(p);
3965 attr.sched_policy = p->policy;
3966 if (p->sched_reset_on_fork)
3967 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3968 if (task_has_dl_policy(p))
3969 __getparam_dl(p, &attr);
3970 else if (task_has_rt_policy(p))
3971 attr.sched_priority = p->rt_priority;
3973 attr.sched_nice = task_nice(p);
3977 retval = sched_read_attr(uattr, &attr, size);
3985 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3987 cpumask_var_t cpus_allowed, new_mask;
3988 struct task_struct *p;
3993 p = find_process_by_pid(pid);
3999 /* Prevent p going away */
4003 if (p->flags & PF_NO_SETAFFINITY) {
4007 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4011 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4013 goto out_free_cpus_allowed;
4016 if (!check_same_owner(p)) {
4018 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4025 retval = security_task_setscheduler(p);
4030 cpuset_cpus_allowed(p, cpus_allowed);
4031 cpumask_and(new_mask, in_mask, cpus_allowed);
4034 * Since bandwidth control happens on root_domain basis,
4035 * if admission test is enabled, we only admit -deadline
4036 * tasks allowed to run on all the CPUs in the task's
4040 if (task_has_dl_policy(p)) {
4041 const struct cpumask *span = task_rq(p)->rd->span;
4043 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
4050 retval = set_cpus_allowed_ptr(p, new_mask);
4053 cpuset_cpus_allowed(p, cpus_allowed);
4054 if (!cpumask_subset(new_mask, cpus_allowed)) {
4056 * We must have raced with a concurrent cpuset
4057 * update. Just reset the cpus_allowed to the
4058 * cpuset's cpus_allowed
4060 cpumask_copy(new_mask, cpus_allowed);
4065 free_cpumask_var(new_mask);
4066 out_free_cpus_allowed:
4067 free_cpumask_var(cpus_allowed);
4073 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4074 struct cpumask *new_mask)
4076 if (len < cpumask_size())
4077 cpumask_clear(new_mask);
4078 else if (len > cpumask_size())
4079 len = cpumask_size();
4081 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4085 * sys_sched_setaffinity - set the cpu affinity of a process
4086 * @pid: pid of the process
4087 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4088 * @user_mask_ptr: user-space pointer to the new cpu mask
4090 * Return: 0 on success. An error code otherwise.
4092 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4093 unsigned long __user *, user_mask_ptr)
4095 cpumask_var_t new_mask;
4098 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4101 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4103 retval = sched_setaffinity(pid, new_mask);
4104 free_cpumask_var(new_mask);
4108 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4110 struct task_struct *p;
4111 unsigned long flags;
4117 p = find_process_by_pid(pid);
4121 retval = security_task_getscheduler(p);
4125 raw_spin_lock_irqsave(&p->pi_lock, flags);
4126 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4127 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4136 * sys_sched_getaffinity - get the cpu affinity of a process
4137 * @pid: pid of the process
4138 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4139 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4141 * Return: 0 on success. An error code otherwise.
4143 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4144 unsigned long __user *, user_mask_ptr)
4149 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4151 if (len & (sizeof(unsigned long)-1))
4154 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4157 ret = sched_getaffinity(pid, mask);
4159 size_t retlen = min_t(size_t, len, cpumask_size());
4161 if (copy_to_user(user_mask_ptr, mask, retlen))
4166 free_cpumask_var(mask);
4172 * sys_sched_yield - yield the current processor to other threads.
4174 * This function yields the current CPU to other tasks. If there are no
4175 * other threads running on this CPU then this function will return.
4179 SYSCALL_DEFINE0(sched_yield)
4181 struct rq *rq = this_rq_lock();
4183 schedstat_inc(rq, yld_count);
4184 current->sched_class->yield_task(rq);
4187 * Since we are going to call schedule() anyway, there's
4188 * no need to preempt or enable interrupts:
4190 __release(rq->lock);
4191 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4192 do_raw_spin_unlock(&rq->lock);
4193 sched_preempt_enable_no_resched();
4200 static void __cond_resched(void)
4202 __preempt_count_add(PREEMPT_ACTIVE);
4204 __preempt_count_sub(PREEMPT_ACTIVE);
4207 int __sched _cond_resched(void)
4209 if (should_resched()) {
4215 EXPORT_SYMBOL(_cond_resched);
4218 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4219 * call schedule, and on return reacquire the lock.
4221 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4222 * operations here to prevent schedule() from being called twice (once via
4223 * spin_unlock(), once by hand).
4225 int __cond_resched_lock(spinlock_t *lock)
4227 int resched = should_resched();
4230 lockdep_assert_held(lock);
4232 if (spin_needbreak(lock) || resched) {
4243 EXPORT_SYMBOL(__cond_resched_lock);
4245 int __sched __cond_resched_softirq(void)
4247 BUG_ON(!in_softirq());
4249 if (should_resched()) {
4257 EXPORT_SYMBOL(__cond_resched_softirq);
4260 * yield - yield the current processor to other threads.
4262 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4264 * The scheduler is at all times free to pick the calling task as the most
4265 * eligible task to run, if removing the yield() call from your code breaks
4266 * it, its already broken.
4268 * Typical broken usage is:
4273 * where one assumes that yield() will let 'the other' process run that will
4274 * make event true. If the current task is a SCHED_FIFO task that will never
4275 * happen. Never use yield() as a progress guarantee!!
4277 * If you want to use yield() to wait for something, use wait_event().
4278 * If you want to use yield() to be 'nice' for others, use cond_resched().
4279 * If you still want to use yield(), do not!
4281 void __sched yield(void)
4283 set_current_state(TASK_RUNNING);
4286 EXPORT_SYMBOL(yield);
4289 * yield_to - yield the current processor to another thread in
4290 * your thread group, or accelerate that thread toward the
4291 * processor it's on.
4293 * @preempt: whether task preemption is allowed or not
4295 * It's the caller's job to ensure that the target task struct
4296 * can't go away on us before we can do any checks.
4299 * true (>0) if we indeed boosted the target task.
4300 * false (0) if we failed to boost the target.
4301 * -ESRCH if there's no task to yield to.
4303 int __sched yield_to(struct task_struct *p, bool preempt)
4305 struct task_struct *curr = current;
4306 struct rq *rq, *p_rq;
4307 unsigned long flags;
4310 local_irq_save(flags);
4316 * If we're the only runnable task on the rq and target rq also
4317 * has only one task, there's absolutely no point in yielding.
4319 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4324 double_rq_lock(rq, p_rq);
4325 if (task_rq(p) != p_rq) {
4326 double_rq_unlock(rq, p_rq);
4330 if (!curr->sched_class->yield_to_task)
4333 if (curr->sched_class != p->sched_class)
4336 if (task_running(p_rq, p) || p->state)
4339 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4341 schedstat_inc(rq, yld_count);
4343 * Make p's CPU reschedule; pick_next_entity takes care of
4346 if (preempt && rq != p_rq)
4351 double_rq_unlock(rq, p_rq);
4353 local_irq_restore(flags);
4360 EXPORT_SYMBOL_GPL(yield_to);
4363 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4364 * that process accounting knows that this is a task in IO wait state.
4366 void __sched io_schedule(void)
4368 struct rq *rq = raw_rq();
4370 delayacct_blkio_start();
4371 atomic_inc(&rq->nr_iowait);
4372 blk_flush_plug(current);
4373 current->in_iowait = 1;
4375 current->in_iowait = 0;
4376 atomic_dec(&rq->nr_iowait);
4377 delayacct_blkio_end();
4379 EXPORT_SYMBOL(io_schedule);
4381 long __sched io_schedule_timeout(long timeout)
4383 struct rq *rq = raw_rq();
4386 delayacct_blkio_start();
4387 atomic_inc(&rq->nr_iowait);
4388 blk_flush_plug(current);
4389 current->in_iowait = 1;
4390 ret = schedule_timeout(timeout);
4391 current->in_iowait = 0;
4392 atomic_dec(&rq->nr_iowait);
4393 delayacct_blkio_end();
4398 * sys_sched_get_priority_max - return maximum RT priority.
4399 * @policy: scheduling class.
4401 * Return: On success, this syscall returns the maximum
4402 * rt_priority that can be used by a given scheduling class.
4403 * On failure, a negative error code is returned.
4405 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4412 ret = MAX_USER_RT_PRIO-1;
4414 case SCHED_DEADLINE:
4425 * sys_sched_get_priority_min - return minimum RT priority.
4426 * @policy: scheduling class.
4428 * Return: On success, this syscall returns the minimum
4429 * rt_priority that can be used by a given scheduling class.
4430 * On failure, a negative error code is returned.
4432 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4441 case SCHED_DEADLINE:
4451 * sys_sched_rr_get_interval - return the default timeslice of a process.
4452 * @pid: pid of the process.
4453 * @interval: userspace pointer to the timeslice value.
4455 * this syscall writes the default timeslice value of a given process
4456 * into the user-space timespec buffer. A value of '0' means infinity.
4458 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4461 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4462 struct timespec __user *, interval)
4464 struct task_struct *p;
4465 unsigned int time_slice;
4466 unsigned long flags;
4476 p = find_process_by_pid(pid);
4480 retval = security_task_getscheduler(p);
4484 rq = task_rq_lock(p, &flags);
4486 if (p->sched_class->get_rr_interval)
4487 time_slice = p->sched_class->get_rr_interval(rq, p);
4488 task_rq_unlock(rq, p, &flags);
4491 jiffies_to_timespec(time_slice, &t);
4492 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4500 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4502 void sched_show_task(struct task_struct *p)
4504 unsigned long free = 0;
4508 state = p->state ? __ffs(p->state) + 1 : 0;
4509 printk(KERN_INFO "%-15.15s %c", p->comm,
4510 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4511 #if BITS_PER_LONG == 32
4512 if (state == TASK_RUNNING)
4513 printk(KERN_CONT " running ");
4515 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4517 if (state == TASK_RUNNING)
4518 printk(KERN_CONT " running task ");
4520 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4522 #ifdef CONFIG_DEBUG_STACK_USAGE
4523 free = stack_not_used(p);
4526 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4528 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4529 task_pid_nr(p), ppid,
4530 (unsigned long)task_thread_info(p)->flags);
4532 print_worker_info(KERN_INFO, p);
4533 show_stack(p, NULL);
4536 void show_state_filter(unsigned long state_filter)
4538 struct task_struct *g, *p;
4540 #if BITS_PER_LONG == 32
4542 " task PC stack pid father\n");
4545 " task PC stack pid father\n");
4548 for_each_process_thread(g, p) {
4550 * reset the NMI-timeout, listing all files on a slow
4551 * console might take a lot of time:
4553 touch_nmi_watchdog();
4554 if (!state_filter || (p->state & state_filter))
4558 touch_all_softlockup_watchdogs();
4560 #ifdef CONFIG_SCHED_DEBUG
4561 sysrq_sched_debug_show();
4565 * Only show locks if all tasks are dumped:
4568 debug_show_all_locks();
4571 void init_idle_bootup_task(struct task_struct *idle)
4573 idle->sched_class = &idle_sched_class;
4577 * init_idle - set up an idle thread for a given CPU
4578 * @idle: task in question
4579 * @cpu: cpu the idle task belongs to
4581 * NOTE: this function does not set the idle thread's NEED_RESCHED
4582 * flag, to make booting more robust.
