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 p->se.nr_migrations++;
956 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
959 __set_task_cpu(p, new_cpu);
962 struct migration_arg {
963 struct task_struct *task;
967 static int migration_cpu_stop(void *data);
970 * wait_task_inactive - wait for a thread to unschedule.
972 * If @match_state is nonzero, it's the @p->state value just checked and
973 * not expected to change. If it changes, i.e. @p might have woken up,
974 * then return zero. When we succeed in waiting for @p to be off its CPU,
975 * we return a positive number (its total switch count). If a second call
976 * a short while later returns the same number, the caller can be sure that
977 * @p has remained unscheduled the whole time.
979 * The caller must ensure that the task *will* unschedule sometime soon,
980 * else this function might spin for a *long* time. This function can't
981 * be called with interrupts off, or it may introduce deadlock with
982 * smp_call_function() if an IPI is sent by the same process we are
983 * waiting to become inactive.
985 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
994 * We do the initial early heuristics without holding
995 * any task-queue locks at all. We'll only try to get
996 * the runqueue lock when things look like they will
1002 * If the task is actively running on another CPU
1003 * still, just relax and busy-wait without holding
1006 * NOTE! Since we don't hold any locks, it's not
1007 * even sure that "rq" stays as the right runqueue!
1008 * But we don't care, since "task_running()" will
1009 * return false if the runqueue has changed and p
1010 * is actually now running somewhere else!
1012 while (task_running(rq, p)) {
1013 if (match_state && unlikely(p->state != match_state))
1019 * Ok, time to look more closely! We need the rq
1020 * lock now, to be *sure*. If we're wrong, we'll
1021 * just go back and repeat.
1023 rq = task_rq_lock(p, &flags);
1024 trace_sched_wait_task(p);
1025 running = task_running(rq, p);
1028 if (!match_state || p->state == match_state)
1029 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1030 task_rq_unlock(rq, p, &flags);
1033 * If it changed from the expected state, bail out now.
1035 if (unlikely(!ncsw))
1039 * Was it really running after all now that we
1040 * checked with the proper locks actually held?
1042 * Oops. Go back and try again..
1044 if (unlikely(running)) {
1050 * It's not enough that it's not actively running,
1051 * it must be off the runqueue _entirely_, and not
1054 * So if it was still runnable (but just not actively
1055 * running right now), it's preempted, and we should
1056 * yield - it could be a while.
1058 if (unlikely(on_rq)) {
1059 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1061 set_current_state(TASK_UNINTERRUPTIBLE);
1062 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1067 * Ahh, all good. It wasn't running, and it wasn't
1068 * runnable, which means that it will never become
1069 * running in the future either. We're all done!
1078 * kick_process - kick a running thread to enter/exit the kernel
1079 * @p: the to-be-kicked thread
1081 * Cause a process which is running on another CPU to enter
1082 * kernel-mode, without any delay. (to get signals handled.)
1084 * NOTE: this function doesn't have to take the runqueue lock,
1085 * because all it wants to ensure is that the remote task enters
1086 * the kernel. If the IPI races and the task has been migrated
1087 * to another CPU then no harm is done and the purpose has been
1090 void kick_process(struct task_struct *p)
1096 if ((cpu != smp_processor_id()) && task_curr(p))
1097 smp_send_reschedule(cpu);
1100 EXPORT_SYMBOL_GPL(kick_process);
1101 #endif /* CONFIG_SMP */
1105 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1107 static int select_fallback_rq(int cpu, struct task_struct *p)
1109 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1110 enum { cpuset, possible, fail } state = cpuset;
1113 /* Look for allowed, online CPU in same node. */
1114 for_each_cpu(dest_cpu, nodemask) {
1115 if (!cpu_online(dest_cpu))
1117 if (!cpu_active(dest_cpu))
1119 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1124 /* Any allowed, online CPU? */
1125 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1126 if (!cpu_online(dest_cpu))
1128 if (!cpu_active(dest_cpu))
1135 /* No more Mr. Nice Guy. */
1136 cpuset_cpus_allowed_fallback(p);
1141 do_set_cpus_allowed(p, cpu_possible_mask);
1152 if (state != cpuset) {
1154 * Don't tell them about moving exiting tasks or
1155 * kernel threads (both mm NULL), since they never
1158 if (p->mm && printk_ratelimit()) {
1159 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1160 task_pid_nr(p), p->comm, cpu);
1168 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1171 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1173 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1176 * In order not to call set_task_cpu() on a blocking task we need
1177 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1180 * Since this is common to all placement strategies, this lives here.
1182 * [ this allows ->select_task() to simply return task_cpu(p) and
1183 * not worry about this generic constraint ]
1185 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1187 cpu = select_fallback_rq(task_cpu(p), p);
1192 static void update_avg(u64 *avg, u64 sample)
1194 s64 diff = sample - *avg;
1200 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1202 #ifdef CONFIG_SCHEDSTATS
1203 struct rq *rq = this_rq();
1206 int this_cpu = smp_processor_id();
1208 if (cpu == this_cpu) {
1209 schedstat_inc(rq, ttwu_local);
1210 schedstat_inc(p, se.statistics.nr_wakeups_local);
1212 struct sched_domain *sd;
1214 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1216 for_each_domain(this_cpu, sd) {
1217 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1218 schedstat_inc(sd, ttwu_wake_remote);
1225 if (wake_flags & WF_MIGRATED)
1226 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1228 #endif /* CONFIG_SMP */
1230 schedstat_inc(rq, ttwu_count);
1231 schedstat_inc(p, se.statistics.nr_wakeups);
1233 if (wake_flags & WF_SYNC)
1234 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1236 #endif /* CONFIG_SCHEDSTATS */
1239 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1241 activate_task(rq, p, en_flags);
1244 /* if a worker is waking up, notify workqueue */
1245 if (p->flags & PF_WQ_WORKER)
1246 wq_worker_waking_up(p, cpu_of(rq));
1250 * Mark the task runnable and perform wakeup-preemption.
1253 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1255 trace_sched_wakeup(p, true);
1256 check_preempt_curr(rq, p, wake_flags);
1258 p->state = TASK_RUNNING;
1260 if (p->sched_class->task_woken)
1261 p->sched_class->task_woken(rq, p);
1263 if (rq->idle_stamp) {
1264 u64 delta = rq->clock - rq->idle_stamp;
1265 u64 max = 2*sysctl_sched_migration_cost;
1270 update_avg(&rq->avg_idle, delta);
1277 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1280 if (p->sched_contributes_to_load)
1281 rq->nr_uninterruptible--;
1284 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1285 ttwu_do_wakeup(rq, p, wake_flags);
1289 * Called in case the task @p isn't fully descheduled from its runqueue,
1290 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1291 * since all we need to do is flip p->state to TASK_RUNNING, since
1292 * the task is still ->on_rq.
1294 static int ttwu_remote(struct task_struct *p, int wake_flags)
1299 rq = __task_rq_lock(p);
1301 ttwu_do_wakeup(rq, p, wake_flags);
1304 __task_rq_unlock(rq);
1310 static void sched_ttwu_pending(void)
1312 struct rq *rq = this_rq();
1313 struct llist_node *llist = llist_del_all(&rq->wake_list);
1314 struct task_struct *p;
1316 raw_spin_lock(&rq->lock);
1319 p = llist_entry(llist, struct task_struct, wake_entry);
1320 llist = llist_next(llist);
1321 ttwu_do_activate(rq, p, 0);
1324 raw_spin_unlock(&rq->lock);
1327 void scheduler_ipi(void)
1329 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1333 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1334 * traditionally all their work was done from the interrupt return
1335 * path. Now that we actually do some work, we need to make sure
1338 * Some archs already do call them, luckily irq_enter/exit nest
1341 * Arguably we should visit all archs and update all handlers,
1342 * however a fair share of IPIs are still resched only so this would
1343 * somewhat pessimize the simple resched case.
1346 sched_ttwu_pending();
1349 * Check if someone kicked us for doing the nohz idle load balance.
1351 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1352 this_rq()->idle_balance = 1;
1353 raise_softirq_irqoff(SCHED_SOFTIRQ);
1358 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1360 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1361 smp_send_reschedule(cpu);
1364 bool cpus_share_cache(int this_cpu, int that_cpu)
1366 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1368 #endif /* CONFIG_SMP */
1370 static void ttwu_queue(struct task_struct *p, int cpu)
1372 struct rq *rq = cpu_rq(cpu);
1374 #if defined(CONFIG_SMP)
1375 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1376 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1377 ttwu_queue_remote(p, cpu);
1382 raw_spin_lock(&rq->lock);
1383 ttwu_do_activate(rq, p, 0);
1384 raw_spin_unlock(&rq->lock);
1388 * try_to_wake_up - wake up a thread
1389 * @p: the thread to be awakened
1390 * @state: the mask of task states that can be woken
1391 * @wake_flags: wake modifier flags (WF_*)
1393 * Put it on the run-queue if it's not already there. The "current"
1394 * thread is always on the run-queue (except when the actual
1395 * re-schedule is in progress), and as such you're allowed to do
1396 * the simpler "current->state = TASK_RUNNING" to mark yourself
1397 * runnable without the overhead of this.
1399 * Returns %true if @p was woken up, %false if it was already running
1400 * or @state didn't match @p's state.
1403 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1405 unsigned long flags;
1406 int cpu, success = 0;
1409 raw_spin_lock_irqsave(&p->pi_lock, flags);
1410 if (!(p->state & state))
1413 success = 1; /* we're going to change ->state */
1416 if (p->on_rq && ttwu_remote(p, wake_flags))
1421 * If the owning (remote) cpu is still in the middle of schedule() with
1422 * this task as prev, wait until its done referencing the task.
1427 * Pairs with the smp_wmb() in finish_lock_switch().
1431 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1432 p->state = TASK_WAKING;
1434 if (p->sched_class->task_waking)
1435 p->sched_class->task_waking(p);
1437 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1438 if (task_cpu(p) != cpu) {
1439 wake_flags |= WF_MIGRATED;
1440 set_task_cpu(p, cpu);
1442 #endif /* CONFIG_SMP */
1446 ttwu_stat(p, cpu, wake_flags);
1448 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1454 * try_to_wake_up_local - try to wake up a local task with rq lock held
1455 * @p: the thread to be awakened
1457 * Put @p on the run-queue if it's not already there. The caller must
1458 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1461 static void try_to_wake_up_local(struct task_struct *p)
1463 struct rq *rq = task_rq(p);
1465 BUG_ON(rq != this_rq());
1466 BUG_ON(p == current);
1467 lockdep_assert_held(&rq->lock);
1469 if (!raw_spin_trylock(&p->pi_lock)) {
1470 raw_spin_unlock(&rq->lock);
1471 raw_spin_lock(&p->pi_lock);
1472 raw_spin_lock(&rq->lock);
1475 if (!(p->state & TASK_NORMAL))
1479 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1481 ttwu_do_wakeup(rq, p, 0);
1482 ttwu_stat(p, smp_processor_id(), 0);
1484 raw_spin_unlock(&p->pi_lock);
1488 * wake_up_process - Wake up a specific process
1489 * @p: The process to be woken up.
1491 * Attempt to wake up the nominated process and move it to the set of runnable
1492 * processes. Returns 1 if the process was woken up, 0 if it was already
1495 * It may be assumed that this function implies a write memory barrier before
1496 * changing the task state if and only if any tasks are woken up.
1498 int wake_up_process(struct task_struct *p)
1500 return try_to_wake_up(p, TASK_ALL, 0);
1502 EXPORT_SYMBOL(wake_up_process);
1504 int wake_up_state(struct task_struct *p, unsigned int state)
1506 return try_to_wake_up(p, state, 0);
1510 * Perform scheduler related setup for a newly forked process p.
1511 * p is forked by current.
1513 * __sched_fork() is basic setup used by init_idle() too:
1515 static void __sched_fork(struct task_struct *p)
1520 p->se.exec_start = 0;
1521 p->se.sum_exec_runtime = 0;
1522 p->se.prev_sum_exec_runtime = 0;
1523 p->se.nr_migrations = 0;
1525 INIT_LIST_HEAD(&p->se.group_node);
1527 #ifdef CONFIG_SCHEDSTATS
1528 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1531 INIT_LIST_HEAD(&p->rt.run_list);
1533 #ifdef CONFIG_PREEMPT_NOTIFIERS
1534 INIT_HLIST_HEAD(&p->preempt_notifiers);
1537 #ifdef CONFIG_NUMA_BALANCING
1538 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1539 p->mm->numa_next_scan = jiffies;
1540 p->mm->numa_next_reset = jiffies;
1541 p->mm->numa_scan_seq = 0;
1544 p->node_stamp = 0ULL;
1545 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1546 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1547 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1548 p->numa_work.next = &p->numa_work;
1549 #endif /* CONFIG_NUMA_BALANCING */
1553 * fork()/clone()-time setup:
1555 void sched_fork(struct task_struct *p)
1557 unsigned long flags;
1558 int cpu = get_cpu();
1562 * We mark the process as running here. This guarantees that
1563 * nobody will actually run it, and a signal or other external
1564 * event cannot wake it up and insert it on the runqueue either.
1566 p->state = TASK_RUNNING;
1569 * Make sure we do not leak PI boosting priority to the child.
1571 p->prio = current->normal_prio;
1574 * Revert to default priority/policy on fork if requested.
1576 if (unlikely(p->sched_reset_on_fork)) {
1577 if (task_has_rt_policy(p)) {
1578 p->policy = SCHED_NORMAL;
1579 p->static_prio = NICE_TO_PRIO(0);
1581 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1582 p->static_prio = NICE_TO_PRIO(0);
1584 p->prio = p->normal_prio = __normal_prio(p);
1588 * We don't need the reset flag anymore after the fork. It has
1589 * fulfilled its duty:
1591 p->sched_reset_on_fork = 0;
1594 if (!rt_prio(p->prio))
1595 p->sched_class = &fair_sched_class;
1597 if (p->sched_class->task_fork)
1598 p->sched_class->task_fork(p);
1601 * The child is not yet in the pid-hash so no cgroup attach races,
1602 * and the cgroup is pinned to this child due to cgroup_fork()
1603 * is ran before sched_fork().
1605 * Silence PROVE_RCU.
1607 raw_spin_lock_irqsave(&p->pi_lock, flags);
1608 set_task_cpu(p, cpu);
1609 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1611 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1612 if (likely(sched_info_on()))
1613 memset(&p->sched_info, 0, sizeof(p->sched_info));
1615 #if defined(CONFIG_SMP)
1618 #ifdef CONFIG_PREEMPT_COUNT
1619 /* Want to start with kernel preemption disabled. */
1620 task_thread_info(p)->preempt_count = 1;
1623 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1630 * wake_up_new_task - wake up a newly created task for the first time.
1632 * This function will do some initial scheduler statistics housekeeping
1633 * that must be done for every newly created context, then puts the task
1634 * on the runqueue and wakes it.
1636 void wake_up_new_task(struct task_struct *p)
1638 unsigned long flags;
1641 raw_spin_lock_irqsave(&p->pi_lock, flags);
1644 * Fork balancing, do it here and not earlier because:
1645 * - cpus_allowed can change in the fork path
1646 * - any previously selected cpu might disappear through hotplug
1648 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1651 rq = __task_rq_lock(p);
1652 activate_task(rq, p, 0);
1654 trace_sched_wakeup_new(p, true);
1655 check_preempt_curr(rq, p, WF_FORK);
1657 if (p->sched_class->task_woken)
1658 p->sched_class->task_woken(rq, p);
1660 task_rq_unlock(rq, p, &flags);
1663 #ifdef CONFIG_PREEMPT_NOTIFIERS
1666 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1667 * @notifier: notifier struct to register
1669 void preempt_notifier_register(struct preempt_notifier *notifier)
1671 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1673 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1676 * preempt_notifier_unregister - no longer interested in preemption notifications
1677 * @notifier: notifier struct to unregister
1679 * This is safe to call from within a preemption notifier.
1681 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1683 hlist_del(¬ifier->link);
1685 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1687 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1689 struct preempt_notifier *notifier;
1690 struct hlist_node *node;
1692 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1693 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1697 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1698 struct task_struct *next)
1700 struct preempt_notifier *notifier;
1701 struct hlist_node *node;
1703 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1704 notifier->ops->sched_out(notifier, next);
1707 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1709 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1714 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1715 struct task_struct *next)
1719 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1722 * prepare_task_switch - prepare to switch tasks
1723 * @rq: the runqueue preparing to switch
1724 * @prev: the current task that is being switched out
1725 * @next: the task we are going to switch to.
