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) 0
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 if (p->sched_class->migrate_task_rq)
956 p->sched_class->migrate_task_rq(p, new_cpu);
957 p->se.nr_migrations++;
958 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
961 __set_task_cpu(p, new_cpu);
964 struct migration_arg {
965 struct task_struct *task;
969 static int migration_cpu_stop(void *data);
972 * wait_task_inactive - wait for a thread to unschedule.
974 * If @match_state is nonzero, it's the @p->state value just checked and
975 * not expected to change. If it changes, i.e. @p might have woken up,
976 * then return zero. When we succeed in waiting for @p to be off its CPU,
977 * we return a positive number (its total switch count). If a second call
978 * a short while later returns the same number, the caller can be sure that
979 * @p has remained unscheduled the whole time.
981 * The caller must ensure that the task *will* unschedule sometime soon,
982 * else this function might spin for a *long* time. This function can't
983 * be called with interrupts off, or it may introduce deadlock with
984 * smp_call_function() if an IPI is sent by the same process we are
985 * waiting to become inactive.
987 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
996 * We do the initial early heuristics without holding
997 * any task-queue locks at all. We'll only try to get
998 * the runqueue lock when things look like they will
1004 * If the task is actively running on another CPU
1005 * still, just relax and busy-wait without holding
1008 * NOTE! Since we don't hold any locks, it's not
1009 * even sure that "rq" stays as the right runqueue!
1010 * But we don't care, since "task_running()" will
1011 * return false if the runqueue has changed and p
1012 * is actually now running somewhere else!
1014 while (task_running(rq, p)) {
1015 if (match_state && unlikely(p->state != match_state))
1021 * Ok, time to look more closely! We need the rq
1022 * lock now, to be *sure*. If we're wrong, we'll
1023 * just go back and repeat.
1025 rq = task_rq_lock(p, &flags);
1026 trace_sched_wait_task(p);
1027 running = task_running(rq, p);
1030 if (!match_state || p->state == match_state)
1031 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1032 task_rq_unlock(rq, p, &flags);
1035 * If it changed from the expected state, bail out now.
1037 if (unlikely(!ncsw))
1041 * Was it really running after all now that we
1042 * checked with the proper locks actually held?
1044 * Oops. Go back and try again..
1046 if (unlikely(running)) {
1052 * It's not enough that it's not actively running,
1053 * it must be off the runqueue _entirely_, and not
1056 * So if it was still runnable (but just not actively
1057 * running right now), it's preempted, and we should
1058 * yield - it could be a while.
1060 if (unlikely(on_rq)) {
1061 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1063 set_current_state(TASK_UNINTERRUPTIBLE);
1064 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1069 * Ahh, all good. It wasn't running, and it wasn't
1070 * runnable, which means that it will never become
1071 * running in the future either. We're all done!
1080 * kick_process - kick a running thread to enter/exit the kernel
1081 * @p: the to-be-kicked thread
1083 * Cause a process which is running on another CPU to enter
1084 * kernel-mode, without any delay. (to get signals handled.)
1086 * NOTE: this function doesn't have to take the runqueue lock,
1087 * because all it wants to ensure is that the remote task enters
1088 * the kernel. If the IPI races and the task has been migrated
1089 * to another CPU then no harm is done and the purpose has been
1092 void kick_process(struct task_struct *p)
1098 if ((cpu != smp_processor_id()) && task_curr(p))
1099 smp_send_reschedule(cpu);
1102 EXPORT_SYMBOL_GPL(kick_process);
1103 #endif /* CONFIG_SMP */
1107 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1109 static int select_fallback_rq(int cpu, struct task_struct *p)
1111 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1112 enum { cpuset, possible, fail } state = cpuset;
1115 /* Look for allowed, online CPU in same node. */
1116 for_each_cpu(dest_cpu, nodemask) {
1117 if (!cpu_online(dest_cpu))
1119 if (!cpu_active(dest_cpu))
1121 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1126 /* Any allowed, online CPU? */
1127 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1128 if (!cpu_online(dest_cpu))
1130 if (!cpu_active(dest_cpu))
1137 /* No more Mr. Nice Guy. */
1138 cpuset_cpus_allowed_fallback(p);
1143 do_set_cpus_allowed(p, cpu_possible_mask);
1154 if (state != cpuset) {
1156 * Don't tell them about moving exiting tasks or
1157 * kernel threads (both mm NULL), since they never
1160 if (p->mm && printk_ratelimit()) {
1161 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1162 task_pid_nr(p), p->comm, cpu);
1170 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1173 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1175 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1178 * In order not to call set_task_cpu() on a blocking task we need
1179 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1182 * Since this is common to all placement strategies, this lives here.
1184 * [ this allows ->select_task() to simply return task_cpu(p) and
1185 * not worry about this generic constraint ]
1187 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1189 cpu = select_fallback_rq(task_cpu(p), p);
1194 static void update_avg(u64 *avg, u64 sample)
1196 s64 diff = sample - *avg;
1202 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1204 #ifdef CONFIG_SCHEDSTATS
1205 struct rq *rq = this_rq();
1208 int this_cpu = smp_processor_id();
1210 if (cpu == this_cpu) {
1211 schedstat_inc(rq, ttwu_local);
1212 schedstat_inc(p, se.statistics.nr_wakeups_local);
1214 struct sched_domain *sd;
1216 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1218 for_each_domain(this_cpu, sd) {
1219 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1220 schedstat_inc(sd, ttwu_wake_remote);
1227 if (wake_flags & WF_MIGRATED)
1228 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1230 #endif /* CONFIG_SMP */
1232 schedstat_inc(rq, ttwu_count);
1233 schedstat_inc(p, se.statistics.nr_wakeups);
1235 if (wake_flags & WF_SYNC)
1236 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1238 #endif /* CONFIG_SCHEDSTATS */
1241 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1243 activate_task(rq, p, en_flags);
1246 /* if a worker is waking up, notify workqueue */
1247 if (p->flags & PF_WQ_WORKER)
1248 wq_worker_waking_up(p, cpu_of(rq));
1252 * Mark the task runnable and perform wakeup-preemption.
1255 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1257 trace_sched_wakeup(p, true);
1258 check_preempt_curr(rq, p, wake_flags);
1260 p->state = TASK_RUNNING;
1262 if (p->sched_class->task_woken)
1263 p->sched_class->task_woken(rq, p);
1265 if (rq->idle_stamp) {
1266 u64 delta = rq->clock - rq->idle_stamp;
1267 u64 max = 2*sysctl_sched_migration_cost;
1272 update_avg(&rq->avg_idle, delta);
1279 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1282 if (p->sched_contributes_to_load)
1283 rq->nr_uninterruptible--;
1286 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1287 ttwu_do_wakeup(rq, p, wake_flags);
1291 * Called in case the task @p isn't fully descheduled from its runqueue,
1292 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1293 * since all we need to do is flip p->state to TASK_RUNNING, since
1294 * the task is still ->on_rq.
1296 static int ttwu_remote(struct task_struct *p, int wake_flags)
1301 rq = __task_rq_lock(p);
1303 ttwu_do_wakeup(rq, p, wake_flags);
1306 __task_rq_unlock(rq);
1312 static void sched_ttwu_pending(void)
1314 struct rq *rq = this_rq();
1315 struct llist_node *llist = llist_del_all(&rq->wake_list);
1316 struct task_struct *p;
1318 raw_spin_lock(&rq->lock);
1321 p = llist_entry(llist, struct task_struct, wake_entry);
1322 llist = llist_next(llist);
1323 ttwu_do_activate(rq, p, 0);
1326 raw_spin_unlock(&rq->lock);
1329 void scheduler_ipi(void)
1331 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1335 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1336 * traditionally all their work was done from the interrupt return
1337 * path. Now that we actually do some work, we need to make sure
1340 * Some archs already do call them, luckily irq_enter/exit nest
1343 * Arguably we should visit all archs and update all handlers,
1344 * however a fair share of IPIs are still resched only so this would
1345 * somewhat pessimize the simple resched case.
1348 sched_ttwu_pending();
1351 * Check if someone kicked us for doing the nohz idle load balance.
1353 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1354 this_rq()->idle_balance = 1;
1355 raise_softirq_irqoff(SCHED_SOFTIRQ);
1360 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1362 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1363 smp_send_reschedule(cpu);
1366 bool cpus_share_cache(int this_cpu, int that_cpu)
1368 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1370 #endif /* CONFIG_SMP */
1372 static void ttwu_queue(struct task_struct *p, int cpu)
1374 struct rq *rq = cpu_rq(cpu);
1376 #if defined(CONFIG_SMP)
1377 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1378 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1379 ttwu_queue_remote(p, cpu);
1384 raw_spin_lock(&rq->lock);
1385 ttwu_do_activate(rq, p, 0);
1386 raw_spin_unlock(&rq->lock);
1390 * try_to_wake_up - wake up a thread
1391 * @p: the thread to be awakened
1392 * @state: the mask of task states that can be woken
1393 * @wake_flags: wake modifier flags (WF_*)
1395 * Put it on the run-queue if it's not already there. The "current"
1396 * thread is always on the run-queue (except when the actual
1397 * re-schedule is in progress), and as such you're allowed to do
1398 * the simpler "current->state = TASK_RUNNING" to mark yourself
1399 * runnable without the overhead of this.
1401 * Returns %true if @p was woken up, %false if it was already running
1402 * or @state didn't match @p's state.
1405 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1407 unsigned long flags;
1408 int cpu, success = 0;
1411 raw_spin_lock_irqsave(&p->pi_lock, flags);
1412 if (!(p->state & state))
1415 success = 1; /* we're going to change ->state */
1418 if (p->on_rq && ttwu_remote(p, wake_flags))
1423 * If the owning (remote) cpu is still in the middle of schedule() with
1424 * this task as prev, wait until its done referencing the task.
1429 * Pairs with the smp_wmb() in finish_lock_switch().
1433 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1434 p->state = TASK_WAKING;
1436 if (p->sched_class->task_waking)
1437 p->sched_class->task_waking(p);
1439 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1440 if (task_cpu(p) != cpu) {
1441 wake_flags |= WF_MIGRATED;
1442 set_task_cpu(p, cpu);
1444 #endif /* CONFIG_SMP */
1448 ttwu_stat(p, cpu, wake_flags);
1450 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1456 * try_to_wake_up_local - try to wake up a local task with rq lock held
1457 * @p: the thread to be awakened
1459 * Put @p on the run-queue if it's not already there. The caller must
1460 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1463 static void try_to_wake_up_local(struct task_struct *p)
1465 struct rq *rq = task_rq(p);
1467 BUG_ON(rq != this_rq());
1468 BUG_ON(p == current);
1469 lockdep_assert_held(&rq->lock);
1471 if (!raw_spin_trylock(&p->pi_lock)) {
1472 raw_spin_unlock(&rq->lock);
1473 raw_spin_lock(&p->pi_lock);
1474 raw_spin_lock(&rq->lock);
1477 if (!(p->state & TASK_NORMAL))
1481 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1483 ttwu_do_wakeup(rq, p, 0);
1484 ttwu_stat(p, smp_processor_id(), 0);
1486 raw_spin_unlock(&p->pi_lock);
1490 * wake_up_process - Wake up a specific process
1491 * @p: The process to be woken up.
1493 * Attempt to wake up the nominated process and move it to the set of runnable
1494 * processes. Returns 1 if the process was woken up, 0 if it was already
1497 * It may be assumed that this function implies a write memory barrier before
1498 * changing the task state if and only if any tasks are woken up.
1500 int wake_up_process(struct task_struct *p)
1502 return try_to_wake_up(p, TASK_ALL, 0);
1504 EXPORT_SYMBOL(wake_up_process);
1506 int wake_up_state(struct task_struct *p, unsigned int state)
1508 return try_to_wake_up(p, state, 0);
1512 * Perform scheduler related setup for a newly forked process p.
1513 * p is forked by current.
1515 * __sched_fork() is basic setup used by init_idle() too:
1517 static void __sched_fork(struct task_struct *p)
1522 p->se.exec_start = 0;
1523 p->se.sum_exec_runtime = 0;
1524 p->se.prev_sum_exec_runtime = 0;
1525 p->se.nr_migrations = 0;
1527 INIT_LIST_HEAD(&p->se.group_node);
1530 p->se.avg.runnable_avg_period = 0;
1531 p->se.avg.runnable_avg_sum = 0;
1533 #ifdef CONFIG_SCHEDSTATS
1534 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1537 INIT_LIST_HEAD(&p->rt.run_list);
1539 #ifdef CONFIG_PREEMPT_NOTIFIERS
1540 INIT_HLIST_HEAD(&p->preempt_notifiers);
1545 * fork()/clone()-time setup:
1547 void sched_fork(struct task_struct *p)
1549 unsigned long flags;
1550 int cpu = get_cpu();
1554 * We mark the process as running here. This guarantees that
1555 * nobody will actually run it, and a signal or other external
1556 * event cannot wake it up and insert it on the runqueue either.
1558 p->state = TASK_RUNNING;
1561 * Make sure we do not leak PI boosting priority to the child.
1563 p->prio = current->normal_prio;
1566 * Revert to default priority/policy on fork if requested.
1568 if (unlikely(p->sched_reset_on_fork)) {
1569 if (task_has_rt_policy(p)) {
1570 p->policy = SCHED_NORMAL;
1571 p->static_prio = NICE_TO_PRIO(0);
1573 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1574 p->static_prio = NICE_TO_PRIO(0);
1576 p->prio = p->normal_prio = __normal_prio(p);
1580 * We don't need the reset flag anymore after the fork. It has
1581 * fulfilled its duty:
1583 p->sched_reset_on_fork = 0;
1586 if (!rt_prio(p->prio))
1587 p->sched_class = &fair_sched_class;
1589 if (p->sched_class->task_fork)
1590 p->sched_class->task_fork(p);
1593 * The child is not yet in the pid-hash so no cgroup attach races,
1594 * and the cgroup is pinned to this child due to cgroup_fork()
1595 * is ran before sched_fork().
1597 * Silence PROVE_RCU.
1599 raw_spin_lock_irqsave(&p->pi_lock, flags);
1600 set_task_cpu(p, cpu);
1601 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1603 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1604 if (likely(sched_info_on()))
1605 memset(&p->sched_info, 0, sizeof(p->sched_info));
1607 #if defined(CONFIG_SMP)
1610 #ifdef CONFIG_PREEMPT_COUNT
1611 /* Want to start with kernel preemption disabled. */
1612 task_thread_info(p)->preempt_count = 1;
1615 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1622 * wake_up_new_task - wake up a newly created task for the first time.
1624 * This function will do some initial scheduler statistics housekeeping
1625 * that must be done for every newly created context, then puts the task
1626 * on the runqueue and wakes it.
1628 void wake_up_new_task(struct task_struct *p)
1630 unsigned long flags;
1633 raw_spin_lock_irqsave(&p->pi_lock, flags);
1636 * Fork balancing, do it here and not earlier because:
1637 * - cpus_allowed can change in the fork path
1638 * - any previously selected cpu might disappear through hotplug
1640 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1643 rq = __task_rq_lock(p);
1644 activate_task(rq, p, 0);
1646 trace_sched_wakeup_new(p, true);
1647 check_preempt_curr(rq, p, WF_FORK);
1649 if (p->sched_class->task_woken)
1650 p->sched_class->task_woken(rq, p);
1652 task_rq_unlock(rq, p, &flags);
1655 #ifdef CONFIG_PREEMPT_NOTIFIERS
1658 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1659 * @notifier: notifier struct to register
1661 void preempt_notifier_register(struct preempt_notifier *notifier)
1663 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1665 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1668 * preempt_notifier_unregister - no longer interested in preemption notifications
1669 * @notifier: notifier struct to unregister
1671 * This is safe to call from within a preemption notifier.
1673 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1675 hlist_del(¬ifier->link);
1677 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1679 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1681 struct preempt_notifier *notifier;
1682 struct hlist_node *node;
1684 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1685 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1689 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1690 struct task_struct *next)
1692 struct preempt_notifier *notifier;
1693 struct hlist_node *node;
1695 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1696 notifier->ops->sched_out(notifier, next);
1699 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1701 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1706 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1707 struct task_struct *next)
1711 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1714 * prepare_task_switch - prepare to switch tasks
1715 * @rq: the runqueue preparing to switch
1716 * @prev: the current task that is being switched out
1717 * @next: the task we are going to switch to.
1719 * This is called with the rq lock held and interrupts off. It must
1720 * be paired with a subsequent finish_task_switch after the context
1723 * prepare_task_switch sets up locking and calls architecture specific
1727 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1728 struct task_struct *next)
1730 trace_sched_switch(prev, next);
1731 sched_info_switch(prev, next);
1732 perf_event_task_sched_out(prev, next);
1733 fire_sched_out_preempt_notifiers(prev, next);
1734 prepare_lock_switch(rq, next);
1735 prepare_arch_switch(next);
1739 * finish_task_switch - clean up after a task-switch
1740 * @rq: runqueue associated with task-switch
1741 * @prev: the thread we just switched away from.
1743 * finish_task_switch must be called after the context switch, paired
1744 * with a prepare_task_switch call before the context switch.
1745 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1746 * and do any other architecture-specific cleanup actions.
1748 * Note that we may have delayed dropping an mm in context_switch(). If
1749 * so, we finish that here outside of the runqueue lock. (Doing it
1750 * with the lock held can cause deadlocks; see schedule() for
1753 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1754 __releases(rq->lock)
1756 struct mm_struct *mm = rq->prev_mm;
1762 * A task struct has one reference for the use as "current".
1763 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1764 * schedule one last time. The schedule call will never return, and
1765 * the scheduled task must drop that reference.
1766 * The test for TASK_DEAD must occur while the runqueue locks are
1767 * still held, otherwise prev could be scheduled on another cpu, die
1768 * there before we look at prev->state, and then the reference would
1770 * Manfred Spraul <manfred@colorfullife.com>
1772 prev_state = prev->state;
1773 vtime_task_switch(prev);
1774 finish_arch_switch(prev);
1775 perf_event_task_sched_in(prev, current);
1776 finish_lock_switch(rq, prev);
1777 finish_arch_post_lock_switch();
1779 fire_sched_in_preempt_notifiers(current);
1782 if (unlikely(prev_state == TASK_DEAD)) {
1784 * Remove function-return probe instances associated with this
1785 * task and put them back on the free list.
1787 kprobe_flush_task(prev);
1788 put_task_struct(prev);
1794 /* assumes rq->lock is held */
1795 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1797 if (prev->sched_class->pre_schedule)
1798 prev->sched_class->pre_schedule(rq, prev);
1801 /* rq->lock is NOT held, but preemption is disabled */
1802 static inline void post_schedule(struct rq *rq)
1804 if (rq->post_schedule) {
1805 unsigned long flags;
1807 raw_spin_lock_irqsave(&rq->lock, flags);
1808 if (rq->curr->sched_class->post_schedule)
1809 rq->curr->sched_class->post_schedule(rq);
1810 raw_spin_unlock_irqrestore(&rq->lock, flags);
1812 rq->post_schedule = 0;
1818 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1822 static inline void post_schedule(struct rq *rq)
1829 * schedule_tail - first thing a freshly forked thread must call.
