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>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
86 #include "../smpboot.h"
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/sched.h>
91 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
94 ktime_t soft, hard, now;
97 if (hrtimer_active(period_timer))
100 now = hrtimer_cb_get_time(period_timer);
101 hrtimer_forward(period_timer, now, period);
103 soft = hrtimer_get_softexpires(period_timer);
104 hard = hrtimer_get_expires(period_timer);
105 delta = ktime_to_ns(ktime_sub(hard, soft));
106 __hrtimer_start_range_ns(period_timer, soft, delta,
107 HRTIMER_MODE_ABS_PINNED, 0);
111 DEFINE_MUTEX(sched_domains_mutex);
112 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
114 static void update_rq_clock_task(struct rq *rq, s64 delta);
116 void update_rq_clock(struct rq *rq)
120 if (rq->skip_clock_update > 0)
123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
125 update_rq_clock_task(rq, delta);
129 * Debugging: various feature bits
132 #define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
135 const_debug unsigned int sysctl_sched_features =
136 #include "features.h"
141 #ifdef CONFIG_SCHED_DEBUG
142 #define SCHED_FEAT(name, enabled) \
145 static const char * const sched_feat_names[] = {
146 #include "features.h"
151 static int sched_feat_show(struct seq_file *m, void *v)
155 for (i = 0; i < __SCHED_FEAT_NR; i++) {
156 if (!(sysctl_sched_features & (1UL << i)))
158 seq_printf(m, "%s ", sched_feat_names[i]);
165 #ifdef HAVE_JUMP_LABEL
167 #define jump_label_key__true STATIC_KEY_INIT_TRUE
168 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 #define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
173 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174 #include "features.h"
179 static void sched_feat_disable(int i)
181 if (static_key_enabled(&sched_feat_keys[i]))
182 static_key_slow_dec(&sched_feat_keys[i]);
185 static void sched_feat_enable(int i)
187 if (!static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_inc(&sched_feat_keys[i]);
191 static void sched_feat_disable(int i) { };
192 static void sched_feat_enable(int i) { };
193 #endif /* HAVE_JUMP_LABEL */
196 sched_feat_write(struct file *filp, const char __user *ubuf,
197 size_t cnt, loff_t *ppos)
207 if (copy_from_user(&buf, ubuf, cnt))
213 if (strncmp(cmp, "NO_", 3) == 0) {
218 for (i = 0; i < __SCHED_FEAT_NR; i++) {
219 if (strcmp(cmp, sched_feat_names[i]) == 0) {
221 sysctl_sched_features &= ~(1UL << i);
222 sched_feat_disable(i);
224 sysctl_sched_features |= (1UL << i);
225 sched_feat_enable(i);
231 if (i == __SCHED_FEAT_NR)
239 static int sched_feat_open(struct inode *inode, struct file *filp)
241 return single_open(filp, sched_feat_show, NULL);
244 static const struct file_operations sched_feat_fops = {
245 .open = sched_feat_open,
246 .write = sched_feat_write,
249 .release = single_release,
252 static __init int sched_init_debug(void)
254 debugfs_create_file("sched_features", 0644, NULL, NULL,
259 late_initcall(sched_init_debug);
260 #endif /* CONFIG_SCHED_DEBUG */
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
266 const_debug unsigned int sysctl_sched_nr_migrate = 32;
269 * period over which we average the RT time consumption, measured
274 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
277 * period over which we measure -rt task cpu usage in us.
280 unsigned int sysctl_sched_rt_period = 1000000;
282 __read_mostly int scheduler_running;
285 * part of the period that we allow rt tasks to run in us.
288 int sysctl_sched_rt_runtime = 950000;
293 * __task_rq_lock - lock the rq @p resides on.
295 static inline struct rq *__task_rq_lock(struct task_struct *p)
300 lockdep_assert_held(&p->pi_lock);
304 raw_spin_lock(&rq->lock);
305 if (likely(rq == task_rq(p)))
307 raw_spin_unlock(&rq->lock);
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
314 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
315 __acquires(p->pi_lock)
321 raw_spin_lock_irqsave(&p->pi_lock, *flags);
323 raw_spin_lock(&rq->lock);
324 if (likely(rq == task_rq(p)))
326 raw_spin_unlock(&rq->lock);
327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
331 static void __task_rq_unlock(struct rq *rq)
334 raw_spin_unlock(&rq->lock);
338 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
340 __releases(p->pi_lock)
342 raw_spin_unlock(&rq->lock);
343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
347 * this_rq_lock - lock this runqueue and disable interrupts.
349 static struct rq *this_rq_lock(void)
356 raw_spin_lock(&rq->lock);
361 #ifdef CONFIG_SCHED_HRTICK
363 * Use HR-timers to deliver accurate preemption points.
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
373 static void hrtick_clear(struct rq *rq)
375 if (hrtimer_active(&rq->hrtick_timer))
376 hrtimer_cancel(&rq->hrtick_timer);
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
383 static enum hrtimer_restart hrtick(struct hrtimer *timer)
385 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
389 raw_spin_lock(&rq->lock);
391 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392 raw_spin_unlock(&rq->lock);
394 return HRTIMER_NORESTART;
399 * called from hardirq (IPI) context
401 static void __hrtick_start(void *arg)
405 raw_spin_lock(&rq->lock);
406 hrtimer_restart(&rq->hrtick_timer);
407 rq->hrtick_csd_pending = 0;
408 raw_spin_unlock(&rq->lock);
412 * Called to set the hrtick timer state.
414 * called with rq->lock held and irqs disabled
416 void hrtick_start(struct rq *rq, u64 delay)
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
421 hrtimer_set_expires(timer, time);
423 if (rq == this_rq()) {
424 hrtimer_restart(timer);
425 } else if (!rq->hrtick_csd_pending) {
426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
427 rq->hrtick_csd_pending = 1;
432 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
434 int cpu = (int)(long)hcpu;
437 case CPU_UP_CANCELED:
438 case CPU_UP_CANCELED_FROZEN:
439 case CPU_DOWN_PREPARE:
440 case CPU_DOWN_PREPARE_FROZEN:
442 case CPU_DEAD_FROZEN:
443 hrtick_clear(cpu_rq(cpu));
450 static __init void init_hrtick(void)
452 hotcpu_notifier(hotplug_hrtick, 0);
456 * Called to set the hrtick timer state.
458 * called with rq->lock held and irqs disabled
460 void hrtick_start(struct rq *rq, u64 delay)
462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
463 HRTIMER_MODE_REL_PINNED, 0);
466 static inline void init_hrtick(void)
469 #endif /* CONFIG_SMP */
471 static void init_rq_hrtick(struct rq *rq)
474 rq->hrtick_csd_pending = 0;
476 rq->hrtick_csd.flags = 0;
477 rq->hrtick_csd.func = __hrtick_start;
478 rq->hrtick_csd.info = rq;
481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
482 rq->hrtick_timer.function = hrtick;
484 #else /* CONFIG_SCHED_HRTICK */
485 static inline void hrtick_clear(struct rq *rq)
489 static inline void init_rq_hrtick(struct rq *rq)
493 static inline void init_hrtick(void)
496 #endif /* CONFIG_SCHED_HRTICK */
499 * resched_task - mark a task 'to be rescheduled now'.
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
507 #ifndef tsk_is_polling
508 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
511 void resched_task(struct task_struct *p)
515 assert_raw_spin_locked(&task_rq(p)->lock);
517 if (test_tsk_need_resched(p))
520 set_tsk_need_resched(p);
523 if (cpu == smp_processor_id())
526 /* NEED_RESCHED must be visible before we test polling */
528 if (!tsk_is_polling(p))
529 smp_send_reschedule(cpu);
532 void resched_cpu(int cpu)
534 struct rq *rq = cpu_rq(cpu);
537 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
539 resched_task(cpu_curr(cpu));
540 raw_spin_unlock_irqrestore(&rq->lock, flags);
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
552 int get_nohz_timer_target(void)
554 int cpu = smp_processor_id();
556 struct sched_domain *sd;
559 for_each_domain(cpu, sd) {
560 for_each_cpu(i, sched_domain_span(sd)) {
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
581 void wake_up_idle_cpu(int cpu)
583 struct rq *rq = cpu_rq(cpu);
585 if (cpu == smp_processor_id())
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
595 if (rq->curr != rq->idle)
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
603 set_tsk_need_resched(rq->idle);
605 /* NEED_RESCHED must be visible before we test polling */
607 if (!tsk_is_polling(rq->idle))
608 smp_send_reschedule(cpu);
611 static inline bool got_nohz_idle_kick(void)
613 int cpu = smp_processor_id();
614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
617 #else /* CONFIG_NO_HZ */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ */
626 void sched_avg_update(struct rq *rq)
628 s64 period = sched_avg_period();
630 while ((s64)(rq->clock - rq->age_stamp) > period) {
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
636 asm("" : "+rm" (rq->age_stamp));
637 rq->age_stamp += period;
642 #else /* !CONFIG_SMP */
643 void resched_task(struct task_struct *p)
645 assert_raw_spin_locked(&task_rq(p)->lock);
646 set_tsk_need_resched(p);
648 #endif /* CONFIG_SMP */
650 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
656 * Caller must hold rcu_lock or sufficient equivalent.
658 int walk_tg_tree_from(struct task_group *from,
659 tg_visitor down, tg_visitor up, void *data)
661 struct task_group *parent, *child;
667 ret = (*down)(parent, data);
670 list_for_each_entry_rcu(child, &parent->children, siblings) {
677 ret = (*up)(parent, data);
678 if (ret || parent == from)
682 parent = parent->parent;
689 int tg_nop(struct task_group *tg, void *data)
695 static void set_load_weight(struct task_struct *p)
697 int prio = p->static_prio - MAX_RT_PRIO;
698 struct load_weight *load = &p->se.load;
701 * SCHED_IDLE tasks get minimal weight:
703 if (p->policy == SCHED_IDLE) {
704 load->weight = scale_load(WEIGHT_IDLEPRIO);
705 load->inv_weight = WMULT_IDLEPRIO;
709 load->weight = scale_load(prio_to_weight[prio]);
710 load->inv_weight = prio_to_wmult[prio];
713 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
716 sched_info_queued(p);
717 p->sched_class->enqueue_task(rq, p, flags);
720 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
723 sched_info_dequeued(p);
724 p->sched_class->dequeue_task(rq, p, flags);
727 void activate_task(struct rq *rq, struct task_struct *p, int flags)
729 if (task_contributes_to_load(p))
730 rq->nr_uninterruptible--;
732 enqueue_task(rq, p, flags);
735 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible++;
740 dequeue_task(rq, p, flags);
743 static void update_rq_clock_task(struct rq *rq, s64 delta)
746 * In theory, the compile should just see 0 here, and optimize out the call
747 * to sched_rt_avg_update. But I don't trust it...
749 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
750 s64 steal = 0, irq_delta = 0;
752 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
753 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
756 * Since irq_time is only updated on {soft,}irq_exit, we might run into
757 * this case when a previous update_rq_clock() happened inside a
760 * When this happens, we stop ->clock_task and only update the
761 * prev_irq_time stamp to account for the part that fit, so that a next
762 * update will consume the rest. This ensures ->clock_task is
765 * It does however cause some slight miss-attribution of {soft,}irq
766 * time, a more accurate solution would be to update the irq_time using
767 * the current rq->clock timestamp, except that would require using
770 if (irq_delta > delta)
773 rq->prev_irq_time += irq_delta;
776 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
777 if (static_key_false((¶virt_steal_rq_enabled))) {
780 steal = paravirt_steal_clock(cpu_of(rq));
781 steal -= rq->prev_steal_time_rq;
783 if (unlikely(steal > delta))
786 st = steal_ticks(steal);
787 steal = st * TICK_NSEC;
789 rq->prev_steal_time_rq += steal;
795 rq->clock_task += delta;
797 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
798 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
799 sched_rt_avg_update(rq, irq_delta + steal);
803 void sched_set_stop_task(int cpu, struct task_struct *stop)
805 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
806 struct task_struct *old_stop = cpu_rq(cpu)->stop;
810 * Make it appear like a SCHED_FIFO task, its something
811 * userspace knows about and won't get confused about.
813 * Also, it will make PI more or less work without too
814 * much confusion -- but then, stop work should not
815 * rely on PI working anyway.
817 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
819 stop->sched_class = &stop_sched_class;
822 cpu_rq(cpu)->stop = stop;
826 * Reset it back to a normal scheduling class so that
827 * it can die in pieces.
829 old_stop->sched_class = &rt_sched_class;
834 * __normal_prio - return the priority that is based on the static prio
836 static inline int __normal_prio(struct task_struct *p)
838 return p->static_prio;
842 * Calculate the expected normal priority: i.e. priority
843 * without taking RT-inheritance into account. Might be
844 * boosted by interactivity modifiers. Changes upon fork,
845 * setprio syscalls, and whenever the interactivity
846 * estimator recalculates.
848 static inline int normal_prio(struct task_struct *p)
852 if (task_has_rt_policy(p))
853 prio = MAX_RT_PRIO-1 - p->rt_priority;
855 prio = __normal_prio(p);
860 * Calculate the current priority, i.e. the priority
861 * taken into account by the scheduler. This value might
862 * be boosted by RT tasks, or might be boosted by
863 * interactivity modifiers. Will be RT if the task got
864 * RT-boosted. If not then it returns p->normal_prio.
866 static int effective_prio(struct task_struct *p)
868 p->normal_prio = normal_prio(p);
870 * If we are RT tasks or we were boosted to RT priority,
871 * keep the priority unchanged. Otherwise, update priority
872 * to the normal priority:
874 if (!rt_prio(p->prio))
875 return p->normal_prio;
880 * task_curr - is this task currently executing on a CPU?
881 * @p: the task in question.
883 inline int task_curr(const struct task_struct *p)
885 return cpu_curr(task_cpu(p)) == p;
888 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
889 const struct sched_class *prev_class,
892 if (prev_class != p->sched_class) {
893 if (prev_class->switched_from)
894 prev_class->switched_from(rq, p);
895 p->sched_class->switched_to(rq, p);
896 } else if (oldprio != p->prio)
897 p->sched_class->prio_changed(rq, p, oldprio);
900 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
902 const struct sched_class *class;
904 if (p->sched_class == rq->curr->sched_class) {
905 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
907 for_each_class(class) {
908 if (class == rq->curr->sched_class)
910 if (class == p->sched_class) {
911 resched_task(rq->curr);
918 * A queue event has occurred, and we're going to schedule. In
919 * this case, we can save a useless back to back clock update.
921 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
922 rq->skip_clock_update = 1;
926 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
928 #ifdef CONFIG_SCHED_DEBUG
930 * We should never call set_task_cpu() on a blocked task,
931 * ttwu() will sort out the placement.
933 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
934 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
936 #ifdef CONFIG_LOCKDEP
938 * The caller should hold either p->pi_lock or rq->lock, when changing
939 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
941 * sched_move_task() holds both and thus holding either pins the cgroup,
944 * Furthermore, all task_rq users should acquire both locks, see
947 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
948 lockdep_is_held(&task_rq(p)->lock)));
952 trace_sched_migrate_task(p, new_cpu);
954 if (task_cpu(p) != new_cpu) {
955 p->se.nr_migrations++;
956 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
959 __set_task_cpu(p, new_cpu);
962 struct migration_arg {
963 struct task_struct *task;
967 static int migration_cpu_stop(void *data);
970 * wait_task_inactive - wait for a thread to unschedule.
972 * If @match_state is nonzero, it's the @p->state value just checked and
973 * not expected to change. If it changes, i.e. @p might have woken up,
974 * then return zero. When we succeed in waiting for @p to be off its CPU,
975 * we return a positive number (its total switch count). If a second call
976 * a short while later returns the same number, the caller can be sure that
977 * @p has remained unscheduled the whole time.
979 * The caller must ensure that the task *will* unschedule sometime soon,
980 * else this function might spin for a *long* time. This function can't
981 * be called with interrupts off, or it may introduce deadlock with
982 * smp_call_function() if an IPI is sent by the same process we are
983 * waiting to become inactive.
985 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
994 * We do the initial early heuristics without holding
995 * any task-queue locks at all. We'll only try to get
996 * the runqueue lock when things look like they will
1002 * If the task is actively running on another CPU
1003 * still, just relax and busy-wait without holding
1006 * NOTE! Since we don't hold any locks, it's not
1007 * even sure that "rq" stays as the right runqueue!
1008 * But we don't care, since "task_running()" will
1009 * return false if the runqueue has changed and p
1010 * is actually now running somewhere else!
1012 while (task_running(rq, p)) {
1013 if (match_state && unlikely(p->state != match_state))
1019 * Ok, time to look more closely! We need the rq
1020 * lock now, to be *sure*. If we're wrong, we'll
1021 * just go back and repeat.
1023 rq = task_rq_lock(p, &flags);
1024 trace_sched_wait_task(p);
1025 running = task_running(rq, p);
1028 if (!match_state || p->state == match_state)
1029 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1030 task_rq_unlock(rq, p, &flags);
1033 * If it changed from the expected state, bail out now.
1035 if (unlikely(!ncsw))
1039 * Was it really running after all now that we
1040 * checked with the proper locks actually held?
1042 * Oops. Go back and try again..
1044 if (unlikely(running)) {
1050 * It's not enough that it's not actively running,
1051 * it must be off the runqueue _entirely_, and not
1054 * So if it was still runnable (but just not actively
1055 * running right now), it's preempted, and we should
1056 * yield - it could be a while.
1058 if (unlikely(on_rq)) {
1059 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1061 set_current_state(TASK_UNINTERRUPTIBLE);
1062 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1067 * Ahh, all good. It wasn't running, and it wasn't
1068 * runnable, which means that it will never become
1069 * running in the future either. We're all done!
1078 * kick_process - kick a running thread to enter/exit the kernel
1079 * @p: the to-be-kicked thread
1081 * Cause a process which is running on another CPU to enter
1082 * kernel-mode, without any delay. (to get signals handled.)
1084 * NOTE: this function doesn't have to take the runqueue lock,
1085 * because all it wants to ensure is that the remote task enters
1086 * the kernel. If the IPI races and the task has been migrated
1087 * to another CPU then no harm is done and the purpose has been
1090 void kick_process(struct task_struct *p)
1096 if ((cpu != smp_processor_id()) && task_curr(p))
1097 smp_send_reschedule(cpu);
1100 EXPORT_SYMBOL_GPL(kick_process);
1101 #endif /* CONFIG_SMP */
1105 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1107 static int select_fallback_rq(int cpu, struct task_struct *p)
1109 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1110 enum { cpuset, possible, fail } state = cpuset;
1113 /* Look for allowed, online CPU in same node. */
1114 for_each_cpu(dest_cpu, nodemask) {
1115 if (!cpu_online(dest_cpu))
1117 if (!cpu_active(dest_cpu))
1119 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1124 /* Any allowed, online CPU? */
1125 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1126 if (!cpu_online(dest_cpu))
1128 if (!cpu_active(dest_cpu))
1135 /* No more Mr. Nice Guy. */
1136 cpuset_cpus_allowed_fallback(p);
1141 do_set_cpus_allowed(p, cpu_possible_mask);
1152 if (state != cpuset) {
1154 * Don't tell them about moving exiting tasks or
1155 * kernel threads (both mm NULL), since they never
1158 if (p->mm && printk_ratelimit()) {
1159 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1160 task_pid_nr(p), p->comm, cpu);
1168 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1171 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1173 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1176 * In order not to call set_task_cpu() on a blocking task we need
1177 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1180 * Since this is common to all placement strategies, this lives here.
1182 * [ this allows ->select_task() to simply return task_cpu(p) and
1183 * not worry about this generic constraint ]
1185 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1187 cpu = select_fallback_rq(task_cpu(p), p);
1192 static void update_avg(u64 *avg, u64 sample)
1194 s64 diff = sample - *avg;
1200 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1202 #ifdef CONFIG_SCHEDSTATS
1203 struct rq *rq = this_rq();
1206 int this_cpu = smp_processor_id();
1208 if (cpu == this_cpu) {
1209 schedstat_inc(rq, ttwu_local);
1210 schedstat_inc(p, se.statistics.nr_wakeups_local);
1212 struct sched_domain *sd;
1214 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1216 for_each_domain(this_cpu, sd) {
1217 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1218 schedstat_inc(sd, ttwu_wake_remote);
1225 if (wake_flags & WF_MIGRATED)
1226 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1228 #endif /* CONFIG_SMP */
1230 schedstat_inc(rq, ttwu_count);
1231 schedstat_inc(p, se.statistics.nr_wakeups);
1233 if (wake_flags & WF_SYNC)
1234 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1236 #endif /* CONFIG_SCHEDSTATS */
1239 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1241 activate_task(rq, p, en_flags);
1244 /* if a worker is waking up, notify workqueue */
1245 if (p->flags & PF_WQ_WORKER)
1246 wq_worker_waking_up(p, cpu_of(rq));
1250 * Mark the task runnable and perform wakeup-preemption.
