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>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_sched.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
197 sched_feat_write(struct file *filp, const char __user *ubuf,
198 size_t cnt, loff_t *ppos)
208 if (copy_from_user(&buf, ubuf, cnt))
214 if (strncmp(cmp, "NO_", 3) == 0) {
219 for (i = 0; i < __SCHED_FEAT_NR; i++) {
220 if (strcmp(cmp, sched_feat_names[i]) == 0) {
222 sysctl_sched_features &= ~(1UL << i);
223 sched_feat_disable(i);
225 sysctl_sched_features |= (1UL << i);
226 sched_feat_enable(i);
232 if (i == __SCHED_FEAT_NR)
240 static int sched_feat_open(struct inode *inode, struct file *filp)
242 return single_open(filp, sched_feat_show, NULL);
245 static const struct file_operations sched_feat_fops = {
246 .open = sched_feat_open,
247 .write = sched_feat_write,
250 .release = single_release,
253 static __init int sched_init_debug(void)
255 debugfs_create_file("sched_features", 0644, NULL, NULL,
260 late_initcall(sched_init_debug);
261 #endif /* CONFIG_SCHED_DEBUG */
264 * Number of tasks to iterate in a single balance run.
265 * Limited because this is done with IRQs disabled.
267 const_debug unsigned int sysctl_sched_nr_migrate = 32;
270 * period over which we average the RT time consumption, measured
275 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
278 * period over which we measure -rt task cpu usage in us.
281 unsigned int sysctl_sched_rt_period = 1000000;
283 __read_mostly int scheduler_running;
286 * part of the period that we allow rt tasks to run in us.
289 int sysctl_sched_rt_runtime = 950000;
294 * __task_rq_lock - lock the rq @p resides on.
296 static inline struct rq *__task_rq_lock(struct task_struct *p)
301 lockdep_assert_held(&p->pi_lock);
305 raw_spin_lock(&rq->lock);
306 if (likely(rq == task_rq(p)))
308 raw_spin_unlock(&rq->lock);
313 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
315 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
316 __acquires(p->pi_lock)
322 raw_spin_lock_irqsave(&p->pi_lock, *flags);
324 raw_spin_lock(&rq->lock);
325 if (likely(rq == task_rq(p)))
327 raw_spin_unlock(&rq->lock);
328 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
332 static void __task_rq_unlock(struct rq *rq)
335 raw_spin_unlock(&rq->lock);
339 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
341 __releases(p->pi_lock)
343 raw_spin_unlock(&rq->lock);
344 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
348 * this_rq_lock - lock this runqueue and disable interrupts.
350 static struct rq *this_rq_lock(void)
357 raw_spin_lock(&rq->lock);
362 #ifdef CONFIG_SCHED_HRTICK
364 * Use HR-timers to deliver accurate preemption points.
366 * Its all a bit involved since we cannot program an hrt while holding the
367 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
370 * When we get rescheduled we reprogram the hrtick_timer outside of the
374 static void hrtick_clear(struct rq *rq)
376 if (hrtimer_active(&rq->hrtick_timer))
377 hrtimer_cancel(&rq->hrtick_timer);
381 * High-resolution timer tick.
382 * Runs from hardirq context with interrupts disabled.
384 static enum hrtimer_restart hrtick(struct hrtimer *timer)
386 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
388 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
390 raw_spin_lock(&rq->lock);
392 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
393 raw_spin_unlock(&rq->lock);
395 return HRTIMER_NORESTART;
400 * called from hardirq (IPI) context
402 static void __hrtick_start(void *arg)
406 raw_spin_lock(&rq->lock);
407 hrtimer_restart(&rq->hrtick_timer);
408 rq->hrtick_csd_pending = 0;
409 raw_spin_unlock(&rq->lock);
413 * Called to set the hrtick timer state.
415 * called with rq->lock held and irqs disabled
417 void hrtick_start(struct rq *rq, u64 delay)
419 struct hrtimer *timer = &rq->hrtick_timer;
420 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
422 hrtimer_set_expires(timer, time);
424 if (rq == this_rq()) {
425 hrtimer_restart(timer);
426 } else if (!rq->hrtick_csd_pending) {
427 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
428 rq->hrtick_csd_pending = 1;
433 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
435 int cpu = (int)(long)hcpu;
438 case CPU_UP_CANCELED:
439 case CPU_UP_CANCELED_FROZEN:
440 case CPU_DOWN_PREPARE:
441 case CPU_DOWN_PREPARE_FROZEN:
443 case CPU_DEAD_FROZEN:
444 hrtick_clear(cpu_rq(cpu));
451 static __init void init_hrtick(void)
453 hotcpu_notifier(hotplug_hrtick, 0);
457 * Called to set the hrtick timer state.
459 * called with rq->lock held and irqs disabled
461 void hrtick_start(struct rq *rq, u64 delay)
463 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
464 HRTIMER_MODE_REL_PINNED, 0);
467 static inline void init_hrtick(void)
470 #endif /* CONFIG_SMP */
472 static void init_rq_hrtick(struct rq *rq)
475 rq->hrtick_csd_pending = 0;
477 rq->hrtick_csd.flags = 0;
478 rq->hrtick_csd.func = __hrtick_start;
479 rq->hrtick_csd.info = rq;
482 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
483 rq->hrtick_timer.function = hrtick;
485 #else /* CONFIG_SCHED_HRTICK */
486 static inline void hrtick_clear(struct rq *rq)
490 static inline void init_rq_hrtick(struct rq *rq)
494 static inline void init_hrtick(void)
497 #endif /* CONFIG_SCHED_HRTICK */
500 * resched_task - mark a task 'to be rescheduled now'.
502 * On UP this means the setting of the need_resched flag, on SMP it
503 * might also involve a cross-CPU call to trigger the scheduler on
508 #ifndef tsk_is_polling
509 #define tsk_is_polling(t) 0
512 void resched_task(struct task_struct *p)
516 assert_raw_spin_locked(&task_rq(p)->lock);
518 if (test_tsk_need_resched(p))
521 set_tsk_need_resched(p);
524 if (cpu == smp_processor_id())
527 /* NEED_RESCHED must be visible before we test polling */
529 if (!tsk_is_polling(p))
530 smp_send_reschedule(cpu);
533 void resched_cpu(int cpu)
535 struct rq *rq = cpu_rq(cpu);
538 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
540 resched_task(cpu_curr(cpu));
541 raw_spin_unlock_irqrestore(&rq->lock, flags);
546 * In the semi idle case, use the nearest busy cpu for migrating timers
547 * from an idle cpu. This is good for power-savings.
549 * We don't do similar optimization for completely idle system, as
550 * selecting an idle cpu will add more delays to the timers than intended
551 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
553 int get_nohz_timer_target(void)
555 int cpu = smp_processor_id();
557 struct sched_domain *sd;
560 for_each_domain(cpu, sd) {
561 for_each_cpu(i, sched_domain_span(sd)) {
573 * When add_timer_on() enqueues a timer into the timer wheel of an
574 * idle CPU then this timer might expire before the next timer event
575 * which is scheduled to wake up that CPU. In case of a completely
576 * idle system the next event might even be infinite time into the
577 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
578 * leaves the inner idle loop so the newly added timer is taken into
579 * account when the CPU goes back to idle and evaluates the timer
580 * wheel for the next timer event.
582 void wake_up_idle_cpu(int cpu)
584 struct rq *rq = cpu_rq(cpu);
586 if (cpu == smp_processor_id())
590 * This is safe, as this function is called with the timer
591 * wheel base lock of (cpu) held. When the CPU is on the way
592 * to idle and has not yet set rq->curr to idle then it will
593 * be serialized on the timer wheel base lock and take the new
594 * timer into account automatically.
596 if (rq->curr != rq->idle)
600 * We can set TIF_RESCHED on the idle task of the other CPU
601 * lockless. The worst case is that the other CPU runs the
602 * idle task through an additional NOOP schedule()
604 set_tsk_need_resched(rq->idle);
606 /* NEED_RESCHED must be visible before we test polling */
608 if (!tsk_is_polling(rq->idle))
609 smp_send_reschedule(cpu);
612 static inline bool got_nohz_idle_kick(void)
614 int cpu = smp_processor_id();
615 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
618 #else /* CONFIG_NO_HZ */
620 static inline bool got_nohz_idle_kick(void)
625 #endif /* CONFIG_NO_HZ */
627 void sched_avg_update(struct rq *rq)
629 s64 period = sched_avg_period();
631 while ((s64)(rq->clock - rq->age_stamp) > period) {
633 * Inline assembly required to prevent the compiler
634 * optimising this loop into a divmod call.
635 * See __iter_div_u64_rem() for another example of this.
637 asm("" : "+rm" (rq->age_stamp));
638 rq->age_stamp += period;
643 #else /* !CONFIG_SMP */
644 void resched_task(struct task_struct *p)
646 assert_raw_spin_locked(&task_rq(p)->lock);
647 set_tsk_need_resched(p);
649 #endif /* CONFIG_SMP */
651 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
652 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
654 * Iterate task_group tree rooted at *from, calling @down when first entering a
655 * node and @up when leaving it for the final time.
657 * Caller must hold rcu_lock or sufficient equivalent.
659 int walk_tg_tree_from(struct task_group *from,
660 tg_visitor down, tg_visitor up, void *data)
662 struct task_group *parent, *child;
668 ret = (*down)(parent, data);
671 list_for_each_entry_rcu(child, &parent->children, siblings) {
678 ret = (*up)(parent, data);
679 if (ret || parent == from)
683 parent = parent->parent;
690 int tg_nop(struct task_group *tg, void *data)
696 static void set_load_weight(struct task_struct *p)
698 int prio = p->static_prio - MAX_RT_PRIO;
699 struct load_weight *load = &p->se.load;
702 * SCHED_IDLE tasks get minimal weight:
704 if (p->policy == SCHED_IDLE) {
705 load->weight = scale_load(WEIGHT_IDLEPRIO);
706 load->inv_weight = WMULT_IDLEPRIO;
710 load->weight = scale_load(prio_to_weight[prio]);
711 load->inv_weight = prio_to_wmult[prio];
714 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
717 sched_info_queued(p);
718 p->sched_class->enqueue_task(rq, p, flags);
721 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
724 sched_info_dequeued(p);
725 p->sched_class->dequeue_task(rq, p, flags);
728 void activate_task(struct rq *rq, struct task_struct *p, int flags)
730 if (task_contributes_to_load(p))
731 rq->nr_uninterruptible--;
733 enqueue_task(rq, p, flags);
736 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
738 if (task_contributes_to_load(p))
739 rq->nr_uninterruptible++;
741 dequeue_task(rq, p, flags);
744 static void update_rq_clock_task(struct rq *rq, s64 delta)
747 * In theory, the compile should just see 0 here, and optimize out the call
748 * to sched_rt_avg_update. But I don't trust it...
750 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
751 s64 steal = 0, irq_delta = 0;
753 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
754 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
757 * Since irq_time is only updated on {soft,}irq_exit, we might run into
758 * this case when a previous update_rq_clock() happened inside a
761 * When this happens, we stop ->clock_task and only update the
762 * prev_irq_time stamp to account for the part that fit, so that a next
763 * update will consume the rest. This ensures ->clock_task is
766 * It does however cause some slight miss-attribution of {soft,}irq
767 * time, a more accurate solution would be to update the irq_time using
768 * the current rq->clock timestamp, except that would require using
771 if (irq_delta > delta)
774 rq->prev_irq_time += irq_delta;
777 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
778 if (static_key_false((¶virt_steal_rq_enabled))) {
781 steal = paravirt_steal_clock(cpu_of(rq));
782 steal -= rq->prev_steal_time_rq;
784 if (unlikely(steal > delta))
787 st = steal_ticks(steal);
788 steal = st * TICK_NSEC;
790 rq->prev_steal_time_rq += steal;
796 rq->clock_task += delta;
798 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
799 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
800 sched_rt_avg_update(rq, irq_delta + steal);
804 void sched_set_stop_task(int cpu, struct task_struct *stop)
806 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
807 struct task_struct *old_stop = cpu_rq(cpu)->stop;
811 * Make it appear like a SCHED_FIFO task, its something
812 * userspace knows about and won't get confused about.
814 * Also, it will make PI more or less work without too
815 * much confusion -- but then, stop work should not
816 * rely on PI working anyway.
818 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
820 stop->sched_class = &stop_sched_class;
823 cpu_rq(cpu)->stop = stop;
827 * Reset it back to a normal scheduling class so that
828 * it can die in pieces.
830 old_stop->sched_class = &rt_sched_class;
835 * __normal_prio - return the priority that is based on the static prio
837 static inline int __normal_prio(struct task_struct *p)
839 return p->static_prio;
843 * Calculate the expected normal priority: i.e. priority
844 * without taking RT-inheritance into account. Might be
845 * boosted by interactivity modifiers. Changes upon fork,
846 * setprio syscalls, and whenever the interactivity
847 * estimator recalculates.
849 static inline int normal_prio(struct task_struct *p)
853 if (task_has_rt_policy(p))
854 prio = MAX_RT_PRIO-1 - p->rt_priority;
856 prio = __normal_prio(p);
861 * Calculate the current priority, i.e. the priority
862 * taken into account by the scheduler. This value might
863 * be boosted by RT tasks, or might be boosted by
864 * interactivity modifiers. Will be RT if the task got
865 * RT-boosted. If not then it returns p->normal_prio.
867 static int effective_prio(struct task_struct *p)
869 p->normal_prio = normal_prio(p);
871 * If we are RT tasks or we were boosted to RT priority,
872 * keep the priority unchanged. Otherwise, update priority
873 * to the normal priority:
875 if (!rt_prio(p->prio))
876 return p->normal_prio;
881 * task_curr - is this task currently executing on a CPU?
882 * @p: the task in question.
884 inline int task_curr(const struct task_struct *p)
886 return cpu_curr(task_cpu(p)) == p;
889 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
890 const struct sched_class *prev_class,
893 if (prev_class != p->sched_class) {
894 if (prev_class->switched_from)
895 prev_class->switched_from(rq, p);
896 p->sched_class->switched_to(rq, p);
897 } else if (oldprio != p->prio)
898 p->sched_class->prio_changed(rq, p, oldprio);
901 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
903 const struct sched_class *class;
905 if (p->sched_class == rq->curr->sched_class) {
906 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
908 for_each_class(class) {
909 if (class == rq->curr->sched_class)
911 if (class == p->sched_class) {
912 resched_task(rq->curr);
919 * A queue event has occurred, and we're going to schedule. In
920 * this case, we can save a useless back to back clock update.
922 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
923 rq->skip_clock_update = 1;
926 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
928 void register_task_migration_notifier(struct notifier_block *n)
930 atomic_notifier_chain_register(&task_migration_notifier, n);
934 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
936 #ifdef CONFIG_SCHED_DEBUG
938 * We should never call set_task_cpu() on a blocked task,
939 * ttwu() will sort out the placement.
941 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
942 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
944 #ifdef CONFIG_LOCKDEP
946 * The caller should hold either p->pi_lock or rq->lock, when changing
947 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
949 * sched_move_task() holds both and thus holding either pins the cgroup,
952 * Furthermore, all task_rq users should acquire both locks, see
955 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
956 lockdep_is_held(&task_rq(p)->lock)));
960 trace_sched_migrate_task(p, new_cpu);
962 if (task_cpu(p) != new_cpu) {
963 struct task_migration_notifier tmn;
965 if (p->sched_class->migrate_task_rq)
966 p->sched_class->migrate_task_rq(p, new_cpu);
967 p->se.nr_migrations++;
968 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
971 tmn.from_cpu = task_cpu(p);
972 tmn.to_cpu = new_cpu;
974 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
977 __set_task_cpu(p, new_cpu);
980 struct migration_arg {
981 struct task_struct *task;
985 static int migration_cpu_stop(void *data);
988 * wait_task_inactive - wait for a thread to unschedule.
990 * If @match_state is nonzero, it's the @p->state value just checked and
991 * not expected to change. If it changes, i.e. @p might have woken up,
992 * then return zero. When we succeed in waiting for @p to be off its CPU,
993 * we return a positive number (its total switch count). If a second call
994 * a short while later returns the same number, the caller can be sure that
995 * @p has remained unscheduled the whole time.
997 * The caller must ensure that the task *will* unschedule sometime soon,
998 * else this function might spin for a *long* time. This function can't
999 * be called with interrupts off, or it may introduce deadlock with
1000 * smp_call_function() if an IPI is sent by the same process we are
1001 * waiting to become inactive.
1003 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1005 unsigned long flags;
1012 * We do the initial early heuristics without holding
1013 * any task-queue locks at all. We'll only try to get
1014 * the runqueue lock when things look like they will
1020 * If the task is actively running on another CPU
1021 * still, just relax and busy-wait without holding
1024 * NOTE! Since we don't hold any locks, it's not
1025 * even sure that "rq" stays as the right runqueue!
1026 * But we don't care, since "task_running()" will
1027 * return false if the runqueue has changed and p
1028 * is actually now running somewhere else!
1030 while (task_running(rq, p)) {
1031 if (match_state && unlikely(p->state != match_state))
1037 * Ok, time to look more closely! We need the rq
1038 * lock now, to be *sure*. If we're wrong, we'll
1039 * just go back and repeat.
1041 rq = task_rq_lock(p, &flags);
1042 trace_sched_wait_task(p);
1043 running = task_running(rq, p);
1046 if (!match_state || p->state == match_state)
1047 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1048 task_rq_unlock(rq, p, &flags);
1051 * If it changed from the expected state, bail out now.
1053 if (unlikely(!ncsw))
1057 * Was it really running after all now that we
1058 * checked with the proper locks actually held?
1060 * Oops. Go back and try again..
1062 if (unlikely(running)) {
1068 * It's not enough that it's not actively running,
1069 * it must be off the runqueue _entirely_, and not
1072 * So if it was still runnable (but just not actively
1073 * running right now), it's preempted, and we should
1074 * yield - it could be a while.
1076 if (unlikely(on_rq)) {
1077 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1079 set_current_state(TASK_UNINTERRUPTIBLE);
1080 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1085 * Ahh, all good. It wasn't running, and it wasn't
1086 * runnable, which means that it will never become
1087 * running in the future either. We're all done!
1096 * kick_process - kick a running thread to enter/exit the kernel
1097 * @p: the to-be-kicked thread
1099 * Cause a process which is running on another CPU to enter
1100 * kernel-mode, without any delay. (to get signals handled.)
1102 * NOTE: this function doesn't have to take the runqueue lock,
1103 * because all it wants to ensure is that the remote task enters
1104 * the kernel. If the IPI races and the task has been migrated
1105 * to another CPU then no harm is done and the purpose has been
1108 void kick_process(struct task_struct *p)
1114 if ((cpu != smp_processor_id()) && task_curr(p))
1115 smp_send_reschedule(cpu);
1118 EXPORT_SYMBOL_GPL(kick_process);
1119 #endif /* CONFIG_SMP */
1123 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1125 static int select_fallback_rq(int cpu, struct task_struct *p)
1127 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1128 enum { cpuset, possible, fail } state = cpuset;
1131 /* Look for allowed, online CPU in same node. */
1132 for_each_cpu(dest_cpu, nodemask) {
1133 if (!cpu_online(dest_cpu))
1135 if (!cpu_active(dest_cpu))
1137 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1142 /* Any allowed, online CPU? */
1143 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1144 if (!cpu_online(dest_cpu))
1146 if (!cpu_active(dest_cpu))
1153 /* No more Mr. Nice Guy. */
1154 cpuset_cpus_allowed_fallback(p);
1159 do_set_cpus_allowed(p, cpu_possible_mask);
1170 if (state != cpuset) {
1172 * Don't tell them about moving exiting tasks or
1173 * kernel threads (both mm NULL), since they never
1176 if (p->mm && printk_ratelimit()) {
1177 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1178 task_pid_nr(p), p->comm, cpu);
1186 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1189 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1191 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1194 * In order not to call set_task_cpu() on a blocking task we need
1195 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1198 * Since this is common to all placement strategies, this lives here.
1200 * [ this allows ->select_task() to simply return task_cpu(p) and
1201 * not worry about this generic constraint ]
1203 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1205 cpu = select_fallback_rq(task_cpu(p), p);
1210 static void update_avg(u64 *avg, u64 sample)
1212 s64 diff = sample - *avg;
1218 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1220 #ifdef CONFIG_SCHEDSTATS
1221 struct rq *rq = this_rq();
1224 int this_cpu = smp_processor_id();
1226 if (cpu == this_cpu) {
1227 schedstat_inc(rq, ttwu_local);
1228 schedstat_inc(p, se.statistics.nr_wakeups_local);
1230 struct sched_domain *sd;
1232 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1234 for_each_domain(this_cpu, sd) {
1235 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1236 schedstat_inc(sd, ttwu_wake_remote);
1243 if (wake_flags & WF_MIGRATED)
1244 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1246 #endif /* CONFIG_SMP */
1248 schedstat_inc(rq, ttwu_count);
1249 schedstat_inc(p, se.statistics.nr_wakeups);
1251 if (wake_flags & WF_SYNC)
1252 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1254 #endif /* CONFIG_SCHEDSTATS */
1257 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1259 activate_task(rq, p, en_flags);
1262 /* if a worker is waking up, notify workqueue */
1263 if (p->flags & PF_WQ_WORKER)
1264 wq_worker_waking_up(p, cpu_of(rq));
1268 * Mark the task runnable and perform wakeup-preemption.