1727 * This is called with the rq lock held and interrupts off. It must
1728 * be paired with a subsequent finish_task_switch after the context
1731 * prepare_task_switch sets up locking and calls architecture specific
1735 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1736 struct task_struct *next)
1738 trace_sched_switch(prev, next);
1739 sched_info_switch(prev, next);
1740 perf_event_task_sched_out(prev, next);
1741 fire_sched_out_preempt_notifiers(prev, next);
1742 prepare_lock_switch(rq, next);
1743 prepare_arch_switch(next);
1747 * finish_task_switch - clean up after a task-switch
1748 * @rq: runqueue associated with task-switch
1749 * @prev: the thread we just switched away from.
1751 * finish_task_switch must be called after the context switch, paired
1752 * with a prepare_task_switch call before the context switch.
1753 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1754 * and do any other architecture-specific cleanup actions.
1756 * Note that we may have delayed dropping an mm in context_switch(). If
1757 * so, we finish that here outside of the runqueue lock. (Doing it
1758 * with the lock held can cause deadlocks; see schedule() for
1761 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1762 __releases(rq->lock)
1764 struct mm_struct *mm = rq->prev_mm;
1770 * A task struct has one reference for the use as "current".
1771 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1772 * schedule one last time. The schedule call will never return, and
1773 * the scheduled task must drop that reference.
1774 * The test for TASK_DEAD must occur while the runqueue locks are
1775 * still held, otherwise prev could be scheduled on another cpu, die
1776 * there before we look at prev->state, and then the reference would
1778 * Manfred Spraul <manfred@colorfullife.com>
1780 prev_state = prev->state;
1781 vtime_task_switch(prev);
1782 finish_arch_switch(prev);
1783 perf_event_task_sched_in(prev, current);
1784 finish_lock_switch(rq, prev);
1785 finish_arch_post_lock_switch();
1787 fire_sched_in_preempt_notifiers(current);
1790 if (unlikely(prev_state == TASK_DEAD)) {
1792 * Remove function-return probe instances associated with this
1793 * task and put them back on the free list.
1795 kprobe_flush_task(prev);
1796 put_task_struct(prev);
1802 /* assumes rq->lock is held */
1803 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1805 if (prev->sched_class->pre_schedule)
1806 prev->sched_class->pre_schedule(rq, prev);
1809 /* rq->lock is NOT held, but preemption is disabled */
1810 static inline void post_schedule(struct rq *rq)
1812 if (rq->post_schedule) {
1813 unsigned long flags;
1815 raw_spin_lock_irqsave(&rq->lock, flags);
1816 if (rq->curr->sched_class->post_schedule)
1817 rq->curr->sched_class->post_schedule(rq);
1818 raw_spin_unlock_irqrestore(&rq->lock, flags);
1820 rq->post_schedule = 0;
1826 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1830 static inline void post_schedule(struct rq *rq)
1837 * schedule_tail - first thing a freshly forked thread must call.
1838 * @prev: the thread we just switched away from.
1840 asmlinkage void schedule_tail(struct task_struct *prev)
1841 __releases(rq->lock)
1843 struct rq *rq = this_rq();
1845 finish_task_switch(rq, prev);
1848 * FIXME: do we need to worry about rq being invalidated by the
1853 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1854 /* In this case, finish_task_switch does not reenable preemption */
1857 if (current->set_child_tid)
1858 put_user(task_pid_vnr(current), current->set_child_tid);
1862 * context_switch - switch to the new MM and the new
1863 * thread's register state.
1866 context_switch(struct rq *rq, struct task_struct *prev,
1867 struct task_struct *next)
1869 struct mm_struct *mm, *oldmm;
1871 prepare_task_switch(rq, prev, next);
1874 oldmm = prev->active_mm;
1876 * For paravirt, this is coupled with an exit in switch_to to
1877 * combine the page table reload and the switch backend into
1880 arch_start_context_switch(prev);
1883 next->active_mm = oldmm;
1884 atomic_inc(&oldmm->mm_count);
1885 enter_lazy_tlb(oldmm, next);
1887 switch_mm(oldmm, mm, next);
1890 prev->active_mm = NULL;
1891 rq->prev_mm = oldmm;
1894 * Since the runqueue lock will be released by the next
1895 * task (which is an invalid locking op but in the case
1896 * of the scheduler it's an obvious special-case), so we
1897 * do an early lockdep release here:
1899 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1900 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1903 /* Here we just switch the register state and the stack. */
1904 rcu_switch(prev, next);
1905 switch_to(prev, next, prev);
1909 * this_rq must be evaluated again because prev may have moved
1910 * CPUs since it called schedule(), thus the 'rq' on its stack
1911 * frame will be invalid.
1913 finish_task_switch(this_rq(), prev);
1917 * nr_running, nr_uninterruptible and nr_context_switches:
1919 * externally visible scheduler statistics: current number of runnable
1920 * threads, current number of uninterruptible-sleeping threads, total
1921 * number of context switches performed since bootup.
1923 unsigned long nr_running(void)
1925 unsigned long i, sum = 0;
1927 for_each_online_cpu(i)
1928 sum += cpu_rq(i)->nr_running;
1933 unsigned long nr_uninterruptible(void)
1935 unsigned long i, sum = 0;
1937 for_each_possible_cpu(i)
1938 sum += cpu_rq(i)->nr_uninterruptible;
1941 * Since we read the counters lockless, it might be slightly
1942 * inaccurate. Do not allow it to go below zero though:
1944 if (unlikely((long)sum < 0))
1950 unsigned long long nr_context_switches(void)
1953 unsigned long long sum = 0;
1955 for_each_possible_cpu(i)
1956 sum += cpu_rq(i)->nr_switches;
1961 unsigned long nr_iowait(void)
1963 unsigned long i, sum = 0;
1965 for_each_possible_cpu(i)
1966 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1971 unsigned long nr_iowait_cpu(int cpu)
1973 struct rq *this = cpu_rq(cpu);
1974 return atomic_read(&this->nr_iowait);
1977 unsigned long this_cpu_load(void)
1979 struct rq *this = this_rq();
1980 return this->cpu_load[0];
1985 * Global load-average calculations
1987 * We take a distributed and async approach to calculating the global load-avg
1988 * in order to minimize overhead.
1990 * The global load average is an exponentially decaying average of nr_running +
1991 * nr_uninterruptible.
1993 * Once every LOAD_FREQ:
1996 * for_each_possible_cpu(cpu)
1997 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
1999 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2001 * Due to a number of reasons the above turns in the mess below:
2003 * - for_each_possible_cpu() is prohibitively expensive on machines with
2004 * serious number of cpus, therefore we need to take a distributed approach
2005 * to calculating nr_active.
2007 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2008 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2010 * So assuming nr_active := 0 when we start out -- true per definition, we
2011 * can simply take per-cpu deltas and fold those into a global accumulate
2012 * to obtain the same result. See calc_load_fold_active().
2014 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2015 * across the machine, we assume 10 ticks is sufficient time for every
2016 * cpu to have completed this task.
2018 * This places an upper-bound on the IRQ-off latency of the machine. Then
2019 * again, being late doesn't loose the delta, just wrecks the sample.
2021 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2022 * this would add another cross-cpu cacheline miss and atomic operation
2023 * to the wakeup path. Instead we increment on whatever cpu the task ran
2024 * when it went into uninterruptible state and decrement on whatever cpu
2025 * did the wakeup. This means that only the sum of nr_uninterruptible over
2026 * all cpus yields the correct result.
2028 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2031 /* Variables and functions for calc_load */
2032 static atomic_long_t calc_load_tasks;
2033 static unsigned long calc_load_update;
2034 unsigned long avenrun[3];
2035 EXPORT_SYMBOL(avenrun); /* should be removed */
2038 * get_avenrun - get the load average array
2039 * @loads: pointer to dest load array
2040 * @offset: offset to add
2041 * @shift: shift count to shift the result left
2043 * These values are estimates at best, so no need for locking.
2045 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2047 loads[0] = (avenrun[0] + offset) << shift;
2048 loads[1] = (avenrun[1] + offset) << shift;
2049 loads[2] = (avenrun[2] + offset) << shift;
2052 static long calc_load_fold_active(struct rq *this_rq)
2054 long nr_active, delta = 0;
2056 nr_active = this_rq->nr_running;
2057 nr_active += (long) this_rq->nr_uninterruptible;
2059 if (nr_active != this_rq->calc_load_active) {
2060 delta = nr_active - this_rq->calc_load_active;
2061 this_rq->calc_load_active = nr_active;
2068 * a1 = a0 * e + a * (1 - e)
2070 static unsigned long
2071 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2074 load += active * (FIXED_1 - exp);
2075 load += 1UL << (FSHIFT - 1);
2076 return load >> FSHIFT;
2081 * Handle NO_HZ for the global load-average.
2083 * Since the above described distributed algorithm to compute the global
2084 * load-average relies on per-cpu sampling from the tick, it is affected by
2087 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2088 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2089 * when we read the global state.
2091 * Obviously reality has to ruin such a delightfully simple scheme:
2093 * - When we go NO_HZ idle during the window, we can negate our sample
2094 * contribution, causing under-accounting.
2096 * We avoid this by keeping two idle-delta counters and flipping them
2097 * when the window starts, thus separating old and new NO_HZ load.
2099 * The only trick is the slight shift in index flip for read vs write.
2103 * |-|-----------|-|-----------|-|-----------|-|
2104 * r:0 0 1 1 0 0 1 1 0
2105 * w:0 1 1 0 0 1 1 0 0
2107 * This ensures we'll fold the old idle contribution in this window while
2108 * accumlating the new one.
2110 * - When we wake up from NO_HZ idle during the window, we push up our
2111 * contribution, since we effectively move our sample point to a known
2114 * This is solved by pushing the window forward, and thus skipping the
2115 * sample, for this cpu (effectively using the idle-delta for this cpu which
2116 * was in effect at the time the window opened). This also solves the issue
2117 * of having to deal with a cpu having been in NOHZ idle for multiple
2118 * LOAD_FREQ intervals.
2120 * When making the ILB scale, we should try to pull this in as well.
2122 static atomic_long_t calc_load_idle[2];
2123 static int calc_load_idx;
2125 static inline int calc_load_write_idx(void)
2127 int idx = calc_load_idx;
2130 * See calc_global_nohz(), if we observe the new index, we also
2131 * need to observe the new update time.
2136 * If the folding window started, make sure we start writing in the
2139 if (!time_before(jiffies, calc_load_update))
2145 static inline int calc_load_read_idx(void)
2147 return calc_load_idx & 1;
2150 void calc_load_enter_idle(void)
2152 struct rq *this_rq = this_rq();
2156 * We're going into NOHZ mode, if there's any pending delta, fold it
2157 * into the pending idle delta.
2159 delta = calc_load_fold_active(this_rq);
2161 int idx = calc_load_write_idx();
2162 atomic_long_add(delta, &calc_load_idle[idx]);
2166 void calc_load_exit_idle(void)
2168 struct rq *this_rq = this_rq();
2171 * If we're still before the sample window, we're done.
2173 if (time_before(jiffies, this_rq->calc_load_update))
2177 * We woke inside or after the sample window, this means we're already
2178 * accounted through the nohz accounting, so skip the entire deal and
2179 * sync up for the next window.
2181 this_rq->calc_load_update = calc_load_update;
2182 if (time_before(jiffies, this_rq->calc_load_update + 10))
2183 this_rq->calc_load_update += LOAD_FREQ;
2186 static long calc_load_fold_idle(void)
2188 int idx = calc_load_read_idx();
2191 if (atomic_long_read(&calc_load_idle[idx]))
2192 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2198 * fixed_power_int - compute: x^n, in O(log n) time
2200 * @x: base of the power
2201 * @frac_bits: fractional bits of @x
2202 * @n: power to raise @x to.
2204 * By exploiting the relation between the definition of the natural power
2205 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2206 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2207 * (where: n_i \elem {0, 1}, the binary vector representing n),
2208 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2209 * of course trivially computable in O(log_2 n), the length of our binary
2212 static unsigned long
2213 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2215 unsigned long result = 1UL << frac_bits;
2220 result += 1UL << (frac_bits - 1);
2221 result >>= frac_bits;
2227 x += 1UL << (frac_bits - 1);
2235 * a1 = a0 * e + a * (1 - e)
2237 * a2 = a1 * e + a * (1 - e)
2238 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2239 * = a0 * e^2 + a * (1 - e) * (1 + e)
2241 * a3 = a2 * e + a * (1 - e)
2242 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2243 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2247 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2248 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2249 * = a0 * e^n + a * (1 - e^n)
2251 * [1] application of the geometric series:
2254 * S_n := \Sum x^i = -------------
2257 static unsigned long
2258 calc_load_n(unsigned long load, unsigned long exp,
2259 unsigned long active, unsigned int n)
2262 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2266 * NO_HZ can leave us missing all per-cpu ticks calling
2267 * calc_load_account_active(), but since an idle CPU folds its delta into
2268 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2269 * in the pending idle delta if our idle period crossed a load cycle boundary.
2271 * Once we've updated the global active value, we need to apply the exponential
2272 * weights adjusted to the number of cycles missed.
2274 static void calc_global_nohz(void)
2276 long delta, active, n;
2278 if (!time_before(jiffies, calc_load_update + 10)) {
2280 * Catch-up, fold however many we are behind still
2282 delta = jiffies - calc_load_update - 10;
2283 n = 1 + (delta / LOAD_FREQ);
2285 active = atomic_long_read(&calc_load_tasks);
2286 active = active > 0 ? active * FIXED_1 : 0;
2288 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2289 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2290 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2292 calc_load_update += n * LOAD_FREQ;
2296 * Flip the idle index...
2298 * Make sure we first write the new time then flip the index, so that
2299 * calc_load_write_idx() will see the new time when it reads the new
2300 * index, this avoids a double flip messing things up.
2305 #else /* !CONFIG_NO_HZ */
2307 static inline long calc_load_fold_idle(void) { return 0; }
2308 static inline void calc_global_nohz(void) { }
2310 #endif /* CONFIG_NO_HZ */
2313 * calc_load - update the avenrun load estimates 10 ticks after the
2314 * CPUs have updated calc_load_tasks.
2316 void calc_global_load(unsigned long ticks)
2320 if (time_before(jiffies, calc_load_update + 10))
2324 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2326 delta = calc_load_fold_idle();
2328 atomic_long_add(delta, &calc_load_tasks);
2330 active = atomic_long_read(&calc_load_tasks);
2331 active = active > 0 ? active * FIXED_1 : 0;
2333 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2334 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2335 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2337 calc_load_update += LOAD_FREQ;
2340 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2346 * Called from update_cpu_load() to periodically update this CPU's
2349 static void calc_load_account_active(struct rq *this_rq)
2353 if (time_before(jiffies, this_rq->calc_load_update))
2356 delta = calc_load_fold_active(this_rq);
2358 atomic_long_add(delta, &calc_load_tasks);
2360 this_rq->calc_load_update += LOAD_FREQ;
2364 * End of global load-average stuff
2368 * The exact cpuload at various idx values, calculated at every tick would be
2369 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2371 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2372 * on nth tick when cpu may be busy, then we have:
2373 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2374 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2376 * decay_load_missed() below does efficient calculation of
2377 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2378 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2380 * The calculation is approximated on a 128 point scale.
2381 * degrade_zero_ticks is the number of ticks after which load at any
2382 * particular idx is approximated to be zero.
2383 * degrade_factor is a precomputed table, a row for each load idx.
2384 * Each column corresponds to degradation factor for a power of two ticks,
2385 * based on 128 point scale.
2387 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2388 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2390 * With this power of 2 load factors, we can degrade the load n times
2391 * by looking at 1 bits in n and doing as many mult/shift instead of
2392 * n mult/shifts needed by the exact degradation.
2394 #define DEGRADE_SHIFT 7
2395 static const unsigned char
2396 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2397 static const unsigned char
2398 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2399 {0, 0, 0, 0, 0, 0, 0, 0},
2400 {64, 32, 8, 0, 0, 0, 0, 0},
2401 {96, 72, 40, 12, 1, 0, 0},
2402 {112, 98, 75, 43, 15, 1, 0},
2403 {120, 112, 98, 76, 45, 16, 2} };
2406 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2407 * would be when CPU is idle and so we just decay the old load without
2408 * adding any new load.
2410 static unsigned long
2411 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2415 if (!missed_updates)
2418 if (missed_updates >= degrade_zero_ticks[idx])
2422 return load >> missed_updates;
2424 while (missed_updates) {
2425 if (missed_updates % 2)
2426 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2428 missed_updates >>= 1;
2435 * Update rq->cpu_load[] statistics. This function is usually called every
2436 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2437 * every tick. We fix it up based on jiffies.