1253 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1255 trace_sched_wakeup(p, true);
1256 check_preempt_curr(rq, p, wake_flags);
1258 p->state = TASK_RUNNING;
1260 if (p->sched_class->task_woken)
1261 p->sched_class->task_woken(rq, p);
1263 if (rq->idle_stamp) {
1264 u64 delta = rq->clock - rq->idle_stamp;
1265 u64 max = 2*sysctl_sched_migration_cost;
1270 update_avg(&rq->avg_idle, delta);
1277 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1280 if (p->sched_contributes_to_load)
1281 rq->nr_uninterruptible--;
1284 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1285 ttwu_do_wakeup(rq, p, wake_flags);
1289 * Called in case the task @p isn't fully descheduled from its runqueue,
1290 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1291 * since all we need to do is flip p->state to TASK_RUNNING, since
1292 * the task is still ->on_rq.
1294 static int ttwu_remote(struct task_struct *p, int wake_flags)
1299 rq = __task_rq_lock(p);
1301 ttwu_do_wakeup(rq, p, wake_flags);
1304 __task_rq_unlock(rq);
1310 static void sched_ttwu_pending(void)
1312 struct rq *rq = this_rq();
1313 struct llist_node *llist = llist_del_all(&rq->wake_list);
1314 struct task_struct *p;
1316 raw_spin_lock(&rq->lock);
1319 p = llist_entry(llist, struct task_struct, wake_entry);
1320 llist = llist_next(llist);
1321 ttwu_do_activate(rq, p, 0);
1324 raw_spin_unlock(&rq->lock);
1327 void scheduler_ipi(void)
1329 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1333 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1334 * traditionally all their work was done from the interrupt return
1335 * path. Now that we actually do some work, we need to make sure
1338 * Some archs already do call them, luckily irq_enter/exit nest
1341 * Arguably we should visit all archs and update all handlers,
1342 * however a fair share of IPIs are still resched only so this would
1343 * somewhat pessimize the simple resched case.
1346 sched_ttwu_pending();
1349 * Check if someone kicked us for doing the nohz idle load balance.
1351 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1352 this_rq()->idle_balance = 1;
1353 raise_softirq_irqoff(SCHED_SOFTIRQ);
1358 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1360 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1361 smp_send_reschedule(cpu);
1364 bool cpus_share_cache(int this_cpu, int that_cpu)
1366 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1368 #endif /* CONFIG_SMP */
1370 static void ttwu_queue(struct task_struct *p, int cpu)
1372 struct rq *rq = cpu_rq(cpu);
1374 #if defined(CONFIG_SMP)
1375 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1376 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1377 ttwu_queue_remote(p, cpu);
1382 raw_spin_lock(&rq->lock);
1383 ttwu_do_activate(rq, p, 0);
1384 raw_spin_unlock(&rq->lock);
1388 * try_to_wake_up - wake up a thread
1389 * @p: the thread to be awakened
1390 * @state: the mask of task states that can be woken
1391 * @wake_flags: wake modifier flags (WF_*)
1393 * Put it on the run-queue if it's not already there. The "current"
1394 * thread is always on the run-queue (except when the actual
1395 * re-schedule is in progress), and as such you're allowed to do
1396 * the simpler "current->state = TASK_RUNNING" to mark yourself
1397 * runnable without the overhead of this.
1399 * Returns %true if @p was woken up, %false if it was already running
1400 * or @state didn't match @p's state.
1403 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1405 unsigned long flags;
1406 int cpu, success = 0;
1409 raw_spin_lock_irqsave(&p->pi_lock, flags);
1410 if (!(p->state & state))
1413 success = 1; /* we're going to change ->state */
1416 if (p->on_rq && ttwu_remote(p, wake_flags))
1421 * If the owning (remote) cpu is still in the middle of schedule() with
1422 * this task as prev, wait until its done referencing the task.
1427 * Pairs with the smp_wmb() in finish_lock_switch().
1431 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1432 p->state = TASK_WAKING;
1434 if (p->sched_class->task_waking)
1435 p->sched_class->task_waking(p);
1437 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1438 if (task_cpu(p) != cpu) {
1439 wake_flags |= WF_MIGRATED;
1440 set_task_cpu(p, cpu);
1442 #endif /* CONFIG_SMP */
1446 ttwu_stat(p, cpu, wake_flags);
1448 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1454 * try_to_wake_up_local - try to wake up a local task with rq lock held
1455 * @p: the thread to be awakened
1457 * Put @p on the run-queue if it's not already there. The caller must
1458 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1461 static void try_to_wake_up_local(struct task_struct *p)
1463 struct rq *rq = task_rq(p);
1465 BUG_ON(rq != this_rq());
1466 BUG_ON(p == current);
1467 lockdep_assert_held(&rq->lock);
1469 if (!raw_spin_trylock(&p->pi_lock)) {
1470 raw_spin_unlock(&rq->lock);
1471 raw_spin_lock(&p->pi_lock);
1472 raw_spin_lock(&rq->lock);
1475 if (!(p->state & TASK_NORMAL))
1479 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1481 ttwu_do_wakeup(rq, p, 0);
1482 ttwu_stat(p, smp_processor_id(), 0);
1484 raw_spin_unlock(&p->pi_lock);
1488 * wake_up_process - Wake up a specific process
1489 * @p: The process to be woken up.
1491 * Attempt to wake up the nominated process and move it to the set of runnable
1492 * processes. Returns 1 if the process was woken up, 0 if it was already
1495 * It may be assumed that this function implies a write memory barrier before
1496 * changing the task state if and only if any tasks are woken up.
1498 int wake_up_process(struct task_struct *p)
1500 return try_to_wake_up(p, TASK_ALL, 0);
1502 EXPORT_SYMBOL(wake_up_process);
1504 int wake_up_state(struct task_struct *p, unsigned int state)
1506 return try_to_wake_up(p, state, 0);
1510 * Perform scheduler related setup for a newly forked process p.
1511 * p is forked by current.
1513 * __sched_fork() is basic setup used by init_idle() too:
1515 static void __sched_fork(struct task_struct *p)
1520 p->se.exec_start = 0;
1521 p->se.sum_exec_runtime = 0;
1522 p->se.prev_sum_exec_runtime = 0;
1523 p->se.nr_migrations = 0;
1525 INIT_LIST_HEAD(&p->se.group_node);
1527 #ifdef CONFIG_SCHEDSTATS
1528 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1531 INIT_LIST_HEAD(&p->rt.run_list);
1533 #ifdef CONFIG_PREEMPT_NOTIFIERS
1534 INIT_HLIST_HEAD(&p->preempt_notifiers);
1539 * fork()/clone()-time setup:
1541 void sched_fork(struct task_struct *p)
1543 unsigned long flags;
1544 int cpu = get_cpu();
1548 * We mark the process as running here. This guarantees that
1549 * nobody will actually run it, and a signal or other external
1550 * event cannot wake it up and insert it on the runqueue either.
1552 p->state = TASK_RUNNING;
1555 * Make sure we do not leak PI boosting priority to the child.
1557 p->prio = current->normal_prio;
1560 * Revert to default priority/policy on fork if requested.
1562 if (unlikely(p->sched_reset_on_fork)) {
1563 if (task_has_rt_policy(p)) {
1564 p->policy = SCHED_NORMAL;
1565 p->static_prio = NICE_TO_PRIO(0);
1567 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1568 p->static_prio = NICE_TO_PRIO(0);
1570 p->prio = p->normal_prio = __normal_prio(p);
1574 * We don't need the reset flag anymore after the fork. It has
1575 * fulfilled its duty:
1577 p->sched_reset_on_fork = 0;
1580 if (!rt_prio(p->prio))
1581 p->sched_class = &fair_sched_class;
1583 if (p->sched_class->task_fork)
1584 p->sched_class->task_fork(p);
1587 * The child is not yet in the pid-hash so no cgroup attach races,
1588 * and the cgroup is pinned to this child due to cgroup_fork()
1589 * is ran before sched_fork().
1591 * Silence PROVE_RCU.
1593 raw_spin_lock_irqsave(&p->pi_lock, flags);
1594 set_task_cpu(p, cpu);
1595 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1597 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1598 if (likely(sched_info_on()))
1599 memset(&p->sched_info, 0, sizeof(p->sched_info));
1601 #if defined(CONFIG_SMP)
1604 #ifdef CONFIG_PREEMPT_COUNT
1605 /* Want to start with kernel preemption disabled. */
1606 task_thread_info(p)->preempt_count = 1;
1609 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1616 * wake_up_new_task - wake up a newly created task for the first time.
1618 * This function will do some initial scheduler statistics housekeeping
1619 * that must be done for every newly created context, then puts the task
1620 * on the runqueue and wakes it.
1622 void wake_up_new_task(struct task_struct *p)
1624 unsigned long flags;
1627 raw_spin_lock_irqsave(&p->pi_lock, flags);
1630 * Fork balancing, do it here and not earlier because:
1631 * - cpus_allowed can change in the fork path
1632 * - any previously selected cpu might disappear through hotplug
1634 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1637 rq = __task_rq_lock(p);
1638 activate_task(rq, p, 0);
1640 trace_sched_wakeup_new(p, true);
1641 check_preempt_curr(rq, p, WF_FORK);
1643 if (p->sched_class->task_woken)
1644 p->sched_class->task_woken(rq, p);
1646 task_rq_unlock(rq, p, &flags);
1649 #ifdef CONFIG_PREEMPT_NOTIFIERS
1652 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1653 * @notifier: notifier struct to register
1655 void preempt_notifier_register(struct preempt_notifier *notifier)
1657 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1659 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1662 * preempt_notifier_unregister - no longer interested in preemption notifications
1663 * @notifier: notifier struct to unregister
1665 * This is safe to call from within a preemption notifier.
1667 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1669 hlist_del(¬ifier->link);
1671 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1673 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1675 struct preempt_notifier *notifier;
1676 struct hlist_node *node;
1678 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1679 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1683 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1684 struct task_struct *next)
1686 struct preempt_notifier *notifier;
1687 struct hlist_node *node;
1689 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1690 notifier->ops->sched_out(notifier, next);
1693 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1695 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1700 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1701 struct task_struct *next)
1705 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1708 * prepare_task_switch - prepare to switch tasks
1709 * @rq: the runqueue preparing to switch
1710 * @prev: the current task that is being switched out
1711 * @next: the task we are going to switch to.
1713 * This is called with the rq lock held and interrupts off. It must
1714 * be paired with a subsequent finish_task_switch after the context
1717 * prepare_task_switch sets up locking and calls architecture specific
1721 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1722 struct task_struct *next)
1724 trace_sched_switch(prev, next);
1725 sched_info_switch(prev, next);
1726 perf_event_task_sched_out(prev, next);
1727 fire_sched_out_preempt_notifiers(prev, next);
1728 prepare_lock_switch(rq, next);
1729 prepare_arch_switch(next);
1733 * finish_task_switch - clean up after a task-switch
1734 * @rq: runqueue associated with task-switch
1735 * @prev: the thread we just switched away from.
1737 * finish_task_switch must be called after the context switch, paired
1738 * with a prepare_task_switch call before the context switch.
1739 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1740 * and do any other architecture-specific cleanup actions.
1742 * Note that we may have delayed dropping an mm in context_switch(). If
1743 * so, we finish that here outside of the runqueue lock. (Doing it
1744 * with the lock held can cause deadlocks; see schedule() for
1747 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1748 __releases(rq->lock)
1750 struct mm_struct *mm = rq->prev_mm;
1756 * A task struct has one reference for the use as "current".
1757 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1758 * schedule one last time. The schedule call will never return, and
1759 * the scheduled task must drop that reference.
1760 * The test for TASK_DEAD must occur while the runqueue locks are
1761 * still held, otherwise prev could be scheduled on another cpu, die
1762 * there before we look at prev->state, and then the reference would
1764 * Manfred Spraul <manfred@colorfullife.com>
1766 prev_state = prev->state;
1767 account_switch_vtime(prev);
1768 finish_arch_switch(prev);
1769 perf_event_task_sched_in(prev, current);
1770 finish_lock_switch(rq, prev);
1771 finish_arch_post_lock_switch();
1773 fire_sched_in_preempt_notifiers(current);
1776 if (unlikely(prev_state == TASK_DEAD)) {
1778 * Remove function-return probe instances associated with this
1779 * task and put them back on the free list.
1781 kprobe_flush_task(prev);
1782 put_task_struct(prev);
1788 /* assumes rq->lock is held */
1789 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1791 if (prev->sched_class->pre_schedule)
1792 prev->sched_class->pre_schedule(rq, prev);
1795 /* rq->lock is NOT held, but preemption is disabled */
1796 static inline void post_schedule(struct rq *rq)
1798 if (rq->post_schedule) {
1799 unsigned long flags;
1801 raw_spin_lock_irqsave(&rq->lock, flags);
1802 if (rq->curr->sched_class->post_schedule)
1803 rq->curr->sched_class->post_schedule(rq);
1804 raw_spin_unlock_irqrestore(&rq->lock, flags);
1806 rq->post_schedule = 0;
1812 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1816 static inline void post_schedule(struct rq *rq)
1823 * schedule_tail - first thing a freshly forked thread must call.
1824 * @prev: the thread we just switched away from.
1826 asmlinkage void schedule_tail(struct task_struct *prev)
1827 __releases(rq->lock)
1829 struct rq *rq = this_rq();
1831 finish_task_switch(rq, prev);
1834 * FIXME: do we need to worry about rq being invalidated by the
1839 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1840 /* In this case, finish_task_switch does not reenable preemption */
1843 if (current->set_child_tid)
1844 put_user(task_pid_vnr(current), current->set_child_tid);
1848 * context_switch - switch to the new MM and the new
1849 * thread's register state.
1852 context_switch(struct rq *rq, struct task_struct *prev,
1853 struct task_struct *next)
1855 struct mm_struct *mm, *oldmm;
1857 prepare_task_switch(rq, prev, next);
1860 oldmm = prev->active_mm;
1862 * For paravirt, this is coupled with an exit in switch_to to
1863 * combine the page table reload and the switch backend into
1866 arch_start_context_switch(prev);
1869 next->active_mm = oldmm;
1870 atomic_inc(&oldmm->mm_count);
1871 enter_lazy_tlb(oldmm, next);
1873 switch_mm(oldmm, mm, next);
1876 prev->active_mm = NULL;
1877 rq->prev_mm = oldmm;
1880 * Since the runqueue lock will be released by the next
1881 * task (which is an invalid locking op but in the case
1882 * of the scheduler it's an obvious special-case), so we
1883 * do an early lockdep release here:
1885 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1886 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1889 /* Here we just switch the register state and the stack. */
1890 switch_to(prev, next, prev);
1894 * this_rq must be evaluated again because prev may have moved
1895 * CPUs since it called schedule(), thus the 'rq' on its stack
1896 * frame will be invalid.
1898 finish_task_switch(this_rq(), prev);
1902 * nr_running, nr_uninterruptible and nr_context_switches:
1904 * externally visible scheduler statistics: current number of runnable
1905 * threads, current number of uninterruptible-sleeping threads, total
1906 * number of context switches performed since bootup.
1908 unsigned long nr_running(void)
1910 unsigned long i, sum = 0;
1912 for_each_online_cpu(i)
1913 sum += cpu_rq(i)->nr_running;
1918 unsigned long nr_uninterruptible(void)
1920 unsigned long i, sum = 0;
1922 for_each_possible_cpu(i)
1923 sum += cpu_rq(i)->nr_uninterruptible;
1926 * Since we read the counters lockless, it might be slightly
1927 * inaccurate. Do not allow it to go below zero though:
1929 if (unlikely((long)sum < 0))
1935 unsigned long long nr_context_switches(void)
1938 unsigned long long sum = 0;
1940 for_each_possible_cpu(i)
1941 sum += cpu_rq(i)->nr_switches;
1946 unsigned long nr_iowait(void)
1948 unsigned long i, sum = 0;
1950 for_each_possible_cpu(i)
1951 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1956 unsigned long nr_iowait_cpu(int cpu)
1958 struct rq *this = cpu_rq(cpu);
1959 return atomic_read(&this->nr_iowait);
1962 unsigned long this_cpu_load(void)
1964 struct rq *this = this_rq();
1965 return this->cpu_load[0];
1970 * Global load-average calculations
1972 * We take a distributed and async approach to calculating the global load-avg
1973 * in order to minimize overhead.
1975 * The global load average is an exponentially decaying average of nr_running +
1976 * nr_uninterruptible.
1978 * Once every LOAD_FREQ:
1981 * for_each_possible_cpu(cpu)
1982 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
1984 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
1986 * Due to a number of reasons the above turns in the mess below:
1988 * - for_each_possible_cpu() is prohibitively expensive on machines with
1989 * serious number of cpus, therefore we need to take a distributed approach
1990 * to calculating nr_active.
1992 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
1993 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
1995 * So assuming nr_active := 0 when we start out -- true per definition, we
1996 * can simply take per-cpu deltas and fold those into a global accumulate
1997 * to obtain the same result. See calc_load_fold_active().
1999 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2000 * across the machine, we assume 10 ticks is sufficient time for every
2001 * cpu to have completed this task.
2003 * This places an upper-bound on the IRQ-off latency of the machine. Then
2004 * again, being late doesn't loose the delta, just wrecks the sample.
2006 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2007 * this would add another cross-cpu cacheline miss and atomic operation
2008 * to the wakeup path. Instead we increment on whatever cpu the task ran
2009 * when it went into uninterruptible state and decrement on whatever cpu
2010 * did the wakeup. This means that only the sum of nr_uninterruptible over
2011 * all cpus yields the correct result.
2013 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2016 /* Variables and functions for calc_load */
2017 static atomic_long_t calc_load_tasks;
2018 static unsigned long calc_load_update;
2019 unsigned long avenrun[3];
2020 EXPORT_SYMBOL(avenrun); /* should be removed */
2023 * get_avenrun - get the load average array
2024 * @loads: pointer to dest load array
2025 * @offset: offset to add
2026 * @shift: shift count to shift the result left
2028 * These values are estimates at best, so no need for locking.
2030 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2032 loads[0] = (avenrun[0] + offset) << shift;
2033 loads[1] = (avenrun[1] + offset) << shift;
2034 loads[2] = (avenrun[2] + offset) << shift;
2037 static long calc_load_fold_active(struct rq *this_rq)
2039 long nr_active, delta = 0;
2041 nr_active = this_rq->nr_running;
2042 nr_active += (long) this_rq->nr_uninterruptible;
2044 if (nr_active != this_rq->calc_load_active) {
2045 delta = nr_active - this_rq->calc_load_active;
2046 this_rq->calc_load_active = nr_active;
2053 * a1 = a0 * e + a * (1 - e)
2055 static unsigned long
2056 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2059 load += active * (FIXED_1 - exp);
2060 load += 1UL << (FSHIFT - 1);
2061 return load >> FSHIFT;
2066 * Handle NO_HZ for the global load-average.
2068 * Since the above described distributed algorithm to compute the global
2069 * load-average relies on per-cpu sampling from the tick, it is affected by
2072 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2073 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2074 * when we read the global state.
2076 * Obviously reality has to ruin such a delightfully simple scheme:
2078 * - When we go NO_HZ idle during the window, we can negate our sample
2079 * contribution, causing under-accounting.
2081 * We avoid this by keeping two idle-delta counters and flipping them
2082 * when the window starts, thus separating old and new NO_HZ load.
2084 * The only trick is the slight shift in index flip for read vs write.
2088 * |-|-----------|-|-----------|-|-----------|-|
2089 * r:0 0 1 1 0 0 1 1 0
2090 * w:0 1 1 0 0 1 1 0 0
2092 * This ensures we'll fold the old idle contribution in this window while
2093 * accumlating the new one.
2095 * - When we wake up from NO_HZ idle during the window, we push up our
2096 * contribution, since we effectively move our sample point to a known
2099 * This is solved by pushing the window forward, and thus skipping the
2100 * sample, for this cpu (effectively using the idle-delta for this cpu which
2101 * was in effect at the time the window opened). This also solves the issue
2102 * of having to deal with a cpu having been in NOHZ idle for multiple
2103 * LOAD_FREQ intervals.
2105 * When making the ILB scale, we should try to pull this in as well.
