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;
458 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
460 hrtimer_set_expires(timer, time);
462 if (rq == this_rq()) {
463 __hrtick_restart(rq);
464 } else if (!rq->hrtick_csd_pending) {
465 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
466 rq->hrtick_csd_pending = 1;
471 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
473 int cpu = (int)(long)hcpu;
476 case CPU_UP_CANCELED:
477 case CPU_UP_CANCELED_FROZEN:
478 case CPU_DOWN_PREPARE:
479 case CPU_DOWN_PREPARE_FROZEN:
481 case CPU_DEAD_FROZEN:
482 hrtick_clear(cpu_rq(cpu));
489 static __init void init_hrtick(void)
491 hotcpu_notifier(hotplug_hrtick, 0);
495 * Called to set the hrtick timer state.
497 * called with rq->lock held and irqs disabled
499 void hrtick_start(struct rq *rq, u64 delay)
501 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
502 HRTIMER_MODE_REL_PINNED, 0);
505 static inline void init_hrtick(void)
508 #endif /* CONFIG_SMP */
510 static void init_rq_hrtick(struct rq *rq)
513 rq->hrtick_csd_pending = 0;
515 rq->hrtick_csd.flags = 0;
516 rq->hrtick_csd.func = __hrtick_start;
517 rq->hrtick_csd.info = rq;
520 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
521 rq->hrtick_timer.function = hrtick;
523 #else /* CONFIG_SCHED_HRTICK */
524 static inline void hrtick_clear(struct rq *rq)
528 static inline void init_rq_hrtick(struct rq *rq)
532 static inline void init_hrtick(void)
535 #endif /* CONFIG_SCHED_HRTICK */
538 * cmpxchg based fetch_or, macro so it works for different integer types
540 #define fetch_or(ptr, val) \
541 ({ typeof(*(ptr)) __old, __val = *(ptr); \
543 __old = cmpxchg((ptr), __val, __val | (val)); \
544 if (__old == __val) \
551 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
553 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
554 * this avoids any races wrt polling state changes and thereby avoids
557 static bool set_nr_and_not_polling(struct task_struct *p)
559 struct thread_info *ti = task_thread_info(p);
560 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
564 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
566 * If this returns true, then the idle task promises to call
567 * sched_ttwu_pending() and reschedule soon.
569 static bool set_nr_if_polling(struct task_struct *p)
571 struct thread_info *ti = task_thread_info(p);
572 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
575 if (!(val & _TIF_POLLING_NRFLAG))
577 if (val & _TIF_NEED_RESCHED)
579 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
588 static bool set_nr_and_not_polling(struct task_struct *p)
590 set_tsk_need_resched(p);
595 static bool set_nr_if_polling(struct task_struct *p)
603 * resched_curr - mark rq's current task 'to be rescheduled now'.
605 * On UP this means the setting of the need_resched flag, on SMP it
606 * might also involve a cross-CPU call to trigger the scheduler on
609 void resched_curr(struct rq *rq)
611 struct task_struct *curr = rq->curr;
614 lockdep_assert_held(&rq->lock);
616 if (test_tsk_need_resched(curr))
621 if (cpu == smp_processor_id()) {
622 set_tsk_need_resched(curr);
623 set_preempt_need_resched();
627 if (set_nr_and_not_polling(curr))
628 smp_send_reschedule(cpu);
630 trace_sched_wake_idle_without_ipi(cpu);
633 void resched_cpu(int cpu)
635 struct rq *rq = cpu_rq(cpu);
638 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
641 raw_spin_unlock_irqrestore(&rq->lock, flags);
645 #ifdef CONFIG_NO_HZ_COMMON
647 * In the semi idle case, use the nearest busy cpu for migrating timers
648 * from an idle cpu. This is good for power-savings.
650 * We don't do similar optimization for completely idle system, as
651 * selecting an idle cpu will add more delays to the timers than intended
652 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
654 int get_nohz_timer_target(int pinned)
656 int cpu = smp_processor_id();
658 struct sched_domain *sd;
660 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
664 for_each_domain(cpu, sd) {
665 for_each_cpu(i, sched_domain_span(sd)) {
677 * When add_timer_on() enqueues a timer into the timer wheel of an
678 * idle CPU then this timer might expire before the next timer event
679 * which is scheduled to wake up that CPU. In case of a completely
680 * idle system the next event might even be infinite time into the
681 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
682 * leaves the inner idle loop so the newly added timer is taken into
683 * account when the CPU goes back to idle and evaluates the timer
684 * wheel for the next timer event.
686 static void wake_up_idle_cpu(int cpu)
688 struct rq *rq = cpu_rq(cpu);
690 if (cpu == smp_processor_id())
693 if (set_nr_and_not_polling(rq->idle))
694 smp_send_reschedule(cpu);
696 trace_sched_wake_idle_without_ipi(cpu);
699 static bool wake_up_full_nohz_cpu(int cpu)
702 * We just need the target to call irq_exit() and re-evaluate
703 * the next tick. The nohz full kick at least implies that.
704 * If needed we can still optimize that later with an
707 if (tick_nohz_full_cpu(cpu)) {
708 if (cpu != smp_processor_id() ||
709 tick_nohz_tick_stopped())
710 tick_nohz_full_kick_cpu(cpu);
717 void wake_up_nohz_cpu(int cpu)
719 if (!wake_up_full_nohz_cpu(cpu))
720 wake_up_idle_cpu(cpu);
723 static inline bool got_nohz_idle_kick(void)
725 int cpu = smp_processor_id();
727 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
730 if (idle_cpu(cpu) && !need_resched())
734 * We can't run Idle Load Balance on this CPU for this time so we
735 * cancel it and clear NOHZ_BALANCE_KICK
737 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
741 #else /* CONFIG_NO_HZ_COMMON */
743 static inline bool got_nohz_idle_kick(void)
748 #endif /* CONFIG_NO_HZ_COMMON */
750 #ifdef CONFIG_NO_HZ_FULL
751 bool sched_can_stop_tick(void)
754 * More than one running task need preemption.
755 * nr_running update is assumed to be visible
756 * after IPI is sent from wakers.
758 if (this_rq()->nr_running > 1)
763 #endif /* CONFIG_NO_HZ_FULL */
765 void sched_avg_update(struct rq *rq)
767 s64 period = sched_avg_period();
769 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
771 * Inline assembly required to prevent the compiler
772 * optimising this loop into a divmod call.
773 * See __iter_div_u64_rem() for another example of this.
775 asm("" : "+rm" (rq->age_stamp));
776 rq->age_stamp += period;
781 #endif /* CONFIG_SMP */
783 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
784 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
786 * Iterate task_group tree rooted at *from, calling @down when first entering a
787 * node and @up when leaving it for the final time.
789 * Caller must hold rcu_lock or sufficient equivalent.
791 int walk_tg_tree_from(struct task_group *from,
792 tg_visitor down, tg_visitor up, void *data)
794 struct task_group *parent, *child;
800 ret = (*down)(parent, data);
803 list_for_each_entry_rcu(child, &parent->children, siblings) {
810 ret = (*up)(parent, data);
811 if (ret || parent == from)
815 parent = parent->parent;
822 int tg_nop(struct task_group *tg, void *data)
828 static void set_load_weight(struct task_struct *p)
830 int prio = p->static_prio - MAX_RT_PRIO;
831 struct load_weight *load = &p->se.load;
834 * SCHED_IDLE tasks get minimal weight:
836 if (p->policy == SCHED_IDLE) {
837 load->weight = scale_load(WEIGHT_IDLEPRIO);
838 load->inv_weight = WMULT_IDLEPRIO;
842 load->weight = scale_load(prio_to_weight[prio]);
843 load->inv_weight = prio_to_wmult[prio];
846 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
849 sched_info_queued(rq, p);
850 p->sched_class->enqueue_task(rq, p, flags);
853 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
856 sched_info_dequeued(rq, p);
857 p->sched_class->dequeue_task(rq, p, flags);
860 void activate_task(struct rq *rq, struct task_struct *p, int flags)
862 if (task_contributes_to_load(p))
863 rq->nr_uninterruptible--;
865 enqueue_task(rq, p, flags);
868 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
870 if (task_contributes_to_load(p))
871 rq->nr_uninterruptible++;
873 dequeue_task(rq, p, flags);
876 static void update_rq_clock_task(struct rq *rq, s64 delta)
879 * In theory, the compile should just see 0 here, and optimize out the call
880 * to sched_rt_avg_update. But I don't trust it...
882 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
883 s64 steal = 0, irq_delta = 0;
885 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
886 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
889 * Since irq_time is only updated on {soft,}irq_exit, we might run into
890 * this case when a previous update_rq_clock() happened inside a
893 * When this happens, we stop ->clock_task and only update the
894 * prev_irq_time stamp to account for the part that fit, so that a next
895 * update will consume the rest. This ensures ->clock_task is
898 * It does however cause some slight miss-attribution of {soft,}irq
899 * time, a more accurate solution would be to update the irq_time using
900 * the current rq->clock timestamp, except that would require using
903 if (irq_delta > delta)
906 rq->prev_irq_time += irq_delta;
909 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
910 if (static_key_false((¶virt_steal_rq_enabled))) {
911 steal = paravirt_steal_clock(cpu_of(rq));
912 steal -= rq->prev_steal_time_rq;
914 if (unlikely(steal > delta))
917 rq->prev_steal_time_rq += steal;
922 rq->clock_task += delta;
924 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
925 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
926 sched_rt_avg_update(rq, irq_delta + steal);
930 void sched_set_stop_task(int cpu, struct task_struct *stop)
932 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
933 struct task_struct *old_stop = cpu_rq(cpu)->stop;
937 * Make it appear like a SCHED_FIFO task, its something
938 * userspace knows about and won't get confused about.
940 * Also, it will make PI more or less work without too
941 * much confusion -- but then, stop work should not
942 * rely on PI working anyway.
944 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
946 stop->sched_class = &stop_sched_class;
949 cpu_rq(cpu)->stop = stop;
953 * Reset it back to a normal scheduling class so that
954 * it can die in pieces.
956 old_stop->sched_class = &rt_sched_class;
961 * __normal_prio - return the priority that is based on the static prio
963 static inline int __normal_prio(struct task_struct *p)
965 return p->static_prio;
969 * Calculate the expected normal priority: i.e. priority
970 * without taking RT-inheritance into account. Might be
971 * boosted by interactivity modifiers. Changes upon fork,
972 * setprio syscalls, and whenever the interactivity
973 * estimator recalculates.
975 static inline int normal_prio(struct task_struct *p)
979 if (task_has_dl_policy(p))
980 prio = MAX_DL_PRIO-1;
981 else if (task_has_rt_policy(p))
982 prio = MAX_RT_PRIO-1 - p->rt_priority;
984 prio = __normal_prio(p);
989 * Calculate the current priority, i.e. the priority
990 * taken into account by the scheduler. This value might
991 * be boosted by RT tasks, or might be boosted by
992 * interactivity modifiers. Will be RT if the task got
993 * RT-boosted. If not then it returns p->normal_prio.
995 static int effective_prio(struct task_struct *p)
997 p->normal_prio = normal_prio(p);
999 * If we are RT tasks or we were boosted to RT priority,
1000 * keep the priority unchanged. Otherwise, update priority
1001 * to the normal priority:
1003 if (!rt_prio(p->prio))
1004 return p->normal_prio;
1009 * task_curr - is this task currently executing on a CPU?
1010 * @p: the task in question.
1012 * Return: 1 if the task is currently executing. 0 otherwise.
1014 inline int task_curr(const struct task_struct *p)
1016 return cpu_curr(task_cpu(p)) == p;
1019 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1020 const struct sched_class *prev_class,
1023 if (prev_class != p->sched_class) {
1024 if (prev_class->switched_from)
1025 prev_class->switched_from(rq, p);
1026 p->sched_class->switched_to(rq, p);
1027 } else if (oldprio != p->prio || dl_task(p))
1028 p->sched_class->prio_changed(rq, p, oldprio);
1031 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1033 const struct sched_class *class;
1035 if (p->sched_class == rq->curr->sched_class) {
1036 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1038 for_each_class(class) {
1039 if (class == rq->curr->sched_class)
1041 if (class == p->sched_class) {
1049 * A queue event has occurred, and we're going to schedule. In
1050 * this case, we can save a useless back to back clock update.
1052 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1053 rq->skip_clock_update = 1;
1057 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1059 #ifdef CONFIG_SCHED_DEBUG
1061 * We should never call set_task_cpu() on a blocked task,
1062 * ttwu() will sort out the placement.
1064 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1065 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1067 #ifdef CONFIG_LOCKDEP
1069 * The caller should hold either p->pi_lock or rq->lock, when changing
1070 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1072 * sched_move_task() holds both and thus holding either pins the cgroup,
1075 * Furthermore, all task_rq users should acquire both locks, see
1078 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1079 lockdep_is_held(&task_rq(p)->lock)));
1083 trace_sched_migrate_task(p, new_cpu);
1085 if (task_cpu(p) != new_cpu) {
1086 if (p->sched_class->migrate_task_rq)
1087 p->sched_class->migrate_task_rq(p, new_cpu);
1088 p->se.nr_migrations++;
1089 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1092 __set_task_cpu(p, new_cpu);
1095 static void __migrate_swap_task(struct task_struct *p, int cpu)
1097 if (task_on_rq_queued(p)) {
1098 struct rq *src_rq, *dst_rq;
1100 src_rq = task_rq(p);
1101 dst_rq = cpu_rq(cpu);
1103 deactivate_task(src_rq, p, 0);
1104 set_task_cpu(p, cpu);
1105 activate_task(dst_rq, p, 0);
1106 check_preempt_curr(dst_rq, p, 0);
1109 * Task isn't running anymore; make it appear like we migrated
1110 * it before it went to sleep. This means on wakeup we make the
1111 * previous cpu our targer instead of where it really is.
1117 struct migration_swap_arg {
1118 struct task_struct *src_task, *dst_task;
1119 int src_cpu, dst_cpu;
1122 static int migrate_swap_stop(void *data)
1124 struct migration_swap_arg *arg = data;
1125 struct rq *src_rq, *dst_rq;
1128 src_rq = cpu_rq(arg->src_cpu);
1129 dst_rq = cpu_rq(arg->dst_cpu);
1131 double_raw_lock(&arg->src_task->pi_lock,
1132 &arg->dst_task->pi_lock);
1133 double_rq_lock(src_rq, dst_rq);
1134 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1137 if (task_cpu(arg->src_task) != arg->src_cpu)
1140 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1143 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1146 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1147 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1152 double_rq_unlock(src_rq, dst_rq);
1153 raw_spin_unlock(&arg->dst_task->pi_lock);
1154 raw_spin_unlock(&arg->src_task->pi_lock);
1160 * Cross migrate two tasks
1162 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1164 struct migration_swap_arg arg;
1167 arg = (struct migration_swap_arg){
1169 .src_cpu = task_cpu(cur),
1171 .dst_cpu = task_cpu(p),
1174 if (arg.src_cpu == arg.dst_cpu)
1178 * These three tests are all lockless; this is OK since all of them
1179 * will be re-checked with proper locks held further down the line.
1181 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1184 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1187 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1190 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1191 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1197 struct migration_arg {
1198 struct task_struct *task;
1202 static int migration_cpu_stop(void *data);
1205 * wait_task_inactive - wait for a thread to unschedule.
1207 * If @match_state is nonzero, it's the @p->state value just checked and
1208 * not expected to change. If it changes, i.e. @p might have woken up,
1209 * then return zero. When we succeed in waiting for @p to be off its CPU,
1210 * we return a positive number (its total switch count). If a second call
1211 * a short while later returns the same number, the caller can be sure that
1212 * @p has remained unscheduled the whole time.
1214 * The caller must ensure that the task *will* unschedule sometime soon,
1215 * else this function might spin for a *long* time. This function can't
1216 * be called with interrupts off, or it may introduce deadlock with
1217 * smp_call_function() if an IPI is sent by the same process we are
1218 * waiting to become inactive.
1220 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1222 unsigned long flags;
1223 int running, queued;
1229 * We do the initial early heuristics without holding
1230 * any task-queue locks at all. We'll only try to get
1231 * the runqueue lock when things look like they will
1237 * If the task is actively running on another CPU
1238 * still, just relax and busy-wait without holding
1241 * NOTE! Since we don't hold any locks, it's not
1242 * even sure that "rq" stays as the right runqueue!
1243 * But we don't care, since "task_running()" will
1244 * return false if the runqueue has changed and p
1245 * is actually now running somewhere else!
1247 while (task_running(rq, p)) {
1248 if (match_state && unlikely(p->state != match_state))
1254 * Ok, time to look more closely! We need the rq
1255 * lock now, to be *sure*. If we're wrong, we'll
1256 * just go back and repeat.
1258 rq = task_rq_lock(p, &flags);
1259 trace_sched_wait_task(p);
1260 running = task_running(rq, p);
1261 queued = task_on_rq_queued(p);
1263 if (!match_state || p->state == match_state)
1264 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1265 task_rq_unlock(rq, p, &flags);
1268 * If it changed from the expected state, bail out now.
1270 if (unlikely(!ncsw))
1274 * Was it really running after all now that we
1275 * checked with the proper locks actually held?
1277 * Oops. Go back and try again..
1279 if (unlikely(running)) {
1285 * It's not enough that it's not actively running,
1286 * it must be off the runqueue _entirely_, and not
1289 * So if it was still runnable (but just not actively
1290 * running right now), it's preempted, and we should
1291 * yield - it could be a while.
1293 if (unlikely(queued)) {
1294 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1296 set_current_state(TASK_UNINTERRUPTIBLE);
1297 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1302 * Ahh, all good. It wasn't running, and it wasn't
1303 * runnable, which means that it will never become
1304 * running in the future either. We're all done!
1313 * kick_process - kick a running thread to enter/exit the kernel
1314 * @p: the to-be-kicked thread
1316 * Cause a process which is running on another CPU to enter
1317 * kernel-mode, without any delay. (to get signals handled.)
1319 * NOTE: this function doesn't have to take the runqueue lock,
1320 * because all it wants to ensure is that the remote task enters
1321 * the kernel. If the IPI races and the task has been migrated
1322 * to another CPU then no harm is done and the purpose has been
1325 void kick_process(struct task_struct *p)
1331 if ((cpu != smp_processor_id()) && task_curr(p))
1332 smp_send_reschedule(cpu);
1335 EXPORT_SYMBOL_GPL(kick_process);
1336 #endif /* CONFIG_SMP */
1340 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1342 static int select_fallback_rq(int cpu, struct task_struct *p)
1344 int nid = cpu_to_node(cpu);
1345 const struct cpumask *nodemask = NULL;
1346 enum { cpuset, possible, fail } state = cpuset;
1350 * If the node that the cpu is on has been offlined, cpu_to_node()
1351 * will return -1. There is no cpu on the node, and we should
1352 * select the cpu on the other node.
1355 nodemask = cpumask_of_node(nid);
1357 /* Look for allowed, online CPU in same node. */
1358 for_each_cpu(dest_cpu, nodemask) {
1359 if (!cpu_online(dest_cpu))
1361 if (!cpu_active(dest_cpu))
1363 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1369 /* Any allowed, online CPU? */
1370 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1371 if (!cpu_online(dest_cpu))
1373 if (!cpu_active(dest_cpu))
1380 /* No more Mr. Nice Guy. */
1381 cpuset_cpus_allowed_fallback(p);
1386 do_set_cpus_allowed(p, cpu_possible_mask);
1397 if (state != cpuset) {
1399 * Don't tell them about moving exiting tasks or
1400 * kernel threads (both mm NULL), since they never
1403 if (p->mm && printk_ratelimit()) {
1404 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1405 task_pid_nr(p), p->comm, cpu);
1413 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1416 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1418 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1421 * In order not to call set_task_cpu() on a blocking task we need
1422 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1425 * Since this is common to all placement strategies, this lives here.
1427 * [ this allows ->select_task() to simply return task_cpu(p) and
1428 * not worry about this generic constraint ]
1430 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1432 cpu = select_fallback_rq(task_cpu(p), p);
1437 static void update_avg(u64 *avg, u64 sample)
1439 s64 diff = sample - *avg;
1445 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1447 #ifdef CONFIG_SCHEDSTATS
1448 struct rq *rq = this_rq();
1451 int this_cpu = smp_processor_id();
1453 if (cpu == this_cpu) {
1454 schedstat_inc(rq, ttwu_local);
1455 schedstat_inc(p, se.statistics.nr_wakeups_local);
1457 struct sched_domain *sd;
1459 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1461 for_each_domain(this_cpu, sd) {
1462 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1463 schedstat_inc(sd, ttwu_wake_remote);
1470 if (wake_flags & WF_MIGRATED)
1471 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1473 #endif /* CONFIG_SMP */
1475 schedstat_inc(rq, ttwu_count);
1476 schedstat_inc(p, se.statistics.nr_wakeups);
1478 if (wake_flags & WF_SYNC)
1479 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1481 #endif /* CONFIG_SCHEDSTATS */
1484 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1486 activate_task(rq, p, en_flags);
1487 p->on_rq = TASK_ON_RQ_QUEUED;
1489 /* if a worker is waking up, notify workqueue */
1490 if (p->flags & PF_WQ_WORKER)
1491 wq_worker_waking_up(p, cpu_of(rq));
1495 * Mark the task runnable and perform wakeup-preemption.