4584 void init_idle(struct task_struct *idle, int cpu)
4586 struct rq *rq = cpu_rq(cpu);
4587 unsigned long flags;
4589 raw_spin_lock_irqsave(&rq->lock, flags);
4591 __sched_fork(0, idle);
4592 idle->state = TASK_RUNNING;
4593 idle->se.exec_start = sched_clock();
4595 do_set_cpus_allowed(idle, cpumask_of(cpu));
4597 * We're having a chicken and egg problem, even though we are
4598 * holding rq->lock, the cpu isn't yet set to this cpu so the
4599 * lockdep check in task_group() will fail.
4601 * Similar case to sched_fork(). / Alternatively we could
4602 * use task_rq_lock() here and obtain the other rq->lock.
4607 __set_task_cpu(idle, cpu);
4610 rq->curr = rq->idle = idle;
4611 idle->on_rq = TASK_ON_RQ_QUEUED;
4612 #if defined(CONFIG_SMP)
4615 raw_spin_unlock_irqrestore(&rq->lock, flags);
4617 /* Set the preempt count _outside_ the spinlocks! */
4618 init_idle_preempt_count(idle, cpu);
4621 * The idle tasks have their own, simple scheduling class:
4623 idle->sched_class = &idle_sched_class;
4624 ftrace_graph_init_idle_task(idle, cpu);
4625 vtime_init_idle(idle, cpu);
4626 #if defined(CONFIG_SMP)
4627 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4632 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4634 if (p->sched_class && p->sched_class->set_cpus_allowed)
4635 p->sched_class->set_cpus_allowed(p, new_mask);
4637 cpumask_copy(&p->cpus_allowed, new_mask);
4638 p->nr_cpus_allowed = cpumask_weight(new_mask);
4642 * This is how migration works:
4644 * 1) we invoke migration_cpu_stop() on the target CPU using
4646 * 2) stopper starts to run (implicitly forcing the migrated thread
4648 * 3) it checks whether the migrated task is still in the wrong runqueue.
4649 * 4) if it's in the wrong runqueue then the migration thread removes
4650 * it and puts it into the right queue.
4651 * 5) stopper completes and stop_one_cpu() returns and the migration
4656 * Change a given task's CPU affinity. Migrate the thread to a
4657 * proper CPU and schedule it away if the CPU it's executing on
4658 * is removed from the allowed bitmask.
4660 * NOTE: the caller must have a valid reference to the task, the
4661 * task must not exit() & deallocate itself prematurely. The
4662 * call is not atomic; no spinlocks may be held.
4664 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4666 unsigned long flags;
4668 unsigned int dest_cpu;
4671 rq = task_rq_lock(p, &flags);
4673 if (cpumask_equal(&p->cpus_allowed, new_mask))
4676 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4681 do_set_cpus_allowed(p, new_mask);
4683 /* Can the task run on the task's current CPU? If so, we're done */
4684 if (cpumask_test_cpu(task_cpu(p), new_mask))
4687 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4688 if (task_on_rq_queued(p) || p->state == TASK_WAKING) {
4689 struct migration_arg arg = { p, dest_cpu };
4690 /* Need help from migration thread: drop lock and wait. */
4691 task_rq_unlock(rq, p, &flags);
4692 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4693 tlb_migrate_finish(p->mm);
4697 task_rq_unlock(rq, p, &flags);
4701 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4704 * Move (not current) task off this cpu, onto dest cpu. We're doing
4705 * this because either it can't run here any more (set_cpus_allowed()
4706 * away from this CPU, or CPU going down), or because we're
4707 * attempting to rebalance this task on exec (sched_exec).
4709 * So we race with normal scheduler movements, but that's OK, as long
4710 * as the task is no longer on this CPU.
4712 * Returns non-zero if task was successfully migrated.
4714 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4719 if (unlikely(!cpu_active(dest_cpu)))
4722 rq = cpu_rq(src_cpu);
4724 raw_spin_lock(&p->pi_lock);
4725 raw_spin_lock(&rq->lock);
4726 /* Already moved. */
4727 if (task_cpu(p) != src_cpu)
4730 /* Affinity changed (again). */
4731 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4735 * If we're not on a rq, the next wake-up will ensure we're
4738 if (task_on_rq_queued(p)) {
4739 dequeue_task(rq, p, 0);
4740 p->on_rq = TASK_ON_RQ_MIGRATING;
4741 set_task_cpu(p, dest_cpu);
4742 raw_spin_unlock(&rq->lock);
4744 rq = cpu_rq(dest_cpu);
4745 raw_spin_lock(&rq->lock);
4746 BUG_ON(task_rq(p) != rq);
4747 p->on_rq = TASK_ON_RQ_QUEUED;
4748 enqueue_task(rq, p, 0);
4749 check_preempt_curr(rq, p, 0);
4754 raw_spin_unlock(&rq->lock);
4755 raw_spin_unlock(&p->pi_lock);
4759 #ifdef CONFIG_NUMA_BALANCING
4760 /* Migrate current task p to target_cpu */
4761 int migrate_task_to(struct task_struct *p, int target_cpu)
4763 struct migration_arg arg = { p, target_cpu };
4764 int curr_cpu = task_cpu(p);
4766 if (curr_cpu == target_cpu)
4769 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4772 /* TODO: This is not properly updating schedstats */
4774 trace_sched_move_numa(p, curr_cpu, target_cpu);
4775 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4779 * Requeue a task on a given node and accurately track the number of NUMA
4780 * tasks on the runqueues
4782 void sched_setnuma(struct task_struct *p, int nid)
4785 unsigned long flags;
4786 bool queued, running;
4788 rq = task_rq_lock(p, &flags);
4789 queued = task_on_rq_queued(p);
4790 running = task_current(rq, p);
4793 dequeue_task(rq, p, 0);
4795 put_prev_task(rq, p);
4797 p->numa_preferred_nid = nid;
4800 p->sched_class->set_curr_task(rq);
4802 enqueue_task(rq, p, 0);
4803 task_rq_unlock(rq, p, &flags);
4808 * migration_cpu_stop - this will be executed by a highprio stopper thread
4809 * and performs thread migration by bumping thread off CPU then
4810 * 'pushing' onto another runqueue.
4812 static int migration_cpu_stop(void *data)
4814 struct migration_arg *arg = data;
4817 * The original target cpu might have gone down and we might
4818 * be on another cpu but it doesn't matter.
4820 local_irq_disable();
4822 * We need to explicitly wake pending tasks before running
4823 * __migrate_task() such that we will not miss enforcing cpus_allowed
4824 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4826 sched_ttwu_pending();
4827 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4832 #ifdef CONFIG_HOTPLUG_CPU
4835 * Ensures that the idle task is using init_mm right before its cpu goes
4838 void idle_task_exit(void)
4840 struct mm_struct *mm = current->active_mm;
4842 BUG_ON(cpu_online(smp_processor_id()));
4844 if (mm != &init_mm) {
4845 switch_mm(mm, &init_mm, current);
4846 finish_arch_post_lock_switch();
4852 * Since this CPU is going 'away' for a while, fold any nr_active delta
4853 * we might have. Assumes we're called after migrate_tasks() so that the
4854 * nr_active count is stable.
4856 * Also see the comment "Global load-average calculations".
4858 static void calc_load_migrate(struct rq *rq)
4860 long delta = calc_load_fold_active(rq);
4862 atomic_long_add(delta, &calc_load_tasks);
4865 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4869 static const struct sched_class fake_sched_class = {
4870 .put_prev_task = put_prev_task_fake,
4873 static struct task_struct fake_task = {
4875 * Avoid pull_{rt,dl}_task()
4877 .prio = MAX_PRIO + 1,
4878 .sched_class = &fake_sched_class,
4882 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4883 * try_to_wake_up()->select_task_rq().
4885 * Called with rq->lock held even though we'er in stop_machine() and
4886 * there's no concurrency possible, we hold the required locks anyway
4887 * because of lock validation efforts.
4889 static void migrate_tasks(unsigned int dead_cpu)
4891 struct rq *rq = cpu_rq(dead_cpu);
4892 struct task_struct *next, *stop = rq->stop;
4896 * Fudge the rq selection such that the below task selection loop
4897 * doesn't get stuck on the currently eligible stop task.
4899 * We're currently inside stop_machine() and the rq is either stuck
4900 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4901 * either way we should never end up calling schedule() until we're
4907 * put_prev_task() and pick_next_task() sched
4908 * class method both need to have an up-to-date
4909 * value of rq->clock[_task]
4911 update_rq_clock(rq);
4915 * There's this thread running, bail when that's the only
4918 if (rq->nr_running == 1)
4921 next = pick_next_task(rq, &fake_task);
4923 next->sched_class->put_prev_task(rq, next);
4925 /* Find suitable destination for @next, with force if needed. */
4926 dest_cpu = select_fallback_rq(dead_cpu, next);
4927 raw_spin_unlock(&rq->lock);
4929 __migrate_task(next, dead_cpu, dest_cpu);
4931 raw_spin_lock(&rq->lock);
4937 #endif /* CONFIG_HOTPLUG_CPU */
4939 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4941 static struct ctl_table sd_ctl_dir[] = {
4943 .procname = "sched_domain",
4949 static struct ctl_table sd_ctl_root[] = {
4951 .procname = "kernel",
4953 .child = sd_ctl_dir,
4958 static struct ctl_table *sd_alloc_ctl_entry(int n)
4960 struct ctl_table *entry =
4961 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4966 static void sd_free_ctl_entry(struct ctl_table **tablep)
4968 struct ctl_table *entry;
4971 * In the intermediate directories, both the child directory and
4972 * procname are dynamically allocated and could fail but the mode
4973 * will always be set. In the lowest directory the names are
4974 * static strings and all have proc handlers.