2439 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2440 unsigned long pending_updates)
2444 this_rq->nr_load_updates++;
2446 /* Update our load: */
2447 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2448 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2449 unsigned long old_load, new_load;
2451 /* scale is effectively 1 << i now, and >> i divides by scale */
2453 old_load = this_rq->cpu_load[i];
2454 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2455 new_load = this_load;
2457 * Round up the averaging division if load is increasing. This
2458 * prevents us from getting stuck on 9 if the load is 10, for
2461 if (new_load > old_load)
2462 new_load += scale - 1;
2464 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2467 sched_avg_update(this_rq);
2472 * There is no sane way to deal with nohz on smp when using jiffies because the
2473 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2474 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2476 * Therefore we cannot use the delta approach from the regular tick since that
2477 * would seriously skew the load calculation. However we'll make do for those
2478 * updates happening while idle (nohz_idle_balance) or coming out of idle
2479 * (tick_nohz_idle_exit).
2481 * This means we might still be one tick off for nohz periods.
2485 * Called from nohz_idle_balance() to update the load ratings before doing the
2488 void update_idle_cpu_load(struct rq *this_rq)
2490 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2491 unsigned long load = this_rq->load.weight;
2492 unsigned long pending_updates;
2495 * bail if there's load or we're actually up-to-date.
2497 if (load || curr_jiffies == this_rq->last_load_update_tick)
2500 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2501 this_rq->last_load_update_tick = curr_jiffies;
2503 __update_cpu_load(this_rq, load, pending_updates);
2507 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2509 void update_cpu_load_nohz(void)
2511 struct rq *this_rq = this_rq();
2512 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2513 unsigned long pending_updates;
2515 if (curr_jiffies == this_rq->last_load_update_tick)
2518 raw_spin_lock(&this_rq->lock);
2519 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2520 if (pending_updates) {
2521 this_rq->last_load_update_tick = curr_jiffies;
2523 * We were idle, this means load 0, the current load might be
2524 * !0 due to remote wakeups and the sort.
2526 __update_cpu_load(this_rq, 0, pending_updates);
2528 raw_spin_unlock(&this_rq->lock);
2530 #endif /* CONFIG_NO_HZ */
2533 * Called from scheduler_tick()
2535 static void update_cpu_load_active(struct rq *this_rq)
2538 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2540 this_rq->last_load_update_tick = jiffies;
2541 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2543 calc_load_account_active(this_rq);
2549 * sched_exec - execve() is a valuable balancing opportunity, because at
2550 * this point the task has the smallest effective memory and cache footprint.
2552 void sched_exec(void)
2554 struct task_struct *p = current;
2555 unsigned long flags;
2558 raw_spin_lock_irqsave(&p->pi_lock, flags);
2559 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2560 if (dest_cpu == smp_processor_id())
2563 if (likely(cpu_active(dest_cpu))) {
2564 struct migration_arg arg = { p, dest_cpu };
2566 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2567 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2571 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2576 DEFINE_PER_CPU(struct kernel_stat, kstat);
2577 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2579 EXPORT_PER_CPU_SYMBOL(kstat);
2580 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2583 * Return any ns on the sched_clock that have not yet been accounted in
2584 * @p in case that task is currently running.
2586 * Called with task_rq_lock() held on @rq.
2588 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2592 if (task_current(rq, p)) {
2593 update_rq_clock(rq);
2594 ns = rq->clock_task - p->se.exec_start;
2602 unsigned long long task_delta_exec(struct task_struct *p)
2604 unsigned long flags;
2608 rq = task_rq_lock(p, &flags);
2609 ns = do_task_delta_exec(p, rq);
2610 task_rq_unlock(rq, p, &flags);
2616 * Return accounted runtime for the task.
2617 * In case the task is currently running, return the runtime plus current's
2618 * pending runtime that have not been accounted yet.
2620 unsigned long long task_sched_runtime(struct task_struct *p)
2622 unsigned long flags;
2626 rq = task_rq_lock(p, &flags);
2627 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2628 task_rq_unlock(rq, p, &flags);
2634 * This function gets called by the timer code, with HZ frequency.
2635 * We call it with interrupts disabled.
2637 void scheduler_tick(void)
2639 int cpu = smp_processor_id();
2640 struct rq *rq = cpu_rq(cpu);
2641 struct task_struct *curr = rq->curr;
2645 raw_spin_lock(&rq->lock);
2646 update_rq_clock(rq);
2647 update_cpu_load_active(rq);
2648 curr->sched_class->task_tick(rq, curr, 0);
2649 raw_spin_unlock(&rq->lock);
2651 perf_event_task_tick();
2654 rq->idle_balance = idle_cpu(cpu);
2655 trigger_load_balance(rq, cpu);
2659 notrace unsigned long get_parent_ip(unsigned long addr)
2661 if (in_lock_functions(addr)) {
2662 addr = CALLER_ADDR2;
2663 if (in_lock_functions(addr))
2664 addr = CALLER_ADDR3;
2669 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2670 defined(CONFIG_PREEMPT_TRACER))
2672 void __kprobes add_preempt_count(int val)
2674 #ifdef CONFIG_DEBUG_PREEMPT
2678 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2681 preempt_count() += val;
2682 #ifdef CONFIG_DEBUG_PREEMPT
2684 * Spinlock count overflowing soon?
2686 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2689 if (preempt_count() == val)
2690 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2692 EXPORT_SYMBOL(add_preempt_count);
2694 void __kprobes sub_preempt_count(int val)
2696 #ifdef CONFIG_DEBUG_PREEMPT
2700 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2703 * Is the spinlock portion underflowing?
2705 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2706 !(preempt_count() & PREEMPT_MASK)))
2710 if (preempt_count() == val)
2711 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2712 preempt_count() -= val;
2714 EXPORT_SYMBOL(sub_preempt_count);
2719 * Print scheduling while atomic bug:
2721 static noinline void __schedule_bug(struct task_struct *prev)
2723 if (oops_in_progress)
2726 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2727 prev->comm, prev->pid, preempt_count());
2729 debug_show_held_locks(prev);
2731 if (irqs_disabled())
2732 print_irqtrace_events(prev);
2734 add_taint(TAINT_WARN);
2738 * Various schedule()-time debugging checks and statistics:
2740 static inline void schedule_debug(struct task_struct *prev)
2743 * Test if we are atomic. Since do_exit() needs to call into
2744 * schedule() atomically, we ignore that path for now.
2745 * Otherwise, whine if we are scheduling when we should not be.
2747 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2748 __schedule_bug(prev);
2751 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2753 schedstat_inc(this_rq(), sched_count);
2756 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2758 if (prev->on_rq || rq->skip_clock_update < 0)
2759 update_rq_clock(rq);
2760 prev->sched_class->put_prev_task(rq, prev);
2764 * Pick up the highest-prio task:
2766 static inline struct task_struct *
2767 pick_next_task(struct rq *rq)
2769 const struct sched_class *class;
2770 struct task_struct *p;
2773 * Optimization: we know that if all tasks are in
2774 * the fair class we can call that function directly:
2776 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2777 p = fair_sched_class.pick_next_task(rq);
2782 for_each_class(class) {
2783 p = class->pick_next_task(rq);
2788 BUG(); /* the idle class will always have a runnable task */
2792 * __schedule() is the main scheduler function.
2794 * The main means of driving the scheduler and thus entering this function are:
2796 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2798 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2799 * paths. For example, see arch/x86/entry_64.S.
2801 * To drive preemption between tasks, the scheduler sets the flag in timer
2802 * interrupt handler scheduler_tick().
2804 * 3. Wakeups don't really cause entry into schedule(). They add a
2805 * task to the run-queue and that's it.
2807 * Now, if the new task added to the run-queue preempts the current
2808 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2809 * called on the nearest possible occasion:
2811 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2813 * - in syscall or exception context, at the next outmost
2814 * preempt_enable(). (this might be as soon as the wake_up()'s
2817 * - in IRQ context, return from interrupt-handler to
2818 * preemptible context
2820 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2823 * - cond_resched() call
2824 * - explicit schedule() call
2825 * - return from syscall or exception to user-space
2826 * - return from interrupt-handler to user-space
2828 static void __sched __schedule(void)
2830 struct task_struct *prev, *next;
2831 unsigned long *switch_count;
2837 cpu = smp_processor_id();
2839 rcu_note_context_switch(cpu);
2842 schedule_debug(prev);
2844 if (sched_feat(HRTICK))
2847 raw_spin_lock_irq(&rq->lock);
2849 switch_count = &prev->nivcsw;
2850 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2851 if (unlikely(signal_pending_state(prev->state, prev))) {
2852 prev->state = TASK_RUNNING;
2854 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2858 * If a worker went to sleep, notify and ask workqueue
2859 * whether it wants to wake up a task to maintain
2862 if (prev->flags & PF_WQ_WORKER) {
2863 struct task_struct *to_wakeup;
2865 to_wakeup = wq_worker_sleeping(prev, cpu);
2867 try_to_wake_up_local(to_wakeup);
2870 switch_count = &prev->nvcsw;
2873 pre_schedule(rq, prev);
2875 if (unlikely(!rq->nr_running))
2876 idle_balance(cpu, rq);
2878 put_prev_task(rq, prev);
2879 next = pick_next_task(rq);
2880 clear_tsk_need_resched(prev);
2881 rq->skip_clock_update = 0;
2883 if (likely(prev != next)) {
2888 context_switch(rq, prev, next); /* unlocks the rq */
2890 * The context switch have flipped the stack from under us
2891 * and restored the local variables which were saved when
2892 * this task called schedule() in the past. prev == current
2893 * is still correct, but it can be moved to another cpu/rq.
2895 cpu = smp_processor_id();
2898 raw_spin_unlock_irq(&rq->lock);
2902 sched_preempt_enable_no_resched();
2907 static inline void sched_submit_work(struct task_struct *tsk)
2909 if (!tsk->state || tsk_is_pi_blocked(tsk))
2912 * If we are going to sleep and we have plugged IO queued,
2913 * make sure to submit it to avoid deadlocks.
2915 if (blk_needs_flush_plug(tsk))
2916 blk_schedule_flush_plug(tsk);
2919 asmlinkage void __sched schedule(void)
2921 struct task_struct *tsk = current;
2923 sched_submit_work(tsk);
2926 EXPORT_SYMBOL(schedule);
2928 #ifdef CONFIG_RCU_USER_QS
2929 asmlinkage void __sched schedule_user(void)
2932 * If we come here after a random call to set_need_resched(),
2933 * or we have been woken up remotely but the IPI has not yet arrived,
2934 * we haven't yet exited the RCU idle mode. Do it here manually until
2935 * we find a better solution.
2944 * schedule_preempt_disabled - called with preemption disabled
2946 * Returns with preemption disabled. Note: preempt_count must be 1
2948 void __sched schedule_preempt_disabled(void)
2950 sched_preempt_enable_no_resched();
2955 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2957 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2959 if (lock->owner != owner)
2963 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2964 * lock->owner still matches owner, if that fails, owner might
2965 * point to free()d memory, if it still matches, the rcu_read_lock()
2966 * ensures the memory stays valid.
2970 return owner->on_cpu;
2974 * Look out! "owner" is an entirely speculative pointer
2975 * access and not reliable.
2977 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
2979 if (!sched_feat(OWNER_SPIN))
2983 while (owner_running(lock, owner)) {
2987 arch_mutex_cpu_relax();
2992 * We break out the loop above on need_resched() and when the
2993 * owner changed, which is a sign for heavy contention. Return
2994 * success only when lock->owner is NULL.
2996 return lock->owner == NULL;
3000 #ifdef CONFIG_PREEMPT
3002 * this is the entry point to schedule() from in-kernel preemption
3003 * off of preempt_enable. Kernel preemptions off return from interrupt
3004 * occur there and call schedule directly.
3006 asmlinkage void __sched notrace preempt_schedule(void)
3008 struct thread_info *ti = current_thread_info();
3011 * If there is a non-zero preempt_count or interrupts are disabled,
3012 * we do not want to preempt the current task. Just return..
3014 if (likely(ti->preempt_count || irqs_disabled()))
3018 add_preempt_count_notrace(PREEMPT_ACTIVE);
3020 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3023 * Check again in case we missed a preemption opportunity
3024 * between schedule and now.
3027 } while (need_resched());
3029 EXPORT_SYMBOL(preempt_schedule);
3032 * this is the entry point to schedule() from kernel preemption
3033 * off of irq context.
3034 * Note, that this is called and return with irqs disabled. This will
3035 * protect us against recursive calling from irq.
3037 asmlinkage void __sched preempt_schedule_irq(void)
3039 struct thread_info *ti = current_thread_info();
3041 /* Catch callers which need to be fixed */
3042 BUG_ON(ti->preempt_count || !irqs_disabled());
3046 add_preempt_count(PREEMPT_ACTIVE);
3049 local_irq_disable();
3050 sub_preempt_count(PREEMPT_ACTIVE);
3053 * Check again in case we missed a preemption opportunity
3054 * between schedule and now.
3057 } while (need_resched());
3060 #endif /* CONFIG_PREEMPT */
3062 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3065 return try_to_wake_up(curr->private, mode, wake_flags);
3067 EXPORT_SYMBOL(default_wake_function);
3070 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3071 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3072 * number) then we wake all the non-exclusive tasks and one exclusive task.
3074 * There are circumstances in which we can try to wake a task which has already
3075 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3076 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3078 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3079 int nr_exclusive, int wake_flags, void *key)
3081 wait_queue_t *curr, *next;
3083 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3084 unsigned flags = curr->flags;
3086 if (curr->func(curr, mode, wake_flags, key) &&
3087 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3093 * __wake_up - wake up threads blocked on a waitqueue.
3095 * @mode: which threads
3096 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3097 * @key: is directly passed to the wakeup function
3099 * It may be assumed that this function implies a write memory barrier before
3100 * changing the task state if and only if any tasks are woken up.
3102 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3103 int nr_exclusive, void *key)
3105 unsigned long flags;
3107 spin_lock_irqsave(&q->lock, flags);
3108 __wake_up_common(q, mode, nr_exclusive, 0, key);
3109 spin_unlock_irqrestore(&q->lock, flags);
3111 EXPORT_SYMBOL(__wake_up);
3114 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3116 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3118 __wake_up_common(q, mode, nr, 0, NULL);
3120 EXPORT_SYMBOL_GPL(__wake_up_locked);
3122 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3124 __wake_up_common(q, mode, 1, 0, key);
3126 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3129 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3131 * @mode: which threads
3132 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3133 * @key: opaque value to be passed to wakeup targets
3135 * The sync wakeup differs that the waker knows that it will schedule
3136 * away soon, so while the target thread will be woken up, it will not
3137 * be migrated to another CPU - ie. the two threads are 'synchronized'
3138 * with each other. This can prevent needless bouncing between CPUs.
3140 * On UP it can prevent extra preemption.
3142 * It may be assumed that this function implies a write memory barrier before
3143 * changing the task state if and only if any tasks are woken up.
3145 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3146 int nr_exclusive, void *key)
3148 unsigned long flags;
3149 int wake_flags = WF_SYNC;
3154 if (unlikely(!nr_exclusive))
3157 spin_lock_irqsave(&q->lock, flags);
3158 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3159 spin_unlock_irqrestore(&q->lock, flags);
3161 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3164 * __wake_up_sync - see __wake_up_sync_key()
3166 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3168 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3170 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3173 * complete: - signals a single thread waiting on this completion
3174 * @x: holds the state of this particular completion
3176 * This will wake up a single thread waiting on this completion. Threads will be
3177 * awakened in the same order in which they were queued.
3179 * See also complete_all(), wait_for_completion() and related routines.
3181 * It may be assumed that this function implies a write memory barrier before
3182 * changing the task state if and only if any tasks are woken up.
3184 void complete(struct completion *x)
3186 unsigned long flags;
3188 spin_lock_irqsave(&x->wait.lock, flags);
3190 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3191 spin_unlock_irqrestore(&x->wait.lock, flags);
3193 EXPORT_SYMBOL(complete);
3196 * complete_all: - signals all threads waiting on this completion
3197 * @x: holds the state of this particular completion
3199 * This will wake up all threads waiting on this particular completion event.
3201 * It may be assumed that this function implies a write memory barrier before
3202 * changing the task state if and only if any tasks are woken up.