1498 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1500 check_preempt_curr(rq, p, wake_flags);
1501 trace_sched_wakeup(p, true);
1503 p->state = TASK_RUNNING;
1505 if (p->sched_class->task_woken)
1506 p->sched_class->task_woken(rq, p);
1508 if (rq->idle_stamp) {
1509 u64 delta = rq_clock(rq) - rq->idle_stamp;
1510 u64 max = 2*rq->max_idle_balance_cost;
1512 update_avg(&rq->avg_idle, delta);
1514 if (rq->avg_idle > max)
1523 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1526 if (p->sched_contributes_to_load)
1527 rq->nr_uninterruptible--;
1530 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1531 ttwu_do_wakeup(rq, p, wake_flags);
1535 * Called in case the task @p isn't fully descheduled from its runqueue,
1536 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1537 * since all we need to do is flip p->state to TASK_RUNNING, since
1538 * the task is still ->on_rq.
1540 static int ttwu_remote(struct task_struct *p, int wake_flags)
1545 rq = __task_rq_lock(p);
1546 if (task_on_rq_queued(p)) {
1547 /* check_preempt_curr() may use rq clock */
1548 update_rq_clock(rq);
1549 ttwu_do_wakeup(rq, p, wake_flags);
1552 __task_rq_unlock(rq);
1558 void sched_ttwu_pending(void)
1560 struct rq *rq = this_rq();
1561 struct llist_node *llist = llist_del_all(&rq->wake_list);
1562 struct task_struct *p;
1563 unsigned long flags;
1568 raw_spin_lock_irqsave(&rq->lock, flags);
1571 p = llist_entry(llist, struct task_struct, wake_entry);
1572 llist = llist_next(llist);
1573 ttwu_do_activate(rq, p, 0);
1576 raw_spin_unlock_irqrestore(&rq->lock, flags);
1579 void scheduler_ipi(void)
1582 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1583 * TIF_NEED_RESCHED remotely (for the first time) will also send
1586 preempt_fold_need_resched();
1588 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1592 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1593 * traditionally all their work was done from the interrupt return
1594 * path. Now that we actually do some work, we need to make sure
1597 * Some archs already do call them, luckily irq_enter/exit nest
1600 * Arguably we should visit all archs and update all handlers,
1601 * however a fair share of IPIs are still resched only so this would
1602 * somewhat pessimize the simple resched case.
1605 sched_ttwu_pending();
1608 * Check if someone kicked us for doing the nohz idle load balance.
1610 if (unlikely(got_nohz_idle_kick())) {
1611 this_rq()->idle_balance = 1;
1612 raise_softirq_irqoff(SCHED_SOFTIRQ);
1617 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1619 struct rq *rq = cpu_rq(cpu);
1621 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1622 if (!set_nr_if_polling(rq->idle))
1623 smp_send_reschedule(cpu);
1625 trace_sched_wake_idle_without_ipi(cpu);
1629 bool cpus_share_cache(int this_cpu, int that_cpu)
1631 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1633 #endif /* CONFIG_SMP */
1635 static void ttwu_queue(struct task_struct *p, int cpu)
1637 struct rq *rq = cpu_rq(cpu);
1639 #if defined(CONFIG_SMP)
1640 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1641 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1642 ttwu_queue_remote(p, cpu);
1647 raw_spin_lock(&rq->lock);
1648 ttwu_do_activate(rq, p, 0);
1649 raw_spin_unlock(&rq->lock);
1653 * try_to_wake_up - wake up a thread
1654 * @p: the thread to be awakened
1655 * @state: the mask of task states that can be woken
1656 * @wake_flags: wake modifier flags (WF_*)
1658 * Put it on the run-queue if it's not already there. The "current"
1659 * thread is always on the run-queue (except when the actual
1660 * re-schedule is in progress), and as such you're allowed to do
1661 * the simpler "current->state = TASK_RUNNING" to mark yourself
1662 * runnable without the overhead of this.
1664 * Return: %true if @p was woken up, %false if it was already running.
1665 * or @state didn't match @p's state.
1668 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1670 unsigned long flags;
1671 int cpu, success = 0;
1674 * If we are going to wake up a thread waiting for CONDITION we
1675 * need to ensure that CONDITION=1 done by the caller can not be
1676 * reordered with p->state check below. This pairs with mb() in
1677 * set_current_state() the waiting thread does.
1679 smp_mb__before_spinlock();
1680 raw_spin_lock_irqsave(&p->pi_lock, flags);
1681 if (!(p->state & state))
1684 success = 1; /* we're going to change ->state */
1687 if (p->on_rq && ttwu_remote(p, wake_flags))
1692 * If the owning (remote) cpu is still in the middle of schedule() with
1693 * this task as prev, wait until its done referencing the task.
1698 * Pairs with the smp_wmb() in finish_lock_switch().
1702 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1703 p->state = TASK_WAKING;
1705 if (p->sched_class->task_waking)
1706 p->sched_class->task_waking(p);
1708 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1709 if (task_cpu(p) != cpu) {
1710 wake_flags |= WF_MIGRATED;
1711 set_task_cpu(p, cpu);
1713 #endif /* CONFIG_SMP */
1717 ttwu_stat(p, cpu, wake_flags);
1719 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1725 * try_to_wake_up_local - try to wake up a local task with rq lock held
1726 * @p: the thread to be awakened
1728 * Put @p on the run-queue if it's not already there. The caller must
1729 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1732 static void try_to_wake_up_local(struct task_struct *p)
1734 struct rq *rq = task_rq(p);
1736 if (WARN_ON_ONCE(rq != this_rq()) ||
1737 WARN_ON_ONCE(p == current))
1740 lockdep_assert_held(&rq->lock);
1742 if (!raw_spin_trylock(&p->pi_lock)) {
1743 raw_spin_unlock(&rq->lock);
1744 raw_spin_lock(&p->pi_lock);
1745 raw_spin_lock(&rq->lock);
1748 if (!(p->state & TASK_NORMAL))
1751 if (!task_on_rq_queued(p))
1752 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1754 ttwu_do_wakeup(rq, p, 0);
1755 ttwu_stat(p, smp_processor_id(), 0);
1757 raw_spin_unlock(&p->pi_lock);
1761 * wake_up_process - Wake up a specific process
1762 * @p: The process to be woken up.
1764 * Attempt to wake up the nominated process and move it to the set of runnable
1767 * Return: 1 if the process was woken up, 0 if it was already running.
1769 * It may be assumed that this function implies a write memory barrier before
1770 * changing the task state if and only if any tasks are woken up.
1772 int wake_up_process(struct task_struct *p)
1774 WARN_ON(task_is_stopped_or_traced(p));
1775 return try_to_wake_up(p, TASK_NORMAL, 0);
1777 EXPORT_SYMBOL(wake_up_process);
1779 int wake_up_state(struct task_struct *p, unsigned int state)
1781 return try_to_wake_up(p, state, 0);
1785 * Perform scheduler related setup for a newly forked process p.
1786 * p is forked by current.
1788 * __sched_fork() is basic setup used by init_idle() too:
1790 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1795 p->se.exec_start = 0;
1796 p->se.sum_exec_runtime = 0;
1797 p->se.prev_sum_exec_runtime = 0;
1798 p->se.nr_migrations = 0;
1800 INIT_LIST_HEAD(&p->se.group_node);
1802 #ifdef CONFIG_SCHEDSTATS
1803 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1806 RB_CLEAR_NODE(&p->dl.rb_node);
1807 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1808 p->dl.dl_runtime = p->dl.runtime = 0;
1809 p->dl.dl_deadline = p->dl.deadline = 0;
1810 p->dl.dl_period = 0;
1813 INIT_LIST_HEAD(&p->rt.run_list);
1815 #ifdef CONFIG_PREEMPT_NOTIFIERS
1816 INIT_HLIST_HEAD(&p->preempt_notifiers);
1819 #ifdef CONFIG_NUMA_BALANCING
1820 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1821 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1822 p->mm->numa_scan_seq = 0;
1825 if (clone_flags & CLONE_VM)
1826 p->numa_preferred_nid = current->numa_preferred_nid;
1828 p->numa_preferred_nid = -1;
1830 p->node_stamp = 0ULL;
1831 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1832 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1833 p->numa_work.next = &p->numa_work;
1834 p->numa_faults_memory = NULL;
1835 p->numa_faults_buffer_memory = NULL;
1836 p->last_task_numa_placement = 0;
1837 p->last_sum_exec_runtime = 0;
1839 INIT_LIST_HEAD(&p->numa_entry);
1840 p->numa_group = NULL;
1841 #endif /* CONFIG_NUMA_BALANCING */
1844 #ifdef CONFIG_NUMA_BALANCING
1845 #ifdef CONFIG_SCHED_DEBUG
1846 void set_numabalancing_state(bool enabled)
1849 sched_feat_set("NUMA");
1851 sched_feat_set("NO_NUMA");
1854 __read_mostly bool numabalancing_enabled;
1856 void set_numabalancing_state(bool enabled)
1858 numabalancing_enabled = enabled;
1860 #endif /* CONFIG_SCHED_DEBUG */
1862 #ifdef CONFIG_PROC_SYSCTL
1863 int sysctl_numa_balancing(struct ctl_table *table, int write,
1864 void __user *buffer, size_t *lenp, loff_t *ppos)
1868 int state = numabalancing_enabled;
1870 if (write && !capable(CAP_SYS_ADMIN))
1875 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1879 set_numabalancing_state(state);
1886 * fork()/clone()-time setup:
1888 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1890 unsigned long flags;
1891 int cpu = get_cpu();
1893 __sched_fork(clone_flags, p);
1895 * We mark the process as running here. This guarantees that
1896 * nobody will actually run it, and a signal or other external
1897 * event cannot wake it up and insert it on the runqueue either.
1899 p->state = TASK_RUNNING;
1902 * Make sure we do not leak PI boosting priority to the child.
1904 p->prio = current->normal_prio;
1907 * Revert to default priority/policy on fork if requested.
1909 if (unlikely(p->sched_reset_on_fork)) {
1910 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1911 p->policy = SCHED_NORMAL;
1912 p->static_prio = NICE_TO_PRIO(0);
1914 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1915 p->static_prio = NICE_TO_PRIO(0);
1917 p->prio = p->normal_prio = __normal_prio(p);
1921 * We don't need the reset flag anymore after the fork. It has
1922 * fulfilled its duty:
1924 p->sched_reset_on_fork = 0;
1927 if (dl_prio(p->prio)) {
1930 } else if (rt_prio(p->prio)) {
1931 p->sched_class = &rt_sched_class;
1933 p->sched_class = &fair_sched_class;
1936 if (p->sched_class->task_fork)
1937 p->sched_class->task_fork(p);
1940 * The child is not yet in the pid-hash so no cgroup attach races,
1941 * and the cgroup is pinned to this child due to cgroup_fork()
1942 * is ran before sched_fork().
1944 * Silence PROVE_RCU.
1946 raw_spin_lock_irqsave(&p->pi_lock, flags);
1947 set_task_cpu(p, cpu);
1948 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1950 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1951 if (likely(sched_info_on()))
1952 memset(&p->sched_info, 0, sizeof(p->sched_info));
1954 #if defined(CONFIG_SMP)
1957 init_task_preempt_count(p);
1959 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1960 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1967 unsigned long to_ratio(u64 period, u64 runtime)
1969 if (runtime == RUNTIME_INF)
1973 * Doing this here saves a lot of checks in all
1974 * the calling paths, and returning zero seems
1975 * safe for them anyway.
1980 return div64_u64(runtime << 20, period);
1984 inline struct dl_bw *dl_bw_of(int i)
1986 return &cpu_rq(i)->rd->dl_bw;
1989 static inline int dl_bw_cpus(int i)
1991 struct root_domain *rd = cpu_rq(i)->rd;
1994 for_each_cpu_and(i, rd->span, cpu_active_mask)
2000 inline struct dl_bw *dl_bw_of(int i)
2002 return &cpu_rq(i)->dl.dl_bw;
2005 static inline int dl_bw_cpus(int i)
2012 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2014 dl_b->total_bw -= tsk_bw;
2018 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2020 dl_b->total_bw += tsk_bw;
2024 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2026 return dl_b->bw != -1 &&
2027 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2031 * We must be sure that accepting a new task (or allowing changing the
2032 * parameters of an existing one) is consistent with the bandwidth
2033 * constraints. If yes, this function also accordingly updates the currently
2034 * allocated bandwidth to reflect the new situation.
2036 * This function is called while holding p's rq->lock.
2038 static int dl_overflow(struct task_struct *p, int policy,
2039 const struct sched_attr *attr)
2042 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2043 u64 period = attr->sched_period ?: attr->sched_deadline;
2044 u64 runtime = attr->sched_runtime;
2045 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2048 if (new_bw == p->dl.dl_bw)
2052 * Either if a task, enters, leave, or stays -deadline but changes
2053 * its parameters, we may need to update accordingly the total
2054 * allocated bandwidth of the container.
2056 raw_spin_lock(&dl_b->lock);
2057 cpus = dl_bw_cpus(task_cpu(p));
2058 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2059 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2060 __dl_add(dl_b, new_bw);
2062 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2063 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2064 __dl_clear(dl_b, p->dl.dl_bw);
2065 __dl_add(dl_b, new_bw);
2067 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2068 __dl_clear(dl_b, p->dl.dl_bw);
2071 raw_spin_unlock(&dl_b->lock);
2076 extern void init_dl_bw(struct dl_bw *dl_b);
2079 * wake_up_new_task - wake up a newly created task for the first time.
2081 * This function will do some initial scheduler statistics housekeeping
2082 * that must be done for every newly created context, then puts the task
2083 * on the runqueue and wakes it.
2085 void wake_up_new_task(struct task_struct *p)
2087 unsigned long flags;
2090 raw_spin_lock_irqsave(&p->pi_lock, flags);
2093 * Fork balancing, do it here and not earlier because:
2094 * - cpus_allowed can change in the fork path
2095 * - any previously selected cpu might disappear through hotplug
2097 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2100 /* Initialize new task's runnable average */
2101 init_task_runnable_average(p);
2102 rq = __task_rq_lock(p);
2103 activate_task(rq, p, 0);
2104 p->on_rq = TASK_ON_RQ_QUEUED;
2105 trace_sched_wakeup_new(p, true);
2106 check_preempt_curr(rq, p, WF_FORK);
2108 if (p->sched_class->task_woken)
2109 p->sched_class->task_woken(rq, p);
2111 task_rq_unlock(rq, p, &flags);
2114 #ifdef CONFIG_PREEMPT_NOTIFIERS
2117 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2118 * @notifier: notifier struct to register
2120 void preempt_notifier_register(struct preempt_notifier *notifier)
2122 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2124 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2127 * preempt_notifier_unregister - no longer interested in preemption notifications
2128 * @notifier: notifier struct to unregister
2130 * This is safe to call from within a preemption notifier.
2132 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2134 hlist_del(¬ifier->link);
2136 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2138 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2140 struct preempt_notifier *notifier;
2142 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2143 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2147 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2148 struct task_struct *next)
2150 struct preempt_notifier *notifier;
2152 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2153 notifier->ops->sched_out(notifier, next);
2156 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2158 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2163 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2164 struct task_struct *next)
2168 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2171 * prepare_task_switch - prepare to switch tasks
2172 * @rq: the runqueue preparing to switch
2173 * @prev: the current task that is being switched out
2174 * @next: the task we are going to switch to.
2176 * This is called with the rq lock held and interrupts off. It must
2177 * be paired with a subsequent finish_task_switch after the context
2180 * prepare_task_switch sets up locking and calls architecture specific
2184 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2185 struct task_struct *next)
2187 trace_sched_switch(prev, next);
2188 sched_info_switch(rq, prev, next);
2189 perf_event_task_sched_out(prev, next);
2190 fire_sched_out_preempt_notifiers(prev, next);
2191 prepare_lock_switch(rq, next);
2192 prepare_arch_switch(next);
2196 * finish_task_switch - clean up after a task-switch
2197 * @rq: runqueue associated with task-switch
2198 * @prev: the thread we just switched away from.
2200 * finish_task_switch must be called after the context switch, paired
2201 * with a prepare_task_switch call before the context switch.
2202 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2203 * and do any other architecture-specific cleanup actions.
2205 * Note that we may have delayed dropping an mm in context_switch(). If
2206 * so, we finish that here outside of the runqueue lock. (Doing it
2207 * with the lock held can cause deadlocks; see schedule() for
2210 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2211 __releases(rq->lock)
2213 struct mm_struct *mm = rq->prev_mm;
2219 * A task struct has one reference for the use as "current".
2220 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2221 * schedule one last time. The schedule call will never return, and
2222 * the scheduled task must drop that reference.
2223 * The test for TASK_DEAD must occur while the runqueue locks are
2224 * still held, otherwise prev could be scheduled on another cpu, die
2225 * there before we look at prev->state, and then the reference would
2227 * Manfred Spraul <manfred@colorfullife.com>
2229 prev_state = prev->state;
2230 vtime_task_switch(prev);
2231 finish_arch_switch(prev);
2232 perf_event_task_sched_in(prev, current);
2233 finish_lock_switch(rq, prev);
2234 finish_arch_post_lock_switch();
2236 fire_sched_in_preempt_notifiers(current);
2239 if (unlikely(prev_state == TASK_DEAD)) {
2240 if (prev->sched_class->task_dead)
2241 prev->sched_class->task_dead(prev);
2244 * Remove function-return probe instances associated with this
2245 * task and put them back on the free list.
2247 kprobe_flush_task(prev);
2248 put_task_struct(prev);
2251 tick_nohz_task_switch(current);
2256 /* rq->lock is NOT held, but preemption is disabled */
2257 static inline void post_schedule(struct rq *rq)
2259 if (rq->post_schedule) {
2260 unsigned long flags;
2262 raw_spin_lock_irqsave(&rq->lock, flags);
2263 if (rq->curr->sched_class->post_schedule)
2264 rq->curr->sched_class->post_schedule(rq);
2265 raw_spin_unlock_irqrestore(&rq->lock, flags);
2267 rq->post_schedule = 0;
2273 static inline void post_schedule(struct rq *rq)
2280 * schedule_tail - first thing a freshly forked thread must call.
2281 * @prev: the thread we just switched away from.
2283 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2284 __releases(rq->lock)
2286 struct rq *rq = this_rq();
2288 finish_task_switch(rq, prev);
2291 * FIXME: do we need to worry about rq being invalidated by the
2296 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2297 /* In this case, finish_task_switch does not reenable preemption */
2300 if (current->set_child_tid)
2301 put_user(task_pid_vnr(current), current->set_child_tid);
2305 * context_switch - switch to the new MM and the new
2306 * thread's register state.
2309 context_switch(struct rq *rq, struct task_struct *prev,
2310 struct task_struct *next)
2312 struct mm_struct *mm, *oldmm;
2314 prepare_task_switch(rq, prev, next);
2317 oldmm = prev->active_mm;
2319 * For paravirt, this is coupled with an exit in switch_to to
2320 * combine the page table reload and the switch backend into
2323 arch_start_context_switch(prev);
2326 next->active_mm = oldmm;
2327 atomic_inc(&oldmm->mm_count);
2328 enter_lazy_tlb(oldmm, next);
2330 switch_mm(oldmm, mm, next);
2333 prev->active_mm = NULL;
2334 rq->prev_mm = oldmm;
2337 * Since the runqueue lock will be released by the next
2338 * task (which is an invalid locking op but in the case
2339 * of the scheduler it's an obvious special-case), so we
2340 * do an early lockdep release here:
2342 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2343 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2346 context_tracking_task_switch(prev, next);
2347 /* Here we just switch the register state and the stack. */
2348 switch_to(prev, next, prev);
2352 * this_rq must be evaluated again because prev may have moved
2353 * CPUs since it called schedule(), thus the 'rq' on its stack
2354 * frame will be invalid.