4976 for (entry = *tablep; entry->mode; entry++) {
4978 sd_free_ctl_entry(&entry->child);
4979 if (entry->proc_handler == NULL)
4980 kfree(entry->procname);
4987 static int min_load_idx = 0;
4988 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4991 set_table_entry(struct ctl_table *entry,
4992 const char *procname, void *data, int maxlen,
4993 umode_t mode, proc_handler *proc_handler,
4996 entry->procname = procname;
4998 entry->maxlen = maxlen;
5000 entry->proc_handler = proc_handler;
5003 entry->extra1 = &min_load_idx;
5004 entry->extra2 = &max_load_idx;
5008 static struct ctl_table *
5009 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5011 struct ctl_table *table = sd_alloc_ctl_entry(14);
5016 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5017 sizeof(long), 0644, proc_doulongvec_minmax, false);
5018 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5019 sizeof(long), 0644, proc_doulongvec_minmax, false);
5020 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5021 sizeof(int), 0644, proc_dointvec_minmax, true);
5022 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5023 sizeof(int), 0644, proc_dointvec_minmax, true);
5024 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5025 sizeof(int), 0644, proc_dointvec_minmax, true);
5026 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5027 sizeof(int), 0644, proc_dointvec_minmax, true);
5028 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5029 sizeof(int), 0644, proc_dointvec_minmax, true);
5030 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5031 sizeof(int), 0644, proc_dointvec_minmax, false);
5032 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5033 sizeof(int), 0644, proc_dointvec_minmax, false);
5034 set_table_entry(&table[9], "cache_nice_tries",
5035 &sd->cache_nice_tries,
5036 sizeof(int), 0644, proc_dointvec_minmax, false);
5037 set_table_entry(&table[10], "flags", &sd->flags,
5038 sizeof(int), 0644, proc_dointvec_minmax, false);
5039 set_table_entry(&table[11], "max_newidle_lb_cost",
5040 &sd->max_newidle_lb_cost,
5041 sizeof(long), 0644, proc_doulongvec_minmax, false);
5042 set_table_entry(&table[12], "name", sd->name,
5043 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5044 /* &table[13] is terminator */
5049 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5051 struct ctl_table *entry, *table;
5052 struct sched_domain *sd;
5053 int domain_num = 0, i;
5056 for_each_domain(cpu, sd)
5058 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5063 for_each_domain(cpu, sd) {
5064 snprintf(buf, 32, "domain%d", i);
5065 entry->procname = kstrdup(buf, GFP_KERNEL);
5067 entry->child = sd_alloc_ctl_domain_table(sd);
5074 static struct ctl_table_header *sd_sysctl_header;
5075 static void register_sched_domain_sysctl(void)
5077 int i, cpu_num = num_possible_cpus();
5078 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5081 WARN_ON(sd_ctl_dir[0].child);
5082 sd_ctl_dir[0].child = entry;
5087 for_each_possible_cpu(i) {
5088 snprintf(buf, 32, "cpu%d", i);
5089 entry->procname = kstrdup(buf, GFP_KERNEL);
5091 entry->child = sd_alloc_ctl_cpu_table(i);
5095 WARN_ON(sd_sysctl_header);
5096 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5099 /* may be called multiple times per register */
5100 static void unregister_sched_domain_sysctl(void)
5102 if (sd_sysctl_header)
5103 unregister_sysctl_table(sd_sysctl_header);
5104 sd_sysctl_header = NULL;
5105 if (sd_ctl_dir[0].child)
5106 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5109 static void register_sched_domain_sysctl(void)
5112 static void unregister_sched_domain_sysctl(void)
5117 static void set_rq_online(struct rq *rq)
5120 const struct sched_class *class;
5122 cpumask_set_cpu(rq->cpu, rq->rd->online);
5125 for_each_class(class) {
5126 if (class->rq_online)
5127 class->rq_online(rq);
5132 static void set_rq_offline(struct rq *rq)
5135 const struct sched_class *class;
5137 for_each_class(class) {
5138 if (class->rq_offline)
5139 class->rq_offline(rq);
5142 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5148 * migration_call - callback that gets triggered when a CPU is added.
5149 * Here we can start up the necessary migration thread for the new CPU.
5152 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5154 int cpu = (long)hcpu;
5155 unsigned long flags;
5156 struct rq *rq = cpu_rq(cpu);
5158 switch (action & ~CPU_TASKS_FROZEN) {
5160 case CPU_UP_PREPARE:
5161 rq->calc_load_update = calc_load_update;
5165 /* Update our root-domain */
5166 raw_spin_lock_irqsave(&rq->lock, flags);
5168 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5172 raw_spin_unlock_irqrestore(&rq->lock, flags);
5175 #ifdef CONFIG_HOTPLUG_CPU
5177 sched_ttwu_pending();
5178 /* Update our root-domain */
5179 raw_spin_lock_irqsave(&rq->lock, flags);
5181 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5185 BUG_ON(rq->nr_running != 1); /* the migration thread */
5186 raw_spin_unlock_irqrestore(&rq->lock, flags);
5190 calc_load_migrate(rq);
5195 update_max_interval();
5201 * Register at high priority so that task migration (migrate_all_tasks)
5202 * happens before everything else. This has to be lower priority than
5203 * the notifier in the perf_event subsystem, though.
5205 static struct notifier_block migration_notifier = {
5206 .notifier_call = migration_call,
5207 .priority = CPU_PRI_MIGRATION,
5210 static void __cpuinit set_cpu_rq_start_time(void)
5212 int cpu = smp_processor_id();
5213 struct rq *rq = cpu_rq(cpu);
5214 rq->age_stamp = sched_clock_cpu(cpu);
5217 static int sched_cpu_active(struct notifier_block *nfb,
5218 unsigned long action, void *hcpu)
5220 switch (action & ~CPU_TASKS_FROZEN) {
5222 set_cpu_rq_start_time();
5224 case CPU_DOWN_FAILED:
5225 set_cpu_active((long)hcpu, true);
5232 static int sched_cpu_inactive(struct notifier_block *nfb,
5233 unsigned long action, void *hcpu)
5235 unsigned long flags;
5236 long cpu = (long)hcpu;
5238 switch (action & ~CPU_TASKS_FROZEN) {
5239 case CPU_DOWN_PREPARE:
5240 set_cpu_active(cpu, false);
5242 /* explicitly allow suspend */
5243 if (!(action & CPU_TASKS_FROZEN)) {
5244 struct dl_bw *dl_b = dl_bw_of(cpu);
5248 raw_spin_lock_irqsave(&dl_b->lock, flags);
5249 cpus = dl_bw_cpus(cpu);
5250 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5251 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5254 return notifier_from_errno(-EBUSY);
5262 static int __init migration_init(void)
5264 void *cpu = (void *)(long)smp_processor_id();
5267 /* Initialize migration for the boot CPU */
5268 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5269 BUG_ON(err == NOTIFY_BAD);
5270 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5271 register_cpu_notifier(&migration_notifier);
5273 /* Register cpu active notifiers */
5274 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5275 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5279 early_initcall(migration_init);
5284 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5286 #ifdef CONFIG_SCHED_DEBUG
5288 static __read_mostly int sched_debug_enabled;
5290 static int __init sched_debug_setup(char *str)
5292 sched_debug_enabled = 1;
5296 early_param("sched_debug", sched_debug_setup);
5298 static inline bool sched_debug(void)
5300 return sched_debug_enabled;
5303 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5304 struct cpumask *groupmask)
5306 struct sched_group *group = sd->groups;
5309 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5310 cpumask_clear(groupmask);
5312 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5314 if (!(sd->flags & SD_LOAD_BALANCE)) {
5315 printk("does not load-balance\n");
5317 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5322 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5324 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5325 printk(KERN_ERR "ERROR: domain->span does not contain "
5328 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5329 printk(KERN_ERR "ERROR: domain->groups does not contain"
5333 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5337 printk(KERN_ERR "ERROR: group is NULL\n");
5342 * Even though we initialize ->capacity to something semi-sane,
5343 * we leave capacity_orig unset. This allows us to detect if
5344 * domain iteration is still funny without causing /0 traps.
5346 if (!group->sgc->capacity_orig) {
5347 printk(KERN_CONT "\n");
5348 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5352 if (!cpumask_weight(sched_group_cpus(group))) {
5353 printk(KERN_CONT "\n");
5354 printk(KERN_ERR "ERROR: empty group\n");
5358 if (!(sd->flags & SD_OVERLAP) &&
5359 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5360 printk(KERN_CONT "\n");
5361 printk(KERN_ERR "ERROR: repeated CPUs\n");
5365 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5367 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5369 printk(KERN_CONT " %s", str);
5370 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5371 printk(KERN_CONT " (cpu_capacity = %d)",
5372 group->sgc->capacity);
5375 group = group->next;
5376 } while (group != sd->groups);
5377 printk(KERN_CONT "\n");
5379 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5380 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5383 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5384 printk(KERN_ERR "ERROR: parent span is not a superset "
5385 "of domain->span\n");
5389 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5393 if (!sched_debug_enabled)
5397 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5401 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5404 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5412 #else /* !CONFIG_SCHED_DEBUG */
5413 # define sched_domain_debug(sd, cpu) do { } while (0)
5414 static inline bool sched_debug(void)
5418 #endif /* CONFIG_SCHED_DEBUG */
5420 static int sd_degenerate(struct sched_domain *sd)
5422 if (cpumask_weight(sched_domain_span(sd)) == 1)
5425 /* Following flags need at least 2 groups */
5426 if (sd->flags & (SD_LOAD_BALANCE |
5427 SD_BALANCE_NEWIDLE |
5430 SD_SHARE_CPUCAPACITY |
5431 SD_SHARE_PKG_RESOURCES |
5432 SD_SHARE_POWERDOMAIN)) {
5433 if (sd->groups != sd->groups->next)
5437 /* Following flags don't use groups */
5438 if (sd->flags & (SD_WAKE_AFFINE))
5445 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5447 unsigned long cflags = sd->flags, pflags = parent->flags;
5449 if (sd_degenerate(parent))
5452 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5455 /* Flags needing groups don't count if only 1 group in parent */
5456 if (parent->groups == parent->groups->next) {
5457 pflags &= ~(SD_LOAD_BALANCE |
5458 SD_BALANCE_NEWIDLE |
5461 SD_SHARE_CPUCAPACITY |
5462 SD_SHARE_PKG_RESOURCES |
5464 SD_SHARE_POWERDOMAIN);
5465 if (nr_node_ids == 1)
5466 pflags &= ~SD_SERIALIZE;
5468 if (~cflags & pflags)
5474 static void free_rootdomain(struct rcu_head *rcu)
5476 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5478 cpupri_cleanup(&rd->cpupri);
5479 cpudl_cleanup(&rd->cpudl);
5480 free_cpumask_var(rd->dlo_mask);
5481 free_cpumask_var(rd->rto_mask);
5482 free_cpumask_var(rd->online);
5483 free_cpumask_var(rd->span);
5487 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5489 struct root_domain *old_rd = NULL;
5490 unsigned long flags;
5492 raw_spin_lock_irqsave(&rq->lock, flags);
5497 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5500 cpumask_clear_cpu(rq->cpu, old_rd->span);
5503 * If we dont want to free the old_rd yet then
5504 * set old_rd to NULL to skip the freeing later
5507 if (!atomic_dec_and_test(&old_rd->refcount))
5511 atomic_inc(&rd->refcount);
5514 cpumask_set_cpu(rq->cpu, rd->span);
5515 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5518 raw_spin_unlock_irqrestore(&rq->lock, flags);
5521 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5524 static int init_rootdomain(struct root_domain *rd)
5526 memset(rd, 0, sizeof(*rd));
5528 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5530 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5532 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5534 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5537 init_dl_bw(&rd->dl_bw);
5538 if (cpudl_init(&rd->cpudl) != 0)
5541 if (cpupri_init(&rd->cpupri) != 0)
5546 free_cpumask_var(rd->rto_mask);
5548 free_cpumask_var(rd->dlo_mask);
5550 free_cpumask_var(rd->online);
5552 free_cpumask_var(rd->span);
5558 * By default the system creates a single root-domain with all cpus as
5559 * members (mimicking the global state we have today).
5561 struct root_domain def_root_domain;
5563 static void init_defrootdomain(void)
5565 init_rootdomain(&def_root_domain);
5567 atomic_set(&def_root_domain.refcount, 1);
5570 static struct root_domain *alloc_rootdomain(void)
5572 struct root_domain *rd;
5574 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5578 if (init_rootdomain(rd) != 0) {
5586 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5588 struct sched_group *tmp, *first;
5597 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5602 } while (sg != first);
5605 static void free_sched_domain(struct rcu_head *rcu)
5607 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5610 * If its an overlapping domain it has private groups, iterate and
5613 if (sd->flags & SD_OVERLAP) {
5614 free_sched_groups(sd->groups, 1);
5615 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5616 kfree(sd->groups->sgc);
5622 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5624 call_rcu(&sd->rcu, free_sched_domain);
5627 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5629 for (; sd; sd = sd->parent)
5630 destroy_sched_domain(sd, cpu);
5634 * Keep a special pointer to the highest sched_domain that has
5635 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5636 * allows us to avoid some pointer chasing select_idle_sibling().