3204 void complete_all(struct completion *x)
3206 unsigned long flags;
3208 spin_lock_irqsave(&x->wait.lock, flags);
3209 x->done += UINT_MAX/2;
3210 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3211 spin_unlock_irqrestore(&x->wait.lock, flags);
3213 EXPORT_SYMBOL(complete_all);
3215 static inline long __sched
3216 do_wait_for_common(struct completion *x, long timeout, int state)
3219 DECLARE_WAITQUEUE(wait, current);
3221 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3223 if (signal_pending_state(state, current)) {
3224 timeout = -ERESTARTSYS;
3227 __set_current_state(state);
3228 spin_unlock_irq(&x->wait.lock);
3229 timeout = schedule_timeout(timeout);
3230 spin_lock_irq(&x->wait.lock);
3231 } while (!x->done && timeout);
3232 __remove_wait_queue(&x->wait, &wait);
3237 return timeout ?: 1;
3241 wait_for_common(struct completion *x, long timeout, int state)
3245 spin_lock_irq(&x->wait.lock);
3246 timeout = do_wait_for_common(x, timeout, state);
3247 spin_unlock_irq(&x->wait.lock);
3252 * wait_for_completion: - waits for completion of a task
3253 * @x: holds the state of this particular completion
3255 * This waits to be signaled for completion of a specific task. It is NOT
3256 * interruptible and there is no timeout.
3258 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3259 * and interrupt capability. Also see complete().
3261 void __sched wait_for_completion(struct completion *x)
3263 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3265 EXPORT_SYMBOL(wait_for_completion);
3268 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3269 * @x: holds the state of this particular completion
3270 * @timeout: timeout value in jiffies
3272 * This waits for either a completion of a specific task to be signaled or for a
3273 * specified timeout to expire. The timeout is in jiffies. It is not
3276 * The return value is 0 if timed out, and positive (at least 1, or number of
3277 * jiffies left till timeout) if completed.
3279 unsigned long __sched
3280 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3282 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3284 EXPORT_SYMBOL(wait_for_completion_timeout);
3287 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3288 * @x: holds the state of this particular completion
3290 * This waits for completion of a specific task to be signaled. It is
3293 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3295 int __sched wait_for_completion_interruptible(struct completion *x)
3297 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3298 if (t == -ERESTARTSYS)
3302 EXPORT_SYMBOL(wait_for_completion_interruptible);
3305 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3306 * @x: holds the state of this particular completion
3307 * @timeout: timeout value in jiffies
3309 * This waits for either a completion of a specific task to be signaled or for a
3310 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3312 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3313 * positive (at least 1, or number of jiffies left till timeout) if completed.
3316 wait_for_completion_interruptible_timeout(struct completion *x,
3317 unsigned long timeout)
3319 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3321 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3324 * wait_for_completion_killable: - waits for completion of a task (killable)
3325 * @x: holds the state of this particular completion
3327 * This waits to be signaled for completion of a specific task. It can be
3328 * interrupted by a kill signal.
3330 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3332 int __sched wait_for_completion_killable(struct completion *x)
3334 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3335 if (t == -ERESTARTSYS)
3339 EXPORT_SYMBOL(wait_for_completion_killable);
3342 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3343 * @x: holds the state of this particular completion
3344 * @timeout: timeout value in jiffies
3346 * This waits for either a completion of a specific task to be
3347 * signaled or for a specified timeout to expire. It can be
3348 * interrupted by a kill signal. The timeout is in jiffies.
3350 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3351 * positive (at least 1, or number of jiffies left till timeout) if completed.
3354 wait_for_completion_killable_timeout(struct completion *x,
3355 unsigned long timeout)
3357 return wait_for_common(x, timeout, TASK_KILLABLE);
3359 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3362 * try_wait_for_completion - try to decrement a completion without blocking
3363 * @x: completion structure
3365 * Returns: 0 if a decrement cannot be done without blocking
3366 * 1 if a decrement succeeded.
3368 * If a completion is being used as a counting completion,
3369 * attempt to decrement the counter without blocking. This
3370 * enables us to avoid waiting if the resource the completion
3371 * is protecting is not available.
3373 bool try_wait_for_completion(struct completion *x)
3375 unsigned long flags;
3378 spin_lock_irqsave(&x->wait.lock, flags);
3383 spin_unlock_irqrestore(&x->wait.lock, flags);
3386 EXPORT_SYMBOL(try_wait_for_completion);
3389 * completion_done - Test to see if a completion has any waiters
3390 * @x: completion structure
3392 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3393 * 1 if there are no waiters.
3396 bool completion_done(struct completion *x)
3398 unsigned long flags;
3401 spin_lock_irqsave(&x->wait.lock, flags);
3404 spin_unlock_irqrestore(&x->wait.lock, flags);
3407 EXPORT_SYMBOL(completion_done);
3410 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3412 unsigned long flags;
3415 init_waitqueue_entry(&wait, current);
3417 __set_current_state(state);
3419 spin_lock_irqsave(&q->lock, flags);
3420 __add_wait_queue(q, &wait);
3421 spin_unlock(&q->lock);
3422 timeout = schedule_timeout(timeout);
3423 spin_lock_irq(&q->lock);
3424 __remove_wait_queue(q, &wait);
3425 spin_unlock_irqrestore(&q->lock, flags);
3430 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3432 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3434 EXPORT_SYMBOL(interruptible_sleep_on);
3437 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3439 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3441 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3443 void __sched sleep_on(wait_queue_head_t *q)
3445 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3447 EXPORT_SYMBOL(sleep_on);
3449 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3451 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3453 EXPORT_SYMBOL(sleep_on_timeout);
3455 #ifdef CONFIG_RT_MUTEXES
3458 * rt_mutex_setprio - set the current priority of a task
3460 * @prio: prio value (kernel-internal form)
3462 * This function changes the 'effective' priority of a task. It does
3463 * not touch ->normal_prio like __setscheduler().
3465 * Used by the rt_mutex code to implement priority inheritance logic.
3467 void rt_mutex_setprio(struct task_struct *p, int prio)
3469 int oldprio, on_rq, running;
3471 const struct sched_class *prev_class;
3473 BUG_ON(prio < 0 || prio > MAX_PRIO);
3475 rq = __task_rq_lock(p);
3478 * Idle task boosting is a nono in general. There is one
3479 * exception, when PREEMPT_RT and NOHZ is active:
3481 * The idle task calls get_next_timer_interrupt() and holds
3482 * the timer wheel base->lock on the CPU and another CPU wants
3483 * to access the timer (probably to cancel it). We can safely
3484 * ignore the boosting request, as the idle CPU runs this code
3485 * with interrupts disabled and will complete the lock
3486 * protected section without being interrupted. So there is no
3487 * real need to boost.
3489 if (unlikely(p == rq->idle)) {
3490 WARN_ON(p != rq->curr);
3491 WARN_ON(p->pi_blocked_on);
3495 trace_sched_pi_setprio(p, prio);
3497 prev_class = p->sched_class;
3499 running = task_current(rq, p);
3501 dequeue_task(rq, p, 0);
3503 p->sched_class->put_prev_task(rq, p);
3506 p->sched_class = &rt_sched_class;
3508 p->sched_class = &fair_sched_class;
3513 p->sched_class->set_curr_task(rq);
3515 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3517 check_class_changed(rq, p, prev_class, oldprio);
3519 __task_rq_unlock(rq);
3522 void set_user_nice(struct task_struct *p, long nice)
3524 int old_prio, delta, on_rq;
3525 unsigned long flags;
3528 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3531 * We have to be careful, if called from sys_setpriority(),
3532 * the task might be in the middle of scheduling on another CPU.
3534 rq = task_rq_lock(p, &flags);
3536 * The RT priorities are set via sched_setscheduler(), but we still
3537 * allow the 'normal' nice value to be set - but as expected
3538 * it wont have any effect on scheduling until the task is
3539 * SCHED_FIFO/SCHED_RR:
3541 if (task_has_rt_policy(p)) {
3542 p->static_prio = NICE_TO_PRIO(nice);
3547 dequeue_task(rq, p, 0);
3549 p->static_prio = NICE_TO_PRIO(nice);
3552 p->prio = effective_prio(p);
3553 delta = p->prio - old_prio;
3556 enqueue_task(rq, p, 0);
3558 * If the task increased its priority or is running and
3559 * lowered its priority, then reschedule its CPU:
3561 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3562 resched_task(rq->curr);
3565 task_rq_unlock(rq, p, &flags);
3567 EXPORT_SYMBOL(set_user_nice);
3570 * can_nice - check if a task can reduce its nice value
3574 int can_nice(const struct task_struct *p, const int nice)
3576 /* convert nice value [19,-20] to rlimit style value [1,40] */
3577 int nice_rlim = 20 - nice;
3579 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3580 capable(CAP_SYS_NICE));
3583 #ifdef __ARCH_WANT_SYS_NICE
3586 * sys_nice - change the priority of the current process.
3587 * @increment: priority increment
3589 * sys_setpriority is a more generic, but much slower function that
3590 * does similar things.
3592 SYSCALL_DEFINE1(nice, int, increment)
3597 * Setpriority might change our priority at the same moment.
3598 * We don't have to worry. Conceptually one call occurs first
3599 * and we have a single winner.
3601 if (increment < -40)
3606 nice = TASK_NICE(current) + increment;
3612 if (increment < 0 && !can_nice(current, nice))
3615 retval = security_task_setnice(current, nice);
3619 set_user_nice(current, nice);
3626 * task_prio - return the priority value of a given task.
3627 * @p: the task in question.
3629 * This is the priority value as seen by users in /proc.
3630 * RT tasks are offset by -200. Normal tasks are centered
3631 * around 0, value goes from -16 to +15.
3633 int task_prio(const struct task_struct *p)
3635 return p->prio - MAX_RT_PRIO;
3639 * task_nice - return the nice value of a given task.
3640 * @p: the task in question.
3642 int task_nice(const struct task_struct *p)
3644 return TASK_NICE(p);
3646 EXPORT_SYMBOL(task_nice);
3649 * idle_cpu - is a given cpu idle currently?
3650 * @cpu: the processor in question.
3652 int idle_cpu(int cpu)
3654 struct rq *rq = cpu_rq(cpu);
3656 if (rq->curr != rq->idle)
3663 if (!llist_empty(&rq->wake_list))
3671 * idle_task - return the idle task for a given cpu.
3672 * @cpu: the processor in question.
3674 struct task_struct *idle_task(int cpu)
3676 return cpu_rq(cpu)->idle;
3680 * find_process_by_pid - find a process with a matching PID value.
3681 * @pid: the pid in question.
3683 static struct task_struct *find_process_by_pid(pid_t pid)
3685 return pid ? find_task_by_vpid(pid) : current;
3688 /* Actually do priority change: must hold rq lock. */
3690 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3693 p->rt_priority = prio;
3694 p->normal_prio = normal_prio(p);
3695 /* we are holding p->pi_lock already */
3696 p->prio = rt_mutex_getprio(p);
3697 if (rt_prio(p->prio))
3698 p->sched_class = &rt_sched_class;
3700 p->sched_class = &fair_sched_class;
3705 * check the target process has a UID that matches the current process's
3707 static bool check_same_owner(struct task_struct *p)
3709 const struct cred *cred = current_cred(), *pcred;
3713 pcred = __task_cred(p);
3714 match = (uid_eq(cred->euid, pcred->euid) ||
3715 uid_eq(cred->euid, pcred->uid));
3720 static int __sched_setscheduler(struct task_struct *p, int policy,
3721 const struct sched_param *param, bool user)
3723 int retval, oldprio, oldpolicy = -1, on_rq, running;
3724 unsigned long flags;
3725 const struct sched_class *prev_class;
3729 /* may grab non-irq protected spin_locks */
3730 BUG_ON(in_interrupt());
3732 /* double check policy once rq lock held */
3734 reset_on_fork = p->sched_reset_on_fork;
3735 policy = oldpolicy = p->policy;
3737 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3738 policy &= ~SCHED_RESET_ON_FORK;
3740 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3741 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3742 policy != SCHED_IDLE)
3747 * Valid priorities for SCHED_FIFO and SCHED_RR are
3748 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3749 * SCHED_BATCH and SCHED_IDLE is 0.
3751 if (param->sched_priority < 0 ||
3752 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3753 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3755 if (rt_policy(policy) != (param->sched_priority != 0))
3759 * Allow unprivileged RT tasks to decrease priority:
3761 if (user && !capable(CAP_SYS_NICE)) {
3762 if (rt_policy(policy)) {
3763 unsigned long rlim_rtprio =
3764 task_rlimit(p, RLIMIT_RTPRIO);
3766 /* can't set/change the rt policy */
3767 if (policy != p->policy && !rlim_rtprio)
3770 /* can't increase priority */
3771 if (param->sched_priority > p->rt_priority &&
3772 param->sched_priority > rlim_rtprio)
3777 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3778 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3780 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3781 if (!can_nice(p, TASK_NICE(p)))
3785 /* can't change other user's priorities */
3786 if (!check_same_owner(p))
3789 /* Normal users shall not reset the sched_reset_on_fork flag */
3790 if (p->sched_reset_on_fork && !reset_on_fork)
3795 retval = security_task_setscheduler(p);
3801 * make sure no PI-waiters arrive (or leave) while we are
3802 * changing the priority of the task:
3804 * To be able to change p->policy safely, the appropriate
3805 * runqueue lock must be held.
3807 rq = task_rq_lock(p, &flags);
3810 * Changing the policy of the stop threads its a very bad idea
3812 if (p == rq->stop) {
3813 task_rq_unlock(rq, p, &flags);
3818 * If not changing anything there's no need to proceed further:
3820 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3821 param->sched_priority == p->rt_priority))) {
3822 task_rq_unlock(rq, p, &flags);
3826 #ifdef CONFIG_RT_GROUP_SCHED
3829 * Do not allow realtime tasks into groups that have no runtime
3832 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3833 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3834 !task_group_is_autogroup(task_group(p))) {
3835 task_rq_unlock(rq, p, &flags);
3841 /* recheck policy now with rq lock held */
3842 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3843 policy = oldpolicy = -1;
3844 task_rq_unlock(rq, p, &flags);
3848 running = task_current(rq, p);
3850 dequeue_task(rq, p, 0);
3852 p->sched_class->put_prev_task(rq, p);
3854 p->sched_reset_on_fork = reset_on_fork;
3857 prev_class = p->sched_class;
3858 __setscheduler(rq, p, policy, param->sched_priority);
3861 p->sched_class->set_curr_task(rq);
3863 enqueue_task(rq, p, 0);
3865 check_class_changed(rq, p, prev_class, oldprio);
3866 task_rq_unlock(rq, p, &flags);
3868 rt_mutex_adjust_pi(p);
3874 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3875 * @p: the task in question.
3876 * @policy: new policy.
3877 * @param: structure containing the new RT priority.
3879 * NOTE that the task may be already dead.
3881 int sched_setscheduler(struct task_struct *p, int policy,
3882 const struct sched_param *param)
3884 return __sched_setscheduler(p, policy, param, true);
3886 EXPORT_SYMBOL_GPL(sched_setscheduler);
3889 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3890 * @p: the task in question.
3891 * @policy: new policy.
3892 * @param: structure containing the new RT priority.
3894 * Just like sched_setscheduler, only don't bother checking if the
3895 * current context has permission. For example, this is needed in
3896 * stop_machine(): we create temporary high priority worker threads,
3897 * but our caller might not have that capability.
3899 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3900 const struct sched_param *param)
3902 return __sched_setscheduler(p, policy, param, false);
3906 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3908 struct sched_param lparam;
3909 struct task_struct *p;
3912 if (!param || pid < 0)
3914 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3919 p = find_process_by_pid(pid);
3921 retval = sched_setscheduler(p, policy, &lparam);
3928 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3929 * @pid: the pid in question.
3930 * @policy: new policy.
3931 * @param: structure containing the new RT priority.
3933 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3934 struct sched_param __user *, param)
3936 /* negative values for policy are not valid */
3940 return do_sched_setscheduler(pid, policy, param);
3944 * sys_sched_setparam - set/change the RT priority of a thread
3945 * @pid: the pid in question.
3946 * @param: structure containing the new RT priority.
3948 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3950 return do_sched_setscheduler(pid, -1, param);
3954 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3955 * @pid: the pid in question.
3957 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3959 struct task_struct *p;
3967 p = find_process_by_pid(pid);
3969 retval = security_task_getscheduler(p);
3972 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3979 * sys_sched_getparam - get the RT priority of a thread
3980 * @pid: the pid in question.
3981 * @param: structure containing the RT priority.