1830 * @prev: the thread we just switched away from.
1832 asmlinkage void schedule_tail(struct task_struct *prev)
1833 __releases(rq->lock)
1835 struct rq *rq = this_rq();
1837 finish_task_switch(rq, prev);
1840 * FIXME: do we need to worry about rq being invalidated by the
1845 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1846 /* In this case, finish_task_switch does not reenable preemption */
1849 if (current->set_child_tid)
1850 put_user(task_pid_vnr(current), current->set_child_tid);
1854 * context_switch - switch to the new MM and the new
1855 * thread's register state.
1858 context_switch(struct rq *rq, struct task_struct *prev,
1859 struct task_struct *next)
1861 struct mm_struct *mm, *oldmm;
1863 prepare_task_switch(rq, prev, next);
1866 oldmm = prev->active_mm;
1868 * For paravirt, this is coupled with an exit in switch_to to
1869 * combine the page table reload and the switch backend into
1872 arch_start_context_switch(prev);
1875 next->active_mm = oldmm;
1876 atomic_inc(&oldmm->mm_count);
1877 enter_lazy_tlb(oldmm, next);
1879 switch_mm(oldmm, mm, next);
1882 prev->active_mm = NULL;
1883 rq->prev_mm = oldmm;
1886 * Since the runqueue lock will be released by the next
1887 * task (which is an invalid locking op but in the case
1888 * of the scheduler it's an obvious special-case), so we
1889 * do an early lockdep release here:
1891 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1892 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1895 /* Here we just switch the register state and the stack. */
1896 rcu_switch(prev, next);
1897 switch_to(prev, next, prev);
1901 * this_rq must be evaluated again because prev may have moved
1902 * CPUs since it called schedule(), thus the 'rq' on its stack
1903 * frame will be invalid.
1905 finish_task_switch(this_rq(), prev);
1909 * nr_running, nr_uninterruptible and nr_context_switches:
1911 * externally visible scheduler statistics: current number of runnable
1912 * threads, current number of uninterruptible-sleeping threads, total
1913 * number of context switches performed since bootup.
1915 unsigned long nr_running(void)
1917 unsigned long i, sum = 0;
1919 for_each_online_cpu(i)
1920 sum += cpu_rq(i)->nr_running;
1925 unsigned long nr_uninterruptible(void)
1927 unsigned long i, sum = 0;
1929 for_each_possible_cpu(i)
1930 sum += cpu_rq(i)->nr_uninterruptible;
1933 * Since we read the counters lockless, it might be slightly
1934 * inaccurate. Do not allow it to go below zero though:
1936 if (unlikely((long)sum < 0))
1942 unsigned long long nr_context_switches(void)
1945 unsigned long long sum = 0;
1947 for_each_possible_cpu(i)
1948 sum += cpu_rq(i)->nr_switches;
1953 unsigned long nr_iowait(void)
1955 unsigned long i, sum = 0;
1957 for_each_possible_cpu(i)
1958 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1963 unsigned long nr_iowait_cpu(int cpu)
1965 struct rq *this = cpu_rq(cpu);
1966 return atomic_read(&this->nr_iowait);
1969 unsigned long this_cpu_load(void)
1971 struct rq *this = this_rq();
1972 return this->cpu_load[0];
1977 * Global load-average calculations
1979 * We take a distributed and async approach to calculating the global load-avg
1980 * in order to minimize overhead.
1982 * The global load average is an exponentially decaying average of nr_running +
1983 * nr_uninterruptible.
1985 * Once every LOAD_FREQ:
1988 * for_each_possible_cpu(cpu)
1989 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
1991 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
1993 * Due to a number of reasons the above turns in the mess below:
1995 * - for_each_possible_cpu() is prohibitively expensive on machines with
1996 * serious number of cpus, therefore we need to take a distributed approach
1997 * to calculating nr_active.
1999 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2000 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2002 * So assuming nr_active := 0 when we start out -- true per definition, we
2003 * can simply take per-cpu deltas and fold those into a global accumulate
2004 * to obtain the same result. See calc_load_fold_active().
2006 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2007 * across the machine, we assume 10 ticks is sufficient time for every
2008 * cpu to have completed this task.
2010 * This places an upper-bound on the IRQ-off latency of the machine. Then
2011 * again, being late doesn't loose the delta, just wrecks the sample.
2013 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2014 * this would add another cross-cpu cacheline miss and atomic operation
2015 * to the wakeup path. Instead we increment on whatever cpu the task ran
2016 * when it went into uninterruptible state and decrement on whatever cpu
2017 * did the wakeup. This means that only the sum of nr_uninterruptible over
2018 * all cpus yields the correct result.
2020 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2023 /* Variables and functions for calc_load */
2024 static atomic_long_t calc_load_tasks;
2025 static unsigned long calc_load_update;
2026 unsigned long avenrun[3];
2027 EXPORT_SYMBOL(avenrun); /* should be removed */
2030 * get_avenrun - get the load average array
2031 * @loads: pointer to dest load array
2032 * @offset: offset to add
2033 * @shift: shift count to shift the result left
2035 * These values are estimates at best, so no need for locking.
2037 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2039 loads[0] = (avenrun[0] + offset) << shift;
2040 loads[1] = (avenrun[1] + offset) << shift;
2041 loads[2] = (avenrun[2] + offset) << shift;
2044 static long calc_load_fold_active(struct rq *this_rq)
2046 long nr_active, delta = 0;
2048 nr_active = this_rq->nr_running;
2049 nr_active += (long) this_rq->nr_uninterruptible;
2051 if (nr_active != this_rq->calc_load_active) {
2052 delta = nr_active - this_rq->calc_load_active;
2053 this_rq->calc_load_active = nr_active;
2060 * a1 = a0 * e + a * (1 - e)
2062 static unsigned long
2063 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2066 load += active * (FIXED_1 - exp);
2067 load += 1UL << (FSHIFT - 1);
2068 return load >> FSHIFT;
2073 * Handle NO_HZ for the global load-average.
2075 * Since the above described distributed algorithm to compute the global
2076 * load-average relies on per-cpu sampling from the tick, it is affected by
2079 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2080 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2081 * when we read the global state.
2083 * Obviously reality has to ruin such a delightfully simple scheme:
2085 * - When we go NO_HZ idle during the window, we can negate our sample
2086 * contribution, causing under-accounting.
2088 * We avoid this by keeping two idle-delta counters and flipping them
2089 * when the window starts, thus separating old and new NO_HZ load.
2091 * The only trick is the slight shift in index flip for read vs write.
2095 * |-|-----------|-|-----------|-|-----------|-|
2096 * r:0 0 1 1 0 0 1 1 0
2097 * w:0 1 1 0 0 1 1 0 0
2099 * This ensures we'll fold the old idle contribution in this window while
2100 * accumlating the new one.
2102 * - When we wake up from NO_HZ idle during the window, we push up our
2103 * contribution, since we effectively move our sample point to a known
2106 * This is solved by pushing the window forward, and thus skipping the
2107 * sample, for this cpu (effectively using the idle-delta for this cpu which
2108 * was in effect at the time the window opened). This also solves the issue
2109 * of having to deal with a cpu having been in NOHZ idle for multiple
2110 * LOAD_FREQ intervals.
2112 * When making the ILB scale, we should try to pull this in as well.
2114 static atomic_long_t calc_load_idle[2];
2115 static int calc_load_idx;
2117 static inline int calc_load_write_idx(void)
2119 int idx = calc_load_idx;
2122 * See calc_global_nohz(), if we observe the new index, we also
2123 * need to observe the new update time.
2128 * If the folding window started, make sure we start writing in the
2131 if (!time_before(jiffies, calc_load_update))
2137 static inline int calc_load_read_idx(void)
2139 return calc_load_idx & 1;
2142 void calc_load_enter_idle(void)
2144 struct rq *this_rq = this_rq();
2148 * We're going into NOHZ mode, if there's any pending delta, fold it
2149 * into the pending idle delta.
2151 delta = calc_load_fold_active(this_rq);
2153 int idx = calc_load_write_idx();
2154 atomic_long_add(delta, &calc_load_idle[idx]);
2158 void calc_load_exit_idle(void)
2160 struct rq *this_rq = this_rq();
2163 * If we're still before the sample window, we're done.
2165 if (time_before(jiffies, this_rq->calc_load_update))
2169 * We woke inside or after the sample window, this means we're already
2170 * accounted through the nohz accounting, so skip the entire deal and
2171 * sync up for the next window.
2173 this_rq->calc_load_update = calc_load_update;
2174 if (time_before(jiffies, this_rq->calc_load_update + 10))
2175 this_rq->calc_load_update += LOAD_FREQ;
2178 static long calc_load_fold_idle(void)
2180 int idx = calc_load_read_idx();
2183 if (atomic_long_read(&calc_load_idle[idx]))
2184 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2190 * fixed_power_int - compute: x^n, in O(log n) time
2192 * @x: base of the power
2193 * @frac_bits: fractional bits of @x
2194 * @n: power to raise @x to.
2196 * By exploiting the relation between the definition of the natural power
2197 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2198 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2199 * (where: n_i \elem {0, 1}, the binary vector representing n),
2200 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2201 * of course trivially computable in O(log_2 n), the length of our binary
2204 static unsigned long
2205 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2207 unsigned long result = 1UL << frac_bits;
2212 result += 1UL << (frac_bits - 1);
2213 result >>= frac_bits;
2219 x += 1UL << (frac_bits - 1);
2227 * a1 = a0 * e + a * (1 - e)
2229 * a2 = a1 * e + a * (1 - e)
2230 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2231 * = a0 * e^2 + a * (1 - e) * (1 + e)
2233 * a3 = a2 * e + a * (1 - e)
2234 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2235 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2239 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2240 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2241 * = a0 * e^n + a * (1 - e^n)
2243 * [1] application of the geometric series:
2246 * S_n := \Sum x^i = -------------
2249 static unsigned long
2250 calc_load_n(unsigned long load, unsigned long exp,
2251 unsigned long active, unsigned int n)
2254 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2258 * NO_HZ can leave us missing all per-cpu ticks calling
2259 * calc_load_account_active(), but since an idle CPU folds its delta into
2260 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2261 * in the pending idle delta if our idle period crossed a load cycle boundary.
2263 * Once we've updated the global active value, we need to apply the exponential
2264 * weights adjusted to the number of cycles missed.
2266 static void calc_global_nohz(void)
2268 long delta, active, n;
2270 if (!time_before(jiffies, calc_load_update + 10)) {
2272 * Catch-up, fold however many we are behind still
2274 delta = jiffies - calc_load_update - 10;
2275 n = 1 + (delta / LOAD_FREQ);
2277 active = atomic_long_read(&calc_load_tasks);
2278 active = active > 0 ? active * FIXED_1 : 0;
2280 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2281 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2282 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2284 calc_load_update += n * LOAD_FREQ;
2288 * Flip the idle index...
2290 * Make sure we first write the new time then flip the index, so that
2291 * calc_load_write_idx() will see the new time when it reads the new
2292 * index, this avoids a double flip messing things up.
2297 #else /* !CONFIG_NO_HZ */
2299 static inline long calc_load_fold_idle(void) { return 0; }
2300 static inline void calc_global_nohz(void) { }
2302 #endif /* CONFIG_NO_HZ */
2305 * calc_load - update the avenrun load estimates 10 ticks after the
2306 * CPUs have updated calc_load_tasks.
2308 void calc_global_load(unsigned long ticks)
2312 if (time_before(jiffies, calc_load_update + 10))
2316 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2318 delta = calc_load_fold_idle();
2320 atomic_long_add(delta, &calc_load_tasks);
2322 active = atomic_long_read(&calc_load_tasks);
2323 active = active > 0 ? active * FIXED_1 : 0;
2325 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2326 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2327 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2329 calc_load_update += LOAD_FREQ;
2332 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2338 * Called from update_cpu_load() to periodically update this CPU's
2341 static void calc_load_account_active(struct rq *this_rq)
2345 if (time_before(jiffies, this_rq->calc_load_update))
2348 delta = calc_load_fold_active(this_rq);
2350 atomic_long_add(delta, &calc_load_tasks);
2352 this_rq->calc_load_update += LOAD_FREQ;
2356 * End of global load-average stuff
2360 * The exact cpuload at various idx values, calculated at every tick would be
2361 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2363 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2364 * on nth tick when cpu may be busy, then we have:
2365 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2366 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2368 * decay_load_missed() below does efficient calculation of
2369 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2370 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2372 * The calculation is approximated on a 128 point scale.
2373 * degrade_zero_ticks is the number of ticks after which load at any
2374 * particular idx is approximated to be zero.
2375 * degrade_factor is a precomputed table, a row for each load idx.
2376 * Each column corresponds to degradation factor for a power of two ticks,
2377 * based on 128 point scale.
2379 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2380 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2382 * With this power of 2 load factors, we can degrade the load n times
2383 * by looking at 1 bits in n and doing as many mult/shift instead of
2384 * n mult/shifts needed by the exact degradation.
2386 #define DEGRADE_SHIFT 7
2387 static const unsigned char
2388 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2389 static const unsigned char
2390 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2391 {0, 0, 0, 0, 0, 0, 0, 0},
2392 {64, 32, 8, 0, 0, 0, 0, 0},
2393 {96, 72, 40, 12, 1, 0, 0},
2394 {112, 98, 75, 43, 15, 1, 0},
2395 {120, 112, 98, 76, 45, 16, 2} };
2398 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2399 * would be when CPU is idle and so we just decay the old load without
2400 * adding any new load.
2402 static unsigned long
2403 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2407 if (!missed_updates)
2410 if (missed_updates >= degrade_zero_ticks[idx])
2414 return load >> missed_updates;
2416 while (missed_updates) {
2417 if (missed_updates % 2)
2418 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2420 missed_updates >>= 1;
2427 * Update rq->cpu_load[] statistics. This function is usually called every
2428 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2429 * every tick. We fix it up based on jiffies.
2431 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2432 unsigned long pending_updates)
2436 this_rq->nr_load_updates++;
2438 /* Update our load: */
2439 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2440 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2441 unsigned long old_load, new_load;
2443 /* scale is effectively 1 << i now, and >> i divides by scale */
2445 old_load = this_rq->cpu_load[i];
2446 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2447 new_load = this_load;
2449 * Round up the averaging division if load is increasing. This
2450 * prevents us from getting stuck on 9 if the load is 10, for
2453 if (new_load > old_load)
2454 new_load += scale - 1;
2456 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2459 sched_avg_update(this_rq);
2464 * There is no sane way to deal with nohz on smp when using jiffies because the
2465 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2466 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2468 * Therefore we cannot use the delta approach from the regular tick since that
2469 * would seriously skew the load calculation. However we'll make do for those
2470 * updates happening while idle (nohz_idle_balance) or coming out of idle
2471 * (tick_nohz_idle_exit).
2473 * This means we might still be one tick off for nohz periods.
2477 * Called from nohz_idle_balance() to update the load ratings before doing the
2480 void update_idle_cpu_load(struct rq *this_rq)
2482 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2483 unsigned long load = this_rq->load.weight;
2484 unsigned long pending_updates;
2487 * bail if there's load or we're actually up-to-date.
2489 if (load || curr_jiffies == this_rq->last_load_update_tick)
2492 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2493 this_rq->last_load_update_tick = curr_jiffies;
2495 __update_cpu_load(this_rq, load, pending_updates);
2499 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2501 void update_cpu_load_nohz(void)
2503 struct rq *this_rq = this_rq();
2504 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2505 unsigned long pending_updates;
2507 if (curr_jiffies == this_rq->last_load_update_tick)
2510 raw_spin_lock(&this_rq->lock);
2511 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2512 if (pending_updates) {
2513 this_rq->last_load_update_tick = curr_jiffies;
2515 * We were idle, this means load 0, the current load might be
2516 * !0 due to remote wakeups and the sort.
2518 __update_cpu_load(this_rq, 0, pending_updates);
2520 raw_spin_unlock(&this_rq->lock);
2522 #endif /* CONFIG_NO_HZ */
2525 * Called from scheduler_tick()
2527 static void update_cpu_load_active(struct rq *this_rq)
2530 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2532 this_rq->last_load_update_tick = jiffies;
2533 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2535 calc_load_account_active(this_rq);
2541 * sched_exec - execve() is a valuable balancing opportunity, because at
2542 * this point the task has the smallest effective memory and cache footprint.
2544 void sched_exec(void)
2546 struct task_struct *p = current;
2547 unsigned long flags;
2550 raw_spin_lock_irqsave(&p->pi_lock, flags);
2551 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2552 if (dest_cpu == smp_processor_id())
2555 if (likely(cpu_active(dest_cpu))) {
2556 struct migration_arg arg = { p, dest_cpu };
2558 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2559 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2563 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2568 DEFINE_PER_CPU(struct kernel_stat, kstat);
2569 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2571 EXPORT_PER_CPU_SYMBOL(kstat);
2572 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2575 * Return any ns on the sched_clock that have not yet been accounted in
2576 * @p in case that task is currently running.
2578 * Called with task_rq_lock() held on @rq.
2580 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2584 if (task_current(rq, p)) {
2585 update_rq_clock(rq);
2586 ns = rq->clock_task - p->se.exec_start;
2594 unsigned long long task_delta_exec(struct task_struct *p)
2596 unsigned long flags;
2600 rq = task_rq_lock(p, &flags);
2601 ns = do_task_delta_exec(p, rq);
2602 task_rq_unlock(rq, p, &flags);
2608 * Return accounted runtime for the task.
2609 * In case the task is currently running, return the runtime plus current's
2610 * pending runtime that have not been accounted yet.
2612 unsigned long long task_sched_runtime(struct task_struct *p)
2614 unsigned long flags;
2618 rq = task_rq_lock(p, &flags);
2619 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2620 task_rq_unlock(rq, p, &flags);
2626 * This function gets called by the timer code, with HZ frequency.
2627 * We call it with interrupts disabled.