5638 * Also keep a unique ID per domain (we use the first cpu number in
5639 * the cpumask of the domain), this allows us to quickly tell if
5640 * two cpus are in the same cache domain, see cpus_share_cache().
5642 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5643 DEFINE_PER_CPU(int, sd_llc_size);
5644 DEFINE_PER_CPU(int, sd_llc_id);
5645 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5646 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5647 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5649 static void update_top_cache_domain(int cpu)
5651 struct sched_domain *sd;
5652 struct sched_domain *busy_sd = NULL;
5656 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5658 id = cpumask_first(sched_domain_span(sd));
5659 size = cpumask_weight(sched_domain_span(sd));
5660 busy_sd = sd->parent; /* sd_busy */
5662 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5664 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5665 per_cpu(sd_llc_size, cpu) = size;
5666 per_cpu(sd_llc_id, cpu) = id;
5668 sd = lowest_flag_domain(cpu, SD_NUMA);
5669 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5671 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5672 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5676 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5677 * hold the hotplug lock.
5680 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5682 struct rq *rq = cpu_rq(cpu);
5683 struct sched_domain *tmp;
5685 /* Remove the sched domains which do not contribute to scheduling. */
5686 for (tmp = sd; tmp; ) {
5687 struct sched_domain *parent = tmp->parent;
5691 if (sd_parent_degenerate(tmp, parent)) {
5692 tmp->parent = parent->parent;
5694 parent->parent->child = tmp;
5696 * Transfer SD_PREFER_SIBLING down in case of a
5697 * degenerate parent; the spans match for this
5698 * so the property transfers.
5700 if (parent->flags & SD_PREFER_SIBLING)
5701 tmp->flags |= SD_PREFER_SIBLING;
5702 destroy_sched_domain(parent, cpu);
5707 if (sd && sd_degenerate(sd)) {
5710 destroy_sched_domain(tmp, cpu);
5715 sched_domain_debug(sd, cpu);
5717 rq_attach_root(rq, rd);
5719 rcu_assign_pointer(rq->sd, sd);
5720 destroy_sched_domains(tmp, cpu);
5722 update_top_cache_domain(cpu);
5725 /* cpus with isolated domains */
5726 static cpumask_var_t cpu_isolated_map;
5728 /* Setup the mask of cpus configured for isolated domains */
5729 static int __init isolated_cpu_setup(char *str)
5731 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5732 cpulist_parse(str, cpu_isolated_map);
5736 __setup("isolcpus=", isolated_cpu_setup);
5739 struct sched_domain ** __percpu sd;
5740 struct root_domain *rd;
5751 * Build an iteration mask that can exclude certain CPUs from the upwards
5754 * Asymmetric node setups can result in situations where the domain tree is of
5755 * unequal depth, make sure to skip domains that already cover the entire
5758 * In that case build_sched_domains() will have terminated the iteration early
5759 * and our sibling sd spans will be empty. Domains should always include the
5760 * cpu they're built on, so check that.
5763 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5765 const struct cpumask *span = sched_domain_span(sd);
5766 struct sd_data *sdd = sd->private;
5767 struct sched_domain *sibling;
5770 for_each_cpu(i, span) {
5771 sibling = *per_cpu_ptr(sdd->sd, i);
5772 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5775 cpumask_set_cpu(i, sched_group_mask(sg));
5780 * Return the canonical balance cpu for this group, this is the first cpu
5781 * of this group that's also in the iteration mask.
5783 int group_balance_cpu(struct sched_group *sg)
5785 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5789 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5791 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5792 const struct cpumask *span = sched_domain_span(sd);
5793 struct cpumask *covered = sched_domains_tmpmask;
5794 struct sd_data *sdd = sd->private;
5795 struct sched_domain *sibling;
5798 cpumask_clear(covered);
5800 for_each_cpu(i, span) {
5801 struct cpumask *sg_span;
5803 if (cpumask_test_cpu(i, covered))
5806 sibling = *per_cpu_ptr(sdd->sd, i);
5808 /* See the comment near build_group_mask(). */
5809 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5812 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5813 GFP_KERNEL, cpu_to_node(cpu));
5818 sg_span = sched_group_cpus(sg);
5820 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5822 cpumask_set_cpu(i, sg_span);
5824 cpumask_or(covered, covered, sg_span);
5826 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5827 if (atomic_inc_return(&sg->sgc->ref) == 1)
5828 build_group_mask(sd, sg);
5831 * Initialize sgc->capacity such that even if we mess up the
5832 * domains and no possible iteration will get us here, we won't
5835 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5836 sg->sgc->capacity_orig = sg->sgc->capacity;
5839 * Make sure the first group of this domain contains the
5840 * canonical balance cpu. Otherwise the sched_domain iteration
5841 * breaks. See update_sg_lb_stats().
5843 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5844 group_balance_cpu(sg) == cpu)
5854 sd->groups = groups;
5859 free_sched_groups(first, 0);
5864 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5866 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5867 struct sched_domain *child = sd->child;
5870 cpu = cpumask_first(sched_domain_span(child));
5873 *sg = *per_cpu_ptr(sdd->sg, cpu);
5874 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5875 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5882 * build_sched_groups will build a circular linked list of the groups
5883 * covered by the given span, and will set each group's ->cpumask correctly,
5884 * and ->cpu_capacity to 0.
5886 * Assumes the sched_domain tree is fully constructed
5889 build_sched_groups(struct sched_domain *sd, int cpu)
5891 struct sched_group *first = NULL, *last = NULL;
5892 struct sd_data *sdd = sd->private;
5893 const struct cpumask *span = sched_domain_span(sd);
5894 struct cpumask *covered;
5897 get_group(cpu, sdd, &sd->groups);
5898 atomic_inc(&sd->groups->ref);
5900 if (cpu != cpumask_first(span))
5903 lockdep_assert_held(&sched_domains_mutex);
5904 covered = sched_domains_tmpmask;
5906 cpumask_clear(covered);
5908 for_each_cpu(i, span) {
5909 struct sched_group *sg;
5912 if (cpumask_test_cpu(i, covered))
5915 group = get_group(i, sdd, &sg);
5916 cpumask_setall(sched_group_mask(sg));
5918 for_each_cpu(j, span) {
5919 if (get_group(j, sdd, NULL) != group)
5922 cpumask_set_cpu(j, covered);
5923 cpumask_set_cpu(j, sched_group_cpus(sg));
5938 * Initialize sched groups cpu_capacity.
5940 * cpu_capacity indicates the capacity of sched group, which is used while
5941 * distributing the load between different sched groups in a sched domain.
5942 * Typically cpu_capacity for all the groups in a sched domain will be same
5943 * unless there are asymmetries in the topology. If there are asymmetries,
5944 * group having more cpu_capacity will pickup more load compared to the
5945 * group having less cpu_capacity.
5947 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5949 struct sched_group *sg = sd->groups;
5954 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5956 } while (sg != sd->groups);
5958 if (cpu != group_balance_cpu(sg))
5961 update_group_capacity(sd, cpu);
5962 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
5966 * Initializers for schedule domains
5967 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5970 static int default_relax_domain_level = -1;
5971 int sched_domain_level_max;
5973 static int __init setup_relax_domain_level(char *str)
5975 if (kstrtoint(str, 0, &default_relax_domain_level))
5976 pr_warn("Unable to set relax_domain_level\n");
5980 __setup("relax_domain_level=", setup_relax_domain_level);
5982 static void set_domain_attribute(struct sched_domain *sd,
5983 struct sched_domain_attr *attr)
5987 if (!attr || attr->relax_domain_level < 0) {
5988 if (default_relax_domain_level < 0)
5991 request = default_relax_domain_level;
5993 request = attr->relax_domain_level;
5994 if (request < sd->level) {
5995 /* turn off idle balance on this domain */
5996 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5998 /* turn on idle balance on this domain */
5999 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6003 static void __sdt_free(const struct cpumask *cpu_map);
6004 static int __sdt_alloc(const struct cpumask *cpu_map);
6006 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6007 const struct cpumask *cpu_map)
6011 if (!atomic_read(&d->rd->refcount))
6012 free_rootdomain(&d->rd->rcu); /* fall through */
6014 free_percpu(d->sd); /* fall through */
6016 __sdt_free(cpu_map); /* fall through */
6022 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6023 const struct cpumask *cpu_map)
6025 memset(d, 0, sizeof(*d));
6027 if (__sdt_alloc(cpu_map))
6028 return sa_sd_storage;
6029 d->sd = alloc_percpu(struct sched_domain *);
6031 return sa_sd_storage;
6032 d->rd = alloc_rootdomain();
6035 return sa_rootdomain;
6039 * NULL the sd_data elements we've used to build the sched_domain and
6040 * sched_group structure so that the subsequent __free_domain_allocs()
6041 * will not free the data we're using.
6043 static void claim_allocations(int cpu, struct sched_domain *sd)
6045 struct sd_data *sdd = sd->private;
6047 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6048 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6050 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6051 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6053 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6054 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6058 static int sched_domains_numa_levels;
6059 static int *sched_domains_numa_distance;
6060 static struct cpumask ***sched_domains_numa_masks;
6061 static int sched_domains_curr_level;
6065 * SD_flags allowed in topology descriptions.
6067 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6068 * SD_SHARE_PKG_RESOURCES - describes shared caches
6069 * SD_NUMA - describes NUMA topologies
6070 * SD_SHARE_POWERDOMAIN - describes shared power domain
6073 * SD_ASYM_PACKING - describes SMT quirks
6075 #define TOPOLOGY_SD_FLAGS \
6076 (SD_SHARE_CPUCAPACITY | \
6077 SD_SHARE_PKG_RESOURCES | \
6080 SD_SHARE_POWERDOMAIN)
6082 static struct sched_domain *
6083 sd_init(struct sched_domain_topology_level *tl, int cpu)
6085 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6086 int sd_weight, sd_flags = 0;
6090 * Ugly hack to pass state to sd_numa_mask()...
6092 sched_domains_curr_level = tl->numa_level;
6095 sd_weight = cpumask_weight(tl->mask(cpu));
6098 sd_flags = (*tl->sd_flags)();
6099 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6100 "wrong sd_flags in topology description\n"))
6101 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6103 *sd = (struct sched_domain){
6104 .min_interval = sd_weight,
6105 .max_interval = 2*sd_weight,
6107 .imbalance_pct = 125,
6109 .cache_nice_tries = 0,
6116 .flags = 1*SD_LOAD_BALANCE
6117 | 1*SD_BALANCE_NEWIDLE
6122 | 0*SD_SHARE_CPUCAPACITY
6123 | 0*SD_SHARE_PKG_RESOURCES
6125 | 0*SD_PREFER_SIBLING
6130 .last_balance = jiffies,
6131 .balance_interval = sd_weight,
6133 .max_newidle_lb_cost = 0,
6134 .next_decay_max_lb_cost = jiffies,
6135 #ifdef CONFIG_SCHED_DEBUG
6141 * Convert topological properties into behaviour.