3983 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3985 struct sched_param lp;
3986 struct task_struct *p;
3989 if (!param || pid < 0)
3993 p = find_process_by_pid(pid);
3998 retval = security_task_getscheduler(p);
4002 lp.sched_priority = p->rt_priority;
4006 * This one might sleep, we cannot do it with a spinlock held ...
4008 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4017 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4019 cpumask_var_t cpus_allowed, new_mask;
4020 struct task_struct *p;
4026 p = find_process_by_pid(pid);
4033 /* Prevent p going away */
4037 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4041 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4043 goto out_free_cpus_allowed;
4046 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4049 retval = security_task_setscheduler(p);
4053 cpuset_cpus_allowed(p, cpus_allowed);
4054 cpumask_and(new_mask, in_mask, cpus_allowed);
4056 retval = set_cpus_allowed_ptr(p, new_mask);
4059 cpuset_cpus_allowed(p, cpus_allowed);
4060 if (!cpumask_subset(new_mask, cpus_allowed)) {
4062 * We must have raced with a concurrent cpuset
4063 * update. Just reset the cpus_allowed to the
4064 * cpuset's cpus_allowed
4066 cpumask_copy(new_mask, cpus_allowed);
4071 free_cpumask_var(new_mask);
4072 out_free_cpus_allowed:
4073 free_cpumask_var(cpus_allowed);
4080 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4081 struct cpumask *new_mask)
4083 if (len < cpumask_size())
4084 cpumask_clear(new_mask);
4085 else if (len > cpumask_size())
4086 len = cpumask_size();
4088 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4092 * sys_sched_setaffinity - set the cpu affinity of a process
4093 * @pid: pid of the process
4094 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4095 * @user_mask_ptr: user-space pointer to the new cpu mask
4097 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4098 unsigned long __user *, user_mask_ptr)
4100 cpumask_var_t new_mask;
4103 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4106 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4108 retval = sched_setaffinity(pid, new_mask);
4109 free_cpumask_var(new_mask);
4113 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4115 struct task_struct *p;
4116 unsigned long flags;
4123 p = find_process_by_pid(pid);
4127 retval = security_task_getscheduler(p);
4131 raw_spin_lock_irqsave(&p->pi_lock, flags);
4132 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4133 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4143 * sys_sched_getaffinity - get the cpu affinity of a process
4144 * @pid: pid of the process
4145 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4146 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4148 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4149 unsigned long __user *, user_mask_ptr)
4154 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4156 if (len & (sizeof(unsigned long)-1))
4159 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4162 ret = sched_getaffinity(pid, mask);
4164 size_t retlen = min_t(size_t, len, cpumask_size());
4166 if (copy_to_user(user_mask_ptr, mask, retlen))
4171 free_cpumask_var(mask);
4177 * sys_sched_yield - yield the current processor to other threads.
4179 * This function yields the current CPU to other tasks. If there are no
4180 * other threads running on this CPU then this function will return.
4182 SYSCALL_DEFINE0(sched_yield)
4184 struct rq *rq = this_rq_lock();
4186 schedstat_inc(rq, yld_count);
4187 current->sched_class->yield_task(rq);
4190 * Since we are going to call schedule() anyway, there's
4191 * no need to preempt or enable interrupts:
4193 __release(rq->lock);
4194 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4195 do_raw_spin_unlock(&rq->lock);
4196 sched_preempt_enable_no_resched();
4203 static inline int should_resched(void)
4205 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4208 static void __cond_resched(void)
4210 add_preempt_count(PREEMPT_ACTIVE);
4212 sub_preempt_count(PREEMPT_ACTIVE);
4215 int __sched _cond_resched(void)
4217 if (should_resched()) {
4223 EXPORT_SYMBOL(_cond_resched);
4226 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4227 * call schedule, and on return reacquire the lock.
4229 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4230 * operations here to prevent schedule() from being called twice (once via
4231 * spin_unlock(), once by hand).
4233 int __cond_resched_lock(spinlock_t *lock)
4235 int resched = should_resched();
4238 lockdep_assert_held(lock);
4240 if (spin_needbreak(lock) || resched) {
4251 EXPORT_SYMBOL(__cond_resched_lock);
4253 int __sched __cond_resched_softirq(void)
4255 BUG_ON(!in_softirq());
4257 if (should_resched()) {
4265 EXPORT_SYMBOL(__cond_resched_softirq);
4268 * yield - yield the current processor to other threads.
4270 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4272 * The scheduler is at all times free to pick the calling task as the most
4273 * eligible task to run, if removing the yield() call from your code breaks
4274 * it, its already broken.
4276 * Typical broken usage is:
4281 * where one assumes that yield() will let 'the other' process run that will
4282 * make event true. If the current task is a SCHED_FIFO task that will never
4283 * happen. Never use yield() as a progress guarantee!!
4285 * If you want to use yield() to wait for something, use wait_event().
4286 * If you want to use yield() to be 'nice' for others, use cond_resched().
4287 * If you still want to use yield(), do not!
4289 void __sched yield(void)
4291 set_current_state(TASK_RUNNING);
4294 EXPORT_SYMBOL(yield);
4297 * yield_to - yield the current processor to another thread in
4298 * your thread group, or accelerate that thread toward the
4299 * processor it's on.
4301 * @preempt: whether task preemption is allowed or not
4303 * It's the caller's job to ensure that the target task struct
4304 * can't go away on us before we can do any checks.
4306 * Returns true if we indeed boosted the target task.
4308 bool __sched yield_to(struct task_struct *p, bool preempt)
4310 struct task_struct *curr = current;
4311 struct rq *rq, *p_rq;
4312 unsigned long flags;
4315 local_irq_save(flags);
4320 double_rq_lock(rq, p_rq);
4321 while (task_rq(p) != p_rq) {
4322 double_rq_unlock(rq, p_rq);
4326 if (!curr->sched_class->yield_to_task)
4329 if (curr->sched_class != p->sched_class)
4332 if (task_running(p_rq, p) || p->state)
4335 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4337 schedstat_inc(rq, yld_count);
4339 * Make p's CPU reschedule; pick_next_entity takes care of
4342 if (preempt && rq != p_rq)
4343 resched_task(p_rq->curr);
4347 double_rq_unlock(rq, p_rq);
4348 local_irq_restore(flags);
4355 EXPORT_SYMBOL_GPL(yield_to);
4358 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4359 * that process accounting knows that this is a task in IO wait state.
4361 void __sched io_schedule(void)
4363 struct rq *rq = raw_rq();
4365 delayacct_blkio_start();
4366 atomic_inc(&rq->nr_iowait);
4367 blk_flush_plug(current);
4368 current->in_iowait = 1;
4370 current->in_iowait = 0;
4371 atomic_dec(&rq->nr_iowait);
4372 delayacct_blkio_end();
4374 EXPORT_SYMBOL(io_schedule);
4376 long __sched io_schedule_timeout(long timeout)
4378 struct rq *rq = raw_rq();
4381 delayacct_blkio_start();
4382 atomic_inc(&rq->nr_iowait);
4383 blk_flush_plug(current);
4384 current->in_iowait = 1;
4385 ret = schedule_timeout(timeout);
4386 current->in_iowait = 0;
4387 atomic_dec(&rq->nr_iowait);
4388 delayacct_blkio_end();
4393 * sys_sched_get_priority_max - return maximum RT priority.
4394 * @policy: scheduling class.
4396 * this syscall returns the maximum rt_priority that can be used
4397 * by a given scheduling class.
4399 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4406 ret = MAX_USER_RT_PRIO-1;
4418 * sys_sched_get_priority_min - return minimum RT priority.
4419 * @policy: scheduling class.
4421 * this syscall returns the minimum rt_priority that can be used
4422 * by a given scheduling class.
4424 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4442 * sys_sched_rr_get_interval - return the default timeslice of a process.
4443 * @pid: pid of the process.
4444 * @interval: userspace pointer to the timeslice value.
4446 * this syscall writes the default timeslice value of a given process
4447 * into the user-space timespec buffer. A value of '0' means infinity.
4449 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4450 struct timespec __user *, interval)
4452 struct task_struct *p;
4453 unsigned int time_slice;
4454 unsigned long flags;
4464 p = find_process_by_pid(pid);
4468 retval = security_task_getscheduler(p);
4472 rq = task_rq_lock(p, &flags);
4473 time_slice = p->sched_class->get_rr_interval(rq, p);
4474 task_rq_unlock(rq, p, &flags);
4477 jiffies_to_timespec(time_slice, &t);
4478 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4486 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4488 void sched_show_task(struct task_struct *p)
4490 unsigned long free = 0;
4493 state = p->state ? __ffs(p->state) + 1 : 0;
4494 printk(KERN_INFO "%-15.15s %c", p->comm,
4495 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4496 #if BITS_PER_LONG == 32
4497 if (state == TASK_RUNNING)
4498 printk(KERN_CONT " running ");
4500 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4502 if (state == TASK_RUNNING)
4503 printk(KERN_CONT " running task ");
4505 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4507 #ifdef CONFIG_DEBUG_STACK_USAGE
4508 free = stack_not_used(p);
4510 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4511 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4512 (unsigned long)task_thread_info(p)->flags);
4514 show_stack(p, NULL);
4517 void show_state_filter(unsigned long state_filter)
4519 struct task_struct *g, *p;
4521 #if BITS_PER_LONG == 32
4523 " task PC stack pid father\n");
4526 " task PC stack pid father\n");
4529 do_each_thread(g, p) {
4531 * reset the NMI-timeout, listing all files on a slow
4532 * console might take a lot of time:
4534 touch_nmi_watchdog();
4535 if (!state_filter || (p->state & state_filter))
4537 } while_each_thread(g, p);
4539 touch_all_softlockup_watchdogs();
4541 #ifdef CONFIG_SCHED_DEBUG
4542 sysrq_sched_debug_show();
4546 * Only show locks if all tasks are dumped:
4549 debug_show_all_locks();
4552 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4554 idle->sched_class = &idle_sched_class;
4558 * init_idle - set up an idle thread for a given CPU
4559 * @idle: task in question
4560 * @cpu: cpu the idle task belongs to
4562 * NOTE: this function does not set the idle thread's NEED_RESCHED
4563 * flag, to make booting more robust.
4565 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4567 struct rq *rq = cpu_rq(cpu);
4568 unsigned long flags;
4570 raw_spin_lock_irqsave(&rq->lock, flags);
4573 idle->state = TASK_RUNNING;
4574 idle->se.exec_start = sched_clock();
4576 do_set_cpus_allowed(idle, cpumask_of(cpu));
4578 * We're having a chicken and egg problem, even though we are
4579 * holding rq->lock, the cpu isn't yet set to this cpu so the
4580 * lockdep check in task_group() will fail.
4582 * Similar case to sched_fork(). / Alternatively we could
4583 * use task_rq_lock() here and obtain the other rq->lock.
4588 __set_task_cpu(idle, cpu);
4591 rq->curr = rq->idle = idle;
4592 #if defined(CONFIG_SMP)
4595 raw_spin_unlock_irqrestore(&rq->lock, flags);
4597 /* Set the preempt count _outside_ the spinlocks! */
4598 task_thread_info(idle)->preempt_count = 0;
4601 * The idle tasks have their own, simple scheduling class:
4603 idle->sched_class = &idle_sched_class;
4604 ftrace_graph_init_idle_task(idle, cpu);
4605 #if defined(CONFIG_SMP)
4606 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4611 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4613 if (p->sched_class && p->sched_class->set_cpus_allowed)
4614 p->sched_class->set_cpus_allowed(p, new_mask);
4616 cpumask_copy(&p->cpus_allowed, new_mask);
4617 p->nr_cpus_allowed = cpumask_weight(new_mask);
4621 * This is how migration works:
4623 * 1) we invoke migration_cpu_stop() on the target CPU using
4625 * 2) stopper starts to run (implicitly forcing the migrated thread
4627 * 3) it checks whether the migrated task is still in the wrong runqueue.
4628 * 4) if it's in the wrong runqueue then the migration thread removes
4629 * it and puts it into the right queue.
4630 * 5) stopper completes and stop_one_cpu() returns and the migration
4635 * Change a given task's CPU affinity. Migrate the thread to a
4636 * proper CPU and schedule it away if the CPU it's executing on
4637 * is removed from the allowed bitmask.
4639 * NOTE: the caller must have a valid reference to the task, the
4640 * task must not exit() & deallocate itself prematurely. The
4641 * call is not atomic; no spinlocks may be held.
4643 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4645 unsigned long flags;
4647 unsigned int dest_cpu;
4650 rq = task_rq_lock(p, &flags);
4652 if (cpumask_equal(&p->cpus_allowed, new_mask))
4655 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4660 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4665 do_set_cpus_allowed(p, new_mask);
4667 /* Can the task run on the task's current CPU? If so, we're done */
4668 if (cpumask_test_cpu(task_cpu(p), new_mask))
4671 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4673 struct migration_arg arg = { p, dest_cpu };
4674 /* Need help from migration thread: drop lock and wait. */
4675 task_rq_unlock(rq, p, &flags);
4676 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4677 tlb_migrate_finish(p->mm);
4681 task_rq_unlock(rq, p, &flags);
4685 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4688 * Move (not current) task off this cpu, onto dest cpu. We're doing
4689 * this because either it can't run here any more (set_cpus_allowed()
4690 * away from this CPU, or CPU going down), or because we're
4691 * attempting to rebalance this task on exec (sched_exec).
4693 * So we race with normal scheduler movements, but that's OK, as long
4694 * as the task is no longer on this CPU.
4696 * Returns non-zero if task was successfully migrated.
4698 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4700 struct rq *rq_dest, *rq_src;
4703 if (unlikely(!cpu_active(dest_cpu)))
4706 rq_src = cpu_rq(src_cpu);
4707 rq_dest = cpu_rq(dest_cpu);
4709 raw_spin_lock(&p->pi_lock);
4710 double_rq_lock(rq_src, rq_dest);
4711 /* Already moved. */
4712 if (task_cpu(p) != src_cpu)
4714 /* Affinity changed (again). */
4715 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4719 * If we're not on a rq, the next wake-up will ensure we're
4723 dequeue_task(rq_src, p, 0);
4724 set_task_cpu(p, dest_cpu);
4725 enqueue_task(rq_dest, p, 0);
4726 check_preempt_curr(rq_dest, p, 0);
4731 double_rq_unlock(rq_src, rq_dest);
4732 raw_spin_unlock(&p->pi_lock);
4737 * migration_cpu_stop - this will be executed by a highprio stopper thread
4738 * and performs thread migration by bumping thread off CPU then
4739 * 'pushing' onto another runqueue.
4741 static int migration_cpu_stop(void *data)
4743 struct migration_arg *arg = data;
4746 * The original target cpu might have gone down and we might
4747 * be on another cpu but it doesn't matter.
4749 local_irq_disable();
4750 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4755 #ifdef CONFIG_HOTPLUG_CPU
4758 * Ensures that the idle task is using init_mm right before its cpu goes
4761 void idle_task_exit(void)
4763 struct mm_struct *mm = current->active_mm;
4765 BUG_ON(cpu_online(smp_processor_id()));
4768 switch_mm(mm, &init_mm, current);
4773 * Since this CPU is going 'away' for a while, fold any nr_active delta
4774 * we might have. Assumes we're called after migrate_tasks() so that the
4775 * nr_active count is stable.
4777 * Also see the comment "Global load-average calculations".
4779 static void calc_load_migrate(struct rq *rq)
4781 long delta = calc_load_fold_active(rq);
4783 atomic_long_add(delta, &calc_load_tasks);
4787 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4788 * try_to_wake_up()->select_task_rq().
4790 * Called with rq->lock held even though we'er in stop_machine() and
4791 * there's no concurrency possible, we hold the required locks anyway
4792 * because of lock validation efforts.
4794 static void migrate_tasks(unsigned int dead_cpu)
4796 struct rq *rq = cpu_rq(dead_cpu);
4797 struct task_struct *next, *stop = rq->stop;
4801 * Fudge the rq selection such that the below task selection loop
4802 * doesn't get stuck on the currently eligible stop task.