1271 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1273 trace_sched_wakeup(p, true);
1274 check_preempt_curr(rq, p, wake_flags);
1276 p->state = TASK_RUNNING;
1278 if (p->sched_class->task_woken)
1279 p->sched_class->task_woken(rq, p);
1281 if (rq->idle_stamp) {
1282 u64 delta = rq->clock - rq->idle_stamp;
1283 u64 max = 2*sysctl_sched_migration_cost;
1288 update_avg(&rq->avg_idle, delta);
1295 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1298 if (p->sched_contributes_to_load)
1299 rq->nr_uninterruptible--;
1302 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1303 ttwu_do_wakeup(rq, p, wake_flags);
1307 * Called in case the task @p isn't fully descheduled from its runqueue,
1308 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1309 * since all we need to do is flip p->state to TASK_RUNNING, since
1310 * the task is still ->on_rq.
1312 static int ttwu_remote(struct task_struct *p, int wake_flags)
1317 rq = __task_rq_lock(p);
1319 ttwu_do_wakeup(rq, p, wake_flags);
1322 __task_rq_unlock(rq);
1328 static void sched_ttwu_pending(void)
1330 struct rq *rq = this_rq();
1331 struct llist_node *llist = llist_del_all(&rq->wake_list);
1332 struct task_struct *p;
1334 raw_spin_lock(&rq->lock);
1337 p = llist_entry(llist, struct task_struct, wake_entry);
1338 llist = llist_next(llist);
1339 ttwu_do_activate(rq, p, 0);
1342 raw_spin_unlock(&rq->lock);
1345 void scheduler_ipi(void)
1347 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1351 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1352 * traditionally all their work was done from the interrupt return
1353 * path. Now that we actually do some work, we need to make sure
1356 * Some archs already do call them, luckily irq_enter/exit nest
1359 * Arguably we should visit all archs and update all handlers,
1360 * however a fair share of IPIs are still resched only so this would
1361 * somewhat pessimize the simple resched case.
1364 sched_ttwu_pending();
1367 * Check if someone kicked us for doing the nohz idle load balance.
1369 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1370 this_rq()->idle_balance = 1;
1371 raise_softirq_irqoff(SCHED_SOFTIRQ);
1376 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1378 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1379 smp_send_reschedule(cpu);
1382 bool cpus_share_cache(int this_cpu, int that_cpu)
1384 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1386 #endif /* CONFIG_SMP */
1388 static void ttwu_queue(struct task_struct *p, int cpu)
1390 struct rq *rq = cpu_rq(cpu);
1392 #if defined(CONFIG_SMP)
1393 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1394 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1395 ttwu_queue_remote(p, cpu);
1400 raw_spin_lock(&rq->lock);
1401 ttwu_do_activate(rq, p, 0);
1402 raw_spin_unlock(&rq->lock);
1406 * try_to_wake_up - wake up a thread
1407 * @p: the thread to be awakened
1408 * @state: the mask of task states that can be woken
1409 * @wake_flags: wake modifier flags (WF_*)
1411 * Put it on the run-queue if it's not already there. The "current"
1412 * thread is always on the run-queue (except when the actual
1413 * re-schedule is in progress), and as such you're allowed to do
1414 * the simpler "current->state = TASK_RUNNING" to mark yourself
1415 * runnable without the overhead of this.
1417 * Returns %true if @p was woken up, %false if it was already running
1418 * or @state didn't match @p's state.
1421 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1423 unsigned long flags;
1424 int cpu, success = 0;
1427 raw_spin_lock_irqsave(&p->pi_lock, flags);
1428 if (!(p->state & state))
1431 success = 1; /* we're going to change ->state */
1434 if (p->on_rq && ttwu_remote(p, wake_flags))
1439 * If the owning (remote) cpu is still in the middle of schedule() with
1440 * this task as prev, wait until its done referencing the task.
1445 * Pairs with the smp_wmb() in finish_lock_switch().
1449 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1450 p->state = TASK_WAKING;
1452 if (p->sched_class->task_waking)
1453 p->sched_class->task_waking(p);
1455 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1456 if (task_cpu(p) != cpu) {
1457 wake_flags |= WF_MIGRATED;
1458 set_task_cpu(p, cpu);
1460 #endif /* CONFIG_SMP */
1464 ttwu_stat(p, cpu, wake_flags);
1466 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1472 * try_to_wake_up_local - try to wake up a local task with rq lock held
1473 * @p: the thread to be awakened
1475 * Put @p on the run-queue if it's not already there. The caller must
1476 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1479 static void try_to_wake_up_local(struct task_struct *p)
1481 struct rq *rq = task_rq(p);
1483 BUG_ON(rq != this_rq());
1484 BUG_ON(p == current);
1485 lockdep_assert_held(&rq->lock);
1487 if (!raw_spin_trylock(&p->pi_lock)) {
1488 raw_spin_unlock(&rq->lock);
1489 raw_spin_lock(&p->pi_lock);
1490 raw_spin_lock(&rq->lock);
1493 if (!(p->state & TASK_NORMAL))
1497 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1499 ttwu_do_wakeup(rq, p, 0);
1500 ttwu_stat(p, smp_processor_id(), 0);
1502 raw_spin_unlock(&p->pi_lock);
1506 * wake_up_process - Wake up a specific process
1507 * @p: The process to be woken up.
1509 * Attempt to wake up the nominated process and move it to the set of runnable
1510 * processes. Returns 1 if the process was woken up, 0 if it was already
1513 * It may be assumed that this function implies a write memory barrier before
1514 * changing the task state if and only if any tasks are woken up.
1516 int wake_up_process(struct task_struct *p)
1518 return try_to_wake_up(p, TASK_ALL, 0);
1520 EXPORT_SYMBOL(wake_up_process);
1522 int wake_up_state(struct task_struct *p, unsigned int state)
1524 return try_to_wake_up(p, state, 0);
1528 * Perform scheduler related setup for a newly forked process p.
1529 * p is forked by current.
1531 * __sched_fork() is basic setup used by init_idle() too:
1533 static void __sched_fork(struct task_struct *p)
1538 p->se.exec_start = 0;
1539 p->se.sum_exec_runtime = 0;
1540 p->se.prev_sum_exec_runtime = 0;
1541 p->se.nr_migrations = 0;
1543 INIT_LIST_HEAD(&p->se.group_node);
1546 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1547 * removed when useful for applications beyond shares distribution (e.g.
1550 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1551 p->se.avg.runnable_avg_period = 0;
1552 p->se.avg.runnable_avg_sum = 0;
1554 #ifdef CONFIG_SCHEDSTATS
1555 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1558 INIT_LIST_HEAD(&p->rt.run_list);
1560 #ifdef CONFIG_PREEMPT_NOTIFIERS
1561 INIT_HLIST_HEAD(&p->preempt_notifiers);
1566 * fork()/clone()-time setup:
1568 void sched_fork(struct task_struct *p)
1570 unsigned long flags;
1571 int cpu = get_cpu();
1575 * We mark the process as running here. This guarantees that
1576 * nobody will actually run it, and a signal or other external
1577 * event cannot wake it up and insert it on the runqueue either.
1579 p->state = TASK_RUNNING;
1582 * Make sure we do not leak PI boosting priority to the child.
1584 p->prio = current->normal_prio;
1587 * Revert to default priority/policy on fork if requested.
1589 if (unlikely(p->sched_reset_on_fork)) {
1590 if (task_has_rt_policy(p)) {
1591 p->policy = SCHED_NORMAL;
1592 p->static_prio = NICE_TO_PRIO(0);
1594 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1595 p->static_prio = NICE_TO_PRIO(0);
1597 p->prio = p->normal_prio = __normal_prio(p);
1601 * We don't need the reset flag anymore after the fork. It has
1602 * fulfilled its duty:
1604 p->sched_reset_on_fork = 0;
1607 if (!rt_prio(p->prio))
1608 p->sched_class = &fair_sched_class;
1610 if (p->sched_class->task_fork)
1611 p->sched_class->task_fork(p);
1614 * The child is not yet in the pid-hash so no cgroup attach races,
1615 * and the cgroup is pinned to this child due to cgroup_fork()
1616 * is ran before sched_fork().
1618 * Silence PROVE_RCU.
1620 raw_spin_lock_irqsave(&p->pi_lock, flags);
1621 set_task_cpu(p, cpu);
1622 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1624 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1625 if (likely(sched_info_on()))
1626 memset(&p->sched_info, 0, sizeof(p->sched_info));
1628 #if defined(CONFIG_SMP)
1631 #ifdef CONFIG_PREEMPT_COUNT
1632 /* Want to start with kernel preemption disabled. */
1633 task_thread_info(p)->preempt_count = 1;
1636 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1643 * wake_up_new_task - wake up a newly created task for the first time.
1645 * This function will do some initial scheduler statistics housekeeping
1646 * that must be done for every newly created context, then puts the task
1647 * on the runqueue and wakes it.
1649 void wake_up_new_task(struct task_struct *p)
1651 unsigned long flags;
1654 raw_spin_lock_irqsave(&p->pi_lock, flags);
1657 * Fork balancing, do it here and not earlier because:
1658 * - cpus_allowed can change in the fork path
1659 * - any previously selected cpu might disappear through hotplug
1661 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1664 rq = __task_rq_lock(p);
1665 activate_task(rq, p, 0);
1667 trace_sched_wakeup_new(p, true);
1668 check_preempt_curr(rq, p, WF_FORK);
1670 if (p->sched_class->task_woken)
1671 p->sched_class->task_woken(rq, p);
1673 task_rq_unlock(rq, p, &flags);
1676 #ifdef CONFIG_PREEMPT_NOTIFIERS
1679 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1680 * @notifier: notifier struct to register
1682 void preempt_notifier_register(struct preempt_notifier *notifier)
1684 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1686 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1689 * preempt_notifier_unregister - no longer interested in preemption notifications
1690 * @notifier: notifier struct to unregister
1692 * This is safe to call from within a preemption notifier.
1694 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1696 hlist_del(¬ifier->link);
1698 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1700 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1702 struct preempt_notifier *notifier;
1703 struct hlist_node *node;
1705 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1706 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1710 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1711 struct task_struct *next)
1713 struct preempt_notifier *notifier;
1714 struct hlist_node *node;
1716 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1717 notifier->ops->sched_out(notifier, next);
1720 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1722 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1727 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1728 struct task_struct *next)
1732 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1735 * prepare_task_switch - prepare to switch tasks
1736 * @rq: the runqueue preparing to switch
1737 * @prev: the current task that is being switched out
1738 * @next: the task we are going to switch to.
1740 * This is called with the rq lock held and interrupts off. It must
1741 * be paired with a subsequent finish_task_switch after the context
1744 * prepare_task_switch sets up locking and calls architecture specific
1748 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1749 struct task_struct *next)
1751 trace_sched_switch(prev, next);
1752 sched_info_switch(prev, next);
1753 perf_event_task_sched_out(prev, next);
1754 fire_sched_out_preempt_notifiers(prev, next);
1755 prepare_lock_switch(rq, next);
1756 prepare_arch_switch(next);
1760 * finish_task_switch - clean up after a task-switch
1761 * @rq: runqueue associated with task-switch
1762 * @prev: the thread we just switched away from.
1764 * finish_task_switch must be called after the context switch, paired
1765 * with a prepare_task_switch call before the context switch.
1766 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1767 * and do any other architecture-specific cleanup actions.
1769 * Note that we may have delayed dropping an mm in context_switch(). If
1770 * so, we finish that here outside of the runqueue lock. (Doing it
1771 * with the lock held can cause deadlocks; see schedule() for
1774 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1775 __releases(rq->lock)
1777 struct mm_struct *mm = rq->prev_mm;
1783 * A task struct has one reference for the use as "current".
1784 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1785 * schedule one last time. The schedule call will never return, and
1786 * the scheduled task must drop that reference.
1787 * The test for TASK_DEAD must occur while the runqueue locks are
1788 * still held, otherwise prev could be scheduled on another cpu, die
1789 * there before we look at prev->state, and then the reference would
1791 * Manfred Spraul <manfred@colorfullife.com>
1793 prev_state = prev->state;
1794 vtime_task_switch(prev);
1795 finish_arch_switch(prev);
1796 perf_event_task_sched_in(prev, current);
1797 finish_lock_switch(rq, prev);
1798 finish_arch_post_lock_switch();
1800 fire_sched_in_preempt_notifiers(current);
1803 if (unlikely(prev_state == TASK_DEAD)) {
1805 * Remove function-return probe instances associated with this
1806 * task and put them back on the free list.
1808 kprobe_flush_task(prev);
1809 put_task_struct(prev);
1815 /* assumes rq->lock is held */
1816 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1818 if (prev->sched_class->pre_schedule)
1819 prev->sched_class->pre_schedule(rq, prev);
1822 /* rq->lock is NOT held, but preemption is disabled */
1823 static inline void post_schedule(struct rq *rq)
1825 if (rq->post_schedule) {
1826 unsigned long flags;
1828 raw_spin_lock_irqsave(&rq->lock, flags);
1829 if (rq->curr->sched_class->post_schedule)
1830 rq->curr->sched_class->post_schedule(rq);
1831 raw_spin_unlock_irqrestore(&rq->lock, flags);
1833 rq->post_schedule = 0;
1839 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1843 static inline void post_schedule(struct rq *rq)
1850 * schedule_tail - first thing a freshly forked thread must call.
1851 * @prev: the thread we just switched away from.
1853 asmlinkage void schedule_tail(struct task_struct *prev)
1854 __releases(rq->lock)
1856 struct rq *rq = this_rq();
1858 finish_task_switch(rq, prev);
1861 * FIXME: do we need to worry about rq being invalidated by the
1866 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1867 /* In this case, finish_task_switch does not reenable preemption */
1870 if (current->set_child_tid)
1871 put_user(task_pid_vnr(current), current->set_child_tid);
1875 * context_switch - switch to the new MM and the new
1876 * thread's register state.
1879 context_switch(struct rq *rq, struct task_struct *prev,
1880 struct task_struct *next)
1882 struct mm_struct *mm, *oldmm;
1884 prepare_task_switch(rq, prev, next);
1887 oldmm = prev->active_mm;
1889 * For paravirt, this is coupled with an exit in switch_to to
1890 * combine the page table reload and the switch backend into
1893 arch_start_context_switch(prev);
1896 next->active_mm = oldmm;
1897 atomic_inc(&oldmm->mm_count);
1898 enter_lazy_tlb(oldmm, next);
1900 switch_mm(oldmm, mm, next);
1903 prev->active_mm = NULL;
1904 rq->prev_mm = oldmm;
1907 * Since the runqueue lock will be released by the next
1908 * task (which is an invalid locking op but in the case
1909 * of the scheduler it's an obvious special-case), so we
1910 * do an early lockdep release here:
1912 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1913 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1916 context_tracking_task_switch(prev, next);
1917 /* Here we just switch the register state and the stack. */
1918 switch_to(prev, next, prev);
1922 * this_rq must be evaluated again because prev may have moved
1923 * CPUs since it called schedule(), thus the 'rq' on its stack
1924 * frame will be invalid.
1926 finish_task_switch(this_rq(), prev);
1930 * nr_running, nr_uninterruptible and nr_context_switches:
1932 * externally visible scheduler statistics: current number of runnable
1933 * threads, current number of uninterruptible-sleeping threads, total
1934 * number of context switches performed since bootup.
1936 unsigned long nr_running(void)
1938 unsigned long i, sum = 0;
1940 for_each_online_cpu(i)
1941 sum += cpu_rq(i)->nr_running;
1946 unsigned long nr_uninterruptible(void)
1948 unsigned long i, sum = 0;
1950 for_each_possible_cpu(i)
1951 sum += cpu_rq(i)->nr_uninterruptible;
1954 * Since we read the counters lockless, it might be slightly
1955 * inaccurate. Do not allow it to go below zero though:
1957 if (unlikely((long)sum < 0))
1963 unsigned long long nr_context_switches(void)
1966 unsigned long long sum = 0;
1968 for_each_possible_cpu(i)
1969 sum += cpu_rq(i)->nr_switches;
1974 unsigned long nr_iowait(void)
1976 unsigned long i, sum = 0;
1978 for_each_possible_cpu(i)
1979 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1984 unsigned long nr_iowait_cpu(int cpu)
1986 struct rq *this = cpu_rq(cpu);
1987 return atomic_read(&this->nr_iowait);
1990 unsigned long this_cpu_load(void)
1992 struct rq *this = this_rq();
1993 return this->cpu_load[0];
1998 * Global load-average calculations
2000 * We take a distributed and async approach to calculating the global load-avg
2001 * in order to minimize overhead.
2003 * The global load average is an exponentially decaying average of nr_running +
2004 * nr_uninterruptible.
2006 * Once every LOAD_FREQ:
2009 * for_each_possible_cpu(cpu)
2010 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2012 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2014 * Due to a number of reasons the above turns in the mess below:
2016 * - for_each_possible_cpu() is prohibitively expensive on machines with
2017 * serious number of cpus, therefore we need to take a distributed approach
2018 * to calculating nr_active.
2020 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2021 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2023 * So assuming nr_active := 0 when we start out -- true per definition, we
2024 * can simply take per-cpu deltas and fold those into a global accumulate
2025 * to obtain the same result. See calc_load_fold_active().
2027 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2028 * across the machine, we assume 10 ticks is sufficient time for every
2029 * cpu to have completed this task.
2031 * This places an upper-bound on the IRQ-off latency of the machine. Then
2032 * again, being late doesn't loose the delta, just wrecks the sample.
2034 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2035 * this would add another cross-cpu cacheline miss and atomic operation
2036 * to the wakeup path. Instead we increment on whatever cpu the task ran
2037 * when it went into uninterruptible state and decrement on whatever cpu
2038 * did the wakeup. This means that only the sum of nr_uninterruptible over
2039 * all cpus yields the correct result.
2041 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2044 /* Variables and functions for calc_load */
2045 static atomic_long_t calc_load_tasks;
2046 static unsigned long calc_load_update;
2047 unsigned long avenrun[3];
2048 EXPORT_SYMBOL(avenrun); /* should be removed */
2051 * get_avenrun - get the load average array
2052 * @loads: pointer to dest load array
2053 * @offset: offset to add
2054 * @shift: shift count to shift the result left
2056 * These values are estimates at best, so no need for locking.
2058 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2060 loads[0] = (avenrun[0] + offset) << shift;
2061 loads[1] = (avenrun[1] + offset) << shift;
2062 loads[2] = (avenrun[2] + offset) << shift;
2065 static long calc_load_fold_active(struct rq *this_rq)
2067 long nr_active, delta = 0;
2069 nr_active = this_rq->nr_running;
2070 nr_active += (long) this_rq->nr_uninterruptible;
2072 if (nr_active != this_rq->calc_load_active) {
2073 delta = nr_active - this_rq->calc_load_active;
2074 this_rq->calc_load_active = nr_active;
2081 * a1 = a0 * e + a * (1 - e)
2083 static unsigned long
2084 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2087 load += active * (FIXED_1 - exp);
2088 load += 1UL << (FSHIFT - 1);
2089 return load >> FSHIFT;
2094 * Handle NO_HZ for the global load-average.
2096 * Since the above described distributed algorithm to compute the global
2097 * load-average relies on per-cpu sampling from the tick, it is affected by
2100 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2101 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2102 * when we read the global state.
2104 * Obviously reality has to ruin such a delightfully simple scheme:
2106 * - When we go NO_HZ idle during the window, we can negate our sample
2107 * contribution, causing under-accounting.
2109 * We avoid this by keeping two idle-delta counters and flipping them
2110 * when the window starts, thus separating old and new NO_HZ load.
2112 * The only trick is the slight shift in index flip for read vs write.
2116 * |-|-----------|-|-----------|-|-----------|-|
2117 * r:0 0 1 1 0 0 1 1 0
2118 * w:0 1 1 0 0 1 1 0 0
2120 * This ensures we'll fold the old idle contribution in this window while
2121 * accumlating the new one.