2356 finish_task_switch(this_rq(), prev);
2360 * nr_running and nr_context_switches:
2362 * externally visible scheduler statistics: current number of runnable
2363 * threads, total number of context switches performed since bootup.
2365 unsigned long nr_running(void)
2367 unsigned long i, sum = 0;
2369 for_each_online_cpu(i)
2370 sum += cpu_rq(i)->nr_running;
2375 unsigned long long nr_context_switches(void)
2378 unsigned long long sum = 0;
2380 for_each_possible_cpu(i)
2381 sum += cpu_rq(i)->nr_switches;
2386 unsigned long nr_iowait(void)
2388 unsigned long i, sum = 0;
2390 for_each_possible_cpu(i)
2391 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2396 unsigned long nr_iowait_cpu(int cpu)
2398 struct rq *this = cpu_rq(cpu);
2399 return atomic_read(&this->nr_iowait);
2405 * sched_exec - execve() is a valuable balancing opportunity, because at
2406 * this point the task has the smallest effective memory and cache footprint.
2408 void sched_exec(void)
2410 struct task_struct *p = current;
2411 unsigned long flags;
2414 raw_spin_lock_irqsave(&p->pi_lock, flags);
2415 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2416 if (dest_cpu == smp_processor_id())
2419 if (likely(cpu_active(dest_cpu))) {
2420 struct migration_arg arg = { p, dest_cpu };
2422 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2423 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2427 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2432 DEFINE_PER_CPU(struct kernel_stat, kstat);
2433 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2435 EXPORT_PER_CPU_SYMBOL(kstat);
2436 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2439 * Return any ns on the sched_clock that have not yet been accounted in
2440 * @p in case that task is currently running.
2442 * Called with task_rq_lock() held on @rq.
2444 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2449 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2450 * project cycles that may never be accounted to this
2451 * thread, breaking clock_gettime().
2453 if (task_current(rq, p) && task_on_rq_queued(p)) {
2454 update_rq_clock(rq);
2455 ns = rq_clock_task(rq) - p->se.exec_start;
2463 unsigned long long task_delta_exec(struct task_struct *p)
2465 unsigned long flags;
2469 rq = task_rq_lock(p, &flags);
2470 ns = do_task_delta_exec(p, rq);
2471 task_rq_unlock(rq, p, &flags);
2477 * Return accounted runtime for the task.
2478 * In case the task is currently running, return the runtime plus current's
2479 * pending runtime that have not been accounted yet.
2481 unsigned long long task_sched_runtime(struct task_struct *p)
2483 unsigned long flags;
2487 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2489 * 64-bit doesn't need locks to atomically read a 64bit value.
2490 * So we have a optimization chance when the task's delta_exec is 0.
2491 * Reading ->on_cpu is racy, but this is ok.
2493 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2494 * If we race with it entering cpu, unaccounted time is 0. This is
2495 * indistinguishable from the read occurring a few cycles earlier.
2496 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2497 * been accounted, so we're correct here as well.
2499 if (!p->on_cpu || !task_on_rq_queued(p))
2500 return p->se.sum_exec_runtime;
2503 rq = task_rq_lock(p, &flags);
2504 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2505 task_rq_unlock(rq, p, &flags);
2511 * This function gets called by the timer code, with HZ frequency.
2512 * We call it with interrupts disabled.
2514 void scheduler_tick(void)
2516 int cpu = smp_processor_id();
2517 struct rq *rq = cpu_rq(cpu);
2518 struct task_struct *curr = rq->curr;
2522 raw_spin_lock(&rq->lock);
2523 update_rq_clock(rq);
2524 curr->sched_class->task_tick(rq, curr, 0);
2525 update_cpu_load_active(rq);
2526 raw_spin_unlock(&rq->lock);
2528 perf_event_task_tick();
2531 rq->idle_balance = idle_cpu(cpu);
2532 trigger_load_balance(rq);
2534 rq_last_tick_reset(rq);
2537 #ifdef CONFIG_NO_HZ_FULL
2539 * scheduler_tick_max_deferment
2541 * Keep at least one tick per second when a single
2542 * active task is running because the scheduler doesn't
2543 * yet completely support full dynticks environment.
2545 * This makes sure that uptime, CFS vruntime, load
2546 * balancing, etc... continue to move forward, even
2547 * with a very low granularity.
2549 * Return: Maximum deferment in nanoseconds.
2551 u64 scheduler_tick_max_deferment(void)
2553 struct rq *rq = this_rq();
2554 unsigned long next, now = ACCESS_ONCE(jiffies);
2556 next = rq->last_sched_tick + HZ;
2558 if (time_before_eq(next, now))
2561 return jiffies_to_nsecs(next - now);
2565 notrace unsigned long get_parent_ip(unsigned long addr)
2567 if (in_lock_functions(addr)) {
2568 addr = CALLER_ADDR2;
2569 if (in_lock_functions(addr))
2570 addr = CALLER_ADDR3;
2575 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2576 defined(CONFIG_PREEMPT_TRACER))
2578 void preempt_count_add(int val)
2580 #ifdef CONFIG_DEBUG_PREEMPT
2584 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2587 __preempt_count_add(val);
2588 #ifdef CONFIG_DEBUG_PREEMPT
2590 * Spinlock count overflowing soon?
2592 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2595 if (preempt_count() == val) {
2596 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2597 #ifdef CONFIG_DEBUG_PREEMPT
2598 current->preempt_disable_ip = ip;
2600 trace_preempt_off(CALLER_ADDR0, ip);
2603 EXPORT_SYMBOL(preempt_count_add);
2604 NOKPROBE_SYMBOL(preempt_count_add);
2606 void preempt_count_sub(int val)
2608 #ifdef CONFIG_DEBUG_PREEMPT
2612 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2615 * Is the spinlock portion underflowing?
2617 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2618 !(preempt_count() & PREEMPT_MASK)))
2622 if (preempt_count() == val)
2623 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2624 __preempt_count_sub(val);
2626 EXPORT_SYMBOL(preempt_count_sub);
2627 NOKPROBE_SYMBOL(preempt_count_sub);
2632 * Print scheduling while atomic bug:
2634 static noinline void __schedule_bug(struct task_struct *prev)
2636 if (oops_in_progress)
2639 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2640 prev->comm, prev->pid, preempt_count());
2642 debug_show_held_locks(prev);
2644 if (irqs_disabled())
2645 print_irqtrace_events(prev);
2646 #ifdef CONFIG_DEBUG_PREEMPT
2647 if (in_atomic_preempt_off()) {
2648 pr_err("Preemption disabled at:");
2649 print_ip_sym(current->preempt_disable_ip);
2654 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2658 * Various schedule()-time debugging checks and statistics:
2660 static inline void schedule_debug(struct task_struct *prev)
2663 * Test if we are atomic. Since do_exit() needs to call into
2664 * schedule() atomically, we ignore that path. Otherwise whine
2665 * if we are scheduling when we should not.
2667 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2668 __schedule_bug(prev);
2671 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2673 schedstat_inc(this_rq(), sched_count);
2677 * Pick up the highest-prio task:
2679 static inline struct task_struct *
2680 pick_next_task(struct rq *rq, struct task_struct *prev)
2682 const struct sched_class *class = &fair_sched_class;
2683 struct task_struct *p;
2686 * Optimization: we know that if all tasks are in
2687 * the fair class we can call that function directly:
2689 if (likely(prev->sched_class == class &&
2690 rq->nr_running == rq->cfs.h_nr_running)) {
2691 p = fair_sched_class.pick_next_task(rq, prev);
2692 if (unlikely(p == RETRY_TASK))
2695 /* assumes fair_sched_class->next == idle_sched_class */
2697 p = idle_sched_class.pick_next_task(rq, prev);
2703 for_each_class(class) {
2704 p = class->pick_next_task(rq, prev);
2706 if (unlikely(p == RETRY_TASK))
2712 BUG(); /* the idle class will always have a runnable task */
2716 * __schedule() is the main scheduler function.
2718 * The main means of driving the scheduler and thus entering this function are:
2720 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2722 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2723 * paths. For example, see arch/x86/entry_64.S.
2725 * To drive preemption between tasks, the scheduler sets the flag in timer
2726 * interrupt handler scheduler_tick().
2728 * 3. Wakeups don't really cause entry into schedule(). They add a
2729 * task to the run-queue and that's it.
2731 * Now, if the new task added to the run-queue preempts the current
2732 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2733 * called on the nearest possible occasion:
2735 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2737 * - in syscall or exception context, at the next outmost
2738 * preempt_enable(). (this might be as soon as the wake_up()'s
2741 * - in IRQ context, return from interrupt-handler to
2742 * preemptible context
2744 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2747 * - cond_resched() call
2748 * - explicit schedule() call
2749 * - return from syscall or exception to user-space
2750 * - return from interrupt-handler to user-space
2752 static void __sched __schedule(void)
2754 struct task_struct *prev, *next;
2755 unsigned long *switch_count;
2761 cpu = smp_processor_id();
2763 rcu_note_context_switch(cpu);
2766 schedule_debug(prev);
2768 if (sched_feat(HRTICK))
2772 * Make sure that signal_pending_state()->signal_pending() below
2773 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2774 * done by the caller to avoid the race with signal_wake_up().
2776 smp_mb__before_spinlock();
2777 raw_spin_lock_irq(&rq->lock);
2779 switch_count = &prev->nivcsw;
2780 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2781 if (unlikely(signal_pending_state(prev->state, prev))) {
2782 prev->state = TASK_RUNNING;
2784 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2788 * If a worker went to sleep, notify and ask workqueue
2789 * whether it wants to wake up a task to maintain
2792 if (prev->flags & PF_WQ_WORKER) {
2793 struct task_struct *to_wakeup;
2795 to_wakeup = wq_worker_sleeping(prev, cpu);
2797 try_to_wake_up_local(to_wakeup);
2800 switch_count = &prev->nvcsw;
2803 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0)
2804 update_rq_clock(rq);
2806 next = pick_next_task(rq, prev);
2807 clear_tsk_need_resched(prev);
2808 clear_preempt_need_resched();
2809 rq->skip_clock_update = 0;
2811 if (likely(prev != next)) {
2816 context_switch(rq, prev, next); /* unlocks the rq */
2818 * The context switch have flipped the stack from under us
2819 * and restored the local variables which were saved when
2820 * this task called schedule() in the past. prev == current
2821 * is still correct, but it can be moved to another cpu/rq.
2823 cpu = smp_processor_id();
2826 raw_spin_unlock_irq(&rq->lock);
2830 sched_preempt_enable_no_resched();
2835 static inline void sched_submit_work(struct task_struct *tsk)
2837 if (!tsk->state || tsk_is_pi_blocked(tsk))
2840 * If we are going to sleep and we have plugged IO queued,
2841 * make sure to submit it to avoid deadlocks.
2843 if (blk_needs_flush_plug(tsk))
2844 blk_schedule_flush_plug(tsk);
2847 asmlinkage __visible void __sched schedule(void)
2849 struct task_struct *tsk = current;
2851 sched_submit_work(tsk);
2854 EXPORT_SYMBOL(schedule);
2856 #ifdef CONFIG_CONTEXT_TRACKING
2857 asmlinkage __visible void __sched schedule_user(void)
2860 * If we come here after a random call to set_need_resched(),
2861 * or we have been woken up remotely but the IPI has not yet arrived,
2862 * we haven't yet exited the RCU idle mode. Do it here manually until
2863 * we find a better solution.
2872 * schedule_preempt_disabled - called with preemption disabled
2874 * Returns with preemption disabled. Note: preempt_count must be 1
2876 void __sched schedule_preempt_disabled(void)
2878 sched_preempt_enable_no_resched();
2883 #ifdef CONFIG_PREEMPT
2885 * this is the entry point to schedule() from in-kernel preemption
2886 * off of preempt_enable. Kernel preemptions off return from interrupt
2887 * occur there and call schedule directly.
2889 asmlinkage __visible void __sched notrace preempt_schedule(void)
2892 * If there is a non-zero preempt_count or interrupts are disabled,
2893 * we do not want to preempt the current task. Just return..
2895 if (likely(!preemptible()))
2899 __preempt_count_add(PREEMPT_ACTIVE);
2901 __preempt_count_sub(PREEMPT_ACTIVE);
2904 * Check again in case we missed a preemption opportunity
2905 * between schedule and now.
2908 } while (need_resched());
2910 NOKPROBE_SYMBOL(preempt_schedule);
2911 EXPORT_SYMBOL(preempt_schedule);
2912 #endif /* CONFIG_PREEMPT */
2915 * this is the entry point to schedule() from kernel preemption
2916 * off of irq context.
2917 * Note, that this is called and return with irqs disabled. This will
2918 * protect us against recursive calling from irq.
2920 asmlinkage __visible void __sched preempt_schedule_irq(void)
2922 enum ctx_state prev_state;
2924 /* Catch callers which need to be fixed */
2925 BUG_ON(preempt_count() || !irqs_disabled());
2927 prev_state = exception_enter();
2930 __preempt_count_add(PREEMPT_ACTIVE);
2933 local_irq_disable();
2934 __preempt_count_sub(PREEMPT_ACTIVE);
2937 * Check again in case we missed a preemption opportunity
2938 * between schedule and now.
2941 } while (need_resched());
2943 exception_exit(prev_state);
2946 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2949 return try_to_wake_up(curr->private, mode, wake_flags);
2951 EXPORT_SYMBOL(default_wake_function);
2953 #ifdef CONFIG_RT_MUTEXES
2956 * rt_mutex_setprio - set the current priority of a task
2958 * @prio: prio value (kernel-internal form)
2960 * This function changes the 'effective' priority of a task. It does
2961 * not touch ->normal_prio like __setscheduler().
2963 * Used by the rt_mutex code to implement priority inheritance
2964 * logic. Call site only calls if the priority of the task changed.
2966 void rt_mutex_setprio(struct task_struct *p, int prio)
2968 int oldprio, queued, running, enqueue_flag = 0;
2970 const struct sched_class *prev_class;
2972 BUG_ON(prio > MAX_PRIO);
2974 rq = __task_rq_lock(p);
2977 * Idle task boosting is a nono in general. There is one
2978 * exception, when PREEMPT_RT and NOHZ is active:
2980 * The idle task calls get_next_timer_interrupt() and holds
2981 * the timer wheel base->lock on the CPU and another CPU wants
2982 * to access the timer (probably to cancel it). We can safely
2983 * ignore the boosting request, as the idle CPU runs this code
2984 * with interrupts disabled and will complete the lock
2985 * protected section without being interrupted. So there is no
2986 * real need to boost.
2988 if (unlikely(p == rq->idle)) {
2989 WARN_ON(p != rq->curr);
2990 WARN_ON(p->pi_blocked_on);
2994 trace_sched_pi_setprio(p, prio);
2996 prev_class = p->sched_class;
2997 queued = task_on_rq_queued(p);
2998 running = task_current(rq, p);
3000 dequeue_task(rq, p, 0);
3002 p->sched_class->put_prev_task(rq, p);
3005 * Boosting condition are:
3006 * 1. -rt task is running and holds mutex A
3007 * --> -dl task blocks on mutex A
3009 * 2. -dl task is running and holds mutex A
3010 * --> -dl task blocks on mutex A and could preempt the
3013 if (dl_prio(prio)) {
3014 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3015 if (!dl_prio(p->normal_prio) ||
3016 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3017 p->dl.dl_boosted = 1;
3018 p->dl.dl_throttled = 0;
3019 enqueue_flag = ENQUEUE_REPLENISH;
3021 p->dl.dl_boosted = 0;
3022 p->sched_class = &dl_sched_class;
3023 } else if (rt_prio(prio)) {
3024 if (dl_prio(oldprio))
3025 p->dl.dl_boosted = 0;
3027 enqueue_flag = ENQUEUE_HEAD;
3028 p->sched_class = &rt_sched_class;
3030 if (dl_prio(oldprio))
3031 p->dl.dl_boosted = 0;
3032 p->sched_class = &fair_sched_class;
3038 p->sched_class->set_curr_task(rq);
3040 enqueue_task(rq, p, enqueue_flag);
3042 check_class_changed(rq, p, prev_class, oldprio);
3044 __task_rq_unlock(rq);
3048 void set_user_nice(struct task_struct *p, long nice)
3050 int old_prio, delta, queued;
3051 unsigned long flags;
3054 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3057 * We have to be careful, if called from sys_setpriority(),
3058 * the task might be in the middle of scheduling on another CPU.
3060 rq = task_rq_lock(p, &flags);
3062 * The RT priorities are set via sched_setscheduler(), but we still
3063 * allow the 'normal' nice value to be set - but as expected
3064 * it wont have any effect on scheduling until the task is
3065 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3067 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3068 p->static_prio = NICE_TO_PRIO(nice);
3071 queued = task_on_rq_queued(p);
3073 dequeue_task(rq, p, 0);
3075 p->static_prio = NICE_TO_PRIO(nice);
3078 p->prio = effective_prio(p);
3079 delta = p->prio - old_prio;
3082 enqueue_task(rq, p, 0);
3084 * If the task increased its priority or is running and
3085 * lowered its priority, then reschedule its CPU:
3087 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3091 task_rq_unlock(rq, p, &flags);
3093 EXPORT_SYMBOL(set_user_nice);
3096 * can_nice - check if a task can reduce its nice value
3100 int can_nice(const struct task_struct *p, const int nice)
3102 /* convert nice value [19,-20] to rlimit style value [1,40] */
3103 int nice_rlim = nice_to_rlimit(nice);
3105 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3106 capable(CAP_SYS_NICE));
3109 #ifdef __ARCH_WANT_SYS_NICE
3112 * sys_nice - change the priority of the current process.
3113 * @increment: priority increment
3115 * sys_setpriority is a more generic, but much slower function that
3116 * does similar things.
3118 SYSCALL_DEFINE1(nice, int, increment)
3123 * Setpriority might change our priority at the same moment.
3124 * We don't have to worry. Conceptually one call occurs first
3125 * and we have a single winner.
3127 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3128 nice = task_nice(current) + increment;
3130 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3131 if (increment < 0 && !can_nice(current, nice))
3134 retval = security_task_setnice(current, nice);
3138 set_user_nice(current, nice);
3145 * task_prio - return the priority value of a given task.
3146 * @p: the task in question.
3148 * Return: The priority value as seen by users in /proc.
3149 * RT tasks are offset by -200. Normal tasks are centered
3150 * around 0, value goes from -16 to +15.
3152 int task_prio(const struct task_struct *p)
3154 return p->prio - MAX_RT_PRIO;
3158 * idle_cpu - is a given cpu idle currently?
3159 * @cpu: the processor in question.
3161 * Return: 1 if the CPU is currently idle. 0 otherwise.
3163 int idle_cpu(int cpu)
3165 struct rq *rq = cpu_rq(cpu);
3167 if (rq->curr != rq->idle)
3174 if (!llist_empty(&rq->wake_list))
3182 * idle_task - return the idle task for a given cpu.
3183 * @cpu: the processor in question.
3185 * Return: The idle task for the cpu @cpu.
3187 struct task_struct *idle_task(int cpu)
3189 return cpu_rq(cpu)->idle;
3193 * find_process_by_pid - find a process with a matching PID value.
3194 * @pid: the pid in question.
3196 * The task of @pid, if found. %NULL otherwise.
3198 static struct task_struct *find_process_by_pid(pid_t pid)
3200 return pid ? find_task_by_vpid(pid) : current;
3204 * This function initializes the sched_dl_entity of a newly becoming
3205 * SCHED_DEADLINE task.
3207 * Only the static values are considered here, the actual runtime and the
3208 * absolute deadline will be properly calculated when the task is enqueued
3209 * for the first time with its new policy.
3212 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3214 struct sched_dl_entity *dl_se = &p->dl;
3216 init_dl_task_timer(dl_se);
3217 dl_se->dl_runtime = attr->sched_runtime;
3218 dl_se->dl_deadline = attr->sched_deadline;
3219 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3220 dl_se->flags = attr->sched_flags;
3221 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3222 dl_se->dl_throttled = 0;
3224 dl_se->dl_yielded = 0;
3228 * sched_setparam() passes in -1 for its policy, to let the functions
3229 * it calls know not to change it.
3231 #define SETPARAM_POLICY -1
3233 static void __setscheduler_params(struct task_struct *p,
3234 const struct sched_attr *attr)
3236 int policy = attr->sched_policy;
3238 if (policy == SETPARAM_POLICY)
3243 if (dl_policy(policy))
3244 __setparam_dl(p, attr);
3245 else if (fair_policy(policy))
3246 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3249 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3250 * !rt_policy. Always setting this ensures that things like
3251 * getparam()/getattr() don't report silly values for !rt tasks.