2107 static atomic_long_t calc_load_idle[2];
2108 static int calc_load_idx;
2110 static inline int calc_load_write_idx(void)
2112 int idx = calc_load_idx;
2115 * See calc_global_nohz(), if we observe the new index, we also
2116 * need to observe the new update time.
2121 * If the folding window started, make sure we start writing in the
2124 if (!time_before(jiffies, calc_load_update))
2130 static inline int calc_load_read_idx(void)
2132 return calc_load_idx & 1;
2135 void calc_load_enter_idle(void)
2137 struct rq *this_rq = this_rq();
2141 * We're going into NOHZ mode, if there's any pending delta, fold it
2142 * into the pending idle delta.
2144 delta = calc_load_fold_active(this_rq);
2146 int idx = calc_load_write_idx();
2147 atomic_long_add(delta, &calc_load_idle[idx]);
2151 void calc_load_exit_idle(void)
2153 struct rq *this_rq = this_rq();
2156 * If we're still before the sample window, we're done.
2158 if (time_before(jiffies, this_rq->calc_load_update))
2162 * We woke inside or after the sample window, this means we're already
2163 * accounted through the nohz accounting, so skip the entire deal and
2164 * sync up for the next window.
2166 this_rq->calc_load_update = calc_load_update;
2167 if (time_before(jiffies, this_rq->calc_load_update + 10))
2168 this_rq->calc_load_update += LOAD_FREQ;
2171 static long calc_load_fold_idle(void)
2173 int idx = calc_load_read_idx();
2176 if (atomic_long_read(&calc_load_idle[idx]))
2177 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2183 * fixed_power_int - compute: x^n, in O(log n) time
2185 * @x: base of the power
2186 * @frac_bits: fractional bits of @x
2187 * @n: power to raise @x to.
2189 * By exploiting the relation between the definition of the natural power
2190 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2191 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2192 * (where: n_i \elem {0, 1}, the binary vector representing n),
2193 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2194 * of course trivially computable in O(log_2 n), the length of our binary
2197 static unsigned long
2198 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2200 unsigned long result = 1UL << frac_bits;
2205 result += 1UL << (frac_bits - 1);
2206 result >>= frac_bits;
2212 x += 1UL << (frac_bits - 1);
2220 * a1 = a0 * e + a * (1 - e)
2222 * a2 = a1 * e + a * (1 - e)
2223 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2224 * = a0 * e^2 + a * (1 - e) * (1 + e)
2226 * a3 = a2 * e + a * (1 - e)
2227 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2228 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2232 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2233 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2234 * = a0 * e^n + a * (1 - e^n)
2236 * [1] application of the geometric series:
2239 * S_n := \Sum x^i = -------------
2242 static unsigned long
2243 calc_load_n(unsigned long load, unsigned long exp,
2244 unsigned long active, unsigned int n)
2247 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2251 * NO_HZ can leave us missing all per-cpu ticks calling
2252 * calc_load_account_active(), but since an idle CPU folds its delta into
2253 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2254 * in the pending idle delta if our idle period crossed a load cycle boundary.
2256 * Once we've updated the global active value, we need to apply the exponential
2257 * weights adjusted to the number of cycles missed.
2259 static void calc_global_nohz(void)
2261 long delta, active, n;
2263 if (!time_before(jiffies, calc_load_update + 10)) {
2265 * Catch-up, fold however many we are behind still
2267 delta = jiffies - calc_load_update - 10;
2268 n = 1 + (delta / LOAD_FREQ);
2270 active = atomic_long_read(&calc_load_tasks);
2271 active = active > 0 ? active * FIXED_1 : 0;
2273 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2274 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2275 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2277 calc_load_update += n * LOAD_FREQ;
2281 * Flip the idle index...
2283 * Make sure we first write the new time then flip the index, so that
2284 * calc_load_write_idx() will see the new time when it reads the new
2285 * index, this avoids a double flip messing things up.
2290 #else /* !CONFIG_NO_HZ */
2292 static inline long calc_load_fold_idle(void) { return 0; }
2293 static inline void calc_global_nohz(void) { }
2295 #endif /* CONFIG_NO_HZ */
2298 * calc_load - update the avenrun load estimates 10 ticks after the
2299 * CPUs have updated calc_load_tasks.
2301 void calc_global_load(unsigned long ticks)
2305 if (time_before(jiffies, calc_load_update + 10))
2309 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2311 delta = calc_load_fold_idle();
2313 atomic_long_add(delta, &calc_load_tasks);
2315 active = atomic_long_read(&calc_load_tasks);
2316 active = active > 0 ? active * FIXED_1 : 0;
2318 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2319 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2320 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2322 calc_load_update += LOAD_FREQ;
2325 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2331 * Called from update_cpu_load() to periodically update this CPU's
2334 static void calc_load_account_active(struct rq *this_rq)
2338 if (time_before(jiffies, this_rq->calc_load_update))
2341 delta = calc_load_fold_active(this_rq);
2343 atomic_long_add(delta, &calc_load_tasks);
2345 this_rq->calc_load_update += LOAD_FREQ;
2349 * End of global load-average stuff
2353 * The exact cpuload at various idx values, calculated at every tick would be
2354 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2356 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2357 * on nth tick when cpu may be busy, then we have:
2358 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2359 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2361 * decay_load_missed() below does efficient calculation of
2362 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2363 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2365 * The calculation is approximated on a 128 point scale.
2366 * degrade_zero_ticks is the number of ticks after which load at any
2367 * particular idx is approximated to be zero.
2368 * degrade_factor is a precomputed table, a row for each load idx.
2369 * Each column corresponds to degradation factor for a power of two ticks,
2370 * based on 128 point scale.
2372 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2373 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2375 * With this power of 2 load factors, we can degrade the load n times
2376 * by looking at 1 bits in n and doing as many mult/shift instead of
2377 * n mult/shifts needed by the exact degradation.
2379 #define DEGRADE_SHIFT 7
2380 static const unsigned char
2381 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2382 static const unsigned char
2383 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2384 {0, 0, 0, 0, 0, 0, 0, 0},
2385 {64, 32, 8, 0, 0, 0, 0, 0},
2386 {96, 72, 40, 12, 1, 0, 0},
2387 {112, 98, 75, 43, 15, 1, 0},
2388 {120, 112, 98, 76, 45, 16, 2} };
2391 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2392 * would be when CPU is idle and so we just decay the old load without
2393 * adding any new load.
2395 static unsigned long
2396 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2400 if (!missed_updates)
2403 if (missed_updates >= degrade_zero_ticks[idx])
2407 return load >> missed_updates;
2409 while (missed_updates) {
2410 if (missed_updates % 2)
2411 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2413 missed_updates >>= 1;
2420 * Update rq->cpu_load[] statistics. This function is usually called every
2421 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2422 * every tick. We fix it up based on jiffies.
2424 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2425 unsigned long pending_updates)
2429 this_rq->nr_load_updates++;
2431 /* Update our load: */
2432 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2433 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2434 unsigned long old_load, new_load;
2436 /* scale is effectively 1 << i now, and >> i divides by scale */
2438 old_load = this_rq->cpu_load[i];
2439 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2440 new_load = this_load;
2442 * Round up the averaging division if load is increasing. This
2443 * prevents us from getting stuck on 9 if the load is 10, for
2446 if (new_load > old_load)
2447 new_load += scale - 1;
2449 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2452 sched_avg_update(this_rq);
2457 * There is no sane way to deal with nohz on smp when using jiffies because the
2458 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2459 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2461 * Therefore we cannot use the delta approach from the regular tick since that
2462 * would seriously skew the load calculation. However we'll make do for those
2463 * updates happening while idle (nohz_idle_balance) or coming out of idle
2464 * (tick_nohz_idle_exit).
2466 * This means we might still be one tick off for nohz periods.
2470 * Called from nohz_idle_balance() to update the load ratings before doing the
2473 void update_idle_cpu_load(struct rq *this_rq)
2475 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2476 unsigned long load = this_rq->load.weight;
2477 unsigned long pending_updates;
2480 * bail if there's load or we're actually up-to-date.
2482 if (load || curr_jiffies == this_rq->last_load_update_tick)
2485 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2486 this_rq->last_load_update_tick = curr_jiffies;
2488 __update_cpu_load(this_rq, load, pending_updates);
2492 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2494 void update_cpu_load_nohz(void)
2496 struct rq *this_rq = this_rq();
2497 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2498 unsigned long pending_updates;
2500 if (curr_jiffies == this_rq->last_load_update_tick)
2503 raw_spin_lock(&this_rq->lock);
2504 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2505 if (pending_updates) {
2506 this_rq->last_load_update_tick = curr_jiffies;
2508 * We were idle, this means load 0, the current load might be
2509 * !0 due to remote wakeups and the sort.
2511 __update_cpu_load(this_rq, 0, pending_updates);
2513 raw_spin_unlock(&this_rq->lock);
2515 #endif /* CONFIG_NO_HZ */
2518 * Called from scheduler_tick()
2520 static void update_cpu_load_active(struct rq *this_rq)
2523 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2525 this_rq->last_load_update_tick = jiffies;
2526 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2528 calc_load_account_active(this_rq);
2534 * sched_exec - execve() is a valuable balancing opportunity, because at
2535 * this point the task has the smallest effective memory and cache footprint.
2537 void sched_exec(void)
2539 struct task_struct *p = current;
2540 unsigned long flags;
2543 raw_spin_lock_irqsave(&p->pi_lock, flags);
2544 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2545 if (dest_cpu == smp_processor_id())
2548 if (likely(cpu_active(dest_cpu))) {
2549 struct migration_arg arg = { p, dest_cpu };
2551 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2552 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2556 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2561 DEFINE_PER_CPU(struct kernel_stat, kstat);
2562 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2564 EXPORT_PER_CPU_SYMBOL(kstat);
2565 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2568 * Return any ns on the sched_clock that have not yet been accounted in
2569 * @p in case that task is currently running.
2571 * Called with task_rq_lock() held on @rq.
2573 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2577 if (task_current(rq, p)) {
2578 update_rq_clock(rq);
2579 ns = rq->clock_task - p->se.exec_start;
2587 unsigned long long task_delta_exec(struct task_struct *p)
2589 unsigned long flags;
2593 rq = task_rq_lock(p, &flags);
2594 ns = do_task_delta_exec(p, rq);
2595 task_rq_unlock(rq, p, &flags);
2601 * Return accounted runtime for the task.
2602 * In case the task is currently running, return the runtime plus current's
2603 * pending runtime that have not been accounted yet.
2605 unsigned long long task_sched_runtime(struct task_struct *p)
2607 unsigned long flags;
2611 rq = task_rq_lock(p, &flags);
2612 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2613 task_rq_unlock(rq, p, &flags);
2619 * This function gets called by the timer code, with HZ frequency.
2620 * We call it with interrupts disabled.
2622 void scheduler_tick(void)
2624 int cpu = smp_processor_id();
2625 struct rq *rq = cpu_rq(cpu);
2626 struct task_struct *curr = rq->curr;
2630 raw_spin_lock(&rq->lock);
2631 update_rq_clock(rq);
2632 update_cpu_load_active(rq);
2633 curr->sched_class->task_tick(rq, curr, 0);
2634 raw_spin_unlock(&rq->lock);
2636 perf_event_task_tick();
2639 rq->idle_balance = idle_cpu(cpu);
2640 trigger_load_balance(rq, cpu);
2644 notrace unsigned long get_parent_ip(unsigned long addr)
2646 if (in_lock_functions(addr)) {
2647 addr = CALLER_ADDR2;
2648 if (in_lock_functions(addr))
2649 addr = CALLER_ADDR3;
2654 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2655 defined(CONFIG_PREEMPT_TRACER))
2657 void __kprobes add_preempt_count(int val)
2659 #ifdef CONFIG_DEBUG_PREEMPT
2663 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2666 preempt_count() += val;
2667 #ifdef CONFIG_DEBUG_PREEMPT
2669 * Spinlock count overflowing soon?
2671 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2674 if (preempt_count() == val)
2675 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2677 EXPORT_SYMBOL(add_preempt_count);
2679 void __kprobes sub_preempt_count(int val)
2681 #ifdef CONFIG_DEBUG_PREEMPT
2685 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2688 * Is the spinlock portion underflowing?
2690 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2691 !(preempt_count() & PREEMPT_MASK)))
2695 if (preempt_count() == val)
2696 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2697 preempt_count() -= val;
2699 EXPORT_SYMBOL(sub_preempt_count);
2704 * Print scheduling while atomic bug:
2706 static noinline void __schedule_bug(struct task_struct *prev)
2708 if (oops_in_progress)
2711 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2712 prev->comm, prev->pid, preempt_count());
2714 debug_show_held_locks(prev);
2716 if (irqs_disabled())
2717 print_irqtrace_events(prev);
2719 add_taint(TAINT_WARN);
2723 * Various schedule()-time debugging checks and statistics:
2725 static inline void schedule_debug(struct task_struct *prev)
2728 * Test if we are atomic. Since do_exit() needs to call into
2729 * schedule() atomically, we ignore that path for now.
2730 * Otherwise, whine if we are scheduling when we should not be.
2732 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2733 __schedule_bug(prev);
2736 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2738 schedstat_inc(this_rq(), sched_count);
2741 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2743 if (prev->on_rq || rq->skip_clock_update < 0)
2744 update_rq_clock(rq);
2745 prev->sched_class->put_prev_task(rq, prev);
2749 * Pick up the highest-prio task:
2751 static inline struct task_struct *
2752 pick_next_task(struct rq *rq)
2754 const struct sched_class *class;
2755 struct task_struct *p;
2758 * Optimization: we know that if all tasks are in
2759 * the fair class we can call that function directly:
2761 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2762 p = fair_sched_class.pick_next_task(rq);
2767 for_each_class(class) {
2768 p = class->pick_next_task(rq);
2773 BUG(); /* the idle class will always have a runnable task */
2777 * __schedule() is the main scheduler function.
2779 * The main means of driving the scheduler and thus entering this function are:
2781 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2783 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2784 * paths. For example, see arch/x86/entry_64.S.
2786 * To drive preemption between tasks, the scheduler sets the flag in timer
2787 * interrupt handler scheduler_tick().
2789 * 3. Wakeups don't really cause entry into schedule(). They add a
2790 * task to the run-queue and that's it.
2792 * Now, if the new task added to the run-queue preempts the current
2793 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2794 * called on the nearest possible occasion:
2796 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2798 * - in syscall or exception context, at the next outmost
2799 * preempt_enable(). (this might be as soon as the wake_up()'s
2802 * - in IRQ context, return from interrupt-handler to
2803 * preemptible context
2805 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2808 * - cond_resched() call
2809 * - explicit schedule() call
2810 * - return from syscall or exception to user-space
2811 * - return from interrupt-handler to user-space
2813 static void __sched __schedule(void)
2815 struct task_struct *prev, *next;
2816 unsigned long *switch_count;
2822 cpu = smp_processor_id();
2824 rcu_note_context_switch(cpu);
2827 schedule_debug(prev);
2829 if (sched_feat(HRTICK))
2832 raw_spin_lock_irq(&rq->lock);
2834 switch_count = &prev->nivcsw;
2835 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2836 if (unlikely(signal_pending_state(prev->state, prev))) {
2837 prev->state = TASK_RUNNING;
2839 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2843 * If a worker went to sleep, notify and ask workqueue
2844 * whether it wants to wake up a task to maintain
2847 if (prev->flags & PF_WQ_WORKER) {
2848 struct task_struct *to_wakeup;
2850 to_wakeup = wq_worker_sleeping(prev, cpu);
2852 try_to_wake_up_local(to_wakeup);
2855 switch_count = &prev->nvcsw;
2858 pre_schedule(rq, prev);
2860 if (unlikely(!rq->nr_running))
2861 idle_balance(cpu, rq);
2863 put_prev_task(rq, prev);
2864 next = pick_next_task(rq);
2865 clear_tsk_need_resched(prev);
2866 rq->skip_clock_update = 0;
2868 if (likely(prev != next)) {
2873 context_switch(rq, prev, next); /* unlocks the rq */
2875 * The context switch have flipped the stack from under us
2876 * and restored the local variables which were saved when
2877 * this task called schedule() in the past. prev == current
2878 * is still correct, but it can be moved to another cpu/rq.
2880 cpu = smp_processor_id();
2883 raw_spin_unlock_irq(&rq->lock);
2887 sched_preempt_enable_no_resched();
2892 static inline void sched_submit_work(struct task_struct *tsk)
2894 if (!tsk->state || tsk_is_pi_blocked(tsk))
2897 * If we are going to sleep and we have plugged IO queued,
2898 * make sure to submit it to avoid deadlocks.
2900 if (blk_needs_flush_plug(tsk))
2901 blk_schedule_flush_plug(tsk);
2904 asmlinkage void __sched schedule(void)
2906 struct task_struct *tsk = current;
2908 sched_submit_work(tsk);
2911 EXPORT_SYMBOL(schedule);
2914 * schedule_preempt_disabled - called with preemption disabled
2916 * Returns with preemption disabled. Note: preempt_count must be 1
2918 void __sched schedule_preempt_disabled(void)
2920 sched_preempt_enable_no_resched();
2925 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2927 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2929 if (lock->owner != owner)
2933 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2934 * lock->owner still matches owner, if that fails, owner might
2935 * point to free()d memory, if it still matches, the rcu_read_lock()
2936 * ensures the memory stays valid.
2940 return owner->on_cpu;
2944 * Look out! "owner" is an entirely speculative pointer
2945 * access and not reliable.
2947 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
2949 if (!sched_feat(OWNER_SPIN))
2953 while (owner_running(lock, owner)) {
2957 arch_mutex_cpu_relax();
2962 * We break out the loop above on need_resched() and when the
2963 * owner changed, which is a sign for heavy contention. Return
2964 * success only when lock->owner is NULL.
2966 return lock->owner == NULL;
2970 #ifdef CONFIG_PREEMPT
2972 * this is the entry point to schedule() from in-kernel preemption
2973 * off of preempt_enable. Kernel preemptions off return from interrupt
2974 * occur there and call schedule directly.
2976 asmlinkage void __sched notrace preempt_schedule(void)
2978 struct thread_info *ti = current_thread_info();
2981 * If there is a non-zero preempt_count or interrupts are disabled,
2982 * we do not want to preempt the current task. Just return..
2984 if (likely(ti->preempt_count || irqs_disabled()))
2988 add_preempt_count_notrace(PREEMPT_ACTIVE);
2990 sub_preempt_count_notrace(PREEMPT_ACTIVE);
2993 * Check again in case we missed a preemption opportunity
2994 * between schedule and now.
2997 } while (need_resched());
2999 EXPORT_SYMBOL(preempt_schedule);
3002 * this is the entry point to schedule() from kernel preemption
3003 * off of irq context.
3004 * Note, that this is called and return with irqs disabled. This will
3005 * protect us against recursive calling from irq.
3007 asmlinkage void __sched preempt_schedule_irq(void)
3009 struct thread_info *ti = current_thread_info();
3011 /* Catch callers which need to be fixed */
3012 BUG_ON(ti->preempt_count || !irqs_disabled());
3015 add_preempt_count(PREEMPT_ACTIVE);
3018 local_irq_disable();
3019 sub_preempt_count(PREEMPT_ACTIVE);
3022 * Check again in case we missed a preemption opportunity
3023 * between schedule and now.
3026 } while (need_resched());
3029 #endif /* CONFIG_PREEMPT */
3031 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3034 return try_to_wake_up(curr->private, mode, wake_flags);
3036 EXPORT_SYMBOL(default_wake_function);
3039 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3040 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3041 * number) then we wake all the non-exclusive tasks and one exclusive task.
3043 * There are circumstances in which we can try to wake a task which has already
3044 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3045 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3047 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3048 int nr_exclusive, int wake_flags, void *key)
3050 wait_queue_t *curr, *next;
3052 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3053 unsigned flags = curr->flags;
3055 if (curr->func(curr, mode, wake_flags, key) &&
3056 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3062 * __wake_up - wake up threads blocked on a waitqueue.
3064 * @mode: which threads
3065 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3066 * @key: is directly passed to the wakeup function
3068 * It may be assumed that this function implies a write memory barrier before
3069 * changing the task state if and only if any tasks are woken up.