4804 * We're currently inside stop_machine() and the rq is either stuck
4805 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4806 * either way we should never end up calling schedule() until we're
4813 * There's this thread running, bail when that's the only
4816 if (rq->nr_running == 1)
4819 next = pick_next_task(rq);
4821 next->sched_class->put_prev_task(rq, next);
4823 /* Find suitable destination for @next, with force if needed. */
4824 dest_cpu = select_fallback_rq(dead_cpu, next);
4825 raw_spin_unlock(&rq->lock);
4827 __migrate_task(next, dead_cpu, dest_cpu);
4829 raw_spin_lock(&rq->lock);
4835 #endif /* CONFIG_HOTPLUG_CPU */
4837 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4839 static struct ctl_table sd_ctl_dir[] = {
4841 .procname = "sched_domain",
4847 static struct ctl_table sd_ctl_root[] = {
4849 .procname = "kernel",
4851 .child = sd_ctl_dir,
4856 static struct ctl_table *sd_alloc_ctl_entry(int n)
4858 struct ctl_table *entry =
4859 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4864 static void sd_free_ctl_entry(struct ctl_table **tablep)
4866 struct ctl_table *entry;
4869 * In the intermediate directories, both the child directory and
4870 * procname are dynamically allocated and could fail but the mode
4871 * will always be set. In the lowest directory the names are
4872 * static strings and all have proc handlers.
4874 for (entry = *tablep; entry->mode; entry++) {
4876 sd_free_ctl_entry(&entry->child);
4877 if (entry->proc_handler == NULL)
4878 kfree(entry->procname);
4885 static int min_load_idx = 0;
4886 static int max_load_idx = CPU_LOAD_IDX_MAX;
4889 set_table_entry(struct ctl_table *entry,
4890 const char *procname, void *data, int maxlen,
4891 umode_t mode, proc_handler *proc_handler,
4894 entry->procname = procname;
4896 entry->maxlen = maxlen;
4898 entry->proc_handler = proc_handler;
4901 entry->extra1 = &min_load_idx;
4902 entry->extra2 = &max_load_idx;
4906 static struct ctl_table *
4907 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4909 struct ctl_table *table = sd_alloc_ctl_entry(13);
4914 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4915 sizeof(long), 0644, proc_doulongvec_minmax, false);
4916 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4917 sizeof(long), 0644, proc_doulongvec_minmax, false);
4918 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4919 sizeof(int), 0644, proc_dointvec_minmax, true);
4920 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4921 sizeof(int), 0644, proc_dointvec_minmax, true);
4922 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4923 sizeof(int), 0644, proc_dointvec_minmax, true);
4924 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4925 sizeof(int), 0644, proc_dointvec_minmax, true);
4926 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4927 sizeof(int), 0644, proc_dointvec_minmax, true);
4928 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4929 sizeof(int), 0644, proc_dointvec_minmax, false);
4930 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4931 sizeof(int), 0644, proc_dointvec_minmax, false);
4932 set_table_entry(&table[9], "cache_nice_tries",
4933 &sd->cache_nice_tries,
4934 sizeof(int), 0644, proc_dointvec_minmax, false);
4935 set_table_entry(&table[10], "flags", &sd->flags,
4936 sizeof(int), 0644, proc_dointvec_minmax, false);
4937 set_table_entry(&table[11], "name", sd->name,
4938 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4939 /* &table[12] is terminator */
4944 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4946 struct ctl_table *entry, *table;
4947 struct sched_domain *sd;
4948 int domain_num = 0, i;
4951 for_each_domain(cpu, sd)
4953 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4958 for_each_domain(cpu, sd) {
4959 snprintf(buf, 32, "domain%d", i);
4960 entry->procname = kstrdup(buf, GFP_KERNEL);
4962 entry->child = sd_alloc_ctl_domain_table(sd);
4969 static struct ctl_table_header *sd_sysctl_header;
4970 static void register_sched_domain_sysctl(void)
4972 int i, cpu_num = num_possible_cpus();
4973 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4976 WARN_ON(sd_ctl_dir[0].child);
4977 sd_ctl_dir[0].child = entry;
4982 for_each_possible_cpu(i) {
4983 snprintf(buf, 32, "cpu%d", i);
4984 entry->procname = kstrdup(buf, GFP_KERNEL);
4986 entry->child = sd_alloc_ctl_cpu_table(i);
4990 WARN_ON(sd_sysctl_header);
4991 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4994 /* may be called multiple times per register */
4995 static void unregister_sched_domain_sysctl(void)
4997 if (sd_sysctl_header)
4998 unregister_sysctl_table(sd_sysctl_header);
4999 sd_sysctl_header = NULL;
5000 if (sd_ctl_dir[0].child)
5001 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5004 static void register_sched_domain_sysctl(void)
5007 static void unregister_sched_domain_sysctl(void)
5012 static void set_rq_online(struct rq *rq)
5015 const struct sched_class *class;
5017 cpumask_set_cpu(rq->cpu, rq->rd->online);
5020 for_each_class(class) {
5021 if (class->rq_online)
5022 class->rq_online(rq);
5027 static void set_rq_offline(struct rq *rq)
5030 const struct sched_class *class;
5032 for_each_class(class) {
5033 if (class->rq_offline)
5034 class->rq_offline(rq);
5037 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5043 * migration_call - callback that gets triggered when a CPU is added.
5044 * Here we can start up the necessary migration thread for the new CPU.
5046 static int __cpuinit
5047 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5049 int cpu = (long)hcpu;
5050 unsigned long flags;
5051 struct rq *rq = cpu_rq(cpu);
5053 switch (action & ~CPU_TASKS_FROZEN) {
5055 case CPU_UP_PREPARE:
5056 rq->calc_load_update = calc_load_update;
5060 /* Update our root-domain */
5061 raw_spin_lock_irqsave(&rq->lock, flags);
5063 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5067 raw_spin_unlock_irqrestore(&rq->lock, flags);
5070 #ifdef CONFIG_HOTPLUG_CPU
5072 sched_ttwu_pending();
5073 /* Update our root-domain */
5074 raw_spin_lock_irqsave(&rq->lock, flags);
5076 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5080 BUG_ON(rq->nr_running != 1); /* the migration thread */
5081 raw_spin_unlock_irqrestore(&rq->lock, flags);
5085 calc_load_migrate(rq);
5090 update_max_interval();
5096 * Register at high priority so that task migration (migrate_all_tasks)
5097 * happens before everything else. This has to be lower priority than
5098 * the notifier in the perf_event subsystem, though.
5100 static struct notifier_block __cpuinitdata migration_notifier = {
5101 .notifier_call = migration_call,
5102 .priority = CPU_PRI_MIGRATION,
5105 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5106 unsigned long action, void *hcpu)
5108 switch (action & ~CPU_TASKS_FROZEN) {
5110 case CPU_DOWN_FAILED:
5111 set_cpu_active((long)hcpu, true);
5118 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5119 unsigned long action, void *hcpu)
5121 switch (action & ~CPU_TASKS_FROZEN) {
5122 case CPU_DOWN_PREPARE:
5123 set_cpu_active((long)hcpu, false);
5130 static int __init migration_init(void)
5132 void *cpu = (void *)(long)smp_processor_id();
5135 /* Initialize migration for the boot CPU */
5136 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5137 BUG_ON(err == NOTIFY_BAD);
5138 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5139 register_cpu_notifier(&migration_notifier);
5141 /* Register cpu active notifiers */
5142 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5143 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5147 early_initcall(migration_init);
5152 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5154 #ifdef CONFIG_SCHED_DEBUG
5156 static __read_mostly int sched_debug_enabled;
5158 static int __init sched_debug_setup(char *str)
5160 sched_debug_enabled = 1;
5164 early_param("sched_debug", sched_debug_setup);
5166 static inline bool sched_debug(void)
5168 return sched_debug_enabled;
5171 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5172 struct cpumask *groupmask)
5174 struct sched_group *group = sd->groups;
5177 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5178 cpumask_clear(groupmask);
5180 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5182 if (!(sd->flags & SD_LOAD_BALANCE)) {
5183 printk("does not load-balance\n");
5185 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5190 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5192 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5193 printk(KERN_ERR "ERROR: domain->span does not contain "
5196 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5197 printk(KERN_ERR "ERROR: domain->groups does not contain"
5201 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5205 printk(KERN_ERR "ERROR: group is NULL\n");
5210 * Even though we initialize ->power to something semi-sane,
5211 * we leave power_orig unset. This allows us to detect if
5212 * domain iteration is still funny without causing /0 traps.
5214 if (!group->sgp->power_orig) {
5215 printk(KERN_CONT "\n");
5216 printk(KERN_ERR "ERROR: domain->cpu_power not "
5221 if (!cpumask_weight(sched_group_cpus(group))) {
5222 printk(KERN_CONT "\n");
5223 printk(KERN_ERR "ERROR: empty group\n");
5227 if (!(sd->flags & SD_OVERLAP) &&
5228 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5229 printk(KERN_CONT "\n");
5230 printk(KERN_ERR "ERROR: repeated CPUs\n");
5234 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5236 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5238 printk(KERN_CONT " %s", str);
5239 if (group->sgp->power != SCHED_POWER_SCALE) {
5240 printk(KERN_CONT " (cpu_power = %d)",
5244 group = group->next;
5245 } while (group != sd->groups);
5246 printk(KERN_CONT "\n");
5248 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5249 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5252 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5253 printk(KERN_ERR "ERROR: parent span is not a superset "
5254 "of domain->span\n");
5258 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5262 if (!sched_debug_enabled)
5266 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5270 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5273 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5281 #else /* !CONFIG_SCHED_DEBUG */
5282 # define sched_domain_debug(sd, cpu) do { } while (0)
5283 static inline bool sched_debug(void)
5287 #endif /* CONFIG_SCHED_DEBUG */
5289 static int sd_degenerate(struct sched_domain *sd)
5291 if (cpumask_weight(sched_domain_span(sd)) == 1)
5294 /* Following flags need at least 2 groups */
5295 if (sd->flags & (SD_LOAD_BALANCE |
5296 SD_BALANCE_NEWIDLE |
5300 SD_SHARE_PKG_RESOURCES)) {
5301 if (sd->groups != sd->groups->next)
5305 /* Following flags don't use groups */
5306 if (sd->flags & (SD_WAKE_AFFINE))
5313 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5315 unsigned long cflags = sd->flags, pflags = parent->flags;
5317 if (sd_degenerate(parent))
5320 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5323 /* Flags needing groups don't count if only 1 group in parent */
5324 if (parent->groups == parent->groups->next) {
5325 pflags &= ~(SD_LOAD_BALANCE |
5326 SD_BALANCE_NEWIDLE |
5330 SD_SHARE_PKG_RESOURCES);
5331 if (nr_node_ids == 1)
5332 pflags &= ~SD_SERIALIZE;
5334 if (~cflags & pflags)
5340 static void free_rootdomain(struct rcu_head *rcu)
5342 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5344 cpupri_cleanup(&rd->cpupri);
5345 free_cpumask_var(rd->rto_mask);
5346 free_cpumask_var(rd->online);
5347 free_cpumask_var(rd->span);
5351 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5353 struct root_domain *old_rd = NULL;
5354 unsigned long flags;
5356 raw_spin_lock_irqsave(&rq->lock, flags);
5361 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5364 cpumask_clear_cpu(rq->cpu, old_rd->span);
5367 * If we dont want to free the old_rt yet then
5368 * set old_rd to NULL to skip the freeing later
5371 if (!atomic_dec_and_test(&old_rd->refcount))
5375 atomic_inc(&rd->refcount);
5378 cpumask_set_cpu(rq->cpu, rd->span);
5379 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5382 raw_spin_unlock_irqrestore(&rq->lock, flags);
5385 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5388 static int init_rootdomain(struct root_domain *rd)
5390 memset(rd, 0, sizeof(*rd));
5392 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5394 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5396 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5399 if (cpupri_init(&rd->cpupri) != 0)
5404 free_cpumask_var(rd->rto_mask);
5406 free_cpumask_var(rd->online);
5408 free_cpumask_var(rd->span);
5414 * By default the system creates a single root-domain with all cpus as
5415 * members (mimicking the global state we have today).
5417 struct root_domain def_root_domain;
5419 static void init_defrootdomain(void)
5421 init_rootdomain(&def_root_domain);
5423 atomic_set(&def_root_domain.refcount, 1);
5426 static struct root_domain *alloc_rootdomain(void)
5428 struct root_domain *rd;
5430 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5434 if (init_rootdomain(rd) != 0) {
5442 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5444 struct sched_group *tmp, *first;
5453 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5458 } while (sg != first);
5461 static void free_sched_domain(struct rcu_head *rcu)
5463 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5466 * If its an overlapping domain it has private groups, iterate and
5469 if (sd->flags & SD_OVERLAP) {
5470 free_sched_groups(sd->groups, 1);
5471 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5472 kfree(sd->groups->sgp);
5478 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5480 call_rcu(&sd->rcu, free_sched_domain);
5483 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5485 for (; sd; sd = sd->parent)
5486 destroy_sched_domain(sd, cpu);
5490 * Keep a special pointer to the highest sched_domain that has
5491 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5492 * allows us to avoid some pointer chasing select_idle_sibling().
5494 * Also keep a unique ID per domain (we use the first cpu number in
5495 * the cpumask of the domain), this allows us to quickly tell if
5496 * two cpus are in the same cache domain, see cpus_share_cache().
5498 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5499 DEFINE_PER_CPU(int, sd_llc_id);
5501 static void update_top_cache_domain(int cpu)
5503 struct sched_domain *sd;
5506 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5508 id = cpumask_first(sched_domain_span(sd));
5510 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5511 per_cpu(sd_llc_id, cpu) = id;
5515 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5516 * hold the hotplug lock.
5519 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5521 struct rq *rq = cpu_rq(cpu);
5522 struct sched_domain *tmp;
5524 /* Remove the sched domains which do not contribute to scheduling. */
5525 for (tmp = sd; tmp; ) {
5526 struct sched_domain *parent = tmp->parent;
5530 if (sd_parent_degenerate(tmp, parent)) {
5531 tmp->parent = parent->parent;
5533 parent->parent->child = tmp;
5534 destroy_sched_domain(parent, cpu);
5539 if (sd && sd_degenerate(sd)) {
5542 destroy_sched_domain(tmp, cpu);
5547 sched_domain_debug(sd, cpu);
5549 rq_attach_root(rq, rd);
5551 rcu_assign_pointer(rq->sd, sd);
5552 destroy_sched_domains(tmp, cpu);
5554 update_top_cache_domain(cpu);
5557 /* cpus with isolated domains */
5558 static cpumask_var_t cpu_isolated_map;
5560 /* Setup the mask of cpus configured for isolated domains */
5561 static int __init isolated_cpu_setup(char *str)
5563 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5564 cpulist_parse(str, cpu_isolated_map);
5568 __setup("isolcpus=", isolated_cpu_setup);
5570 static const struct cpumask *cpu_cpu_mask(int cpu)
5572 return cpumask_of_node(cpu_to_node(cpu));
5576 struct sched_domain **__percpu sd;
5577 struct sched_group **__percpu sg;
5578 struct sched_group_power **__percpu sgp;
5582 struct sched_domain ** __percpu sd;
5583 struct root_domain *rd;
5593 struct sched_domain_topology_level;
5595 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5596 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5598 #define SDTL_OVERLAP 0x01
5600 struct sched_domain_topology_level {
5601 sched_domain_init_f init;
5602 sched_domain_mask_f mask;
5605 struct sd_data data;
5609 * Build an iteration mask that can exclude certain CPUs from the upwards
5612 * Asymmetric node setups can result in situations where the domain tree is of
5613 * unequal depth, make sure to skip domains that already cover the entire
5616 * In that case build_sched_domains() will have terminated the iteration early
5617 * and our sibling sd spans will be empty. Domains should always include the
5618 * cpu they're built on, so check that.
5621 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5623 const struct cpumask *span = sched_domain_span(sd);
5624 struct sd_data *sdd = sd->private;
5625 struct sched_domain *sibling;
5628 for_each_cpu(i, span) {
5629 sibling = *per_cpu_ptr(sdd->sd, i);
5630 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5633 cpumask_set_cpu(i, sched_group_mask(sg));
5638 * Return the canonical balance cpu for this group, this is the first cpu
5639 * of this group that's also in the iteration mask.
5641 int group_balance_cpu(struct sched_group *sg)
5643 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5647 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5649 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5650 const struct cpumask *span = sched_domain_span(sd);
5651 struct cpumask *covered = sched_domains_tmpmask;
5652 struct sd_data *sdd = sd->private;
5653 struct sched_domain *child;
5656 cpumask_clear(covered);
5658 for_each_cpu(i, span) {
5659 struct cpumask *sg_span;
5661 if (cpumask_test_cpu(i, covered))
5664 child = *per_cpu_ptr(sdd->sd, i);
5666 /* See the comment near build_group_mask(). */
5667 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5670 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5671 GFP_KERNEL, cpu_to_node(cpu));
5676 sg_span = sched_group_cpus(sg);
5678 child = child->child;
5679 cpumask_copy(sg_span, sched_domain_span(child));
5681 cpumask_set_cpu(i, sg_span);
5683 cpumask_or(covered, covered, sg_span);
5685 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5686 if (atomic_inc_return(&sg->sgp->ref) == 1)
5687 build_group_mask(sd, sg);
5690 * Initialize sgp->power such that even if we mess up the
5691 * domains and no possible iteration will get us here, we won't
5694 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5697 * Make sure the first group of this domain contains the
5698 * canonical balance cpu. Otherwise the sched_domain iteration
5699 * breaks. See update_sg_lb_stats().