3253 p->rt_priority = attr->sched_priority;
3254 p->normal_prio = normal_prio(p);
3258 /* Actually do priority change: must hold pi & rq lock. */
3259 static void __setscheduler(struct rq *rq, struct task_struct *p,
3260 const struct sched_attr *attr)
3262 __setscheduler_params(p, attr);
3265 * If we get here, there was no pi waiters boosting the
3266 * task. It is safe to use the normal prio.
3268 p->prio = normal_prio(p);
3270 if (dl_prio(p->prio))
3271 p->sched_class = &dl_sched_class;
3272 else if (rt_prio(p->prio))
3273 p->sched_class = &rt_sched_class;
3275 p->sched_class = &fair_sched_class;
3279 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3281 struct sched_dl_entity *dl_se = &p->dl;
3283 attr->sched_priority = p->rt_priority;
3284 attr->sched_runtime = dl_se->dl_runtime;
3285 attr->sched_deadline = dl_se->dl_deadline;
3286 attr->sched_period = dl_se->dl_period;
3287 attr->sched_flags = dl_se->flags;
3291 * This function validates the new parameters of a -deadline task.
3292 * We ask for the deadline not being zero, and greater or equal
3293 * than the runtime, as well as the period of being zero or
3294 * greater than deadline. Furthermore, we have to be sure that
3295 * user parameters are above the internal resolution of 1us (we
3296 * check sched_runtime only since it is always the smaller one) and
3297 * below 2^63 ns (we have to check both sched_deadline and
3298 * sched_period, as the latter can be zero).
3301 __checkparam_dl(const struct sched_attr *attr)
3304 if (attr->sched_deadline == 0)
3308 * Since we truncate DL_SCALE bits, make sure we're at least
3311 if (attr->sched_runtime < (1ULL << DL_SCALE))
3315 * Since we use the MSB for wrap-around and sign issues, make
3316 * sure it's not set (mind that period can be equal to zero).
3318 if (attr->sched_deadline & (1ULL << 63) ||
3319 attr->sched_period & (1ULL << 63))
3322 /* runtime <= deadline <= period (if period != 0) */
3323 if ((attr->sched_period != 0 &&
3324 attr->sched_period < attr->sched_deadline) ||
3325 attr->sched_deadline < attr->sched_runtime)
3332 * check the target process has a UID that matches the current process's
3334 static bool check_same_owner(struct task_struct *p)
3336 const struct cred *cred = current_cred(), *pcred;
3340 pcred = __task_cred(p);
3341 match = (uid_eq(cred->euid, pcred->euid) ||
3342 uid_eq(cred->euid, pcred->uid));
3347 static int __sched_setscheduler(struct task_struct *p,
3348 const struct sched_attr *attr,
3351 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3352 MAX_RT_PRIO - 1 - attr->sched_priority;
3353 int retval, oldprio, oldpolicy = -1, queued, running;
3354 int policy = attr->sched_policy;
3355 unsigned long flags;
3356 const struct sched_class *prev_class;
3360 /* may grab non-irq protected spin_locks */
3361 BUG_ON(in_interrupt());
3363 /* double check policy once rq lock held */
3365 reset_on_fork = p->sched_reset_on_fork;
3366 policy = oldpolicy = p->policy;
3368 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3370 if (policy != SCHED_DEADLINE &&
3371 policy != SCHED_FIFO && policy != SCHED_RR &&
3372 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3373 policy != SCHED_IDLE)
3377 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3381 * Valid priorities for SCHED_FIFO and SCHED_RR are
3382 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3383 * SCHED_BATCH and SCHED_IDLE is 0.
3385 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3386 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3388 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3389 (rt_policy(policy) != (attr->sched_priority != 0)))
3393 * Allow unprivileged RT tasks to decrease priority:
3395 if (user && !capable(CAP_SYS_NICE)) {
3396 if (fair_policy(policy)) {
3397 if (attr->sched_nice < task_nice(p) &&
3398 !can_nice(p, attr->sched_nice))
3402 if (rt_policy(policy)) {
3403 unsigned long rlim_rtprio =
3404 task_rlimit(p, RLIMIT_RTPRIO);
3406 /* can't set/change the rt policy */
3407 if (policy != p->policy && !rlim_rtprio)
3410 /* can't increase priority */
3411 if (attr->sched_priority > p->rt_priority &&
3412 attr->sched_priority > rlim_rtprio)
3417 * Can't set/change SCHED_DEADLINE policy at all for now
3418 * (safest behavior); in the future we would like to allow
3419 * unprivileged DL tasks to increase their relative deadline
3420 * or reduce their runtime (both ways reducing utilization)
3422 if (dl_policy(policy))
3426 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3427 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3429 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3430 if (!can_nice(p, task_nice(p)))
3434 /* can't change other user's priorities */
3435 if (!check_same_owner(p))
3438 /* Normal users shall not reset the sched_reset_on_fork flag */
3439 if (p->sched_reset_on_fork && !reset_on_fork)
3444 retval = security_task_setscheduler(p);
3450 * make sure no PI-waiters arrive (or leave) while we are
3451 * changing the priority of the task:
3453 * To be able to change p->policy safely, the appropriate
3454 * runqueue lock must be held.
3456 rq = task_rq_lock(p, &flags);
3459 * Changing the policy of the stop threads its a very bad idea
3461 if (p == rq->stop) {
3462 task_rq_unlock(rq, p, &flags);
3467 * If not changing anything there's no need to proceed further,
3468 * but store a possible modification of reset_on_fork.
3470 if (unlikely(policy == p->policy)) {
3471 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3473 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3475 if (dl_policy(policy))
3478 p->sched_reset_on_fork = reset_on_fork;
3479 task_rq_unlock(rq, p, &flags);
3485 #ifdef CONFIG_RT_GROUP_SCHED
3487 * Do not allow realtime tasks into groups that have no runtime
3490 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3491 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3492 !task_group_is_autogroup(task_group(p))) {
3493 task_rq_unlock(rq, p, &flags);
3498 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3499 cpumask_t *span = rq->rd->span;
3502 * Don't allow tasks with an affinity mask smaller than
3503 * the entire root_domain to become SCHED_DEADLINE. We
3504 * will also fail if there's no bandwidth available.
3506 if (!cpumask_subset(span, &p->cpus_allowed) ||
3507 rq->rd->dl_bw.bw == 0) {
3508 task_rq_unlock(rq, p, &flags);
3515 /* recheck policy now with rq lock held */
3516 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3517 policy = oldpolicy = -1;
3518 task_rq_unlock(rq, p, &flags);
3523 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3524 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3527 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3528 task_rq_unlock(rq, p, &flags);
3532 p->sched_reset_on_fork = reset_on_fork;
3536 * Special case for priority boosted tasks.
3538 * If the new priority is lower or equal (user space view)
3539 * than the current (boosted) priority, we just store the new
3540 * normal parameters and do not touch the scheduler class and
3541 * the runqueue. This will be done when the task deboost
3544 if (rt_mutex_check_prio(p, newprio)) {
3545 __setscheduler_params(p, attr);
3546 task_rq_unlock(rq, p, &flags);
3550 queued = task_on_rq_queued(p);
3551 running = task_current(rq, p);
3553 dequeue_task(rq, p, 0);
3555 p->sched_class->put_prev_task(rq, p);
3557 prev_class = p->sched_class;
3558 __setscheduler(rq, p, attr);
3561 p->sched_class->set_curr_task(rq);
3564 * We enqueue to tail when the priority of a task is
3565 * increased (user space view).
3567 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3570 check_class_changed(rq, p, prev_class, oldprio);
3571 task_rq_unlock(rq, p, &flags);
3573 rt_mutex_adjust_pi(p);
3578 static int _sched_setscheduler(struct task_struct *p, int policy,
3579 const struct sched_param *param, bool check)
3581 struct sched_attr attr = {
3582 .sched_policy = policy,
3583 .sched_priority = param->sched_priority,
3584 .sched_nice = PRIO_TO_NICE(p->static_prio),
3587 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3588 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3589 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3590 policy &= ~SCHED_RESET_ON_FORK;
3591 attr.sched_policy = policy;
3594 return __sched_setscheduler(p, &attr, check);
3597 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3598 * @p: the task in question.
3599 * @policy: new policy.
3600 * @param: structure containing the new RT priority.
3602 * Return: 0 on success. An error code otherwise.
3604 * NOTE that the task may be already dead.
3606 int sched_setscheduler(struct task_struct *p, int policy,
3607 const struct sched_param *param)
3609 return _sched_setscheduler(p, policy, param, true);
3611 EXPORT_SYMBOL_GPL(sched_setscheduler);
3613 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3615 return __sched_setscheduler(p, attr, true);
3617 EXPORT_SYMBOL_GPL(sched_setattr);
3620 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3621 * @p: the task in question.
3622 * @policy: new policy.
3623 * @param: structure containing the new RT priority.
3625 * Just like sched_setscheduler, only don't bother checking if the
3626 * current context has permission. For example, this is needed in
3627 * stop_machine(): we create temporary high priority worker threads,
3628 * but our caller might not have that capability.
3630 * Return: 0 on success. An error code otherwise.
3632 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3633 const struct sched_param *param)
3635 return _sched_setscheduler(p, policy, param, false);
3639 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3641 struct sched_param lparam;
3642 struct task_struct *p;
3645 if (!param || pid < 0)
3647 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3652 p = find_process_by_pid(pid);
3654 retval = sched_setscheduler(p, policy, &lparam);
3661 * Mimics kernel/events/core.c perf_copy_attr().
3663 static int sched_copy_attr(struct sched_attr __user *uattr,
3664 struct sched_attr *attr)
3669 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3673 * zero the full structure, so that a short copy will be nice.
3675 memset(attr, 0, sizeof(*attr));
3677 ret = get_user(size, &uattr->size);
3681 if (size > PAGE_SIZE) /* silly large */
3684 if (!size) /* abi compat */
3685 size = SCHED_ATTR_SIZE_VER0;
3687 if (size < SCHED_ATTR_SIZE_VER0)
3691 * If we're handed a bigger struct than we know of,
3692 * ensure all the unknown bits are 0 - i.e. new
3693 * user-space does not rely on any kernel feature
3694 * extensions we dont know about yet.
3696 if (size > sizeof(*attr)) {
3697 unsigned char __user *addr;
3698 unsigned char __user *end;
3701 addr = (void __user *)uattr + sizeof(*attr);
3702 end = (void __user *)uattr + size;
3704 for (; addr < end; addr++) {
3705 ret = get_user(val, addr);
3711 size = sizeof(*attr);
3714 ret = copy_from_user(attr, uattr, size);
3719 * XXX: do we want to be lenient like existing syscalls; or do we want
3720 * to be strict and return an error on out-of-bounds values?
3722 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3727 put_user(sizeof(*attr), &uattr->size);
3732 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3733 * @pid: the pid in question.
3734 * @policy: new policy.
3735 * @param: structure containing the new RT priority.
3737 * Return: 0 on success. An error code otherwise.
3739 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3740 struct sched_param __user *, param)
3742 /* negative values for policy are not valid */
3746 return do_sched_setscheduler(pid, policy, param);
3750 * sys_sched_setparam - set/change the RT priority of a thread
3751 * @pid: the pid in question.
3752 * @param: structure containing the new RT priority.
3754 * Return: 0 on success. An error code otherwise.
3756 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3758 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3762 * sys_sched_setattr - same as above, but with extended sched_attr
3763 * @pid: the pid in question.
3764 * @uattr: structure containing the extended parameters.
3765 * @flags: for future extension.
3767 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3768 unsigned int, flags)
3770 struct sched_attr attr;
3771 struct task_struct *p;
3774 if (!uattr || pid < 0 || flags)
3777 retval = sched_copy_attr(uattr, &attr);
3781 if ((int)attr.sched_policy < 0)
3786 p = find_process_by_pid(pid);
3788 retval = sched_setattr(p, &attr);
3795 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3796 * @pid: the pid in question.
3798 * Return: On success, the policy of the thread. Otherwise, a negative error
3801 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3803 struct task_struct *p;
3811 p = find_process_by_pid(pid);
3813 retval = security_task_getscheduler(p);
3816 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3823 * sys_sched_getparam - get the RT priority of a thread
3824 * @pid: the pid in question.
3825 * @param: structure containing the RT priority.
3827 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3830 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3832 struct sched_param lp = { .sched_priority = 0 };
3833 struct task_struct *p;
3836 if (!param || pid < 0)
3840 p = find_process_by_pid(pid);
3845 retval = security_task_getscheduler(p);
3849 if (task_has_rt_policy(p))
3850 lp.sched_priority = p->rt_priority;
3854 * This one might sleep, we cannot do it with a spinlock held ...
3856 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3865 static int sched_read_attr(struct sched_attr __user *uattr,
3866 struct sched_attr *attr,
3871 if (!access_ok(VERIFY_WRITE, uattr, usize))
3875 * If we're handed a smaller struct than we know of,
3876 * ensure all the unknown bits are 0 - i.e. old
3877 * user-space does not get uncomplete information.
3879 if (usize < sizeof(*attr)) {
3880 unsigned char *addr;
3883 addr = (void *)attr + usize;
3884 end = (void *)attr + sizeof(*attr);
3886 for (; addr < end; addr++) {
3894 ret = copy_to_user(uattr, attr, attr->size);
3902 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3903 * @pid: the pid in question.
3904 * @uattr: structure containing the extended parameters.
3905 * @size: sizeof(attr) for fwd/bwd comp.
3906 * @flags: for future extension.
3908 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3909 unsigned int, size, unsigned int, flags)
3911 struct sched_attr attr = {
3912 .size = sizeof(struct sched_attr),
3914 struct task_struct *p;
3917 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3918 size < SCHED_ATTR_SIZE_VER0 || flags)
3922 p = find_process_by_pid(pid);
3927 retval = security_task_getscheduler(p);
3931 attr.sched_policy = p->policy;
3932 if (p->sched_reset_on_fork)
3933 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3934 if (task_has_dl_policy(p))
3935 __getparam_dl(p, &attr);
3936 else if (task_has_rt_policy(p))
3937 attr.sched_priority = p->rt_priority;
3939 attr.sched_nice = task_nice(p);
3943 retval = sched_read_attr(uattr, &attr, size);
3951 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3953 cpumask_var_t cpus_allowed, new_mask;
3954 struct task_struct *p;
3959 p = find_process_by_pid(pid);
3965 /* Prevent p going away */
3969 if (p->flags & PF_NO_SETAFFINITY) {
3973 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3977 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3979 goto out_free_cpus_allowed;
3982 if (!check_same_owner(p)) {
3984 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3991 retval = security_task_setscheduler(p);
3996 cpuset_cpus_allowed(p, cpus_allowed);
3997 cpumask_and(new_mask, in_mask, cpus_allowed);
4000 * Since bandwidth control happens on root_domain basis,
4001 * if admission test is enabled, we only admit -deadline
4002 * tasks allowed to run on all the CPUs in the task's
4006 if (task_has_dl_policy(p)) {
4007 const struct cpumask *span = task_rq(p)->rd->span;
4009 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
4016 retval = set_cpus_allowed_ptr(p, new_mask);
4019 cpuset_cpus_allowed(p, cpus_allowed);
4020 if (!cpumask_subset(new_mask, cpus_allowed)) {
4022 * We must have raced with a concurrent cpuset
4023 * update. Just reset the cpus_allowed to the
4024 * cpuset's cpus_allowed
4026 cpumask_copy(new_mask, cpus_allowed);
4031 free_cpumask_var(new_mask);
4032 out_free_cpus_allowed:
4033 free_cpumask_var(cpus_allowed);
4039 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4040 struct cpumask *new_mask)
4042 if (len < cpumask_size())
4043 cpumask_clear(new_mask);
4044 else if (len > cpumask_size())
4045 len = cpumask_size();
4047 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4051 * sys_sched_setaffinity - set the cpu affinity of a process
4052 * @pid: pid of the process
4053 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4054 * @user_mask_ptr: user-space pointer to the new cpu mask
4056 * Return: 0 on success. An error code otherwise.
4058 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4059 unsigned long __user *, user_mask_ptr)
4061 cpumask_var_t new_mask;
4064 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4067 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4069 retval = sched_setaffinity(pid, new_mask);
4070 free_cpumask_var(new_mask);
4074 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4076 struct task_struct *p;
4077 unsigned long flags;
4083 p = find_process_by_pid(pid);
4087 retval = security_task_getscheduler(p);
4091 raw_spin_lock_irqsave(&p->pi_lock, flags);
4092 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4093 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4102 * sys_sched_getaffinity - get the cpu affinity of a process
4103 * @pid: pid of the process
4104 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4105 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4107 * Return: 0 on success. An error code otherwise.
4109 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4110 unsigned long __user *, user_mask_ptr)
4115 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4117 if (len & (sizeof(unsigned long)-1))
4120 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4123 ret = sched_getaffinity(pid, mask);
4125 size_t retlen = min_t(size_t, len, cpumask_size());
4127 if (copy_to_user(user_mask_ptr, mask, retlen))
4132 free_cpumask_var(mask);
4138 * sys_sched_yield - yield the current processor to other threads.
4140 * This function yields the current CPU to other tasks. If there are no
4141 * other threads running on this CPU then this function will return.
4145 SYSCALL_DEFINE0(sched_yield)
4147 struct rq *rq = this_rq_lock();
4149 schedstat_inc(rq, yld_count);
4150 current->sched_class->yield_task(rq);
4153 * Since we are going to call schedule() anyway, there's
4154 * no need to preempt or enable interrupts:
4156 __release(rq->lock);
4157 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4158 do_raw_spin_unlock(&rq->lock);
4159 sched_preempt_enable_no_resched();
4166 static void __cond_resched(void)
4168 __preempt_count_add(PREEMPT_ACTIVE);
4170 __preempt_count_sub(PREEMPT_ACTIVE);
4173 int __sched _cond_resched(void)
4175 if (should_resched()) {
4181 EXPORT_SYMBOL(_cond_resched);
4184 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4185 * call schedule, and on return reacquire the lock.
4187 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4188 * operations here to prevent schedule() from being called twice (once via
4189 * spin_unlock(), once by hand).
4191 int __cond_resched_lock(spinlock_t *lock)
4193 int resched = should_resched();
4196 lockdep_assert_held(lock);
4198 if (spin_needbreak(lock) || resched) {
4209 EXPORT_SYMBOL(__cond_resched_lock);
4211 int __sched __cond_resched_softirq(void)
4213 BUG_ON(!in_softirq());
4215 if (should_resched()) {
4223 EXPORT_SYMBOL(__cond_resched_softirq);
4226 * yield - yield the current processor to other threads.
4228 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4230 * The scheduler is at all times free to pick the calling task as the most
4231 * eligible task to run, if removing the yield() call from your code breaks
4232 * it, its already broken.
4234 * Typical broken usage is:
4239 * where one assumes that yield() will let 'the other' process run that will
4240 * make event true. If the current task is a SCHED_FIFO task that will never
4241 * happen. Never use yield() as a progress guarantee!!
4243 * If you want to use yield() to wait for something, use wait_event().
4244 * If you want to use yield() to be 'nice' for others, use cond_resched().
4245 * If you still want to use yield(), do not!
4247 void __sched yield(void)
4249 set_current_state(TASK_RUNNING);
4252 EXPORT_SYMBOL(yield);
4255 * yield_to - yield the current processor to another thread in
4256 * your thread group, or accelerate that thread toward the
4257 * processor it's on.
4259 * @preempt: whether task preemption is allowed or not
4261 * It's the caller's job to ensure that the target task struct
4262 * can't go away on us before we can do any checks.
4265 * true (>0) if we indeed boosted the target task.
4266 * false (0) if we failed to boost the target.
4267 * -ESRCH if there's no task to yield to.
4269 int __sched yield_to(struct task_struct *p, bool preempt)
4271 struct task_struct *curr = current;
4272 struct rq *rq, *p_rq;
4273 unsigned long flags;
4276 local_irq_save(flags);
4282 * If we're the only runnable task on the rq and target rq also
4283 * has only one task, there's absolutely no point in yielding.
4285 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4290 double_rq_lock(rq, p_rq);
4291 if (task_rq(p) != p_rq) {
4292 double_rq_unlock(rq, p_rq);
4296 if (!curr->sched_class->yield_to_task)
4299 if (curr->sched_class != p->sched_class)
4302 if (task_running(p_rq, p) || p->state)
4305 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4307 schedstat_inc(rq, yld_count);
4309 * Make p's CPU reschedule; pick_next_entity takes care of
4312 if (preempt && rq != p_rq)
4317 double_rq_unlock(rq, p_rq);
4319 local_irq_restore(flags);
4326 EXPORT_SYMBOL_GPL(yield_to);
4329 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4330 * that process accounting knows that this is a task in IO wait state.