3071 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3072 int nr_exclusive, void *key)
3074 unsigned long flags;
3076 spin_lock_irqsave(&q->lock, flags);
3077 __wake_up_common(q, mode, nr_exclusive, 0, key);
3078 spin_unlock_irqrestore(&q->lock, flags);
3080 EXPORT_SYMBOL(__wake_up);
3083 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3085 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3087 __wake_up_common(q, mode, nr, 0, NULL);
3089 EXPORT_SYMBOL_GPL(__wake_up_locked);
3091 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3093 __wake_up_common(q, mode, 1, 0, key);
3095 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3098 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3100 * @mode: which threads
3101 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3102 * @key: opaque value to be passed to wakeup targets
3104 * The sync wakeup differs that the waker knows that it will schedule
3105 * away soon, so while the target thread will be woken up, it will not
3106 * be migrated to another CPU - ie. the two threads are 'synchronized'
3107 * with each other. This can prevent needless bouncing between CPUs.
3109 * On UP it can prevent extra preemption.
3111 * It may be assumed that this function implies a write memory barrier before
3112 * changing the task state if and only if any tasks are woken up.
3114 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3115 int nr_exclusive, void *key)
3117 unsigned long flags;
3118 int wake_flags = WF_SYNC;
3123 if (unlikely(!nr_exclusive))
3126 spin_lock_irqsave(&q->lock, flags);
3127 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3128 spin_unlock_irqrestore(&q->lock, flags);
3130 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3133 * __wake_up_sync - see __wake_up_sync_key()
3135 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3137 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3139 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3142 * complete: - signals a single thread waiting on this completion
3143 * @x: holds the state of this particular completion
3145 * This will wake up a single thread waiting on this completion. Threads will be
3146 * awakened in the same order in which they were queued.
3148 * See also complete_all(), wait_for_completion() and related routines.
3150 * It may be assumed that this function implies a write memory barrier before
3151 * changing the task state if and only if any tasks are woken up.
3153 void complete(struct completion *x)
3155 unsigned long flags;
3157 spin_lock_irqsave(&x->wait.lock, flags);
3159 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3160 spin_unlock_irqrestore(&x->wait.lock, flags);
3162 EXPORT_SYMBOL(complete);
3165 * complete_all: - signals all threads waiting on this completion
3166 * @x: holds the state of this particular completion
3168 * This will wake up all threads waiting on this particular completion event.
3170 * It may be assumed that this function implies a write memory barrier before
3171 * changing the task state if and only if any tasks are woken up.
3173 void complete_all(struct completion *x)
3175 unsigned long flags;
3177 spin_lock_irqsave(&x->wait.lock, flags);
3178 x->done += UINT_MAX/2;
3179 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3180 spin_unlock_irqrestore(&x->wait.lock, flags);
3182 EXPORT_SYMBOL(complete_all);
3184 static inline long __sched
3185 do_wait_for_common(struct completion *x, long timeout, int state)
3188 DECLARE_WAITQUEUE(wait, current);
3190 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3192 if (signal_pending_state(state, current)) {
3193 timeout = -ERESTARTSYS;
3196 __set_current_state(state);
3197 spin_unlock_irq(&x->wait.lock);
3198 timeout = schedule_timeout(timeout);
3199 spin_lock_irq(&x->wait.lock);
3200 } while (!x->done && timeout);
3201 __remove_wait_queue(&x->wait, &wait);
3206 return timeout ?: 1;
3210 wait_for_common(struct completion *x, long timeout, int state)
3214 spin_lock_irq(&x->wait.lock);
3215 timeout = do_wait_for_common(x, timeout, state);
3216 spin_unlock_irq(&x->wait.lock);
3221 * wait_for_completion: - waits for completion of a task
3222 * @x: holds the state of this particular completion
3224 * This waits to be signaled for completion of a specific task. It is NOT
3225 * interruptible and there is no timeout.
3227 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3228 * and interrupt capability. Also see complete().
3230 void __sched wait_for_completion(struct completion *x)
3232 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3234 EXPORT_SYMBOL(wait_for_completion);
3237 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3238 * @x: holds the state of this particular completion
3239 * @timeout: timeout value in jiffies
3241 * This waits for either a completion of a specific task to be signaled or for a
3242 * specified timeout to expire. The timeout is in jiffies. It is not
3245 * The return value is 0 if timed out, and positive (at least 1, or number of
3246 * jiffies left till timeout) if completed.
3248 unsigned long __sched
3249 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3251 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3253 EXPORT_SYMBOL(wait_for_completion_timeout);
3256 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3257 * @x: holds the state of this particular completion
3259 * This waits for completion of a specific task to be signaled. It is
3262 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3264 int __sched wait_for_completion_interruptible(struct completion *x)
3266 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3267 if (t == -ERESTARTSYS)
3271 EXPORT_SYMBOL(wait_for_completion_interruptible);
3274 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3275 * @x: holds the state of this particular completion
3276 * @timeout: timeout value in jiffies
3278 * This waits for either a completion of a specific task to be signaled or for a
3279 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3281 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3282 * positive (at least 1, or number of jiffies left till timeout) if completed.
3285 wait_for_completion_interruptible_timeout(struct completion *x,
3286 unsigned long timeout)
3288 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3290 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3293 * wait_for_completion_killable: - waits for completion of a task (killable)
3294 * @x: holds the state of this particular completion
3296 * This waits to be signaled for completion of a specific task. It can be
3297 * interrupted by a kill signal.
3299 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3301 int __sched wait_for_completion_killable(struct completion *x)
3303 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3304 if (t == -ERESTARTSYS)
3308 EXPORT_SYMBOL(wait_for_completion_killable);
3311 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3312 * @x: holds the state of this particular completion
3313 * @timeout: timeout value in jiffies
3315 * This waits for either a completion of a specific task to be
3316 * signaled or for a specified timeout to expire. It can be
3317 * interrupted by a kill signal. The timeout is in jiffies.
3319 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3320 * positive (at least 1, or number of jiffies left till timeout) if completed.
3323 wait_for_completion_killable_timeout(struct completion *x,
3324 unsigned long timeout)
3326 return wait_for_common(x, timeout, TASK_KILLABLE);
3328 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3331 * try_wait_for_completion - try to decrement a completion without blocking
3332 * @x: completion structure
3334 * Returns: 0 if a decrement cannot be done without blocking
3335 * 1 if a decrement succeeded.
3337 * If a completion is being used as a counting completion,
3338 * attempt to decrement the counter without blocking. This
3339 * enables us to avoid waiting if the resource the completion
3340 * is protecting is not available.
3342 bool try_wait_for_completion(struct completion *x)
3344 unsigned long flags;
3347 spin_lock_irqsave(&x->wait.lock, flags);
3352 spin_unlock_irqrestore(&x->wait.lock, flags);
3355 EXPORT_SYMBOL(try_wait_for_completion);
3358 * completion_done - Test to see if a completion has any waiters
3359 * @x: completion structure
3361 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3362 * 1 if there are no waiters.
3365 bool completion_done(struct completion *x)
3367 unsigned long flags;
3370 spin_lock_irqsave(&x->wait.lock, flags);
3373 spin_unlock_irqrestore(&x->wait.lock, flags);
3376 EXPORT_SYMBOL(completion_done);
3379 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3381 unsigned long flags;
3384 init_waitqueue_entry(&wait, current);
3386 __set_current_state(state);
3388 spin_lock_irqsave(&q->lock, flags);
3389 __add_wait_queue(q, &wait);
3390 spin_unlock(&q->lock);
3391 timeout = schedule_timeout(timeout);
3392 spin_lock_irq(&q->lock);
3393 __remove_wait_queue(q, &wait);
3394 spin_unlock_irqrestore(&q->lock, flags);
3399 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3401 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3403 EXPORT_SYMBOL(interruptible_sleep_on);
3406 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3408 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3410 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3412 void __sched sleep_on(wait_queue_head_t *q)
3414 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3416 EXPORT_SYMBOL(sleep_on);
3418 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3420 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3422 EXPORT_SYMBOL(sleep_on_timeout);
3424 #ifdef CONFIG_RT_MUTEXES
3427 * rt_mutex_setprio - set the current priority of a task
3429 * @prio: prio value (kernel-internal form)
3431 * This function changes the 'effective' priority of a task. It does
3432 * not touch ->normal_prio like __setscheduler().
3434 * Used by the rt_mutex code to implement priority inheritance logic.
3436 void rt_mutex_setprio(struct task_struct *p, int prio)
3438 int oldprio, on_rq, running;
3440 const struct sched_class *prev_class;
3442 BUG_ON(prio < 0 || prio > MAX_PRIO);
3444 rq = __task_rq_lock(p);
3447 * Idle task boosting is a nono in general. There is one
3448 * exception, when PREEMPT_RT and NOHZ is active:
3450 * The idle task calls get_next_timer_interrupt() and holds
3451 * the timer wheel base->lock on the CPU and another CPU wants
3452 * to access the timer (probably to cancel it). We can safely
3453 * ignore the boosting request, as the idle CPU runs this code
3454 * with interrupts disabled and will complete the lock
3455 * protected section without being interrupted. So there is no
3456 * real need to boost.
3458 if (unlikely(p == rq->idle)) {
3459 WARN_ON(p != rq->curr);
3460 WARN_ON(p->pi_blocked_on);
3464 trace_sched_pi_setprio(p, prio);
3466 prev_class = p->sched_class;
3468 running = task_current(rq, p);
3470 dequeue_task(rq, p, 0);
3472 p->sched_class->put_prev_task(rq, p);
3475 p->sched_class = &rt_sched_class;
3477 p->sched_class = &fair_sched_class;
3482 p->sched_class->set_curr_task(rq);
3484 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3486 check_class_changed(rq, p, prev_class, oldprio);
3488 __task_rq_unlock(rq);
3491 void set_user_nice(struct task_struct *p, long nice)
3493 int old_prio, delta, on_rq;
3494 unsigned long flags;
3497 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3500 * We have to be careful, if called from sys_setpriority(),
3501 * the task might be in the middle of scheduling on another CPU.
3503 rq = task_rq_lock(p, &flags);
3505 * The RT priorities are set via sched_setscheduler(), but we still
3506 * allow the 'normal' nice value to be set - but as expected
3507 * it wont have any effect on scheduling until the task is
3508 * SCHED_FIFO/SCHED_RR:
3510 if (task_has_rt_policy(p)) {
3511 p->static_prio = NICE_TO_PRIO(nice);
3516 dequeue_task(rq, p, 0);
3518 p->static_prio = NICE_TO_PRIO(nice);
3521 p->prio = effective_prio(p);
3522 delta = p->prio - old_prio;
3525 enqueue_task(rq, p, 0);
3527 * If the task increased its priority or is running and
3528 * lowered its priority, then reschedule its CPU:
3530 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3531 resched_task(rq->curr);
3534 task_rq_unlock(rq, p, &flags);
3536 EXPORT_SYMBOL(set_user_nice);
3539 * can_nice - check if a task can reduce its nice value
3543 int can_nice(const struct task_struct *p, const int nice)
3545 /* convert nice value [19,-20] to rlimit style value [1,40] */
3546 int nice_rlim = 20 - nice;
3548 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3549 capable(CAP_SYS_NICE));
3552 #ifdef __ARCH_WANT_SYS_NICE
3555 * sys_nice - change the priority of the current process.
3556 * @increment: priority increment
3558 * sys_setpriority is a more generic, but much slower function that
3559 * does similar things.
3561 SYSCALL_DEFINE1(nice, int, increment)
3566 * Setpriority might change our priority at the same moment.
3567 * We don't have to worry. Conceptually one call occurs first
3568 * and we have a single winner.
3570 if (increment < -40)
3575 nice = TASK_NICE(current) + increment;
3581 if (increment < 0 && !can_nice(current, nice))
3584 retval = security_task_setnice(current, nice);
3588 set_user_nice(current, nice);
3595 * task_prio - return the priority value of a given task.
3596 * @p: the task in question.
3598 * This is the priority value as seen by users in /proc.
3599 * RT tasks are offset by -200. Normal tasks are centered
3600 * around 0, value goes from -16 to +15.
3602 int task_prio(const struct task_struct *p)
3604 return p->prio - MAX_RT_PRIO;
3608 * task_nice - return the nice value of a given task.
3609 * @p: the task in question.
3611 int task_nice(const struct task_struct *p)
3613 return TASK_NICE(p);
3615 EXPORT_SYMBOL(task_nice);
3618 * idle_cpu - is a given cpu idle currently?
3619 * @cpu: the processor in question.
3621 int idle_cpu(int cpu)
3623 struct rq *rq = cpu_rq(cpu);
3625 if (rq->curr != rq->idle)
3632 if (!llist_empty(&rq->wake_list))
3640 * idle_task - return the idle task for a given cpu.
3641 * @cpu: the processor in question.
3643 struct task_struct *idle_task(int cpu)
3645 return cpu_rq(cpu)->idle;
3649 * find_process_by_pid - find a process with a matching PID value.
3650 * @pid: the pid in question.
3652 static struct task_struct *find_process_by_pid(pid_t pid)
3654 return pid ? find_task_by_vpid(pid) : current;
3657 /* Actually do priority change: must hold rq lock. */
3659 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3662 p->rt_priority = prio;
3663 p->normal_prio = normal_prio(p);
3664 /* we are holding p->pi_lock already */
3665 p->prio = rt_mutex_getprio(p);
3666 if (rt_prio(p->prio))
3667 p->sched_class = &rt_sched_class;
3669 p->sched_class = &fair_sched_class;
3674 * check the target process has a UID that matches the current process's
3676 static bool check_same_owner(struct task_struct *p)
3678 const struct cred *cred = current_cred(), *pcred;
3682 pcred = __task_cred(p);
3683 match = (uid_eq(cred->euid, pcred->euid) ||
3684 uid_eq(cred->euid, pcred->uid));
3689 static int __sched_setscheduler(struct task_struct *p, int policy,
3690 const struct sched_param *param, bool user)
3692 int retval, oldprio, oldpolicy = -1, on_rq, running;
3693 unsigned long flags;
3694 const struct sched_class *prev_class;
3698 /* may grab non-irq protected spin_locks */
3699 BUG_ON(in_interrupt());
3701 /* double check policy once rq lock held */
3703 reset_on_fork = p->sched_reset_on_fork;
3704 policy = oldpolicy = p->policy;
3706 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3707 policy &= ~SCHED_RESET_ON_FORK;
3709 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3710 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3711 policy != SCHED_IDLE)
3716 * Valid priorities for SCHED_FIFO and SCHED_RR are
3717 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3718 * SCHED_BATCH and SCHED_IDLE is 0.
3720 if (param->sched_priority < 0 ||
3721 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3722 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3724 if (rt_policy(policy) != (param->sched_priority != 0))
3728 * Allow unprivileged RT tasks to decrease priority:
3730 if (user && !capable(CAP_SYS_NICE)) {
3731 if (rt_policy(policy)) {
3732 unsigned long rlim_rtprio =
3733 task_rlimit(p, RLIMIT_RTPRIO);
3735 /* can't set/change the rt policy */
3736 if (policy != p->policy && !rlim_rtprio)
3739 /* can't increase priority */
3740 if (param->sched_priority > p->rt_priority &&
3741 param->sched_priority > rlim_rtprio)
3746 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3747 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3749 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3750 if (!can_nice(p, TASK_NICE(p)))
3754 /* can't change other user's priorities */
3755 if (!check_same_owner(p))
3758 /* Normal users shall not reset the sched_reset_on_fork flag */
3759 if (p->sched_reset_on_fork && !reset_on_fork)
3764 retval = security_task_setscheduler(p);
3770 * make sure no PI-waiters arrive (or leave) while we are
3771 * changing the priority of the task:
3773 * To be able to change p->policy safely, the appropriate
3774 * runqueue lock must be held.
3776 rq = task_rq_lock(p, &flags);
3779 * Changing the policy of the stop threads its a very bad idea
3781 if (p == rq->stop) {
3782 task_rq_unlock(rq, p, &flags);
3787 * If not changing anything there's no need to proceed further:
3789 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3790 param->sched_priority == p->rt_priority))) {
3791 task_rq_unlock(rq, p, &flags);
3795 #ifdef CONFIG_RT_GROUP_SCHED
3798 * Do not allow realtime tasks into groups that have no runtime
3801 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3802 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3803 !task_group_is_autogroup(task_group(p))) {
3804 task_rq_unlock(rq, p, &flags);
3810 /* recheck policy now with rq lock held */
3811 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3812 policy = oldpolicy = -1;
3813 task_rq_unlock(rq, p, &flags);
3817 running = task_current(rq, p);
3819 dequeue_task(rq, p, 0);
3821 p->sched_class->put_prev_task(rq, p);
3823 p->sched_reset_on_fork = reset_on_fork;
3826 prev_class = p->sched_class;
3827 __setscheduler(rq, p, policy, param->sched_priority);
3830 p->sched_class->set_curr_task(rq);
3832 enqueue_task(rq, p, 0);
3834 check_class_changed(rq, p, prev_class, oldprio);
3835 task_rq_unlock(rq, p, &flags);
3837 rt_mutex_adjust_pi(p);
3843 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3844 * @p: the task in question.
3845 * @policy: new policy.
3846 * @param: structure containing the new RT priority.
3848 * NOTE that the task may be already dead.
3850 int sched_setscheduler(struct task_struct *p, int policy,
3851 const struct sched_param *param)
3853 return __sched_setscheduler(p, policy, param, true);
3855 EXPORT_SYMBOL_GPL(sched_setscheduler);
3858 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3859 * @p: the task in question.
3860 * @policy: new policy.
3861 * @param: structure containing the new RT priority.
3863 * Just like sched_setscheduler, only don't bother checking if the
3864 * current context has permission. For example, this is needed in
3865 * stop_machine(): we create temporary high priority worker threads,
3866 * but our caller might not have that capability.
3868 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3869 const struct sched_param *param)
3871 return __sched_setscheduler(p, policy, param, false);
3875 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3877 struct sched_param lparam;
3878 struct task_struct *p;
3881 if (!param || pid < 0)
3883 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3888 p = find_process_by_pid(pid);
3890 retval = sched_setscheduler(p, policy, &lparam);
3897 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3898 * @pid: the pid in question.
3899 * @policy: new policy.
3900 * @param: structure containing the new RT priority.
3902 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3903 struct sched_param __user *, param)
3905 /* negative values for policy are not valid */
3909 return do_sched_setscheduler(pid, policy, param);
3913 * sys_sched_setparam - set/change the RT priority of a thread
3914 * @pid: the pid in question.
3915 * @param: structure containing the new RT priority.
3917 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3919 return do_sched_setscheduler(pid, -1, param);
3923 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3924 * @pid: the pid in question.
3926 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3928 struct task_struct *p;
3936 p = find_process_by_pid(pid);
3938 retval = security_task_getscheduler(p);
3941 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3948 * sys_sched_getparam - get the RT priority of a thread
3949 * @pid: the pid in question.
3950 * @param: structure containing the RT priority.
3952 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3954 struct sched_param lp;
3955 struct task_struct *p;
3958 if (!param || pid < 0)
3962 p = find_process_by_pid(pid);
3967 retval = security_task_getscheduler(p);
3971 lp.sched_priority = p->rt_priority;
3975 * This one might sleep, we cannot do it with a spinlock held ...
3977 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3986 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3988 cpumask_var_t cpus_allowed, new_mask;
3989 struct task_struct *p;
3995 p = find_process_by_pid(pid);
4002 /* Prevent p going away */
4006 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4010 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4012 goto out_free_cpus_allowed;
4015 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4018 retval = security_task_setscheduler(p);
4022 cpuset_cpus_allowed(p, cpus_allowed);
4023 cpumask_and(new_mask, in_mask, cpus_allowed);
4025 retval = set_cpus_allowed_ptr(p, new_mask);
4028 cpuset_cpus_allowed(p, cpus_allowed);
4029 if (!cpumask_subset(new_mask, cpus_allowed)) {
4031 * We must have raced with a concurrent cpuset
4032 * update. Just reset the cpus_allowed to the
4033 * cpuset's cpus_allowed
4035 cpumask_copy(new_mask, cpus_allowed);
4040 free_cpumask_var(new_mask);
4041 out_free_cpus_allowed:
4042 free_cpumask_var(cpus_allowed);
4049 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4050 struct cpumask *new_mask)
4052 if (len < cpumask_size())
4053 cpumask_clear(new_mask);
4054 else if (len > cpumask_size())
4055 len = cpumask_size();
4057 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4061 * sys_sched_setaffinity - set the cpu affinity of a process
4062 * @pid: pid of the process
4063 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4064 * @user_mask_ptr: user-space pointer to the new cpu mask
4066 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4067 unsigned long __user *, user_mask_ptr)
4069 cpumask_var_t new_mask;
4072 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4075 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4077 retval = sched_setaffinity(pid, new_mask);
4078 free_cpumask_var(new_mask);
4082 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4084 struct task_struct *p;
4085 unsigned long flags;
4092 p = find_process_by_pid(pid);
4096 retval = security_task_getscheduler(p);
4100 raw_spin_lock_irqsave(&p->pi_lock, flags);
4101 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4102 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4112 * sys_sched_getaffinity - get the cpu affinity of a process
4113 * @pid: pid of the process
4114 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4115 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4117 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4118 unsigned long __user *, user_mask_ptr)
4123 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4125 if (len & (sizeof(unsigned long)-1))
4128 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4131 ret = sched_getaffinity(pid, mask);
4133 size_t retlen = min_t(size_t, len, cpumask_size());
4135 if (copy_to_user(user_mask_ptr, mask, retlen))
4140 free_cpumask_var(mask);
4146 * sys_sched_yield - yield the current processor to other threads.