2629 void scheduler_tick(void)
2631 int cpu = smp_processor_id();
2632 struct rq *rq = cpu_rq(cpu);
2633 struct task_struct *curr = rq->curr;
2637 raw_spin_lock(&rq->lock);
2638 update_rq_clock(rq);
2639 update_cpu_load_active(rq);
2640 curr->sched_class->task_tick(rq, curr, 0);
2641 raw_spin_unlock(&rq->lock);
2643 perf_event_task_tick();
2646 rq->idle_balance = idle_cpu(cpu);
2647 trigger_load_balance(rq, cpu);
2651 notrace unsigned long get_parent_ip(unsigned long addr)
2653 if (in_lock_functions(addr)) {
2654 addr = CALLER_ADDR2;
2655 if (in_lock_functions(addr))
2656 addr = CALLER_ADDR3;
2661 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2662 defined(CONFIG_PREEMPT_TRACER))
2664 void __kprobes add_preempt_count(int val)
2666 #ifdef CONFIG_DEBUG_PREEMPT
2670 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2673 preempt_count() += val;
2674 #ifdef CONFIG_DEBUG_PREEMPT
2676 * Spinlock count overflowing soon?
2678 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2681 if (preempt_count() == val)
2682 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2684 EXPORT_SYMBOL(add_preempt_count);
2686 void __kprobes sub_preempt_count(int val)
2688 #ifdef CONFIG_DEBUG_PREEMPT
2692 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2695 * Is the spinlock portion underflowing?
2697 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2698 !(preempt_count() & PREEMPT_MASK)))
2702 if (preempt_count() == val)
2703 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2704 preempt_count() -= val;
2706 EXPORT_SYMBOL(sub_preempt_count);
2711 * Print scheduling while atomic bug:
2713 static noinline void __schedule_bug(struct task_struct *prev)
2715 if (oops_in_progress)
2718 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2719 prev->comm, prev->pid, preempt_count());
2721 debug_show_held_locks(prev);
2723 if (irqs_disabled())
2724 print_irqtrace_events(prev);
2726 add_taint(TAINT_WARN);
2730 * Various schedule()-time debugging checks and statistics:
2732 static inline void schedule_debug(struct task_struct *prev)
2735 * Test if we are atomic. Since do_exit() needs to call into
2736 * schedule() atomically, we ignore that path for now.
2737 * Otherwise, whine if we are scheduling when we should not be.
2739 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2740 __schedule_bug(prev);
2743 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2745 schedstat_inc(this_rq(), sched_count);
2748 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2750 if (prev->on_rq || rq->skip_clock_update < 0)
2751 update_rq_clock(rq);
2752 prev->sched_class->put_prev_task(rq, prev);
2756 * Pick up the highest-prio task:
2758 static inline struct task_struct *
2759 pick_next_task(struct rq *rq)
2761 const struct sched_class *class;
2762 struct task_struct *p;
2765 * Optimization: we know that if all tasks are in
2766 * the fair class we can call that function directly:
2768 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2769 p = fair_sched_class.pick_next_task(rq);
2774 for_each_class(class) {
2775 p = class->pick_next_task(rq);
2780 BUG(); /* the idle class will always have a runnable task */
2784 * __schedule() is the main scheduler function.
2786 * The main means of driving the scheduler and thus entering this function are:
2788 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2790 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2791 * paths. For example, see arch/x86/entry_64.S.
2793 * To drive preemption between tasks, the scheduler sets the flag in timer
2794 * interrupt handler scheduler_tick().
2796 * 3. Wakeups don't really cause entry into schedule(). They add a
2797 * task to the run-queue and that's it.
2799 * Now, if the new task added to the run-queue preempts the current
2800 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2801 * called on the nearest possible occasion:
2803 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2805 * - in syscall or exception context, at the next outmost
2806 * preempt_enable(). (this might be as soon as the wake_up()'s
2809 * - in IRQ context, return from interrupt-handler to
2810 * preemptible context
2812 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2815 * - cond_resched() call
2816 * - explicit schedule() call
2817 * - return from syscall or exception to user-space
2818 * - return from interrupt-handler to user-space
2820 static void __sched __schedule(void)
2822 struct task_struct *prev, *next;
2823 unsigned long *switch_count;
2829 cpu = smp_processor_id();
2831 rcu_note_context_switch(cpu);
2834 schedule_debug(prev);
2836 if (sched_feat(HRTICK))
2839 raw_spin_lock_irq(&rq->lock);
2841 switch_count = &prev->nivcsw;
2842 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2843 if (unlikely(signal_pending_state(prev->state, prev))) {
2844 prev->state = TASK_RUNNING;
2846 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2850 * If a worker went to sleep, notify and ask workqueue
2851 * whether it wants to wake up a task to maintain
2854 if (prev->flags & PF_WQ_WORKER) {
2855 struct task_struct *to_wakeup;
2857 to_wakeup = wq_worker_sleeping(prev, cpu);
2859 try_to_wake_up_local(to_wakeup);
2862 switch_count = &prev->nvcsw;
2865 pre_schedule(rq, prev);
2867 if (unlikely(!rq->nr_running))
2868 idle_balance(cpu, rq);
2870 put_prev_task(rq, prev);
2871 next = pick_next_task(rq);
2872 clear_tsk_need_resched(prev);
2873 rq->skip_clock_update = 0;
2875 if (likely(prev != next)) {
2880 context_switch(rq, prev, next); /* unlocks the rq */
2882 * The context switch have flipped the stack from under us
2883 * and restored the local variables which were saved when
2884 * this task called schedule() in the past. prev == current
2885 * is still correct, but it can be moved to another cpu/rq.
2887 cpu = smp_processor_id();
2890 raw_spin_unlock_irq(&rq->lock);
2894 sched_preempt_enable_no_resched();
2899 static inline void sched_submit_work(struct task_struct *tsk)
2901 if (!tsk->state || tsk_is_pi_blocked(tsk))
2904 * If we are going to sleep and we have plugged IO queued,
2905 * make sure to submit it to avoid deadlocks.
2907 if (blk_needs_flush_plug(tsk))
2908 blk_schedule_flush_plug(tsk);
2911 asmlinkage void __sched schedule(void)
2913 struct task_struct *tsk = current;
2915 sched_submit_work(tsk);
2918 EXPORT_SYMBOL(schedule);
2920 #ifdef CONFIG_RCU_USER_QS
2921 asmlinkage void __sched schedule_user(void)
2924 * If we come here after a random call to set_need_resched(),
2925 * or we have been woken up remotely but the IPI has not yet arrived,
2926 * we haven't yet exited the RCU idle mode. Do it here manually until
2927 * we find a better solution.
2936 * schedule_preempt_disabled - called with preemption disabled
2938 * Returns with preemption disabled. Note: preempt_count must be 1
2940 void __sched schedule_preempt_disabled(void)
2942 sched_preempt_enable_no_resched();
2947 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2949 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2951 if (lock->owner != owner)
2955 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2956 * lock->owner still matches owner, if that fails, owner might
2957 * point to free()d memory, if it still matches, the rcu_read_lock()
2958 * ensures the memory stays valid.
2962 return owner->on_cpu;
2966 * Look out! "owner" is an entirely speculative pointer
2967 * access and not reliable.
2969 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
2971 if (!sched_feat(OWNER_SPIN))
2975 while (owner_running(lock, owner)) {
2979 arch_mutex_cpu_relax();
2984 * We break out the loop above on need_resched() and when the
2985 * owner changed, which is a sign for heavy contention. Return
2986 * success only when lock->owner is NULL.
2988 return lock->owner == NULL;
2992 #ifdef CONFIG_PREEMPT
2994 * this is the entry point to schedule() from in-kernel preemption
2995 * off of preempt_enable. Kernel preemptions off return from interrupt
2996 * occur there and call schedule directly.
2998 asmlinkage void __sched notrace preempt_schedule(void)
3000 struct thread_info *ti = current_thread_info();
3003 * If there is a non-zero preempt_count or interrupts are disabled,
3004 * we do not want to preempt the current task. Just return..
3006 if (likely(ti->preempt_count || irqs_disabled()))
3010 add_preempt_count_notrace(PREEMPT_ACTIVE);
3012 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3015 * Check again in case we missed a preemption opportunity
3016 * between schedule and now.
3019 } while (need_resched());
3021 EXPORT_SYMBOL(preempt_schedule);
3024 * this is the entry point to schedule() from kernel preemption
3025 * off of irq context.
3026 * Note, that this is called and return with irqs disabled. This will
3027 * protect us against recursive calling from irq.
3029 asmlinkage void __sched preempt_schedule_irq(void)
3031 struct thread_info *ti = current_thread_info();
3033 /* Catch callers which need to be fixed */
3034 BUG_ON(ti->preempt_count || !irqs_disabled());
3038 add_preempt_count(PREEMPT_ACTIVE);
3041 local_irq_disable();
3042 sub_preempt_count(PREEMPT_ACTIVE);
3045 * Check again in case we missed a preemption opportunity
3046 * between schedule and now.
3049 } while (need_resched());
3052 #endif /* CONFIG_PREEMPT */
3054 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3057 return try_to_wake_up(curr->private, mode, wake_flags);
3059 EXPORT_SYMBOL(default_wake_function);
3062 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3063 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3064 * number) then we wake all the non-exclusive tasks and one exclusive task.
3066 * There are circumstances in which we can try to wake a task which has already
3067 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3068 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3070 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3071 int nr_exclusive, int wake_flags, void *key)
3073 wait_queue_t *curr, *next;
3075 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3076 unsigned flags = curr->flags;
3078 if (curr->func(curr, mode, wake_flags, key) &&
3079 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3085 * __wake_up - wake up threads blocked on a waitqueue.
3087 * @mode: which threads
3088 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3089 * @key: is directly passed to the wakeup function
3091 * It may be assumed that this function implies a write memory barrier before
3092 * changing the task state if and only if any tasks are woken up.
3094 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3095 int nr_exclusive, void *key)
3097 unsigned long flags;
3099 spin_lock_irqsave(&q->lock, flags);
3100 __wake_up_common(q, mode, nr_exclusive, 0, key);
3101 spin_unlock_irqrestore(&q->lock, flags);
3103 EXPORT_SYMBOL(__wake_up);
3106 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3108 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3110 __wake_up_common(q, mode, nr, 0, NULL);
3112 EXPORT_SYMBOL_GPL(__wake_up_locked);
3114 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3116 __wake_up_common(q, mode, 1, 0, key);
3118 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3121 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3123 * @mode: which threads
3124 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3125 * @key: opaque value to be passed to wakeup targets
3127 * The sync wakeup differs that the waker knows that it will schedule
3128 * away soon, so while the target thread will be woken up, it will not
3129 * be migrated to another CPU - ie. the two threads are 'synchronized'
3130 * with each other. This can prevent needless bouncing between CPUs.
3132 * On UP it can prevent extra preemption.
3134 * It may be assumed that this function implies a write memory barrier before
3135 * changing the task state if and only if any tasks are woken up.
3137 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3138 int nr_exclusive, void *key)
3140 unsigned long flags;
3141 int wake_flags = WF_SYNC;
3146 if (unlikely(!nr_exclusive))
3149 spin_lock_irqsave(&q->lock, flags);
3150 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3151 spin_unlock_irqrestore(&q->lock, flags);
3153 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3156 * __wake_up_sync - see __wake_up_sync_key()
3158 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3160 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3162 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3165 * complete: - signals a single thread waiting on this completion
3166 * @x: holds the state of this particular completion
3168 * This will wake up a single thread waiting on this completion. Threads will be
3169 * awakened in the same order in which they were queued.
3171 * See also complete_all(), wait_for_completion() and related routines.
3173 * It may be assumed that this function implies a write memory barrier before
3174 * changing the task state if and only if any tasks are woken up.
3176 void complete(struct completion *x)
3178 unsigned long flags;
3180 spin_lock_irqsave(&x->wait.lock, flags);
3182 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3183 spin_unlock_irqrestore(&x->wait.lock, flags);
3185 EXPORT_SYMBOL(complete);
3188 * complete_all: - signals all threads waiting on this completion
3189 * @x: holds the state of this particular completion
3191 * This will wake up all threads waiting on this particular completion event.
3193 * It may be assumed that this function implies a write memory barrier before
3194 * changing the task state if and only if any tasks are woken up.
3196 void complete_all(struct completion *x)
3198 unsigned long flags;
3200 spin_lock_irqsave(&x->wait.lock, flags);
3201 x->done += UINT_MAX/2;
3202 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3203 spin_unlock_irqrestore(&x->wait.lock, flags);
3205 EXPORT_SYMBOL(complete_all);
3207 static inline long __sched
3208 do_wait_for_common(struct completion *x, long timeout, int state)
3211 DECLARE_WAITQUEUE(wait, current);
3213 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3215 if (signal_pending_state(state, current)) {
3216 timeout = -ERESTARTSYS;
3219 __set_current_state(state);
3220 spin_unlock_irq(&x->wait.lock);
3221 timeout = schedule_timeout(timeout);
3222 spin_lock_irq(&x->wait.lock);
3223 } while (!x->done && timeout);
3224 __remove_wait_queue(&x->wait, &wait);
3229 return timeout ?: 1;
3233 wait_for_common(struct completion *x, long timeout, int state)
3237 spin_lock_irq(&x->wait.lock);
3238 timeout = do_wait_for_common(x, timeout, state);
3239 spin_unlock_irq(&x->wait.lock);
3244 * wait_for_completion: - waits for completion of a task
3245 * @x: holds the state of this particular completion
3247 * This waits to be signaled for completion of a specific task. It is NOT
3248 * interruptible and there is no timeout.
3250 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3251 * and interrupt capability. Also see complete().
3253 void __sched wait_for_completion(struct completion *x)
3255 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3257 EXPORT_SYMBOL(wait_for_completion);
3260 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3261 * @x: holds the state of this particular completion
3262 * @timeout: timeout value in jiffies
3264 * This waits for either a completion of a specific task to be signaled or for a
3265 * specified timeout to expire. The timeout is in jiffies. It is not
3268 * The return value is 0 if timed out, and positive (at least 1, or number of
3269 * jiffies left till timeout) if completed.
3271 unsigned long __sched
3272 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3274 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3276 EXPORT_SYMBOL(wait_for_completion_timeout);
3279 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3280 * @x: holds the state of this particular completion
3282 * This waits for completion of a specific task to be signaled. It is
3285 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3287 int __sched wait_for_completion_interruptible(struct completion *x)
3289 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3290 if (t == -ERESTARTSYS)
3294 EXPORT_SYMBOL(wait_for_completion_interruptible);
3297 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3298 * @x: holds the state of this particular completion
3299 * @timeout: timeout value in jiffies
3301 * This waits for either a completion of a specific task to be signaled or for a
3302 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3304 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3305 * positive (at least 1, or number of jiffies left till timeout) if completed.
3308 wait_for_completion_interruptible_timeout(struct completion *x,
3309 unsigned long timeout)
3311 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3313 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3316 * wait_for_completion_killable: - waits for completion of a task (killable)
3317 * @x: holds the state of this particular completion
3319 * This waits to be signaled for completion of a specific task. It can be
3320 * interrupted by a kill signal.
3322 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3324 int __sched wait_for_completion_killable(struct completion *x)
3326 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3327 if (t == -ERESTARTSYS)
3331 EXPORT_SYMBOL(wait_for_completion_killable);
3334 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3335 * @x: holds the state of this particular completion
3336 * @timeout: timeout value in jiffies
3338 * This waits for either a completion of a specific task to be
3339 * signaled or for a specified timeout to expire. It can be
3340 * interrupted by a kill signal. The timeout is in jiffies.
3342 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3343 * positive (at least 1, or number of jiffies left till timeout) if completed.
3346 wait_for_completion_killable_timeout(struct completion *x,
3347 unsigned long timeout)
3349 return wait_for_common(x, timeout, TASK_KILLABLE);
3351 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3354 * try_wait_for_completion - try to decrement a completion without blocking
3355 * @x: completion structure
3357 * Returns: 0 if a decrement cannot be done without blocking
3358 * 1 if a decrement succeeded.
3360 * If a completion is being used as a counting completion,
3361 * attempt to decrement the counter without blocking. This
3362 * enables us to avoid waiting if the resource the completion
3363 * is protecting is not available.
3365 bool try_wait_for_completion(struct completion *x)
3367 unsigned long flags;
3370 spin_lock_irqsave(&x->wait.lock, flags);
3375 spin_unlock_irqrestore(&x->wait.lock, flags);
3378 EXPORT_SYMBOL(try_wait_for_completion);
3381 * completion_done - Test to see if a completion has any waiters
3382 * @x: completion structure
3384 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3385 * 1 if there are no waiters.
3388 bool completion_done(struct completion *x)
3390 unsigned long flags;
3393 spin_lock_irqsave(&x->wait.lock, flags);
3396 spin_unlock_irqrestore(&x->wait.lock, flags);
3399 EXPORT_SYMBOL(completion_done);
3402 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3404 unsigned long flags;
3407 init_waitqueue_entry(&wait, current);
3409 __set_current_state(state);
3411 spin_lock_irqsave(&q->lock, flags);
3412 __add_wait_queue(q, &wait);
3413 spin_unlock(&q->lock);
3414 timeout = schedule_timeout(timeout);
3415 spin_lock_irq(&q->lock);
3416 __remove_wait_queue(q, &wait);
3417 spin_unlock_irqrestore(&q->lock, flags);
3422 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3424 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3426 EXPORT_SYMBOL(interruptible_sleep_on);
3429 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3431 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3433 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3435 void __sched sleep_on(wait_queue_head_t *q)
3437 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3439 EXPORT_SYMBOL(sleep_on);
3441 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3443 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3445 EXPORT_SYMBOL(sleep_on_timeout);
3447 #ifdef CONFIG_RT_MUTEXES
3450 * rt_mutex_setprio - set the current priority of a task
3452 * @prio: prio value (kernel-internal form)
3454 * This function changes the 'effective' priority of a task. It does
3455 * not touch ->normal_prio like __setscheduler().
3457 * Used by the rt_mutex code to implement priority inheritance logic.
3459 void rt_mutex_setprio(struct task_struct *p, int prio)
3461 int oldprio, on_rq, running;
3463 const struct sched_class *prev_class;
3465 BUG_ON(prio < 0 || prio > MAX_PRIO);
3467 rq = __task_rq_lock(p);
3470 * Idle task boosting is a nono in general. There is one
3471 * exception, when PREEMPT_RT and NOHZ is active:
3473 * The idle task calls get_next_timer_interrupt() and holds
3474 * the timer wheel base->lock on the CPU and another CPU wants
3475 * to access the timer (probably to cancel it). We can safely
3476 * ignore the boosting request, as the idle CPU runs this code
3477 * with interrupts disabled and will complete the lock
3478 * protected section without being interrupted. So there is no
3479 * real need to boost.
3481 if (unlikely(p == rq->idle)) {
3482 WARN_ON(p != rq->curr);
3483 WARN_ON(p->pi_blocked_on);
3487 trace_sched_pi_setprio(p, prio);
3489 prev_class = p->sched_class;
3491 running = task_current(rq, p);
3493 dequeue_task(rq, p, 0);
3495 p->sched_class->put_prev_task(rq, p);
3498 p->sched_class = &rt_sched_class;
3500 p->sched_class = &fair_sched_class;
3505 p->sched_class->set_curr_task(rq);
3507 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3509 check_class_changed(rq, p, prev_class, oldprio);
3511 __task_rq_unlock(rq);
3514 void set_user_nice(struct task_struct *p, long nice)
3516 int old_prio, delta, on_rq;
3517 unsigned long flags;
3520 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3523 * We have to be careful, if called from sys_setpriority(),
3524 * the task might be in the middle of scheduling on another CPU.