4332 void __sched io_schedule(void)
4334 struct rq *rq = raw_rq();
4336 delayacct_blkio_start();
4337 atomic_inc(&rq->nr_iowait);
4338 blk_flush_plug(current);
4339 current->in_iowait = 1;
4341 current->in_iowait = 0;
4342 atomic_dec(&rq->nr_iowait);
4343 delayacct_blkio_end();
4345 EXPORT_SYMBOL(io_schedule);
4347 long __sched io_schedule_timeout(long timeout)
4349 struct rq *rq = raw_rq();
4352 delayacct_blkio_start();
4353 atomic_inc(&rq->nr_iowait);
4354 blk_flush_plug(current);
4355 current->in_iowait = 1;
4356 ret = schedule_timeout(timeout);
4357 current->in_iowait = 0;
4358 atomic_dec(&rq->nr_iowait);
4359 delayacct_blkio_end();
4364 * sys_sched_get_priority_max - return maximum RT priority.
4365 * @policy: scheduling class.
4367 * Return: On success, this syscall returns the maximum
4368 * rt_priority that can be used by a given scheduling class.
4369 * On failure, a negative error code is returned.
4371 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4378 ret = MAX_USER_RT_PRIO-1;
4380 case SCHED_DEADLINE:
4391 * sys_sched_get_priority_min - return minimum RT priority.
4392 * @policy: scheduling class.
4394 * Return: On success, this syscall returns the minimum
4395 * rt_priority that can be used by a given scheduling class.
4396 * On failure, a negative error code is returned.
4398 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4407 case SCHED_DEADLINE:
4417 * sys_sched_rr_get_interval - return the default timeslice of a process.
4418 * @pid: pid of the process.
4419 * @interval: userspace pointer to the timeslice value.
4421 * this syscall writes the default timeslice value of a given process
4422 * into the user-space timespec buffer. A value of '0' means infinity.
4424 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4427 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4428 struct timespec __user *, interval)
4430 struct task_struct *p;
4431 unsigned int time_slice;
4432 unsigned long flags;
4442 p = find_process_by_pid(pid);
4446 retval = security_task_getscheduler(p);
4450 rq = task_rq_lock(p, &flags);
4452 if (p->sched_class->get_rr_interval)
4453 time_slice = p->sched_class->get_rr_interval(rq, p);
4454 task_rq_unlock(rq, p, &flags);
4457 jiffies_to_timespec(time_slice, &t);
4458 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4466 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4468 void sched_show_task(struct task_struct *p)
4470 unsigned long free = 0;
4474 state = p->state ? __ffs(p->state) + 1 : 0;
4475 printk(KERN_INFO "%-15.15s %c", p->comm,
4476 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4477 #if BITS_PER_LONG == 32
4478 if (state == TASK_RUNNING)
4479 printk(KERN_CONT " running ");
4481 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4483 if (state == TASK_RUNNING)
4484 printk(KERN_CONT " running task ");
4486 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4488 #ifdef CONFIG_DEBUG_STACK_USAGE
4489 free = stack_not_used(p);
4492 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4494 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4495 task_pid_nr(p), ppid,
4496 (unsigned long)task_thread_info(p)->flags);
4498 print_worker_info(KERN_INFO, p);
4499 show_stack(p, NULL);
4502 void show_state_filter(unsigned long state_filter)
4504 struct task_struct *g, *p;
4506 #if BITS_PER_LONG == 32
4508 " task PC stack pid father\n");
4511 " task PC stack pid father\n");
4514 for_each_process_thread(g, p) {
4516 * reset the NMI-timeout, listing all files on a slow
4517 * console might take a lot of time:
4519 touch_nmi_watchdog();
4520 if (!state_filter || (p->state & state_filter))
4524 touch_all_softlockup_watchdogs();
4526 #ifdef CONFIG_SCHED_DEBUG
4527 sysrq_sched_debug_show();
4531 * Only show locks if all tasks are dumped:
4534 debug_show_all_locks();
4537 void init_idle_bootup_task(struct task_struct *idle)
4539 idle->sched_class = &idle_sched_class;
4543 * init_idle - set up an idle thread for a given CPU
4544 * @idle: task in question
4545 * @cpu: cpu the idle task belongs to
4547 * NOTE: this function does not set the idle thread's NEED_RESCHED
4548 * flag, to make booting more robust.
4550 void init_idle(struct task_struct *idle, int cpu)
4552 struct rq *rq = cpu_rq(cpu);
4553 unsigned long flags;
4555 raw_spin_lock_irqsave(&rq->lock, flags);
4557 __sched_fork(0, idle);
4558 idle->state = TASK_RUNNING;
4559 idle->se.exec_start = sched_clock();
4561 do_set_cpus_allowed(idle, cpumask_of(cpu));
4563 * We're having a chicken and egg problem, even though we are
4564 * holding rq->lock, the cpu isn't yet set to this cpu so the
4565 * lockdep check in task_group() will fail.
4567 * Similar case to sched_fork(). / Alternatively we could
4568 * use task_rq_lock() here and obtain the other rq->lock.
4573 __set_task_cpu(idle, cpu);
4576 rq->curr = rq->idle = idle;
4577 idle->on_rq = TASK_ON_RQ_QUEUED;
4578 #if defined(CONFIG_SMP)
4581 raw_spin_unlock_irqrestore(&rq->lock, flags);
4583 /* Set the preempt count _outside_ the spinlocks! */
4584 init_idle_preempt_count(idle, cpu);
4587 * The idle tasks have their own, simple scheduling class:
4589 idle->sched_class = &idle_sched_class;
4590 ftrace_graph_init_idle_task(idle, cpu);
4591 vtime_init_idle(idle, cpu);
4592 #if defined(CONFIG_SMP)
4593 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4598 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4600 if (p->sched_class && p->sched_class->set_cpus_allowed)
4601 p->sched_class->set_cpus_allowed(p, new_mask);
4603 cpumask_copy(&p->cpus_allowed, new_mask);
4604 p->nr_cpus_allowed = cpumask_weight(new_mask);
4608 * This is how migration works:
4610 * 1) we invoke migration_cpu_stop() on the target CPU using
4612 * 2) stopper starts to run (implicitly forcing the migrated thread
4614 * 3) it checks whether the migrated task is still in the wrong runqueue.
4615 * 4) if it's in the wrong runqueue then the migration thread removes
4616 * it and puts it into the right queue.
4617 * 5) stopper completes and stop_one_cpu() returns and the migration
4622 * Change a given task's CPU affinity. Migrate the thread to a
4623 * proper CPU and schedule it away if the CPU it's executing on
4624 * is removed from the allowed bitmask.
4626 * NOTE: the caller must have a valid reference to the task, the
4627 * task must not exit() & deallocate itself prematurely. The
4628 * call is not atomic; no spinlocks may be held.
4630 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4632 unsigned long flags;
4634 unsigned int dest_cpu;
4637 rq = task_rq_lock(p, &flags);
4639 if (cpumask_equal(&p->cpus_allowed, new_mask))
4642 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4647 do_set_cpus_allowed(p, new_mask);
4649 /* Can the task run on the task's current CPU? If so, we're done */
4650 if (cpumask_test_cpu(task_cpu(p), new_mask))
4653 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4654 if (task_on_rq_queued(p)) {
4655 struct migration_arg arg = { p, dest_cpu };
4656 /* Need help from migration thread: drop lock and wait. */
4657 task_rq_unlock(rq, p, &flags);
4658 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4659 tlb_migrate_finish(p->mm);
4663 task_rq_unlock(rq, p, &flags);
4667 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4670 * Move (not current) task off this cpu, onto dest cpu. We're doing
4671 * this because either it can't run here any more (set_cpus_allowed()
4672 * away from this CPU, or CPU going down), or because we're
4673 * attempting to rebalance this task on exec (sched_exec).
4675 * So we race with normal scheduler movements, but that's OK, as long
4676 * as the task is no longer on this CPU.
4678 * Returns non-zero if task was successfully migrated.
4680 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4682 struct rq *rq_dest, *rq_src;
4685 if (unlikely(!cpu_active(dest_cpu)))
4688 rq_src = cpu_rq(src_cpu);
4689 rq_dest = cpu_rq(dest_cpu);
4691 raw_spin_lock(&p->pi_lock);
4692 double_rq_lock(rq_src, rq_dest);
4693 /* Already moved. */
4694 if (task_cpu(p) != src_cpu)
4696 /* Affinity changed (again). */
4697 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4701 * If we're not on a rq, the next wake-up will ensure we're
4704 if (task_on_rq_queued(p)) {
4705 dequeue_task(rq_src, p, 0);
4706 set_task_cpu(p, dest_cpu);
4707 enqueue_task(rq_dest, p, 0);
4708 check_preempt_curr(rq_dest, p, 0);
4713 double_rq_unlock(rq_src, rq_dest);
4714 raw_spin_unlock(&p->pi_lock);
4718 #ifdef CONFIG_NUMA_BALANCING
4719 /* Migrate current task p to target_cpu */
4720 int migrate_task_to(struct task_struct *p, int target_cpu)
4722 struct migration_arg arg = { p, target_cpu };
4723 int curr_cpu = task_cpu(p);
4725 if (curr_cpu == target_cpu)
4728 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4731 /* TODO: This is not properly updating schedstats */
4733 trace_sched_move_numa(p, curr_cpu, target_cpu);
4734 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4738 * Requeue a task on a given node and accurately track the number of NUMA
4739 * tasks on the runqueues
4741 void sched_setnuma(struct task_struct *p, int nid)
4744 unsigned long flags;
4745 bool queued, running;
4747 rq = task_rq_lock(p, &flags);
4748 queued = task_on_rq_queued(p);
4749 running = task_current(rq, p);
4752 dequeue_task(rq, p, 0);
4754 p->sched_class->put_prev_task(rq, p);
4756 p->numa_preferred_nid = nid;
4759 p->sched_class->set_curr_task(rq);
4761 enqueue_task(rq, p, 0);
4762 task_rq_unlock(rq, p, &flags);
4767 * migration_cpu_stop - this will be executed by a highprio stopper thread
4768 * and performs thread migration by bumping thread off CPU then
4769 * 'pushing' onto another runqueue.
4771 static int migration_cpu_stop(void *data)
4773 struct migration_arg *arg = data;
4776 * The original target cpu might have gone down and we might
4777 * be on another cpu but it doesn't matter.
4779 local_irq_disable();
4780 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4785 #ifdef CONFIG_HOTPLUG_CPU
4788 * Ensures that the idle task is using init_mm right before its cpu goes
4791 void idle_task_exit(void)
4793 struct mm_struct *mm = current->active_mm;
4795 BUG_ON(cpu_online(smp_processor_id()));
4797 if (mm != &init_mm) {
4798 switch_mm(mm, &init_mm, current);
4799 finish_arch_post_lock_switch();
4805 * Since this CPU is going 'away' for a while, fold any nr_active delta
4806 * we might have. Assumes we're called after migrate_tasks() so that the
4807 * nr_active count is stable.
4809 * Also see the comment "Global load-average calculations".
4811 static void calc_load_migrate(struct rq *rq)
4813 long delta = calc_load_fold_active(rq);
4815 atomic_long_add(delta, &calc_load_tasks);
4818 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4822 static const struct sched_class fake_sched_class = {
4823 .put_prev_task = put_prev_task_fake,
4826 static struct task_struct fake_task = {
4828 * Avoid pull_{rt,dl}_task()
4830 .prio = MAX_PRIO + 1,
4831 .sched_class = &fake_sched_class,
4835 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4836 * try_to_wake_up()->select_task_rq().
4838 * Called with rq->lock held even though we'er in stop_machine() and
4839 * there's no concurrency possible, we hold the required locks anyway
4840 * because of lock validation efforts.
4842 static void migrate_tasks(unsigned int dead_cpu)
4844 struct rq *rq = cpu_rq(dead_cpu);
4845 struct task_struct *next, *stop = rq->stop;
4849 * Fudge the rq selection such that the below task selection loop
4850 * doesn't get stuck on the currently eligible stop task.
4852 * We're currently inside stop_machine() and the rq is either stuck
4853 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4854 * either way we should never end up calling schedule() until we're
4860 * put_prev_task() and pick_next_task() sched
4861 * class method both need to have an up-to-date
4862 * value of rq->clock[_task]
4864 update_rq_clock(rq);
4868 * There's this thread running, bail when that's the only
4871 if (rq->nr_running == 1)
4874 next = pick_next_task(rq, &fake_task);
4876 next->sched_class->put_prev_task(rq, next);
4878 /* Find suitable destination for @next, with force if needed. */
4879 dest_cpu = select_fallback_rq(dead_cpu, next);
4880 raw_spin_unlock(&rq->lock);
4882 __migrate_task(next, dead_cpu, dest_cpu);
4884 raw_spin_lock(&rq->lock);
4890 #endif /* CONFIG_HOTPLUG_CPU */
4892 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4894 static struct ctl_table sd_ctl_dir[] = {
4896 .procname = "sched_domain",
4902 static struct ctl_table sd_ctl_root[] = {
4904 .procname = "kernel",
4906 .child = sd_ctl_dir,
4911 static struct ctl_table *sd_alloc_ctl_entry(int n)
4913 struct ctl_table *entry =
4914 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4919 static void sd_free_ctl_entry(struct ctl_table **tablep)
4921 struct ctl_table *entry;
4924 * In the intermediate directories, both the child directory and
4925 * procname are dynamically allocated and could fail but the mode
4926 * will always be set. In the lowest directory the names are
4927 * static strings and all have proc handlers.
4929 for (entry = *tablep; entry->mode; entry++) {
4931 sd_free_ctl_entry(&entry->child);
4932 if (entry->proc_handler == NULL)
4933 kfree(entry->procname);
4940 static int min_load_idx = 0;
4941 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4944 set_table_entry(struct ctl_table *entry,
4945 const char *procname, void *data, int maxlen,
4946 umode_t mode, proc_handler *proc_handler,
4949 entry->procname = procname;
4951 entry->maxlen = maxlen;
4953 entry->proc_handler = proc_handler;
4956 entry->extra1 = &min_load_idx;
4957 entry->extra2 = &max_load_idx;
4961 static struct ctl_table *
4962 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4964 struct ctl_table *table = sd_alloc_ctl_entry(14);
4969 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4970 sizeof(long), 0644, proc_doulongvec_minmax, false);
4971 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4972 sizeof(long), 0644, proc_doulongvec_minmax, false);
4973 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4974 sizeof(int), 0644, proc_dointvec_minmax, true);
4975 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4976 sizeof(int), 0644, proc_dointvec_minmax, true);
4977 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4978 sizeof(int), 0644, proc_dointvec_minmax, true);
4979 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4980 sizeof(int), 0644, proc_dointvec_minmax, true);
4981 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4982 sizeof(int), 0644, proc_dointvec_minmax, true);
4983 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4984 sizeof(int), 0644, proc_dointvec_minmax, false);
4985 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4986 sizeof(int), 0644, proc_dointvec_minmax, false);
4987 set_table_entry(&table[9], "cache_nice_tries",
4988 &sd->cache_nice_tries,
4989 sizeof(int), 0644, proc_dointvec_minmax, false);
4990 set_table_entry(&table[10], "flags", &sd->flags,
4991 sizeof(int), 0644, proc_dointvec_minmax, false);
4992 set_table_entry(&table[11], "max_newidle_lb_cost",
4993 &sd->max_newidle_lb_cost,
4994 sizeof(long), 0644, proc_doulongvec_minmax, false);
4995 set_table_entry(&table[12], "name", sd->name,
4996 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4997 /* &table[13] is terminator */
5002 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5004 struct ctl_table *entry, *table;
5005 struct sched_domain *sd;
5006 int domain_num = 0, i;
5009 for_each_domain(cpu, sd)
5011 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5016 for_each_domain(cpu, sd) {
5017 snprintf(buf, 32, "domain%d", i);
5018 entry->procname = kstrdup(buf, GFP_KERNEL);
5020 entry->child = sd_alloc_ctl_domain_table(sd);
5027 static struct ctl_table_header *sd_sysctl_header;
5028 static void register_sched_domain_sysctl(void)
5030 int i, cpu_num = num_possible_cpus();
5031 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5034 WARN_ON(sd_ctl_dir[0].child);
5035 sd_ctl_dir[0].child = entry;
5040 for_each_possible_cpu(i) {
5041 snprintf(buf, 32, "cpu%d", i);
5042 entry->procname = kstrdup(buf, GFP_KERNEL);
5044 entry->child = sd_alloc_ctl_cpu_table(i);
5048 WARN_ON(sd_sysctl_header);
5049 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5052 /* may be called multiple times per register */
5053 static void unregister_sched_domain_sysctl(void)
5055 if (sd_sysctl_header)
5056 unregister_sysctl_table(sd_sysctl_header);
5057 sd_sysctl_header = NULL;
5058 if (sd_ctl_dir[0].child)
5059 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5062 static void register_sched_domain_sysctl(void)
5065 static void unregister_sched_domain_sysctl(void)
5070 static void set_rq_online(struct rq *rq)
5073 const struct sched_class *class;
5075 cpumask_set_cpu(rq->cpu, rq->rd->online);
5078 for_each_class(class) {
5079 if (class->rq_online)
5080 class->rq_online(rq);
5085 static void set_rq_offline(struct rq *rq)
5088 const struct sched_class *class;
5090 for_each_class(class) {
5091 if (class->rq_offline)
5092 class->rq_offline(rq);
5095 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5101 * migration_call - callback that gets triggered when a CPU is added.
5102 * Here we can start up the necessary migration thread for the new CPU.
5105 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5107 int cpu = (long)hcpu;
5108 unsigned long flags;
5109 struct rq *rq = cpu_rq(cpu);
5111 switch (action & ~CPU_TASKS_FROZEN) {
5113 case CPU_UP_PREPARE:
5114 rq->calc_load_update = calc_load_update;
5118 /* Update our root-domain */
5119 raw_spin_lock_irqsave(&rq->lock, flags);
5121 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5125 raw_spin_unlock_irqrestore(&rq->lock, flags);
5128 #ifdef CONFIG_HOTPLUG_CPU
5130 sched_ttwu_pending();
5131 /* Update our root-domain */
5132 raw_spin_lock_irqsave(&rq->lock, flags);
5134 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5138 BUG_ON(rq->nr_running != 1); /* the migration thread */
5139 raw_spin_unlock_irqrestore(&rq->lock, flags);
5143 calc_load_migrate(rq);
5148 update_max_interval();
5154 * Register at high priority so that task migration (migrate_all_tasks)
5155 * happens before everything else. This has to be lower priority than
5156 * the notifier in the perf_event subsystem, though.
5158 static struct notifier_block migration_notifier = {
5159 .notifier_call = migration_call,
5160 .priority = CPU_PRI_MIGRATION,
5163 static void __cpuinit set_cpu_rq_start_time(void)
5165 int cpu = smp_processor_id();
5166 struct rq *rq = cpu_rq(cpu);
5167 rq->age_stamp = sched_clock_cpu(cpu);
5170 static int sched_cpu_active(struct notifier_block *nfb,
5171 unsigned long action, void *hcpu)
5173 switch (action & ~CPU_TASKS_FROZEN) {
5175 set_cpu_rq_start_time();
5177 case CPU_DOWN_FAILED:
5178 set_cpu_active((long)hcpu, true);
5185 static int sched_cpu_inactive(struct notifier_block *nfb,
5186 unsigned long action, void *hcpu)
5188 unsigned long flags;
5189 long cpu = (long)hcpu;
5191 switch (action & ~CPU_TASKS_FROZEN) {
5192 case CPU_DOWN_PREPARE:
5193 set_cpu_active(cpu, false);
5195 /* explicitly allow suspend */
5196 if (!(action & CPU_TASKS_FROZEN)) {
5197 struct dl_bw *dl_b = dl_bw_of(cpu);
5201 raw_spin_lock_irqsave(&dl_b->lock, flags);
5202 cpus = dl_bw_cpus(cpu);
5203 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5204 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5207 return notifier_from_errno(-EBUSY);
5215 static int __init migration_init(void)
5217 void *cpu = (void *)(long)smp_processor_id();
5220 /* Initialize migration for the boot CPU */
5221 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5222 BUG_ON(err == NOTIFY_BAD);
5223 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5224 register_cpu_notifier(&migration_notifier);
5226 /* Register cpu active notifiers */
5227 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5228 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5232 early_initcall(migration_init);
5237 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5239 #ifdef CONFIG_SCHED_DEBUG
5241 static __read_mostly int sched_debug_enabled;
5243 static int __init sched_debug_setup(char *str)
5245 sched_debug_enabled = 1;
5249 early_param("sched_debug", sched_debug_setup);
5251 static inline bool sched_debug(void)
5253 return sched_debug_enabled;
5256 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5257 struct cpumask *groupmask)
5259 struct sched_group *group = sd->groups;
5262 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5263 cpumask_clear(groupmask);
5265 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5267 if (!(sd->flags & SD_LOAD_BALANCE)) {
5268 printk("does not load-balance\n");
5270 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5275 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5277 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5278 printk(KERN_ERR "ERROR: domain->span does not contain "
5281 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5282 printk(KERN_ERR "ERROR: domain->groups does not contain"
5286 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5290 printk(KERN_ERR "ERROR: group is NULL\n");
5295 * Even though we initialize ->capacity to something semi-sane,
5296 * we leave capacity_orig unset. This allows us to detect if
5297 * domain iteration is still funny without causing /0 traps.