6144 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6145 sd->imbalance_pct = 110;
6146 sd->smt_gain = 1178; /* ~15% */
6148 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6149 sd->imbalance_pct = 117;
6150 sd->cache_nice_tries = 1;
6154 } else if (sd->flags & SD_NUMA) {
6155 sd->cache_nice_tries = 2;
6159 sd->flags |= SD_SERIALIZE;
6160 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6161 sd->flags &= ~(SD_BALANCE_EXEC |
6168 sd->flags |= SD_PREFER_SIBLING;
6169 sd->cache_nice_tries = 1;
6174 sd->private = &tl->data;
6180 * Topology list, bottom-up.
6182 static struct sched_domain_topology_level default_topology[] = {
6183 #ifdef CONFIG_SCHED_SMT
6184 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6186 #ifdef CONFIG_SCHED_MC
6187 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6189 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6193 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6195 #define for_each_sd_topology(tl) \
6196 for (tl = sched_domain_topology; tl->mask; tl++)
6198 void set_sched_topology(struct sched_domain_topology_level *tl)
6200 sched_domain_topology = tl;
6205 static const struct cpumask *sd_numa_mask(int cpu)
6207 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6210 static void sched_numa_warn(const char *str)
6212 static int done = false;
6220 printk(KERN_WARNING "ERROR: %s\n\n", str);
6222 for (i = 0; i < nr_node_ids; i++) {
6223 printk(KERN_WARNING " ");
6224 for (j = 0; j < nr_node_ids; j++)
6225 printk(KERN_CONT "%02d ", node_distance(i,j));
6226 printk(KERN_CONT "\n");
6228 printk(KERN_WARNING "\n");
6231 static bool find_numa_distance(int distance)
6235 if (distance == node_distance(0, 0))
6238 for (i = 0; i < sched_domains_numa_levels; i++) {
6239 if (sched_domains_numa_distance[i] == distance)
6246 static void sched_init_numa(void)
6248 int next_distance, curr_distance = node_distance(0, 0);
6249 struct sched_domain_topology_level *tl;
6253 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6254 if (!sched_domains_numa_distance)
6258 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6259 * unique distances in the node_distance() table.
6261 * Assumes node_distance(0,j) includes all distances in
6262 * node_distance(i,j) in order to avoid cubic time.
6264 next_distance = curr_distance;
6265 for (i = 0; i < nr_node_ids; i++) {
6266 for (j = 0; j < nr_node_ids; j++) {
6267 for (k = 0; k < nr_node_ids; k++) {
6268 int distance = node_distance(i, k);
6270 if (distance > curr_distance &&
6271 (distance < next_distance ||
6272 next_distance == curr_distance))
6273 next_distance = distance;
6276 * While not a strong assumption it would be nice to know
6277 * about cases where if node A is connected to B, B is not
6278 * equally connected to A.
6280 if (sched_debug() && node_distance(k, i) != distance)
6281 sched_numa_warn("Node-distance not symmetric");
6283 if (sched_debug() && i && !find_numa_distance(distance))
6284 sched_numa_warn("Node-0 not representative");
6286 if (next_distance != curr_distance) {
6287 sched_domains_numa_distance[level++] = next_distance;
6288 sched_domains_numa_levels = level;
6289 curr_distance = next_distance;
6294 * In case of sched_debug() we verify the above assumption.
6300 * 'level' contains the number of unique distances, excluding the
6301 * identity distance node_distance(i,i).
6303 * The sched_domains_numa_distance[] array includes the actual distance
6308 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6309 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6310 * the array will contain less then 'level' members. This could be
6311 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6312 * in other functions.
6314 * We reset it to 'level' at the end of this function.
6316 sched_domains_numa_levels = 0;
6318 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6319 if (!sched_domains_numa_masks)
6323 * Now for each level, construct a mask per node which contains all
6324 * cpus of nodes that are that many hops away from us.
6326 for (i = 0; i < level; i++) {
6327 sched_domains_numa_masks[i] =
6328 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6329 if (!sched_domains_numa_masks[i])
6332 for (j = 0; j < nr_node_ids; j++) {
6333 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6337 sched_domains_numa_masks[i][j] = mask;
6339 for (k = 0; k < nr_node_ids; k++) {
6340 if (node_distance(j, k) > sched_domains_numa_distance[i])
6343 cpumask_or(mask, mask, cpumask_of_node(k));
6348 /* Compute default topology size */
6349 for (i = 0; sched_domain_topology[i].mask; i++);
6351 tl = kzalloc((i + level + 1) *
6352 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6357 * Copy the default topology bits..
6359 for (i = 0; sched_domain_topology[i].mask; i++)
6360 tl[i] = sched_domain_topology[i];
6363 * .. and append 'j' levels of NUMA goodness.
6365 for (j = 0; j < level; i++, j++) {
6366 tl[i] = (struct sched_domain_topology_level){
6367 .mask = sd_numa_mask,
6368 .sd_flags = cpu_numa_flags,
6369 .flags = SDTL_OVERLAP,
6375 sched_domain_topology = tl;
6377 sched_domains_numa_levels = level;
6380 static void sched_domains_numa_masks_set(int cpu)
6383 int node = cpu_to_node(cpu);
6385 for (i = 0; i < sched_domains_numa_levels; i++) {
6386 for (j = 0; j < nr_node_ids; j++) {
6387 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6388 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6393 static void sched_domains_numa_masks_clear(int cpu)
6396 for (i = 0; i < sched_domains_numa_levels; i++) {
6397 for (j = 0; j < nr_node_ids; j++)
6398 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6403 * Update sched_domains_numa_masks[level][node] array when new cpus
6406 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6407 unsigned long action,
6410 int cpu = (long)hcpu;
6412 switch (action & ~CPU_TASKS_FROZEN) {
6414 sched_domains_numa_masks_set(cpu);
6418 sched_domains_numa_masks_clear(cpu);
6428 static inline void sched_init_numa(void)
6432 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6433 unsigned long action,
6438 #endif /* CONFIG_NUMA */
6440 static int __sdt_alloc(const struct cpumask *cpu_map)
6442 struct sched_domain_topology_level *tl;
6445 for_each_sd_topology(tl) {
6446 struct sd_data *sdd = &tl->data;
6448 sdd->sd = alloc_percpu(struct sched_domain *);
6452 sdd->sg = alloc_percpu(struct sched_group *);
6456 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6460 for_each_cpu(j, cpu_map) {
6461 struct sched_domain *sd;
6462 struct sched_group *sg;
6463 struct sched_group_capacity *sgc;
6465 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6466 GFP_KERNEL, cpu_to_node(j));
6470 *per_cpu_ptr(sdd->sd, j) = sd;
6472 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6473 GFP_KERNEL, cpu_to_node(j));
6479 *per_cpu_ptr(sdd->sg, j) = sg;
6481 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6482 GFP_KERNEL, cpu_to_node(j));
6486 *per_cpu_ptr(sdd->sgc, j) = sgc;
6493 static void __sdt_free(const struct cpumask *cpu_map)
6495 struct sched_domain_topology_level *tl;
6498 for_each_sd_topology(tl) {
6499 struct sd_data *sdd = &tl->data;
6501 for_each_cpu(j, cpu_map) {
6502 struct sched_domain *sd;
6505 sd = *per_cpu_ptr(sdd->sd, j);
6506 if (sd && (sd->flags & SD_OVERLAP))
6507 free_sched_groups(sd->groups, 0);
6508 kfree(*per_cpu_ptr(sdd->sd, j));
6512 kfree(*per_cpu_ptr(sdd->sg, j));
6514 kfree(*per_cpu_ptr(sdd->sgc, j));
6516 free_percpu(sdd->sd);
6518 free_percpu(sdd->sg);
6520 free_percpu(sdd->sgc);
6525 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6526 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6527 struct sched_domain *child, int cpu)
6529 struct sched_domain *sd = sd_init(tl, cpu);
6533 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6535 sd->level = child->level + 1;
6536 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6540 if (!cpumask_subset(sched_domain_span(child),
6541 sched_domain_span(sd))) {
6542 pr_err("BUG: arch topology borken\n");
6543 #ifdef CONFIG_SCHED_DEBUG
6544 pr_err(" the %s domain not a subset of the %s domain\n",
6545 child->name, sd->name);
6547 /* Fixup, ensure @sd has at least @child cpus. */
6548 cpumask_or(sched_domain_span(sd),
6549 sched_domain_span(sd),
6550 sched_domain_span(child));
6554 set_domain_attribute(sd, attr);
6560 * Build sched domains for a given set of cpus and attach the sched domains
6561 * to the individual cpus
6563 static int build_sched_domains(const struct cpumask *cpu_map,
6564 struct sched_domain_attr *attr)
6566 enum s_alloc alloc_state;
6567 struct sched_domain *sd;
6569 int i, ret = -ENOMEM;
6571 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6572 if (alloc_state != sa_rootdomain)
6575 /* Set up domains for cpus specified by the cpu_map. */
6576 for_each_cpu(i, cpu_map) {
6577 struct sched_domain_topology_level *tl;
6580 for_each_sd_topology(tl) {
6581 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6582 if (tl == sched_domain_topology)
6583 *per_cpu_ptr(d.sd, i) = sd;
6584 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6585 sd->flags |= SD_OVERLAP;
6586 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6591 /* Build the groups for the domains */
6592 for_each_cpu(i, cpu_map) {
6593 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6594 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6595 if (sd->flags & SD_OVERLAP) {
6596 if (build_overlap_sched_groups(sd, i))
6599 if (build_sched_groups(sd, i))
6605 /* Calculate CPU capacity for physical packages and nodes */
6606 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6607 if (!cpumask_test_cpu(i, cpu_map))
6610 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6611 claim_allocations(i, sd);
6612 init_sched_groups_capacity(i, sd);
6616 /* Attach the domains */
6618 for_each_cpu(i, cpu_map) {
6619 sd = *per_cpu_ptr(d.sd, i);
6620 cpu_attach_domain(sd, d.rd, i);
6626 __free_domain_allocs(&d, alloc_state, cpu_map);
6630 static cpumask_var_t *doms_cur; /* current sched domains */
6631 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6632 static struct sched_domain_attr *dattr_cur;
6633 /* attribues of custom domains in 'doms_cur' */
6636 * Special case: If a kmalloc of a doms_cur partition (array of
6637 * cpumask) fails, then fallback to a single sched domain,
6638 * as determined by the single cpumask fallback_doms.
6640 static cpumask_var_t fallback_doms;
6643 * arch_update_cpu_topology lets virtualized architectures update the
6644 * cpu core maps. It is supposed to return 1 if the topology changed
6645 * or 0 if it stayed the same.
6647 int __weak arch_update_cpu_topology(void)
6652 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6655 cpumask_var_t *doms;
6657 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6660 for (i = 0; i < ndoms; i++) {
6661 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6662 free_sched_domains(doms, i);
6669 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6672 for (i = 0; i < ndoms; i++)
6673 free_cpumask_var(doms[i]);
6678 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6679 * For now this just excludes isolated cpus, but could be used to
6680 * exclude other special cases in the future.