3526 rq = task_rq_lock(p, &flags);
3528 * The RT priorities are set via sched_setscheduler(), but we still
3529 * allow the 'normal' nice value to be set - but as expected
3530 * it wont have any effect on scheduling until the task is
3531 * SCHED_FIFO/SCHED_RR:
3533 if (task_has_rt_policy(p)) {
3534 p->static_prio = NICE_TO_PRIO(nice);
3539 dequeue_task(rq, p, 0);
3541 p->static_prio = NICE_TO_PRIO(nice);
3544 p->prio = effective_prio(p);
3545 delta = p->prio - old_prio;
3548 enqueue_task(rq, p, 0);
3550 * If the task increased its priority or is running and
3551 * lowered its priority, then reschedule its CPU:
3553 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3554 resched_task(rq->curr);
3557 task_rq_unlock(rq, p, &flags);
3559 EXPORT_SYMBOL(set_user_nice);
3562 * can_nice - check if a task can reduce its nice value
3566 int can_nice(const struct task_struct *p, const int nice)
3568 /* convert nice value [19,-20] to rlimit style value [1,40] */
3569 int nice_rlim = 20 - nice;
3571 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3572 capable(CAP_SYS_NICE));
3575 #ifdef __ARCH_WANT_SYS_NICE
3578 * sys_nice - change the priority of the current process.
3579 * @increment: priority increment
3581 * sys_setpriority is a more generic, but much slower function that
3582 * does similar things.
3584 SYSCALL_DEFINE1(nice, int, increment)
3589 * Setpriority might change our priority at the same moment.
3590 * We don't have to worry. Conceptually one call occurs first
3591 * and we have a single winner.
3593 if (increment < -40)
3598 nice = TASK_NICE(current) + increment;
3604 if (increment < 0 && !can_nice(current, nice))
3607 retval = security_task_setnice(current, nice);
3611 set_user_nice(current, nice);
3618 * task_prio - return the priority value of a given task.
3619 * @p: the task in question.
3621 * This is the priority value as seen by users in /proc.
3622 * RT tasks are offset by -200. Normal tasks are centered
3623 * around 0, value goes from -16 to +15.
3625 int task_prio(const struct task_struct *p)
3627 return p->prio - MAX_RT_PRIO;
3631 * task_nice - return the nice value of a given task.
3632 * @p: the task in question.
3634 int task_nice(const struct task_struct *p)
3636 return TASK_NICE(p);
3638 EXPORT_SYMBOL(task_nice);
3641 * idle_cpu - is a given cpu idle currently?
3642 * @cpu: the processor in question.
3644 int idle_cpu(int cpu)
3646 struct rq *rq = cpu_rq(cpu);
3648 if (rq->curr != rq->idle)
3655 if (!llist_empty(&rq->wake_list))
3663 * idle_task - return the idle task for a given cpu.
3664 * @cpu: the processor in question.
3666 struct task_struct *idle_task(int cpu)
3668 return cpu_rq(cpu)->idle;
3672 * find_process_by_pid - find a process with a matching PID value.
3673 * @pid: the pid in question.
3675 static struct task_struct *find_process_by_pid(pid_t pid)
3677 return pid ? find_task_by_vpid(pid) : current;
3680 /* Actually do priority change: must hold rq lock. */
3682 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3685 p->rt_priority = prio;
3686 p->normal_prio = normal_prio(p);
3687 /* we are holding p->pi_lock already */
3688 p->prio = rt_mutex_getprio(p);
3689 if (rt_prio(p->prio))
3690 p->sched_class = &rt_sched_class;
3692 p->sched_class = &fair_sched_class;
3697 * check the target process has a UID that matches the current process's
3699 static bool check_same_owner(struct task_struct *p)
3701 const struct cred *cred = current_cred(), *pcred;
3705 pcred = __task_cred(p);
3706 match = (uid_eq(cred->euid, pcred->euid) ||
3707 uid_eq(cred->euid, pcred->uid));
3712 static int __sched_setscheduler(struct task_struct *p, int policy,
3713 const struct sched_param *param, bool user)
3715 int retval, oldprio, oldpolicy = -1, on_rq, running;
3716 unsigned long flags;
3717 const struct sched_class *prev_class;
3721 /* may grab non-irq protected spin_locks */
3722 BUG_ON(in_interrupt());
3724 /* double check policy once rq lock held */
3726 reset_on_fork = p->sched_reset_on_fork;
3727 policy = oldpolicy = p->policy;
3729 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3730 policy &= ~SCHED_RESET_ON_FORK;
3732 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3733 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3734 policy != SCHED_IDLE)
3739 * Valid priorities for SCHED_FIFO and SCHED_RR are
3740 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3741 * SCHED_BATCH and SCHED_IDLE is 0.
3743 if (param->sched_priority < 0 ||
3744 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3745 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3747 if (rt_policy(policy) != (param->sched_priority != 0))
3751 * Allow unprivileged RT tasks to decrease priority:
3753 if (user && !capable(CAP_SYS_NICE)) {
3754 if (rt_policy(policy)) {
3755 unsigned long rlim_rtprio =
3756 task_rlimit(p, RLIMIT_RTPRIO);
3758 /* can't set/change the rt policy */
3759 if (policy != p->policy && !rlim_rtprio)
3762 /* can't increase priority */
3763 if (param->sched_priority > p->rt_priority &&
3764 param->sched_priority > rlim_rtprio)
3769 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3770 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3772 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3773 if (!can_nice(p, TASK_NICE(p)))
3777 /* can't change other user's priorities */
3778 if (!check_same_owner(p))
3781 /* Normal users shall not reset the sched_reset_on_fork flag */
3782 if (p->sched_reset_on_fork && !reset_on_fork)
3787 retval = security_task_setscheduler(p);
3793 * make sure no PI-waiters arrive (or leave) while we are
3794 * changing the priority of the task:
3796 * To be able to change p->policy safely, the appropriate
3797 * runqueue lock must be held.
3799 rq = task_rq_lock(p, &flags);
3802 * Changing the policy of the stop threads its a very bad idea
3804 if (p == rq->stop) {
3805 task_rq_unlock(rq, p, &flags);
3810 * If not changing anything there's no need to proceed further:
3812 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3813 param->sched_priority == p->rt_priority))) {
3814 task_rq_unlock(rq, p, &flags);
3818 #ifdef CONFIG_RT_GROUP_SCHED
3821 * Do not allow realtime tasks into groups that have no runtime
3824 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3825 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3826 !task_group_is_autogroup(task_group(p))) {
3827 task_rq_unlock(rq, p, &flags);
3833 /* recheck policy now with rq lock held */
3834 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3835 policy = oldpolicy = -1;
3836 task_rq_unlock(rq, p, &flags);
3840 running = task_current(rq, p);
3842 dequeue_task(rq, p, 0);
3844 p->sched_class->put_prev_task(rq, p);
3846 p->sched_reset_on_fork = reset_on_fork;
3849 prev_class = p->sched_class;
3850 __setscheduler(rq, p, policy, param->sched_priority);
3853 p->sched_class->set_curr_task(rq);
3855 enqueue_task(rq, p, 0);
3857 check_class_changed(rq, p, prev_class, oldprio);
3858 task_rq_unlock(rq, p, &flags);
3860 rt_mutex_adjust_pi(p);
3866 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3867 * @p: the task in question.
3868 * @policy: new policy.
3869 * @param: structure containing the new RT priority.
3871 * NOTE that the task may be already dead.
3873 int sched_setscheduler(struct task_struct *p, int policy,
3874 const struct sched_param *param)
3876 return __sched_setscheduler(p, policy, param, true);
3878 EXPORT_SYMBOL_GPL(sched_setscheduler);
3881 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3882 * @p: the task in question.
3883 * @policy: new policy.
3884 * @param: structure containing the new RT priority.
3886 * Just like sched_setscheduler, only don't bother checking if the
3887 * current context has permission. For example, this is needed in
3888 * stop_machine(): we create temporary high priority worker threads,
3889 * but our caller might not have that capability.
3891 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3892 const struct sched_param *param)
3894 return __sched_setscheduler(p, policy, param, false);
3898 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3900 struct sched_param lparam;
3901 struct task_struct *p;
3904 if (!param || pid < 0)
3906 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3911 p = find_process_by_pid(pid);
3913 retval = sched_setscheduler(p, policy, &lparam);
3920 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3921 * @pid: the pid in question.
3922 * @policy: new policy.
3923 * @param: structure containing the new RT priority.
3925 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3926 struct sched_param __user *, param)
3928 /* negative values for policy are not valid */
3932 return do_sched_setscheduler(pid, policy, param);
3936 * sys_sched_setparam - set/change the RT priority of a thread
3937 * @pid: the pid in question.
3938 * @param: structure containing the new RT priority.
3940 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3942 return do_sched_setscheduler(pid, -1, param);
3946 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3947 * @pid: the pid in question.
3949 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3951 struct task_struct *p;
3959 p = find_process_by_pid(pid);
3961 retval = security_task_getscheduler(p);
3964 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3971 * sys_sched_getparam - get the RT priority of a thread
3972 * @pid: the pid in question.
3973 * @param: structure containing the RT priority.
3975 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3977 struct sched_param lp;
3978 struct task_struct *p;
3981 if (!param || pid < 0)
3985 p = find_process_by_pid(pid);
3990 retval = security_task_getscheduler(p);
3994 lp.sched_priority = p->rt_priority;
3998 * This one might sleep, we cannot do it with a spinlock held ...
4000 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4009 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4011 cpumask_var_t cpus_allowed, new_mask;
4012 struct task_struct *p;
4018 p = find_process_by_pid(pid);
4025 /* Prevent p going away */
4029 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4033 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4035 goto out_free_cpus_allowed;
4038 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4041 retval = security_task_setscheduler(p);
4045 cpuset_cpus_allowed(p, cpus_allowed);
4046 cpumask_and(new_mask, in_mask, cpus_allowed);
4048 retval = set_cpus_allowed_ptr(p, new_mask);
4051 cpuset_cpus_allowed(p, cpus_allowed);
4052 if (!cpumask_subset(new_mask, cpus_allowed)) {
4054 * We must have raced with a concurrent cpuset
4055 * update. Just reset the cpus_allowed to the
4056 * cpuset's cpus_allowed
4058 cpumask_copy(new_mask, cpus_allowed);
4063 free_cpumask_var(new_mask);
4064 out_free_cpus_allowed:
4065 free_cpumask_var(cpus_allowed);
4072 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4073 struct cpumask *new_mask)
4075 if (len < cpumask_size())
4076 cpumask_clear(new_mask);
4077 else if (len > cpumask_size())
4078 len = cpumask_size();
4080 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4084 * sys_sched_setaffinity - set the cpu affinity of a process
4085 * @pid: pid of the process
4086 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4087 * @user_mask_ptr: user-space pointer to the new cpu mask
4089 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4090 unsigned long __user *, user_mask_ptr)
4092 cpumask_var_t new_mask;
4095 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4098 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4100 retval = sched_setaffinity(pid, new_mask);
4101 free_cpumask_var(new_mask);
4105 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4107 struct task_struct *p;
4108 unsigned long flags;
4115 p = find_process_by_pid(pid);
4119 retval = security_task_getscheduler(p);
4123 raw_spin_lock_irqsave(&p->pi_lock, flags);
4124 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4125 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4135 * sys_sched_getaffinity - get the cpu affinity of a process
4136 * @pid: pid of the process
4137 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4138 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4140 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4141 unsigned long __user *, user_mask_ptr)
4146 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4148 if (len & (sizeof(unsigned long)-1))
4151 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4154 ret = sched_getaffinity(pid, mask);
4156 size_t retlen = min_t(size_t, len, cpumask_size());
4158 if (copy_to_user(user_mask_ptr, mask, retlen))
4163 free_cpumask_var(mask);
4169 * sys_sched_yield - yield the current processor to other threads.
4171 * This function yields the current CPU to other tasks. If there are no
4172 * other threads running on this CPU then this function will return.
4174 SYSCALL_DEFINE0(sched_yield)
4176 struct rq *rq = this_rq_lock();
4178 schedstat_inc(rq, yld_count);
4179 current->sched_class->yield_task(rq);
4182 * Since we are going to call schedule() anyway, there's
4183 * no need to preempt or enable interrupts:
4185 __release(rq->lock);
4186 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4187 do_raw_spin_unlock(&rq->lock);
4188 sched_preempt_enable_no_resched();
4195 static inline int should_resched(void)
4197 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4200 static void __cond_resched(void)
4202 add_preempt_count(PREEMPT_ACTIVE);
4204 sub_preempt_count(PREEMPT_ACTIVE);
4207 int __sched _cond_resched(void)
4209 if (should_resched()) {
4215 EXPORT_SYMBOL(_cond_resched);
4218 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4219 * call schedule, and on return reacquire the lock.
4221 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4222 * operations here to prevent schedule() from being called twice (once via
4223 * spin_unlock(), once by hand).
4225 int __cond_resched_lock(spinlock_t *lock)
4227 int resched = should_resched();
4230 lockdep_assert_held(lock);
4232 if (spin_needbreak(lock) || resched) {
4243 EXPORT_SYMBOL(__cond_resched_lock);
4245 int __sched __cond_resched_softirq(void)
4247 BUG_ON(!in_softirq());
4249 if (should_resched()) {
4257 EXPORT_SYMBOL(__cond_resched_softirq);
4260 * yield - yield the current processor to other threads.
4262 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4264 * The scheduler is at all times free to pick the calling task as the most
4265 * eligible task to run, if removing the yield() call from your code breaks
4266 * it, its already broken.
4268 * Typical broken usage is:
4273 * where one assumes that yield() will let 'the other' process run that will
4274 * make event true. If the current task is a SCHED_FIFO task that will never
4275 * happen. Never use yield() as a progress guarantee!!
4277 * If you want to use yield() to wait for something, use wait_event().
4278 * If you want to use yield() to be 'nice' for others, use cond_resched().
4279 * If you still want to use yield(), do not!
4281 void __sched yield(void)
4283 set_current_state(TASK_RUNNING);
4286 EXPORT_SYMBOL(yield);
4289 * yield_to - yield the current processor to another thread in
4290 * your thread group, or accelerate that thread toward the
4291 * processor it's on.
4293 * @preempt: whether task preemption is allowed or not
4295 * It's the caller's job to ensure that the target task struct
4296 * can't go away on us before we can do any checks.
4298 * Returns true if we indeed boosted the target task.
4300 bool __sched yield_to(struct task_struct *p, bool preempt)
4302 struct task_struct *curr = current;
4303 struct rq *rq, *p_rq;
4304 unsigned long flags;
4307 local_irq_save(flags);
4312 double_rq_lock(rq, p_rq);
4313 while (task_rq(p) != p_rq) {
4314 double_rq_unlock(rq, p_rq);
4318 if (!curr->sched_class->yield_to_task)
4321 if (curr->sched_class != p->sched_class)
4324 if (task_running(p_rq, p) || p->state)
4327 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4329 schedstat_inc(rq, yld_count);
4331 * Make p's CPU reschedule; pick_next_entity takes care of
4334 if (preempt && rq != p_rq)
4335 resched_task(p_rq->curr);
4339 double_rq_unlock(rq, p_rq);
4340 local_irq_restore(flags);
4347 EXPORT_SYMBOL_GPL(yield_to);
4350 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4351 * that process accounting knows that this is a task in IO wait state.
4353 void __sched io_schedule(void)
4355 struct rq *rq = raw_rq();
4357 delayacct_blkio_start();
4358 atomic_inc(&rq->nr_iowait);
4359 blk_flush_plug(current);
4360 current->in_iowait = 1;
4362 current->in_iowait = 0;
4363 atomic_dec(&rq->nr_iowait);
4364 delayacct_blkio_end();
4366 EXPORT_SYMBOL(io_schedule);
4368 long __sched io_schedule_timeout(long timeout)
4370 struct rq *rq = raw_rq();
4373 delayacct_blkio_start();
4374 atomic_inc(&rq->nr_iowait);
4375 blk_flush_plug(current);
4376 current->in_iowait = 1;
4377 ret = schedule_timeout(timeout);
4378 current->in_iowait = 0;
4379 atomic_dec(&rq->nr_iowait);
4380 delayacct_blkio_end();
4385 * sys_sched_get_priority_max - return maximum RT priority.
4386 * @policy: scheduling class.
4388 * this syscall returns the maximum rt_priority that can be used
4389 * by a given scheduling class.
4391 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4398 ret = MAX_USER_RT_PRIO-1;
4410 * sys_sched_get_priority_min - return minimum RT priority.
4411 * @policy: scheduling class.
4413 * this syscall returns the minimum rt_priority that can be used
4414 * by a given scheduling class.
4416 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4434 * sys_sched_rr_get_interval - return the default timeslice of a process.
4435 * @pid: pid of the process.
4436 * @interval: userspace pointer to the timeslice value.
4438 * this syscall writes the default timeslice value of a given process
4439 * into the user-space timespec buffer. A value of '0' means infinity.
4441 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4442 struct timespec __user *, interval)
4444 struct task_struct *p;
4445 unsigned int time_slice;
4446 unsigned long flags;
4456 p = find_process_by_pid(pid);
4460 retval = security_task_getscheduler(p);
4464 rq = task_rq_lock(p, &flags);
4465 time_slice = p->sched_class->get_rr_interval(rq, p);
4466 task_rq_unlock(rq, p, &flags);
4469 jiffies_to_timespec(time_slice, &t);
4470 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4478 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4480 void sched_show_task(struct task_struct *p)
4482 unsigned long free = 0;
4485 state = p->state ? __ffs(p->state) + 1 : 0;
4486 printk(KERN_INFO "%-15.15s %c", p->comm,
4487 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4488 #if BITS_PER_LONG == 32
4489 if (state == TASK_RUNNING)
4490 printk(KERN_CONT " running ");
4492 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4494 if (state == TASK_RUNNING)
4495 printk(KERN_CONT " running task ");
4497 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4499 #ifdef CONFIG_DEBUG_STACK_USAGE
4500 free = stack_not_used(p);
4502 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4503 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4504 (unsigned long)task_thread_info(p)->flags);
4506 show_stack(p, NULL);
4509 void show_state_filter(unsigned long state_filter)
4511 struct task_struct *g, *p;
4513 #if BITS_PER_LONG == 32
4515 " task PC stack pid father\n");
4518 " task PC stack pid father\n");
4521 do_each_thread(g, p) {
4523 * reset the NMI-timeout, listing all files on a slow
4524 * console might take a lot of time:
4526 touch_nmi_watchdog();
4527 if (!state_filter || (p->state & state_filter))
4529 } while_each_thread(g, p);
4531 touch_all_softlockup_watchdogs();
4533 #ifdef CONFIG_SCHED_DEBUG
4534 sysrq_sched_debug_show();
4538 * Only show locks if all tasks are dumped:
4541 debug_show_all_locks();
4544 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4546 idle->sched_class = &idle_sched_class;
4550 * init_idle - set up an idle thread for a given CPU
4551 * @idle: task in question
4552 * @cpu: cpu the idle task belongs to
4554 * NOTE: this function does not set the idle thread's NEED_RESCHED
4555 * flag, to make booting more robust.