5701 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5702 group_balance_cpu(sg) == cpu)
5712 sd->groups = groups;
5717 free_sched_groups(first, 0);
5722 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5724 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5725 struct sched_domain *child = sd->child;
5728 cpu = cpumask_first(sched_domain_span(child));
5731 *sg = *per_cpu_ptr(sdd->sg, cpu);
5732 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5733 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5740 * build_sched_groups will build a circular linked list of the groups
5741 * covered by the given span, and will set each group's ->cpumask correctly,
5742 * and ->cpu_power to 0.
5744 * Assumes the sched_domain tree is fully constructed
5747 build_sched_groups(struct sched_domain *sd, int cpu)
5749 struct sched_group *first = NULL, *last = NULL;
5750 struct sd_data *sdd = sd->private;
5751 const struct cpumask *span = sched_domain_span(sd);
5752 struct cpumask *covered;
5755 get_group(cpu, sdd, &sd->groups);
5756 atomic_inc(&sd->groups->ref);
5758 if (cpu != cpumask_first(sched_domain_span(sd)))
5761 lockdep_assert_held(&sched_domains_mutex);
5762 covered = sched_domains_tmpmask;
5764 cpumask_clear(covered);
5766 for_each_cpu(i, span) {
5767 struct sched_group *sg;
5768 int group = get_group(i, sdd, &sg);
5771 if (cpumask_test_cpu(i, covered))
5774 cpumask_clear(sched_group_cpus(sg));
5776 cpumask_setall(sched_group_mask(sg));
5778 for_each_cpu(j, span) {
5779 if (get_group(j, sdd, NULL) != group)
5782 cpumask_set_cpu(j, covered);
5783 cpumask_set_cpu(j, sched_group_cpus(sg));
5798 * Initialize sched groups cpu_power.
5800 * cpu_power indicates the capacity of sched group, which is used while
5801 * distributing the load between different sched groups in a sched domain.
5802 * Typically cpu_power for all the groups in a sched domain will be same unless
5803 * there are asymmetries in the topology. If there are asymmetries, group
5804 * having more cpu_power will pickup more load compared to the group having
5807 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5809 struct sched_group *sg = sd->groups;
5811 WARN_ON(!sd || !sg);
5814 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5816 } while (sg != sd->groups);
5818 if (cpu != group_balance_cpu(sg))
5821 update_group_power(sd, cpu);
5822 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5825 int __weak arch_sd_sibling_asym_packing(void)
5827 return 0*SD_ASYM_PACKING;
5831 * Initializers for schedule domains
5832 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5835 #ifdef CONFIG_SCHED_DEBUG
5836 # define SD_INIT_NAME(sd, type) sd->name = #type
5838 # define SD_INIT_NAME(sd, type) do { } while (0)
5841 #define SD_INIT_FUNC(type) \
5842 static noinline struct sched_domain * \
5843 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5845 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5846 *sd = SD_##type##_INIT; \
5847 SD_INIT_NAME(sd, type); \
5848 sd->private = &tl->data; \
5853 #ifdef CONFIG_SCHED_SMT
5854 SD_INIT_FUNC(SIBLING)
5856 #ifdef CONFIG_SCHED_MC
5859 #ifdef CONFIG_SCHED_BOOK
5863 static int default_relax_domain_level = -1;
5864 int sched_domain_level_max;
5866 static int __init setup_relax_domain_level(char *str)
5868 if (kstrtoint(str, 0, &default_relax_domain_level))
5869 pr_warn("Unable to set relax_domain_level\n");
5873 __setup("relax_domain_level=", setup_relax_domain_level);
5875 static void set_domain_attribute(struct sched_domain *sd,
5876 struct sched_domain_attr *attr)
5880 if (!attr || attr->relax_domain_level < 0) {
5881 if (default_relax_domain_level < 0)
5884 request = default_relax_domain_level;
5886 request = attr->relax_domain_level;
5887 if (request < sd->level) {
5888 /* turn off idle balance on this domain */
5889 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5891 /* turn on idle balance on this domain */
5892 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5896 static void __sdt_free(const struct cpumask *cpu_map);
5897 static int __sdt_alloc(const struct cpumask *cpu_map);
5899 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5900 const struct cpumask *cpu_map)
5904 if (!atomic_read(&d->rd->refcount))
5905 free_rootdomain(&d->rd->rcu); /* fall through */
5907 free_percpu(d->sd); /* fall through */
5909 __sdt_free(cpu_map); /* fall through */
5915 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5916 const struct cpumask *cpu_map)
5918 memset(d, 0, sizeof(*d));
5920 if (__sdt_alloc(cpu_map))
5921 return sa_sd_storage;
5922 d->sd = alloc_percpu(struct sched_domain *);
5924 return sa_sd_storage;
5925 d->rd = alloc_rootdomain();
5928 return sa_rootdomain;
5932 * NULL the sd_data elements we've used to build the sched_domain and
5933 * sched_group structure so that the subsequent __free_domain_allocs()
5934 * will not free the data we're using.
5936 static void claim_allocations(int cpu, struct sched_domain *sd)
5938 struct sd_data *sdd = sd->private;
5940 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5941 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5943 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5944 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5946 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5947 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5950 #ifdef CONFIG_SCHED_SMT
5951 static const struct cpumask *cpu_smt_mask(int cpu)
5953 return topology_thread_cpumask(cpu);
5958 * Topology list, bottom-up.
5960 static struct sched_domain_topology_level default_topology[] = {
5961 #ifdef CONFIG_SCHED_SMT
5962 { sd_init_SIBLING, cpu_smt_mask, },
5964 #ifdef CONFIG_SCHED_MC
5965 { sd_init_MC, cpu_coregroup_mask, },
5967 #ifdef CONFIG_SCHED_BOOK
5968 { sd_init_BOOK, cpu_book_mask, },
5970 { sd_init_CPU, cpu_cpu_mask, },
5974 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5978 static int sched_domains_numa_levels;
5979 static int *sched_domains_numa_distance;
5980 static struct cpumask ***sched_domains_numa_masks;
5981 static int sched_domains_curr_level;
5983 static inline int sd_local_flags(int level)
5985 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5988 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5991 static struct sched_domain *
5992 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5994 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5995 int level = tl->numa_level;
5996 int sd_weight = cpumask_weight(
5997 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5999 *sd = (struct sched_domain){
6000 .min_interval = sd_weight,
6001 .max_interval = 2*sd_weight,
6003 .imbalance_pct = 125,
6004 .cache_nice_tries = 2,
6011 .flags = 1*SD_LOAD_BALANCE
6012 | 1*SD_BALANCE_NEWIDLE
6017 | 0*SD_SHARE_CPUPOWER
6018 | 0*SD_SHARE_PKG_RESOURCES
6020 | 0*SD_PREFER_SIBLING
6021 | sd_local_flags(level)
6023 .last_balance = jiffies,
6024 .balance_interval = sd_weight,
6026 SD_INIT_NAME(sd, NUMA);
6027 sd->private = &tl->data;
6030 * Ugly hack to pass state to sd_numa_mask()...
6032 sched_domains_curr_level = tl->numa_level;
6037 static const struct cpumask *sd_numa_mask(int cpu)
6039 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6042 static void sched_numa_warn(const char *str)
6044 static int done = false;
6052 printk(KERN_WARNING "ERROR: %s\n\n", str);
6054 for (i = 0; i < nr_node_ids; i++) {
6055 printk(KERN_WARNING " ");
6056 for (j = 0; j < nr_node_ids; j++)
6057 printk(KERN_CONT "%02d ", node_distance(i,j));
6058 printk(KERN_CONT "\n");
6060 printk(KERN_WARNING "\n");
6063 static bool find_numa_distance(int distance)
6067 if (distance == node_distance(0, 0))
6070 for (i = 0; i < sched_domains_numa_levels; i++) {
6071 if (sched_domains_numa_distance[i] == distance)
6078 static void sched_init_numa(void)
6080 int next_distance, curr_distance = node_distance(0, 0);
6081 struct sched_domain_topology_level *tl;
6085 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6086 if (!sched_domains_numa_distance)
6090 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6091 * unique distances in the node_distance() table.
6093 * Assumes node_distance(0,j) includes all distances in
6094 * node_distance(i,j) in order to avoid cubic time.
6096 next_distance = curr_distance;
6097 for (i = 0; i < nr_node_ids; i++) {
6098 for (j = 0; j < nr_node_ids; j++) {
6099 for (k = 0; k < nr_node_ids; k++) {
6100 int distance = node_distance(i, k);
6102 if (distance > curr_distance &&
6103 (distance < next_distance ||
6104 next_distance == curr_distance))
6105 next_distance = distance;
6108 * While not a strong assumption it would be nice to know
6109 * about cases where if node A is connected to B, B is not
6110 * equally connected to A.
6112 if (sched_debug() && node_distance(k, i) != distance)
6113 sched_numa_warn("Node-distance not symmetric");
6115 if (sched_debug() && i && !find_numa_distance(distance))
6116 sched_numa_warn("Node-0 not representative");
6118 if (next_distance != curr_distance) {
6119 sched_domains_numa_distance[level++] = next_distance;
6120 sched_domains_numa_levels = level;
6121 curr_distance = next_distance;
6126 * In case of sched_debug() we verify the above assumption.
6132 * 'level' contains the number of unique distances, excluding the
6133 * identity distance node_distance(i,i).
6135 * The sched_domains_nume_distance[] array includes the actual distance
6140 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6141 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6142 * the array will contain less then 'level' members. This could be
6143 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6144 * in other functions.
6146 * We reset it to 'level' at the end of this function.
6148 sched_domains_numa_levels = 0;
6150 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6151 if (!sched_domains_numa_masks)
6155 * Now for each level, construct a mask per node which contains all
6156 * cpus of nodes that are that many hops away from us.
6158 for (i = 0; i < level; i++) {
6159 sched_domains_numa_masks[i] =
6160 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6161 if (!sched_domains_numa_masks[i])
6164 for (j = 0; j < nr_node_ids; j++) {
6165 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6169 sched_domains_numa_masks[i][j] = mask;
6171 for (k = 0; k < nr_node_ids; k++) {
6172 if (node_distance(j, k) > sched_domains_numa_distance[i])
6175 cpumask_or(mask, mask, cpumask_of_node(k));
6180 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6181 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6186 * Copy the default topology bits..
6188 for (i = 0; default_topology[i].init; i++)
6189 tl[i] = default_topology[i];
6192 * .. and append 'j' levels of NUMA goodness.
6194 for (j = 0; j < level; i++, j++) {
6195 tl[i] = (struct sched_domain_topology_level){
6196 .init = sd_numa_init,
6197 .mask = sd_numa_mask,
6198 .flags = SDTL_OVERLAP,
6203 sched_domain_topology = tl;
6205 sched_domains_numa_levels = level;
6208 static void sched_domains_numa_masks_set(int cpu)
6211 int node = cpu_to_node(cpu);
6213 for (i = 0; i < sched_domains_numa_levels; i++) {
6214 for (j = 0; j < nr_node_ids; j++) {
6215 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6216 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6221 static void sched_domains_numa_masks_clear(int cpu)
6224 for (i = 0; i < sched_domains_numa_levels; i++) {
6225 for (j = 0; j < nr_node_ids; j++)
6226 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6231 * Update sched_domains_numa_masks[level][node] array when new cpus
6234 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6235 unsigned long action,
6238 int cpu = (long)hcpu;
6240 switch (action & ~CPU_TASKS_FROZEN) {
6242 sched_domains_numa_masks_set(cpu);
6246 sched_domains_numa_masks_clear(cpu);
6256 static inline void sched_init_numa(void)
6260 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6261 unsigned long action,
6266 #endif /* CONFIG_NUMA */
6268 static int __sdt_alloc(const struct cpumask *cpu_map)
6270 struct sched_domain_topology_level *tl;
6273 for (tl = sched_domain_topology; tl->init; tl++) {
6274 struct sd_data *sdd = &tl->data;
6276 sdd->sd = alloc_percpu(struct sched_domain *);
6280 sdd->sg = alloc_percpu(struct sched_group *);
6284 sdd->sgp = alloc_percpu(struct sched_group_power *);
6288 for_each_cpu(j, cpu_map) {
6289 struct sched_domain *sd;
6290 struct sched_group *sg;
6291 struct sched_group_power *sgp;
6293 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6294 GFP_KERNEL, cpu_to_node(j));
6298 *per_cpu_ptr(sdd->sd, j) = sd;
6300 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6301 GFP_KERNEL, cpu_to_node(j));
6307 *per_cpu_ptr(sdd->sg, j) = sg;
6309 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6310 GFP_KERNEL, cpu_to_node(j));
6314 *per_cpu_ptr(sdd->sgp, j) = sgp;
6321 static void __sdt_free(const struct cpumask *cpu_map)
6323 struct sched_domain_topology_level *tl;
6326 for (tl = sched_domain_topology; tl->init; tl++) {
6327 struct sd_data *sdd = &tl->data;
6329 for_each_cpu(j, cpu_map) {
6330 struct sched_domain *sd;
6333 sd = *per_cpu_ptr(sdd->sd, j);
6334 if (sd && (sd->flags & SD_OVERLAP))
6335 free_sched_groups(sd->groups, 0);
6336 kfree(*per_cpu_ptr(sdd->sd, j));
6340 kfree(*per_cpu_ptr(sdd->sg, j));
6342 kfree(*per_cpu_ptr(sdd->sgp, j));
6344 free_percpu(sdd->sd);
6346 free_percpu(sdd->sg);
6348 free_percpu(sdd->sgp);
6353 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6354 struct s_data *d, const struct cpumask *cpu_map,
6355 struct sched_domain_attr *attr, struct sched_domain *child,
6358 struct sched_domain *sd = tl->init(tl, cpu);
6362 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6364 sd->level = child->level + 1;
6365 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6369 set_domain_attribute(sd, attr);
6375 * Build sched domains for a given set of cpus and attach the sched domains
6376 * to the individual cpus
6378 static int build_sched_domains(const struct cpumask *cpu_map,
6379 struct sched_domain_attr *attr)
6381 enum s_alloc alloc_state = sa_none;
6382 struct sched_domain *sd;
6384 int i, ret = -ENOMEM;
6386 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6387 if (alloc_state != sa_rootdomain)
6390 /* Set up domains for cpus specified by the cpu_map. */
6391 for_each_cpu(i, cpu_map) {
6392 struct sched_domain_topology_level *tl;
6395 for (tl = sched_domain_topology; tl->init; tl++) {
6396 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6397 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6398 sd->flags |= SD_OVERLAP;
6399 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6406 *per_cpu_ptr(d.sd, i) = sd;
6409 /* Build the groups for the domains */
6410 for_each_cpu(i, cpu_map) {
6411 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6412 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6413 if (sd->flags & SD_OVERLAP) {
6414 if (build_overlap_sched_groups(sd, i))
6417 if (build_sched_groups(sd, i))
6423 /* Calculate CPU power for physical packages and nodes */
6424 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6425 if (!cpumask_test_cpu(i, cpu_map))
6428 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6429 claim_allocations(i, sd);
6430 init_sched_groups_power(i, sd);
6434 /* Attach the domains */
6436 for_each_cpu(i, cpu_map) {
6437 sd = *per_cpu_ptr(d.sd, i);
6438 cpu_attach_domain(sd, d.rd, i);
6444 __free_domain_allocs(&d, alloc_state, cpu_map);
6448 static cpumask_var_t *doms_cur; /* current sched domains */
6449 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6450 static struct sched_domain_attr *dattr_cur;
6451 /* attribues of custom domains in 'doms_cur' */
6454 * Special case: If a kmalloc of a doms_cur partition (array of
6455 * cpumask) fails, then fallback to a single sched domain,
6456 * as determined by the single cpumask fallback_doms.
6458 static cpumask_var_t fallback_doms;
6461 * arch_update_cpu_topology lets virtualized architectures update the
6462 * cpu core maps. It is supposed to return 1 if the topology changed
6463 * or 0 if it stayed the same.