6682 static int init_sched_domains(const struct cpumask *cpu_map)
6686 arch_update_cpu_topology();
6688 doms_cur = alloc_sched_domains(ndoms_cur);
6690 doms_cur = &fallback_doms;
6691 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6692 err = build_sched_domains(doms_cur[0], NULL);
6693 register_sched_domain_sysctl();
6699 * Detach sched domains from a group of cpus specified in cpu_map
6700 * These cpus will now be attached to the NULL domain
6702 static void detach_destroy_domains(const struct cpumask *cpu_map)
6707 for_each_cpu(i, cpu_map)
6708 cpu_attach_domain(NULL, &def_root_domain, i);
6712 /* handle null as "default" */
6713 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6714 struct sched_domain_attr *new, int idx_new)
6716 struct sched_domain_attr tmp;
6723 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6724 new ? (new + idx_new) : &tmp,
6725 sizeof(struct sched_domain_attr));
6729 * Partition sched domains as specified by the 'ndoms_new'
6730 * cpumasks in the array doms_new[] of cpumasks. This compares
6731 * doms_new[] to the current sched domain partitioning, doms_cur[].
6732 * It destroys each deleted domain and builds each new domain.
6734 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6735 * The masks don't intersect (don't overlap.) We should setup one
6736 * sched domain for each mask. CPUs not in any of the cpumasks will
6737 * not be load balanced. If the same cpumask appears both in the
6738 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6741 * The passed in 'doms_new' should be allocated using
6742 * alloc_sched_domains. This routine takes ownership of it and will
6743 * free_sched_domains it when done with it. If the caller failed the
6744 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6745 * and partition_sched_domains() will fallback to the single partition
6746 * 'fallback_doms', it also forces the domains to be rebuilt.
6748 * If doms_new == NULL it will be replaced with cpu_online_mask.
6749 * ndoms_new == 0 is a special case for destroying existing domains,
6750 * and it will not create the default domain.
6752 * Call with hotplug lock held
6754 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6755 struct sched_domain_attr *dattr_new)
6760 mutex_lock(&sched_domains_mutex);
6762 /* always unregister in case we don't destroy any domains */
6763 unregister_sched_domain_sysctl();
6765 /* Let architecture update cpu core mappings. */
6766 new_topology = arch_update_cpu_topology();
6768 n = doms_new ? ndoms_new : 0;
6770 /* Destroy deleted domains */
6771 for (i = 0; i < ndoms_cur; i++) {
6772 for (j = 0; j < n && !new_topology; j++) {
6773 if (cpumask_equal(doms_cur[i], doms_new[j])
6774 && dattrs_equal(dattr_cur, i, dattr_new, j))
6777 /* no match - a current sched domain not in new doms_new[] */
6778 detach_destroy_domains(doms_cur[i]);
6784 if (doms_new == NULL) {
6786 doms_new = &fallback_doms;
6787 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6788 WARN_ON_ONCE(dattr_new);
6791 /* Build new domains */
6792 for (i = 0; i < ndoms_new; i++) {
6793 for (j = 0; j < n && !new_topology; j++) {
6794 if (cpumask_equal(doms_new[i], doms_cur[j])
6795 && dattrs_equal(dattr_new, i, dattr_cur, j))
6798 /* no match - add a new doms_new */
6799 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6804 /* Remember the new sched domains */
6805 if (doms_cur != &fallback_doms)
6806 free_sched_domains(doms_cur, ndoms_cur);
6807 kfree(dattr_cur); /* kfree(NULL) is safe */
6808 doms_cur = doms_new;
6809 dattr_cur = dattr_new;
6810 ndoms_cur = ndoms_new;
6812 register_sched_domain_sysctl();
6814 mutex_unlock(&sched_domains_mutex);
6817 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6820 * Update cpusets according to cpu_active mask. If cpusets are
6821 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6822 * around partition_sched_domains().
6824 * If we come here as part of a suspend/resume, don't touch cpusets because we
6825 * want to restore it back to its original state upon resume anyway.
6827 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6831 case CPU_ONLINE_FROZEN:
6832 case CPU_DOWN_FAILED_FROZEN:
6835 * num_cpus_frozen tracks how many CPUs are involved in suspend
6836 * resume sequence. As long as this is not the last online
6837 * operation in the resume sequence, just build a single sched
6838 * domain, ignoring cpusets.
6841 if (likely(num_cpus_frozen)) {
6842 partition_sched_domains(1, NULL, NULL);
6847 * This is the last CPU online operation. So fall through and
6848 * restore the original sched domains by considering the
6849 * cpuset configurations.
6853 case CPU_DOWN_FAILED:
6854 cpuset_update_active_cpus(true);
6862 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6866 case CPU_DOWN_PREPARE:
6867 cpuset_update_active_cpus(false);
6869 case CPU_DOWN_PREPARE_FROZEN:
6871 partition_sched_domains(1, NULL, NULL);
6879 void __init sched_init_smp(void)
6881 cpumask_var_t non_isolated_cpus;
6883 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6884 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6889 * There's no userspace yet to cause hotplug operations; hence all the
6890 * cpu masks are stable and all blatant races in the below code cannot
6893 mutex_lock(&sched_domains_mutex);
6894 init_sched_domains(cpu_active_mask);
6895 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6896 if (cpumask_empty(non_isolated_cpus))
6897 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6898 mutex_unlock(&sched_domains_mutex);
6900 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6901 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6902 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6906 /* Move init over to a non-isolated CPU */
6907 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6909 sched_init_granularity();
6910 free_cpumask_var(non_isolated_cpus);
6912 init_sched_rt_class();
6913 init_sched_dl_class();
6916 void __init sched_init_smp(void)
6918 sched_init_granularity();
6920 #endif /* CONFIG_SMP */
6922 const_debug unsigned int sysctl_timer_migration = 1;
6924 int in_sched_functions(unsigned long addr)
6926 return in_lock_functions(addr) ||
6927 (addr >= (unsigned long)__sched_text_start
6928 && addr < (unsigned long)__sched_text_end);
6931 #ifdef CONFIG_CGROUP_SCHED
6933 * Default task group.
6934 * Every task in system belongs to this group at bootup.
6936 struct task_group root_task_group;
6937 LIST_HEAD(task_groups);
6940 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6942 void __init sched_init(void)
6945 unsigned long alloc_size = 0, ptr;
6947 #ifdef CONFIG_FAIR_GROUP_SCHED
6948 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6950 #ifdef CONFIG_RT_GROUP_SCHED
6951 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6953 #ifdef CONFIG_CPUMASK_OFFSTACK
6954 alloc_size += num_possible_cpus() * cpumask_size();
6957 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6959 #ifdef CONFIG_FAIR_GROUP_SCHED
6960 root_task_group.se = (struct sched_entity **)ptr;
6961 ptr += nr_cpu_ids * sizeof(void **);
6963 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6964 ptr += nr_cpu_ids * sizeof(void **);
6966 #endif /* CONFIG_FAIR_GROUP_SCHED */
6967 #ifdef CONFIG_RT_GROUP_SCHED
6968 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6969 ptr += nr_cpu_ids * sizeof(void **);
6971 root_task_group.rt_rq = (struct rt_rq **)ptr;
6972 ptr += nr_cpu_ids * sizeof(void **);
6974 #endif /* CONFIG_RT_GROUP_SCHED */
6975 #ifdef CONFIG_CPUMASK_OFFSTACK
6976 for_each_possible_cpu(i) {
6977 per_cpu(load_balance_mask, i) = (void *)ptr;
6978 ptr += cpumask_size();
6980 #endif /* CONFIG_CPUMASK_OFFSTACK */
6983 init_rt_bandwidth(&def_rt_bandwidth,
6984 global_rt_period(), global_rt_runtime());
6985 init_dl_bandwidth(&def_dl_bandwidth,
6986 global_rt_period(), global_rt_runtime());
6989 init_defrootdomain();
6992 #ifdef CONFIG_RT_GROUP_SCHED
6993 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6994 global_rt_period(), global_rt_runtime());
6995 #endif /* CONFIG_RT_GROUP_SCHED */
6997 #ifdef CONFIG_CGROUP_SCHED
6998 list_add(&root_task_group.list, &task_groups);
6999 INIT_LIST_HEAD(&root_task_group.children);
7000 INIT_LIST_HEAD(&root_task_group.siblings);
7001 autogroup_init(&init_task);
7003 #endif /* CONFIG_CGROUP_SCHED */
7005 for_each_possible_cpu(i) {
7009 raw_spin_lock_init(&rq->lock);
7011 rq->calc_load_active = 0;
7012 rq->calc_load_update = jiffies + LOAD_FREQ;
7013 init_cfs_rq(&rq->cfs);
7014 init_rt_rq(&rq->rt, rq);
7015 init_dl_rq(&rq->dl, rq);
7016 #ifdef CONFIG_FAIR_GROUP_SCHED
7017 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7018 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7020 * How much cpu bandwidth does root_task_group get?
7022 * In case of task-groups formed thr' the cgroup filesystem, it
7023 * gets 100% of the cpu resources in the system. This overall
7024 * system cpu resource is divided among the tasks of
7025 * root_task_group and its child task-groups in a fair manner,
7026 * based on each entity's (task or task-group's) weight
7027 * (se->load.weight).
7029 * In other words, if root_task_group has 10 tasks of weight
7030 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7031 * then A0's share of the cpu resource is:
7033 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7035 * We achieve this by letting root_task_group's tasks sit
7036 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7038 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7039 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7040 #endif /* CONFIG_FAIR_GROUP_SCHED */
7042 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7043 #ifdef CONFIG_RT_GROUP_SCHED
7044 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7047 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7048 rq->cpu_load[j] = 0;
7050 rq->last_load_update_tick = jiffies;
7055 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7056 rq->post_schedule = 0;
7057 rq->active_balance = 0;
7058 rq->next_balance = jiffies;
7063 rq->avg_idle = 2*sysctl_sched_migration_cost;
7064 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7066 INIT_LIST_HEAD(&rq->cfs_tasks);
7068 rq_attach_root(rq, &def_root_domain);
7069 #ifdef CONFIG_NO_HZ_COMMON
7072 #ifdef CONFIG_NO_HZ_FULL
7073 rq->last_sched_tick = 0;
7077 atomic_set(&rq->nr_iowait, 0);
7080 set_load_weight(&init_task);
7082 #ifdef CONFIG_PREEMPT_NOTIFIERS
7083 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7087 * The boot idle thread does lazy MMU switching as well:
7089 atomic_inc(&init_mm.mm_count);
7090 enter_lazy_tlb(&init_mm, current);
7093 * Make us the idle thread. Technically, schedule() should not be
7094 * called from this thread, however somewhere below it might be,
7095 * but because we are the idle thread, we just pick up running again
7096 * when this runqueue becomes "idle".