2123 * - When we wake up from NO_HZ idle during the window, we push up our
2124 * contribution, since we effectively move our sample point to a known
2127 * This is solved by pushing the window forward, and thus skipping the
2128 * sample, for this cpu (effectively using the idle-delta for this cpu which
2129 * was in effect at the time the window opened). This also solves the issue
2130 * of having to deal with a cpu having been in NOHZ idle for multiple
2131 * LOAD_FREQ intervals.
2133 * When making the ILB scale, we should try to pull this in as well.
2135 static atomic_long_t calc_load_idle[2];
2136 static int calc_load_idx;
2138 static inline int calc_load_write_idx(void)
2140 int idx = calc_load_idx;
2143 * See calc_global_nohz(), if we observe the new index, we also
2144 * need to observe the new update time.
2149 * If the folding window started, make sure we start writing in the
2152 if (!time_before(jiffies, calc_load_update))
2158 static inline int calc_load_read_idx(void)
2160 return calc_load_idx & 1;
2163 void calc_load_enter_idle(void)
2165 struct rq *this_rq = this_rq();
2169 * We're going into NOHZ mode, if there's any pending delta, fold it
2170 * into the pending idle delta.
2172 delta = calc_load_fold_active(this_rq);
2174 int idx = calc_load_write_idx();
2175 atomic_long_add(delta, &calc_load_idle[idx]);
2179 void calc_load_exit_idle(void)
2181 struct rq *this_rq = this_rq();
2184 * If we're still before the sample window, we're done.
2186 if (time_before(jiffies, this_rq->calc_load_update))
2190 * We woke inside or after the sample window, this means we're already
2191 * accounted through the nohz accounting, so skip the entire deal and
2192 * sync up for the next window.
2194 this_rq->calc_load_update = calc_load_update;
2195 if (time_before(jiffies, this_rq->calc_load_update + 10))
2196 this_rq->calc_load_update += LOAD_FREQ;
2199 static long calc_load_fold_idle(void)
2201 int idx = calc_load_read_idx();
2204 if (atomic_long_read(&calc_load_idle[idx]))
2205 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2211 * fixed_power_int - compute: x^n, in O(log n) time
2213 * @x: base of the power
2214 * @frac_bits: fractional bits of @x
2215 * @n: power to raise @x to.
2217 * By exploiting the relation between the definition of the natural power
2218 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2219 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2220 * (where: n_i \elem {0, 1}, the binary vector representing n),
2221 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2222 * of course trivially computable in O(log_2 n), the length of our binary
2225 static unsigned long
2226 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2228 unsigned long result = 1UL << frac_bits;
2233 result += 1UL << (frac_bits - 1);
2234 result >>= frac_bits;
2240 x += 1UL << (frac_bits - 1);
2248 * a1 = a0 * e + a * (1 - e)
2250 * a2 = a1 * e + a * (1 - e)
2251 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2252 * = a0 * e^2 + a * (1 - e) * (1 + e)
2254 * a3 = a2 * e + a * (1 - e)
2255 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2256 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2260 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2261 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2262 * = a0 * e^n + a * (1 - e^n)
2264 * [1] application of the geometric series:
2267 * S_n := \Sum x^i = -------------
2270 static unsigned long
2271 calc_load_n(unsigned long load, unsigned long exp,
2272 unsigned long active, unsigned int n)
2275 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2279 * NO_HZ can leave us missing all per-cpu ticks calling
2280 * calc_load_account_active(), but since an idle CPU folds its delta into
2281 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2282 * in the pending idle delta if our idle period crossed a load cycle boundary.
2284 * Once we've updated the global active value, we need to apply the exponential
2285 * weights adjusted to the number of cycles missed.
2287 static void calc_global_nohz(void)
2289 long delta, active, n;
2291 if (!time_before(jiffies, calc_load_update + 10)) {
2293 * Catch-up, fold however many we are behind still
2295 delta = jiffies - calc_load_update - 10;
2296 n = 1 + (delta / LOAD_FREQ);
2298 active = atomic_long_read(&calc_load_tasks);
2299 active = active > 0 ? active * FIXED_1 : 0;
2301 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2302 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2303 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2305 calc_load_update += n * LOAD_FREQ;
2309 * Flip the idle index...
2311 * Make sure we first write the new time then flip the index, so that
2312 * calc_load_write_idx() will see the new time when it reads the new
2313 * index, this avoids a double flip messing things up.
2318 #else /* !CONFIG_NO_HZ */
2320 static inline long calc_load_fold_idle(void) { return 0; }
2321 static inline void calc_global_nohz(void) { }
2323 #endif /* CONFIG_NO_HZ */
2326 * calc_load - update the avenrun load estimates 10 ticks after the
2327 * CPUs have updated calc_load_tasks.
2329 void calc_global_load(unsigned long ticks)
2333 if (time_before(jiffies, calc_load_update + 10))
2337 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2339 delta = calc_load_fold_idle();
2341 atomic_long_add(delta, &calc_load_tasks);
2343 active = atomic_long_read(&calc_load_tasks);
2344 active = active > 0 ? active * FIXED_1 : 0;
2346 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2347 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2348 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2350 calc_load_update += LOAD_FREQ;
2353 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2359 * Called from update_cpu_load() to periodically update this CPU's
2362 static void calc_load_account_active(struct rq *this_rq)
2366 if (time_before(jiffies, this_rq->calc_load_update))
2369 delta = calc_load_fold_active(this_rq);
2371 atomic_long_add(delta, &calc_load_tasks);
2373 this_rq->calc_load_update += LOAD_FREQ;
2377 * End of global load-average stuff
2381 * The exact cpuload at various idx values, calculated at every tick would be
2382 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2384 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2385 * on nth tick when cpu may be busy, then we have:
2386 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2387 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2389 * decay_load_missed() below does efficient calculation of
2390 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2391 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2393 * The calculation is approximated on a 128 point scale.
2394 * degrade_zero_ticks is the number of ticks after which load at any
2395 * particular idx is approximated to be zero.
2396 * degrade_factor is a precomputed table, a row for each load idx.
2397 * Each column corresponds to degradation factor for a power of two ticks,
2398 * based on 128 point scale.
2400 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2401 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2403 * With this power of 2 load factors, we can degrade the load n times
2404 * by looking at 1 bits in n and doing as many mult/shift instead of
2405 * n mult/shifts needed by the exact degradation.
2407 #define DEGRADE_SHIFT 7
2408 static const unsigned char
2409 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2410 static const unsigned char
2411 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2412 {0, 0, 0, 0, 0, 0, 0, 0},
2413 {64, 32, 8, 0, 0, 0, 0, 0},
2414 {96, 72, 40, 12, 1, 0, 0},
2415 {112, 98, 75, 43, 15, 1, 0},
2416 {120, 112, 98, 76, 45, 16, 2} };
2419 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2420 * would be when CPU is idle and so we just decay the old load without
2421 * adding any new load.
2423 static unsigned long
2424 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2428 if (!missed_updates)
2431 if (missed_updates >= degrade_zero_ticks[idx])
2435 return load >> missed_updates;
2437 while (missed_updates) {
2438 if (missed_updates % 2)
2439 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2441 missed_updates >>= 1;
2448 * Update rq->cpu_load[] statistics. This function is usually called every
2449 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2450 * every tick. We fix it up based on jiffies.
2452 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2453 unsigned long pending_updates)
2457 this_rq->nr_load_updates++;
2459 /* Update our load: */
2460 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2461 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2462 unsigned long old_load, new_load;
2464 /* scale is effectively 1 << i now, and >> i divides by scale */
2466 old_load = this_rq->cpu_load[i];
2467 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2468 new_load = this_load;
2470 * Round up the averaging division if load is increasing. This
2471 * prevents us from getting stuck on 9 if the load is 10, for
2474 if (new_load > old_load)
2475 new_load += scale - 1;
2477 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2480 sched_avg_update(this_rq);
2485 * There is no sane way to deal with nohz on smp when using jiffies because the
2486 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2487 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2489 * Therefore we cannot use the delta approach from the regular tick since that
2490 * would seriously skew the load calculation. However we'll make do for those
2491 * updates happening while idle (nohz_idle_balance) or coming out of idle
2492 * (tick_nohz_idle_exit).
2494 * This means we might still be one tick off for nohz periods.
2498 * Called from nohz_idle_balance() to update the load ratings before doing the
2501 void update_idle_cpu_load(struct rq *this_rq)
2503 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2504 unsigned long load = this_rq->load.weight;
2505 unsigned long pending_updates;
2508 * bail if there's load or we're actually up-to-date.
2510 if (load || curr_jiffies == this_rq->last_load_update_tick)
2513 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2514 this_rq->last_load_update_tick = curr_jiffies;
2516 __update_cpu_load(this_rq, load, pending_updates);
2520 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2522 void update_cpu_load_nohz(void)
2524 struct rq *this_rq = this_rq();
2525 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2526 unsigned long pending_updates;
2528 if (curr_jiffies == this_rq->last_load_update_tick)
2531 raw_spin_lock(&this_rq->lock);
2532 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2533 if (pending_updates) {
2534 this_rq->last_load_update_tick = curr_jiffies;
2536 * We were idle, this means load 0, the current load might be
2537 * !0 due to remote wakeups and the sort.
2539 __update_cpu_load(this_rq, 0, pending_updates);
2541 raw_spin_unlock(&this_rq->lock);
2543 #endif /* CONFIG_NO_HZ */
2546 * Called from scheduler_tick()
2548 static void update_cpu_load_active(struct rq *this_rq)
2551 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2553 this_rq->last_load_update_tick = jiffies;
2554 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2556 calc_load_account_active(this_rq);
2562 * sched_exec - execve() is a valuable balancing opportunity, because at
2563 * this point the task has the smallest effective memory and cache footprint.
2565 void sched_exec(void)
2567 struct task_struct *p = current;
2568 unsigned long flags;
2571 raw_spin_lock_irqsave(&p->pi_lock, flags);
2572 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2573 if (dest_cpu == smp_processor_id())
2576 if (likely(cpu_active(dest_cpu))) {
2577 struct migration_arg arg = { p, dest_cpu };
2579 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2580 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2584 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2589 DEFINE_PER_CPU(struct kernel_stat, kstat);
2590 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2592 EXPORT_PER_CPU_SYMBOL(kstat);
2593 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2596 * Return any ns on the sched_clock that have not yet been accounted in
2597 * @p in case that task is currently running.
2599 * Called with task_rq_lock() held on @rq.
2601 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2605 if (task_current(rq, p)) {
2606 update_rq_clock(rq);
2607 ns = rq->clock_task - p->se.exec_start;
2615 unsigned long long task_delta_exec(struct task_struct *p)
2617 unsigned long flags;
2621 rq = task_rq_lock(p, &flags);
2622 ns = do_task_delta_exec(p, rq);
2623 task_rq_unlock(rq, p, &flags);
2629 * Return accounted runtime for the task.
2630 * In case the task is currently running, return the runtime plus current's
2631 * pending runtime that have not been accounted yet.
2633 unsigned long long task_sched_runtime(struct task_struct *p)
2635 unsigned long flags;
2639 rq = task_rq_lock(p, &flags);
2640 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2641 task_rq_unlock(rq, p, &flags);
2647 * This function gets called by the timer code, with HZ frequency.
2648 * We call it with interrupts disabled.
2650 void scheduler_tick(void)
2652 int cpu = smp_processor_id();
2653 struct rq *rq = cpu_rq(cpu);
2654 struct task_struct *curr = rq->curr;
2658 raw_spin_lock(&rq->lock);
2659 update_rq_clock(rq);
2660 update_cpu_load_active(rq);
2661 curr->sched_class->task_tick(rq, curr, 0);
2662 raw_spin_unlock(&rq->lock);
2664 perf_event_task_tick();
2667 rq->idle_balance = idle_cpu(cpu);
2668 trigger_load_balance(rq, cpu);
2672 notrace unsigned long get_parent_ip(unsigned long addr)
2674 if (in_lock_functions(addr)) {
2675 addr = CALLER_ADDR2;
2676 if (in_lock_functions(addr))
2677 addr = CALLER_ADDR3;
2682 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2683 defined(CONFIG_PREEMPT_TRACER))
2685 void __kprobes add_preempt_count(int val)
2687 #ifdef CONFIG_DEBUG_PREEMPT
2691 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2694 preempt_count() += val;
2695 #ifdef CONFIG_DEBUG_PREEMPT
2697 * Spinlock count overflowing soon?
2699 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2702 if (preempt_count() == val)
2703 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2705 EXPORT_SYMBOL(add_preempt_count);
2707 void __kprobes sub_preempt_count(int val)
2709 #ifdef CONFIG_DEBUG_PREEMPT
2713 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2716 * Is the spinlock portion underflowing?
2718 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2719 !(preempt_count() & PREEMPT_MASK)))
2723 if (preempt_count() == val)
2724 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2725 preempt_count() -= val;
2727 EXPORT_SYMBOL(sub_preempt_count);
2732 * Print scheduling while atomic bug:
2734 static noinline void __schedule_bug(struct task_struct *prev)
2736 if (oops_in_progress)
2739 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2740 prev->comm, prev->pid, preempt_count());
2742 debug_show_held_locks(prev);
2744 if (irqs_disabled())
2745 print_irqtrace_events(prev);
2747 add_taint(TAINT_WARN);
2751 * Various schedule()-time debugging checks and statistics:
2753 static inline void schedule_debug(struct task_struct *prev)
2756 * Test if we are atomic. Since do_exit() needs to call into
2757 * schedule() atomically, we ignore that path for now.
2758 * Otherwise, whine if we are scheduling when we should not be.
2760 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2761 __schedule_bug(prev);
2764 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2766 schedstat_inc(this_rq(), sched_count);
2769 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2771 if (prev->on_rq || rq->skip_clock_update < 0)
2772 update_rq_clock(rq);
2773 prev->sched_class->put_prev_task(rq, prev);
2777 * Pick up the highest-prio task:
2779 static inline struct task_struct *
2780 pick_next_task(struct rq *rq)
2782 const struct sched_class *class;
2783 struct task_struct *p;
2786 * Optimization: we know that if all tasks are in
2787 * the fair class we can call that function directly:
2789 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2790 p = fair_sched_class.pick_next_task(rq);
2795 for_each_class(class) {
2796 p = class->pick_next_task(rq);
2801 BUG(); /* the idle class will always have a runnable task */
2805 * __schedule() is the main scheduler function.
2807 * The main means of driving the scheduler and thus entering this function are:
2809 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2811 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2812 * paths. For example, see arch/x86/entry_64.S.
2814 * To drive preemption between tasks, the scheduler sets the flag in timer
2815 * interrupt handler scheduler_tick().
2817 * 3. Wakeups don't really cause entry into schedule(). They add a
2818 * task to the run-queue and that's it.
2820 * Now, if the new task added to the run-queue preempts the current
2821 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2822 * called on the nearest possible occasion:
2824 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2826 * - in syscall or exception context, at the next outmost
2827 * preempt_enable(). (this might be as soon as the wake_up()'s
2830 * - in IRQ context, return from interrupt-handler to
2831 * preemptible context
2833 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2836 * - cond_resched() call
2837 * - explicit schedule() call
2838 * - return from syscall or exception to user-space
2839 * - return from interrupt-handler to user-space
2841 static void __sched __schedule(void)
2843 struct task_struct *prev, *next;
2844 unsigned long *switch_count;
2850 cpu = smp_processor_id();
2852 rcu_note_context_switch(cpu);
2855 schedule_debug(prev);
2857 if (sched_feat(HRTICK))
2860 raw_spin_lock_irq(&rq->lock);
2862 switch_count = &prev->nivcsw;
2863 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2864 if (unlikely(signal_pending_state(prev->state, prev))) {
2865 prev->state = TASK_RUNNING;
2867 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2871 * If a worker went to sleep, notify and ask workqueue
2872 * whether it wants to wake up a task to maintain
2875 if (prev->flags & PF_WQ_WORKER) {
2876 struct task_struct *to_wakeup;
2878 to_wakeup = wq_worker_sleeping(prev, cpu);
2880 try_to_wake_up_local(to_wakeup);
2883 switch_count = &prev->nvcsw;
2886 pre_schedule(rq, prev);
2888 if (unlikely(!rq->nr_running))
2889 idle_balance(cpu, rq);
2891 put_prev_task(rq, prev);
2892 next = pick_next_task(rq);
2893 clear_tsk_need_resched(prev);
2894 rq->skip_clock_update = 0;
2896 if (likely(prev != next)) {
2901 context_switch(rq, prev, next); /* unlocks the rq */
2903 * The context switch have flipped the stack from under us
2904 * and restored the local variables which were saved when
2905 * this task called schedule() in the past. prev == current
2906 * is still correct, but it can be moved to another cpu/rq.
2908 cpu = smp_processor_id();
2911 raw_spin_unlock_irq(&rq->lock);
2915 sched_preempt_enable_no_resched();
2920 static inline void sched_submit_work(struct task_struct *tsk)
2922 if (!tsk->state || tsk_is_pi_blocked(tsk))
2925 * If we are going to sleep and we have plugged IO queued,
2926 * make sure to submit it to avoid deadlocks.
2928 if (blk_needs_flush_plug(tsk))
2929 blk_schedule_flush_plug(tsk);
2932 asmlinkage void __sched schedule(void)
2934 struct task_struct *tsk = current;
2936 sched_submit_work(tsk);
2939 EXPORT_SYMBOL(schedule);
2941 #ifdef CONFIG_CONTEXT_TRACKING
2942 asmlinkage void __sched schedule_user(void)
2945 * If we come here after a random call to set_need_resched(),
2946 * or we have been woken up remotely but the IPI has not yet arrived,
2947 * we haven't yet exited the RCU idle mode. Do it here manually until
2948 * we find a better solution.
2957 * schedule_preempt_disabled - called with preemption disabled
2959 * Returns with preemption disabled. Note: preempt_count must be 1
2961 void __sched schedule_preempt_disabled(void)
2963 sched_preempt_enable_no_resched();
2968 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2970 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2972 if (lock->owner != owner)
2976 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2977 * lock->owner still matches owner, if that fails, owner might
2978 * point to free()d memory, if it still matches, the rcu_read_lock()
2979 * ensures the memory stays valid.
2983 return owner->on_cpu;
2987 * Look out! "owner" is an entirely speculative pointer
2988 * access and not reliable.
2990 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
2992 if (!sched_feat(OWNER_SPIN))
2996 while (owner_running(lock, owner)) {
3000 arch_mutex_cpu_relax();
3005 * We break out the loop above on need_resched() and when the
3006 * owner changed, which is a sign for heavy contention. Return
3007 * success only when lock->owner is NULL.
3009 return lock->owner == NULL;
3013 #ifdef CONFIG_PREEMPT
3015 * this is the entry point to schedule() from in-kernel preemption
3016 * off of preempt_enable. Kernel preemptions off return from interrupt
3017 * occur there and call schedule directly.
3019 asmlinkage void __sched notrace preempt_schedule(void)
3021 struct thread_info *ti = current_thread_info();
3024 * If there is a non-zero preempt_count or interrupts are disabled,
3025 * we do not want to preempt the current task. Just return..
3027 if (likely(ti->preempt_count || irqs_disabled()))
3031 add_preempt_count_notrace(PREEMPT_ACTIVE);
3033 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3036 * Check again in case we missed a preemption opportunity
3037 * between schedule and now.
3040 } while (need_resched());
3042 EXPORT_SYMBOL(preempt_schedule);
3045 * this is the entry point to schedule() from kernel preemption
3046 * off of irq context.
3047 * Note, that this is called and return with irqs disabled. This will
3048 * protect us against recursive calling from irq.
3050 asmlinkage void __sched preempt_schedule_irq(void)
3052 struct thread_info *ti = current_thread_info();
3054 /* Catch callers which need to be fixed */
3055 BUG_ON(ti->preempt_count || !irqs_disabled());
3059 add_preempt_count(PREEMPT_ACTIVE);
3062 local_irq_disable();
3063 sub_preempt_count(PREEMPT_ACTIVE);
3066 * Check again in case we missed a preemption opportunity
3067 * between schedule and now.
3070 } while (need_resched());
3073 #endif /* CONFIG_PREEMPT */
3075 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3078 return try_to_wake_up(curr->private, mode, wake_flags);
3080 EXPORT_SYMBOL(default_wake_function);
3083 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3084 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3085 * number) then we wake all the non-exclusive tasks and one exclusive task.
3087 * There are circumstances in which we can try to wake a task which has already
3088 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3089 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3091 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3092 int nr_exclusive, int wake_flags, void *key)
3094 wait_queue_t *curr, *next;
3096 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3097 unsigned flags = curr->flags;
3099 if (curr->func(curr, mode, wake_flags, key) &&
3100 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3106 * __wake_up - wake up threads blocked on a waitqueue.