4148 * This function yields the current CPU to other tasks. If there are no
4149 * other threads running on this CPU then this function will return.
4151 SYSCALL_DEFINE0(sched_yield)
4153 struct rq *rq = this_rq_lock();
4155 schedstat_inc(rq, yld_count);
4156 current->sched_class->yield_task(rq);
4159 * Since we are going to call schedule() anyway, there's
4160 * no need to preempt or enable interrupts:
4162 __release(rq->lock);
4163 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4164 do_raw_spin_unlock(&rq->lock);
4165 sched_preempt_enable_no_resched();
4172 static inline int should_resched(void)
4174 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4177 static void __cond_resched(void)
4179 add_preempt_count(PREEMPT_ACTIVE);
4181 sub_preempt_count(PREEMPT_ACTIVE);
4184 int __sched _cond_resched(void)
4186 if (should_resched()) {
4192 EXPORT_SYMBOL(_cond_resched);
4195 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4196 * call schedule, and on return reacquire the lock.
4198 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4199 * operations here to prevent schedule() from being called twice (once via
4200 * spin_unlock(), once by hand).
4202 int __cond_resched_lock(spinlock_t *lock)
4204 int resched = should_resched();
4207 lockdep_assert_held(lock);
4209 if (spin_needbreak(lock) || resched) {
4220 EXPORT_SYMBOL(__cond_resched_lock);
4222 int __sched __cond_resched_softirq(void)
4224 BUG_ON(!in_softirq());
4226 if (should_resched()) {
4234 EXPORT_SYMBOL(__cond_resched_softirq);
4237 * yield - yield the current processor to other threads.
4239 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4241 * The scheduler is at all times free to pick the calling task as the most
4242 * eligible task to run, if removing the yield() call from your code breaks
4243 * it, its already broken.
4245 * Typical broken usage is:
4250 * where one assumes that yield() will let 'the other' process run that will
4251 * make event true. If the current task is a SCHED_FIFO task that will never
4252 * happen. Never use yield() as a progress guarantee!!
4254 * If you want to use yield() to wait for something, use wait_event().
4255 * If you want to use yield() to be 'nice' for others, use cond_resched().
4256 * If you still want to use yield(), do not!
4258 void __sched yield(void)
4260 set_current_state(TASK_RUNNING);
4263 EXPORT_SYMBOL(yield);
4266 * yield_to - yield the current processor to another thread in
4267 * your thread group, or accelerate that thread toward the
4268 * processor it's on.
4270 * @preempt: whether task preemption is allowed or not
4272 * It's the caller's job to ensure that the target task struct
4273 * can't go away on us before we can do any checks.
4275 * Returns true if we indeed boosted the target task.
4277 bool __sched yield_to(struct task_struct *p, bool preempt)
4279 struct task_struct *curr = current;
4280 struct rq *rq, *p_rq;
4281 unsigned long flags;
4284 local_irq_save(flags);
4289 double_rq_lock(rq, p_rq);
4290 while (task_rq(p) != p_rq) {
4291 double_rq_unlock(rq, p_rq);
4295 if (!curr->sched_class->yield_to_task)
4298 if (curr->sched_class != p->sched_class)
4301 if (task_running(p_rq, p) || p->state)
4304 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4306 schedstat_inc(rq, yld_count);
4308 * Make p's CPU reschedule; pick_next_entity takes care of
4311 if (preempt && rq != p_rq)
4312 resched_task(p_rq->curr);
4316 double_rq_unlock(rq, p_rq);
4317 local_irq_restore(flags);
4324 EXPORT_SYMBOL_GPL(yield_to);
4327 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4328 * that process accounting knows that this is a task in IO wait state.
4330 void __sched io_schedule(void)
4332 struct rq *rq = raw_rq();
4334 delayacct_blkio_start();
4335 atomic_inc(&rq->nr_iowait);
4336 blk_flush_plug(current);
4337 current->in_iowait = 1;
4339 current->in_iowait = 0;
4340 atomic_dec(&rq->nr_iowait);
4341 delayacct_blkio_end();
4343 EXPORT_SYMBOL(io_schedule);
4345 long __sched io_schedule_timeout(long timeout)
4347 struct rq *rq = raw_rq();
4350 delayacct_blkio_start();
4351 atomic_inc(&rq->nr_iowait);
4352 blk_flush_plug(current);
4353 current->in_iowait = 1;
4354 ret = schedule_timeout(timeout);
4355 current->in_iowait = 0;
4356 atomic_dec(&rq->nr_iowait);
4357 delayacct_blkio_end();
4362 * sys_sched_get_priority_max - return maximum RT priority.
4363 * @policy: scheduling class.
4365 * this syscall returns the maximum rt_priority that can be used
4366 * by a given scheduling class.
4368 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4375 ret = MAX_USER_RT_PRIO-1;
4387 * sys_sched_get_priority_min - return minimum RT priority.
4388 * @policy: scheduling class.
4390 * this syscall returns the minimum rt_priority that can be used
4391 * by a given scheduling class.
4393 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4411 * sys_sched_rr_get_interval - return the default timeslice of a process.
4412 * @pid: pid of the process.
4413 * @interval: userspace pointer to the timeslice value.
4415 * this syscall writes the default timeslice value of a given process
4416 * into the user-space timespec buffer. A value of '0' means infinity.
4418 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4419 struct timespec __user *, interval)
4421 struct task_struct *p;
4422 unsigned int time_slice;
4423 unsigned long flags;
4433 p = find_process_by_pid(pid);
4437 retval = security_task_getscheduler(p);
4441 rq = task_rq_lock(p, &flags);
4442 time_slice = p->sched_class->get_rr_interval(rq, p);
4443 task_rq_unlock(rq, p, &flags);
4446 jiffies_to_timespec(time_slice, &t);
4447 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4455 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4457 void sched_show_task(struct task_struct *p)
4459 unsigned long free = 0;
4462 state = p->state ? __ffs(p->state) + 1 : 0;
4463 printk(KERN_INFO "%-15.15s %c", p->comm,
4464 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4465 #if BITS_PER_LONG == 32
4466 if (state == TASK_RUNNING)
4467 printk(KERN_CONT " running ");
4469 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4471 if (state == TASK_RUNNING)
4472 printk(KERN_CONT " running task ");
4474 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4476 #ifdef CONFIG_DEBUG_STACK_USAGE
4477 free = stack_not_used(p);
4479 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4480 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4481 (unsigned long)task_thread_info(p)->flags);
4483 show_stack(p, NULL);
4486 void show_state_filter(unsigned long state_filter)
4488 struct task_struct *g, *p;
4490 #if BITS_PER_LONG == 32
4492 " task PC stack pid father\n");
4495 " task PC stack pid father\n");
4498 do_each_thread(g, p) {
4500 * reset the NMI-timeout, listing all files on a slow
4501 * console might take a lot of time:
4503 touch_nmi_watchdog();
4504 if (!state_filter || (p->state & state_filter))
4506 } while_each_thread(g, p);
4508 touch_all_softlockup_watchdogs();
4510 #ifdef CONFIG_SCHED_DEBUG
4511 sysrq_sched_debug_show();
4515 * Only show locks if all tasks are dumped:
4518 debug_show_all_locks();
4521 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4523 idle->sched_class = &idle_sched_class;
4527 * init_idle - set up an idle thread for a given CPU
4528 * @idle: task in question
4529 * @cpu: cpu the idle task belongs to
4531 * NOTE: this function does not set the idle thread's NEED_RESCHED
4532 * flag, to make booting more robust.
4534 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4536 struct rq *rq = cpu_rq(cpu);
4537 unsigned long flags;
4539 raw_spin_lock_irqsave(&rq->lock, flags);
4542 idle->state = TASK_RUNNING;
4543 idle->se.exec_start = sched_clock();
4545 do_set_cpus_allowed(idle, cpumask_of(cpu));
4547 * We're having a chicken and egg problem, even though we are
4548 * holding rq->lock, the cpu isn't yet set to this cpu so the
4549 * lockdep check in task_group() will fail.
4551 * Similar case to sched_fork(). / Alternatively we could
4552 * use task_rq_lock() here and obtain the other rq->lock.
4557 __set_task_cpu(idle, cpu);
4560 rq->curr = rq->idle = idle;
4561 #if defined(CONFIG_SMP)
4564 raw_spin_unlock_irqrestore(&rq->lock, flags);
4566 /* Set the preempt count _outside_ the spinlocks! */
4567 task_thread_info(idle)->preempt_count = 0;
4570 * The idle tasks have their own, simple scheduling class:
4572 idle->sched_class = &idle_sched_class;
4573 ftrace_graph_init_idle_task(idle, cpu);
4574 #if defined(CONFIG_SMP)
4575 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4580 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4582 if (p->sched_class && p->sched_class->set_cpus_allowed)
4583 p->sched_class->set_cpus_allowed(p, new_mask);
4585 cpumask_copy(&p->cpus_allowed, new_mask);
4586 p->nr_cpus_allowed = cpumask_weight(new_mask);
4590 * This is how migration works:
4592 * 1) we invoke migration_cpu_stop() on the target CPU using
4594 * 2) stopper starts to run (implicitly forcing the migrated thread
4596 * 3) it checks whether the migrated task is still in the wrong runqueue.
4597 * 4) if it's in the wrong runqueue then the migration thread removes
4598 * it and puts it into the right queue.
4599 * 5) stopper completes and stop_one_cpu() returns and the migration
4604 * Change a given task's CPU affinity. Migrate the thread to a
4605 * proper CPU and schedule it away if the CPU it's executing on
4606 * is removed from the allowed bitmask.
4608 * NOTE: the caller must have a valid reference to the task, the
4609 * task must not exit() & deallocate itself prematurely. The
4610 * call is not atomic; no spinlocks may be held.
4612 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4614 unsigned long flags;
4616 unsigned int dest_cpu;
4619 rq = task_rq_lock(p, &flags);
4621 if (cpumask_equal(&p->cpus_allowed, new_mask))
4624 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4629 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4634 do_set_cpus_allowed(p, new_mask);
4636 /* Can the task run on the task's current CPU? If so, we're done */
4637 if (cpumask_test_cpu(task_cpu(p), new_mask))
4640 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4642 struct migration_arg arg = { p, dest_cpu };
4643 /* Need help from migration thread: drop lock and wait. */
4644 task_rq_unlock(rq, p, &flags);
4645 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4646 tlb_migrate_finish(p->mm);
4650 task_rq_unlock(rq, p, &flags);
4654 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4657 * Move (not current) task off this cpu, onto dest cpu. We're doing
4658 * this because either it can't run here any more (set_cpus_allowed()
4659 * away from this CPU, or CPU going down), or because we're
4660 * attempting to rebalance this task on exec (sched_exec).
4662 * So we race with normal scheduler movements, but that's OK, as long
4663 * as the task is no longer on this CPU.
4665 * Returns non-zero if task was successfully migrated.
4667 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4669 struct rq *rq_dest, *rq_src;
4672 if (unlikely(!cpu_active(dest_cpu)))
4675 rq_src = cpu_rq(src_cpu);
4676 rq_dest = cpu_rq(dest_cpu);
4678 raw_spin_lock(&p->pi_lock);
4679 double_rq_lock(rq_src, rq_dest);
4680 /* Already moved. */
4681 if (task_cpu(p) != src_cpu)
4683 /* Affinity changed (again). */
4684 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4688 * If we're not on a rq, the next wake-up will ensure we're
4692 dequeue_task(rq_src, p, 0);
4693 set_task_cpu(p, dest_cpu);
4694 enqueue_task(rq_dest, p, 0);
4695 check_preempt_curr(rq_dest, p, 0);
4700 double_rq_unlock(rq_src, rq_dest);
4701 raw_spin_unlock(&p->pi_lock);
4706 * migration_cpu_stop - this will be executed by a highprio stopper thread
4707 * and performs thread migration by bumping thread off CPU then
4708 * 'pushing' onto another runqueue.
4710 static int migration_cpu_stop(void *data)
4712 struct migration_arg *arg = data;
4715 * The original target cpu might have gone down and we might
4716 * be on another cpu but it doesn't matter.
4718 local_irq_disable();
4719 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4724 #ifdef CONFIG_HOTPLUG_CPU
4727 * Ensures that the idle task is using init_mm right before its cpu goes
4730 void idle_task_exit(void)
4732 struct mm_struct *mm = current->active_mm;
4734 BUG_ON(cpu_online(smp_processor_id()));
4737 switch_mm(mm, &init_mm, current);
4742 * Since this CPU is going 'away' for a while, fold any nr_active delta
4743 * we might have. Assumes we're called after migrate_tasks() so that the
4744 * nr_active count is stable.
4746 * Also see the comment "Global load-average calculations".
4748 static void calc_load_migrate(struct rq *rq)
4750 long delta = calc_load_fold_active(rq);
4752 atomic_long_add(delta, &calc_load_tasks);
4756 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4757 * try_to_wake_up()->select_task_rq().
4759 * Called with rq->lock held even though we'er in stop_machine() and
4760 * there's no concurrency possible, we hold the required locks anyway
4761 * because of lock validation efforts.
4763 static void migrate_tasks(unsigned int dead_cpu)
4765 struct rq *rq = cpu_rq(dead_cpu);
4766 struct task_struct *next, *stop = rq->stop;
4770 * Fudge the rq selection such that the below task selection loop
4771 * doesn't get stuck on the currently eligible stop task.
4773 * We're currently inside stop_machine() and the rq is either stuck
4774 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4775 * either way we should never end up calling schedule() until we're
4782 * There's this thread running, bail when that's the only
4785 if (rq->nr_running == 1)
4788 next = pick_next_task(rq);
4790 next->sched_class->put_prev_task(rq, next);
4792 /* Find suitable destination for @next, with force if needed. */
4793 dest_cpu = select_fallback_rq(dead_cpu, next);
4794 raw_spin_unlock(&rq->lock);
4796 __migrate_task(next, dead_cpu, dest_cpu);
4798 raw_spin_lock(&rq->lock);
4804 #endif /* CONFIG_HOTPLUG_CPU */
4806 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4808 static struct ctl_table sd_ctl_dir[] = {
4810 .procname = "sched_domain",
4816 static struct ctl_table sd_ctl_root[] = {
4818 .procname = "kernel",
4820 .child = sd_ctl_dir,
4825 static struct ctl_table *sd_alloc_ctl_entry(int n)
4827 struct ctl_table *entry =
4828 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4833 static void sd_free_ctl_entry(struct ctl_table **tablep)
4835 struct ctl_table *entry;
4838 * In the intermediate directories, both the child directory and
4839 * procname are dynamically allocated and could fail but the mode
4840 * will always be set. In the lowest directory the names are
4841 * static strings and all have proc handlers.
4843 for (entry = *tablep; entry->mode; entry++) {
4845 sd_free_ctl_entry(&entry->child);
4846 if (entry->proc_handler == NULL)
4847 kfree(entry->procname);
4854 static int min_load_idx = 0;
4855 static int max_load_idx = CPU_LOAD_IDX_MAX;
4858 set_table_entry(struct ctl_table *entry,
4859 const char *procname, void *data, int maxlen,
4860 umode_t mode, proc_handler *proc_handler,
4863 entry->procname = procname;
4865 entry->maxlen = maxlen;
4867 entry->proc_handler = proc_handler;
4870 entry->extra1 = &min_load_idx;
4871 entry->extra2 = &max_load_idx;
4875 static struct ctl_table *
4876 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4878 struct ctl_table *table = sd_alloc_ctl_entry(13);
4883 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4884 sizeof(long), 0644, proc_doulongvec_minmax, false);
4885 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4886 sizeof(long), 0644, proc_doulongvec_minmax, false);
4887 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4888 sizeof(int), 0644, proc_dointvec_minmax, true);
4889 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4890 sizeof(int), 0644, proc_dointvec_minmax, true);
4891 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4892 sizeof(int), 0644, proc_dointvec_minmax, true);
4893 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4894 sizeof(int), 0644, proc_dointvec_minmax, true);
4895 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4896 sizeof(int), 0644, proc_dointvec_minmax, true);
4897 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4898 sizeof(int), 0644, proc_dointvec_minmax, false);
4899 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4900 sizeof(int), 0644, proc_dointvec_minmax, false);
4901 set_table_entry(&table[9], "cache_nice_tries",
4902 &sd->cache_nice_tries,
4903 sizeof(int), 0644, proc_dointvec_minmax, false);
4904 set_table_entry(&table[10], "flags", &sd->flags,
4905 sizeof(int), 0644, proc_dointvec_minmax, false);
4906 set_table_entry(&table[11], "name", sd->name,
4907 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4908 /* &table[12] is terminator */
4913 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4915 struct ctl_table *entry, *table;
4916 struct sched_domain *sd;
4917 int domain_num = 0, i;
4920 for_each_domain(cpu, sd)
4922 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4927 for_each_domain(cpu, sd) {
4928 snprintf(buf, 32, "domain%d", i);
4929 entry->procname = kstrdup(buf, GFP_KERNEL);
4931 entry->child = sd_alloc_ctl_domain_table(sd);
4938 static struct ctl_table_header *sd_sysctl_header;
4939 static void register_sched_domain_sysctl(void)
4941 int i, cpu_num = num_possible_cpus();
4942 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4945 WARN_ON(sd_ctl_dir[0].child);
4946 sd_ctl_dir[0].child = entry;
4951 for_each_possible_cpu(i) {
4952 snprintf(buf, 32, "cpu%d", i);
4953 entry->procname = kstrdup(buf, GFP_KERNEL);
4955 entry->child = sd_alloc_ctl_cpu_table(i);
4959 WARN_ON(sd_sysctl_header);
4960 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4963 /* may be called multiple times per register */
4964 static void unregister_sched_domain_sysctl(void)
4966 if (sd_sysctl_header)
4967 unregister_sysctl_table(sd_sysctl_header);
4968 sd_sysctl_header = NULL;
4969 if (sd_ctl_dir[0].child)
4970 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4973 static void register_sched_domain_sysctl(void)
4976 static void unregister_sched_domain_sysctl(void)
4981 static void set_rq_online(struct rq *rq)
4984 const struct sched_class *class;
4986 cpumask_set_cpu(rq->cpu, rq->rd->online);
4989 for_each_class(class) {
4990 if (class->rq_online)
4991 class->rq_online(rq);
4996 static void set_rq_offline(struct rq *rq)
4999 const struct sched_class *class;
5001 for_each_class(class) {
5002 if (class->rq_offline)
5003 class->rq_offline(rq);
5006 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5012 * migration_call - callback that gets triggered when a CPU is added.
5013 * Here we can start up the necessary migration thread for the new CPU.
5015 static int __cpuinit
5016 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5018 int cpu = (long)hcpu;
5019 unsigned long flags;
5020 struct rq *rq = cpu_rq(cpu);
5022 switch (action & ~CPU_TASKS_FROZEN) {
5024 case CPU_UP_PREPARE:
5025 rq->calc_load_update = calc_load_update;
5029 /* Update our root-domain */
5030 raw_spin_lock_irqsave(&rq->lock, flags);
5032 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5036 raw_spin_unlock_irqrestore(&rq->lock, flags);
5039 #ifdef CONFIG_HOTPLUG_CPU
5041 sched_ttwu_pending();
5042 /* Update our root-domain */
5043 raw_spin_lock_irqsave(&rq->lock, flags);
5045 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5049 BUG_ON(rq->nr_running != 1); /* the migration thread */
5050 raw_spin_unlock_irqrestore(&rq->lock, flags);
5052 calc_load_migrate(rq);
5057 update_max_interval();
5063 * Register at high priority so that task migration (migrate_all_tasks)
5064 * happens before everything else. This has to be lower priority than
5065 * the notifier in the perf_event subsystem, though.