4557 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4559 struct rq *rq = cpu_rq(cpu);
4560 unsigned long flags;
4562 raw_spin_lock_irqsave(&rq->lock, flags);
4565 idle->state = TASK_RUNNING;
4566 idle->se.exec_start = sched_clock();
4568 do_set_cpus_allowed(idle, cpumask_of(cpu));
4570 * We're having a chicken and egg problem, even though we are
4571 * holding rq->lock, the cpu isn't yet set to this cpu so the
4572 * lockdep check in task_group() will fail.
4574 * Similar case to sched_fork(). / Alternatively we could
4575 * use task_rq_lock() here and obtain the other rq->lock.
4580 __set_task_cpu(idle, cpu);
4583 rq->curr = rq->idle = idle;
4584 #if defined(CONFIG_SMP)
4587 raw_spin_unlock_irqrestore(&rq->lock, flags);
4589 /* Set the preempt count _outside_ the spinlocks! */
4590 task_thread_info(idle)->preempt_count = 0;
4593 * The idle tasks have their own, simple scheduling class:
4595 idle->sched_class = &idle_sched_class;
4596 ftrace_graph_init_idle_task(idle, cpu);
4597 #if defined(CONFIG_SMP)
4598 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4603 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4605 if (p->sched_class && p->sched_class->set_cpus_allowed)
4606 p->sched_class->set_cpus_allowed(p, new_mask);
4608 cpumask_copy(&p->cpus_allowed, new_mask);
4609 p->nr_cpus_allowed = cpumask_weight(new_mask);
4613 * This is how migration works:
4615 * 1) we invoke migration_cpu_stop() on the target CPU using
4617 * 2) stopper starts to run (implicitly forcing the migrated thread
4619 * 3) it checks whether the migrated task is still in the wrong runqueue.
4620 * 4) if it's in the wrong runqueue then the migration thread removes
4621 * it and puts it into the right queue.
4622 * 5) stopper completes and stop_one_cpu() returns and the migration
4627 * Change a given task's CPU affinity. Migrate the thread to a
4628 * proper CPU and schedule it away if the CPU it's executing on
4629 * is removed from the allowed bitmask.
4631 * NOTE: the caller must have a valid reference to the task, the
4632 * task must not exit() & deallocate itself prematurely. The
4633 * call is not atomic; no spinlocks may be held.
4635 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4637 unsigned long flags;
4639 unsigned int dest_cpu;
4642 rq = task_rq_lock(p, &flags);
4644 if (cpumask_equal(&p->cpus_allowed, new_mask))
4647 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4652 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4657 do_set_cpus_allowed(p, new_mask);
4659 /* Can the task run on the task's current CPU? If so, we're done */
4660 if (cpumask_test_cpu(task_cpu(p), new_mask))
4663 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4665 struct migration_arg arg = { p, dest_cpu };
4666 /* Need help from migration thread: drop lock and wait. */
4667 task_rq_unlock(rq, p, &flags);
4668 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4669 tlb_migrate_finish(p->mm);
4673 task_rq_unlock(rq, p, &flags);
4677 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4680 * Move (not current) task off this cpu, onto dest cpu. We're doing
4681 * this because either it can't run here any more (set_cpus_allowed()
4682 * away from this CPU, or CPU going down), or because we're
4683 * attempting to rebalance this task on exec (sched_exec).
4685 * So we race with normal scheduler movements, but that's OK, as long
4686 * as the task is no longer on this CPU.
4688 * Returns non-zero if task was successfully migrated.
4690 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4692 struct rq *rq_dest, *rq_src;
4695 if (unlikely(!cpu_active(dest_cpu)))
4698 rq_src = cpu_rq(src_cpu);
4699 rq_dest = cpu_rq(dest_cpu);
4701 raw_spin_lock(&p->pi_lock);
4702 double_rq_lock(rq_src, rq_dest);
4703 /* Already moved. */
4704 if (task_cpu(p) != src_cpu)
4706 /* Affinity changed (again). */
4707 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4711 * If we're not on a rq, the next wake-up will ensure we're
4715 dequeue_task(rq_src, p, 0);
4716 set_task_cpu(p, dest_cpu);
4717 enqueue_task(rq_dest, p, 0);
4718 check_preempt_curr(rq_dest, p, 0);
4723 double_rq_unlock(rq_src, rq_dest);
4724 raw_spin_unlock(&p->pi_lock);
4729 * migration_cpu_stop - this will be executed by a highprio stopper thread
4730 * and performs thread migration by bumping thread off CPU then
4731 * 'pushing' onto another runqueue.
4733 static int migration_cpu_stop(void *data)
4735 struct migration_arg *arg = data;
4738 * The original target cpu might have gone down and we might
4739 * be on another cpu but it doesn't matter.
4741 local_irq_disable();
4742 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4747 #ifdef CONFIG_HOTPLUG_CPU
4750 * Ensures that the idle task is using init_mm right before its cpu goes
4753 void idle_task_exit(void)
4755 struct mm_struct *mm = current->active_mm;
4757 BUG_ON(cpu_online(smp_processor_id()));
4760 switch_mm(mm, &init_mm, current);
4765 * Since this CPU is going 'away' for a while, fold any nr_active delta
4766 * we might have. Assumes we're called after migrate_tasks() so that the
4767 * nr_active count is stable.
4769 * Also see the comment "Global load-average calculations".
4771 static void calc_load_migrate(struct rq *rq)
4773 long delta = calc_load_fold_active(rq);
4775 atomic_long_add(delta, &calc_load_tasks);
4779 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4780 * try_to_wake_up()->select_task_rq().
4782 * Called with rq->lock held even though we'er in stop_machine() and
4783 * there's no concurrency possible, we hold the required locks anyway
4784 * because of lock validation efforts.
4786 static void migrate_tasks(unsigned int dead_cpu)
4788 struct rq *rq = cpu_rq(dead_cpu);
4789 struct task_struct *next, *stop = rq->stop;
4793 * Fudge the rq selection such that the below task selection loop
4794 * doesn't get stuck on the currently eligible stop task.
4796 * We're currently inside stop_machine() and the rq is either stuck
4797 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4798 * either way we should never end up calling schedule() until we're
4805 * There's this thread running, bail when that's the only
4808 if (rq->nr_running == 1)
4811 next = pick_next_task(rq);
4813 next->sched_class->put_prev_task(rq, next);
4815 /* Find suitable destination for @next, with force if needed. */
4816 dest_cpu = select_fallback_rq(dead_cpu, next);
4817 raw_spin_unlock(&rq->lock);
4819 __migrate_task(next, dead_cpu, dest_cpu);
4821 raw_spin_lock(&rq->lock);
4827 #endif /* CONFIG_HOTPLUG_CPU */
4829 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4831 static struct ctl_table sd_ctl_dir[] = {
4833 .procname = "sched_domain",
4839 static struct ctl_table sd_ctl_root[] = {
4841 .procname = "kernel",
4843 .child = sd_ctl_dir,
4848 static struct ctl_table *sd_alloc_ctl_entry(int n)
4850 struct ctl_table *entry =
4851 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4856 static void sd_free_ctl_entry(struct ctl_table **tablep)
4858 struct ctl_table *entry;
4861 * In the intermediate directories, both the child directory and
4862 * procname are dynamically allocated and could fail but the mode
4863 * will always be set. In the lowest directory the names are
4864 * static strings and all have proc handlers.
4866 for (entry = *tablep; entry->mode; entry++) {
4868 sd_free_ctl_entry(&entry->child);
4869 if (entry->proc_handler == NULL)
4870 kfree(entry->procname);
4877 static int min_load_idx = 0;
4878 static int max_load_idx = CPU_LOAD_IDX_MAX;
4881 set_table_entry(struct ctl_table *entry,
4882 const char *procname, void *data, int maxlen,
4883 umode_t mode, proc_handler *proc_handler,
4886 entry->procname = procname;
4888 entry->maxlen = maxlen;
4890 entry->proc_handler = proc_handler;
4893 entry->extra1 = &min_load_idx;
4894 entry->extra2 = &max_load_idx;
4898 static struct ctl_table *
4899 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4901 struct ctl_table *table = sd_alloc_ctl_entry(13);
4906 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4907 sizeof(long), 0644, proc_doulongvec_minmax, false);
4908 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4909 sizeof(long), 0644, proc_doulongvec_minmax, false);
4910 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4911 sizeof(int), 0644, proc_dointvec_minmax, true);
4912 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4913 sizeof(int), 0644, proc_dointvec_minmax, true);
4914 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4915 sizeof(int), 0644, proc_dointvec_minmax, true);
4916 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4917 sizeof(int), 0644, proc_dointvec_minmax, true);
4918 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4919 sizeof(int), 0644, proc_dointvec_minmax, true);
4920 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4921 sizeof(int), 0644, proc_dointvec_minmax, false);
4922 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4923 sizeof(int), 0644, proc_dointvec_minmax, false);
4924 set_table_entry(&table[9], "cache_nice_tries",
4925 &sd->cache_nice_tries,
4926 sizeof(int), 0644, proc_dointvec_minmax, false);
4927 set_table_entry(&table[10], "flags", &sd->flags,
4928 sizeof(int), 0644, proc_dointvec_minmax, false);
4929 set_table_entry(&table[11], "name", sd->name,
4930 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4931 /* &table[12] is terminator */
4936 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4938 struct ctl_table *entry, *table;
4939 struct sched_domain *sd;
4940 int domain_num = 0, i;
4943 for_each_domain(cpu, sd)
4945 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4950 for_each_domain(cpu, sd) {
4951 snprintf(buf, 32, "domain%d", i);
4952 entry->procname = kstrdup(buf, GFP_KERNEL);
4954 entry->child = sd_alloc_ctl_domain_table(sd);
4961 static struct ctl_table_header *sd_sysctl_header;
4962 static void register_sched_domain_sysctl(void)
4964 int i, cpu_num = num_possible_cpus();
4965 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4968 WARN_ON(sd_ctl_dir[0].child);
4969 sd_ctl_dir[0].child = entry;
4974 for_each_possible_cpu(i) {
4975 snprintf(buf, 32, "cpu%d", i);
4976 entry->procname = kstrdup(buf, GFP_KERNEL);
4978 entry->child = sd_alloc_ctl_cpu_table(i);
4982 WARN_ON(sd_sysctl_header);
4983 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4986 /* may be called multiple times per register */
4987 static void unregister_sched_domain_sysctl(void)
4989 if (sd_sysctl_header)
4990 unregister_sysctl_table(sd_sysctl_header);
4991 sd_sysctl_header = NULL;
4992 if (sd_ctl_dir[0].child)
4993 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4996 static void register_sched_domain_sysctl(void)
4999 static void unregister_sched_domain_sysctl(void)
5004 static void set_rq_online(struct rq *rq)
5007 const struct sched_class *class;
5009 cpumask_set_cpu(rq->cpu, rq->rd->online);
5012 for_each_class(class) {
5013 if (class->rq_online)
5014 class->rq_online(rq);
5019 static void set_rq_offline(struct rq *rq)
5022 const struct sched_class *class;
5024 for_each_class(class) {
5025 if (class->rq_offline)
5026 class->rq_offline(rq);
5029 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5035 * migration_call - callback that gets triggered when a CPU is added.
5036 * Here we can start up the necessary migration thread for the new CPU.
5038 static int __cpuinit
5039 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5041 int cpu = (long)hcpu;
5042 unsigned long flags;
5043 struct rq *rq = cpu_rq(cpu);
5045 switch (action & ~CPU_TASKS_FROZEN) {
5047 case CPU_UP_PREPARE:
5048 rq->calc_load_update = calc_load_update;
5052 /* Update our root-domain */
5053 raw_spin_lock_irqsave(&rq->lock, flags);
5055 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5059 raw_spin_unlock_irqrestore(&rq->lock, flags);
5062 #ifdef CONFIG_HOTPLUG_CPU
5064 sched_ttwu_pending();
5065 /* Update our root-domain */
5066 raw_spin_lock_irqsave(&rq->lock, flags);
5068 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5072 BUG_ON(rq->nr_running != 1); /* the migration thread */
5073 raw_spin_unlock_irqrestore(&rq->lock, flags);
5077 calc_load_migrate(rq);
5082 update_max_interval();
5088 * Register at high priority so that task migration (migrate_all_tasks)
5089 * happens before everything else. This has to be lower priority than
5090 * the notifier in the perf_event subsystem, though.
5092 static struct notifier_block __cpuinitdata migration_notifier = {
5093 .notifier_call = migration_call,
5094 .priority = CPU_PRI_MIGRATION,
5097 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5098 unsigned long action, void *hcpu)
5100 switch (action & ~CPU_TASKS_FROZEN) {
5102 case CPU_DOWN_FAILED:
5103 set_cpu_active((long)hcpu, true);
5110 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5111 unsigned long action, void *hcpu)
5113 switch (action & ~CPU_TASKS_FROZEN) {
5114 case CPU_DOWN_PREPARE:
5115 set_cpu_active((long)hcpu, false);
5122 static int __init migration_init(void)
5124 void *cpu = (void *)(long)smp_processor_id();
5127 /* Initialize migration for the boot CPU */
5128 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5129 BUG_ON(err == NOTIFY_BAD);
5130 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5131 register_cpu_notifier(&migration_notifier);
5133 /* Register cpu active notifiers */
5134 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5135 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5139 early_initcall(migration_init);
5144 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5146 #ifdef CONFIG_SCHED_DEBUG
5148 static __read_mostly int sched_debug_enabled;
5150 static int __init sched_debug_setup(char *str)
5152 sched_debug_enabled = 1;
5156 early_param("sched_debug", sched_debug_setup);
5158 static inline bool sched_debug(void)
5160 return sched_debug_enabled;
5163 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5164 struct cpumask *groupmask)
5166 struct sched_group *group = sd->groups;
5169 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5170 cpumask_clear(groupmask);
5172 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5174 if (!(sd->flags & SD_LOAD_BALANCE)) {
5175 printk("does not load-balance\n");
5177 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5182 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5184 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5185 printk(KERN_ERR "ERROR: domain->span does not contain "
5188 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5189 printk(KERN_ERR "ERROR: domain->groups does not contain"
5193 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5197 printk(KERN_ERR "ERROR: group is NULL\n");
5202 * Even though we initialize ->power to something semi-sane,
5203 * we leave power_orig unset. This allows us to detect if
5204 * domain iteration is still funny without causing /0 traps.
5206 if (!group->sgp->power_orig) {
5207 printk(KERN_CONT "\n");
5208 printk(KERN_ERR "ERROR: domain->cpu_power not "
5213 if (!cpumask_weight(sched_group_cpus(group))) {
5214 printk(KERN_CONT "\n");
5215 printk(KERN_ERR "ERROR: empty group\n");
5219 if (!(sd->flags & SD_OVERLAP) &&
5220 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5221 printk(KERN_CONT "\n");
5222 printk(KERN_ERR "ERROR: repeated CPUs\n");
5226 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5228 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5230 printk(KERN_CONT " %s", str);
5231 if (group->sgp->power != SCHED_POWER_SCALE) {
5232 printk(KERN_CONT " (cpu_power = %d)",
5236 group = group->next;
5237 } while (group != sd->groups);
5238 printk(KERN_CONT "\n");
5240 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5241 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5244 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5245 printk(KERN_ERR "ERROR: parent span is not a superset "
5246 "of domain->span\n");
5250 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5254 if (!sched_debug_enabled)
5258 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5262 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5265 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5273 #else /* !CONFIG_SCHED_DEBUG */
5274 # define sched_domain_debug(sd, cpu) do { } while (0)
5275 static inline bool sched_debug(void)
5279 #endif /* CONFIG_SCHED_DEBUG */
5281 static int sd_degenerate(struct sched_domain *sd)
5283 if (cpumask_weight(sched_domain_span(sd)) == 1)
5286 /* Following flags need at least 2 groups */
5287 if (sd->flags & (SD_LOAD_BALANCE |
5288 SD_BALANCE_NEWIDLE |
5292 SD_SHARE_PKG_RESOURCES)) {
5293 if (sd->groups != sd->groups->next)
5297 /* Following flags don't use groups */
5298 if (sd->flags & (SD_WAKE_AFFINE))
5305 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5307 unsigned long cflags = sd->flags, pflags = parent->flags;
5309 if (sd_degenerate(parent))
5312 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5315 /* Flags needing groups don't count if only 1 group in parent */
5316 if (parent->groups == parent->groups->next) {
5317 pflags &= ~(SD_LOAD_BALANCE |
5318 SD_BALANCE_NEWIDLE |
5322 SD_SHARE_PKG_RESOURCES);
5323 if (nr_node_ids == 1)
5324 pflags &= ~SD_SERIALIZE;
5326 if (~cflags & pflags)
5332 static void free_rootdomain(struct rcu_head *rcu)
5334 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5336 cpupri_cleanup(&rd->cpupri);
5337 free_cpumask_var(rd->rto_mask);
5338 free_cpumask_var(rd->online);
5339 free_cpumask_var(rd->span);
5343 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5345 struct root_domain *old_rd = NULL;
5346 unsigned long flags;
5348 raw_spin_lock_irqsave(&rq->lock, flags);
5353 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5356 cpumask_clear_cpu(rq->cpu, old_rd->span);
5359 * If we dont want to free the old_rt yet then
5360 * set old_rd to NULL to skip the freeing later
5363 if (!atomic_dec_and_test(&old_rd->refcount))
5367 atomic_inc(&rd->refcount);
5370 cpumask_set_cpu(rq->cpu, rd->span);
5371 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5374 raw_spin_unlock_irqrestore(&rq->lock, flags);
5377 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5380 static int init_rootdomain(struct root_domain *rd)
5382 memset(rd, 0, sizeof(*rd));
5384 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5386 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5388 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5391 if (cpupri_init(&rd->cpupri) != 0)
5396 free_cpumask_var(rd->rto_mask);
5398 free_cpumask_var(rd->online);
5400 free_cpumask_var(rd->span);
5406 * By default the system creates a single root-domain with all cpus as
5407 * members (mimicking the global state we have today).