3108 * @mode: which threads
3109 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3110 * @key: is directly passed to the wakeup function
3112 * It may be assumed that this function implies a write memory barrier before
3113 * changing the task state if and only if any tasks are woken up.
3115 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3116 int nr_exclusive, void *key)
3118 unsigned long flags;
3120 spin_lock_irqsave(&q->lock, flags);
3121 __wake_up_common(q, mode, nr_exclusive, 0, key);
3122 spin_unlock_irqrestore(&q->lock, flags);
3124 EXPORT_SYMBOL(__wake_up);
3127 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3129 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3131 __wake_up_common(q, mode, nr, 0, NULL);
3133 EXPORT_SYMBOL_GPL(__wake_up_locked);
3135 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3137 __wake_up_common(q, mode, 1, 0, key);
3139 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3142 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3144 * @mode: which threads
3145 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3146 * @key: opaque value to be passed to wakeup targets
3148 * The sync wakeup differs that the waker knows that it will schedule
3149 * away soon, so while the target thread will be woken up, it will not
3150 * be migrated to another CPU - ie. the two threads are 'synchronized'
3151 * with each other. This can prevent needless bouncing between CPUs.
3153 * On UP it can prevent extra preemption.
3155 * It may be assumed that this function implies a write memory barrier before
3156 * changing the task state if and only if any tasks are woken up.
3158 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3159 int nr_exclusive, void *key)
3161 unsigned long flags;
3162 int wake_flags = WF_SYNC;
3167 if (unlikely(!nr_exclusive))
3170 spin_lock_irqsave(&q->lock, flags);
3171 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3172 spin_unlock_irqrestore(&q->lock, flags);
3174 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3177 * __wake_up_sync - see __wake_up_sync_key()
3179 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3181 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3183 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3186 * complete: - signals a single thread waiting on this completion
3187 * @x: holds the state of this particular completion
3189 * This will wake up a single thread waiting on this completion. Threads will be
3190 * awakened in the same order in which they were queued.
3192 * See also complete_all(), wait_for_completion() and related routines.
3194 * It may be assumed that this function implies a write memory barrier before
3195 * changing the task state if and only if any tasks are woken up.
3197 void complete(struct completion *x)
3199 unsigned long flags;
3201 spin_lock_irqsave(&x->wait.lock, flags);
3203 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3204 spin_unlock_irqrestore(&x->wait.lock, flags);
3206 EXPORT_SYMBOL(complete);
3209 * complete_all: - signals all threads waiting on this completion
3210 * @x: holds the state of this particular completion
3212 * This will wake up all threads waiting on this particular completion event.
3214 * It may be assumed that this function implies a write memory barrier before
3215 * changing the task state if and only if any tasks are woken up.
3217 void complete_all(struct completion *x)
3219 unsigned long flags;
3221 spin_lock_irqsave(&x->wait.lock, flags);
3222 x->done += UINT_MAX/2;
3223 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3224 spin_unlock_irqrestore(&x->wait.lock, flags);
3226 EXPORT_SYMBOL(complete_all);
3228 static inline long __sched
3229 do_wait_for_common(struct completion *x, long timeout, int state)
3232 DECLARE_WAITQUEUE(wait, current);
3234 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3236 if (signal_pending_state(state, current)) {
3237 timeout = -ERESTARTSYS;
3240 __set_current_state(state);
3241 spin_unlock_irq(&x->wait.lock);
3242 timeout = schedule_timeout(timeout);
3243 spin_lock_irq(&x->wait.lock);
3244 } while (!x->done && timeout);
3245 __remove_wait_queue(&x->wait, &wait);
3250 return timeout ?: 1;
3254 wait_for_common(struct completion *x, long timeout, int state)
3258 spin_lock_irq(&x->wait.lock);
3259 timeout = do_wait_for_common(x, timeout, state);
3260 spin_unlock_irq(&x->wait.lock);
3265 * wait_for_completion: - waits for completion of a task
3266 * @x: holds the state of this particular completion
3268 * This waits to be signaled for completion of a specific task. It is NOT
3269 * interruptible and there is no timeout.
3271 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3272 * and interrupt capability. Also see complete().
3274 void __sched wait_for_completion(struct completion *x)
3276 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3278 EXPORT_SYMBOL(wait_for_completion);
3281 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3282 * @x: holds the state of this particular completion
3283 * @timeout: timeout value in jiffies
3285 * This waits for either a completion of a specific task to be signaled or for a
3286 * specified timeout to expire. The timeout is in jiffies. It is not
3289 * The return value is 0 if timed out, and positive (at least 1, or number of
3290 * jiffies left till timeout) if completed.
3292 unsigned long __sched
3293 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3295 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3297 EXPORT_SYMBOL(wait_for_completion_timeout);
3300 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3301 * @x: holds the state of this particular completion
3303 * This waits for completion of a specific task to be signaled. It is
3306 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3308 int __sched wait_for_completion_interruptible(struct completion *x)
3310 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3311 if (t == -ERESTARTSYS)
3315 EXPORT_SYMBOL(wait_for_completion_interruptible);
3318 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3319 * @x: holds the state of this particular completion
3320 * @timeout: timeout value in jiffies
3322 * This waits for either a completion of a specific task to be signaled or for a
3323 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3325 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3326 * positive (at least 1, or number of jiffies left till timeout) if completed.
3329 wait_for_completion_interruptible_timeout(struct completion *x,
3330 unsigned long timeout)
3332 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3334 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3337 * wait_for_completion_killable: - waits for completion of a task (killable)
3338 * @x: holds the state of this particular completion
3340 * This waits to be signaled for completion of a specific task. It can be
3341 * interrupted by a kill signal.
3343 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3345 int __sched wait_for_completion_killable(struct completion *x)
3347 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3348 if (t == -ERESTARTSYS)
3352 EXPORT_SYMBOL(wait_for_completion_killable);
3355 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3356 * @x: holds the state of this particular completion
3357 * @timeout: timeout value in jiffies
3359 * This waits for either a completion of a specific task to be
3360 * signaled or for a specified timeout to expire. It can be
3361 * interrupted by a kill signal. The timeout is in jiffies.
3363 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3364 * positive (at least 1, or number of jiffies left till timeout) if completed.
3367 wait_for_completion_killable_timeout(struct completion *x,
3368 unsigned long timeout)
3370 return wait_for_common(x, timeout, TASK_KILLABLE);
3372 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3375 * try_wait_for_completion - try to decrement a completion without blocking
3376 * @x: completion structure
3378 * Returns: 0 if a decrement cannot be done without blocking
3379 * 1 if a decrement succeeded.
3381 * If a completion is being used as a counting completion,
3382 * attempt to decrement the counter without blocking. This
3383 * enables us to avoid waiting if the resource the completion
3384 * is protecting is not available.
3386 bool try_wait_for_completion(struct completion *x)
3388 unsigned long flags;
3391 spin_lock_irqsave(&x->wait.lock, flags);
3396 spin_unlock_irqrestore(&x->wait.lock, flags);
3399 EXPORT_SYMBOL(try_wait_for_completion);
3402 * completion_done - Test to see if a completion has any waiters
3403 * @x: completion structure
3405 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3406 * 1 if there are no waiters.
3409 bool completion_done(struct completion *x)
3411 unsigned long flags;
3414 spin_lock_irqsave(&x->wait.lock, flags);
3417 spin_unlock_irqrestore(&x->wait.lock, flags);
3420 EXPORT_SYMBOL(completion_done);
3423 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3425 unsigned long flags;
3428 init_waitqueue_entry(&wait, current);
3430 __set_current_state(state);
3432 spin_lock_irqsave(&q->lock, flags);
3433 __add_wait_queue(q, &wait);
3434 spin_unlock(&q->lock);
3435 timeout = schedule_timeout(timeout);
3436 spin_lock_irq(&q->lock);
3437 __remove_wait_queue(q, &wait);
3438 spin_unlock_irqrestore(&q->lock, flags);
3443 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3445 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3447 EXPORT_SYMBOL(interruptible_sleep_on);
3450 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3452 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3454 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3456 void __sched sleep_on(wait_queue_head_t *q)
3458 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3460 EXPORT_SYMBOL(sleep_on);
3462 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3464 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3466 EXPORT_SYMBOL(sleep_on_timeout);
3468 #ifdef CONFIG_RT_MUTEXES
3471 * rt_mutex_setprio - set the current priority of a task
3473 * @prio: prio value (kernel-internal form)
3475 * This function changes the 'effective' priority of a task. It does
3476 * not touch ->normal_prio like __setscheduler().
3478 * Used by the rt_mutex code to implement priority inheritance logic.
3480 void rt_mutex_setprio(struct task_struct *p, int prio)
3482 int oldprio, on_rq, running;
3484 const struct sched_class *prev_class;
3486 BUG_ON(prio < 0 || prio > MAX_PRIO);
3488 rq = __task_rq_lock(p);
3491 * Idle task boosting is a nono in general. There is one
3492 * exception, when PREEMPT_RT and NOHZ is active:
3494 * The idle task calls get_next_timer_interrupt() and holds
3495 * the timer wheel base->lock on the CPU and another CPU wants
3496 * to access the timer (probably to cancel it). We can safely
3497 * ignore the boosting request, as the idle CPU runs this code
3498 * with interrupts disabled and will complete the lock
3499 * protected section without being interrupted. So there is no
3500 * real need to boost.
3502 if (unlikely(p == rq->idle)) {
3503 WARN_ON(p != rq->curr);
3504 WARN_ON(p->pi_blocked_on);
3508 trace_sched_pi_setprio(p, prio);
3510 prev_class = p->sched_class;
3512 running = task_current(rq, p);
3514 dequeue_task(rq, p, 0);
3516 p->sched_class->put_prev_task(rq, p);
3519 p->sched_class = &rt_sched_class;
3521 p->sched_class = &fair_sched_class;
3526 p->sched_class->set_curr_task(rq);
3528 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3530 check_class_changed(rq, p, prev_class, oldprio);
3532 __task_rq_unlock(rq);
3535 void set_user_nice(struct task_struct *p, long nice)
3537 int old_prio, delta, on_rq;
3538 unsigned long flags;
3541 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3544 * We have to be careful, if called from sys_setpriority(),
3545 * the task might be in the middle of scheduling on another CPU.
3547 rq = task_rq_lock(p, &flags);
3549 * The RT priorities are set via sched_setscheduler(), but we still
3550 * allow the 'normal' nice value to be set - but as expected
3551 * it wont have any effect on scheduling until the task is
3552 * SCHED_FIFO/SCHED_RR:
3554 if (task_has_rt_policy(p)) {
3555 p->static_prio = NICE_TO_PRIO(nice);
3560 dequeue_task(rq, p, 0);
3562 p->static_prio = NICE_TO_PRIO(nice);
3565 p->prio = effective_prio(p);
3566 delta = p->prio - old_prio;
3569 enqueue_task(rq, p, 0);
3571 * If the task increased its priority or is running and
3572 * lowered its priority, then reschedule its CPU:
3574 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3575 resched_task(rq->curr);
3578 task_rq_unlock(rq, p, &flags);
3580 EXPORT_SYMBOL(set_user_nice);
3583 * can_nice - check if a task can reduce its nice value
3587 int can_nice(const struct task_struct *p, const int nice)
3589 /* convert nice value [19,-20] to rlimit style value [1,40] */
3590 int nice_rlim = 20 - nice;
3592 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3593 capable(CAP_SYS_NICE));
3596 #ifdef __ARCH_WANT_SYS_NICE
3599 * sys_nice - change the priority of the current process.
3600 * @increment: priority increment
3602 * sys_setpriority is a more generic, but much slower function that
3603 * does similar things.
3605 SYSCALL_DEFINE1(nice, int, increment)
3610 * Setpriority might change our priority at the same moment.
3611 * We don't have to worry. Conceptually one call occurs first
3612 * and we have a single winner.
3614 if (increment < -40)
3619 nice = TASK_NICE(current) + increment;
3625 if (increment < 0 && !can_nice(current, nice))
3628 retval = security_task_setnice(current, nice);
3632 set_user_nice(current, nice);
3639 * task_prio - return the priority value of a given task.
3640 * @p: the task in question.
3642 * This is the priority value as seen by users in /proc.
3643 * RT tasks are offset by -200. Normal tasks are centered
3644 * around 0, value goes from -16 to +15.
3646 int task_prio(const struct task_struct *p)
3648 return p->prio - MAX_RT_PRIO;
3652 * task_nice - return the nice value of a given task.
3653 * @p: the task in question.
3655 int task_nice(const struct task_struct *p)
3657 return TASK_NICE(p);
3659 EXPORT_SYMBOL(task_nice);
3662 * idle_cpu - is a given cpu idle currently?
3663 * @cpu: the processor in question.
3665 int idle_cpu(int cpu)
3667 struct rq *rq = cpu_rq(cpu);
3669 if (rq->curr != rq->idle)
3676 if (!llist_empty(&rq->wake_list))
3684 * idle_task - return the idle task for a given cpu.
3685 * @cpu: the processor in question.
3687 struct task_struct *idle_task(int cpu)
3689 return cpu_rq(cpu)->idle;
3693 * find_process_by_pid - find a process with a matching PID value.
3694 * @pid: the pid in question.
3696 static struct task_struct *find_process_by_pid(pid_t pid)
3698 return pid ? find_task_by_vpid(pid) : current;
3701 /* Actually do priority change: must hold rq lock. */
3703 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3706 p->rt_priority = prio;
3707 p->normal_prio = normal_prio(p);
3708 /* we are holding p->pi_lock already */
3709 p->prio = rt_mutex_getprio(p);
3710 if (rt_prio(p->prio))
3711 p->sched_class = &rt_sched_class;
3713 p->sched_class = &fair_sched_class;
3718 * check the target process has a UID that matches the current process's
3720 static bool check_same_owner(struct task_struct *p)
3722 const struct cred *cred = current_cred(), *pcred;
3726 pcred = __task_cred(p);
3727 match = (uid_eq(cred->euid, pcred->euid) ||
3728 uid_eq(cred->euid, pcred->uid));
3733 static int __sched_setscheduler(struct task_struct *p, int policy,
3734 const struct sched_param *param, bool user)
3736 int retval, oldprio, oldpolicy = -1, on_rq, running;
3737 unsigned long flags;
3738 const struct sched_class *prev_class;
3742 /* may grab non-irq protected spin_locks */
3743 BUG_ON(in_interrupt());
3745 /* double check policy once rq lock held */
3747 reset_on_fork = p->sched_reset_on_fork;
3748 policy = oldpolicy = p->policy;
3750 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3751 policy &= ~SCHED_RESET_ON_FORK;
3753 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3754 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3755 policy != SCHED_IDLE)
3760 * Valid priorities for SCHED_FIFO and SCHED_RR are
3761 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3762 * SCHED_BATCH and SCHED_IDLE is 0.
3764 if (param->sched_priority < 0 ||
3765 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3766 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3768 if (rt_policy(policy) != (param->sched_priority != 0))
3772 * Allow unprivileged RT tasks to decrease priority:
3774 if (user && !capable(CAP_SYS_NICE)) {
3775 if (rt_policy(policy)) {
3776 unsigned long rlim_rtprio =
3777 task_rlimit(p, RLIMIT_RTPRIO);
3779 /* can't set/change the rt policy */
3780 if (policy != p->policy && !rlim_rtprio)
3783 /* can't increase priority */
3784 if (param->sched_priority > p->rt_priority &&
3785 param->sched_priority > rlim_rtprio)
3790 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3791 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3793 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3794 if (!can_nice(p, TASK_NICE(p)))
3798 /* can't change other user's priorities */
3799 if (!check_same_owner(p))
3802 /* Normal users shall not reset the sched_reset_on_fork flag */
3803 if (p->sched_reset_on_fork && !reset_on_fork)
3808 retval = security_task_setscheduler(p);
3814 * make sure no PI-waiters arrive (or leave) while we are
3815 * changing the priority of the task:
3817 * To be able to change p->policy safely, the appropriate
3818 * runqueue lock must be held.
3820 rq = task_rq_lock(p, &flags);
3823 * Changing the policy of the stop threads its a very bad idea
3825 if (p == rq->stop) {
3826 task_rq_unlock(rq, p, &flags);
3831 * If not changing anything there's no need to proceed further:
3833 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3834 param->sched_priority == p->rt_priority))) {
3835 task_rq_unlock(rq, p, &flags);
3839 #ifdef CONFIG_RT_GROUP_SCHED
3842 * Do not allow realtime tasks into groups that have no runtime
3845 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3846 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3847 !task_group_is_autogroup(task_group(p))) {
3848 task_rq_unlock(rq, p, &flags);
3854 /* recheck policy now with rq lock held */
3855 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3856 policy = oldpolicy = -1;
3857 task_rq_unlock(rq, p, &flags);
3861 running = task_current(rq, p);
3863 dequeue_task(rq, p, 0);
3865 p->sched_class->put_prev_task(rq, p);
3867 p->sched_reset_on_fork = reset_on_fork;
3870 prev_class = p->sched_class;
3871 __setscheduler(rq, p, policy, param->sched_priority);
3874 p->sched_class->set_curr_task(rq);
3876 enqueue_task(rq, p, 0);
3878 check_class_changed(rq, p, prev_class, oldprio);
3879 task_rq_unlock(rq, p, &flags);
3881 rt_mutex_adjust_pi(p);
3887 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3888 * @p: the task in question.
3889 * @policy: new policy.
3890 * @param: structure containing the new RT priority.
3892 * NOTE that the task may be already dead.
3894 int sched_setscheduler(struct task_struct *p, int policy,
3895 const struct sched_param *param)
3897 return __sched_setscheduler(p, policy, param, true);
3899 EXPORT_SYMBOL_GPL(sched_setscheduler);
3902 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3903 * @p: the task in question.
3904 * @policy: new policy.
3905 * @param: structure containing the new RT priority.
3907 * Just like sched_setscheduler, only don't bother checking if the
3908 * current context has permission. For example, this is needed in
3909 * stop_machine(): we create temporary high priority worker threads,
3910 * but our caller might not have that capability.
3912 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3913 const struct sched_param *param)
3915 return __sched_setscheduler(p, policy, param, false);
3919 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3921 struct sched_param lparam;
3922 struct task_struct *p;
3925 if (!param || pid < 0)
3927 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3932 p = find_process_by_pid(pid);
3934 retval = sched_setscheduler(p, policy, &lparam);
3941 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3942 * @pid: the pid in question.
3943 * @policy: new policy.
3944 * @param: structure containing the new RT priority.
3946 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3947 struct sched_param __user *, param)
3949 /* negative values for policy are not valid */
3953 return do_sched_setscheduler(pid, policy, param);
3957 * sys_sched_setparam - set/change the RT priority of a thread
3958 * @pid: the pid in question.
3959 * @param: structure containing the new RT priority.
3961 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3963 return do_sched_setscheduler(pid, -1, param);
3967 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3968 * @pid: the pid in question.
3970 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3972 struct task_struct *p;
3980 p = find_process_by_pid(pid);
3982 retval = security_task_getscheduler(p);
3985 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3992 * sys_sched_getparam - get the RT priority of a thread
3993 * @pid: the pid in question.
3994 * @param: structure containing the RT priority.
3996 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3998 struct sched_param lp;
3999 struct task_struct *p;
4002 if (!param || pid < 0)
4006 p = find_process_by_pid(pid);
4011 retval = security_task_getscheduler(p);
4015 lp.sched_priority = p->rt_priority;
4019 * This one might sleep, we cannot do it with a spinlock held ...
4021 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4030 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4032 cpumask_var_t cpus_allowed, new_mask;
4033 struct task_struct *p;
4039 p = find_process_by_pid(pid);
4046 /* Prevent p going away */
4050 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4054 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4056 goto out_free_cpus_allowed;
4059 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4062 retval = security_task_setscheduler(p);
4066 cpuset_cpus_allowed(p, cpus_allowed);
4067 cpumask_and(new_mask, in_mask, cpus_allowed);
4069 retval = set_cpus_allowed_ptr(p, new_mask);
4072 cpuset_cpus_allowed(p, cpus_allowed);
4073 if (!cpumask_subset(new_mask, cpus_allowed)) {
4075 * We must have raced with a concurrent cpuset
4076 * update. Just reset the cpus_allowed to the
4077 * cpuset's cpus_allowed
4079 cpumask_copy(new_mask, cpus_allowed);
4084 free_cpumask_var(new_mask);
4085 out_free_cpus_allowed:
4086 free_cpumask_var(cpus_allowed);
4093 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4094 struct cpumask *new_mask)
4096 if (len < cpumask_size())
4097 cpumask_clear(new_mask);
4098 else if (len > cpumask_size())
4099 len = cpumask_size();
4101 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4105 * sys_sched_setaffinity - set the cpu affinity of a process
4106 * @pid: pid of the process
4107 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4108 * @user_mask_ptr: user-space pointer to the new cpu mask
4110 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4111 unsigned long __user *, user_mask_ptr)
4113 cpumask_var_t new_mask;
4116 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4119 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4121 retval = sched_setaffinity(pid, new_mask);
4122 free_cpumask_var(new_mask);
4126 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4128 struct task_struct *p;
4129 unsigned long flags;
4136 p = find_process_by_pid(pid);
4140 retval = security_task_getscheduler(p);
4144 raw_spin_lock_irqsave(&p->pi_lock, flags);
4145 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4146 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4156 * sys_sched_getaffinity - get the cpu affinity of a process
4157 * @pid: pid of the process
4158 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4159 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4161 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4162 unsigned long __user *, user_mask_ptr)
4167 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4169 if (len & (sizeof(unsigned long)-1))
4172 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4175 ret = sched_getaffinity(pid, mask);
4177 size_t retlen = min_t(size_t, len, cpumask_size());
4179 if (copy_to_user(user_mask_ptr, mask, retlen))
4184 free_cpumask_var(mask);
4190 * sys_sched_yield - yield the current processor to other threads.