5067 static struct notifier_block __cpuinitdata migration_notifier = {
5068 .notifier_call = migration_call,
5069 .priority = CPU_PRI_MIGRATION,
5072 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5073 unsigned long action, void *hcpu)
5075 switch (action & ~CPU_TASKS_FROZEN) {
5077 case CPU_DOWN_FAILED:
5078 set_cpu_active((long)hcpu, true);
5085 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5086 unsigned long action, void *hcpu)
5088 switch (action & ~CPU_TASKS_FROZEN) {
5089 case CPU_DOWN_PREPARE:
5090 set_cpu_active((long)hcpu, false);
5097 static int __init migration_init(void)
5099 void *cpu = (void *)(long)smp_processor_id();
5102 /* Initialize migration for the boot CPU */
5103 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5104 BUG_ON(err == NOTIFY_BAD);
5105 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5106 register_cpu_notifier(&migration_notifier);
5108 /* Register cpu active notifiers */
5109 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5110 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5114 early_initcall(migration_init);
5119 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5121 #ifdef CONFIG_SCHED_DEBUG
5123 static __read_mostly int sched_debug_enabled;
5125 static int __init sched_debug_setup(char *str)
5127 sched_debug_enabled = 1;
5131 early_param("sched_debug", sched_debug_setup);
5133 static inline bool sched_debug(void)
5135 return sched_debug_enabled;
5138 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5139 struct cpumask *groupmask)
5141 struct sched_group *group = sd->groups;
5144 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5145 cpumask_clear(groupmask);
5147 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5149 if (!(sd->flags & SD_LOAD_BALANCE)) {
5150 printk("does not load-balance\n");
5152 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5157 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5159 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5160 printk(KERN_ERR "ERROR: domain->span does not contain "
5163 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5164 printk(KERN_ERR "ERROR: domain->groups does not contain"
5168 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5172 printk(KERN_ERR "ERROR: group is NULL\n");
5177 * Even though we initialize ->power to something semi-sane,
5178 * we leave power_orig unset. This allows us to detect if
5179 * domain iteration is still funny without causing /0 traps.
5181 if (!group->sgp->power_orig) {
5182 printk(KERN_CONT "\n");
5183 printk(KERN_ERR "ERROR: domain->cpu_power not "
5188 if (!cpumask_weight(sched_group_cpus(group))) {
5189 printk(KERN_CONT "\n");
5190 printk(KERN_ERR "ERROR: empty group\n");
5194 if (!(sd->flags & SD_OVERLAP) &&
5195 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5196 printk(KERN_CONT "\n");
5197 printk(KERN_ERR "ERROR: repeated CPUs\n");
5201 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5203 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5205 printk(KERN_CONT " %s", str);
5206 if (group->sgp->power != SCHED_POWER_SCALE) {
5207 printk(KERN_CONT " (cpu_power = %d)",
5211 group = group->next;
5212 } while (group != sd->groups);
5213 printk(KERN_CONT "\n");
5215 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5216 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5219 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5220 printk(KERN_ERR "ERROR: parent span is not a superset "
5221 "of domain->span\n");
5225 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5229 if (!sched_debug_enabled)
5233 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5237 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5240 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5248 #else /* !CONFIG_SCHED_DEBUG */
5249 # define sched_domain_debug(sd, cpu) do { } while (0)
5250 static inline bool sched_debug(void)
5254 #endif /* CONFIG_SCHED_DEBUG */
5256 static int sd_degenerate(struct sched_domain *sd)
5258 if (cpumask_weight(sched_domain_span(sd)) == 1)
5261 /* Following flags need at least 2 groups */
5262 if (sd->flags & (SD_LOAD_BALANCE |
5263 SD_BALANCE_NEWIDLE |
5267 SD_SHARE_PKG_RESOURCES)) {
5268 if (sd->groups != sd->groups->next)
5272 /* Following flags don't use groups */
5273 if (sd->flags & (SD_WAKE_AFFINE))
5280 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5282 unsigned long cflags = sd->flags, pflags = parent->flags;
5284 if (sd_degenerate(parent))
5287 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5290 /* Flags needing groups don't count if only 1 group in parent */
5291 if (parent->groups == parent->groups->next) {
5292 pflags &= ~(SD_LOAD_BALANCE |
5293 SD_BALANCE_NEWIDLE |
5297 SD_SHARE_PKG_RESOURCES);
5298 if (nr_node_ids == 1)
5299 pflags &= ~SD_SERIALIZE;
5301 if (~cflags & pflags)
5307 static void free_rootdomain(struct rcu_head *rcu)
5309 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5311 cpupri_cleanup(&rd->cpupri);
5312 free_cpumask_var(rd->rto_mask);
5313 free_cpumask_var(rd->online);
5314 free_cpumask_var(rd->span);
5318 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5320 struct root_domain *old_rd = NULL;
5321 unsigned long flags;
5323 raw_spin_lock_irqsave(&rq->lock, flags);
5328 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5331 cpumask_clear_cpu(rq->cpu, old_rd->span);
5334 * If we dont want to free the old_rt yet then
5335 * set old_rd to NULL to skip the freeing later
5338 if (!atomic_dec_and_test(&old_rd->refcount))
5342 atomic_inc(&rd->refcount);
5345 cpumask_set_cpu(rq->cpu, rd->span);
5346 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5349 raw_spin_unlock_irqrestore(&rq->lock, flags);
5352 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5355 static int init_rootdomain(struct root_domain *rd)
5357 memset(rd, 0, sizeof(*rd));
5359 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5361 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5363 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5366 if (cpupri_init(&rd->cpupri) != 0)
5371 free_cpumask_var(rd->rto_mask);
5373 free_cpumask_var(rd->online);
5375 free_cpumask_var(rd->span);
5381 * By default the system creates a single root-domain with all cpus as
5382 * members (mimicking the global state we have today).
5384 struct root_domain def_root_domain;
5386 static void init_defrootdomain(void)
5388 init_rootdomain(&def_root_domain);
5390 atomic_set(&def_root_domain.refcount, 1);
5393 static struct root_domain *alloc_rootdomain(void)
5395 struct root_domain *rd;
5397 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5401 if (init_rootdomain(rd) != 0) {
5409 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5411 struct sched_group *tmp, *first;
5420 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5425 } while (sg != first);
5428 static void free_sched_domain(struct rcu_head *rcu)
5430 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5433 * If its an overlapping domain it has private groups, iterate and
5436 if (sd->flags & SD_OVERLAP) {
5437 free_sched_groups(sd->groups, 1);
5438 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5439 kfree(sd->groups->sgp);
5445 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5447 call_rcu(&sd->rcu, free_sched_domain);
5450 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5452 for (; sd; sd = sd->parent)
5453 destroy_sched_domain(sd, cpu);
5457 * Keep a special pointer to the highest sched_domain that has
5458 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5459 * allows us to avoid some pointer chasing select_idle_sibling().
5461 * Iterate domains and sched_groups downward, assigning CPUs to be
5462 * select_idle_sibling() hw buddy. Cross-wiring hw makes bouncing
5463 * due to random perturbation self canceling, ie sw buddies pull
5464 * their counterpart to their CPU's hw counterpart.
5466 * Also keep a unique ID per domain (we use the first cpu number in
5467 * the cpumask of the domain), this allows us to quickly tell if
5468 * two cpus are in the same cache domain, see cpus_share_cache().
5470 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5471 DEFINE_PER_CPU(int, sd_llc_id);
5473 static void update_top_cache_domain(int cpu)
5475 struct sched_domain *sd;
5478 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5480 struct sched_domain *tmp = sd;
5481 struct sched_group *sg, *prev;
5485 * Traverse to first CPU in group, and count hops
5486 * to cpu from there, switching direction on each
5487 * hop, never ever pointing the last CPU rightward.
5490 id = cpumask_first(sched_domain_span(tmp));
5491 prev = sg = tmp->groups;
5494 while (cpumask_first(sched_group_cpus(sg)) != id)
5497 while (!cpumask_test_cpu(cpu, sched_group_cpus(sg))) {
5503 /* A CPU went down, never point back to domain start. */
5504 if (right && cpumask_first(sched_group_cpus(sg->next)) == id)
5507 sg = right ? sg->next : prev;
5508 tmp->idle_buddy = cpumask_first(sched_group_cpus(sg));
5509 } while ((tmp = tmp->child));
5511 id = cpumask_first(sched_domain_span(sd));
5514 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5515 per_cpu(sd_llc_id, cpu) = id;
5519 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5520 * hold the hotplug lock.
5523 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5525 struct rq *rq = cpu_rq(cpu);
5526 struct sched_domain *tmp;
5528 /* Remove the sched domains which do not contribute to scheduling. */
5529 for (tmp = sd; tmp; ) {
5530 struct sched_domain *parent = tmp->parent;
5534 if (sd_parent_degenerate(tmp, parent)) {
5535 tmp->parent = parent->parent;
5537 parent->parent->child = tmp;
5538 destroy_sched_domain(parent, cpu);
5543 if (sd && sd_degenerate(sd)) {
5546 destroy_sched_domain(tmp, cpu);
5551 sched_domain_debug(sd, cpu);
5553 rq_attach_root(rq, rd);
5555 rcu_assign_pointer(rq->sd, sd);
5556 destroy_sched_domains(tmp, cpu);
5558 update_top_cache_domain(cpu);
5561 /* cpus with isolated domains */
5562 static cpumask_var_t cpu_isolated_map;
5564 /* Setup the mask of cpus configured for isolated domains */
5565 static int __init isolated_cpu_setup(char *str)
5567 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5568 cpulist_parse(str, cpu_isolated_map);
5572 __setup("isolcpus=", isolated_cpu_setup);
5574 static const struct cpumask *cpu_cpu_mask(int cpu)
5576 return cpumask_of_node(cpu_to_node(cpu));
5580 struct sched_domain **__percpu sd;
5581 struct sched_group **__percpu sg;
5582 struct sched_group_power **__percpu sgp;
5586 struct sched_domain ** __percpu sd;
5587 struct root_domain *rd;
5597 struct sched_domain_topology_level;
5599 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5600 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5602 #define SDTL_OVERLAP 0x01
5604 struct sched_domain_topology_level {
5605 sched_domain_init_f init;
5606 sched_domain_mask_f mask;
5609 struct sd_data data;
5613 * Build an iteration mask that can exclude certain CPUs from the upwards
5616 * Asymmetric node setups can result in situations where the domain tree is of
5617 * unequal depth, make sure to skip domains that already cover the entire
5620 * In that case build_sched_domains() will have terminated the iteration early
5621 * and our sibling sd spans will be empty. Domains should always include the
5622 * cpu they're built on, so check that.
5625 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5627 const struct cpumask *span = sched_domain_span(sd);
5628 struct sd_data *sdd = sd->private;
5629 struct sched_domain *sibling;
5632 for_each_cpu(i, span) {
5633 sibling = *per_cpu_ptr(sdd->sd, i);
5634 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5637 cpumask_set_cpu(i, sched_group_mask(sg));
5642 * Return the canonical balance cpu for this group, this is the first cpu
5643 * of this group that's also in the iteration mask.
5645 int group_balance_cpu(struct sched_group *sg)
5647 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5651 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5653 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5654 const struct cpumask *span = sched_domain_span(sd);
5655 struct cpumask *covered = sched_domains_tmpmask;
5656 struct sd_data *sdd = sd->private;
5657 struct sched_domain *child;
5660 cpumask_clear(covered);
5662 for_each_cpu(i, span) {
5663 struct cpumask *sg_span;
5665 if (cpumask_test_cpu(i, covered))
5668 child = *per_cpu_ptr(sdd->sd, i);
5670 /* See the comment near build_group_mask(). */
5671 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5674 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5675 GFP_KERNEL, cpu_to_node(cpu));
5680 sg_span = sched_group_cpus(sg);
5682 child = child->child;
5683 cpumask_copy(sg_span, sched_domain_span(child));
5685 cpumask_set_cpu(i, sg_span);
5687 cpumask_or(covered, covered, sg_span);
5689 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5690 if (atomic_inc_return(&sg->sgp->ref) == 1)
5691 build_group_mask(sd, sg);
5694 * Initialize sgp->power such that even if we mess up the
5695 * domains and no possible iteration will get us here, we won't
5698 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5701 * Make sure the first group of this domain contains the
5702 * canonical balance cpu. Otherwise the sched_domain iteration
5703 * breaks. See update_sg_lb_stats().
5705 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5706 group_balance_cpu(sg) == cpu)
5716 sd->groups = groups;
5721 free_sched_groups(first, 0);
5726 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5728 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5729 struct sched_domain *child = sd->child;
5732 cpu = cpumask_first(sched_domain_span(child));
5735 *sg = *per_cpu_ptr(sdd->sg, cpu);
5736 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5737 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5744 * build_sched_groups will build a circular linked list of the groups
5745 * covered by the given span, and will set each group's ->cpumask correctly,
5746 * and ->cpu_power to 0.
5748 * Assumes the sched_domain tree is fully constructed
5751 build_sched_groups(struct sched_domain *sd, int cpu)
5753 struct sched_group *first = NULL, *last = NULL;
5754 struct sd_data *sdd = sd->private;
5755 const struct cpumask *span = sched_domain_span(sd);
5756 struct cpumask *covered;
5759 get_group(cpu, sdd, &sd->groups);
5760 atomic_inc(&sd->groups->ref);
5762 if (cpu != cpumask_first(sched_domain_span(sd)))
5765 lockdep_assert_held(&sched_domains_mutex);
5766 covered = sched_domains_tmpmask;
5768 cpumask_clear(covered);
5770 for_each_cpu(i, span) {
5771 struct sched_group *sg;
5772 int group = get_group(i, sdd, &sg);
5775 if (cpumask_test_cpu(i, covered))
5778 cpumask_clear(sched_group_cpus(sg));
5780 cpumask_setall(sched_group_mask(sg));
5782 for_each_cpu(j, span) {
5783 if (get_group(j, sdd, NULL) != group)
5786 cpumask_set_cpu(j, covered);
5787 cpumask_set_cpu(j, sched_group_cpus(sg));
5802 * Initialize sched groups cpu_power.
5804 * cpu_power indicates the capacity of sched group, which is used while
5805 * distributing the load between different sched groups in a sched domain.
5806 * Typically cpu_power for all the groups in a sched domain will be same unless
5807 * there are asymmetries in the topology. If there are asymmetries, group
5808 * having more cpu_power will pickup more load compared to the group having
5811 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5813 struct sched_group *sg = sd->groups;
5815 WARN_ON(!sd || !sg);
5818 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5820 } while (sg != sd->groups);
5822 if (cpu != group_balance_cpu(sg))
5825 update_group_power(sd, cpu);
5826 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5829 int __weak arch_sd_sibling_asym_packing(void)
5831 return 0*SD_ASYM_PACKING;
5835 * Initializers for schedule domains
5836 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5839 #ifdef CONFIG_SCHED_DEBUG
5840 # define SD_INIT_NAME(sd, type) sd->name = #type
5842 # define SD_INIT_NAME(sd, type) do { } while (0)
5845 #define SD_INIT_FUNC(type) \
5846 static noinline struct sched_domain * \
5847 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5849 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5850 *sd = SD_##type##_INIT; \
5851 SD_INIT_NAME(sd, type); \
5852 sd->private = &tl->data; \
5857 #ifdef CONFIG_SCHED_SMT
5858 SD_INIT_FUNC(SIBLING)
5860 #ifdef CONFIG_SCHED_MC
5863 #ifdef CONFIG_SCHED_BOOK
5867 static int default_relax_domain_level = -1;
5868 int sched_domain_level_max;
5870 static int __init setup_relax_domain_level(char *str)
5872 if (kstrtoint(str, 0, &default_relax_domain_level))
5873 pr_warn("Unable to set relax_domain_level\n");
5877 __setup("relax_domain_level=", setup_relax_domain_level);
5879 static void set_domain_attribute(struct sched_domain *sd,
5880 struct sched_domain_attr *attr)
5884 if (!attr || attr->relax_domain_level < 0) {
5885 if (default_relax_domain_level < 0)
5888 request = default_relax_domain_level;
5890 request = attr->relax_domain_level;
5891 if (request < sd->level) {
5892 /* turn off idle balance on this domain */
5893 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5895 /* turn on idle balance on this domain */
5896 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5900 static void __sdt_free(const struct cpumask *cpu_map);
5901 static int __sdt_alloc(const struct cpumask *cpu_map);
5903 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5904 const struct cpumask *cpu_map)
5908 if (!atomic_read(&d->rd->refcount))
5909 free_rootdomain(&d->rd->rcu); /* fall through */
5911 free_percpu(d->sd); /* fall through */
5913 __sdt_free(cpu_map); /* fall through */
5919 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5920 const struct cpumask *cpu_map)
5922 memset(d, 0, sizeof(*d));
5924 if (__sdt_alloc(cpu_map))
5925 return sa_sd_storage;
5926 d->sd = alloc_percpu(struct sched_domain *);
5928 return sa_sd_storage;
5929 d->rd = alloc_rootdomain();
5932 return sa_rootdomain;
5936 * NULL the sd_data elements we've used to build the sched_domain and
5937 * sched_group structure so that the subsequent __free_domain_allocs()
5938 * will not free the data we're using.
5940 static void claim_allocations(int cpu, struct sched_domain *sd)
5942 struct sd_data *sdd = sd->private;
5944 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5945 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5947 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5948 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5950 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5951 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5954 #ifdef CONFIG_SCHED_SMT
5955 static const struct cpumask *cpu_smt_mask(int cpu)
5957 return topology_thread_cpumask(cpu);
5962 * Topology list, bottom-up.
5964 static struct sched_domain_topology_level default_topology[] = {
5965 #ifdef CONFIG_SCHED_SMT
5966 { sd_init_SIBLING, cpu_smt_mask, },
5968 #ifdef CONFIG_SCHED_MC
5969 { sd_init_MC, cpu_coregroup_mask, },
5971 #ifdef CONFIG_SCHED_BOOK
5972 { sd_init_BOOK, cpu_book_mask, },
5974 { sd_init_CPU, cpu_cpu_mask, },
5978 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5982 static int sched_domains_numa_levels;
5983 static int *sched_domains_numa_distance;
5984 static struct cpumask ***sched_domains_numa_masks;
5985 static int sched_domains_curr_level;
5987 static inline int sd_local_flags(int level)
5989 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5992 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5995 static struct sched_domain *
5996 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5998 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5999 int level = tl->numa_level;
6000 int sd_weight = cpumask_weight(
6001 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6003 *sd = (struct sched_domain){
6004 .min_interval = sd_weight,
6005 .max_interval = 2*sd_weight,
6007 .imbalance_pct = 125,
6008 .cache_nice_tries = 2,
6015 .flags = 1*SD_LOAD_BALANCE
6016 | 1*SD_BALANCE_NEWIDLE
6021 | 0*SD_SHARE_CPUPOWER
6022 | 0*SD_SHARE_PKG_RESOURCES
6024 | 0*SD_PREFER_SIBLING
6025 | sd_local_flags(level)
6027 .last_balance = jiffies,
6028 .balance_interval = sd_weight,
6030 SD_INIT_NAME(sd, NUMA);
6031 sd->private = &tl->data;
6034 * Ugly hack to pass state to sd_numa_mask()...
6036 sched_domains_curr_level = tl->numa_level;
6041 static const struct cpumask *sd_numa_mask(int cpu)
6043 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6046 static void sched_numa_warn(const char *str)
6048 static int done = false;
6056 printk(KERN_WARNING "ERROR: %s\n\n", str);
6058 for (i = 0; i < nr_node_ids; i++) {
6059 printk(KERN_WARNING " ");
6060 for (j = 0; j < nr_node_ids; j++)
6061 printk(KERN_CONT "%02d ", node_distance(i,j));
6062 printk(KERN_CONT "\n");
6064 printk(KERN_WARNING "\n");
6067 static bool find_numa_distance(int distance)
6071 if (distance == node_distance(0, 0))
6074 for (i = 0; i < sched_domains_numa_levels; i++) {
6075 if (sched_domains_numa_distance[i] == distance)
6082 static void sched_init_numa(void)
6084 int next_distance, curr_distance = node_distance(0, 0);
6085 struct sched_domain_topology_level *tl;
6089 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6090 if (!sched_domains_numa_distance)
6094 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6095 * unique distances in the node_distance() table.
6097 * Assumes node_distance(0,j) includes all distances in
6098 * node_distance(i,j) in order to avoid cubic time.
6100 next_distance = curr_distance;
6101 for (i = 0; i < nr_node_ids; i++) {
6102 for (j = 0; j < nr_node_ids; j++) {
6103 for (k = 0; k < nr_node_ids; k++) {
6104 int distance = node_distance(i, k);
6106 if (distance > curr_distance &&
6107 (distance < next_distance ||
6108 next_distance == curr_distance))
6109 next_distance = distance;
6112 * While not a strong assumption it would be nice to know
6113 * about cases where if node A is connected to B, B is not
6114 * equally connected to A.