5409 struct root_domain def_root_domain;
5411 static void init_defrootdomain(void)
5413 init_rootdomain(&def_root_domain);
5415 atomic_set(&def_root_domain.refcount, 1);
5418 static struct root_domain *alloc_rootdomain(void)
5420 struct root_domain *rd;
5422 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5426 if (init_rootdomain(rd) != 0) {
5434 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5436 struct sched_group *tmp, *first;
5445 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5450 } while (sg != first);
5453 static void free_sched_domain(struct rcu_head *rcu)
5455 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5458 * If its an overlapping domain it has private groups, iterate and
5461 if (sd->flags & SD_OVERLAP) {
5462 free_sched_groups(sd->groups, 1);
5463 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5464 kfree(sd->groups->sgp);
5470 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5472 call_rcu(&sd->rcu, free_sched_domain);
5475 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5477 for (; sd; sd = sd->parent)
5478 destroy_sched_domain(sd, cpu);
5482 * Keep a special pointer to the highest sched_domain that has
5483 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5484 * allows us to avoid some pointer chasing select_idle_sibling().
5486 * Also keep a unique ID per domain (we use the first cpu number in
5487 * the cpumask of the domain), this allows us to quickly tell if
5488 * two cpus are in the same cache domain, see cpus_share_cache().
5490 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5491 DEFINE_PER_CPU(int, sd_llc_id);
5493 static void update_top_cache_domain(int cpu)
5495 struct sched_domain *sd;
5498 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5500 id = cpumask_first(sched_domain_span(sd));
5502 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5503 per_cpu(sd_llc_id, cpu) = id;
5507 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5508 * hold the hotplug lock.
5511 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5513 struct rq *rq = cpu_rq(cpu);
5514 struct sched_domain *tmp;
5516 /* Remove the sched domains which do not contribute to scheduling. */
5517 for (tmp = sd; tmp; ) {
5518 struct sched_domain *parent = tmp->parent;
5522 if (sd_parent_degenerate(tmp, parent)) {
5523 tmp->parent = parent->parent;
5525 parent->parent->child = tmp;
5526 destroy_sched_domain(parent, cpu);
5531 if (sd && sd_degenerate(sd)) {
5534 destroy_sched_domain(tmp, cpu);
5539 sched_domain_debug(sd, cpu);
5541 rq_attach_root(rq, rd);
5543 rcu_assign_pointer(rq->sd, sd);
5544 destroy_sched_domains(tmp, cpu);
5546 update_top_cache_domain(cpu);
5549 /* cpus with isolated domains */
5550 static cpumask_var_t cpu_isolated_map;
5552 /* Setup the mask of cpus configured for isolated domains */
5553 static int __init isolated_cpu_setup(char *str)
5555 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5556 cpulist_parse(str, cpu_isolated_map);
5560 __setup("isolcpus=", isolated_cpu_setup);
5562 static const struct cpumask *cpu_cpu_mask(int cpu)
5564 return cpumask_of_node(cpu_to_node(cpu));
5568 struct sched_domain **__percpu sd;
5569 struct sched_group **__percpu sg;
5570 struct sched_group_power **__percpu sgp;
5574 struct sched_domain ** __percpu sd;
5575 struct root_domain *rd;
5585 struct sched_domain_topology_level;
5587 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5588 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5590 #define SDTL_OVERLAP 0x01
5592 struct sched_domain_topology_level {
5593 sched_domain_init_f init;
5594 sched_domain_mask_f mask;
5597 struct sd_data data;
5601 * Build an iteration mask that can exclude certain CPUs from the upwards
5604 * Asymmetric node setups can result in situations where the domain tree is of
5605 * unequal depth, make sure to skip domains that already cover the entire
5608 * In that case build_sched_domains() will have terminated the iteration early
5609 * and our sibling sd spans will be empty. Domains should always include the
5610 * cpu they're built on, so check that.
5613 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5615 const struct cpumask *span = sched_domain_span(sd);
5616 struct sd_data *sdd = sd->private;
5617 struct sched_domain *sibling;
5620 for_each_cpu(i, span) {
5621 sibling = *per_cpu_ptr(sdd->sd, i);
5622 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5625 cpumask_set_cpu(i, sched_group_mask(sg));
5630 * Return the canonical balance cpu for this group, this is the first cpu
5631 * of this group that's also in the iteration mask.
5633 int group_balance_cpu(struct sched_group *sg)
5635 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5639 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5641 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5642 const struct cpumask *span = sched_domain_span(sd);
5643 struct cpumask *covered = sched_domains_tmpmask;
5644 struct sd_data *sdd = sd->private;
5645 struct sched_domain *child;
5648 cpumask_clear(covered);
5650 for_each_cpu(i, span) {
5651 struct cpumask *sg_span;
5653 if (cpumask_test_cpu(i, covered))
5656 child = *per_cpu_ptr(sdd->sd, i);
5658 /* See the comment near build_group_mask(). */
5659 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5662 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5663 GFP_KERNEL, cpu_to_node(cpu));
5668 sg_span = sched_group_cpus(sg);
5670 child = child->child;
5671 cpumask_copy(sg_span, sched_domain_span(child));
5673 cpumask_set_cpu(i, sg_span);
5675 cpumask_or(covered, covered, sg_span);
5677 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5678 if (atomic_inc_return(&sg->sgp->ref) == 1)
5679 build_group_mask(sd, sg);
5682 * Initialize sgp->power such that even if we mess up the
5683 * domains and no possible iteration will get us here, we won't
5686 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5689 * Make sure the first group of this domain contains the
5690 * canonical balance cpu. Otherwise the sched_domain iteration
5691 * breaks. See update_sg_lb_stats().
5693 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5694 group_balance_cpu(sg) == cpu)
5704 sd->groups = groups;
5709 free_sched_groups(first, 0);
5714 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5716 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5717 struct sched_domain *child = sd->child;
5720 cpu = cpumask_first(sched_domain_span(child));
5723 *sg = *per_cpu_ptr(sdd->sg, cpu);
5724 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5725 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5732 * build_sched_groups will build a circular linked list of the groups
5733 * covered by the given span, and will set each group's ->cpumask correctly,
5734 * and ->cpu_power to 0.
5736 * Assumes the sched_domain tree is fully constructed
5739 build_sched_groups(struct sched_domain *sd, int cpu)
5741 struct sched_group *first = NULL, *last = NULL;
5742 struct sd_data *sdd = sd->private;
5743 const struct cpumask *span = sched_domain_span(sd);
5744 struct cpumask *covered;
5747 get_group(cpu, sdd, &sd->groups);
5748 atomic_inc(&sd->groups->ref);
5750 if (cpu != cpumask_first(sched_domain_span(sd)))
5753 lockdep_assert_held(&sched_domains_mutex);
5754 covered = sched_domains_tmpmask;
5756 cpumask_clear(covered);
5758 for_each_cpu(i, span) {
5759 struct sched_group *sg;
5760 int group = get_group(i, sdd, &sg);
5763 if (cpumask_test_cpu(i, covered))
5766 cpumask_clear(sched_group_cpus(sg));
5768 cpumask_setall(sched_group_mask(sg));
5770 for_each_cpu(j, span) {
5771 if (get_group(j, sdd, NULL) != group)
5774 cpumask_set_cpu(j, covered);
5775 cpumask_set_cpu(j, sched_group_cpus(sg));
5790 * Initialize sched groups cpu_power.
5792 * cpu_power indicates the capacity of sched group, which is used while
5793 * distributing the load between different sched groups in a sched domain.
5794 * Typically cpu_power for all the groups in a sched domain will be same unless
5795 * there are asymmetries in the topology. If there are asymmetries, group
5796 * having more cpu_power will pickup more load compared to the group having
5799 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5801 struct sched_group *sg = sd->groups;
5803 WARN_ON(!sd || !sg);
5806 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5808 } while (sg != sd->groups);
5810 if (cpu != group_balance_cpu(sg))
5813 update_group_power(sd, cpu);
5814 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5817 int __weak arch_sd_sibling_asym_packing(void)
5819 return 0*SD_ASYM_PACKING;
5823 * Initializers for schedule domains
5824 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5827 #ifdef CONFIG_SCHED_DEBUG
5828 # define SD_INIT_NAME(sd, type) sd->name = #type
5830 # define SD_INIT_NAME(sd, type) do { } while (0)
5833 #define SD_INIT_FUNC(type) \
5834 static noinline struct sched_domain * \
5835 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5837 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5838 *sd = SD_##type##_INIT; \
5839 SD_INIT_NAME(sd, type); \
5840 sd->private = &tl->data; \
5845 #ifdef CONFIG_SCHED_SMT
5846 SD_INIT_FUNC(SIBLING)
5848 #ifdef CONFIG_SCHED_MC
5851 #ifdef CONFIG_SCHED_BOOK
5855 static int default_relax_domain_level = -1;
5856 int sched_domain_level_max;
5858 static int __init setup_relax_domain_level(char *str)
5860 if (kstrtoint(str, 0, &default_relax_domain_level))
5861 pr_warn("Unable to set relax_domain_level\n");
5865 __setup("relax_domain_level=", setup_relax_domain_level);
5867 static void set_domain_attribute(struct sched_domain *sd,
5868 struct sched_domain_attr *attr)
5872 if (!attr || attr->relax_domain_level < 0) {
5873 if (default_relax_domain_level < 0)
5876 request = default_relax_domain_level;
5878 request = attr->relax_domain_level;
5879 if (request < sd->level) {
5880 /* turn off idle balance on this domain */
5881 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5883 /* turn on idle balance on this domain */
5884 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5888 static void __sdt_free(const struct cpumask *cpu_map);
5889 static int __sdt_alloc(const struct cpumask *cpu_map);
5891 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5892 const struct cpumask *cpu_map)
5896 if (!atomic_read(&d->rd->refcount))
5897 free_rootdomain(&d->rd->rcu); /* fall through */
5899 free_percpu(d->sd); /* fall through */
5901 __sdt_free(cpu_map); /* fall through */
5907 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5908 const struct cpumask *cpu_map)
5910 memset(d, 0, sizeof(*d));
5912 if (__sdt_alloc(cpu_map))
5913 return sa_sd_storage;
5914 d->sd = alloc_percpu(struct sched_domain *);
5916 return sa_sd_storage;
5917 d->rd = alloc_rootdomain();
5920 return sa_rootdomain;
5924 * NULL the sd_data elements we've used to build the sched_domain and
5925 * sched_group structure so that the subsequent __free_domain_allocs()
5926 * will not free the data we're using.
5928 static void claim_allocations(int cpu, struct sched_domain *sd)
5930 struct sd_data *sdd = sd->private;
5932 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5933 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5935 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5936 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5938 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5939 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5942 #ifdef CONFIG_SCHED_SMT
5943 static const struct cpumask *cpu_smt_mask(int cpu)
5945 return topology_thread_cpumask(cpu);
5950 * Topology list, bottom-up.
5952 static struct sched_domain_topology_level default_topology[] = {
5953 #ifdef CONFIG_SCHED_SMT
5954 { sd_init_SIBLING, cpu_smt_mask, },
5956 #ifdef CONFIG_SCHED_MC
5957 { sd_init_MC, cpu_coregroup_mask, },
5959 #ifdef CONFIG_SCHED_BOOK
5960 { sd_init_BOOK, cpu_book_mask, },
5962 { sd_init_CPU, cpu_cpu_mask, },
5966 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5970 static int sched_domains_numa_levels;
5971 static int *sched_domains_numa_distance;
5972 static struct cpumask ***sched_domains_numa_masks;
5973 static int sched_domains_curr_level;
5975 static inline int sd_local_flags(int level)
5977 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5980 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5983 static struct sched_domain *
5984 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5986 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5987 int level = tl->numa_level;
5988 int sd_weight = cpumask_weight(
5989 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5991 *sd = (struct sched_domain){
5992 .min_interval = sd_weight,
5993 .max_interval = 2*sd_weight,
5995 .imbalance_pct = 125,
5996 .cache_nice_tries = 2,
6003 .flags = 1*SD_LOAD_BALANCE
6004 | 1*SD_BALANCE_NEWIDLE
6009 | 0*SD_SHARE_CPUPOWER
6010 | 0*SD_SHARE_PKG_RESOURCES
6012 | 0*SD_PREFER_SIBLING
6013 | sd_local_flags(level)
6015 .last_balance = jiffies,
6016 .balance_interval = sd_weight,
6018 SD_INIT_NAME(sd, NUMA);
6019 sd->private = &tl->data;
6022 * Ugly hack to pass state to sd_numa_mask()...
6024 sched_domains_curr_level = tl->numa_level;
6029 static const struct cpumask *sd_numa_mask(int cpu)
6031 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6034 static void sched_numa_warn(const char *str)
6036 static int done = false;
6044 printk(KERN_WARNING "ERROR: %s\n\n", str);
6046 for (i = 0; i < nr_node_ids; i++) {
6047 printk(KERN_WARNING " ");
6048 for (j = 0; j < nr_node_ids; j++)
6049 printk(KERN_CONT "%02d ", node_distance(i,j));
6050 printk(KERN_CONT "\n");
6052 printk(KERN_WARNING "\n");
6055 static bool find_numa_distance(int distance)
6059 if (distance == node_distance(0, 0))
6062 for (i = 0; i < sched_domains_numa_levels; i++) {
6063 if (sched_domains_numa_distance[i] == distance)
6070 static void sched_init_numa(void)
6072 int next_distance, curr_distance = node_distance(0, 0);
6073 struct sched_domain_topology_level *tl;
6077 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6078 if (!sched_domains_numa_distance)
6082 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6083 * unique distances in the node_distance() table.
6085 * Assumes node_distance(0,j) includes all distances in
6086 * node_distance(i,j) in order to avoid cubic time.
6088 next_distance = curr_distance;
6089 for (i = 0; i < nr_node_ids; i++) {
6090 for (j = 0; j < nr_node_ids; j++) {
6091 for (k = 0; k < nr_node_ids; k++) {
6092 int distance = node_distance(i, k);
6094 if (distance > curr_distance &&
6095 (distance < next_distance ||
6096 next_distance == curr_distance))
6097 next_distance = distance;
6100 * While not a strong assumption it would be nice to know
6101 * about cases where if node A is connected to B, B is not
6102 * equally connected to A.
6104 if (sched_debug() && node_distance(k, i) != distance)
6105 sched_numa_warn("Node-distance not symmetric");
6107 if (sched_debug() && i && !find_numa_distance(distance))
6108 sched_numa_warn("Node-0 not representative");
6110 if (next_distance != curr_distance) {
6111 sched_domains_numa_distance[level++] = next_distance;
6112 sched_domains_numa_levels = level;
6113 curr_distance = next_distance;
6118 * In case of sched_debug() we verify the above assumption.
6124 * 'level' contains the number of unique distances, excluding the
6125 * identity distance node_distance(i,i).
6127 * The sched_domains_nume_distance[] array includes the actual distance
6132 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6133 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6134 * the array will contain less then 'level' members. This could be
6135 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6136 * in other functions.
6138 * We reset it to 'level' at the end of this function.
6140 sched_domains_numa_levels = 0;
6142 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6143 if (!sched_domains_numa_masks)
6147 * Now for each level, construct a mask per node which contains all
6148 * cpus of nodes that are that many hops away from us.
6150 for (i = 0; i < level; i++) {
6151 sched_domains_numa_masks[i] =
6152 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6153 if (!sched_domains_numa_masks[i])
6156 for (j = 0; j < nr_node_ids; j++) {
6157 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6161 sched_domains_numa_masks[i][j] = mask;
6163 for (k = 0; k < nr_node_ids; k++) {
6164 if (node_distance(j, k) > sched_domains_numa_distance[i])
6167 cpumask_or(mask, mask, cpumask_of_node(k));
6172 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6173 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6178 * Copy the default topology bits..
6180 for (i = 0; default_topology[i].init; i++)
6181 tl[i] = default_topology[i];
6184 * .. and append 'j' levels of NUMA goodness.
6186 for (j = 0; j < level; i++, j++) {
6187 tl[i] = (struct sched_domain_topology_level){
6188 .init = sd_numa_init,
6189 .mask = sd_numa_mask,
6190 .flags = SDTL_OVERLAP,
6195 sched_domain_topology = tl;
6197 sched_domains_numa_levels = level;
6200 static void sched_domains_numa_masks_set(int cpu)
6203 int node = cpu_to_node(cpu);
6205 for (i = 0; i < sched_domains_numa_levels; i++) {
6206 for (j = 0; j < nr_node_ids; j++) {
6207 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6208 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6213 static void sched_domains_numa_masks_clear(int cpu)
6216 for (i = 0; i < sched_domains_numa_levels; i++) {
6217 for (j = 0; j < nr_node_ids; j++)
6218 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6223 * Update sched_domains_numa_masks[level][node] array when new cpus
6226 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6227 unsigned long action,
6230 int cpu = (long)hcpu;
6232 switch (action & ~CPU_TASKS_FROZEN) {
6234 sched_domains_numa_masks_set(cpu);
6238 sched_domains_numa_masks_clear(cpu);
6248 static inline void sched_init_numa(void)
6252 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6253 unsigned long action,
6258 #endif /* CONFIG_NUMA */
6260 static int __sdt_alloc(const struct cpumask *cpu_map)
6262 struct sched_domain_topology_level *tl;
6265 for (tl = sched_domain_topology; tl->init; tl++) {
6266 struct sd_data *sdd = &tl->data;
6268 sdd->sd = alloc_percpu(struct sched_domain *);
6272 sdd->sg = alloc_percpu(struct sched_group *);
6276 sdd->sgp = alloc_percpu(struct sched_group_power *);
6280 for_each_cpu(j, cpu_map) {
6281 struct sched_domain *sd;
6282 struct sched_group *sg;
6283 struct sched_group_power *sgp;
6285 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6286 GFP_KERNEL, cpu_to_node(j));
6290 *per_cpu_ptr(sdd->sd, j) = sd;
6292 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6293 GFP_KERNEL, cpu_to_node(j));
6299 *per_cpu_ptr(sdd->sg, j) = sg;
6301 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6302 GFP_KERNEL, cpu_to_node(j));
6306 *per_cpu_ptr(sdd->sgp, j) = sgp;
6313 static void __sdt_free(const struct cpumask *cpu_map)
6315 struct sched_domain_topology_level *tl;
6318 for (tl = sched_domain_topology; tl->init; tl++) {
6319 struct sd_data *sdd = &tl->data;
6321 for_each_cpu(j, cpu_map) {
6322 struct sched_domain *sd;
6325 sd = *per_cpu_ptr(sdd->sd, j);
6326 if (sd && (sd->flags & SD_OVERLAP))
6327 free_sched_groups(sd->groups, 0);
6328 kfree(*per_cpu_ptr(sdd->sd, j));
6332 kfree(*per_cpu_ptr(sdd->sg, j));
6334 kfree(*per_cpu_ptr(sdd->sgp, j));
6336 free_percpu(sdd->sd);
6338 free_percpu(sdd->sg);
6340 free_percpu(sdd->sgp);
6345 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6346 struct s_data *d, const struct cpumask *cpu_map,
6347 struct sched_domain_attr *attr, struct sched_domain *child,
6350 struct sched_domain *sd = tl->init(tl, cpu);
6354 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6356 sd->level = child->level + 1;
6357 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6361 set_domain_attribute(sd, attr);
6367 * Build sched domains for a given set of cpus and attach the sched domains
6368 * to the individual cpus
6370 static int build_sched_domains(const struct cpumask *cpu_map,
6371 struct sched_domain_attr *attr)
6373 enum s_alloc alloc_state = sa_none;
6374 struct sched_domain *sd;
6376 int i, ret = -ENOMEM;
6378 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6379 if (alloc_state != sa_rootdomain)
6382 /* Set up domains for cpus specified by the cpu_map. */
6383 for_each_cpu(i, cpu_map) {
6384 struct sched_domain_topology_level *tl;
6387 for (tl = sched_domain_topology; tl->init; tl++) {
6388 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6389 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6390 sd->flags |= SD_OVERLAP;
6391 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6398 *per_cpu_ptr(d.sd, i) = sd;
6401 /* Build the groups for the domains */
6402 for_each_cpu(i, cpu_map) {
6403 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6404 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6405 if (sd->flags & SD_OVERLAP) {
6406 if (build_overlap_sched_groups(sd, i))
6409 if (build_sched_groups(sd, i))
6415 /* Calculate CPU power for physical packages and nodes */
6416 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6417 if (!cpumask_test_cpu(i, cpu_map))
6420 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6421 claim_allocations(i, sd);
6422 init_sched_groups_power(i, sd);
6426 /* Attach the domains */
6428 for_each_cpu(i, cpu_map) {
6429 sd = *per_cpu_ptr(d.sd, i);
6430 cpu_attach_domain(sd, d.rd, i);
6436 __free_domain_allocs(&d, alloc_state, cpu_map);
6440 static cpumask_var_t *doms_cur; /* current sched domains */
6441 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6442 static struct sched_domain_attr *dattr_cur;
6443 /* attribues of custom domains in 'doms_cur' */
6446 * Special case: If a kmalloc of a doms_cur partition (array of
6447 * cpumask) fails, then fallback to a single sched domain,
6448 * as determined by the single cpumask fallback_doms.