4192 * This function yields the current CPU to other tasks. If there are no
4193 * other threads running on this CPU then this function will return.
4195 SYSCALL_DEFINE0(sched_yield)
4197 struct rq *rq = this_rq_lock();
4199 schedstat_inc(rq, yld_count);
4200 current->sched_class->yield_task(rq);
4203 * Since we are going to call schedule() anyway, there's
4204 * no need to preempt or enable interrupts:
4206 __release(rq->lock);
4207 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4208 do_raw_spin_unlock(&rq->lock);
4209 sched_preempt_enable_no_resched();
4216 static inline int should_resched(void)
4218 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4221 static void __cond_resched(void)
4223 add_preempt_count(PREEMPT_ACTIVE);
4225 sub_preempt_count(PREEMPT_ACTIVE);
4228 int __sched _cond_resched(void)
4230 if (should_resched()) {
4236 EXPORT_SYMBOL(_cond_resched);
4239 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4240 * call schedule, and on return reacquire the lock.
4242 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4243 * operations here to prevent schedule() from being called twice (once via
4244 * spin_unlock(), once by hand).
4246 int __cond_resched_lock(spinlock_t *lock)
4248 int resched = should_resched();
4251 lockdep_assert_held(lock);
4253 if (spin_needbreak(lock) || resched) {
4264 EXPORT_SYMBOL(__cond_resched_lock);
4266 int __sched __cond_resched_softirq(void)
4268 BUG_ON(!in_softirq());
4270 if (should_resched()) {
4278 EXPORT_SYMBOL(__cond_resched_softirq);
4281 * yield - yield the current processor to other threads.
4283 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4285 * The scheduler is at all times free to pick the calling task as the most
4286 * eligible task to run, if removing the yield() call from your code breaks
4287 * it, its already broken.
4289 * Typical broken usage is:
4294 * where one assumes that yield() will let 'the other' process run that will
4295 * make event true. If the current task is a SCHED_FIFO task that will never
4296 * happen. Never use yield() as a progress guarantee!!
4298 * If you want to use yield() to wait for something, use wait_event().
4299 * If you want to use yield() to be 'nice' for others, use cond_resched().
4300 * If you still want to use yield(), do not!
4302 void __sched yield(void)
4304 set_current_state(TASK_RUNNING);
4307 EXPORT_SYMBOL(yield);
4310 * yield_to - yield the current processor to another thread in
4311 * your thread group, or accelerate that thread toward the
4312 * processor it's on.
4314 * @preempt: whether task preemption is allowed or not
4316 * It's the caller's job to ensure that the target task struct
4317 * can't go away on us before we can do any checks.
4319 * Returns true if we indeed boosted the target task.
4321 bool __sched yield_to(struct task_struct *p, bool preempt)
4323 struct task_struct *curr = current;
4324 struct rq *rq, *p_rq;
4325 unsigned long flags;
4328 local_irq_save(flags);
4333 double_rq_lock(rq, p_rq);
4334 while (task_rq(p) != p_rq) {
4335 double_rq_unlock(rq, p_rq);
4339 if (!curr->sched_class->yield_to_task)
4342 if (curr->sched_class != p->sched_class)
4345 if (task_running(p_rq, p) || p->state)
4348 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4350 schedstat_inc(rq, yld_count);
4352 * Make p's CPU reschedule; pick_next_entity takes care of
4355 if (preempt && rq != p_rq)
4356 resched_task(p_rq->curr);
4360 double_rq_unlock(rq, p_rq);
4361 local_irq_restore(flags);
4368 EXPORT_SYMBOL_GPL(yield_to);
4371 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4372 * that process accounting knows that this is a task in IO wait state.
4374 void __sched io_schedule(void)
4376 struct rq *rq = raw_rq();
4378 delayacct_blkio_start();
4379 atomic_inc(&rq->nr_iowait);
4380 blk_flush_plug(current);
4381 current->in_iowait = 1;
4383 current->in_iowait = 0;
4384 atomic_dec(&rq->nr_iowait);
4385 delayacct_blkio_end();
4387 EXPORT_SYMBOL(io_schedule);
4389 long __sched io_schedule_timeout(long timeout)
4391 struct rq *rq = raw_rq();
4394 delayacct_blkio_start();
4395 atomic_inc(&rq->nr_iowait);
4396 blk_flush_plug(current);
4397 current->in_iowait = 1;
4398 ret = schedule_timeout(timeout);
4399 current->in_iowait = 0;
4400 atomic_dec(&rq->nr_iowait);
4401 delayacct_blkio_end();
4406 * sys_sched_get_priority_max - return maximum RT priority.
4407 * @policy: scheduling class.
4409 * this syscall returns the maximum rt_priority that can be used
4410 * by a given scheduling class.
4412 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4419 ret = MAX_USER_RT_PRIO-1;
4431 * sys_sched_get_priority_min - return minimum RT priority.
4432 * @policy: scheduling class.
4434 * this syscall returns the minimum rt_priority that can be used
4435 * by a given scheduling class.
4437 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4455 * sys_sched_rr_get_interval - return the default timeslice of a process.
4456 * @pid: pid of the process.
4457 * @interval: userspace pointer to the timeslice value.
4459 * this syscall writes the default timeslice value of a given process
4460 * into the user-space timespec buffer. A value of '0' means infinity.
4462 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4463 struct timespec __user *, interval)
4465 struct task_struct *p;
4466 unsigned int time_slice;
4467 unsigned long flags;
4477 p = find_process_by_pid(pid);
4481 retval = security_task_getscheduler(p);
4485 rq = task_rq_lock(p, &flags);
4486 time_slice = p->sched_class->get_rr_interval(rq, p);
4487 task_rq_unlock(rq, p, &flags);
4490 jiffies_to_timespec(time_slice, &t);
4491 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4499 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4501 void sched_show_task(struct task_struct *p)
4503 unsigned long free = 0;
4507 state = p->state ? __ffs(p->state) + 1 : 0;
4508 printk(KERN_INFO "%-15.15s %c", p->comm,
4509 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4510 #if BITS_PER_LONG == 32
4511 if (state == TASK_RUNNING)
4512 printk(KERN_CONT " running ");
4514 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4516 if (state == TASK_RUNNING)
4517 printk(KERN_CONT " running task ");
4519 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4521 #ifdef CONFIG_DEBUG_STACK_USAGE
4522 free = stack_not_used(p);
4525 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4527 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4528 task_pid_nr(p), ppid,
4529 (unsigned long)task_thread_info(p)->flags);
4531 show_stack(p, NULL);
4534 void show_state_filter(unsigned long state_filter)
4536 struct task_struct *g, *p;
4538 #if BITS_PER_LONG == 32
4540 " task PC stack pid father\n");
4543 " task PC stack pid father\n");
4546 do_each_thread(g, p) {
4548 * reset the NMI-timeout, listing all files on a slow
4549 * console might take a lot of time:
4551 touch_nmi_watchdog();
4552 if (!state_filter || (p->state & state_filter))
4554 } while_each_thread(g, p);
4556 touch_all_softlockup_watchdogs();
4558 #ifdef CONFIG_SCHED_DEBUG
4559 sysrq_sched_debug_show();
4563 * Only show locks if all tasks are dumped:
4566 debug_show_all_locks();
4569 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4571 idle->sched_class = &idle_sched_class;
4575 * init_idle - set up an idle thread for a given CPU
4576 * @idle: task in question
4577 * @cpu: cpu the idle task belongs to
4579 * NOTE: this function does not set the idle thread's NEED_RESCHED
4580 * flag, to make booting more robust.
4582 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4584 struct rq *rq = cpu_rq(cpu);
4585 unsigned long flags;
4587 raw_spin_lock_irqsave(&rq->lock, flags);
4590 idle->state = TASK_RUNNING;
4591 idle->se.exec_start = sched_clock();
4593 do_set_cpus_allowed(idle, cpumask_of(cpu));
4595 * We're having a chicken and egg problem, even though we are
4596 * holding rq->lock, the cpu isn't yet set to this cpu so the
4597 * lockdep check in task_group() will fail.
4599 * Similar case to sched_fork(). / Alternatively we could
4600 * use task_rq_lock() here and obtain the other rq->lock.
4605 __set_task_cpu(idle, cpu);
4608 rq->curr = rq->idle = idle;
4609 #if defined(CONFIG_SMP)
4612 raw_spin_unlock_irqrestore(&rq->lock, flags);
4614 /* Set the preempt count _outside_ the spinlocks! */
4615 task_thread_info(idle)->preempt_count = 0;
4618 * The idle tasks have their own, simple scheduling class:
4620 idle->sched_class = &idle_sched_class;
4621 ftrace_graph_init_idle_task(idle, cpu);
4622 #if defined(CONFIG_SMP)
4623 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4628 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4630 if (p->sched_class && p->sched_class->set_cpus_allowed)
4631 p->sched_class->set_cpus_allowed(p, new_mask);
4633 cpumask_copy(&p->cpus_allowed, new_mask);
4634 p->nr_cpus_allowed = cpumask_weight(new_mask);
4638 * This is how migration works:
4640 * 1) we invoke migration_cpu_stop() on the target CPU using
4642 * 2) stopper starts to run (implicitly forcing the migrated thread
4644 * 3) it checks whether the migrated task is still in the wrong runqueue.
4645 * 4) if it's in the wrong runqueue then the migration thread removes
4646 * it and puts it into the right queue.
4647 * 5) stopper completes and stop_one_cpu() returns and the migration
4652 * Change a given task's CPU affinity. Migrate the thread to a
4653 * proper CPU and schedule it away if the CPU it's executing on
4654 * is removed from the allowed bitmask.
4656 * NOTE: the caller must have a valid reference to the task, the
4657 * task must not exit() & deallocate itself prematurely. The
4658 * call is not atomic; no spinlocks may be held.
4660 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4662 unsigned long flags;
4664 unsigned int dest_cpu;
4667 rq = task_rq_lock(p, &flags);
4669 if (cpumask_equal(&p->cpus_allowed, new_mask))
4672 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4677 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4682 do_set_cpus_allowed(p, new_mask);
4684 /* Can the task run on the task's current CPU? If so, we're done */
4685 if (cpumask_test_cpu(task_cpu(p), new_mask))
4688 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4690 struct migration_arg arg = { p, dest_cpu };
4691 /* Need help from migration thread: drop lock and wait. */
4692 task_rq_unlock(rq, p, &flags);
4693 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4694 tlb_migrate_finish(p->mm);
4698 task_rq_unlock(rq, p, &flags);
4702 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4705 * Move (not current) task off this cpu, onto dest cpu. We're doing
4706 * this because either it can't run here any more (set_cpus_allowed()
4707 * away from this CPU, or CPU going down), or because we're
4708 * attempting to rebalance this task on exec (sched_exec).
4710 * So we race with normal scheduler movements, but that's OK, as long
4711 * as the task is no longer on this CPU.
4713 * Returns non-zero if task was successfully migrated.
4715 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4717 struct rq *rq_dest, *rq_src;
4720 if (unlikely(!cpu_active(dest_cpu)))
4723 rq_src = cpu_rq(src_cpu);
4724 rq_dest = cpu_rq(dest_cpu);
4726 raw_spin_lock(&p->pi_lock);
4727 double_rq_lock(rq_src, rq_dest);
4728 /* Already moved. */
4729 if (task_cpu(p) != src_cpu)
4731 /* Affinity changed (again). */
4732 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4736 * If we're not on a rq, the next wake-up will ensure we're
4740 dequeue_task(rq_src, p, 0);
4741 set_task_cpu(p, dest_cpu);
4742 enqueue_task(rq_dest, p, 0);
4743 check_preempt_curr(rq_dest, p, 0);
4748 double_rq_unlock(rq_src, rq_dest);
4749 raw_spin_unlock(&p->pi_lock);
4754 * migration_cpu_stop - this will be executed by a highprio stopper thread
4755 * and performs thread migration by bumping thread off CPU then
4756 * 'pushing' onto another runqueue.
4758 static int migration_cpu_stop(void *data)
4760 struct migration_arg *arg = data;
4763 * The original target cpu might have gone down and we might
4764 * be on another cpu but it doesn't matter.
4766 local_irq_disable();
4767 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4772 #ifdef CONFIG_HOTPLUG_CPU
4775 * Ensures that the idle task is using init_mm right before its cpu goes
4778 void idle_task_exit(void)
4780 struct mm_struct *mm = current->active_mm;
4782 BUG_ON(cpu_online(smp_processor_id()));
4785 switch_mm(mm, &init_mm, current);
4790 * Since this CPU is going 'away' for a while, fold any nr_active delta
4791 * we might have. Assumes we're called after migrate_tasks() so that the
4792 * nr_active count is stable.
4794 * Also see the comment "Global load-average calculations".
4796 static void calc_load_migrate(struct rq *rq)
4798 long delta = calc_load_fold_active(rq);
4800 atomic_long_add(delta, &calc_load_tasks);
4804 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4805 * try_to_wake_up()->select_task_rq().
4807 * Called with rq->lock held even though we'er in stop_machine() and
4808 * there's no concurrency possible, we hold the required locks anyway
4809 * because of lock validation efforts.
4811 static void migrate_tasks(unsigned int dead_cpu)
4813 struct rq *rq = cpu_rq(dead_cpu);
4814 struct task_struct *next, *stop = rq->stop;
4818 * Fudge the rq selection such that the below task selection loop
4819 * doesn't get stuck on the currently eligible stop task.
4821 * We're currently inside stop_machine() and the rq is either stuck
4822 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4823 * either way we should never end up calling schedule() until we're
4830 * There's this thread running, bail when that's the only
4833 if (rq->nr_running == 1)
4836 next = pick_next_task(rq);
4838 next->sched_class->put_prev_task(rq, next);
4840 /* Find suitable destination for @next, with force if needed. */
4841 dest_cpu = select_fallback_rq(dead_cpu, next);
4842 raw_spin_unlock(&rq->lock);
4844 __migrate_task(next, dead_cpu, dest_cpu);
4846 raw_spin_lock(&rq->lock);
4852 #endif /* CONFIG_HOTPLUG_CPU */
4854 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4856 static struct ctl_table sd_ctl_dir[] = {
4858 .procname = "sched_domain",
4864 static struct ctl_table sd_ctl_root[] = {
4866 .procname = "kernel",
4868 .child = sd_ctl_dir,
4873 static struct ctl_table *sd_alloc_ctl_entry(int n)
4875 struct ctl_table *entry =
4876 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4881 static void sd_free_ctl_entry(struct ctl_table **tablep)
4883 struct ctl_table *entry;
4886 * In the intermediate directories, both the child directory and
4887 * procname are dynamically allocated and could fail but the mode
4888 * will always be set. In the lowest directory the names are
4889 * static strings and all have proc handlers.
4891 for (entry = *tablep; entry->mode; entry++) {
4893 sd_free_ctl_entry(&entry->child);
4894 if (entry->proc_handler == NULL)
4895 kfree(entry->procname);
4902 static int min_load_idx = 0;
4903 static int max_load_idx = CPU_LOAD_IDX_MAX;
4906 set_table_entry(struct ctl_table *entry,
4907 const char *procname, void *data, int maxlen,
4908 umode_t mode, proc_handler *proc_handler,
4911 entry->procname = procname;
4913 entry->maxlen = maxlen;
4915 entry->proc_handler = proc_handler;
4918 entry->extra1 = &min_load_idx;
4919 entry->extra2 = &max_load_idx;
4923 static struct ctl_table *
4924 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4926 struct ctl_table *table = sd_alloc_ctl_entry(13);
4931 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4932 sizeof(long), 0644, proc_doulongvec_minmax, false);
4933 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4934 sizeof(long), 0644, proc_doulongvec_minmax, false);
4935 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4936 sizeof(int), 0644, proc_dointvec_minmax, true);
4937 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4938 sizeof(int), 0644, proc_dointvec_minmax, true);
4939 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4940 sizeof(int), 0644, proc_dointvec_minmax, true);
4941 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4942 sizeof(int), 0644, proc_dointvec_minmax, true);
4943 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4944 sizeof(int), 0644, proc_dointvec_minmax, true);
4945 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4946 sizeof(int), 0644, proc_dointvec_minmax, false);
4947 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4948 sizeof(int), 0644, proc_dointvec_minmax, false);
4949 set_table_entry(&table[9], "cache_nice_tries",
4950 &sd->cache_nice_tries,
4951 sizeof(int), 0644, proc_dointvec_minmax, false);
4952 set_table_entry(&table[10], "flags", &sd->flags,
4953 sizeof(int), 0644, proc_dointvec_minmax, false);
4954 set_table_entry(&table[11], "name", sd->name,
4955 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4956 /* &table[12] is terminator */
4961 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4963 struct ctl_table *entry, *table;
4964 struct sched_domain *sd;
4965 int domain_num = 0, i;
4968 for_each_domain(cpu, sd)
4970 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4975 for_each_domain(cpu, sd) {
4976 snprintf(buf, 32, "domain%d", i);
4977 entry->procname = kstrdup(buf, GFP_KERNEL);
4979 entry->child = sd_alloc_ctl_domain_table(sd);
4986 static struct ctl_table_header *sd_sysctl_header;
4987 static void register_sched_domain_sysctl(void)
4989 int i, cpu_num = num_possible_cpus();
4990 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4993 WARN_ON(sd_ctl_dir[0].child);
4994 sd_ctl_dir[0].child = entry;
4999 for_each_possible_cpu(i) {
5000 snprintf(buf, 32, "cpu%d", i);
5001 entry->procname = kstrdup(buf, GFP_KERNEL);
5003 entry->child = sd_alloc_ctl_cpu_table(i);
5007 WARN_ON(sd_sysctl_header);
5008 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5011 /* may be called multiple times per register */
5012 static void unregister_sched_domain_sysctl(void)
5014 if (sd_sysctl_header)
5015 unregister_sysctl_table(sd_sysctl_header);
5016 sd_sysctl_header = NULL;
5017 if (sd_ctl_dir[0].child)
5018 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5021 static void register_sched_domain_sysctl(void)
5024 static void unregister_sched_domain_sysctl(void)
5029 static void set_rq_online(struct rq *rq)
5032 const struct sched_class *class;
5034 cpumask_set_cpu(rq->cpu, rq->rd->online);
5037 for_each_class(class) {
5038 if (class->rq_online)
5039 class->rq_online(rq);
5044 static void set_rq_offline(struct rq *rq)
5047 const struct sched_class *class;
5049 for_each_class(class) {
5050 if (class->rq_offline)
5051 class->rq_offline(rq);
5054 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5060 * migration_call - callback that gets triggered when a CPU is added.
5061 * Here we can start up the necessary migration thread for the new CPU.
5063 static int __cpuinit
5064 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5066 int cpu = (long)hcpu;
5067 unsigned long flags;
5068 struct rq *rq = cpu_rq(cpu);
5070 switch (action & ~CPU_TASKS_FROZEN) {
5072 case CPU_UP_PREPARE:
5073 rq->calc_load_update = calc_load_update;
5077 /* Update our root-domain */
5078 raw_spin_lock_irqsave(&rq->lock, flags);
5080 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5084 raw_spin_unlock_irqrestore(&rq->lock, flags);
5087 #ifdef CONFIG_HOTPLUG_CPU
5089 sched_ttwu_pending();
5090 /* Update our root-domain */
5091 raw_spin_lock_irqsave(&rq->lock, flags);
5093 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5097 BUG_ON(rq->nr_running != 1); /* the migration thread */
5098 raw_spin_unlock_irqrestore(&rq->lock, flags);
5102 calc_load_migrate(rq);
5107 update_max_interval();
5113 * Register at high priority so that task migration (migrate_all_tasks)
5114 * happens before everything else. This has to be lower priority than
5115 * the notifier in the perf_event subsystem, though.