6116 if (sched_debug() && node_distance(k, i) != distance)
6117 sched_numa_warn("Node-distance not symmetric");
6119 if (sched_debug() && i && !find_numa_distance(distance))
6120 sched_numa_warn("Node-0 not representative");
6122 if (next_distance != curr_distance) {
6123 sched_domains_numa_distance[level++] = next_distance;
6124 sched_domains_numa_levels = level;
6125 curr_distance = next_distance;
6130 * In case of sched_debug() we verify the above assumption.
6136 * 'level' contains the number of unique distances, excluding the
6137 * identity distance node_distance(i,i).
6139 * The sched_domains_nume_distance[] array includes the actual distance
6143 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6144 if (!sched_domains_numa_masks)
6148 * Now for each level, construct a mask per node which contains all
6149 * cpus of nodes that are that many hops away from us.
6151 for (i = 0; i < level; i++) {
6152 sched_domains_numa_masks[i] =
6153 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6154 if (!sched_domains_numa_masks[i])
6157 for (j = 0; j < nr_node_ids; j++) {
6158 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6162 sched_domains_numa_masks[i][j] = mask;
6164 for (k = 0; k < nr_node_ids; k++) {
6165 if (node_distance(j, k) > sched_domains_numa_distance[i])
6168 cpumask_or(mask, mask, cpumask_of_node(k));
6173 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6174 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6179 * Copy the default topology bits..
6181 for (i = 0; default_topology[i].init; i++)
6182 tl[i] = default_topology[i];
6185 * .. and append 'j' levels of NUMA goodness.
6187 for (j = 0; j < level; i++, j++) {
6188 tl[i] = (struct sched_domain_topology_level){
6189 .init = sd_numa_init,
6190 .mask = sd_numa_mask,
6191 .flags = SDTL_OVERLAP,
6196 sched_domain_topology = tl;
6199 static inline void sched_init_numa(void)
6202 #endif /* CONFIG_NUMA */
6204 static int __sdt_alloc(const struct cpumask *cpu_map)
6206 struct sched_domain_topology_level *tl;
6209 for (tl = sched_domain_topology; tl->init; tl++) {
6210 struct sd_data *sdd = &tl->data;
6212 sdd->sd = alloc_percpu(struct sched_domain *);
6216 sdd->sg = alloc_percpu(struct sched_group *);
6220 sdd->sgp = alloc_percpu(struct sched_group_power *);
6224 for_each_cpu(j, cpu_map) {
6225 struct sched_domain *sd;
6226 struct sched_group *sg;
6227 struct sched_group_power *sgp;
6229 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6230 GFP_KERNEL, cpu_to_node(j));
6234 *per_cpu_ptr(sdd->sd, j) = sd;
6236 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6237 GFP_KERNEL, cpu_to_node(j));
6243 *per_cpu_ptr(sdd->sg, j) = sg;
6245 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6246 GFP_KERNEL, cpu_to_node(j));
6250 *per_cpu_ptr(sdd->sgp, j) = sgp;
6257 static void __sdt_free(const struct cpumask *cpu_map)
6259 struct sched_domain_topology_level *tl;
6262 for (tl = sched_domain_topology; tl->init; tl++) {
6263 struct sd_data *sdd = &tl->data;
6265 for_each_cpu(j, cpu_map) {
6266 struct sched_domain *sd;
6269 sd = *per_cpu_ptr(sdd->sd, j);
6270 if (sd && (sd->flags & SD_OVERLAP))
6271 free_sched_groups(sd->groups, 0);
6272 kfree(*per_cpu_ptr(sdd->sd, j));
6276 kfree(*per_cpu_ptr(sdd->sg, j));
6278 kfree(*per_cpu_ptr(sdd->sgp, j));
6280 free_percpu(sdd->sd);
6282 free_percpu(sdd->sg);
6284 free_percpu(sdd->sgp);
6289 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6290 struct s_data *d, const struct cpumask *cpu_map,
6291 struct sched_domain_attr *attr, struct sched_domain *child,
6294 struct sched_domain *sd = tl->init(tl, cpu);
6298 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6300 sd->level = child->level + 1;
6301 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6305 set_domain_attribute(sd, attr);
6311 * Build sched domains for a given set of cpus and attach the sched domains
6312 * to the individual cpus
6314 static int build_sched_domains(const struct cpumask *cpu_map,
6315 struct sched_domain_attr *attr)
6317 enum s_alloc alloc_state = sa_none;
6318 struct sched_domain *sd;
6320 int i, ret = -ENOMEM;
6322 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6323 if (alloc_state != sa_rootdomain)
6326 /* Set up domains for cpus specified by the cpu_map. */
6327 for_each_cpu(i, cpu_map) {
6328 struct sched_domain_topology_level *tl;
6331 for (tl = sched_domain_topology; tl->init; tl++) {
6332 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6333 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6334 sd->flags |= SD_OVERLAP;
6335 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6342 *per_cpu_ptr(d.sd, i) = sd;
6345 /* Build the groups for the domains */
6346 for_each_cpu(i, cpu_map) {
6347 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6348 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6349 if (sd->flags & SD_OVERLAP) {
6350 if (build_overlap_sched_groups(sd, i))
6353 if (build_sched_groups(sd, i))
6359 /* Calculate CPU power for physical packages and nodes */
6360 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6361 if (!cpumask_test_cpu(i, cpu_map))
6364 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6365 claim_allocations(i, sd);
6366 init_sched_groups_power(i, sd);
6370 /* Attach the domains */
6372 for_each_cpu(i, cpu_map) {
6373 sd = *per_cpu_ptr(d.sd, i);
6374 cpu_attach_domain(sd, d.rd, i);
6380 __free_domain_allocs(&d, alloc_state, cpu_map);
6384 static cpumask_var_t *doms_cur; /* current sched domains */
6385 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6386 static struct sched_domain_attr *dattr_cur;
6387 /* attribues of custom domains in 'doms_cur' */
6390 * Special case: If a kmalloc of a doms_cur partition (array of
6391 * cpumask) fails, then fallback to a single sched domain,
6392 * as determined by the single cpumask fallback_doms.
6394 static cpumask_var_t fallback_doms;
6397 * arch_update_cpu_topology lets virtualized architectures update the
6398 * cpu core maps. It is supposed to return 1 if the topology changed
6399 * or 0 if it stayed the same.
6401 int __attribute__((weak)) arch_update_cpu_topology(void)
6406 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6409 cpumask_var_t *doms;
6411 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6414 for (i = 0; i < ndoms; i++) {
6415 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6416 free_sched_domains(doms, i);
6423 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6426 for (i = 0; i < ndoms; i++)
6427 free_cpumask_var(doms[i]);
6432 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6433 * For now this just excludes isolated cpus, but could be used to
6434 * exclude other special cases in the future.
6436 static int init_sched_domains(const struct cpumask *cpu_map)
6440 arch_update_cpu_topology();
6442 doms_cur = alloc_sched_domains(ndoms_cur);
6444 doms_cur = &fallback_doms;
6445 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6446 err = build_sched_domains(doms_cur[0], NULL);
6447 register_sched_domain_sysctl();
6453 * Detach sched domains from a group of cpus specified in cpu_map
6454 * These cpus will now be attached to the NULL domain
6456 static void detach_destroy_domains(const struct cpumask *cpu_map)
6461 for_each_cpu(i, cpu_map)
6462 cpu_attach_domain(NULL, &def_root_domain, i);
6466 /* handle null as "default" */
6467 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6468 struct sched_domain_attr *new, int idx_new)
6470 struct sched_domain_attr tmp;
6477 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6478 new ? (new + idx_new) : &tmp,
6479 sizeof(struct sched_domain_attr));
6483 * Partition sched domains as specified by the 'ndoms_new'
6484 * cpumasks in the array doms_new[] of cpumasks. This compares
6485 * doms_new[] to the current sched domain partitioning, doms_cur[].
6486 * It destroys each deleted domain and builds each new domain.
6488 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6489 * The masks don't intersect (don't overlap.) We should setup one
6490 * sched domain for each mask. CPUs not in any of the cpumasks will
6491 * not be load balanced. If the same cpumask appears both in the
6492 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6495 * The passed in 'doms_new' should be allocated using
6496 * alloc_sched_domains. This routine takes ownership of it and will
6497 * free_sched_domains it when done with it. If the caller failed the
6498 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6499 * and partition_sched_domains() will fallback to the single partition
6500 * 'fallback_doms', it also forces the domains to be rebuilt.
6502 * If doms_new == NULL it will be replaced with cpu_online_mask.
6503 * ndoms_new == 0 is a special case for destroying existing domains,
6504 * and it will not create the default domain.
6506 * Call with hotplug lock held
6508 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6509 struct sched_domain_attr *dattr_new)
6514 mutex_lock(&sched_domains_mutex);
6516 /* always unregister in case we don't destroy any domains */
6517 unregister_sched_domain_sysctl();
6519 /* Let architecture update cpu core mappings. */
6520 new_topology = arch_update_cpu_topology();
6522 n = doms_new ? ndoms_new : 0;
6524 /* Destroy deleted domains */
6525 for (i = 0; i < ndoms_cur; i++) {
6526 for (j = 0; j < n && !new_topology; j++) {
6527 if (cpumask_equal(doms_cur[i], doms_new[j])
6528 && dattrs_equal(dattr_cur, i, dattr_new, j))
6531 /* no match - a current sched domain not in new doms_new[] */
6532 detach_destroy_domains(doms_cur[i]);
6537 if (doms_new == NULL) {
6539 doms_new = &fallback_doms;
6540 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6541 WARN_ON_ONCE(dattr_new);
6544 /* Build new domains */
6545 for (i = 0; i < ndoms_new; i++) {
6546 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6547 if (cpumask_equal(doms_new[i], doms_cur[j])
6548 && dattrs_equal(dattr_new, i, dattr_cur, j))
6551 /* no match - add a new doms_new */
6552 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6557 /* Remember the new sched domains */
6558 if (doms_cur != &fallback_doms)
6559 free_sched_domains(doms_cur, ndoms_cur);
6560 kfree(dattr_cur); /* kfree(NULL) is safe */
6561 doms_cur = doms_new;
6562 dattr_cur = dattr_new;
6563 ndoms_cur = ndoms_new;
6565 register_sched_domain_sysctl();
6567 mutex_unlock(&sched_domains_mutex);
6570 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6573 * Update cpusets according to cpu_active mask. If cpusets are
6574 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6575 * around partition_sched_domains().
6577 * If we come here as part of a suspend/resume, don't touch cpusets because we
6578 * want to restore it back to its original state upon resume anyway.
6580 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6584 case CPU_ONLINE_FROZEN:
6585 case CPU_DOWN_FAILED_FROZEN:
6588 * num_cpus_frozen tracks how many CPUs are involved in suspend
6589 * resume sequence. As long as this is not the last online
6590 * operation in the resume sequence, just build a single sched
6591 * domain, ignoring cpusets.
6594 if (likely(num_cpus_frozen)) {
6595 partition_sched_domains(1, NULL, NULL);
6600 * This is the last CPU online operation. So fall through and
6601 * restore the original sched domains by considering the
6602 * cpuset configurations.
6606 case CPU_DOWN_FAILED:
6607 cpuset_update_active_cpus(true);
6615 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6619 case CPU_DOWN_PREPARE:
6620 cpuset_update_active_cpus(false);
6622 case CPU_DOWN_PREPARE_FROZEN:
6624 partition_sched_domains(1, NULL, NULL);
6632 void __init sched_init_smp(void)
6634 cpumask_var_t non_isolated_cpus;
6636 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6637 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6642 mutex_lock(&sched_domains_mutex);
6643 init_sched_domains(cpu_active_mask);
6644 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6645 if (cpumask_empty(non_isolated_cpus))
6646 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6647 mutex_unlock(&sched_domains_mutex);
6650 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6651 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6653 /* RT runtime code needs to handle some hotplug events */
6654 hotcpu_notifier(update_runtime, 0);
6658 /* Move init over to a non-isolated CPU */
6659 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6661 sched_init_granularity();
6662 free_cpumask_var(non_isolated_cpus);
6664 init_sched_rt_class();
6667 void __init sched_init_smp(void)
6669 sched_init_granularity();
6671 #endif /* CONFIG_SMP */
6673 const_debug unsigned int sysctl_timer_migration = 1;
6675 int in_sched_functions(unsigned long addr)
6677 return in_lock_functions(addr) ||
6678 (addr >= (unsigned long)__sched_text_start
6679 && addr < (unsigned long)__sched_text_end);
6682 #ifdef CONFIG_CGROUP_SCHED
6683 struct task_group root_task_group;
6684 LIST_HEAD(task_groups);
6687 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6689 void __init sched_init(void)
6692 unsigned long alloc_size = 0, ptr;
6694 #ifdef CONFIG_FAIR_GROUP_SCHED
6695 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6697 #ifdef CONFIG_RT_GROUP_SCHED
6698 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6700 #ifdef CONFIG_CPUMASK_OFFSTACK
6701 alloc_size += num_possible_cpus() * cpumask_size();
6704 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6706 #ifdef CONFIG_FAIR_GROUP_SCHED
6707 root_task_group.se = (struct sched_entity **)ptr;
6708 ptr += nr_cpu_ids * sizeof(void **);
6710 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6711 ptr += nr_cpu_ids * sizeof(void **);
6713 #endif /* CONFIG_FAIR_GROUP_SCHED */
6714 #ifdef CONFIG_RT_GROUP_SCHED
6715 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6716 ptr += nr_cpu_ids * sizeof(void **);
6718 root_task_group.rt_rq = (struct rt_rq **)ptr;
6719 ptr += nr_cpu_ids * sizeof(void **);
6721 #endif /* CONFIG_RT_GROUP_SCHED */
6722 #ifdef CONFIG_CPUMASK_OFFSTACK
6723 for_each_possible_cpu(i) {
6724 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6725 ptr += cpumask_size();
6727 #endif /* CONFIG_CPUMASK_OFFSTACK */
6731 init_defrootdomain();
6734 init_rt_bandwidth(&def_rt_bandwidth,
6735 global_rt_period(), global_rt_runtime());
6737 #ifdef CONFIG_RT_GROUP_SCHED
6738 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6739 global_rt_period(), global_rt_runtime());
6740 #endif /* CONFIG_RT_GROUP_SCHED */
6742 #ifdef CONFIG_CGROUP_SCHED
6743 list_add(&root_task_group.list, &task_groups);
6744 INIT_LIST_HEAD(&root_task_group.children);
6745 INIT_LIST_HEAD(&root_task_group.siblings);
6746 autogroup_init(&init_task);
6748 #endif /* CONFIG_CGROUP_SCHED */
6750 #ifdef CONFIG_CGROUP_CPUACCT
6751 root_cpuacct.cpustat = &kernel_cpustat;
6752 root_cpuacct.cpuusage = alloc_percpu(u64);
6753 /* Too early, not expected to fail */
6754 BUG_ON(!root_cpuacct.cpuusage);
6756 for_each_possible_cpu(i) {
6760 raw_spin_lock_init(&rq->lock);
6762 rq->calc_load_active = 0;
6763 rq->calc_load_update = jiffies + LOAD_FREQ;
6764 init_cfs_rq(&rq->cfs);
6765 init_rt_rq(&rq->rt, rq);
6766 #ifdef CONFIG_FAIR_GROUP_SCHED
6767 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6768 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6770 * How much cpu bandwidth does root_task_group get?
6772 * In case of task-groups formed thr' the cgroup filesystem, it
6773 * gets 100% of the cpu resources in the system. This overall
6774 * system cpu resource is divided among the tasks of
6775 * root_task_group and its child task-groups in a fair manner,
6776 * based on each entity's (task or task-group's) weight
6777 * (se->load.weight).
6779 * In other words, if root_task_group has 10 tasks of weight
6780 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6781 * then A0's share of the cpu resource is:
6783 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6785 * We achieve this by letting root_task_group's tasks sit
6786 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6788 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6789 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6790 #endif /* CONFIG_FAIR_GROUP_SCHED */
6792 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6793 #ifdef CONFIG_RT_GROUP_SCHED
6794 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6795 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6798 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6799 rq->cpu_load[j] = 0;
6801 rq->last_load_update_tick = jiffies;
6806 rq->cpu_power = SCHED_POWER_SCALE;
6807 rq->post_schedule = 0;
6808 rq->active_balance = 0;
6809 rq->next_balance = jiffies;
6814 rq->avg_idle = 2*sysctl_sched_migration_cost;
6816 INIT_LIST_HEAD(&rq->cfs_tasks);
6818 rq_attach_root(rq, &def_root_domain);
6824 atomic_set(&rq->nr_iowait, 0);
6827 set_load_weight(&init_task);
6829 #ifdef CONFIG_PREEMPT_NOTIFIERS
6830 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6833 #ifdef CONFIG_RT_MUTEXES
6834 plist_head_init(&init_task.pi_waiters);
6838 * The boot idle thread does lazy MMU switching as well:
6840 atomic_inc(&init_mm.mm_count);
6841 enter_lazy_tlb(&init_mm, current);
6844 * Make us the idle thread. Technically, schedule() should not be
6845 * called from this thread, however somewhere below it might be,
6846 * but because we are the idle thread, we just pick up running again
6847 * when this runqueue becomes "idle".
6849 init_idle(current, smp_processor_id());
6851 calc_load_update = jiffies + LOAD_FREQ;
6854 * During early bootup we pretend to be a normal task:
6856 current->sched_class = &fair_sched_class;
6859 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6860 /* May be allocated at isolcpus cmdline parse time */
6861 if (cpu_isolated_map == NULL)
6862 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6863 idle_thread_set_boot_cpu();
6865 init_sched_fair_class();
6867 scheduler_running = 1;
6870 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6871 static inline int preempt_count_equals(int preempt_offset)
6873 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6875 return (nested == preempt_offset);
6878 void __might_sleep(const char *file, int line, int preempt_offset)
6880 static unsigned long prev_jiffy; /* ratelimiting */
6882 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6883 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6884 system_state != SYSTEM_RUNNING || oops_in_progress)
6886 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6888 prev_jiffy = jiffies;
6891 "BUG: sleeping function called from invalid context at %s:%d\n",
6894 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6895 in_atomic(), irqs_disabled(),
6896 current->pid, current->comm);
6898 debug_show_held_locks(current);
6899 if (irqs_disabled())
6900 print_irqtrace_events(current);
6903 EXPORT_SYMBOL(__might_sleep);
6906 #ifdef CONFIG_MAGIC_SYSRQ
6907 static void normalize_task(struct rq *rq, struct task_struct *p)
6909 const struct sched_class *prev_class = p->sched_class;
6910 int old_prio = p->prio;
6915 dequeue_task(rq, p, 0);
6916 __setscheduler(rq, p, SCHED_NORMAL, 0);
6918 enqueue_task(rq, p, 0);
6919 resched_task(rq->curr);
6922 check_class_changed(rq, p, prev_class, old_prio);
6925 void normalize_rt_tasks(void)
6927 struct task_struct *g, *p;
6928 unsigned long flags;
6931 read_lock_irqsave(&tasklist_lock, flags);
6932 do_each_thread(g, p) {
6934 * Only normalize user tasks:
6939 p->se.exec_start = 0;
6940 #ifdef CONFIG_SCHEDSTATS
6941 p->se.statistics.wait_start = 0;
6942 p->se.statistics.sleep_start = 0;
6943 p->se.statistics.block_start = 0;
6948 * Renice negative nice level userspace
6951 if (TASK_NICE(p) < 0 && p->mm)
6952 set_user_nice(p, 0);
6956 raw_spin_lock(&p->pi_lock);
6957 rq = __task_rq_lock(p);
6959 normalize_task(rq, p);
6961 __task_rq_unlock(rq);
6962 raw_spin_unlock(&p->pi_lock);
6963 } while_each_thread(g, p);
6965 read_unlock_irqrestore(&tasklist_lock, flags);
6968 #endif /* CONFIG_MAGIC_SYSRQ */
6970 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6972 * These functions are only useful for the IA64 MCA handling, or kdb.
6974 * They can only be called when the whole system has been
6975 * stopped - every CPU needs to be quiescent, and no scheduling
6976 * activity can take place. Using them for anything else would
6977 * be a serious bug, and as a result, they aren't even visible
6978 * under any other configuration.
6982 * curr_task - return the current task for a given cpu.
6983 * @cpu: the processor in question.
6985 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6987 struct task_struct *curr_task(int cpu)
6989 return cpu_curr(cpu);
6992 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6996 * set_curr_task - set the current task for a given cpu.