6450 static cpumask_var_t fallback_doms;
6453 * arch_update_cpu_topology lets virtualized architectures update the
6454 * cpu core maps. It is supposed to return 1 if the topology changed
6455 * or 0 if it stayed the same.
6457 int __attribute__((weak)) arch_update_cpu_topology(void)
6462 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6465 cpumask_var_t *doms;
6467 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6470 for (i = 0; i < ndoms; i++) {
6471 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6472 free_sched_domains(doms, i);
6479 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6482 for (i = 0; i < ndoms; i++)
6483 free_cpumask_var(doms[i]);
6488 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6489 * For now this just excludes isolated cpus, but could be used to
6490 * exclude other special cases in the future.
6492 static int init_sched_domains(const struct cpumask *cpu_map)
6496 arch_update_cpu_topology();
6498 doms_cur = alloc_sched_domains(ndoms_cur);
6500 doms_cur = &fallback_doms;
6501 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6502 err = build_sched_domains(doms_cur[0], NULL);
6503 register_sched_domain_sysctl();
6509 * Detach sched domains from a group of cpus specified in cpu_map
6510 * These cpus will now be attached to the NULL domain
6512 static void detach_destroy_domains(const struct cpumask *cpu_map)
6517 for_each_cpu(i, cpu_map)
6518 cpu_attach_domain(NULL, &def_root_domain, i);
6522 /* handle null as "default" */
6523 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6524 struct sched_domain_attr *new, int idx_new)
6526 struct sched_domain_attr tmp;
6533 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6534 new ? (new + idx_new) : &tmp,
6535 sizeof(struct sched_domain_attr));
6539 * Partition sched domains as specified by the 'ndoms_new'
6540 * cpumasks in the array doms_new[] of cpumasks. This compares
6541 * doms_new[] to the current sched domain partitioning, doms_cur[].
6542 * It destroys each deleted domain and builds each new domain.
6544 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6545 * The masks don't intersect (don't overlap.) We should setup one
6546 * sched domain for each mask. CPUs not in any of the cpumasks will
6547 * not be load balanced. If the same cpumask appears both in the
6548 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6551 * The passed in 'doms_new' should be allocated using
6552 * alloc_sched_domains. This routine takes ownership of it and will
6553 * free_sched_domains it when done with it. If the caller failed the
6554 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6555 * and partition_sched_domains() will fallback to the single partition
6556 * 'fallback_doms', it also forces the domains to be rebuilt.
6558 * If doms_new == NULL it will be replaced with cpu_online_mask.
6559 * ndoms_new == 0 is a special case for destroying existing domains,
6560 * and it will not create the default domain.
6562 * Call with hotplug lock held
6564 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6565 struct sched_domain_attr *dattr_new)
6570 mutex_lock(&sched_domains_mutex);
6572 /* always unregister in case we don't destroy any domains */
6573 unregister_sched_domain_sysctl();
6575 /* Let architecture update cpu core mappings. */
6576 new_topology = arch_update_cpu_topology();
6578 n = doms_new ? ndoms_new : 0;
6580 /* Destroy deleted domains */
6581 for (i = 0; i < ndoms_cur; i++) {
6582 for (j = 0; j < n && !new_topology; j++) {
6583 if (cpumask_equal(doms_cur[i], doms_new[j])
6584 && dattrs_equal(dattr_cur, i, dattr_new, j))
6587 /* no match - a current sched domain not in new doms_new[] */
6588 detach_destroy_domains(doms_cur[i]);
6593 if (doms_new == NULL) {
6595 doms_new = &fallback_doms;
6596 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6597 WARN_ON_ONCE(dattr_new);
6600 /* Build new domains */
6601 for (i = 0; i < ndoms_new; i++) {
6602 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6603 if (cpumask_equal(doms_new[i], doms_cur[j])
6604 && dattrs_equal(dattr_new, i, dattr_cur, j))
6607 /* no match - add a new doms_new */
6608 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6613 /* Remember the new sched domains */
6614 if (doms_cur != &fallback_doms)
6615 free_sched_domains(doms_cur, ndoms_cur);
6616 kfree(dattr_cur); /* kfree(NULL) is safe */
6617 doms_cur = doms_new;
6618 dattr_cur = dattr_new;
6619 ndoms_cur = ndoms_new;
6621 register_sched_domain_sysctl();
6623 mutex_unlock(&sched_domains_mutex);
6626 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6629 * Update cpusets according to cpu_active mask. If cpusets are
6630 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6631 * around partition_sched_domains().
6633 * If we come here as part of a suspend/resume, don't touch cpusets because we
6634 * want to restore it back to its original state upon resume anyway.
6636 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6640 case CPU_ONLINE_FROZEN:
6641 case CPU_DOWN_FAILED_FROZEN:
6644 * num_cpus_frozen tracks how many CPUs are involved in suspend
6645 * resume sequence. As long as this is not the last online
6646 * operation in the resume sequence, just build a single sched
6647 * domain, ignoring cpusets.
6650 if (likely(num_cpus_frozen)) {
6651 partition_sched_domains(1, NULL, NULL);
6656 * This is the last CPU online operation. So fall through and
6657 * restore the original sched domains by considering the
6658 * cpuset configurations.
6662 case CPU_DOWN_FAILED:
6663 cpuset_update_active_cpus(true);
6671 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6675 case CPU_DOWN_PREPARE:
6676 cpuset_update_active_cpus(false);
6678 case CPU_DOWN_PREPARE_FROZEN:
6680 partition_sched_domains(1, NULL, NULL);
6688 void __init sched_init_smp(void)
6690 cpumask_var_t non_isolated_cpus;
6692 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6693 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6698 mutex_lock(&sched_domains_mutex);
6699 init_sched_domains(cpu_active_mask);
6700 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6701 if (cpumask_empty(non_isolated_cpus))
6702 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6703 mutex_unlock(&sched_domains_mutex);
6706 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6707 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6708 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6710 /* RT runtime code needs to handle some hotplug events */
6711 hotcpu_notifier(update_runtime, 0);
6715 /* Move init over to a non-isolated CPU */
6716 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6718 sched_init_granularity();
6719 free_cpumask_var(non_isolated_cpus);
6721 init_sched_rt_class();
6724 void __init sched_init_smp(void)
6726 sched_init_granularity();
6728 #endif /* CONFIG_SMP */
6730 const_debug unsigned int sysctl_timer_migration = 1;
6732 int in_sched_functions(unsigned long addr)
6734 return in_lock_functions(addr) ||
6735 (addr >= (unsigned long)__sched_text_start
6736 && addr < (unsigned long)__sched_text_end);
6739 #ifdef CONFIG_CGROUP_SCHED
6740 struct task_group root_task_group;
6741 LIST_HEAD(task_groups);
6744 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6746 void __init sched_init(void)
6749 unsigned long alloc_size = 0, ptr;
6751 #ifdef CONFIG_FAIR_GROUP_SCHED
6752 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6754 #ifdef CONFIG_RT_GROUP_SCHED
6755 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6757 #ifdef CONFIG_CPUMASK_OFFSTACK
6758 alloc_size += num_possible_cpus() * cpumask_size();
6761 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6763 #ifdef CONFIG_FAIR_GROUP_SCHED
6764 root_task_group.se = (struct sched_entity **)ptr;
6765 ptr += nr_cpu_ids * sizeof(void **);
6767 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6768 ptr += nr_cpu_ids * sizeof(void **);
6770 #endif /* CONFIG_FAIR_GROUP_SCHED */
6771 #ifdef CONFIG_RT_GROUP_SCHED
6772 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6773 ptr += nr_cpu_ids * sizeof(void **);
6775 root_task_group.rt_rq = (struct rt_rq **)ptr;
6776 ptr += nr_cpu_ids * sizeof(void **);
6778 #endif /* CONFIG_RT_GROUP_SCHED */
6779 #ifdef CONFIG_CPUMASK_OFFSTACK
6780 for_each_possible_cpu(i) {
6781 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6782 ptr += cpumask_size();
6784 #endif /* CONFIG_CPUMASK_OFFSTACK */
6788 init_defrootdomain();
6791 init_rt_bandwidth(&def_rt_bandwidth,
6792 global_rt_period(), global_rt_runtime());
6794 #ifdef CONFIG_RT_GROUP_SCHED
6795 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6796 global_rt_period(), global_rt_runtime());
6797 #endif /* CONFIG_RT_GROUP_SCHED */
6799 #ifdef CONFIG_CGROUP_SCHED
6800 list_add(&root_task_group.list, &task_groups);
6801 INIT_LIST_HEAD(&root_task_group.children);
6802 INIT_LIST_HEAD(&root_task_group.siblings);
6803 autogroup_init(&init_task);
6805 #endif /* CONFIG_CGROUP_SCHED */
6807 #ifdef CONFIG_CGROUP_CPUACCT
6808 root_cpuacct.cpustat = &kernel_cpustat;
6809 root_cpuacct.cpuusage = alloc_percpu(u64);
6810 /* Too early, not expected to fail */
6811 BUG_ON(!root_cpuacct.cpuusage);
6813 for_each_possible_cpu(i) {
6817 raw_spin_lock_init(&rq->lock);
6819 rq->calc_load_active = 0;
6820 rq->calc_load_update = jiffies + LOAD_FREQ;
6821 init_cfs_rq(&rq->cfs);
6822 init_rt_rq(&rq->rt, rq);
6823 #ifdef CONFIG_FAIR_GROUP_SCHED
6824 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6825 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6827 * How much cpu bandwidth does root_task_group get?
6829 * In case of task-groups formed thr' the cgroup filesystem, it
6830 * gets 100% of the cpu resources in the system. This overall
6831 * system cpu resource is divided among the tasks of
6832 * root_task_group and its child task-groups in a fair manner,
6833 * based on each entity's (task or task-group's) weight
6834 * (se->load.weight).
6836 * In other words, if root_task_group has 10 tasks of weight
6837 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6838 * then A0's share of the cpu resource is:
6840 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6842 * We achieve this by letting root_task_group's tasks sit
6843 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6845 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6846 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6847 #endif /* CONFIG_FAIR_GROUP_SCHED */
6849 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6850 #ifdef CONFIG_RT_GROUP_SCHED
6851 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6852 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6855 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6856 rq->cpu_load[j] = 0;
6858 rq->last_load_update_tick = jiffies;
6863 rq->cpu_power = SCHED_POWER_SCALE;
6864 rq->post_schedule = 0;
6865 rq->active_balance = 0;
6866 rq->next_balance = jiffies;
6871 rq->avg_idle = 2*sysctl_sched_migration_cost;
6873 INIT_LIST_HEAD(&rq->cfs_tasks);
6875 rq_attach_root(rq, &def_root_domain);
6881 atomic_set(&rq->nr_iowait, 0);
6884 set_load_weight(&init_task);
6886 #ifdef CONFIG_PREEMPT_NOTIFIERS
6887 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6890 #ifdef CONFIG_RT_MUTEXES
6891 plist_head_init(&init_task.pi_waiters);
6895 * The boot idle thread does lazy MMU switching as well:
6897 atomic_inc(&init_mm.mm_count);
6898 enter_lazy_tlb(&init_mm, current);
6901 * Make us the idle thread. Technically, schedule() should not be
6902 * called from this thread, however somewhere below it might be,
6903 * but because we are the idle thread, we just pick up running again
6904 * when this runqueue becomes "idle".
6906 init_idle(current, smp_processor_id());
6908 calc_load_update = jiffies + LOAD_FREQ;
6911 * During early bootup we pretend to be a normal task:
6913 current->sched_class = &fair_sched_class;
6916 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6917 /* May be allocated at isolcpus cmdline parse time */
6918 if (cpu_isolated_map == NULL)
6919 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6920 idle_thread_set_boot_cpu();
6922 init_sched_fair_class();
6924 scheduler_running = 1;
6927 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6928 static inline int preempt_count_equals(int preempt_offset)
6930 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6932 return (nested == preempt_offset);
6935 void __might_sleep(const char *file, int line, int preempt_offset)
6937 static unsigned long prev_jiffy; /* ratelimiting */
6939 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6940 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6941 system_state != SYSTEM_RUNNING || oops_in_progress)
6943 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6945 prev_jiffy = jiffies;
6948 "BUG: sleeping function called from invalid context at %s:%d\n",
6951 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6952 in_atomic(), irqs_disabled(),
6953 current->pid, current->comm);
6955 debug_show_held_locks(current);
6956 if (irqs_disabled())
6957 print_irqtrace_events(current);
6960 EXPORT_SYMBOL(__might_sleep);
6963 #ifdef CONFIG_MAGIC_SYSRQ
6964 static void normalize_task(struct rq *rq, struct task_struct *p)
6966 const struct sched_class *prev_class = p->sched_class;
6967 int old_prio = p->prio;
6972 dequeue_task(rq, p, 0);
6973 __setscheduler(rq, p, SCHED_NORMAL, 0);
6975 enqueue_task(rq, p, 0);
6976 resched_task(rq->curr);
6979 check_class_changed(rq, p, prev_class, old_prio);
6982 void normalize_rt_tasks(void)
6984 struct task_struct *g, *p;
6985 unsigned long flags;
6988 read_lock_irqsave(&tasklist_lock, flags);
6989 do_each_thread(g, p) {
6991 * Only normalize user tasks:
6996 p->se.exec_start = 0;
6997 #ifdef CONFIG_SCHEDSTATS
6998 p->se.statistics.wait_start = 0;
6999 p->se.statistics.sleep_start = 0;
7000 p->se.statistics.block_start = 0;
7005 * Renice negative nice level userspace
7008 if (TASK_NICE(p) < 0 && p->mm)
7009 set_user_nice(p, 0);
7013 raw_spin_lock(&p->pi_lock);
7014 rq = __task_rq_lock(p);
7016 normalize_task(rq, p);
7018 __task_rq_unlock(rq);
7019 raw_spin_unlock(&p->pi_lock);
7020 } while_each_thread(g, p);
7022 read_unlock_irqrestore(&tasklist_lock, flags);
7025 #endif /* CONFIG_MAGIC_SYSRQ */
7027 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7029 * These functions are only useful for the IA64 MCA handling, or kdb.
7031 * They can only be called when the whole system has been
7032 * stopped - every CPU needs to be quiescent, and no scheduling
7033 * activity can take place. Using them for anything else would
7034 * be a serious bug, and as a result, they aren't even visible
7035 * under any other configuration.
7039 * curr_task - return the current task for a given cpu.
7040 * @cpu: the processor in question.
7042 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7044 struct task_struct *curr_task(int cpu)
7046 return cpu_curr(cpu);
7049 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7053 * set_curr_task - set the current task for a given cpu.
7054 * @cpu: the processor in question.
7055 * @p: the task pointer to set.
7057 * Description: This function must only be used when non-maskable interrupts
7058 * are serviced on a separate stack. It allows the architecture to switch the
7059 * notion of the current task on a cpu in a non-blocking manner. This function
7060 * must be called with all CPU's synchronized, and interrupts disabled, the
7061 * and caller must save the original value of the current task (see
7062 * curr_task() above) and restore that value before reenabling interrupts and
7063 * re-starting the system.
7065 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7067 void set_curr_task(int cpu, struct task_struct *p)
7074 #ifdef CONFIG_CGROUP_SCHED
7075 /* task_group_lock serializes the addition/removal of task groups */
7076 static DEFINE_SPINLOCK(task_group_lock);
7078 static void free_sched_group(struct task_group *tg)
7080 free_fair_sched_group(tg);
7081 free_rt_sched_group(tg);
7086 /* allocate runqueue etc for a new task group */
7087 struct task_group *sched_create_group(struct task_group *parent)
7089 struct task_group *tg;
7090 unsigned long flags;
7092 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7094 return ERR_PTR(-ENOMEM);
7096 if (!alloc_fair_sched_group(tg, parent))
7099 if (!alloc_rt_sched_group(tg, parent))
7102 spin_lock_irqsave(&task_group_lock, flags);
7103 list_add_rcu(&tg->list, &task_groups);
7105 WARN_ON(!parent); /* root should already exist */
7107 tg->parent = parent;
7108 INIT_LIST_HEAD(&tg->children);
7109 list_add_rcu(&tg->siblings, &parent->children);
7110 spin_unlock_irqrestore(&task_group_lock, flags);
7115 free_sched_group(tg);
7116 return ERR_PTR(-ENOMEM);
7119 /* rcu callback to free various structures associated with a task group */
7120 static void free_sched_group_rcu(struct rcu_head *rhp)
7122 /* now it should be safe to free those cfs_rqs */
7123 free_sched_group(container_of(rhp, struct task_group, rcu));
7126 /* Destroy runqueue etc associated with a task group */
7127 void sched_destroy_group(struct task_group *tg)
7129 unsigned long flags;
7132 /* end participation in shares distribution */
7133 for_each_possible_cpu(i)
7134 unregister_fair_sched_group(tg, i);
7136 spin_lock_irqsave(&task_group_lock, flags);
7137 list_del_rcu(&tg->list);
7138 list_del_rcu(&tg->siblings);
7139 spin_unlock_irqrestore(&task_group_lock, flags);
7141 /* wait for possible concurrent references to cfs_rqs complete */
7142 call_rcu(&tg->rcu, free_sched_group_rcu);
7145 /* change task's runqueue when it moves between groups.