5117 static struct notifier_block __cpuinitdata migration_notifier = {
5118 .notifier_call = migration_call,
5119 .priority = CPU_PRI_MIGRATION,
5122 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5123 unsigned long action, void *hcpu)
5125 switch (action & ~CPU_TASKS_FROZEN) {
5127 case CPU_DOWN_FAILED:
5128 set_cpu_active((long)hcpu, true);
5135 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5136 unsigned long action, void *hcpu)
5138 switch (action & ~CPU_TASKS_FROZEN) {
5139 case CPU_DOWN_PREPARE:
5140 set_cpu_active((long)hcpu, false);
5147 static int __init migration_init(void)
5149 void *cpu = (void *)(long)smp_processor_id();
5152 /* Initialize migration for the boot CPU */
5153 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5154 BUG_ON(err == NOTIFY_BAD);
5155 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5156 register_cpu_notifier(&migration_notifier);
5158 /* Register cpu active notifiers */
5159 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5160 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5164 early_initcall(migration_init);
5169 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5171 #ifdef CONFIG_SCHED_DEBUG
5173 static __read_mostly int sched_debug_enabled;
5175 static int __init sched_debug_setup(char *str)
5177 sched_debug_enabled = 1;
5181 early_param("sched_debug", sched_debug_setup);
5183 static inline bool sched_debug(void)
5185 return sched_debug_enabled;
5188 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5189 struct cpumask *groupmask)
5191 struct sched_group *group = sd->groups;
5194 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5195 cpumask_clear(groupmask);
5197 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5199 if (!(sd->flags & SD_LOAD_BALANCE)) {
5200 printk("does not load-balance\n");
5202 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5207 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5209 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5210 printk(KERN_ERR "ERROR: domain->span does not contain "
5213 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5214 printk(KERN_ERR "ERROR: domain->groups does not contain"
5218 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5222 printk(KERN_ERR "ERROR: group is NULL\n");
5227 * Even though we initialize ->power to something semi-sane,
5228 * we leave power_orig unset. This allows us to detect if
5229 * domain iteration is still funny without causing /0 traps.
5231 if (!group->sgp->power_orig) {
5232 printk(KERN_CONT "\n");
5233 printk(KERN_ERR "ERROR: domain->cpu_power not "
5238 if (!cpumask_weight(sched_group_cpus(group))) {
5239 printk(KERN_CONT "\n");
5240 printk(KERN_ERR "ERROR: empty group\n");
5244 if (!(sd->flags & SD_OVERLAP) &&
5245 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5246 printk(KERN_CONT "\n");
5247 printk(KERN_ERR "ERROR: repeated CPUs\n");
5251 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5253 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5255 printk(KERN_CONT " %s", str);
5256 if (group->sgp->power != SCHED_POWER_SCALE) {
5257 printk(KERN_CONT " (cpu_power = %d)",
5261 group = group->next;
5262 } while (group != sd->groups);
5263 printk(KERN_CONT "\n");
5265 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5266 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5269 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5270 printk(KERN_ERR "ERROR: parent span is not a superset "
5271 "of domain->span\n");
5275 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5279 if (!sched_debug_enabled)
5283 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5287 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5290 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5298 #else /* !CONFIG_SCHED_DEBUG */
5299 # define sched_domain_debug(sd, cpu) do { } while (0)
5300 static inline bool sched_debug(void)
5304 #endif /* CONFIG_SCHED_DEBUG */
5306 static int sd_degenerate(struct sched_domain *sd)
5308 if (cpumask_weight(sched_domain_span(sd)) == 1)
5311 /* Following flags need at least 2 groups */
5312 if (sd->flags & (SD_LOAD_BALANCE |
5313 SD_BALANCE_NEWIDLE |
5317 SD_SHARE_PKG_RESOURCES)) {
5318 if (sd->groups != sd->groups->next)
5322 /* Following flags don't use groups */
5323 if (sd->flags & (SD_WAKE_AFFINE))
5330 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5332 unsigned long cflags = sd->flags, pflags = parent->flags;
5334 if (sd_degenerate(parent))
5337 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5340 /* Flags needing groups don't count if only 1 group in parent */
5341 if (parent->groups == parent->groups->next) {
5342 pflags &= ~(SD_LOAD_BALANCE |
5343 SD_BALANCE_NEWIDLE |
5347 SD_SHARE_PKG_RESOURCES);
5348 if (nr_node_ids == 1)
5349 pflags &= ~SD_SERIALIZE;
5351 if (~cflags & pflags)
5357 static void free_rootdomain(struct rcu_head *rcu)
5359 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5361 cpupri_cleanup(&rd->cpupri);
5362 free_cpumask_var(rd->rto_mask);
5363 free_cpumask_var(rd->online);
5364 free_cpumask_var(rd->span);
5368 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5370 struct root_domain *old_rd = NULL;
5371 unsigned long flags;
5373 raw_spin_lock_irqsave(&rq->lock, flags);
5378 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5381 cpumask_clear_cpu(rq->cpu, old_rd->span);
5384 * If we dont want to free the old_rt yet then
5385 * set old_rd to NULL to skip the freeing later
5388 if (!atomic_dec_and_test(&old_rd->refcount))
5392 atomic_inc(&rd->refcount);
5395 cpumask_set_cpu(rq->cpu, rd->span);
5396 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5399 raw_spin_unlock_irqrestore(&rq->lock, flags);
5402 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5405 static int init_rootdomain(struct root_domain *rd)
5407 memset(rd, 0, sizeof(*rd));
5409 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5411 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5413 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5416 if (cpupri_init(&rd->cpupri) != 0)
5421 free_cpumask_var(rd->rto_mask);
5423 free_cpumask_var(rd->online);
5425 free_cpumask_var(rd->span);
5431 * By default the system creates a single root-domain with all cpus as
5432 * members (mimicking the global state we have today).
5434 struct root_domain def_root_domain;
5436 static void init_defrootdomain(void)
5438 init_rootdomain(&def_root_domain);
5440 atomic_set(&def_root_domain.refcount, 1);
5443 static struct root_domain *alloc_rootdomain(void)
5445 struct root_domain *rd;
5447 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5451 if (init_rootdomain(rd) != 0) {
5459 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5461 struct sched_group *tmp, *first;
5470 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5475 } while (sg != first);
5478 static void free_sched_domain(struct rcu_head *rcu)
5480 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5483 * If its an overlapping domain it has private groups, iterate and
5486 if (sd->flags & SD_OVERLAP) {
5487 free_sched_groups(sd->groups, 1);
5488 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5489 kfree(sd->groups->sgp);
5495 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5497 call_rcu(&sd->rcu, free_sched_domain);
5500 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5502 for (; sd; sd = sd->parent)
5503 destroy_sched_domain(sd, cpu);
5507 * Keep a special pointer to the highest sched_domain that has
5508 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5509 * allows us to avoid some pointer chasing select_idle_sibling().
5511 * Also keep a unique ID per domain (we use the first cpu number in
5512 * the cpumask of the domain), this allows us to quickly tell if
5513 * two cpus are in the same cache domain, see cpus_share_cache().
5515 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5516 DEFINE_PER_CPU(int, sd_llc_id);
5518 static void update_top_cache_domain(int cpu)
5520 struct sched_domain *sd;
5523 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5525 id = cpumask_first(sched_domain_span(sd));
5527 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5528 per_cpu(sd_llc_id, cpu) = id;
5532 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5533 * hold the hotplug lock.
5536 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5538 struct rq *rq = cpu_rq(cpu);
5539 struct sched_domain *tmp;
5541 /* Remove the sched domains which do not contribute to scheduling. */
5542 for (tmp = sd; tmp; ) {
5543 struct sched_domain *parent = tmp->parent;
5547 if (sd_parent_degenerate(tmp, parent)) {
5548 tmp->parent = parent->parent;
5550 parent->parent->child = tmp;
5551 destroy_sched_domain(parent, cpu);
5556 if (sd && sd_degenerate(sd)) {
5559 destroy_sched_domain(tmp, cpu);
5564 sched_domain_debug(sd, cpu);
5566 rq_attach_root(rq, rd);
5568 rcu_assign_pointer(rq->sd, sd);
5569 destroy_sched_domains(tmp, cpu);
5571 update_top_cache_domain(cpu);
5574 /* cpus with isolated domains */
5575 static cpumask_var_t cpu_isolated_map;
5577 /* Setup the mask of cpus configured for isolated domains */
5578 static int __init isolated_cpu_setup(char *str)
5580 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5581 cpulist_parse(str, cpu_isolated_map);
5585 __setup("isolcpus=", isolated_cpu_setup);
5587 static const struct cpumask *cpu_cpu_mask(int cpu)
5589 return cpumask_of_node(cpu_to_node(cpu));
5593 struct sched_domain **__percpu sd;
5594 struct sched_group **__percpu sg;
5595 struct sched_group_power **__percpu sgp;
5599 struct sched_domain ** __percpu sd;
5600 struct root_domain *rd;
5610 struct sched_domain_topology_level;
5612 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5613 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5615 #define SDTL_OVERLAP 0x01
5617 struct sched_domain_topology_level {
5618 sched_domain_init_f init;
5619 sched_domain_mask_f mask;
5622 struct sd_data data;
5626 * Build an iteration mask that can exclude certain CPUs from the upwards
5629 * Asymmetric node setups can result in situations where the domain tree is of
5630 * unequal depth, make sure to skip domains that already cover the entire
5633 * In that case build_sched_domains() will have terminated the iteration early
5634 * and our sibling sd spans will be empty. Domains should always include the
5635 * cpu they're built on, so check that.
5638 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5640 const struct cpumask *span = sched_domain_span(sd);
5641 struct sd_data *sdd = sd->private;
5642 struct sched_domain *sibling;
5645 for_each_cpu(i, span) {
5646 sibling = *per_cpu_ptr(sdd->sd, i);
5647 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5650 cpumask_set_cpu(i, sched_group_mask(sg));
5655 * Return the canonical balance cpu for this group, this is the first cpu
5656 * of this group that's also in the iteration mask.
5658 int group_balance_cpu(struct sched_group *sg)
5660 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5664 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5666 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5667 const struct cpumask *span = sched_domain_span(sd);
5668 struct cpumask *covered = sched_domains_tmpmask;
5669 struct sd_data *sdd = sd->private;
5670 struct sched_domain *child;
5673 cpumask_clear(covered);
5675 for_each_cpu(i, span) {
5676 struct cpumask *sg_span;
5678 if (cpumask_test_cpu(i, covered))
5681 child = *per_cpu_ptr(sdd->sd, i);
5683 /* See the comment near build_group_mask(). */
5684 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5687 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5688 GFP_KERNEL, cpu_to_node(cpu));
5693 sg_span = sched_group_cpus(sg);
5695 child = child->child;
5696 cpumask_copy(sg_span, sched_domain_span(child));
5698 cpumask_set_cpu(i, sg_span);
5700 cpumask_or(covered, covered, sg_span);
5702 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5703 if (atomic_inc_return(&sg->sgp->ref) == 1)
5704 build_group_mask(sd, sg);
5707 * Initialize sgp->power such that even if we mess up the
5708 * domains and no possible iteration will get us here, we won't
5711 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5714 * Make sure the first group of this domain contains the
5715 * canonical balance cpu. Otherwise the sched_domain iteration
5716 * breaks. See update_sg_lb_stats().
5718 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5719 group_balance_cpu(sg) == cpu)
5729 sd->groups = groups;
5734 free_sched_groups(first, 0);
5739 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5741 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5742 struct sched_domain *child = sd->child;
5745 cpu = cpumask_first(sched_domain_span(child));
5748 *sg = *per_cpu_ptr(sdd->sg, cpu);
5749 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5750 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5757 * build_sched_groups will build a circular linked list of the groups
5758 * covered by the given span, and will set each group's ->cpumask correctly,
5759 * and ->cpu_power to 0.
5761 * Assumes the sched_domain tree is fully constructed
5764 build_sched_groups(struct sched_domain *sd, int cpu)
5766 struct sched_group *first = NULL, *last = NULL;
5767 struct sd_data *sdd = sd->private;
5768 const struct cpumask *span = sched_domain_span(sd);
5769 struct cpumask *covered;
5772 get_group(cpu, sdd, &sd->groups);
5773 atomic_inc(&sd->groups->ref);
5775 if (cpu != cpumask_first(sched_domain_span(sd)))
5778 lockdep_assert_held(&sched_domains_mutex);
5779 covered = sched_domains_tmpmask;
5781 cpumask_clear(covered);
5783 for_each_cpu(i, span) {
5784 struct sched_group *sg;
5785 int group = get_group(i, sdd, &sg);
5788 if (cpumask_test_cpu(i, covered))
5791 cpumask_clear(sched_group_cpus(sg));
5793 cpumask_setall(sched_group_mask(sg));
5795 for_each_cpu(j, span) {
5796 if (get_group(j, sdd, NULL) != group)
5799 cpumask_set_cpu(j, covered);
5800 cpumask_set_cpu(j, sched_group_cpus(sg));
5815 * Initialize sched groups cpu_power.
5817 * cpu_power indicates the capacity of sched group, which is used while
5818 * distributing the load between different sched groups in a sched domain.
5819 * Typically cpu_power for all the groups in a sched domain will be same unless
5820 * there are asymmetries in the topology. If there are asymmetries, group
5821 * having more cpu_power will pickup more load compared to the group having
5824 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5826 struct sched_group *sg = sd->groups;
5828 WARN_ON(!sd || !sg);
5831 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5833 } while (sg != sd->groups);
5835 if (cpu != group_balance_cpu(sg))
5838 update_group_power(sd, cpu);
5839 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5842 int __weak arch_sd_sibling_asym_packing(void)
5844 return 0*SD_ASYM_PACKING;
5848 * Initializers for schedule domains
5849 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5852 #ifdef CONFIG_SCHED_DEBUG
5853 # define SD_INIT_NAME(sd, type) sd->name = #type
5855 # define SD_INIT_NAME(sd, type) do { } while (0)
5858 #define SD_INIT_FUNC(type) \
5859 static noinline struct sched_domain * \
5860 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5862 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5863 *sd = SD_##type##_INIT; \
5864 SD_INIT_NAME(sd, type); \
5865 sd->private = &tl->data; \
5870 #ifdef CONFIG_SCHED_SMT
5871 SD_INIT_FUNC(SIBLING)
5873 #ifdef CONFIG_SCHED_MC
5876 #ifdef CONFIG_SCHED_BOOK
5880 static int default_relax_domain_level = -1;
5881 int sched_domain_level_max;
5883 static int __init setup_relax_domain_level(char *str)
5885 if (kstrtoint(str, 0, &default_relax_domain_level))
5886 pr_warn("Unable to set relax_domain_level\n");
5890 __setup("relax_domain_level=", setup_relax_domain_level);
5892 static void set_domain_attribute(struct sched_domain *sd,
5893 struct sched_domain_attr *attr)
5897 if (!attr || attr->relax_domain_level < 0) {
5898 if (default_relax_domain_level < 0)
5901 request = default_relax_domain_level;
5903 request = attr->relax_domain_level;
5904 if (request < sd->level) {
5905 /* turn off idle balance on this domain */
5906 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5908 /* turn on idle balance on this domain */
5909 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5913 static void __sdt_free(const struct cpumask *cpu_map);
5914 static int __sdt_alloc(const struct cpumask *cpu_map);
5916 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5917 const struct cpumask *cpu_map)
5921 if (!atomic_read(&d->rd->refcount))
5922 free_rootdomain(&d->rd->rcu); /* fall through */
5924 free_percpu(d->sd); /* fall through */
5926 __sdt_free(cpu_map); /* fall through */
5932 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5933 const struct cpumask *cpu_map)
5935 memset(d, 0, sizeof(*d));
5937 if (__sdt_alloc(cpu_map))
5938 return sa_sd_storage;
5939 d->sd = alloc_percpu(struct sched_domain *);
5941 return sa_sd_storage;
5942 d->rd = alloc_rootdomain();
5945 return sa_rootdomain;
5949 * NULL the sd_data elements we've used to build the sched_domain and
5950 * sched_group structure so that the subsequent __free_domain_allocs()
5951 * will not free the data we're using.
5953 static void claim_allocations(int cpu, struct sched_domain *sd)
5955 struct sd_data *sdd = sd->private;
5957 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5958 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5960 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5961 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5963 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5964 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5967 #ifdef CONFIG_SCHED_SMT
5968 static const struct cpumask *cpu_smt_mask(int cpu)
5970 return topology_thread_cpumask(cpu);
5975 * Topology list, bottom-up.
5977 static struct sched_domain_topology_level default_topology[] = {
5978 #ifdef CONFIG_SCHED_SMT
5979 { sd_init_SIBLING, cpu_smt_mask, },
5981 #ifdef CONFIG_SCHED_MC
5982 { sd_init_MC, cpu_coregroup_mask, },
5984 #ifdef CONFIG_SCHED_BOOK
5985 { sd_init_BOOK, cpu_book_mask, },
5987 { sd_init_CPU, cpu_cpu_mask, },
5991 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5995 static int sched_domains_numa_levels;
5996 static int *sched_domains_numa_distance;
5997 static struct cpumask ***sched_domains_numa_masks;
5998 static int sched_domains_curr_level;
6000 static inline int sd_local_flags(int level)
6002 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6005 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6008 static struct sched_domain *
6009 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6011 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6012 int level = tl->numa_level;
6013 int sd_weight = cpumask_weight(
6014 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6016 *sd = (struct sched_domain){
6017 .min_interval = sd_weight,
6018 .max_interval = 2*sd_weight,
6020 .imbalance_pct = 125,
6021 .cache_nice_tries = 2,
6028 .flags = 1*SD_LOAD_BALANCE
6029 | 1*SD_BALANCE_NEWIDLE
6034 | 0*SD_SHARE_CPUPOWER
6035 | 0*SD_SHARE_PKG_RESOURCES
6037 | 0*SD_PREFER_SIBLING
6038 | sd_local_flags(level)
6040 .last_balance = jiffies,
6041 .balance_interval = sd_weight,
6043 SD_INIT_NAME(sd, NUMA);
6044 sd->private = &tl->data;
6047 * Ugly hack to pass state to sd_numa_mask()...
6049 sched_domains_curr_level = tl->numa_level;
6054 static const struct cpumask *sd_numa_mask(int cpu)
6056 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6059 static void sched_numa_warn(const char *str)
6061 static int done = false;
6069 printk(KERN_WARNING "ERROR: %s\n\n", str);
6071 for (i = 0; i < nr_node_ids; i++) {
6072 printk(KERN_WARNING " ");
6073 for (j = 0; j < nr_node_ids; j++)
6074 printk(KERN_CONT "%02d ", node_distance(i,j));
6075 printk(KERN_CONT "\n");
6077 printk(KERN_WARNING "\n");
6080 static bool find_numa_distance(int distance)
6084 if (distance == node_distance(0, 0))
6087 for (i = 0; i < sched_domains_numa_levels; i++) {
6088 if (sched_domains_numa_distance[i] == distance)
6095 static void sched_init_numa(void)
6097 int next_distance, curr_distance = node_distance(0, 0);
6098 struct sched_domain_topology_level *tl;
6102 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6103 if (!sched_domains_numa_distance)
6107 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6108 * unique distances in the node_distance() table.
6110 * Assumes node_distance(0,j) includes all distances in
6111 * node_distance(i,j) in order to avoid cubic time.
6113 next_distance = curr_distance;
6114 for (i = 0; i < nr_node_ids; i++) {
6115 for (j = 0; j < nr_node_ids; j++) {
6116 for (k = 0; k < nr_node_ids; k++) {
6117 int distance = node_distance(i, k);
6119 if (distance > curr_distance &&
6120 (distance < next_distance ||
6121 next_distance == curr_distance))
6122 next_distance = distance;
6125 * While not a strong assumption it would be nice to know
6126 * about cases where if node A is connected to B, B is not
6127 * equally connected to A.
6129 if (sched_debug() && node_distance(k, i) != distance)
6130 sched_numa_warn("Node-distance not symmetric");
6132 if (sched_debug() && i && !find_numa_distance(distance))
6133 sched_numa_warn("Node-0 not representative");
6135 if (next_distance != curr_distance) {
6136 sched_domains_numa_distance[level++] = next_distance;
6137 sched_domains_numa_levels = level;
6138 curr_distance = next_distance;
6143 * In case of sched_debug() we verify the above assumption.
6149 * 'level' contains the number of unique distances, excluding the
6150 * identity distance node_distance(i,i).
6152 * The sched_domains_nume_distance[] array includes the actual distance
6157 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6158 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6159 * the array will contain less then 'level' members. This could be
6160 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6161 * in other functions.
6163 * We reset it to 'level' at the end of this function.