7098 init_idle(current, smp_processor_id());
7100 calc_load_update = jiffies + LOAD_FREQ;
7103 * During early bootup we pretend to be a normal task:
7105 current->sched_class = &fair_sched_class;
7108 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7109 /* May be allocated at isolcpus cmdline parse time */
7110 if (cpu_isolated_map == NULL)
7111 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7112 idle_thread_set_boot_cpu();
7113 set_cpu_rq_start_time();
7115 init_sched_fair_class();
7117 scheduler_running = 1;
7120 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7121 static inline int preempt_count_equals(int preempt_offset)
7123 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7125 return (nested == preempt_offset);
7128 void __might_sleep(const char *file, int line, int preempt_offset)
7130 static unsigned long prev_jiffy; /* ratelimiting */
7132 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7133 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7134 !is_idle_task(current)) ||
7135 system_state != SYSTEM_RUNNING || oops_in_progress)
7137 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7139 prev_jiffy = jiffies;
7142 "BUG: sleeping function called from invalid context at %s:%d\n",
7145 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7146 in_atomic(), irqs_disabled(),
7147 current->pid, current->comm);
7149 debug_show_held_locks(current);
7150 if (irqs_disabled())
7151 print_irqtrace_events(current);
7152 #ifdef CONFIG_DEBUG_PREEMPT
7153 if (!preempt_count_equals(preempt_offset)) {
7154 pr_err("Preemption disabled at:");
7155 print_ip_sym(current->preempt_disable_ip);
7161 EXPORT_SYMBOL(__might_sleep);
7164 #ifdef CONFIG_MAGIC_SYSRQ
7165 static void normalize_task(struct rq *rq, struct task_struct *p)
7167 const struct sched_class *prev_class = p->sched_class;
7168 struct sched_attr attr = {
7169 .sched_policy = SCHED_NORMAL,
7171 int old_prio = p->prio;
7174 queued = task_on_rq_queued(p);
7176 dequeue_task(rq, p, 0);
7177 __setscheduler(rq, p, &attr);
7179 enqueue_task(rq, p, 0);
7183 check_class_changed(rq, p, prev_class, old_prio);
7186 void normalize_rt_tasks(void)
7188 struct task_struct *g, *p;
7189 unsigned long flags;
7192 read_lock_irqsave(&tasklist_lock, flags);
7193 for_each_process_thread(g, p) {
7195 * Only normalize user tasks:
7200 p->se.exec_start = 0;
7201 #ifdef CONFIG_SCHEDSTATS
7202 p->se.statistics.wait_start = 0;
7203 p->se.statistics.sleep_start = 0;
7204 p->se.statistics.block_start = 0;
7207 if (!dl_task(p) && !rt_task(p)) {
7209 * Renice negative nice level userspace
7212 if (task_nice(p) < 0 && p->mm)
7213 set_user_nice(p, 0);
7217 raw_spin_lock(&p->pi_lock);
7218 rq = __task_rq_lock(p);
7220 normalize_task(rq, p);
7222 __task_rq_unlock(rq);
7223 raw_spin_unlock(&p->pi_lock);
7225 read_unlock_irqrestore(&tasklist_lock, flags);
7228 #endif /* CONFIG_MAGIC_SYSRQ */
7230 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7232 * These functions are only useful for the IA64 MCA handling, or kdb.
7234 * They can only be called when the whole system has been
7235 * stopped - every CPU needs to be quiescent, and no scheduling
7236 * activity can take place. Using them for anything else would
7237 * be a serious bug, and as a result, they aren't even visible
7238 * under any other configuration.
7242 * curr_task - return the current task for a given cpu.
7243 * @cpu: the processor in question.
7245 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7247 * Return: The current task for @cpu.
7249 struct task_struct *curr_task(int cpu)
7251 return cpu_curr(cpu);
7254 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7258 * set_curr_task - set the current task for a given cpu.
7259 * @cpu: the processor in question.
7260 * @p: the task pointer to set.
7262 * Description: This function must only be used when non-maskable interrupts
7263 * are serviced on a separate stack. It allows the architecture to switch the
7264 * notion of the current task on a cpu in a non-blocking manner. This function
7265 * must be called with all CPU's synchronized, and interrupts disabled, the
7266 * and caller must save the original value of the current task (see
7267 * curr_task() above) and restore that value before reenabling interrupts and
7268 * re-starting the system.
7270 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7272 void set_curr_task(int cpu, struct task_struct *p)
7279 #ifdef CONFIG_CGROUP_SCHED
7280 /* task_group_lock serializes the addition/removal of task groups */
7281 static DEFINE_SPINLOCK(task_group_lock);
7283 static void free_sched_group(struct task_group *tg)
7285 free_fair_sched_group(tg);
7286 free_rt_sched_group(tg);
7291 /* allocate runqueue etc for a new task group */
7292 struct task_group *sched_create_group(struct task_group *parent)
7294 struct task_group *tg;
7296 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7298 return ERR_PTR(-ENOMEM);
7300 if (!alloc_fair_sched_group(tg, parent))
7303 if (!alloc_rt_sched_group(tg, parent))
7309 free_sched_group(tg);
7310 return ERR_PTR(-ENOMEM);
7313 void sched_online_group(struct task_group *tg, struct task_group *parent)
7315 unsigned long flags;
7317 spin_lock_irqsave(&task_group_lock, flags);
7318 list_add_rcu(&tg->list, &task_groups);
7320 WARN_ON(!parent); /* root should already exist */
7322 tg->parent = parent;
7323 INIT_LIST_HEAD(&tg->children);
7324 list_add_rcu(&tg->siblings, &parent->children);
7325 spin_unlock_irqrestore(&task_group_lock, flags);
7328 /* rcu callback to free various structures associated with a task group */
7329 static void free_sched_group_rcu(struct rcu_head *rhp)
7331 /* now it should be safe to free those cfs_rqs */
7332 free_sched_group(container_of(rhp, struct task_group, rcu));
7335 /* Destroy runqueue etc associated with a task group */
7336 void sched_destroy_group(struct task_group *tg)
7338 /* wait for possible concurrent references to cfs_rqs complete */
7339 call_rcu(&tg->rcu, free_sched_group_rcu);
7342 void sched_offline_group(struct task_group *tg)
7344 unsigned long flags;
7347 /* end participation in shares distribution */
7348 for_each_possible_cpu(i)
7349 unregister_fair_sched_group(tg, i);
7351 spin_lock_irqsave(&task_group_lock, flags);
7352 list_del_rcu(&tg->list);
7353 list_del_rcu(&tg->siblings);
7354 spin_unlock_irqrestore(&task_group_lock, flags);
7357 /* change task's runqueue when it moves between groups.
7358 * The caller of this function should have put the task in its new group
7359 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7360 * reflect its new group.
7362 void sched_move_task(struct task_struct *tsk)
7364 struct task_group *tg;
7365 int queued, running;
7366 unsigned long flags;
7369 rq = task_rq_lock(tsk, &flags);
7371 running = task_current(rq, tsk);
7372 queued = task_on_rq_queued(tsk);
7375 dequeue_task(rq, tsk, 0);
7376 if (unlikely(running))
7377 put_prev_task(rq, tsk);
7379 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7380 lockdep_is_held(&tsk->sighand->siglock)),
7381 struct task_group, css);
7382 tg = autogroup_task_group(tsk, tg);
7383 tsk->sched_task_group = tg;
7385 #ifdef CONFIG_FAIR_GROUP_SCHED
7386 if (tsk->sched_class->task_move_group)
7387 tsk->sched_class->task_move_group(tsk, queued);
7390 set_task_rq(tsk, task_cpu(tsk));
7392 if (unlikely(running))
7393 tsk->sched_class->set_curr_task(rq);
7395 enqueue_task(rq, tsk, 0);
7397 task_rq_unlock(rq, tsk, &flags);
7399 #endif /* CONFIG_CGROUP_SCHED */
7401 #ifdef CONFIG_RT_GROUP_SCHED
7403 * Ensure that the real time constraints are schedulable.
7405 static DEFINE_MUTEX(rt_constraints_mutex);
7407 /* Must be called with tasklist_lock held */
7408 static inline int tg_has_rt_tasks(struct task_group *tg)
7410 struct task_struct *g, *p;
7412 for_each_process_thread(g, p) {
7413 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7420 struct rt_schedulable_data {
7421 struct task_group *tg;
7426 static int tg_rt_schedulable(struct task_group *tg, void *data)
7428 struct rt_schedulable_data *d = data;
7429 struct task_group *child;
7430 unsigned long total, sum = 0;
7431 u64 period, runtime;
7433 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7434 runtime = tg->rt_bandwidth.rt_runtime;
7437 period = d->rt_period;
7438 runtime = d->rt_runtime;
7442 * Cannot have more runtime than the period.
7444 if (runtime > period && runtime != RUNTIME_INF)
7448 * Ensure we don't starve existing RT tasks.
7450 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7453 total = to_ratio(period, runtime);
7456 * Nobody can have more than the global setting allows.
7458 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7462 * The sum of our children's runtime should not exceed our own.
7464 list_for_each_entry_rcu(child, &tg->children, siblings) {
7465 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7466 runtime = child->rt_bandwidth.rt_runtime;
7468 if (child == d->tg) {
7469 period = d->rt_period;
7470 runtime = d->rt_runtime;
7473 sum += to_ratio(period, runtime);
7482 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7486 struct rt_schedulable_data data = {
7488 .rt_period = period,
7489 .rt_runtime = runtime,
7493 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7499 static int tg_set_rt_bandwidth(struct task_group *tg,
7500 u64 rt_period, u64 rt_runtime)
7504 mutex_lock(&rt_constraints_mutex);
7505 read_lock(&tasklist_lock);
7506 err = __rt_schedulable(tg, rt_period, rt_runtime);
7510 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7511 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7512 tg->rt_bandwidth.rt_runtime = rt_runtime;
7514 for_each_possible_cpu(i) {
7515 struct rt_rq *rt_rq = tg->rt_rq[i];
7517 raw_spin_lock(&rt_rq->rt_runtime_lock);
7518 rt_rq->rt_runtime = rt_runtime;
7519 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7521 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7523 read_unlock(&tasklist_lock);
7524 mutex_unlock(&rt_constraints_mutex);
7529 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7531 u64 rt_runtime, rt_period;
7533 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7534 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7535 if (rt_runtime_us < 0)
7536 rt_runtime = RUNTIME_INF;
7538 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7541 static long sched_group_rt_runtime(struct task_group *tg)
7545 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7548 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7549 do_div(rt_runtime_us, NSEC_PER_USEC);
7550 return rt_runtime_us;
7553 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7555 u64 rt_runtime, rt_period;
7557 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7558 rt_runtime = tg->rt_bandwidth.rt_runtime;
7563 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7566 static long sched_group_rt_period(struct task_group *tg)
7570 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7571 do_div(rt_period_us, NSEC_PER_USEC);
7572 return rt_period_us;
7574 #endif /* CONFIG_RT_GROUP_SCHED */
7576 #ifdef CONFIG_RT_GROUP_SCHED
7577 static int sched_rt_global_constraints(void)
7581 mutex_lock(&rt_constraints_mutex);
7582 read_lock(&tasklist_lock);
7583 ret = __rt_schedulable(NULL, 0, 0);
7584 read_unlock(&tasklist_lock);
7585 mutex_unlock(&rt_constraints_mutex);
7590 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7592 /* Don't accept realtime tasks when there is no way for them to run */
7593 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7599 #else /* !CONFIG_RT_GROUP_SCHED */
7600 static int sched_rt_global_constraints(void)
7602 unsigned long flags;
7605 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7606 for_each_possible_cpu(i) {
7607 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7609 raw_spin_lock(&rt_rq->rt_runtime_lock);
7610 rt_rq->rt_runtime = global_rt_runtime();
7611 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7613 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7617 #endif /* CONFIG_RT_GROUP_SCHED */
7619 static int sched_dl_global_constraints(void)
7621 u64 runtime = global_rt_runtime();
7622 u64 period = global_rt_period();
7623 u64 new_bw = to_ratio(period, runtime);
7625 unsigned long flags;
7628 * Here we want to check the bandwidth not being set to some
7629 * value smaller than the currently allocated bandwidth in
7630 * any of the root_domains.