5299 if (!group->sgc->capacity_orig) {
5300 printk(KERN_CONT "\n");
5301 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5305 if (!cpumask_weight(sched_group_cpus(group))) {
5306 printk(KERN_CONT "\n");
5307 printk(KERN_ERR "ERROR: empty group\n");
5311 if (!(sd->flags & SD_OVERLAP) &&
5312 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5313 printk(KERN_CONT "\n");
5314 printk(KERN_ERR "ERROR: repeated CPUs\n");
5318 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5320 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5322 printk(KERN_CONT " %s", str);
5323 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5324 printk(KERN_CONT " (cpu_capacity = %d)",
5325 group->sgc->capacity);
5328 group = group->next;
5329 } while (group != sd->groups);
5330 printk(KERN_CONT "\n");
5332 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5333 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5336 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5337 printk(KERN_ERR "ERROR: parent span is not a superset "
5338 "of domain->span\n");
5342 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5346 if (!sched_debug_enabled)
5350 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5354 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5357 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5365 #else /* !CONFIG_SCHED_DEBUG */
5366 # define sched_domain_debug(sd, cpu) do { } while (0)
5367 static inline bool sched_debug(void)
5371 #endif /* CONFIG_SCHED_DEBUG */
5373 static int sd_degenerate(struct sched_domain *sd)
5375 if (cpumask_weight(sched_domain_span(sd)) == 1)
5378 /* Following flags need at least 2 groups */
5379 if (sd->flags & (SD_LOAD_BALANCE |
5380 SD_BALANCE_NEWIDLE |
5383 SD_SHARE_CPUCAPACITY |
5384 SD_SHARE_PKG_RESOURCES |
5385 SD_SHARE_POWERDOMAIN)) {
5386 if (sd->groups != sd->groups->next)
5390 /* Following flags don't use groups */
5391 if (sd->flags & (SD_WAKE_AFFINE))
5398 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5400 unsigned long cflags = sd->flags, pflags = parent->flags;
5402 if (sd_degenerate(parent))
5405 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5408 /* Flags needing groups don't count if only 1 group in parent */
5409 if (parent->groups == parent->groups->next) {
5410 pflags &= ~(SD_LOAD_BALANCE |
5411 SD_BALANCE_NEWIDLE |
5414 SD_SHARE_CPUCAPACITY |
5415 SD_SHARE_PKG_RESOURCES |
5417 SD_SHARE_POWERDOMAIN);
5418 if (nr_node_ids == 1)
5419 pflags &= ~SD_SERIALIZE;
5421 if (~cflags & pflags)
5427 static void free_rootdomain(struct rcu_head *rcu)
5429 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5431 cpupri_cleanup(&rd->cpupri);
5432 cpudl_cleanup(&rd->cpudl);
5433 free_cpumask_var(rd->dlo_mask);
5434 free_cpumask_var(rd->rto_mask);
5435 free_cpumask_var(rd->online);
5436 free_cpumask_var(rd->span);
5440 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5442 struct root_domain *old_rd = NULL;
5443 unsigned long flags;
5445 raw_spin_lock_irqsave(&rq->lock, flags);
5450 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5453 cpumask_clear_cpu(rq->cpu, old_rd->span);
5456 * If we dont want to free the old_rd yet then
5457 * set old_rd to NULL to skip the freeing later
5460 if (!atomic_dec_and_test(&old_rd->refcount))
5464 atomic_inc(&rd->refcount);
5467 cpumask_set_cpu(rq->cpu, rd->span);
5468 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5471 raw_spin_unlock_irqrestore(&rq->lock, flags);
5474 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5477 static int init_rootdomain(struct root_domain *rd)
5479 memset(rd, 0, sizeof(*rd));
5481 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5483 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5485 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5487 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5490 init_dl_bw(&rd->dl_bw);
5491 if (cpudl_init(&rd->cpudl) != 0)
5494 if (cpupri_init(&rd->cpupri) != 0)
5499 free_cpumask_var(rd->rto_mask);
5501 free_cpumask_var(rd->dlo_mask);
5503 free_cpumask_var(rd->online);
5505 free_cpumask_var(rd->span);
5511 * By default the system creates a single root-domain with all cpus as
5512 * members (mimicking the global state we have today).
5514 struct root_domain def_root_domain;
5516 static void init_defrootdomain(void)
5518 init_rootdomain(&def_root_domain);
5520 atomic_set(&def_root_domain.refcount, 1);
5523 static struct root_domain *alloc_rootdomain(void)
5525 struct root_domain *rd;
5527 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5531 if (init_rootdomain(rd) != 0) {
5539 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5541 struct sched_group *tmp, *first;
5550 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5555 } while (sg != first);
5558 static void free_sched_domain(struct rcu_head *rcu)
5560 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5563 * If its an overlapping domain it has private groups, iterate and
5566 if (sd->flags & SD_OVERLAP) {
5567 free_sched_groups(sd->groups, 1);
5568 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5569 kfree(sd->groups->sgc);
5575 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5577 call_rcu(&sd->rcu, free_sched_domain);
5580 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5582 for (; sd; sd = sd->parent)
5583 destroy_sched_domain(sd, cpu);
5587 * Keep a special pointer to the highest sched_domain that has
5588 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5589 * allows us to avoid some pointer chasing select_idle_sibling().
5591 * Also keep a unique ID per domain (we use the first cpu number in
5592 * the cpumask of the domain), this allows us to quickly tell if
5593 * two cpus are in the same cache domain, see cpus_share_cache().
5595 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5596 DEFINE_PER_CPU(int, sd_llc_size);
5597 DEFINE_PER_CPU(int, sd_llc_id);
5598 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5599 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5600 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5602 static void update_top_cache_domain(int cpu)
5604 struct sched_domain *sd;
5605 struct sched_domain *busy_sd = NULL;
5609 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5611 id = cpumask_first(sched_domain_span(sd));
5612 size = cpumask_weight(sched_domain_span(sd));
5613 busy_sd = sd->parent; /* sd_busy */
5615 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5617 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5618 per_cpu(sd_llc_size, cpu) = size;
5619 per_cpu(sd_llc_id, cpu) = id;
5621 sd = lowest_flag_domain(cpu, SD_NUMA);
5622 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5624 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5625 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5629 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5630 * hold the hotplug lock.
5633 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5635 struct rq *rq = cpu_rq(cpu);
5636 struct sched_domain *tmp;
5638 /* Remove the sched domains which do not contribute to scheduling. */
5639 for (tmp = sd; tmp; ) {
5640 struct sched_domain *parent = tmp->parent;
5644 if (sd_parent_degenerate(tmp, parent)) {
5645 tmp->parent = parent->parent;
5647 parent->parent->child = tmp;
5649 * Transfer SD_PREFER_SIBLING down in case of a
5650 * degenerate parent; the spans match for this
5651 * so the property transfers.
5653 if (parent->flags & SD_PREFER_SIBLING)
5654 tmp->flags |= SD_PREFER_SIBLING;
5655 destroy_sched_domain(parent, cpu);
5660 if (sd && sd_degenerate(sd)) {
5663 destroy_sched_domain(tmp, cpu);
5668 sched_domain_debug(sd, cpu);
5670 rq_attach_root(rq, rd);
5672 rcu_assign_pointer(rq->sd, sd);
5673 destroy_sched_domains(tmp, cpu);
5675 update_top_cache_domain(cpu);
5678 /* cpus with isolated domains */
5679 static cpumask_var_t cpu_isolated_map;
5681 /* Setup the mask of cpus configured for isolated domains */
5682 static int __init isolated_cpu_setup(char *str)
5684 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5685 cpulist_parse(str, cpu_isolated_map);
5689 __setup("isolcpus=", isolated_cpu_setup);
5692 struct sched_domain ** __percpu sd;
5693 struct root_domain *rd;
5704 * Build an iteration mask that can exclude certain CPUs from the upwards
5707 * Asymmetric node setups can result in situations where the domain tree is of
5708 * unequal depth, make sure to skip domains that already cover the entire
5711 * In that case build_sched_domains() will have terminated the iteration early
5712 * and our sibling sd spans will be empty. Domains should always include the
5713 * cpu they're built on, so check that.
5716 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5718 const struct cpumask *span = sched_domain_span(sd);
5719 struct sd_data *sdd = sd->private;
5720 struct sched_domain *sibling;
5723 for_each_cpu(i, span) {
5724 sibling = *per_cpu_ptr(sdd->sd, i);
5725 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5728 cpumask_set_cpu(i, sched_group_mask(sg));
5733 * Return the canonical balance cpu for this group, this is the first cpu
5734 * of this group that's also in the iteration mask.
5736 int group_balance_cpu(struct sched_group *sg)
5738 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5742 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5744 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5745 const struct cpumask *span = sched_domain_span(sd);
5746 struct cpumask *covered = sched_domains_tmpmask;
5747 struct sd_data *sdd = sd->private;
5748 struct sched_domain *sibling;
5751 cpumask_clear(covered);
5753 for_each_cpu(i, span) {
5754 struct cpumask *sg_span;
5756 if (cpumask_test_cpu(i, covered))
5759 sibling = *per_cpu_ptr(sdd->sd, i);
5761 /* See the comment near build_group_mask(). */
5762 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5765 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5766 GFP_KERNEL, cpu_to_node(cpu));
5771 sg_span = sched_group_cpus(sg);
5773 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5775 cpumask_set_cpu(i, sg_span);
5777 cpumask_or(covered, covered, sg_span);
5779 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5780 if (atomic_inc_return(&sg->sgc->ref) == 1)
5781 build_group_mask(sd, sg);
5784 * Initialize sgc->capacity such that even if we mess up the
5785 * domains and no possible iteration will get us here, we won't
5788 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5789 sg->sgc->capacity_orig = sg->sgc->capacity;
5792 * Make sure the first group of this domain contains the
5793 * canonical balance cpu. Otherwise the sched_domain iteration
5794 * breaks. See update_sg_lb_stats().
5796 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5797 group_balance_cpu(sg) == cpu)
5807 sd->groups = groups;
5812 free_sched_groups(first, 0);
5817 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5819 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5820 struct sched_domain *child = sd->child;
5823 cpu = cpumask_first(sched_domain_span(child));
5826 *sg = *per_cpu_ptr(sdd->sg, cpu);
5827 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5828 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5835 * build_sched_groups will build a circular linked list of the groups
5836 * covered by the given span, and will set each group's ->cpumask correctly,
5837 * and ->cpu_capacity to 0.
5839 * Assumes the sched_domain tree is fully constructed
5842 build_sched_groups(struct sched_domain *sd, int cpu)
5844 struct sched_group *first = NULL, *last = NULL;
5845 struct sd_data *sdd = sd->private;
5846 const struct cpumask *span = sched_domain_span(sd);
5847 struct cpumask *covered;
5850 get_group(cpu, sdd, &sd->groups);
5851 atomic_inc(&sd->groups->ref);
5853 if (cpu != cpumask_first(span))
5856 lockdep_assert_held(&sched_domains_mutex);
5857 covered = sched_domains_tmpmask;
5859 cpumask_clear(covered);
5861 for_each_cpu(i, span) {
5862 struct sched_group *sg;
5865 if (cpumask_test_cpu(i, covered))
5868 group = get_group(i, sdd, &sg);
5869 cpumask_setall(sched_group_mask(sg));
5871 for_each_cpu(j, span) {
5872 if (get_group(j, sdd, NULL) != group)
5875 cpumask_set_cpu(j, covered);
5876 cpumask_set_cpu(j, sched_group_cpus(sg));
5891 * Initialize sched groups cpu_capacity.
5893 * cpu_capacity indicates the capacity of sched group, which is used while
5894 * distributing the load between different sched groups in a sched domain.
5895 * Typically cpu_capacity for all the groups in a sched domain will be same
5896 * unless there are asymmetries in the topology. If there are asymmetries,
5897 * group having more cpu_capacity will pickup more load compared to the
5898 * group having less cpu_capacity.
5900 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5902 struct sched_group *sg = sd->groups;
5907 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5909 } while (sg != sd->groups);
5911 if (cpu != group_balance_cpu(sg))
5914 update_group_capacity(sd, cpu);
5915 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
5919 * Initializers for schedule domains
5920 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5923 static int default_relax_domain_level = -1;
5924 int sched_domain_level_max;
5926 static int __init setup_relax_domain_level(char *str)
5928 if (kstrtoint(str, 0, &default_relax_domain_level))
5929 pr_warn("Unable to set relax_domain_level\n");
5933 __setup("relax_domain_level=", setup_relax_domain_level);
5935 static void set_domain_attribute(struct sched_domain *sd,
5936 struct sched_domain_attr *attr)
5940 if (!attr || attr->relax_domain_level < 0) {
5941 if (default_relax_domain_level < 0)
5944 request = default_relax_domain_level;
5946 request = attr->relax_domain_level;
5947 if (request < sd->level) {
5948 /* turn off idle balance on this domain */
5949 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5951 /* turn on idle balance on this domain */
5952 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5956 static void __sdt_free(const struct cpumask *cpu_map);
5957 static int __sdt_alloc(const struct cpumask *cpu_map);
5959 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5960 const struct cpumask *cpu_map)
5964 if (!atomic_read(&d->rd->refcount))
5965 free_rootdomain(&d->rd->rcu); /* fall through */
5967 free_percpu(d->sd); /* fall through */
5969 __sdt_free(cpu_map); /* fall through */
5975 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5976 const struct cpumask *cpu_map)
5978 memset(d, 0, sizeof(*d));
5980 if (__sdt_alloc(cpu_map))
5981 return sa_sd_storage;
5982 d->sd = alloc_percpu(struct sched_domain *);
5984 return sa_sd_storage;
5985 d->rd = alloc_rootdomain();
5988 return sa_rootdomain;
5992 * NULL the sd_data elements we've used to build the sched_domain and
5993 * sched_group structure so that the subsequent __free_domain_allocs()
5994 * will not free the data we're using.
5996 static void claim_allocations(int cpu, struct sched_domain *sd)
5998 struct sd_data *sdd = sd->private;
6000 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6001 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6003 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6004 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6006 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6007 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6011 static int sched_domains_numa_levels;
6012 static int *sched_domains_numa_distance;
6013 static struct cpumask ***sched_domains_numa_masks;
6014 static int sched_domains_curr_level;
6018 * SD_flags allowed in topology descriptions.
6020 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6021 * SD_SHARE_PKG_RESOURCES - describes shared caches
6022 * SD_NUMA - describes NUMA topologies
6023 * SD_SHARE_POWERDOMAIN - describes shared power domain
6026 * SD_ASYM_PACKING - describes SMT quirks
6028 #define TOPOLOGY_SD_FLAGS \
6029 (SD_SHARE_CPUCAPACITY | \
6030 SD_SHARE_PKG_RESOURCES | \
6033 SD_SHARE_POWERDOMAIN)
6035 static struct sched_domain *
6036 sd_init(struct sched_domain_topology_level *tl, int cpu)
6038 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6039 int sd_weight, sd_flags = 0;
6043 * Ugly hack to pass state to sd_numa_mask()...
6045 sched_domains_curr_level = tl->numa_level;
6048 sd_weight = cpumask_weight(tl->mask(cpu));
6051 sd_flags = (*tl->sd_flags)();
6052 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6053 "wrong sd_flags in topology description\n"))
6054 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6056 *sd = (struct sched_domain){
6057 .min_interval = sd_weight,
6058 .max_interval = 2*sd_weight,
6060 .imbalance_pct = 125,
6062 .cache_nice_tries = 0,
6069 .flags = 1*SD_LOAD_BALANCE
6070 | 1*SD_BALANCE_NEWIDLE
6075 | 0*SD_SHARE_CPUCAPACITY
6076 | 0*SD_SHARE_PKG_RESOURCES
6078 | 0*SD_PREFER_SIBLING
6083 .last_balance = jiffies,
6084 .balance_interval = sd_weight,
6086 .max_newidle_lb_cost = 0,
6087 .next_decay_max_lb_cost = jiffies,
6088 #ifdef CONFIG_SCHED_DEBUG
6094 * Convert topological properties into behaviour.
6097 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6098 sd->imbalance_pct = 110;
6099 sd->smt_gain = 1178; /* ~15% */
6101 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6102 sd->imbalance_pct = 117;
6103 sd->cache_nice_tries = 1;
6107 } else if (sd->flags & SD_NUMA) {
6108 sd->cache_nice_tries = 2;
6112 sd->flags |= SD_SERIALIZE;
6113 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6114 sd->flags &= ~(SD_BALANCE_EXEC |
6121 sd->flags |= SD_PREFER_SIBLING;
6122 sd->cache_nice_tries = 1;
6127 sd->private = &tl->data;
6133 * Topology list, bottom-up.
6135 static struct sched_domain_topology_level default_topology[] = {
6136 #ifdef CONFIG_SCHED_SMT
6137 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6139 #ifdef CONFIG_SCHED_MC
6140 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6142 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6146 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6148 #define for_each_sd_topology(tl) \
6149 for (tl = sched_domain_topology; tl->mask; tl++)
6151 void set_sched_topology(struct sched_domain_topology_level *tl)
6153 sched_domain_topology = tl;
6158 static const struct cpumask *sd_numa_mask(int cpu)
6160 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6163 static void sched_numa_warn(const char *str)
6165 static int done = false;
6173 printk(KERN_WARNING "ERROR: %s\n\n", str);
6175 for (i = 0; i < nr_node_ids; i++) {
6176 printk(KERN_WARNING " ");
6177 for (j = 0; j < nr_node_ids; j++)
6178 printk(KERN_CONT "%02d ", node_distance(i,j));
6179 printk(KERN_CONT "\n");
6181 printk(KERN_WARNING "\n");
6184 static bool find_numa_distance(int distance)
6188 if (distance == node_distance(0, 0))
6191 for (i = 0; i < sched_domains_numa_levels; i++) {
6192 if (sched_domains_numa_distance[i] == distance)
6199 static void sched_init_numa(void)
6201 int next_distance, curr_distance = node_distance(0, 0);
6202 struct sched_domain_topology_level *tl;
6206 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6207 if (!sched_domains_numa_distance)
6211 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6212 * unique distances in the node_distance() table.
6214 * Assumes node_distance(0,j) includes all distances in
6215 * node_distance(i,j) in order to avoid cubic time.
6217 next_distance = curr_distance;
6218 for (i = 0; i < nr_node_ids; i++) {
6219 for (j = 0; j < nr_node_ids; j++) {
6220 for (k = 0; k < nr_node_ids; k++) {
6221 int distance = node_distance(i, k);
6223 if (distance > curr_distance &&
6224 (distance < next_distance ||
6225 next_distance == curr_distance))
6226 next_distance = distance;
6229 * While not a strong assumption it would be nice to know
6230 * about cases where if node A is connected to B, B is not
6231 * equally connected to A.
6233 if (sched_debug() && node_distance(k, i) != distance)
6234 sched_numa_warn("Node-distance not symmetric");
6236 if (sched_debug() && i && !find_numa_distance(distance))
6237 sched_numa_warn("Node-0 not representative");
6239 if (next_distance != curr_distance) {
6240 sched_domains_numa_distance[level++] = next_distance;
6241 sched_domains_numa_levels = level;
6242 curr_distance = next_distance;
6247 * In case of sched_debug() we verify the above assumption.
6253 * 'level' contains the number of unique distances, excluding the
6254 * identity distance node_distance(i,i).
6256 * The sched_domains_numa_distance[] array includes the actual distance
6261 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6262 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6263 * the array will contain less then 'level' members. This could be
6264 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6265 * in other functions.
6267 * We reset it to 'level' at the end of this function.
6269 sched_domains_numa_levels = 0;
6271 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6272 if (!sched_domains_numa_masks)
6276 * Now for each level, construct a mask per node which contains all
6277 * cpus of nodes that are that many hops away from us.