6165 sched_domains_numa_levels = 0;
6167 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6168 if (!sched_domains_numa_masks)
6172 * Now for each level, construct a mask per node which contains all
6173 * cpus of nodes that are that many hops away from us.
6175 for (i = 0; i < level; i++) {
6176 sched_domains_numa_masks[i] =
6177 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6178 if (!sched_domains_numa_masks[i])
6181 for (j = 0; j < nr_node_ids; j++) {
6182 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6186 sched_domains_numa_masks[i][j] = mask;
6188 for (k = 0; k < nr_node_ids; k++) {
6189 if (node_distance(j, k) > sched_domains_numa_distance[i])
6192 cpumask_or(mask, mask, cpumask_of_node(k));
6197 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6198 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6203 * Copy the default topology bits..
6205 for (i = 0; default_topology[i].init; i++)
6206 tl[i] = default_topology[i];
6209 * .. and append 'j' levels of NUMA goodness.
6211 for (j = 0; j < level; i++, j++) {
6212 tl[i] = (struct sched_domain_topology_level){
6213 .init = sd_numa_init,
6214 .mask = sd_numa_mask,
6215 .flags = SDTL_OVERLAP,
6220 sched_domain_topology = tl;
6222 sched_domains_numa_levels = level;
6225 static void sched_domains_numa_masks_set(int cpu)
6228 int node = cpu_to_node(cpu);
6230 for (i = 0; i < sched_domains_numa_levels; i++) {
6231 for (j = 0; j < nr_node_ids; j++) {
6232 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6233 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6238 static void sched_domains_numa_masks_clear(int cpu)
6241 for (i = 0; i < sched_domains_numa_levels; i++) {
6242 for (j = 0; j < nr_node_ids; j++)
6243 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6248 * Update sched_domains_numa_masks[level][node] array when new cpus
6251 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6252 unsigned long action,
6255 int cpu = (long)hcpu;
6257 switch (action & ~CPU_TASKS_FROZEN) {
6259 sched_domains_numa_masks_set(cpu);
6263 sched_domains_numa_masks_clear(cpu);
6273 static inline void sched_init_numa(void)
6277 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6278 unsigned long action,
6283 #endif /* CONFIG_NUMA */
6285 static int __sdt_alloc(const struct cpumask *cpu_map)
6287 struct sched_domain_topology_level *tl;
6290 for (tl = sched_domain_topology; tl->init; tl++) {
6291 struct sd_data *sdd = &tl->data;
6293 sdd->sd = alloc_percpu(struct sched_domain *);
6297 sdd->sg = alloc_percpu(struct sched_group *);
6301 sdd->sgp = alloc_percpu(struct sched_group_power *);
6305 for_each_cpu(j, cpu_map) {
6306 struct sched_domain *sd;
6307 struct sched_group *sg;
6308 struct sched_group_power *sgp;
6310 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6311 GFP_KERNEL, cpu_to_node(j));
6315 *per_cpu_ptr(sdd->sd, j) = sd;
6317 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6318 GFP_KERNEL, cpu_to_node(j));
6324 *per_cpu_ptr(sdd->sg, j) = sg;
6326 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6327 GFP_KERNEL, cpu_to_node(j));
6331 *per_cpu_ptr(sdd->sgp, j) = sgp;
6338 static void __sdt_free(const struct cpumask *cpu_map)
6340 struct sched_domain_topology_level *tl;
6343 for (tl = sched_domain_topology; tl->init; tl++) {
6344 struct sd_data *sdd = &tl->data;
6346 for_each_cpu(j, cpu_map) {
6347 struct sched_domain *sd;
6350 sd = *per_cpu_ptr(sdd->sd, j);
6351 if (sd && (sd->flags & SD_OVERLAP))
6352 free_sched_groups(sd->groups, 0);
6353 kfree(*per_cpu_ptr(sdd->sd, j));
6357 kfree(*per_cpu_ptr(sdd->sg, j));
6359 kfree(*per_cpu_ptr(sdd->sgp, j));
6361 free_percpu(sdd->sd);
6363 free_percpu(sdd->sg);
6365 free_percpu(sdd->sgp);
6370 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6371 struct s_data *d, const struct cpumask *cpu_map,
6372 struct sched_domain_attr *attr, struct sched_domain *child,
6375 struct sched_domain *sd = tl->init(tl, cpu);
6379 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6381 sd->level = child->level + 1;
6382 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6386 set_domain_attribute(sd, attr);
6392 * Build sched domains for a given set of cpus and attach the sched domains
6393 * to the individual cpus
6395 static int build_sched_domains(const struct cpumask *cpu_map,
6396 struct sched_domain_attr *attr)
6398 enum s_alloc alloc_state = sa_none;
6399 struct sched_domain *sd;
6401 int i, ret = -ENOMEM;
6403 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6404 if (alloc_state != sa_rootdomain)
6407 /* Set up domains for cpus specified by the cpu_map. */
6408 for_each_cpu(i, cpu_map) {
6409 struct sched_domain_topology_level *tl;
6412 for (tl = sched_domain_topology; tl->init; tl++) {
6413 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6414 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6415 sd->flags |= SD_OVERLAP;
6416 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6423 *per_cpu_ptr(d.sd, i) = sd;
6426 /* Build the groups for the domains */
6427 for_each_cpu(i, cpu_map) {
6428 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6429 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6430 if (sd->flags & SD_OVERLAP) {
6431 if (build_overlap_sched_groups(sd, i))
6434 if (build_sched_groups(sd, i))
6440 /* Calculate CPU power for physical packages and nodes */
6441 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6442 if (!cpumask_test_cpu(i, cpu_map))
6445 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6446 claim_allocations(i, sd);
6447 init_sched_groups_power(i, sd);
6451 /* Attach the domains */
6453 for_each_cpu(i, cpu_map) {
6454 sd = *per_cpu_ptr(d.sd, i);
6455 cpu_attach_domain(sd, d.rd, i);
6461 __free_domain_allocs(&d, alloc_state, cpu_map);
6465 static cpumask_var_t *doms_cur; /* current sched domains */
6466 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6467 static struct sched_domain_attr *dattr_cur;
6468 /* attribues of custom domains in 'doms_cur' */
6471 * Special case: If a kmalloc of a doms_cur partition (array of
6472 * cpumask) fails, then fallback to a single sched domain,
6473 * as determined by the single cpumask fallback_doms.
6475 static cpumask_var_t fallback_doms;
6478 * arch_update_cpu_topology lets virtualized architectures update the
6479 * cpu core maps. It is supposed to return 1 if the topology changed
6480 * or 0 if it stayed the same.
6482 int __attribute__((weak)) arch_update_cpu_topology(void)
6487 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6490 cpumask_var_t *doms;
6492 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6495 for (i = 0; i < ndoms; i++) {
6496 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6497 free_sched_domains(doms, i);
6504 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6507 for (i = 0; i < ndoms; i++)
6508 free_cpumask_var(doms[i]);
6513 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6514 * For now this just excludes isolated cpus, but could be used to
6515 * exclude other special cases in the future.
6517 static int init_sched_domains(const struct cpumask *cpu_map)
6521 arch_update_cpu_topology();
6523 doms_cur = alloc_sched_domains(ndoms_cur);
6525 doms_cur = &fallback_doms;
6526 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6527 err = build_sched_domains(doms_cur[0], NULL);
6528 register_sched_domain_sysctl();
6534 * Detach sched domains from a group of cpus specified in cpu_map
6535 * These cpus will now be attached to the NULL domain
6537 static void detach_destroy_domains(const struct cpumask *cpu_map)
6542 for_each_cpu(i, cpu_map)
6543 cpu_attach_domain(NULL, &def_root_domain, i);
6547 /* handle null as "default" */
6548 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6549 struct sched_domain_attr *new, int idx_new)
6551 struct sched_domain_attr tmp;
6558 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6559 new ? (new + idx_new) : &tmp,
6560 sizeof(struct sched_domain_attr));
6564 * Partition sched domains as specified by the 'ndoms_new'
6565 * cpumasks in the array doms_new[] of cpumasks. This compares
6566 * doms_new[] to the current sched domain partitioning, doms_cur[].
6567 * It destroys each deleted domain and builds each new domain.
6569 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6570 * The masks don't intersect (don't overlap.) We should setup one
6571 * sched domain for each mask. CPUs not in any of the cpumasks will
6572 * not be load balanced. If the same cpumask appears both in the
6573 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6576 * The passed in 'doms_new' should be allocated using
6577 * alloc_sched_domains. This routine takes ownership of it and will
6578 * free_sched_domains it when done with it. If the caller failed the
6579 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6580 * and partition_sched_domains() will fallback to the single partition
6581 * 'fallback_doms', it also forces the domains to be rebuilt.
6583 * If doms_new == NULL it will be replaced with cpu_online_mask.
6584 * ndoms_new == 0 is a special case for destroying existing domains,
6585 * and it will not create the default domain.
6587 * Call with hotplug lock held
6589 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6590 struct sched_domain_attr *dattr_new)
6595 mutex_lock(&sched_domains_mutex);
6597 /* always unregister in case we don't destroy any domains */
6598 unregister_sched_domain_sysctl();
6600 /* Let architecture update cpu core mappings. */
6601 new_topology = arch_update_cpu_topology();
6603 n = doms_new ? ndoms_new : 0;
6605 /* Destroy deleted domains */
6606 for (i = 0; i < ndoms_cur; i++) {
6607 for (j = 0; j < n && !new_topology; j++) {
6608 if (cpumask_equal(doms_cur[i], doms_new[j])
6609 && dattrs_equal(dattr_cur, i, dattr_new, j))
6612 /* no match - a current sched domain not in new doms_new[] */
6613 detach_destroy_domains(doms_cur[i]);
6618 if (doms_new == NULL) {
6620 doms_new = &fallback_doms;
6621 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6622 WARN_ON_ONCE(dattr_new);
6625 /* Build new domains */
6626 for (i = 0; i < ndoms_new; i++) {
6627 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6628 if (cpumask_equal(doms_new[i], doms_cur[j])
6629 && dattrs_equal(dattr_new, i, dattr_cur, j))
6632 /* no match - add a new doms_new */
6633 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6638 /* Remember the new sched domains */
6639 if (doms_cur != &fallback_doms)
6640 free_sched_domains(doms_cur, ndoms_cur);
6641 kfree(dattr_cur); /* kfree(NULL) is safe */
6642 doms_cur = doms_new;
6643 dattr_cur = dattr_new;
6644 ndoms_cur = ndoms_new;
6646 register_sched_domain_sysctl();
6648 mutex_unlock(&sched_domains_mutex);
6651 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6654 * Update cpusets according to cpu_active mask. If cpusets are
6655 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6656 * around partition_sched_domains().
6658 * If we come here as part of a suspend/resume, don't touch cpusets because we
6659 * want to restore it back to its original state upon resume anyway.
6661 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6665 case CPU_ONLINE_FROZEN:
6666 case CPU_DOWN_FAILED_FROZEN:
6669 * num_cpus_frozen tracks how many CPUs are involved in suspend
6670 * resume sequence. As long as this is not the last online
6671 * operation in the resume sequence, just build a single sched
6672 * domain, ignoring cpusets.
6675 if (likely(num_cpus_frozen)) {
6676 partition_sched_domains(1, NULL, NULL);
6681 * This is the last CPU online operation. So fall through and
6682 * restore the original sched domains by considering the
6683 * cpuset configurations.
6687 case CPU_DOWN_FAILED:
6688 cpuset_update_active_cpus(true);
6696 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6700 case CPU_DOWN_PREPARE:
6701 cpuset_update_active_cpus(false);
6703 case CPU_DOWN_PREPARE_FROZEN:
6705 partition_sched_domains(1, NULL, NULL);
6713 void __init sched_init_smp(void)
6715 cpumask_var_t non_isolated_cpus;
6717 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6718 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6723 mutex_lock(&sched_domains_mutex);
6724 init_sched_domains(cpu_active_mask);
6725 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6726 if (cpumask_empty(non_isolated_cpus))
6727 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6728 mutex_unlock(&sched_domains_mutex);
6731 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6732 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6733 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6735 /* RT runtime code needs to handle some hotplug events */
6736 hotcpu_notifier(update_runtime, 0);
6740 /* Move init over to a non-isolated CPU */
6741 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6743 sched_init_granularity();
6744 free_cpumask_var(non_isolated_cpus);
6746 init_sched_rt_class();
6749 void __init sched_init_smp(void)
6751 sched_init_granularity();
6753 #endif /* CONFIG_SMP */
6755 const_debug unsigned int sysctl_timer_migration = 1;
6757 int in_sched_functions(unsigned long addr)
6759 return in_lock_functions(addr) ||
6760 (addr >= (unsigned long)__sched_text_start
6761 && addr < (unsigned long)__sched_text_end);
6764 #ifdef CONFIG_CGROUP_SCHED
6765 struct task_group root_task_group;
6766 LIST_HEAD(task_groups);
6769 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6771 void __init sched_init(void)
6774 unsigned long alloc_size = 0, ptr;
6776 #ifdef CONFIG_FAIR_GROUP_SCHED
6777 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6779 #ifdef CONFIG_RT_GROUP_SCHED
6780 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6782 #ifdef CONFIG_CPUMASK_OFFSTACK
6783 alloc_size += num_possible_cpus() * cpumask_size();
6786 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6788 #ifdef CONFIG_FAIR_GROUP_SCHED
6789 root_task_group.se = (struct sched_entity **)ptr;
6790 ptr += nr_cpu_ids * sizeof(void **);
6792 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6793 ptr += nr_cpu_ids * sizeof(void **);
6795 #endif /* CONFIG_FAIR_GROUP_SCHED */
6796 #ifdef CONFIG_RT_GROUP_SCHED
6797 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6798 ptr += nr_cpu_ids * sizeof(void **);
6800 root_task_group.rt_rq = (struct rt_rq **)ptr;
6801 ptr += nr_cpu_ids * sizeof(void **);
6803 #endif /* CONFIG_RT_GROUP_SCHED */
6804 #ifdef CONFIG_CPUMASK_OFFSTACK
6805 for_each_possible_cpu(i) {
6806 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6807 ptr += cpumask_size();
6809 #endif /* CONFIG_CPUMASK_OFFSTACK */
6813 init_defrootdomain();
6816 init_rt_bandwidth(&def_rt_bandwidth,
6817 global_rt_period(), global_rt_runtime());
6819 #ifdef CONFIG_RT_GROUP_SCHED
6820 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6821 global_rt_period(), global_rt_runtime());
6822 #endif /* CONFIG_RT_GROUP_SCHED */
6824 #ifdef CONFIG_CGROUP_SCHED
6825 list_add(&root_task_group.list, &task_groups);
6826 INIT_LIST_HEAD(&root_task_group.children);
6827 INIT_LIST_HEAD(&root_task_group.siblings);
6828 autogroup_init(&init_task);
6830 #endif /* CONFIG_CGROUP_SCHED */
6832 #ifdef CONFIG_CGROUP_CPUACCT
6833 root_cpuacct.cpustat = &kernel_cpustat;
6834 root_cpuacct.cpuusage = alloc_percpu(u64);
6835 /* Too early, not expected to fail */
6836 BUG_ON(!root_cpuacct.cpuusage);
6838 for_each_possible_cpu(i) {
6842 raw_spin_lock_init(&rq->lock);
6844 rq->calc_load_active = 0;
6845 rq->calc_load_update = jiffies + LOAD_FREQ;
6846 init_cfs_rq(&rq->cfs);
6847 init_rt_rq(&rq->rt, rq);
6848 #ifdef CONFIG_FAIR_GROUP_SCHED
6849 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6850 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6852 * How much cpu bandwidth does root_task_group get?
6854 * In case of task-groups formed thr' the cgroup filesystem, it
6855 * gets 100% of the cpu resources in the system. This overall
6856 * system cpu resource is divided among the tasks of
6857 * root_task_group and its child task-groups in a fair manner,
6858 * based on each entity's (task or task-group's) weight
6859 * (se->load.weight).
6861 * In other words, if root_task_group has 10 tasks of weight
6862 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6863 * then A0's share of the cpu resource is:
6865 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6867 * We achieve this by letting root_task_group's tasks sit
6868 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6870 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6871 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6872 #endif /* CONFIG_FAIR_GROUP_SCHED */
6874 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6875 #ifdef CONFIG_RT_GROUP_SCHED
6876 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6877 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6880 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6881 rq->cpu_load[j] = 0;
6883 rq->last_load_update_tick = jiffies;
6888 rq->cpu_power = SCHED_POWER_SCALE;
6889 rq->post_schedule = 0;
6890 rq->active_balance = 0;
6891 rq->next_balance = jiffies;
6896 rq->avg_idle = 2*sysctl_sched_migration_cost;
6898 INIT_LIST_HEAD(&rq->cfs_tasks);
6900 rq_attach_root(rq, &def_root_domain);
6906 atomic_set(&rq->nr_iowait, 0);
6909 set_load_weight(&init_task);
6911 #ifdef CONFIG_PREEMPT_NOTIFIERS
6912 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6915 #ifdef CONFIG_RT_MUTEXES
6916 plist_head_init(&init_task.pi_waiters);
6920 * The boot idle thread does lazy MMU switching as well:
6922 atomic_inc(&init_mm.mm_count);
6923 enter_lazy_tlb(&init_mm, current);
6926 * Make us the idle thread. Technically, schedule() should not be
6927 * called from this thread, however somewhere below it might be,
6928 * but because we are the idle thread, we just pick up running again
6929 * when this runqueue becomes "idle".
6931 init_idle(current, smp_processor_id());
6933 calc_load_update = jiffies + LOAD_FREQ;
6936 * During early bootup we pretend to be a normal task:
6938 current->sched_class = &fair_sched_class;
6941 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6942 /* May be allocated at isolcpus cmdline parse time */
6943 if (cpu_isolated_map == NULL)
6944 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6945 idle_thread_set_boot_cpu();
6947 init_sched_fair_class();
6949 scheduler_running = 1;
6952 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6953 static inline int preempt_count_equals(int preempt_offset)
6955 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6957 return (nested == preempt_offset);
6960 void __might_sleep(const char *file, int line, int preempt_offset)
6962 static unsigned long prev_jiffy; /* ratelimiting */
6964 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6965 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6966 system_state != SYSTEM_RUNNING || oops_in_progress)
6968 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6970 prev_jiffy = jiffies;
6973 "BUG: sleeping function called from invalid context at %s:%d\n",
6976 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6977 in_atomic(), irqs_disabled(),
6978 current->pid, current->comm);
6980 debug_show_held_locks(current);
6981 if (irqs_disabled())
6982 print_irqtrace_events(current);
6985 EXPORT_SYMBOL(__might_sleep);
6988 #ifdef CONFIG_MAGIC_SYSRQ
6989 static void normalize_task(struct rq *rq, struct task_struct *p)
6991 const struct sched_class *prev_class = p->sched_class;
6992 int old_prio = p->prio;
6997 dequeue_task(rq, p, 0);
6998 __setscheduler(rq, p, SCHED_NORMAL, 0);
7000 enqueue_task(rq, p, 0);
7001 resched_task(rq->curr);
7004 check_class_changed(rq, p, prev_class, old_prio);
7007 void normalize_rt_tasks(void)
7009 struct task_struct *g, *p;
7010 unsigned long flags;
7013 read_lock_irqsave(&tasklist_lock, flags);
7014 do_each_thread(g, p) {
7016 * Only normalize user tasks:
7021 p->se.exec_start = 0;
7022 #ifdef CONFIG_SCHEDSTATS
7023 p->se.statistics.wait_start = 0;
7024 p->se.statistics.sleep_start = 0;
7025 p->se.statistics.block_start = 0;
7030 * Renice negative nice level userspace
7033 if (TASK_NICE(p) < 0 && p->mm)
7034 set_user_nice(p, 0);
7038 raw_spin_lock(&p->pi_lock);
7039 rq = __task_rq_lock(p);
7041 normalize_task(rq, p);
7043 __task_rq_unlock(rq);
7044 raw_spin_unlock(&p->pi_lock);
7045 } while_each_thread(g, p);
7047 read_unlock_irqrestore(&tasklist_lock, flags);
7050 #endif /* CONFIG_MAGIC_SYSRQ */
7052 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7054 * These functions are only useful for the IA64 MCA handling, or kdb.
7056 * They can only be called when the whole system has been
7057 * stopped - every CPU needs to be quiescent, and no scheduling
7058 * activity can take place. Using them for anything else would
7059 * be a serious bug, and as a result, they aren't even visible
7060 * under any other configuration.
7064 * curr_task - return the current task for a given cpu.
7065 * @cpu: the processor in question.
7067 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7069 struct task_struct *curr_task(int cpu)
7071 return cpu_curr(cpu);
7074 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7078 * set_curr_task - set the current task for a given cpu.