7146 * The caller of this function should have put the task in its new group
7147 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7148 * reflect its new group.
7150 void sched_move_task(struct task_struct *tsk)
7152 struct task_group *tg;
7154 unsigned long flags;
7157 rq = task_rq_lock(tsk, &flags);
7159 running = task_current(rq, tsk);
7163 dequeue_task(rq, tsk, 0);
7164 if (unlikely(running))
7165 tsk->sched_class->put_prev_task(rq, tsk);
7167 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7168 lockdep_is_held(&tsk->sighand->siglock)),
7169 struct task_group, css);
7170 tg = autogroup_task_group(tsk, tg);
7171 tsk->sched_task_group = tg;
7173 #ifdef CONFIG_FAIR_GROUP_SCHED
7174 if (tsk->sched_class->task_move_group)
7175 tsk->sched_class->task_move_group(tsk, on_rq);
7178 set_task_rq(tsk, task_cpu(tsk));
7180 if (unlikely(running))
7181 tsk->sched_class->set_curr_task(rq);
7183 enqueue_task(rq, tsk, 0);
7185 task_rq_unlock(rq, tsk, &flags);
7187 #endif /* CONFIG_CGROUP_SCHED */
7189 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7190 static unsigned long to_ratio(u64 period, u64 runtime)
7192 if (runtime == RUNTIME_INF)
7195 return div64_u64(runtime << 20, period);
7199 #ifdef CONFIG_RT_GROUP_SCHED
7201 * Ensure that the real time constraints are schedulable.
7203 static DEFINE_MUTEX(rt_constraints_mutex);
7205 /* Must be called with tasklist_lock held */
7206 static inline int tg_has_rt_tasks(struct task_group *tg)
7208 struct task_struct *g, *p;
7210 do_each_thread(g, p) {
7211 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7213 } while_each_thread(g, p);
7218 struct rt_schedulable_data {
7219 struct task_group *tg;
7224 static int tg_rt_schedulable(struct task_group *tg, void *data)
7226 struct rt_schedulable_data *d = data;
7227 struct task_group *child;
7228 unsigned long total, sum = 0;
7229 u64 period, runtime;
7231 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7232 runtime = tg->rt_bandwidth.rt_runtime;
7235 period = d->rt_period;
7236 runtime = d->rt_runtime;
7240 * Cannot have more runtime than the period.
7242 if (runtime > period && runtime != RUNTIME_INF)
7246 * Ensure we don't starve existing RT tasks.
7248 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7251 total = to_ratio(period, runtime);
7254 * Nobody can have more than the global setting allows.
7256 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7260 * The sum of our children's runtime should not exceed our own.
7262 list_for_each_entry_rcu(child, &tg->children, siblings) {
7263 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7264 runtime = child->rt_bandwidth.rt_runtime;
7266 if (child == d->tg) {
7267 period = d->rt_period;
7268 runtime = d->rt_runtime;
7271 sum += to_ratio(period, runtime);
7280 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7284 struct rt_schedulable_data data = {
7286 .rt_period = period,
7287 .rt_runtime = runtime,
7291 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7297 static int tg_set_rt_bandwidth(struct task_group *tg,
7298 u64 rt_period, u64 rt_runtime)
7302 mutex_lock(&rt_constraints_mutex);
7303 read_lock(&tasklist_lock);
7304 err = __rt_schedulable(tg, rt_period, rt_runtime);
7308 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7309 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7310 tg->rt_bandwidth.rt_runtime = rt_runtime;
7312 for_each_possible_cpu(i) {
7313 struct rt_rq *rt_rq = tg->rt_rq[i];
7315 raw_spin_lock(&rt_rq->rt_runtime_lock);
7316 rt_rq->rt_runtime = rt_runtime;
7317 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7319 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7321 read_unlock(&tasklist_lock);
7322 mutex_unlock(&rt_constraints_mutex);
7327 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7329 u64 rt_runtime, rt_period;
7331 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7332 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7333 if (rt_runtime_us < 0)
7334 rt_runtime = RUNTIME_INF;
7336 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7339 long sched_group_rt_runtime(struct task_group *tg)
7343 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7346 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7347 do_div(rt_runtime_us, NSEC_PER_USEC);
7348 return rt_runtime_us;
7351 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7353 u64 rt_runtime, rt_period;
7355 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7356 rt_runtime = tg->rt_bandwidth.rt_runtime;
7361 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7364 long sched_group_rt_period(struct task_group *tg)
7368 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7369 do_div(rt_period_us, NSEC_PER_USEC);
7370 return rt_period_us;
7373 static int sched_rt_global_constraints(void)
7375 u64 runtime, period;
7378 if (sysctl_sched_rt_period <= 0)
7381 runtime = global_rt_runtime();
7382 period = global_rt_period();
7385 * Sanity check on the sysctl variables.
7387 if (runtime > period && runtime != RUNTIME_INF)
7390 mutex_lock(&rt_constraints_mutex);
7391 read_lock(&tasklist_lock);
7392 ret = __rt_schedulable(NULL, 0, 0);
7393 read_unlock(&tasklist_lock);
7394 mutex_unlock(&rt_constraints_mutex);
7399 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7401 /* Don't accept realtime tasks when there is no way for them to run */
7402 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7408 #else /* !CONFIG_RT_GROUP_SCHED */
7409 static int sched_rt_global_constraints(void)
7411 unsigned long flags;
7414 if (sysctl_sched_rt_period <= 0)
7418 * There's always some RT tasks in the root group
7419 * -- migration, kstopmachine etc..
7421 if (sysctl_sched_rt_runtime == 0)
7424 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7425 for_each_possible_cpu(i) {
7426 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7428 raw_spin_lock(&rt_rq->rt_runtime_lock);
7429 rt_rq->rt_runtime = global_rt_runtime();
7430 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7432 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7436 #endif /* CONFIG_RT_GROUP_SCHED */
7438 int sched_rt_handler(struct ctl_table *table, int write,
7439 void __user *buffer, size_t *lenp,
7443 int old_period, old_runtime;
7444 static DEFINE_MUTEX(mutex);
7447 old_period = sysctl_sched_rt_period;
7448 old_runtime = sysctl_sched_rt_runtime;
7450 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7452 if (!ret && write) {
7453 ret = sched_rt_global_constraints();
7455 sysctl_sched_rt_period = old_period;
7456 sysctl_sched_rt_runtime = old_runtime;
7458 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7459 def_rt_bandwidth.rt_period =
7460 ns_to_ktime(global_rt_period());
7463 mutex_unlock(&mutex);
7468 #ifdef CONFIG_CGROUP_SCHED
7470 /* return corresponding task_group object of a cgroup */
7471 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7473 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7474 struct task_group, css);
7477 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7479 struct task_group *tg, *parent;
7481 if (!cgrp->parent) {
7482 /* This is early initialization for the top cgroup */
7483 return &root_task_group.css;
7486 parent = cgroup_tg(cgrp->parent);
7487 tg = sched_create_group(parent);
7489 return ERR_PTR(-ENOMEM);
7494 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7496 struct task_group *tg = cgroup_tg(cgrp);
7498 sched_destroy_group(tg);
7501 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7502 struct cgroup_taskset *tset)
7504 struct task_struct *task;
7506 cgroup_taskset_for_each(task, cgrp, tset) {
7507 #ifdef CONFIG_RT_GROUP_SCHED
7508 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7511 /* We don't support RT-tasks being in separate groups */
7512 if (task->sched_class != &fair_sched_class)
7519 static void cpu_cgroup_attach(struct cgroup *cgrp,
7520 struct cgroup_taskset *tset)
7522 struct task_struct *task;
7524 cgroup_taskset_for_each(task, cgrp, tset)
7525 sched_move_task(task);
7529 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7530 struct task_struct *task)
7533 * cgroup_exit() is called in the copy_process() failure path.
7534 * Ignore this case since the task hasn't ran yet, this avoids
7535 * trying to poke a half freed task state from generic code.
7537 if (!(task->flags & PF_EXITING))
7540 sched_move_task(task);
7543 #ifdef CONFIG_FAIR_GROUP_SCHED
7544 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7547 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7550 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7552 struct task_group *tg = cgroup_tg(cgrp);
7554 return (u64) scale_load_down(tg->shares);
7557 #ifdef CONFIG_CFS_BANDWIDTH
7558 static DEFINE_MUTEX(cfs_constraints_mutex);
7560 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7561 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7563 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7565 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7567 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7568 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7570 if (tg == &root_task_group)
7574 * Ensure we have at some amount of bandwidth every period. This is
7575 * to prevent reaching a state of large arrears when throttled via
7576 * entity_tick() resulting in prolonged exit starvation.
7578 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7582 * Likewise, bound things on the otherside by preventing insane quota
7583 * periods. This also allows us to normalize in computing quota
7586 if (period > max_cfs_quota_period)
7589 mutex_lock(&cfs_constraints_mutex);
7590 ret = __cfs_schedulable(tg, period, quota);
7594 runtime_enabled = quota != RUNTIME_INF;
7595 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7596 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7597 raw_spin_lock_irq(&cfs_b->lock);
7598 cfs_b->period = ns_to_ktime(period);
7599 cfs_b->quota = quota;
7601 __refill_cfs_bandwidth_runtime(cfs_b);
7602 /* restart the period timer (if active) to handle new period expiry */
7603 if (runtime_enabled && cfs_b->timer_active) {
7604 /* force a reprogram */
7605 cfs_b->timer_active = 0;
7606 __start_cfs_bandwidth(cfs_b);
7608 raw_spin_unlock_irq(&cfs_b->lock);
7610 for_each_possible_cpu(i) {
7611 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7612 struct rq *rq = cfs_rq->rq;
7614 raw_spin_lock_irq(&rq->lock);
7615 cfs_rq->runtime_enabled = runtime_enabled;
7616 cfs_rq->runtime_remaining = 0;
7618 if (cfs_rq->throttled)
7619 unthrottle_cfs_rq(cfs_rq);
7620 raw_spin_unlock_irq(&rq->lock);
7623 mutex_unlock(&cfs_constraints_mutex);
7628 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7632 period = ktime_to_ns(tg->cfs_bandwidth.period);
7633 if (cfs_quota_us < 0)
7634 quota = RUNTIME_INF;
7636 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7638 return tg_set_cfs_bandwidth(tg, period, quota);
7641 long tg_get_cfs_quota(struct task_group *tg)
7645 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7648 quota_us = tg->cfs_bandwidth.quota;
7649 do_div(quota_us, NSEC_PER_USEC);
7654 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7658 period = (u64)cfs_period_us * NSEC_PER_USEC;
7659 quota = tg->cfs_bandwidth.quota;
7661 return tg_set_cfs_bandwidth(tg, period, quota);
7664 long tg_get_cfs_period(struct task_group *tg)
7668 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7669 do_div(cfs_period_us, NSEC_PER_USEC);
7671 return cfs_period_us;
7674 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7676 return tg_get_cfs_quota(cgroup_tg(cgrp));
7679 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7682 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7685 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7687 return tg_get_cfs_period(cgroup_tg(cgrp));
7690 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7693 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7696 struct cfs_schedulable_data {
7697 struct task_group *tg;
7702 * normalize group quota/period to be quota/max_period
7703 * note: units are usecs
7705 static u64 normalize_cfs_quota(struct task_group *tg,
7706 struct cfs_schedulable_data *d)
7714 period = tg_get_cfs_period(tg);
7715 quota = tg_get_cfs_quota(tg);
7718 /* note: these should typically be equivalent */
7719 if (quota == RUNTIME_INF || quota == -1)
7722 return to_ratio(period, quota);
7725 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7727 struct cfs_schedulable_data *d = data;
7728 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7729 s64 quota = 0, parent_quota = -1;
7732 quota = RUNTIME_INF;
7734 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7736 quota = normalize_cfs_quota(tg, d);
7737 parent_quota = parent_b->hierarchal_quota;
7740 * ensure max(child_quota) <= parent_quota, inherit when no
7743 if (quota == RUNTIME_INF)
7744 quota = parent_quota;
7745 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7748 cfs_b->hierarchal_quota = quota;
7753 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7756 struct cfs_schedulable_data data = {
7762 if (quota != RUNTIME_INF) {
7763 do_div(data.period, NSEC_PER_USEC);
7764 do_div(data.quota, NSEC_PER_USEC);
7768 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7774 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7775 struct cgroup_map_cb *cb)
7777 struct task_group *tg = cgroup_tg(cgrp);
7778 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7780 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7781 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7782 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7786 #endif /* CONFIG_CFS_BANDWIDTH */
7787 #endif /* CONFIG_FAIR_GROUP_SCHED */
7789 #ifdef CONFIG_RT_GROUP_SCHED
7790 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7793 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7796 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7798 return sched_group_rt_runtime(cgroup_tg(cgrp));
7801 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7804 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7807 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7809 return sched_group_rt_period(cgroup_tg(cgrp));
7811 #endif /* CONFIG_RT_GROUP_SCHED */
7813 static struct cftype cpu_files[] = {
7814 #ifdef CONFIG_FAIR_GROUP_SCHED
7817 .read_u64 = cpu_shares_read_u64,
7818 .write_u64 = cpu_shares_write_u64,
7821 #ifdef CONFIG_CFS_BANDWIDTH
7823 .name = "cfs_quota_us",
7824 .read_s64 = cpu_cfs_quota_read_s64,
7825 .write_s64 = cpu_cfs_quota_write_s64,
7828 .name = "cfs_period_us",
7829 .read_u64 = cpu_cfs_period_read_u64,
7830 .write_u64 = cpu_cfs_period_write_u64,
7834 .read_map = cpu_stats_show,
7837 #ifdef CONFIG_RT_GROUP_SCHED
7839 .name = "rt_runtime_us",
7840 .read_s64 = cpu_rt_runtime_read,
7841 .write_s64 = cpu_rt_runtime_write,
7844 .name = "rt_period_us",
7845 .read_u64 = cpu_rt_period_read_uint,
7846 .write_u64 = cpu_rt_period_write_uint,
7852 struct cgroup_subsys cpu_cgroup_subsys = {
7854 .create = cpu_cgroup_create,
7855 .destroy = cpu_cgroup_destroy,
7856 .can_attach = cpu_cgroup_can_attach,
7857 .attach = cpu_cgroup_attach,
7858 .exit = cpu_cgroup_exit,
7859 .subsys_id = cpu_cgroup_subsys_id,
7860 .base_cftypes = cpu_files,
7864 #endif /* CONFIG_CGROUP_SCHED */
7866 #ifdef CONFIG_CGROUP_CPUACCT
7869 * CPU accounting code for task groups.
7871 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7872 * (balbir@in.ibm.com).
7875 struct cpuacct root_cpuacct;
7877 /* create a new cpu accounting group */
7878 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
7883 return &root_cpuacct.css;
7885 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7889 ca->cpuusage = alloc_percpu(u64);
7893 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7895 goto out_free_cpuusage;
7900 free_percpu(ca->cpuusage);
7904 return ERR_PTR(-ENOMEM);
7907 /* destroy an existing cpu accounting group */
7908 static void cpuacct_destroy(struct cgroup *cgrp)
7910 struct cpuacct *ca = cgroup_ca(cgrp);
7912 free_percpu(ca->cpustat);
7913 free_percpu(ca->cpuusage);
7917 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7919 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7922 #ifndef CONFIG_64BIT
7924 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7926 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7928 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7936 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7938 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7940 #ifndef CONFIG_64BIT
7942 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7944 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7946 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7952 /* return total cpu usage (in nanoseconds) of a group */
7953 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7955 struct cpuacct *ca = cgroup_ca(cgrp);
7956 u64 totalcpuusage = 0;
7959 for_each_present_cpu(i)
7960 totalcpuusage += cpuacct_cpuusage_read(ca, i);
7962 return totalcpuusage;
7965 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
7968 struct cpuacct *ca = cgroup_ca(cgrp);
7977 for_each_present_cpu(i)
7978 cpuacct_cpuusage_write(ca, i, 0);
7984 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
7987 struct cpuacct *ca = cgroup_ca(cgroup);
7991 for_each_present_cpu(i) {
7992 percpu = cpuacct_cpuusage_read(ca, i);
7993 seq_printf(m, "%llu ", (unsigned long long) percpu);
7995 seq_printf(m, "\n");
7999 static const char *cpuacct_stat_desc[] = {
8000 [CPUACCT_STAT_USER] = "user",
8001 [CPUACCT_STAT_SYSTEM] = "system",
8004 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8005 struct cgroup_map_cb *cb)
8007 struct cpuacct *ca = cgroup_ca(cgrp);
8011 for_each_online_cpu(cpu) {
8012 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8013 val += kcpustat->cpustat[CPUTIME_USER];
8014 val += kcpustat->cpustat[CPUTIME_NICE];
8016 val = cputime64_to_clock_t(val);
8017 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8020 for_each_online_cpu(cpu) {
8021 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8022 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8023 val += kcpustat->cpustat[CPUTIME_IRQ];
8024 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8027 val = cputime64_to_clock_t(val);
8028 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8033 static struct cftype files[] = {
8036 .read_u64 = cpuusage_read,
8037 .write_u64 = cpuusage_write,
8040 .name = "usage_percpu",
8041 .read_seq_string = cpuacct_percpu_seq_read,
8045 .read_map = cpuacct_stats_show,
8051 * charge this task's execution time to its accounting group.
8053 * called with rq->lock held.
8055 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8060 if (unlikely(!cpuacct_subsys.active))
8063 cpu = task_cpu(tsk);
8069 for (; ca; ca = parent_ca(ca)) {
8070 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8071 *cpuusage += cputime;
8077 struct cgroup_subsys cpuacct_subsys = {
8079 .create = cpuacct_create,
8080 .destroy = cpuacct_destroy,
8081 .subsys_id = cpuacct_subsys_id,
8082 .base_cftypes = files,
8084 #endif /* CONFIG_CGROUP_CPUACCT */