6279 for (i = 0; i < level; i++) {
6280 sched_domains_numa_masks[i] =
6281 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6282 if (!sched_domains_numa_masks[i])
6285 for (j = 0; j < nr_node_ids; j++) {
6286 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6290 sched_domains_numa_masks[i][j] = mask;
6292 for (k = 0; k < nr_node_ids; k++) {
6293 if (node_distance(j, k) > sched_domains_numa_distance[i])
6296 cpumask_or(mask, mask, cpumask_of_node(k));
6301 /* Compute default topology size */
6302 for (i = 0; sched_domain_topology[i].mask; i++);
6304 tl = kzalloc((i + level + 1) *
6305 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6310 * Copy the default topology bits..
6312 for (i = 0; sched_domain_topology[i].mask; i++)
6313 tl[i] = sched_domain_topology[i];
6316 * .. and append 'j' levels of NUMA goodness.
6318 for (j = 0; j < level; i++, j++) {
6319 tl[i] = (struct sched_domain_topology_level){
6320 .mask = sd_numa_mask,
6321 .sd_flags = cpu_numa_flags,
6322 .flags = SDTL_OVERLAP,
6328 sched_domain_topology = tl;
6330 sched_domains_numa_levels = level;
6333 static void sched_domains_numa_masks_set(int cpu)
6336 int node = cpu_to_node(cpu);
6338 for (i = 0; i < sched_domains_numa_levels; i++) {
6339 for (j = 0; j < nr_node_ids; j++) {
6340 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6341 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6346 static void sched_domains_numa_masks_clear(int cpu)
6349 for (i = 0; i < sched_domains_numa_levels; i++) {
6350 for (j = 0; j < nr_node_ids; j++)
6351 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6356 * Update sched_domains_numa_masks[level][node] array when new cpus
6359 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6360 unsigned long action,
6363 int cpu = (long)hcpu;
6365 switch (action & ~CPU_TASKS_FROZEN) {
6367 sched_domains_numa_masks_set(cpu);
6371 sched_domains_numa_masks_clear(cpu);
6381 static inline void sched_init_numa(void)
6385 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6386 unsigned long action,
6391 #endif /* CONFIG_NUMA */
6393 static int __sdt_alloc(const struct cpumask *cpu_map)
6395 struct sched_domain_topology_level *tl;
6398 for_each_sd_topology(tl) {
6399 struct sd_data *sdd = &tl->data;
6401 sdd->sd = alloc_percpu(struct sched_domain *);
6405 sdd->sg = alloc_percpu(struct sched_group *);
6409 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6413 for_each_cpu(j, cpu_map) {
6414 struct sched_domain *sd;
6415 struct sched_group *sg;
6416 struct sched_group_capacity *sgc;
6418 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6419 GFP_KERNEL, cpu_to_node(j));
6423 *per_cpu_ptr(sdd->sd, j) = sd;
6425 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6426 GFP_KERNEL, cpu_to_node(j));
6432 *per_cpu_ptr(sdd->sg, j) = sg;
6434 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6435 GFP_KERNEL, cpu_to_node(j));
6439 *per_cpu_ptr(sdd->sgc, j) = sgc;
6446 static void __sdt_free(const struct cpumask *cpu_map)
6448 struct sched_domain_topology_level *tl;
6451 for_each_sd_topology(tl) {
6452 struct sd_data *sdd = &tl->data;
6454 for_each_cpu(j, cpu_map) {
6455 struct sched_domain *sd;
6458 sd = *per_cpu_ptr(sdd->sd, j);
6459 if (sd && (sd->flags & SD_OVERLAP))
6460 free_sched_groups(sd->groups, 0);
6461 kfree(*per_cpu_ptr(sdd->sd, j));
6465 kfree(*per_cpu_ptr(sdd->sg, j));
6467 kfree(*per_cpu_ptr(sdd->sgc, j));
6469 free_percpu(sdd->sd);
6471 free_percpu(sdd->sg);
6473 free_percpu(sdd->sgc);
6478 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6479 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6480 struct sched_domain *child, int cpu)
6482 struct sched_domain *sd = sd_init(tl, cpu);
6486 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6488 sd->level = child->level + 1;
6489 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6493 if (!cpumask_subset(sched_domain_span(child),
6494 sched_domain_span(sd))) {
6495 pr_err("BUG: arch topology borken\n");
6496 #ifdef CONFIG_SCHED_DEBUG
6497 pr_err(" the %s domain not a subset of the %s domain\n",
6498 child->name, sd->name);
6500 /* Fixup, ensure @sd has at least @child cpus. */
6501 cpumask_or(sched_domain_span(sd),
6502 sched_domain_span(sd),
6503 sched_domain_span(child));
6507 set_domain_attribute(sd, attr);
6513 * Build sched domains for a given set of cpus and attach the sched domains
6514 * to the individual cpus
6516 static int build_sched_domains(const struct cpumask *cpu_map,
6517 struct sched_domain_attr *attr)
6519 enum s_alloc alloc_state;
6520 struct sched_domain *sd;
6522 int i, ret = -ENOMEM;
6524 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6525 if (alloc_state != sa_rootdomain)
6528 /* Set up domains for cpus specified by the cpu_map. */
6529 for_each_cpu(i, cpu_map) {
6530 struct sched_domain_topology_level *tl;
6533 for_each_sd_topology(tl) {
6534 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6535 if (tl == sched_domain_topology)
6536 *per_cpu_ptr(d.sd, i) = sd;
6537 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6538 sd->flags |= SD_OVERLAP;
6539 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6544 /* Build the groups for the domains */
6545 for_each_cpu(i, cpu_map) {
6546 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6547 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6548 if (sd->flags & SD_OVERLAP) {
6549 if (build_overlap_sched_groups(sd, i))
6552 if (build_sched_groups(sd, i))
6558 /* Calculate CPU capacity for physical packages and nodes */
6559 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6560 if (!cpumask_test_cpu(i, cpu_map))
6563 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6564 claim_allocations(i, sd);
6565 init_sched_groups_capacity(i, sd);
6569 /* Attach the domains */
6571 for_each_cpu(i, cpu_map) {
6572 sd = *per_cpu_ptr(d.sd, i);
6573 cpu_attach_domain(sd, d.rd, i);
6579 __free_domain_allocs(&d, alloc_state, cpu_map);
6583 static cpumask_var_t *doms_cur; /* current sched domains */
6584 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6585 static struct sched_domain_attr *dattr_cur;
6586 /* attribues of custom domains in 'doms_cur' */
6589 * Special case: If a kmalloc of a doms_cur partition (array of
6590 * cpumask) fails, then fallback to a single sched domain,
6591 * as determined by the single cpumask fallback_doms.
6593 static cpumask_var_t fallback_doms;
6596 * arch_update_cpu_topology lets virtualized architectures update the
6597 * cpu core maps. It is supposed to return 1 if the topology changed
6598 * or 0 if it stayed the same.
6600 int __weak arch_update_cpu_topology(void)
6605 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6608 cpumask_var_t *doms;
6610 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6613 for (i = 0; i < ndoms; i++) {
6614 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6615 free_sched_domains(doms, i);
6622 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6625 for (i = 0; i < ndoms; i++)
6626 free_cpumask_var(doms[i]);
6631 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6632 * For now this just excludes isolated cpus, but could be used to
6633 * exclude other special cases in the future.
6635 static int init_sched_domains(const struct cpumask *cpu_map)
6639 arch_update_cpu_topology();
6641 doms_cur = alloc_sched_domains(ndoms_cur);
6643 doms_cur = &fallback_doms;
6644 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6645 err = build_sched_domains(doms_cur[0], NULL);
6646 register_sched_domain_sysctl();
6652 * Detach sched domains from a group of cpus specified in cpu_map
6653 * These cpus will now be attached to the NULL domain
6655 static void detach_destroy_domains(const struct cpumask *cpu_map)
6660 for_each_cpu(i, cpu_map)
6661 cpu_attach_domain(NULL, &def_root_domain, i);
6665 /* handle null as "default" */
6666 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6667 struct sched_domain_attr *new, int idx_new)
6669 struct sched_domain_attr tmp;
6676 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6677 new ? (new + idx_new) : &tmp,
6678 sizeof(struct sched_domain_attr));
6682 * Partition sched domains as specified by the 'ndoms_new'
6683 * cpumasks in the array doms_new[] of cpumasks. This compares
6684 * doms_new[] to the current sched domain partitioning, doms_cur[].
6685 * It destroys each deleted domain and builds each new domain.
6687 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6688 * The masks don't intersect (don't overlap.) We should setup one
6689 * sched domain for each mask. CPUs not in any of the cpumasks will
6690 * not be load balanced. If the same cpumask appears both in the
6691 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6694 * The passed in 'doms_new' should be allocated using
6695 * alloc_sched_domains. This routine takes ownership of it and will
6696 * free_sched_domains it when done with it. If the caller failed the
6697 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6698 * and partition_sched_domains() will fallback to the single partition
6699 * 'fallback_doms', it also forces the domains to be rebuilt.
6701 * If doms_new == NULL it will be replaced with cpu_online_mask.
6702 * ndoms_new == 0 is a special case for destroying existing domains,
6703 * and it will not create the default domain.
6705 * Call with hotplug lock held
6707 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6708 struct sched_domain_attr *dattr_new)
6713 mutex_lock(&sched_domains_mutex);
6715 /* always unregister in case we don't destroy any domains */
6716 unregister_sched_domain_sysctl();
6718 /* Let architecture update cpu core mappings. */
6719 new_topology = arch_update_cpu_topology();
6721 n = doms_new ? ndoms_new : 0;
6723 /* Destroy deleted domains */
6724 for (i = 0; i < ndoms_cur; i++) {
6725 for (j = 0; j < n && !new_topology; j++) {
6726 if (cpumask_equal(doms_cur[i], doms_new[j])
6727 && dattrs_equal(dattr_cur, i, dattr_new, j))
6730 /* no match - a current sched domain not in new doms_new[] */
6731 detach_destroy_domains(doms_cur[i]);
6737 if (doms_new == NULL) {
6739 doms_new = &fallback_doms;
6740 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6741 WARN_ON_ONCE(dattr_new);
6744 /* Build new domains */
6745 for (i = 0; i < ndoms_new; i++) {
6746 for (j = 0; j < n && !new_topology; j++) {
6747 if (cpumask_equal(doms_new[i], doms_cur[j])
6748 && dattrs_equal(dattr_new, i, dattr_cur, j))
6751 /* no match - add a new doms_new */
6752 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6757 /* Remember the new sched domains */
6758 if (doms_cur != &fallback_doms)
6759 free_sched_domains(doms_cur, ndoms_cur);
6760 kfree(dattr_cur); /* kfree(NULL) is safe */
6761 doms_cur = doms_new;
6762 dattr_cur = dattr_new;
6763 ndoms_cur = ndoms_new;
6765 register_sched_domain_sysctl();
6767 mutex_unlock(&sched_domains_mutex);
6770 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6773 * Update cpusets according to cpu_active mask. If cpusets are
6774 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6775 * around partition_sched_domains().
6777 * If we come here as part of a suspend/resume, don't touch cpusets because we
6778 * want to restore it back to its original state upon resume anyway.
6780 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6784 case CPU_ONLINE_FROZEN:
6785 case CPU_DOWN_FAILED_FROZEN:
6788 * num_cpus_frozen tracks how many CPUs are involved in suspend
6789 * resume sequence. As long as this is not the last online
6790 * operation in the resume sequence, just build a single sched
6791 * domain, ignoring cpusets.
6794 if (likely(num_cpus_frozen)) {
6795 partition_sched_domains(1, NULL, NULL);
6800 * This is the last CPU online operation. So fall through and
6801 * restore the original sched domains by considering the
6802 * cpuset configurations.
6806 case CPU_DOWN_FAILED:
6807 cpuset_update_active_cpus(true);
6815 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6819 case CPU_DOWN_PREPARE:
6820 cpuset_update_active_cpus(false);
6822 case CPU_DOWN_PREPARE_FROZEN:
6824 partition_sched_domains(1, NULL, NULL);
6832 void __init sched_init_smp(void)
6834 cpumask_var_t non_isolated_cpus;
6836 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6837 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6842 * There's no userspace yet to cause hotplug operations; hence all the
6843 * cpu masks are stable and all blatant races in the below code cannot
6846 mutex_lock(&sched_domains_mutex);
6847 init_sched_domains(cpu_active_mask);
6848 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6849 if (cpumask_empty(non_isolated_cpus))
6850 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6851 mutex_unlock(&sched_domains_mutex);
6853 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6854 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6855 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6859 /* Move init over to a non-isolated CPU */
6860 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6862 sched_init_granularity();
6863 free_cpumask_var(non_isolated_cpus);
6865 init_sched_rt_class();
6866 init_sched_dl_class();
6869 void __init sched_init_smp(void)
6871 sched_init_granularity();
6873 #endif /* CONFIG_SMP */
6875 const_debug unsigned int sysctl_timer_migration = 1;
6877 int in_sched_functions(unsigned long addr)
6879 return in_lock_functions(addr) ||
6880 (addr >= (unsigned long)__sched_text_start
6881 && addr < (unsigned long)__sched_text_end);
6884 #ifdef CONFIG_CGROUP_SCHED
6886 * Default task group.
6887 * Every task in system belongs to this group at bootup.
6889 struct task_group root_task_group;
6890 LIST_HEAD(task_groups);
6893 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6895 void __init sched_init(void)
6898 unsigned long alloc_size = 0, ptr;
6900 #ifdef CONFIG_FAIR_GROUP_SCHED
6901 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6903 #ifdef CONFIG_RT_GROUP_SCHED
6904 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6906 #ifdef CONFIG_CPUMASK_OFFSTACK
6907 alloc_size += num_possible_cpus() * cpumask_size();
6910 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6912 #ifdef CONFIG_FAIR_GROUP_SCHED
6913 root_task_group.se = (struct sched_entity **)ptr;
6914 ptr += nr_cpu_ids * sizeof(void **);
6916 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6917 ptr += nr_cpu_ids * sizeof(void **);
6919 #endif /* CONFIG_FAIR_GROUP_SCHED */
6920 #ifdef CONFIG_RT_GROUP_SCHED
6921 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6922 ptr += nr_cpu_ids * sizeof(void **);
6924 root_task_group.rt_rq = (struct rt_rq **)ptr;
6925 ptr += nr_cpu_ids * sizeof(void **);
6927 #endif /* CONFIG_RT_GROUP_SCHED */
6928 #ifdef CONFIG_CPUMASK_OFFSTACK
6929 for_each_possible_cpu(i) {
6930 per_cpu(load_balance_mask, i) = (void *)ptr;
6931 ptr += cpumask_size();
6933 #endif /* CONFIG_CPUMASK_OFFSTACK */
6936 init_rt_bandwidth(&def_rt_bandwidth,
6937 global_rt_period(), global_rt_runtime());
6938 init_dl_bandwidth(&def_dl_bandwidth,
6939 global_rt_period(), global_rt_runtime());
6942 init_defrootdomain();
6945 #ifdef CONFIG_RT_GROUP_SCHED
6946 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6947 global_rt_period(), global_rt_runtime());
6948 #endif /* CONFIG_RT_GROUP_SCHED */
6950 #ifdef CONFIG_CGROUP_SCHED
6951 list_add(&root_task_group.list, &task_groups);
6952 INIT_LIST_HEAD(&root_task_group.children);
6953 INIT_LIST_HEAD(&root_task_group.siblings);
6954 autogroup_init(&init_task);
6956 #endif /* CONFIG_CGROUP_SCHED */
6958 for_each_possible_cpu(i) {
6962 raw_spin_lock_init(&rq->lock);
6964 rq->calc_load_active = 0;
6965 rq->calc_load_update = jiffies + LOAD_FREQ;
6966 init_cfs_rq(&rq->cfs);
6967 init_rt_rq(&rq->rt, rq);
6968 init_dl_rq(&rq->dl, rq);
6969 #ifdef CONFIG_FAIR_GROUP_SCHED
6970 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6971 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6973 * How much cpu bandwidth does root_task_group get?
6975 * In case of task-groups formed thr' the cgroup filesystem, it
6976 * gets 100% of the cpu resources in the system. This overall
6977 * system cpu resource is divided among the tasks of
6978 * root_task_group and its child task-groups in a fair manner,
6979 * based on each entity's (task or task-group's) weight
6980 * (se->load.weight).
6982 * In other words, if root_task_group has 10 tasks of weight
6983 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6984 * then A0's share of the cpu resource is:
6986 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6988 * We achieve this by letting root_task_group's tasks sit
6989 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6991 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6992 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6993 #endif /* CONFIG_FAIR_GROUP_SCHED */
6995 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6996 #ifdef CONFIG_RT_GROUP_SCHED
6997 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7000 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7001 rq->cpu_load[j] = 0;
7003 rq->last_load_update_tick = jiffies;
7008 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7009 rq->post_schedule = 0;
7010 rq->active_balance = 0;
7011 rq->next_balance = jiffies;
7016 rq->avg_idle = 2*sysctl_sched_migration_cost;
7017 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7019 INIT_LIST_HEAD(&rq->cfs_tasks);
7021 rq_attach_root(rq, &def_root_domain);
7022 #ifdef CONFIG_NO_HZ_COMMON
7025 #ifdef CONFIG_NO_HZ_FULL
7026 rq->last_sched_tick = 0;
7030 atomic_set(&rq->nr_iowait, 0);
7033 set_load_weight(&init_task);
7035 #ifdef CONFIG_PREEMPT_NOTIFIERS
7036 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7040 * The boot idle thread does lazy MMU switching as well:
7042 atomic_inc(&init_mm.mm_count);
7043 enter_lazy_tlb(&init_mm, current);
7046 * Make us the idle thread. Technically, schedule() should not be
7047 * called from this thread, however somewhere below it might be,
7048 * but because we are the idle thread, we just pick up running again
7049 * when this runqueue becomes "idle".
7051 init_idle(current, smp_processor_id());
7053 calc_load_update = jiffies + LOAD_FREQ;
7056 * During early bootup we pretend to be a normal task:
7058 current->sched_class = &fair_sched_class;
7061 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7062 /* May be allocated at isolcpus cmdline parse time */
7063 if (cpu_isolated_map == NULL)
7064 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7065 idle_thread_set_boot_cpu();
7066 set_cpu_rq_start_time();
7068 init_sched_fair_class();
7070 scheduler_running = 1;
7073 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7074 static inline int preempt_count_equals(int preempt_offset)
7076 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7078 return (nested == preempt_offset);
7081 void __might_sleep(const char *file, int line, int preempt_offset)
7083 static unsigned long prev_jiffy; /* ratelimiting */
7085 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7086 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7087 !is_idle_task(current)) ||
7088 system_state != SYSTEM_RUNNING || oops_in_progress)
7090 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7092 prev_jiffy = jiffies;
7095 "BUG: sleeping function called from invalid context at %s:%d\n",
7098 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7099 in_atomic(), irqs_disabled(),
7100 current->pid, current->comm);
7102 debug_show_held_locks(current);
7103 if (irqs_disabled())
7104 print_irqtrace_events(current);
7105 #ifdef CONFIG_DEBUG_PREEMPT
7106 if (!preempt_count_equals(preempt_offset)) {
7107 pr_err("Preemption disabled at:");
7108 print_ip_sym(current->preempt_disable_ip);
7114 EXPORT_SYMBOL(__might_sleep);
7117 #ifdef CONFIG_MAGIC_SYSRQ
7118 static void normalize_task(struct rq *rq, struct task_struct *p)
7120 const struct sched_class *prev_class = p->sched_class;
7121 struct sched_attr attr = {
7122 .sched_policy = SCHED_NORMAL,
7124 int old_prio = p->prio;
7127 queued = task_on_rq_queued(p);
7129 dequeue_task(rq, p, 0);
7130 __setscheduler(rq, p, &attr);
7132 enqueue_task(rq, p, 0);
7136 check_class_changed(rq, p, prev_class, old_prio);
7139 void normalize_rt_tasks(void)
7141 struct task_struct *g, *p;
7142 unsigned long flags;
7145 read_lock_irqsave(&tasklist_lock, flags);
7146 for_each_process_thread(g, p) {
7148 * Only normalize user tasks:
7153 p->se.exec_start = 0;
7154 #ifdef CONFIG_SCHEDSTATS
7155 p->se.statistics.wait_start = 0;
7156 p->se.statistics.sleep_start = 0;
7157 p->se.statistics.block_start = 0;
7160 if (!dl_task(p) && !rt_task(p)) {
7162 * Renice negative nice level userspace
7165 if (task_nice(p) < 0 && p->mm)
7166 set_user_nice(p, 0);
7170 raw_spin_lock(&p->pi_lock);
7171 rq = __task_rq_lock(p);
7173 normalize_task(rq, p);
7175 __task_rq_unlock(rq);
7176 raw_spin_unlock(&p->pi_lock);
7178 read_unlock_irqrestore(&tasklist_lock, flags);
7181 #endif /* CONFIG_MAGIC_SYSRQ */
7183 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7185 * These functions are only useful for the IA64 MCA handling, or kdb.