7632 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7633 * cycling on root_domains... Discussion on different/better
7634 * solutions is welcome!
7636 for_each_possible_cpu(cpu) {
7637 struct dl_bw *dl_b = dl_bw_of(cpu);
7639 raw_spin_lock_irqsave(&dl_b->lock, flags);
7640 if (new_bw < dl_b->total_bw)
7642 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7651 static void sched_dl_do_global(void)
7655 unsigned long flags;
7657 def_dl_bandwidth.dl_period = global_rt_period();
7658 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7660 if (global_rt_runtime() != RUNTIME_INF)
7661 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7664 * FIXME: As above...
7666 for_each_possible_cpu(cpu) {
7667 struct dl_bw *dl_b = dl_bw_of(cpu);
7669 raw_spin_lock_irqsave(&dl_b->lock, flags);
7671 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7675 static int sched_rt_global_validate(void)
7677 if (sysctl_sched_rt_period <= 0)
7680 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7681 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7687 static void sched_rt_do_global(void)
7689 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7690 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7693 int sched_rt_handler(struct ctl_table *table, int write,
7694 void __user *buffer, size_t *lenp,
7697 int old_period, old_runtime;
7698 static DEFINE_MUTEX(mutex);
7702 old_period = sysctl_sched_rt_period;
7703 old_runtime = sysctl_sched_rt_runtime;
7705 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7707 if (!ret && write) {
7708 ret = sched_rt_global_validate();
7712 ret = sched_rt_global_constraints();
7716 ret = sched_dl_global_constraints();
7720 sched_rt_do_global();
7721 sched_dl_do_global();
7725 sysctl_sched_rt_period = old_period;
7726 sysctl_sched_rt_runtime = old_runtime;
7728 mutex_unlock(&mutex);
7733 int sched_rr_handler(struct ctl_table *table, int write,
7734 void __user *buffer, size_t *lenp,
7738 static DEFINE_MUTEX(mutex);
7741 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7742 /* make sure that internally we keep jiffies */
7743 /* also, writing zero resets timeslice to default */
7744 if (!ret && write) {
7745 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7746 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7748 mutex_unlock(&mutex);
7752 #ifdef CONFIG_CGROUP_SCHED
7754 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7756 return css ? container_of(css, struct task_group, css) : NULL;
7759 static struct cgroup_subsys_state *
7760 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7762 struct task_group *parent = css_tg(parent_css);
7763 struct task_group *tg;
7766 /* This is early initialization for the top cgroup */
7767 return &root_task_group.css;
7770 tg = sched_create_group(parent);
7772 return ERR_PTR(-ENOMEM);
7777 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7779 struct task_group *tg = css_tg(css);
7780 struct task_group *parent = css_tg(css->parent);
7783 sched_online_group(tg, parent);
7787 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7789 struct task_group *tg = css_tg(css);
7791 sched_destroy_group(tg);
7794 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7796 struct task_group *tg = css_tg(css);
7798 sched_offline_group(tg);
7801 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7802 struct cgroup_taskset *tset)
7804 struct task_struct *task;
7806 cgroup_taskset_for_each(task, tset) {
7807 #ifdef CONFIG_RT_GROUP_SCHED
7808 if (!sched_rt_can_attach(css_tg(css), task))
7811 /* We don't support RT-tasks being in separate groups */
7812 if (task->sched_class != &fair_sched_class)
7819 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7820 struct cgroup_taskset *tset)
7822 struct task_struct *task;
7824 cgroup_taskset_for_each(task, tset)
7825 sched_move_task(task);
7828 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7829 struct cgroup_subsys_state *old_css,
7830 struct task_struct *task)
7833 * cgroup_exit() is called in the copy_process() failure path.
7834 * Ignore this case since the task hasn't ran yet, this avoids
7835 * trying to poke a half freed task state from generic code.
7837 if (!(task->flags & PF_EXITING))
7840 sched_move_task(task);
7843 #ifdef CONFIG_FAIR_GROUP_SCHED
7844 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7845 struct cftype *cftype, u64 shareval)
7847 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7850 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7853 struct task_group *tg = css_tg(css);
7855 return (u64) scale_load_down(tg->shares);
7858 #ifdef CONFIG_CFS_BANDWIDTH
7859 static DEFINE_MUTEX(cfs_constraints_mutex);
7861 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7862 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7864 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7866 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7868 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7869 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7871 if (tg == &root_task_group)
7875 * Ensure we have at some amount of bandwidth every period. This is
7876 * to prevent reaching a state of large arrears when throttled via
7877 * entity_tick() resulting in prolonged exit starvation.
7879 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7883 * Likewise, bound things on the otherside by preventing insane quota
7884 * periods. This also allows us to normalize in computing quota
7887 if (period > max_cfs_quota_period)
7891 * Prevent race between setting of cfs_rq->runtime_enabled and
7892 * unthrottle_offline_cfs_rqs().
7895 mutex_lock(&cfs_constraints_mutex);
7896 ret = __cfs_schedulable(tg, period, quota);
7900 runtime_enabled = quota != RUNTIME_INF;
7901 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7903 * If we need to toggle cfs_bandwidth_used, off->on must occur
7904 * before making related changes, and on->off must occur afterwards
7906 if (runtime_enabled && !runtime_was_enabled)
7907 cfs_bandwidth_usage_inc();
7908 raw_spin_lock_irq(&cfs_b->lock);
7909 cfs_b->period = ns_to_ktime(period);
7910 cfs_b->quota = quota;
7912 __refill_cfs_bandwidth_runtime(cfs_b);
7913 /* restart the period timer (if active) to handle new period expiry */
7914 if (runtime_enabled && cfs_b->timer_active) {
7915 /* force a reprogram */
7916 __start_cfs_bandwidth(cfs_b, true);
7918 raw_spin_unlock_irq(&cfs_b->lock);
7920 for_each_online_cpu(i) {
7921 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7922 struct rq *rq = cfs_rq->rq;
7924 raw_spin_lock_irq(&rq->lock);
7925 cfs_rq->runtime_enabled = runtime_enabled;
7926 cfs_rq->runtime_remaining = 0;
7928 if (cfs_rq->throttled)
7929 unthrottle_cfs_rq(cfs_rq);
7930 raw_spin_unlock_irq(&rq->lock);
7932 if (runtime_was_enabled && !runtime_enabled)
7933 cfs_bandwidth_usage_dec();
7935 mutex_unlock(&cfs_constraints_mutex);
7941 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7945 period = ktime_to_ns(tg->cfs_bandwidth.period);
7946 if (cfs_quota_us < 0)
7947 quota = RUNTIME_INF;
7949 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7951 return tg_set_cfs_bandwidth(tg, period, quota);
7954 long tg_get_cfs_quota(struct task_group *tg)
7958 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7961 quota_us = tg->cfs_bandwidth.quota;
7962 do_div(quota_us, NSEC_PER_USEC);
7967 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7971 period = (u64)cfs_period_us * NSEC_PER_USEC;
7972 quota = tg->cfs_bandwidth.quota;
7974 return tg_set_cfs_bandwidth(tg, period, quota);
7977 long tg_get_cfs_period(struct task_group *tg)
7981 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7982 do_div(cfs_period_us, NSEC_PER_USEC);
7984 return cfs_period_us;
7987 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7990 return tg_get_cfs_quota(css_tg(css));
7993 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7994 struct cftype *cftype, s64 cfs_quota_us)
7996 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7999 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8002 return tg_get_cfs_period(css_tg(css));
8005 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8006 struct cftype *cftype, u64 cfs_period_us)
8008 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8011 struct cfs_schedulable_data {
8012 struct task_group *tg;
8017 * normalize group quota/period to be quota/max_period
8018 * note: units are usecs
8020 static u64 normalize_cfs_quota(struct task_group *tg,
8021 struct cfs_schedulable_data *d)
8029 period = tg_get_cfs_period(tg);
8030 quota = tg_get_cfs_quota(tg);
8033 /* note: these should typically be equivalent */
8034 if (quota == RUNTIME_INF || quota == -1)
8037 return to_ratio(period, quota);
8040 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8042 struct cfs_schedulable_data *d = data;
8043 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8044 s64 quota = 0, parent_quota = -1;
8047 quota = RUNTIME_INF;
8049 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8051 quota = normalize_cfs_quota(tg, d);
8052 parent_quota = parent_b->hierarchal_quota;
8055 * ensure max(child_quota) <= parent_quota, inherit when no
8058 if (quota == RUNTIME_INF)
8059 quota = parent_quota;
8060 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8063 cfs_b->hierarchal_quota = quota;
8068 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8071 struct cfs_schedulable_data data = {
8077 if (quota != RUNTIME_INF) {
8078 do_div(data.period, NSEC_PER_USEC);
8079 do_div(data.quota, NSEC_PER_USEC);
8083 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8089 static int cpu_stats_show(struct seq_file *sf, void *v)
8091 struct task_group *tg = css_tg(seq_css(sf));
8092 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8094 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8095 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8096 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8100 #endif /* CONFIG_CFS_BANDWIDTH */
8101 #endif /* CONFIG_FAIR_GROUP_SCHED */
8103 #ifdef CONFIG_RT_GROUP_SCHED
8104 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8105 struct cftype *cft, s64 val)
8107 return sched_group_set_rt_runtime(css_tg(css), val);
8110 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8113 return sched_group_rt_runtime(css_tg(css));
8116 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8117 struct cftype *cftype, u64 rt_period_us)
8119 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8122 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8125 return sched_group_rt_period(css_tg(css));
8127 #endif /* CONFIG_RT_GROUP_SCHED */
8129 static struct cftype cpu_files[] = {
8130 #ifdef CONFIG_FAIR_GROUP_SCHED
8133 .read_u64 = cpu_shares_read_u64,
8134 .write_u64 = cpu_shares_write_u64,
8137 #ifdef CONFIG_CFS_BANDWIDTH
8139 .name = "cfs_quota_us",
8140 .read_s64 = cpu_cfs_quota_read_s64,
8141 .write_s64 = cpu_cfs_quota_write_s64,
8144 .name = "cfs_period_us",
8145 .read_u64 = cpu_cfs_period_read_u64,
8146 .write_u64 = cpu_cfs_period_write_u64,
8150 .seq_show = cpu_stats_show,
8153 #ifdef CONFIG_RT_GROUP_SCHED
8155 .name = "rt_runtime_us",
8156 .read_s64 = cpu_rt_runtime_read,
8157 .write_s64 = cpu_rt_runtime_write,
8160 .name = "rt_period_us",
8161 .read_u64 = cpu_rt_period_read_uint,
8162 .write_u64 = cpu_rt_period_write_uint,
8168 struct cgroup_subsys cpu_cgrp_subsys = {
8169 .css_alloc = cpu_cgroup_css_alloc,
8170 .css_free = cpu_cgroup_css_free,
8171 .css_online = cpu_cgroup_css_online,
8172 .css_offline = cpu_cgroup_css_offline,
8173 .can_attach = cpu_cgroup_can_attach,
8174 .attach = cpu_cgroup_attach,
8175 .exit = cpu_cgroup_exit,
8176 .legacy_cftypes = cpu_files,
8180 #endif /* CONFIG_CGROUP_SCHED */
8182 void dump_cpu_task(int cpu)
8184 pr_info("Task dump for CPU %d:\n", cpu);
8185 sched_show_task(cpu_curr(cpu));