6997 * @cpu: the processor in question.
6998 * @p: the task pointer to set.
7000 * Description: This function must only be used when non-maskable interrupts
7001 * are serviced on a separate stack. It allows the architecture to switch the
7002 * notion of the current task on a cpu in a non-blocking manner. This function
7003 * must be called with all CPU's synchronized, and interrupts disabled, the
7004 * and caller must save the original value of the current task (see
7005 * curr_task() above) and restore that value before reenabling interrupts and
7006 * re-starting the system.
7008 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7010 void set_curr_task(int cpu, struct task_struct *p)
7017 #ifdef CONFIG_CGROUP_SCHED
7018 /* task_group_lock serializes the addition/removal of task groups */
7019 static DEFINE_SPINLOCK(task_group_lock);
7021 static void free_sched_group(struct task_group *tg)
7023 free_fair_sched_group(tg);
7024 free_rt_sched_group(tg);
7029 /* allocate runqueue etc for a new task group */
7030 struct task_group *sched_create_group(struct task_group *parent)
7032 struct task_group *tg;
7033 unsigned long flags;
7035 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7037 return ERR_PTR(-ENOMEM);
7039 if (!alloc_fair_sched_group(tg, parent))
7042 if (!alloc_rt_sched_group(tg, parent))
7045 spin_lock_irqsave(&task_group_lock, flags);
7046 list_add_rcu(&tg->list, &task_groups);
7048 WARN_ON(!parent); /* root should already exist */
7050 tg->parent = parent;
7051 INIT_LIST_HEAD(&tg->children);
7052 list_add_rcu(&tg->siblings, &parent->children);
7053 spin_unlock_irqrestore(&task_group_lock, flags);
7058 free_sched_group(tg);
7059 return ERR_PTR(-ENOMEM);
7062 /* rcu callback to free various structures associated with a task group */
7063 static void free_sched_group_rcu(struct rcu_head *rhp)
7065 /* now it should be safe to free those cfs_rqs */
7066 free_sched_group(container_of(rhp, struct task_group, rcu));
7069 /* Destroy runqueue etc associated with a task group */
7070 void sched_destroy_group(struct task_group *tg)
7072 unsigned long flags;
7075 /* end participation in shares distribution */
7076 for_each_possible_cpu(i)
7077 unregister_fair_sched_group(tg, i);
7079 spin_lock_irqsave(&task_group_lock, flags);
7080 list_del_rcu(&tg->list);
7081 list_del_rcu(&tg->siblings);
7082 spin_unlock_irqrestore(&task_group_lock, flags);
7084 /* wait for possible concurrent references to cfs_rqs complete */
7085 call_rcu(&tg->rcu, free_sched_group_rcu);
7088 /* change task's runqueue when it moves between groups.
7089 * The caller of this function should have put the task in its new group
7090 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7091 * reflect its new group.
7093 void sched_move_task(struct task_struct *tsk)
7095 struct task_group *tg;
7097 unsigned long flags;
7100 rq = task_rq_lock(tsk, &flags);
7102 running = task_current(rq, tsk);
7106 dequeue_task(rq, tsk, 0);
7107 if (unlikely(running))
7108 tsk->sched_class->put_prev_task(rq, tsk);
7110 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7111 lockdep_is_held(&tsk->sighand->siglock)),
7112 struct task_group, css);
7113 tg = autogroup_task_group(tsk, tg);
7114 tsk->sched_task_group = tg;
7116 #ifdef CONFIG_FAIR_GROUP_SCHED
7117 if (tsk->sched_class->task_move_group)
7118 tsk->sched_class->task_move_group(tsk, on_rq);
7121 set_task_rq(tsk, task_cpu(tsk));
7123 if (unlikely(running))
7124 tsk->sched_class->set_curr_task(rq);
7126 enqueue_task(rq, tsk, 0);
7128 task_rq_unlock(rq, tsk, &flags);
7130 #endif /* CONFIG_CGROUP_SCHED */
7132 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7133 static unsigned long to_ratio(u64 period, u64 runtime)
7135 if (runtime == RUNTIME_INF)
7138 return div64_u64(runtime << 20, period);
7142 #ifdef CONFIG_RT_GROUP_SCHED
7144 * Ensure that the real time constraints are schedulable.
7146 static DEFINE_MUTEX(rt_constraints_mutex);
7148 /* Must be called with tasklist_lock held */
7149 static inline int tg_has_rt_tasks(struct task_group *tg)
7151 struct task_struct *g, *p;
7153 do_each_thread(g, p) {
7154 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7156 } while_each_thread(g, p);
7161 struct rt_schedulable_data {
7162 struct task_group *tg;
7167 static int tg_rt_schedulable(struct task_group *tg, void *data)
7169 struct rt_schedulable_data *d = data;
7170 struct task_group *child;
7171 unsigned long total, sum = 0;
7172 u64 period, runtime;
7174 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7175 runtime = tg->rt_bandwidth.rt_runtime;
7178 period = d->rt_period;
7179 runtime = d->rt_runtime;
7183 * Cannot have more runtime than the period.
7185 if (runtime > period && runtime != RUNTIME_INF)
7189 * Ensure we don't starve existing RT tasks.
7191 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7194 total = to_ratio(period, runtime);
7197 * Nobody can have more than the global setting allows.
7199 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7203 * The sum of our children's runtime should not exceed our own.
7205 list_for_each_entry_rcu(child, &tg->children, siblings) {
7206 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7207 runtime = child->rt_bandwidth.rt_runtime;
7209 if (child == d->tg) {
7210 period = d->rt_period;
7211 runtime = d->rt_runtime;
7214 sum += to_ratio(period, runtime);
7223 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7227 struct rt_schedulable_data data = {
7229 .rt_period = period,
7230 .rt_runtime = runtime,
7234 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7240 static int tg_set_rt_bandwidth(struct task_group *tg,
7241 u64 rt_period, u64 rt_runtime)
7245 mutex_lock(&rt_constraints_mutex);
7246 read_lock(&tasklist_lock);
7247 err = __rt_schedulable(tg, rt_period, rt_runtime);
7251 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7252 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7253 tg->rt_bandwidth.rt_runtime = rt_runtime;
7255 for_each_possible_cpu(i) {
7256 struct rt_rq *rt_rq = tg->rt_rq[i];
7258 raw_spin_lock(&rt_rq->rt_runtime_lock);
7259 rt_rq->rt_runtime = rt_runtime;
7260 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7262 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7264 read_unlock(&tasklist_lock);
7265 mutex_unlock(&rt_constraints_mutex);
7270 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7272 u64 rt_runtime, rt_period;
7274 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7275 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7276 if (rt_runtime_us < 0)
7277 rt_runtime = RUNTIME_INF;
7279 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7282 long sched_group_rt_runtime(struct task_group *tg)
7286 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7289 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7290 do_div(rt_runtime_us, NSEC_PER_USEC);
7291 return rt_runtime_us;
7294 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7296 u64 rt_runtime, rt_period;
7298 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7299 rt_runtime = tg->rt_bandwidth.rt_runtime;
7304 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7307 long sched_group_rt_period(struct task_group *tg)
7311 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7312 do_div(rt_period_us, NSEC_PER_USEC);
7313 return rt_period_us;
7316 static int sched_rt_global_constraints(void)
7318 u64 runtime, period;
7321 if (sysctl_sched_rt_period <= 0)
7324 runtime = global_rt_runtime();
7325 period = global_rt_period();
7328 * Sanity check on the sysctl variables.
7330 if (runtime > period && runtime != RUNTIME_INF)
7333 mutex_lock(&rt_constraints_mutex);
7334 read_lock(&tasklist_lock);
7335 ret = __rt_schedulable(NULL, 0, 0);
7336 read_unlock(&tasklist_lock);
7337 mutex_unlock(&rt_constraints_mutex);
7342 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7344 /* Don't accept realtime tasks when there is no way for them to run */
7345 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7351 #else /* !CONFIG_RT_GROUP_SCHED */
7352 static int sched_rt_global_constraints(void)
7354 unsigned long flags;
7357 if (sysctl_sched_rt_period <= 0)
7361 * There's always some RT tasks in the root group
7362 * -- migration, kstopmachine etc..
7364 if (sysctl_sched_rt_runtime == 0)
7367 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7368 for_each_possible_cpu(i) {
7369 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7371 raw_spin_lock(&rt_rq->rt_runtime_lock);
7372 rt_rq->rt_runtime = global_rt_runtime();
7373 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7375 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7379 #endif /* CONFIG_RT_GROUP_SCHED */
7381 int sched_rt_handler(struct ctl_table *table, int write,
7382 void __user *buffer, size_t *lenp,
7386 int old_period, old_runtime;
7387 static DEFINE_MUTEX(mutex);
7390 old_period = sysctl_sched_rt_period;
7391 old_runtime = sysctl_sched_rt_runtime;
7393 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7395 if (!ret && write) {
7396 ret = sched_rt_global_constraints();
7398 sysctl_sched_rt_period = old_period;
7399 sysctl_sched_rt_runtime = old_runtime;
7401 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7402 def_rt_bandwidth.rt_period =
7403 ns_to_ktime(global_rt_period());
7406 mutex_unlock(&mutex);
7411 #ifdef CONFIG_CGROUP_SCHED
7413 /* return corresponding task_group object of a cgroup */
7414 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7416 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7417 struct task_group, css);
7420 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7422 struct task_group *tg, *parent;
7424 if (!cgrp->parent) {
7425 /* This is early initialization for the top cgroup */
7426 return &root_task_group.css;
7429 parent = cgroup_tg(cgrp->parent);
7430 tg = sched_create_group(parent);
7432 return ERR_PTR(-ENOMEM);
7437 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7439 struct task_group *tg = cgroup_tg(cgrp);
7441 sched_destroy_group(tg);
7444 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7445 struct cgroup_taskset *tset)
7447 struct task_struct *task;
7449 cgroup_taskset_for_each(task, cgrp, tset) {
7450 #ifdef CONFIG_RT_GROUP_SCHED
7451 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7454 /* We don't support RT-tasks being in separate groups */
7455 if (task->sched_class != &fair_sched_class)
7462 static void cpu_cgroup_attach(struct cgroup *cgrp,
7463 struct cgroup_taskset *tset)
7465 struct task_struct *task;
7467 cgroup_taskset_for_each(task, cgrp, tset)
7468 sched_move_task(task);
7472 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7473 struct task_struct *task)
7476 * cgroup_exit() is called in the copy_process() failure path.
7477 * Ignore this case since the task hasn't ran yet, this avoids
7478 * trying to poke a half freed task state from generic code.
7480 if (!(task->flags & PF_EXITING))
7483 sched_move_task(task);
7486 #ifdef CONFIG_FAIR_GROUP_SCHED
7487 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7490 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7493 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7495 struct task_group *tg = cgroup_tg(cgrp);
7497 return (u64) scale_load_down(tg->shares);
7500 #ifdef CONFIG_CFS_BANDWIDTH
7501 static DEFINE_MUTEX(cfs_constraints_mutex);
7503 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7504 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7506 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7508 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7510 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7511 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7513 if (tg == &root_task_group)
7517 * Ensure we have at some amount of bandwidth every period. This is
7518 * to prevent reaching a state of large arrears when throttled via
7519 * entity_tick() resulting in prolonged exit starvation.
7521 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7525 * Likewise, bound things on the otherside by preventing insane quota
7526 * periods. This also allows us to normalize in computing quota
7529 if (period > max_cfs_quota_period)
7532 mutex_lock(&cfs_constraints_mutex);
7533 ret = __cfs_schedulable(tg, period, quota);
7537 runtime_enabled = quota != RUNTIME_INF;
7538 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7539 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7540 raw_spin_lock_irq(&cfs_b->lock);
7541 cfs_b->period = ns_to_ktime(period);
7542 cfs_b->quota = quota;
7544 __refill_cfs_bandwidth_runtime(cfs_b);
7545 /* restart the period timer (if active) to handle new period expiry */
7546 if (runtime_enabled && cfs_b->timer_active) {
7547 /* force a reprogram */
7548 cfs_b->timer_active = 0;
7549 __start_cfs_bandwidth(cfs_b);
7551 raw_spin_unlock_irq(&cfs_b->lock);
7553 for_each_possible_cpu(i) {
7554 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7555 struct rq *rq = cfs_rq->rq;
7557 raw_spin_lock_irq(&rq->lock);
7558 cfs_rq->runtime_enabled = runtime_enabled;
7559 cfs_rq->runtime_remaining = 0;
7561 if (cfs_rq->throttled)
7562 unthrottle_cfs_rq(cfs_rq);
7563 raw_spin_unlock_irq(&rq->lock);
7566 mutex_unlock(&cfs_constraints_mutex);
7571 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7575 period = ktime_to_ns(tg->cfs_bandwidth.period);
7576 if (cfs_quota_us < 0)
7577 quota = RUNTIME_INF;
7579 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7581 return tg_set_cfs_bandwidth(tg, period, quota);
7584 long tg_get_cfs_quota(struct task_group *tg)
7588 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7591 quota_us = tg->cfs_bandwidth.quota;
7592 do_div(quota_us, NSEC_PER_USEC);
7597 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7601 period = (u64)cfs_period_us * NSEC_PER_USEC;
7602 quota = tg->cfs_bandwidth.quota;
7604 return tg_set_cfs_bandwidth(tg, period, quota);
7607 long tg_get_cfs_period(struct task_group *tg)
7611 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7612 do_div(cfs_period_us, NSEC_PER_USEC);
7614 return cfs_period_us;
7617 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7619 return tg_get_cfs_quota(cgroup_tg(cgrp));
7622 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7625 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7628 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7630 return tg_get_cfs_period(cgroup_tg(cgrp));
7633 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7636 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7639 struct cfs_schedulable_data {
7640 struct task_group *tg;
7645 * normalize group quota/period to be quota/max_period
7646 * note: units are usecs
7648 static u64 normalize_cfs_quota(struct task_group *tg,
7649 struct cfs_schedulable_data *d)
7657 period = tg_get_cfs_period(tg);
7658 quota = tg_get_cfs_quota(tg);
7661 /* note: these should typically be equivalent */
7662 if (quota == RUNTIME_INF || quota == -1)
7665 return to_ratio(period, quota);
7668 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7670 struct cfs_schedulable_data *d = data;
7671 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7672 s64 quota = 0, parent_quota = -1;
7675 quota = RUNTIME_INF;
7677 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7679 quota = normalize_cfs_quota(tg, d);
7680 parent_quota = parent_b->hierarchal_quota;
7683 * ensure max(child_quota) <= parent_quota, inherit when no
7686 if (quota == RUNTIME_INF)
7687 quota = parent_quota;
7688 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7691 cfs_b->hierarchal_quota = quota;
7696 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7699 struct cfs_schedulable_data data = {
7705 if (quota != RUNTIME_INF) {
7706 do_div(data.period, NSEC_PER_USEC);
7707 do_div(data.quota, NSEC_PER_USEC);
7711 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7717 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7718 struct cgroup_map_cb *cb)
7720 struct task_group *tg = cgroup_tg(cgrp);
7721 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7723 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7724 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7725 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7729 #endif /* CONFIG_CFS_BANDWIDTH */
7730 #endif /* CONFIG_FAIR_GROUP_SCHED */
7732 #ifdef CONFIG_RT_GROUP_SCHED
7733 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7736 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7739 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7741 return sched_group_rt_runtime(cgroup_tg(cgrp));
7744 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7747 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7750 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7752 return sched_group_rt_period(cgroup_tg(cgrp));
7754 #endif /* CONFIG_RT_GROUP_SCHED */
7756 static struct cftype cpu_files[] = {
7757 #ifdef CONFIG_FAIR_GROUP_SCHED
7760 .read_u64 = cpu_shares_read_u64,
7761 .write_u64 = cpu_shares_write_u64,
7764 #ifdef CONFIG_CFS_BANDWIDTH
7766 .name = "cfs_quota_us",
7767 .read_s64 = cpu_cfs_quota_read_s64,
7768 .write_s64 = cpu_cfs_quota_write_s64,
7771 .name = "cfs_period_us",
7772 .read_u64 = cpu_cfs_period_read_u64,
7773 .write_u64 = cpu_cfs_period_write_u64,
7777 .read_map = cpu_stats_show,
7780 #ifdef CONFIG_RT_GROUP_SCHED
7782 .name = "rt_runtime_us",
7783 .read_s64 = cpu_rt_runtime_read,
7784 .write_s64 = cpu_rt_runtime_write,
7787 .name = "rt_period_us",
7788 .read_u64 = cpu_rt_period_read_uint,
7789 .write_u64 = cpu_rt_period_write_uint,
7795 struct cgroup_subsys cpu_cgroup_subsys = {
7797 .create = cpu_cgroup_create,
7798 .destroy = cpu_cgroup_destroy,
7799 .can_attach = cpu_cgroup_can_attach,
7800 .attach = cpu_cgroup_attach,
7801 .exit = cpu_cgroup_exit,
7802 .subsys_id = cpu_cgroup_subsys_id,
7803 .base_cftypes = cpu_files,
7807 #endif /* CONFIG_CGROUP_SCHED */
7809 #ifdef CONFIG_CGROUP_CPUACCT
7812 * CPU accounting code for task groups.
7814 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7815 * (balbir@in.ibm.com).
7818 struct cpuacct root_cpuacct;
7820 /* create a new cpu accounting group */
7821 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
7826 return &root_cpuacct.css;
7828 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7832 ca->cpuusage = alloc_percpu(u64);
7836 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7838 goto out_free_cpuusage;
7843 free_percpu(ca->cpuusage);
7847 return ERR_PTR(-ENOMEM);
7850 /* destroy an existing cpu accounting group */
7851 static void cpuacct_destroy(struct cgroup *cgrp)
7853 struct cpuacct *ca = cgroup_ca(cgrp);
7855 free_percpu(ca->cpustat);
7856 free_percpu(ca->cpuusage);
7860 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7862 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7865 #ifndef CONFIG_64BIT
7867 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7869 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7871 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7879 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7881 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7883 #ifndef CONFIG_64BIT
7885 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7887 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7889 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7895 /* return total cpu usage (in nanoseconds) of a group */
7896 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7898 struct cpuacct *ca = cgroup_ca(cgrp);
7899 u64 totalcpuusage = 0;
7902 for_each_present_cpu(i)
7903 totalcpuusage += cpuacct_cpuusage_read(ca, i);
7905 return totalcpuusage;
7908 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
7911 struct cpuacct *ca = cgroup_ca(cgrp);
7920 for_each_present_cpu(i)
7921 cpuacct_cpuusage_write(ca, i, 0);
7927 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
7930 struct cpuacct *ca = cgroup_ca(cgroup);
7934 for_each_present_cpu(i) {
7935 percpu = cpuacct_cpuusage_read(ca, i);
7936 seq_printf(m, "%llu ", (unsigned long long) percpu);
7938 seq_printf(m, "\n");
7942 static const char *cpuacct_stat_desc[] = {
7943 [CPUACCT_STAT_USER] = "user",
7944 [CPUACCT_STAT_SYSTEM] = "system",
7947 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
7948 struct cgroup_map_cb *cb)
7950 struct cpuacct *ca = cgroup_ca(cgrp);
7954 for_each_online_cpu(cpu) {
7955 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
7956 val += kcpustat->cpustat[CPUTIME_USER];
7957 val += kcpustat->cpustat[CPUTIME_NICE];
7959 val = cputime64_to_clock_t(val);
7960 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
7963 for_each_online_cpu(cpu) {
7964 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
7965 val += kcpustat->cpustat[CPUTIME_SYSTEM];
7966 val += kcpustat->cpustat[CPUTIME_IRQ];
7967 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
7970 val = cputime64_to_clock_t(val);
7971 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
7976 static struct cftype files[] = {
7979 .read_u64 = cpuusage_read,
7980 .write_u64 = cpuusage_write,
7983 .name = "usage_percpu",
7984 .read_seq_string = cpuacct_percpu_seq_read,
7988 .read_map = cpuacct_stats_show,
7994 * charge this task's execution time to its accounting group.
7996 * called with rq->lock held.
7998 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8003 if (unlikely(!cpuacct_subsys.active))
8006 cpu = task_cpu(tsk);
8012 for (; ca; ca = parent_ca(ca)) {
8013 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8014 *cpuusage += cputime;
8020 struct cgroup_subsys cpuacct_subsys = {
8022 .create = cpuacct_create,
8023 .destroy = cpuacct_destroy,
8024 .subsys_id = cpuacct_subsys_id,
8025 .base_cftypes = files,
8027 #endif /* CONFIG_CGROUP_CPUACCT */