7187 * They can only be called when the whole system has been
7188 * stopped - every CPU needs to be quiescent, and no scheduling
7189 * activity can take place. Using them for anything else would
7190 * be a serious bug, and as a result, they aren't even visible
7191 * under any other configuration.
7195 * curr_task - return the current task for a given cpu.
7196 * @cpu: the processor in question.
7198 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7200 * Return: The current task for @cpu.
7202 struct task_struct *curr_task(int cpu)
7204 return cpu_curr(cpu);
7207 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7211 * set_curr_task - set the current task for a given cpu.
7212 * @cpu: the processor in question.
7213 * @p: the task pointer to set.
7215 * Description: This function must only be used when non-maskable interrupts
7216 * are serviced on a separate stack. It allows the architecture to switch the
7217 * notion of the current task on a cpu in a non-blocking manner. This function
7218 * must be called with all CPU's synchronized, and interrupts disabled, the
7219 * and caller must save the original value of the current task (see
7220 * curr_task() above) and restore that value before reenabling interrupts and
7221 * re-starting the system.
7223 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7225 void set_curr_task(int cpu, struct task_struct *p)
7232 #ifdef CONFIG_CGROUP_SCHED
7233 /* task_group_lock serializes the addition/removal of task groups */
7234 static DEFINE_SPINLOCK(task_group_lock);
7236 static void free_sched_group(struct task_group *tg)
7238 free_fair_sched_group(tg);
7239 free_rt_sched_group(tg);
7244 /* allocate runqueue etc for a new task group */
7245 struct task_group *sched_create_group(struct task_group *parent)
7247 struct task_group *tg;
7249 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7251 return ERR_PTR(-ENOMEM);
7253 if (!alloc_fair_sched_group(tg, parent))
7256 if (!alloc_rt_sched_group(tg, parent))
7262 free_sched_group(tg);
7263 return ERR_PTR(-ENOMEM);
7266 void sched_online_group(struct task_group *tg, struct task_group *parent)
7268 unsigned long flags;
7270 spin_lock_irqsave(&task_group_lock, flags);
7271 list_add_rcu(&tg->list, &task_groups);
7273 WARN_ON(!parent); /* root should already exist */
7275 tg->parent = parent;
7276 INIT_LIST_HEAD(&tg->children);
7277 list_add_rcu(&tg->siblings, &parent->children);
7278 spin_unlock_irqrestore(&task_group_lock, flags);
7281 /* rcu callback to free various structures associated with a task group */
7282 static void free_sched_group_rcu(struct rcu_head *rhp)
7284 /* now it should be safe to free those cfs_rqs */
7285 free_sched_group(container_of(rhp, struct task_group, rcu));
7288 /* Destroy runqueue etc associated with a task group */
7289 void sched_destroy_group(struct task_group *tg)
7291 /* wait for possible concurrent references to cfs_rqs complete */
7292 call_rcu(&tg->rcu, free_sched_group_rcu);
7295 void sched_offline_group(struct task_group *tg)
7297 unsigned long flags;
7300 /* end participation in shares distribution */
7301 for_each_possible_cpu(i)
7302 unregister_fair_sched_group(tg, i);
7304 spin_lock_irqsave(&task_group_lock, flags);
7305 list_del_rcu(&tg->list);
7306 list_del_rcu(&tg->siblings);
7307 spin_unlock_irqrestore(&task_group_lock, flags);
7310 /* change task's runqueue when it moves between groups.
7311 * The caller of this function should have put the task in its new group
7312 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7313 * reflect its new group.
7315 void sched_move_task(struct task_struct *tsk)
7317 struct task_group *tg;
7318 int queued, running;
7319 unsigned long flags;
7322 rq = task_rq_lock(tsk, &flags);
7324 running = task_current(rq, tsk);
7325 queued = task_on_rq_queued(tsk);
7328 dequeue_task(rq, tsk, 0);
7329 if (unlikely(running))
7330 tsk->sched_class->put_prev_task(rq, tsk);
7332 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7333 lockdep_is_held(&tsk->sighand->siglock)),
7334 struct task_group, css);
7335 tg = autogroup_task_group(tsk, tg);
7336 tsk->sched_task_group = tg;
7338 #ifdef CONFIG_FAIR_GROUP_SCHED
7339 if (tsk->sched_class->task_move_group)
7340 tsk->sched_class->task_move_group(tsk, queued);
7343 set_task_rq(tsk, task_cpu(tsk));
7345 if (unlikely(running))
7346 tsk->sched_class->set_curr_task(rq);
7348 enqueue_task(rq, tsk, 0);
7350 task_rq_unlock(rq, tsk, &flags);
7352 #endif /* CONFIG_CGROUP_SCHED */
7354 #ifdef CONFIG_RT_GROUP_SCHED
7356 * Ensure that the real time constraints are schedulable.
7358 static DEFINE_MUTEX(rt_constraints_mutex);
7360 /* Must be called with tasklist_lock held */
7361 static inline int tg_has_rt_tasks(struct task_group *tg)
7363 struct task_struct *g, *p;
7365 for_each_process_thread(g, p) {
7366 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7373 struct rt_schedulable_data {
7374 struct task_group *tg;
7379 static int tg_rt_schedulable(struct task_group *tg, void *data)
7381 struct rt_schedulable_data *d = data;
7382 struct task_group *child;
7383 unsigned long total, sum = 0;
7384 u64 period, runtime;
7386 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7387 runtime = tg->rt_bandwidth.rt_runtime;
7390 period = d->rt_period;
7391 runtime = d->rt_runtime;
7395 * Cannot have more runtime than the period.
7397 if (runtime > period && runtime != RUNTIME_INF)
7401 * Ensure we don't starve existing RT tasks.
7403 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7406 total = to_ratio(period, runtime);
7409 * Nobody can have more than the global setting allows.
7411 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7415 * The sum of our children's runtime should not exceed our own.
7417 list_for_each_entry_rcu(child, &tg->children, siblings) {
7418 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7419 runtime = child->rt_bandwidth.rt_runtime;
7421 if (child == d->tg) {
7422 period = d->rt_period;
7423 runtime = d->rt_runtime;
7426 sum += to_ratio(period, runtime);
7435 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7439 struct rt_schedulable_data data = {
7441 .rt_period = period,
7442 .rt_runtime = runtime,
7446 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7452 static int tg_set_rt_bandwidth(struct task_group *tg,
7453 u64 rt_period, u64 rt_runtime)
7457 mutex_lock(&rt_constraints_mutex);
7458 read_lock(&tasklist_lock);
7459 err = __rt_schedulable(tg, rt_period, rt_runtime);
7463 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7464 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7465 tg->rt_bandwidth.rt_runtime = rt_runtime;
7467 for_each_possible_cpu(i) {
7468 struct rt_rq *rt_rq = tg->rt_rq[i];
7470 raw_spin_lock(&rt_rq->rt_runtime_lock);
7471 rt_rq->rt_runtime = rt_runtime;
7472 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7474 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7476 read_unlock(&tasklist_lock);
7477 mutex_unlock(&rt_constraints_mutex);
7482 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7484 u64 rt_runtime, rt_period;
7486 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7487 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7488 if (rt_runtime_us < 0)
7489 rt_runtime = RUNTIME_INF;
7491 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7494 static long sched_group_rt_runtime(struct task_group *tg)
7498 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7501 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7502 do_div(rt_runtime_us, NSEC_PER_USEC);
7503 return rt_runtime_us;
7506 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7508 u64 rt_runtime, rt_period;
7510 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7511 rt_runtime = tg->rt_bandwidth.rt_runtime;
7516 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7519 static long sched_group_rt_period(struct task_group *tg)
7523 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7524 do_div(rt_period_us, NSEC_PER_USEC);
7525 return rt_period_us;
7527 #endif /* CONFIG_RT_GROUP_SCHED */
7529 #ifdef CONFIG_RT_GROUP_SCHED
7530 static int sched_rt_global_constraints(void)
7534 mutex_lock(&rt_constraints_mutex);
7535 read_lock(&tasklist_lock);
7536 ret = __rt_schedulable(NULL, 0, 0);
7537 read_unlock(&tasklist_lock);
7538 mutex_unlock(&rt_constraints_mutex);
7543 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7545 /* Don't accept realtime tasks when there is no way for them to run */
7546 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7552 #else /* !CONFIG_RT_GROUP_SCHED */
7553 static int sched_rt_global_constraints(void)
7555 unsigned long flags;
7558 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7559 for_each_possible_cpu(i) {
7560 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7562 raw_spin_lock(&rt_rq->rt_runtime_lock);
7563 rt_rq->rt_runtime = global_rt_runtime();
7564 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7566 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7570 #endif /* CONFIG_RT_GROUP_SCHED */
7572 static int sched_dl_global_constraints(void)
7574 u64 runtime = global_rt_runtime();
7575 u64 period = global_rt_period();
7576 u64 new_bw = to_ratio(period, runtime);
7578 unsigned long flags;
7581 * Here we want to check the bandwidth not being set to some
7582 * value smaller than the currently allocated bandwidth in
7583 * any of the root_domains.
7585 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7586 * cycling on root_domains... Discussion on different/better
7587 * solutions is welcome!
7589 for_each_possible_cpu(cpu) {
7590 struct dl_bw *dl_b = dl_bw_of(cpu);
7592 raw_spin_lock_irqsave(&dl_b->lock, flags);
7593 if (new_bw < dl_b->total_bw)
7595 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7604 static void sched_dl_do_global(void)
7608 unsigned long flags;
7610 def_dl_bandwidth.dl_period = global_rt_period();
7611 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7613 if (global_rt_runtime() != RUNTIME_INF)
7614 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7617 * FIXME: As above...
7619 for_each_possible_cpu(cpu) {
7620 struct dl_bw *dl_b = dl_bw_of(cpu);
7622 raw_spin_lock_irqsave(&dl_b->lock, flags);
7624 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7628 static int sched_rt_global_validate(void)
7630 if (sysctl_sched_rt_period <= 0)
7633 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7634 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7640 static void sched_rt_do_global(void)
7642 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7643 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7646 int sched_rt_handler(struct ctl_table *table, int write,
7647 void __user *buffer, size_t *lenp,
7650 int old_period, old_runtime;
7651 static DEFINE_MUTEX(mutex);
7655 old_period = sysctl_sched_rt_period;
7656 old_runtime = sysctl_sched_rt_runtime;
7658 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7660 if (!ret && write) {
7661 ret = sched_rt_global_validate();
7665 ret = sched_rt_global_constraints();
7669 ret = sched_dl_global_constraints();
7673 sched_rt_do_global();
7674 sched_dl_do_global();
7678 sysctl_sched_rt_period = old_period;
7679 sysctl_sched_rt_runtime = old_runtime;
7681 mutex_unlock(&mutex);
7686 int sched_rr_handler(struct ctl_table *table, int write,
7687 void __user *buffer, size_t *lenp,
7691 static DEFINE_MUTEX(mutex);
7694 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7695 /* make sure that internally we keep jiffies */
7696 /* also, writing zero resets timeslice to default */
7697 if (!ret && write) {
7698 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7699 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7701 mutex_unlock(&mutex);
7705 #ifdef CONFIG_CGROUP_SCHED
7707 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7709 return css ? container_of(css, struct task_group, css) : NULL;
7712 static struct cgroup_subsys_state *
7713 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7715 struct task_group *parent = css_tg(parent_css);
7716 struct task_group *tg;
7719 /* This is early initialization for the top cgroup */
7720 return &root_task_group.css;
7723 tg = sched_create_group(parent);
7725 return ERR_PTR(-ENOMEM);
7730 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7732 struct task_group *tg = css_tg(css);
7733 struct task_group *parent = css_tg(css->parent);
7736 sched_online_group(tg, parent);
7740 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7742 struct task_group *tg = css_tg(css);
7744 sched_destroy_group(tg);
7747 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7749 struct task_group *tg = css_tg(css);
7751 sched_offline_group(tg);
7754 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7755 struct cgroup_taskset *tset)
7757 struct task_struct *task;
7759 cgroup_taskset_for_each(task, tset) {
7760 #ifdef CONFIG_RT_GROUP_SCHED
7761 if (!sched_rt_can_attach(css_tg(css), task))
7764 /* We don't support RT-tasks being in separate groups */
7765 if (task->sched_class != &fair_sched_class)
7772 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7773 struct cgroup_taskset *tset)
7775 struct task_struct *task;
7777 cgroup_taskset_for_each(task, tset)
7778 sched_move_task(task);
7781 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7782 struct cgroup_subsys_state *old_css,
7783 struct task_struct *task)
7786 * cgroup_exit() is called in the copy_process() failure path.
7787 * Ignore this case since the task hasn't ran yet, this avoids
7788 * trying to poke a half freed task state from generic code.
7790 if (!(task->flags & PF_EXITING))
7793 sched_move_task(task);
7796 #ifdef CONFIG_FAIR_GROUP_SCHED
7797 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7798 struct cftype *cftype, u64 shareval)
7800 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7803 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7806 struct task_group *tg = css_tg(css);
7808 return (u64) scale_load_down(tg->shares);
7811 #ifdef CONFIG_CFS_BANDWIDTH
7812 static DEFINE_MUTEX(cfs_constraints_mutex);
7814 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7815 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7817 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7819 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7821 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7822 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7824 if (tg == &root_task_group)
7828 * Ensure we have at some amount of bandwidth every period. This is
7829 * to prevent reaching a state of large arrears when throttled via
7830 * entity_tick() resulting in prolonged exit starvation.
7832 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7836 * Likewise, bound things on the otherside by preventing insane quota
7837 * periods. This also allows us to normalize in computing quota
7840 if (period > max_cfs_quota_period)
7844 * Prevent race between setting of cfs_rq->runtime_enabled and
7845 * unthrottle_offline_cfs_rqs().
7848 mutex_lock(&cfs_constraints_mutex);
7849 ret = __cfs_schedulable(tg, period, quota);
7853 runtime_enabled = quota != RUNTIME_INF;
7854 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7856 * If we need to toggle cfs_bandwidth_used, off->on must occur
7857 * before making related changes, and on->off must occur afterwards
7859 if (runtime_enabled && !runtime_was_enabled)
7860 cfs_bandwidth_usage_inc();
7861 raw_spin_lock_irq(&cfs_b->lock);
7862 cfs_b->period = ns_to_ktime(period);
7863 cfs_b->quota = quota;
7865 __refill_cfs_bandwidth_runtime(cfs_b);
7866 /* restart the period timer (if active) to handle new period expiry */
7867 if (runtime_enabled && cfs_b->timer_active) {
7868 /* force a reprogram */
7869 __start_cfs_bandwidth(cfs_b, true);
7871 raw_spin_unlock_irq(&cfs_b->lock);
7873 for_each_online_cpu(i) {
7874 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7875 struct rq *rq = cfs_rq->rq;
7877 raw_spin_lock_irq(&rq->lock);
7878 cfs_rq->runtime_enabled = runtime_enabled;
7879 cfs_rq->runtime_remaining = 0;
7881 if (cfs_rq->throttled)
7882 unthrottle_cfs_rq(cfs_rq);
7883 raw_spin_unlock_irq(&rq->lock);
7885 if (runtime_was_enabled && !runtime_enabled)
7886 cfs_bandwidth_usage_dec();
7888 mutex_unlock(&cfs_constraints_mutex);
7894 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7898 period = ktime_to_ns(tg->cfs_bandwidth.period);
7899 if (cfs_quota_us < 0)
7900 quota = RUNTIME_INF;
7902 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7904 return tg_set_cfs_bandwidth(tg, period, quota);
7907 long tg_get_cfs_quota(struct task_group *tg)
7911 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7914 quota_us = tg->cfs_bandwidth.quota;
7915 do_div(quota_us, NSEC_PER_USEC);
7920 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7924 period = (u64)cfs_period_us * NSEC_PER_USEC;
7925 quota = tg->cfs_bandwidth.quota;
7927 return tg_set_cfs_bandwidth(tg, period, quota);
7930 long tg_get_cfs_period(struct task_group *tg)
7934 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7935 do_div(cfs_period_us, NSEC_PER_USEC);
7937 return cfs_period_us;
7940 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7943 return tg_get_cfs_quota(css_tg(css));
7946 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7947 struct cftype *cftype, s64 cfs_quota_us)
7949 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7952 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7955 return tg_get_cfs_period(css_tg(css));
7958 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7959 struct cftype *cftype, u64 cfs_period_us)
7961 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7964 struct cfs_schedulable_data {
7965 struct task_group *tg;
7970 * normalize group quota/period to be quota/max_period
7971 * note: units are usecs
7973 static u64 normalize_cfs_quota(struct task_group *tg,
7974 struct cfs_schedulable_data *d)
7982 period = tg_get_cfs_period(tg);
7983 quota = tg_get_cfs_quota(tg);
7986 /* note: these should typically be equivalent */
7987 if (quota == RUNTIME_INF || quota == -1)
7990 return to_ratio(period, quota);
7993 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7995 struct cfs_schedulable_data *d = data;
7996 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7997 s64 quota = 0, parent_quota = -1;
8000 quota = RUNTIME_INF;
8002 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8004 quota = normalize_cfs_quota(tg, d);
8005 parent_quota = parent_b->hierarchal_quota;
8008 * ensure max(child_quota) <= parent_quota, inherit when no
8011 if (quota == RUNTIME_INF)
8012 quota = parent_quota;
8013 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8016 cfs_b->hierarchal_quota = quota;
8021 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8024 struct cfs_schedulable_data data = {
8030 if (quota != RUNTIME_INF) {
8031 do_div(data.period, NSEC_PER_USEC);
8032 do_div(data.quota, NSEC_PER_USEC);
8036 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8042 static int cpu_stats_show(struct seq_file *sf, void *v)
8044 struct task_group *tg = css_tg(seq_css(sf));
8045 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8047 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8048 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8049 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8053 #endif /* CONFIG_CFS_BANDWIDTH */
8054 #endif /* CONFIG_FAIR_GROUP_SCHED */
8056 #ifdef CONFIG_RT_GROUP_SCHED
8057 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8058 struct cftype *cft, s64 val)
8060 return sched_group_set_rt_runtime(css_tg(css), val);
8063 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8066 return sched_group_rt_runtime(css_tg(css));
8069 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8070 struct cftype *cftype, u64 rt_period_us)
8072 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8075 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8078 return sched_group_rt_period(css_tg(css));
8080 #endif /* CONFIG_RT_GROUP_SCHED */
8082 static struct cftype cpu_files[] = {
8083 #ifdef CONFIG_FAIR_GROUP_SCHED
8086 .read_u64 = cpu_shares_read_u64,
8087 .write_u64 = cpu_shares_write_u64,
8090 #ifdef CONFIG_CFS_BANDWIDTH
8092 .name = "cfs_quota_us",
8093 .read_s64 = cpu_cfs_quota_read_s64,
8094 .write_s64 = cpu_cfs_quota_write_s64,
8097 .name = "cfs_period_us",
8098 .read_u64 = cpu_cfs_period_read_u64,
8099 .write_u64 = cpu_cfs_period_write_u64,
8103 .seq_show = cpu_stats_show,
8106 #ifdef CONFIG_RT_GROUP_SCHED
8108 .name = "rt_runtime_us",
8109 .read_s64 = cpu_rt_runtime_read,
8110 .write_s64 = cpu_rt_runtime_write,
8113 .name = "rt_period_us",
8114 .read_u64 = cpu_rt_period_read_uint,
8115 .write_u64 = cpu_rt_period_write_uint,
8121 struct cgroup_subsys cpu_cgrp_subsys = {
8122 .css_alloc = cpu_cgroup_css_alloc,
8123 .css_free = cpu_cgroup_css_free,
8124 .css_online = cpu_cgroup_css_online,
8125 .css_offline = cpu_cgroup_css_offline,
8126 .can_attach = cpu_cgroup_can_attach,
8127 .attach = cpu_cgroup_attach,
8128 .exit = cpu_cgroup_exit,
8129 .legacy_cftypes = cpu_files,
8133 #endif /* CONFIG_CGROUP_SCHED */
8135 void dump_cpu_task(int cpu)
8137 pr_info("Task dump for CPU %d:\n", cpu);
8138 sched_show_task(cpu_curr(cpu));