7079 * @cpu: the processor in question.
7080 * @p: the task pointer to set.
7082 * Description: This function must only be used when non-maskable interrupts
7083 * are serviced on a separate stack. It allows the architecture to switch the
7084 * notion of the current task on a cpu in a non-blocking manner. This function
7085 * must be called with all CPU's synchronized, and interrupts disabled, the
7086 * and caller must save the original value of the current task (see
7087 * curr_task() above) and restore that value before reenabling interrupts and
7088 * re-starting the system.
7090 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7092 void set_curr_task(int cpu, struct task_struct *p)
7099 #ifdef CONFIG_CGROUP_SCHED
7100 /* task_group_lock serializes the addition/removal of task groups */
7101 static DEFINE_SPINLOCK(task_group_lock);
7103 static void free_sched_group(struct task_group *tg)
7105 free_fair_sched_group(tg);
7106 free_rt_sched_group(tg);
7111 /* allocate runqueue etc for a new task group */
7112 struct task_group *sched_create_group(struct task_group *parent)
7114 struct task_group *tg;
7115 unsigned long flags;
7117 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7119 return ERR_PTR(-ENOMEM);
7121 if (!alloc_fair_sched_group(tg, parent))
7124 if (!alloc_rt_sched_group(tg, parent))
7127 spin_lock_irqsave(&task_group_lock, flags);
7128 list_add_rcu(&tg->list, &task_groups);
7130 WARN_ON(!parent); /* root should already exist */
7132 tg->parent = parent;
7133 INIT_LIST_HEAD(&tg->children);
7134 list_add_rcu(&tg->siblings, &parent->children);
7135 spin_unlock_irqrestore(&task_group_lock, flags);
7140 free_sched_group(tg);
7141 return ERR_PTR(-ENOMEM);
7144 /* rcu callback to free various structures associated with a task group */
7145 static void free_sched_group_rcu(struct rcu_head *rhp)
7147 /* now it should be safe to free those cfs_rqs */
7148 free_sched_group(container_of(rhp, struct task_group, rcu));
7151 /* Destroy runqueue etc associated with a task group */
7152 void sched_destroy_group(struct task_group *tg)
7154 unsigned long flags;
7157 /* end participation in shares distribution */
7158 for_each_possible_cpu(i)
7159 unregister_fair_sched_group(tg, i);
7161 spin_lock_irqsave(&task_group_lock, flags);
7162 list_del_rcu(&tg->list);
7163 list_del_rcu(&tg->siblings);
7164 spin_unlock_irqrestore(&task_group_lock, flags);
7166 /* wait for possible concurrent references to cfs_rqs complete */
7167 call_rcu(&tg->rcu, free_sched_group_rcu);
7170 /* change task's runqueue when it moves between groups.
7171 * The caller of this function should have put the task in its new group
7172 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7173 * reflect its new group.
7175 void sched_move_task(struct task_struct *tsk)
7177 struct task_group *tg;
7179 unsigned long flags;
7182 rq = task_rq_lock(tsk, &flags);
7184 running = task_current(rq, tsk);
7188 dequeue_task(rq, tsk, 0);
7189 if (unlikely(running))
7190 tsk->sched_class->put_prev_task(rq, tsk);
7192 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7193 lockdep_is_held(&tsk->sighand->siglock)),
7194 struct task_group, css);
7195 tg = autogroup_task_group(tsk, tg);
7196 tsk->sched_task_group = tg;
7198 #ifdef CONFIG_FAIR_GROUP_SCHED
7199 if (tsk->sched_class->task_move_group)
7200 tsk->sched_class->task_move_group(tsk, on_rq);
7203 set_task_rq(tsk, task_cpu(tsk));
7205 if (unlikely(running))
7206 tsk->sched_class->set_curr_task(rq);
7208 enqueue_task(rq, tsk, 0);
7210 task_rq_unlock(rq, tsk, &flags);
7212 #endif /* CONFIG_CGROUP_SCHED */
7214 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7215 static unsigned long to_ratio(u64 period, u64 runtime)
7217 if (runtime == RUNTIME_INF)
7220 return div64_u64(runtime << 20, period);
7224 #ifdef CONFIG_RT_GROUP_SCHED
7226 * Ensure that the real time constraints are schedulable.
7228 static DEFINE_MUTEX(rt_constraints_mutex);
7230 /* Must be called with tasklist_lock held */
7231 static inline int tg_has_rt_tasks(struct task_group *tg)
7233 struct task_struct *g, *p;
7235 do_each_thread(g, p) {
7236 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7238 } while_each_thread(g, p);
7243 struct rt_schedulable_data {
7244 struct task_group *tg;
7249 static int tg_rt_schedulable(struct task_group *tg, void *data)
7251 struct rt_schedulable_data *d = data;
7252 struct task_group *child;
7253 unsigned long total, sum = 0;
7254 u64 period, runtime;
7256 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7257 runtime = tg->rt_bandwidth.rt_runtime;
7260 period = d->rt_period;
7261 runtime = d->rt_runtime;
7265 * Cannot have more runtime than the period.
7267 if (runtime > period && runtime != RUNTIME_INF)
7271 * Ensure we don't starve existing RT tasks.
7273 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7276 total = to_ratio(period, runtime);
7279 * Nobody can have more than the global setting allows.
7281 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7285 * The sum of our children's runtime should not exceed our own.
7287 list_for_each_entry_rcu(child, &tg->children, siblings) {
7288 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7289 runtime = child->rt_bandwidth.rt_runtime;
7291 if (child == d->tg) {
7292 period = d->rt_period;
7293 runtime = d->rt_runtime;
7296 sum += to_ratio(period, runtime);
7305 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7309 struct rt_schedulable_data data = {
7311 .rt_period = period,
7312 .rt_runtime = runtime,
7316 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7322 static int tg_set_rt_bandwidth(struct task_group *tg,
7323 u64 rt_period, u64 rt_runtime)
7327 mutex_lock(&rt_constraints_mutex);
7328 read_lock(&tasklist_lock);
7329 err = __rt_schedulable(tg, rt_period, rt_runtime);
7333 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7334 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7335 tg->rt_bandwidth.rt_runtime = rt_runtime;
7337 for_each_possible_cpu(i) {
7338 struct rt_rq *rt_rq = tg->rt_rq[i];
7340 raw_spin_lock(&rt_rq->rt_runtime_lock);
7341 rt_rq->rt_runtime = rt_runtime;
7342 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7344 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7346 read_unlock(&tasklist_lock);
7347 mutex_unlock(&rt_constraints_mutex);
7352 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7354 u64 rt_runtime, rt_period;
7356 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7357 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7358 if (rt_runtime_us < 0)
7359 rt_runtime = RUNTIME_INF;
7361 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7364 long sched_group_rt_runtime(struct task_group *tg)
7368 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7371 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7372 do_div(rt_runtime_us, NSEC_PER_USEC);
7373 return rt_runtime_us;
7376 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7378 u64 rt_runtime, rt_period;
7380 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7381 rt_runtime = tg->rt_bandwidth.rt_runtime;
7386 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7389 long sched_group_rt_period(struct task_group *tg)
7393 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7394 do_div(rt_period_us, NSEC_PER_USEC);
7395 return rt_period_us;
7398 static int sched_rt_global_constraints(void)
7400 u64 runtime, period;
7403 if (sysctl_sched_rt_period <= 0)
7406 runtime = global_rt_runtime();
7407 period = global_rt_period();
7410 * Sanity check on the sysctl variables.
7412 if (runtime > period && runtime != RUNTIME_INF)
7415 mutex_lock(&rt_constraints_mutex);
7416 read_lock(&tasklist_lock);
7417 ret = __rt_schedulable(NULL, 0, 0);
7418 read_unlock(&tasklist_lock);
7419 mutex_unlock(&rt_constraints_mutex);
7424 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7426 /* Don't accept realtime tasks when there is no way for them to run */
7427 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7433 #else /* !CONFIG_RT_GROUP_SCHED */
7434 static int sched_rt_global_constraints(void)
7436 unsigned long flags;
7439 if (sysctl_sched_rt_period <= 0)
7443 * There's always some RT tasks in the root group
7444 * -- migration, kstopmachine etc..
7446 if (sysctl_sched_rt_runtime == 0)
7449 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7450 for_each_possible_cpu(i) {
7451 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7453 raw_spin_lock(&rt_rq->rt_runtime_lock);
7454 rt_rq->rt_runtime = global_rt_runtime();
7455 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7457 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7461 #endif /* CONFIG_RT_GROUP_SCHED */
7463 int sched_rt_handler(struct ctl_table *table, int write,
7464 void __user *buffer, size_t *lenp,
7468 int old_period, old_runtime;
7469 static DEFINE_MUTEX(mutex);
7472 old_period = sysctl_sched_rt_period;
7473 old_runtime = sysctl_sched_rt_runtime;
7475 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7477 if (!ret && write) {
7478 ret = sched_rt_global_constraints();
7480 sysctl_sched_rt_period = old_period;
7481 sysctl_sched_rt_runtime = old_runtime;
7483 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7484 def_rt_bandwidth.rt_period =
7485 ns_to_ktime(global_rt_period());
7488 mutex_unlock(&mutex);
7493 #ifdef CONFIG_CGROUP_SCHED
7495 /* return corresponding task_group object of a cgroup */
7496 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7498 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7499 struct task_group, css);
7502 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7504 struct task_group *tg, *parent;
7506 if (!cgrp->parent) {
7507 /* This is early initialization for the top cgroup */
7508 return &root_task_group.css;
7511 parent = cgroup_tg(cgrp->parent);
7512 tg = sched_create_group(parent);
7514 return ERR_PTR(-ENOMEM);
7519 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7521 struct task_group *tg = cgroup_tg(cgrp);
7523 sched_destroy_group(tg);
7526 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7527 struct cgroup_taskset *tset)
7529 struct task_struct *task;
7531 cgroup_taskset_for_each(task, cgrp, tset) {
7532 #ifdef CONFIG_RT_GROUP_SCHED
7533 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7536 /* We don't support RT-tasks being in separate groups */
7537 if (task->sched_class != &fair_sched_class)
7544 static void cpu_cgroup_attach(struct cgroup *cgrp,
7545 struct cgroup_taskset *tset)
7547 struct task_struct *task;
7549 cgroup_taskset_for_each(task, cgrp, tset)
7550 sched_move_task(task);
7554 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7555 struct task_struct *task)
7558 * cgroup_exit() is called in the copy_process() failure path.
7559 * Ignore this case since the task hasn't ran yet, this avoids
7560 * trying to poke a half freed task state from generic code.
7562 if (!(task->flags & PF_EXITING))
7565 sched_move_task(task);
7568 #ifdef CONFIG_FAIR_GROUP_SCHED
7569 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7572 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7575 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7577 struct task_group *tg = cgroup_tg(cgrp);
7579 return (u64) scale_load_down(tg->shares);
7582 #ifdef CONFIG_CFS_BANDWIDTH
7583 static DEFINE_MUTEX(cfs_constraints_mutex);
7585 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7586 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7588 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7590 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7592 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7593 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7595 if (tg == &root_task_group)
7599 * Ensure we have at some amount of bandwidth every period. This is
7600 * to prevent reaching a state of large arrears when throttled via
7601 * entity_tick() resulting in prolonged exit starvation.
7603 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7607 * Likewise, bound things on the otherside by preventing insane quota
7608 * periods. This also allows us to normalize in computing quota
7611 if (period > max_cfs_quota_period)
7614 mutex_lock(&cfs_constraints_mutex);
7615 ret = __cfs_schedulable(tg, period, quota);
7619 runtime_enabled = quota != RUNTIME_INF;
7620 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7621 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7622 raw_spin_lock_irq(&cfs_b->lock);
7623 cfs_b->period = ns_to_ktime(period);
7624 cfs_b->quota = quota;
7626 __refill_cfs_bandwidth_runtime(cfs_b);
7627 /* restart the period timer (if active) to handle new period expiry */
7628 if (runtime_enabled && cfs_b->timer_active) {
7629 /* force a reprogram */
7630 cfs_b->timer_active = 0;
7631 __start_cfs_bandwidth(cfs_b);
7633 raw_spin_unlock_irq(&cfs_b->lock);
7635 for_each_possible_cpu(i) {
7636 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7637 struct rq *rq = cfs_rq->rq;
7639 raw_spin_lock_irq(&rq->lock);
7640 cfs_rq->runtime_enabled = runtime_enabled;
7641 cfs_rq->runtime_remaining = 0;
7643 if (cfs_rq->throttled)
7644 unthrottle_cfs_rq(cfs_rq);
7645 raw_spin_unlock_irq(&rq->lock);
7648 mutex_unlock(&cfs_constraints_mutex);
7653 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7657 period = ktime_to_ns(tg->cfs_bandwidth.period);
7658 if (cfs_quota_us < 0)
7659 quota = RUNTIME_INF;
7661 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7663 return tg_set_cfs_bandwidth(tg, period, quota);
7666 long tg_get_cfs_quota(struct task_group *tg)
7670 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7673 quota_us = tg->cfs_bandwidth.quota;
7674 do_div(quota_us, NSEC_PER_USEC);
7679 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7683 period = (u64)cfs_period_us * NSEC_PER_USEC;
7684 quota = tg->cfs_bandwidth.quota;
7686 return tg_set_cfs_bandwidth(tg, period, quota);
7689 long tg_get_cfs_period(struct task_group *tg)
7693 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7694 do_div(cfs_period_us, NSEC_PER_USEC);
7696 return cfs_period_us;
7699 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7701 return tg_get_cfs_quota(cgroup_tg(cgrp));
7704 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7707 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7710 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7712 return tg_get_cfs_period(cgroup_tg(cgrp));
7715 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7718 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7721 struct cfs_schedulable_data {
7722 struct task_group *tg;
7727 * normalize group quota/period to be quota/max_period
7728 * note: units are usecs
7730 static u64 normalize_cfs_quota(struct task_group *tg,
7731 struct cfs_schedulable_data *d)
7739 period = tg_get_cfs_period(tg);
7740 quota = tg_get_cfs_quota(tg);
7743 /* note: these should typically be equivalent */
7744 if (quota == RUNTIME_INF || quota == -1)
7747 return to_ratio(period, quota);
7750 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7752 struct cfs_schedulable_data *d = data;
7753 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7754 s64 quota = 0, parent_quota = -1;
7757 quota = RUNTIME_INF;
7759 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7761 quota = normalize_cfs_quota(tg, d);
7762 parent_quota = parent_b->hierarchal_quota;
7765 * ensure max(child_quota) <= parent_quota, inherit when no
7768 if (quota == RUNTIME_INF)
7769 quota = parent_quota;
7770 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7773 cfs_b->hierarchal_quota = quota;
7778 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7781 struct cfs_schedulable_data data = {
7787 if (quota != RUNTIME_INF) {
7788 do_div(data.period, NSEC_PER_USEC);
7789 do_div(data.quota, NSEC_PER_USEC);
7793 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7799 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7800 struct cgroup_map_cb *cb)
7802 struct task_group *tg = cgroup_tg(cgrp);
7803 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7805 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7806 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7807 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7811 #endif /* CONFIG_CFS_BANDWIDTH */
7812 #endif /* CONFIG_FAIR_GROUP_SCHED */
7814 #ifdef CONFIG_RT_GROUP_SCHED
7815 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7818 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7821 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7823 return sched_group_rt_runtime(cgroup_tg(cgrp));
7826 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7829 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7832 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7834 return sched_group_rt_period(cgroup_tg(cgrp));
7836 #endif /* CONFIG_RT_GROUP_SCHED */
7838 static struct cftype cpu_files[] = {
7839 #ifdef CONFIG_FAIR_GROUP_SCHED
7842 .read_u64 = cpu_shares_read_u64,
7843 .write_u64 = cpu_shares_write_u64,
7846 #ifdef CONFIG_CFS_BANDWIDTH
7848 .name = "cfs_quota_us",
7849 .read_s64 = cpu_cfs_quota_read_s64,
7850 .write_s64 = cpu_cfs_quota_write_s64,
7853 .name = "cfs_period_us",
7854 .read_u64 = cpu_cfs_period_read_u64,
7855 .write_u64 = cpu_cfs_period_write_u64,
7859 .read_map = cpu_stats_show,
7862 #ifdef CONFIG_RT_GROUP_SCHED
7864 .name = "rt_runtime_us",
7865 .read_s64 = cpu_rt_runtime_read,
7866 .write_s64 = cpu_rt_runtime_write,
7869 .name = "rt_period_us",
7870 .read_u64 = cpu_rt_period_read_uint,
7871 .write_u64 = cpu_rt_period_write_uint,
7877 struct cgroup_subsys cpu_cgroup_subsys = {
7879 .css_alloc = cpu_cgroup_css_alloc,
7880 .css_free = cpu_cgroup_css_free,
7881 .can_attach = cpu_cgroup_can_attach,
7882 .attach = cpu_cgroup_attach,
7883 .exit = cpu_cgroup_exit,
7884 .subsys_id = cpu_cgroup_subsys_id,
7885 .base_cftypes = cpu_files,
7889 #endif /* CONFIG_CGROUP_SCHED */
7891 #ifdef CONFIG_CGROUP_CPUACCT
7894 * CPU accounting code for task groups.
7896 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7897 * (balbir@in.ibm.com).
7900 struct cpuacct root_cpuacct;
7902 /* create a new cpu accounting group */
7903 static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp)
7908 return &root_cpuacct.css;
7910 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7914 ca->cpuusage = alloc_percpu(u64);
7918 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7920 goto out_free_cpuusage;
7925 free_percpu(ca->cpuusage);
7929 return ERR_PTR(-ENOMEM);
7932 /* destroy an existing cpu accounting group */
7933 static void cpuacct_css_free(struct cgroup *cgrp)
7935 struct cpuacct *ca = cgroup_ca(cgrp);
7937 free_percpu(ca->cpustat);
7938 free_percpu(ca->cpuusage);
7942 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7944 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7947 #ifndef CONFIG_64BIT
7949 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7951 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7953 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7961 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7963 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7965 #ifndef CONFIG_64BIT
7967 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7969 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7971 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7977 /* return total cpu usage (in nanoseconds) of a group */
7978 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7980 struct cpuacct *ca = cgroup_ca(cgrp);
7981 u64 totalcpuusage = 0;
7984 for_each_present_cpu(i)
7985 totalcpuusage += cpuacct_cpuusage_read(ca, i);
7987 return totalcpuusage;
7990 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
7993 struct cpuacct *ca = cgroup_ca(cgrp);
8002 for_each_present_cpu(i)
8003 cpuacct_cpuusage_write(ca, i, 0);
8009 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8012 struct cpuacct *ca = cgroup_ca(cgroup);
8016 for_each_present_cpu(i) {
8017 percpu = cpuacct_cpuusage_read(ca, i);
8018 seq_printf(m, "%llu ", (unsigned long long) percpu);
8020 seq_printf(m, "\n");
8024 static const char *cpuacct_stat_desc[] = {
8025 [CPUACCT_STAT_USER] = "user",
8026 [CPUACCT_STAT_SYSTEM] = "system",
8029 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8030 struct cgroup_map_cb *cb)
8032 struct cpuacct *ca = cgroup_ca(cgrp);
8036 for_each_online_cpu(cpu) {
8037 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8038 val += kcpustat->cpustat[CPUTIME_USER];
8039 val += kcpustat->cpustat[CPUTIME_NICE];
8041 val = cputime64_to_clock_t(val);
8042 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8045 for_each_online_cpu(cpu) {
8046 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8047 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8048 val += kcpustat->cpustat[CPUTIME_IRQ];
8049 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8052 val = cputime64_to_clock_t(val);
8053 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8058 static struct cftype files[] = {
8061 .read_u64 = cpuusage_read,
8062 .write_u64 = cpuusage_write,
8065 .name = "usage_percpu",
8066 .read_seq_string = cpuacct_percpu_seq_read,
8070 .read_map = cpuacct_stats_show,
8076 * charge this task's execution time to its accounting group.
8078 * called with rq->lock held.
8080 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8085 if (unlikely(!cpuacct_subsys.active))
8088 cpu = task_cpu(tsk);
8094 for (; ca; ca = parent_ca(ca)) {
8095 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8096 *cpuusage += cputime;
8102 struct cgroup_subsys cpuacct_subsys = {
8104 .css_alloc = cpuacct_css_alloc,
8105 .css_free = cpuacct_css_free,
8106 .subsys_id = cpuacct_subsys_id,
8107 .base_cftypes = files,
8109 #endif /* CONFIG_CGROUP_CPUACCT */
8111 void dump_cpu_task(int cpu)
8113 pr_info("Task dump for CPU %d:\n", cpu);
8114 sched_show_task(cpu_curr(cpu));