6465 int __attribute__((weak)) arch_update_cpu_topology(void)
6470 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6473 cpumask_var_t *doms;
6475 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6478 for (i = 0; i < ndoms; i++) {
6479 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6480 free_sched_domains(doms, i);
6487 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6490 for (i = 0; i < ndoms; i++)
6491 free_cpumask_var(doms[i]);
6496 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6497 * For now this just excludes isolated cpus, but could be used to
6498 * exclude other special cases in the future.
6500 static int init_sched_domains(const struct cpumask *cpu_map)
6504 arch_update_cpu_topology();
6506 doms_cur = alloc_sched_domains(ndoms_cur);
6508 doms_cur = &fallback_doms;
6509 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6510 err = build_sched_domains(doms_cur[0], NULL);
6511 register_sched_domain_sysctl();
6517 * Detach sched domains from a group of cpus specified in cpu_map
6518 * These cpus will now be attached to the NULL domain
6520 static void detach_destroy_domains(const struct cpumask *cpu_map)
6525 for_each_cpu(i, cpu_map)
6526 cpu_attach_domain(NULL, &def_root_domain, i);
6530 /* handle null as "default" */
6531 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6532 struct sched_domain_attr *new, int idx_new)
6534 struct sched_domain_attr tmp;
6541 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6542 new ? (new + idx_new) : &tmp,
6543 sizeof(struct sched_domain_attr));
6547 * Partition sched domains as specified by the 'ndoms_new'
6548 * cpumasks in the array doms_new[] of cpumasks. This compares
6549 * doms_new[] to the current sched domain partitioning, doms_cur[].
6550 * It destroys each deleted domain and builds each new domain.
6552 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6553 * The masks don't intersect (don't overlap.) We should setup one
6554 * sched domain for each mask. CPUs not in any of the cpumasks will
6555 * not be load balanced. If the same cpumask appears both in the
6556 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6559 * The passed in 'doms_new' should be allocated using
6560 * alloc_sched_domains. This routine takes ownership of it and will
6561 * free_sched_domains it when done with it. If the caller failed the
6562 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6563 * and partition_sched_domains() will fallback to the single partition
6564 * 'fallback_doms', it also forces the domains to be rebuilt.
6566 * If doms_new == NULL it will be replaced with cpu_online_mask.
6567 * ndoms_new == 0 is a special case for destroying existing domains,
6568 * and it will not create the default domain.
6570 * Call with hotplug lock held
6572 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6573 struct sched_domain_attr *dattr_new)
6578 mutex_lock(&sched_domains_mutex);
6580 /* always unregister in case we don't destroy any domains */
6581 unregister_sched_domain_sysctl();
6583 /* Let architecture update cpu core mappings. */
6584 new_topology = arch_update_cpu_topology();
6586 n = doms_new ? ndoms_new : 0;
6588 /* Destroy deleted domains */
6589 for (i = 0; i < ndoms_cur; i++) {
6590 for (j = 0; j < n && !new_topology; j++) {
6591 if (cpumask_equal(doms_cur[i], doms_new[j])
6592 && dattrs_equal(dattr_cur, i, dattr_new, j))
6595 /* no match - a current sched domain not in new doms_new[] */
6596 detach_destroy_domains(doms_cur[i]);
6601 if (doms_new == NULL) {
6603 doms_new = &fallback_doms;
6604 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6605 WARN_ON_ONCE(dattr_new);
6608 /* Build new domains */
6609 for (i = 0; i < ndoms_new; i++) {
6610 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6611 if (cpumask_equal(doms_new[i], doms_cur[j])
6612 && dattrs_equal(dattr_new, i, dattr_cur, j))
6615 /* no match - add a new doms_new */
6616 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6621 /* Remember the new sched domains */
6622 if (doms_cur != &fallback_doms)
6623 free_sched_domains(doms_cur, ndoms_cur);
6624 kfree(dattr_cur); /* kfree(NULL) is safe */
6625 doms_cur = doms_new;
6626 dattr_cur = dattr_new;
6627 ndoms_cur = ndoms_new;
6629 register_sched_domain_sysctl();
6631 mutex_unlock(&sched_domains_mutex);
6634 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6637 * Update cpusets according to cpu_active mask. If cpusets are
6638 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6639 * around partition_sched_domains().
6641 * If we come here as part of a suspend/resume, don't touch cpusets because we
6642 * want to restore it back to its original state upon resume anyway.
6644 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6648 case CPU_ONLINE_FROZEN:
6649 case CPU_DOWN_FAILED_FROZEN:
6652 * num_cpus_frozen tracks how many CPUs are involved in suspend
6653 * resume sequence. As long as this is not the last online
6654 * operation in the resume sequence, just build a single sched
6655 * domain, ignoring cpusets.
6658 if (likely(num_cpus_frozen)) {
6659 partition_sched_domains(1, NULL, NULL);
6664 * This is the last CPU online operation. So fall through and
6665 * restore the original sched domains by considering the
6666 * cpuset configurations.
6670 case CPU_DOWN_FAILED:
6671 cpuset_update_active_cpus(true);
6679 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6683 case CPU_DOWN_PREPARE:
6684 cpuset_update_active_cpus(false);
6686 case CPU_DOWN_PREPARE_FROZEN:
6688 partition_sched_domains(1, NULL, NULL);
6696 void __init sched_init_smp(void)
6698 cpumask_var_t non_isolated_cpus;
6700 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6701 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6706 mutex_lock(&sched_domains_mutex);
6707 init_sched_domains(cpu_active_mask);
6708 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6709 if (cpumask_empty(non_isolated_cpus))
6710 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6711 mutex_unlock(&sched_domains_mutex);
6714 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6715 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6716 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6718 /* RT runtime code needs to handle some hotplug events */
6719 hotcpu_notifier(update_runtime, 0);
6723 /* Move init over to a non-isolated CPU */
6724 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6726 sched_init_granularity();
6727 free_cpumask_var(non_isolated_cpus);
6729 init_sched_rt_class();
6732 void __init sched_init_smp(void)
6734 sched_init_granularity();
6736 #endif /* CONFIG_SMP */
6738 const_debug unsigned int sysctl_timer_migration = 1;
6740 int in_sched_functions(unsigned long addr)
6742 return in_lock_functions(addr) ||
6743 (addr >= (unsigned long)__sched_text_start
6744 && addr < (unsigned long)__sched_text_end);
6747 #ifdef CONFIG_CGROUP_SCHED
6748 struct task_group root_task_group;
6749 LIST_HEAD(task_groups);
6752 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6754 void __init sched_init(void)
6757 unsigned long alloc_size = 0, ptr;
6759 #ifdef CONFIG_FAIR_GROUP_SCHED
6760 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6762 #ifdef CONFIG_RT_GROUP_SCHED
6763 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6765 #ifdef CONFIG_CPUMASK_OFFSTACK
6766 alloc_size += num_possible_cpus() * cpumask_size();
6769 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6771 #ifdef CONFIG_FAIR_GROUP_SCHED
6772 root_task_group.se = (struct sched_entity **)ptr;
6773 ptr += nr_cpu_ids * sizeof(void **);
6775 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6776 ptr += nr_cpu_ids * sizeof(void **);
6778 #endif /* CONFIG_FAIR_GROUP_SCHED */
6779 #ifdef CONFIG_RT_GROUP_SCHED
6780 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6781 ptr += nr_cpu_ids * sizeof(void **);
6783 root_task_group.rt_rq = (struct rt_rq **)ptr;
6784 ptr += nr_cpu_ids * sizeof(void **);
6786 #endif /* CONFIG_RT_GROUP_SCHED */
6787 #ifdef CONFIG_CPUMASK_OFFSTACK
6788 for_each_possible_cpu(i) {
6789 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6790 ptr += cpumask_size();
6792 #endif /* CONFIG_CPUMASK_OFFSTACK */
6796 init_defrootdomain();
6799 init_rt_bandwidth(&def_rt_bandwidth,
6800 global_rt_period(), global_rt_runtime());
6802 #ifdef CONFIG_RT_GROUP_SCHED
6803 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6804 global_rt_period(), global_rt_runtime());
6805 #endif /* CONFIG_RT_GROUP_SCHED */
6807 #ifdef CONFIG_CGROUP_SCHED
6808 list_add(&root_task_group.list, &task_groups);
6809 INIT_LIST_HEAD(&root_task_group.children);
6810 INIT_LIST_HEAD(&root_task_group.siblings);
6811 autogroup_init(&init_task);
6813 #endif /* CONFIG_CGROUP_SCHED */
6815 #ifdef CONFIG_CGROUP_CPUACCT
6816 root_cpuacct.cpustat = &kernel_cpustat;
6817 root_cpuacct.cpuusage = alloc_percpu(u64);
6818 /* Too early, not expected to fail */
6819 BUG_ON(!root_cpuacct.cpuusage);
6821 for_each_possible_cpu(i) {
6825 raw_spin_lock_init(&rq->lock);
6827 rq->calc_load_active = 0;
6828 rq->calc_load_update = jiffies + LOAD_FREQ;
6829 init_cfs_rq(&rq->cfs);
6830 init_rt_rq(&rq->rt, rq);
6831 #ifdef CONFIG_FAIR_GROUP_SCHED
6832 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6833 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6835 * How much cpu bandwidth does root_task_group get?
6837 * In case of task-groups formed thr' the cgroup filesystem, it
6838 * gets 100% of the cpu resources in the system. This overall
6839 * system cpu resource is divided among the tasks of
6840 * root_task_group and its child task-groups in a fair manner,
6841 * based on each entity's (task or task-group's) weight
6842 * (se->load.weight).
6844 * In other words, if root_task_group has 10 tasks of weight
6845 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6846 * then A0's share of the cpu resource is:
6848 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6850 * We achieve this by letting root_task_group's tasks sit
6851 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6853 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6854 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6855 #endif /* CONFIG_FAIR_GROUP_SCHED */
6857 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6858 #ifdef CONFIG_RT_GROUP_SCHED
6859 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6860 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6863 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6864 rq->cpu_load[j] = 0;
6866 rq->last_load_update_tick = jiffies;
6871 rq->cpu_power = SCHED_POWER_SCALE;
6872 rq->post_schedule = 0;
6873 rq->active_balance = 0;
6874 rq->next_balance = jiffies;
6879 rq->avg_idle = 2*sysctl_sched_migration_cost;
6881 INIT_LIST_HEAD(&rq->cfs_tasks);
6883 rq_attach_root(rq, &def_root_domain);
6889 atomic_set(&rq->nr_iowait, 0);
6892 set_load_weight(&init_task);
6894 #ifdef CONFIG_PREEMPT_NOTIFIERS
6895 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6898 #ifdef CONFIG_RT_MUTEXES
6899 plist_head_init(&init_task.pi_waiters);
6903 * The boot idle thread does lazy MMU switching as well:
6905 atomic_inc(&init_mm.mm_count);
6906 enter_lazy_tlb(&init_mm, current);
6909 * Make us the idle thread. Technically, schedule() should not be
6910 * called from this thread, however somewhere below it might be,
6911 * but because we are the idle thread, we just pick up running again
6912 * when this runqueue becomes "idle".
6914 init_idle(current, smp_processor_id());
6916 calc_load_update = jiffies + LOAD_FREQ;
6919 * During early bootup we pretend to be a normal task:
6921 current->sched_class = &fair_sched_class;
6924 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6925 /* May be allocated at isolcpus cmdline parse time */
6926 if (cpu_isolated_map == NULL)
6927 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6928 idle_thread_set_boot_cpu();
6930 init_sched_fair_class();
6932 scheduler_running = 1;
6935 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6936 static inline int preempt_count_equals(int preempt_offset)
6938 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6940 return (nested == preempt_offset);
6943 void __might_sleep(const char *file, int line, int preempt_offset)
6945 static unsigned long prev_jiffy; /* ratelimiting */
6947 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6948 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6949 system_state != SYSTEM_RUNNING || oops_in_progress)
6951 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6953 prev_jiffy = jiffies;
6956 "BUG: sleeping function called from invalid context at %s:%d\n",
6959 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6960 in_atomic(), irqs_disabled(),
6961 current->pid, current->comm);
6963 debug_show_held_locks(current);
6964 if (irqs_disabled())
6965 print_irqtrace_events(current);
6968 EXPORT_SYMBOL(__might_sleep);
6971 #ifdef CONFIG_MAGIC_SYSRQ
6972 static void normalize_task(struct rq *rq, struct task_struct *p)
6974 const struct sched_class *prev_class = p->sched_class;
6975 int old_prio = p->prio;
6980 dequeue_task(rq, p, 0);
6981 __setscheduler(rq, p, SCHED_NORMAL, 0);
6983 enqueue_task(rq, p, 0);
6984 resched_task(rq->curr);
6987 check_class_changed(rq, p, prev_class, old_prio);
6990 void normalize_rt_tasks(void)
6992 struct task_struct *g, *p;
6993 unsigned long flags;
6996 read_lock_irqsave(&tasklist_lock, flags);
6997 do_each_thread(g, p) {
6999 * Only normalize user tasks:
7004 p->se.exec_start = 0;
7005 #ifdef CONFIG_SCHEDSTATS
7006 p->se.statistics.wait_start = 0;
7007 p->se.statistics.sleep_start = 0;
7008 p->se.statistics.block_start = 0;
7013 * Renice negative nice level userspace
7016 if (TASK_NICE(p) < 0 && p->mm)
7017 set_user_nice(p, 0);
7021 raw_spin_lock(&p->pi_lock);
7022 rq = __task_rq_lock(p);
7024 normalize_task(rq, p);
7026 __task_rq_unlock(rq);
7027 raw_spin_unlock(&p->pi_lock);
7028 } while_each_thread(g, p);
7030 read_unlock_irqrestore(&tasklist_lock, flags);
7033 #endif /* CONFIG_MAGIC_SYSRQ */
7035 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7037 * These functions are only useful for the IA64 MCA handling, or kdb.
7039 * They can only be called when the whole system has been
7040 * stopped - every CPU needs to be quiescent, and no scheduling
7041 * activity can take place. Using them for anything else would
7042 * be a serious bug, and as a result, they aren't even visible
7043 * under any other configuration.
7047 * curr_task - return the current task for a given cpu.
7048 * @cpu: the processor in question.
7050 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7052 struct task_struct *curr_task(int cpu)
7054 return cpu_curr(cpu);
7057 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7061 * set_curr_task - set the current task for a given cpu.
7062 * @cpu: the processor in question.
7063 * @p: the task pointer to set.
7065 * Description: This function must only be used when non-maskable interrupts
7066 * are serviced on a separate stack. It allows the architecture to switch the
7067 * notion of the current task on a cpu in a non-blocking manner. This function
7068 * must be called with all CPU's synchronized, and interrupts disabled, the
7069 * and caller must save the original value of the current task (see
7070 * curr_task() above) and restore that value before reenabling interrupts and
7071 * re-starting the system.
7073 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7075 void set_curr_task(int cpu, struct task_struct *p)
7082 #ifdef CONFIG_CGROUP_SCHED
7083 /* task_group_lock serializes the addition/removal of task groups */
7084 static DEFINE_SPINLOCK(task_group_lock);
7086 static void free_sched_group(struct task_group *tg)
7088 free_fair_sched_group(tg);
7089 free_rt_sched_group(tg);
7094 /* allocate runqueue etc for a new task group */
7095 struct task_group *sched_create_group(struct task_group *parent)
7097 struct task_group *tg;
7098 unsigned long flags;
7100 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7102 return ERR_PTR(-ENOMEM);
7104 if (!alloc_fair_sched_group(tg, parent))
7107 if (!alloc_rt_sched_group(tg, parent))
7110 spin_lock_irqsave(&task_group_lock, flags);
7111 list_add_rcu(&tg->list, &task_groups);
7113 WARN_ON(!parent); /* root should already exist */
7115 tg->parent = parent;
7116 INIT_LIST_HEAD(&tg->children);
7117 list_add_rcu(&tg->siblings, &parent->children);
7118 spin_unlock_irqrestore(&task_group_lock, flags);
7123 free_sched_group(tg);
7124 return ERR_PTR(-ENOMEM);
7127 /* rcu callback to free various structures associated with a task group */
7128 static void free_sched_group_rcu(struct rcu_head *rhp)
7130 /* now it should be safe to free those cfs_rqs */
7131 free_sched_group(container_of(rhp, struct task_group, rcu));
7134 /* Destroy runqueue etc associated with a task group */
7135 void sched_destroy_group(struct task_group *tg)
7137 unsigned long flags;
7140 /* end participation in shares distribution */
7141 for_each_possible_cpu(i)
7142 unregister_fair_sched_group(tg, i);
7144 spin_lock_irqsave(&task_group_lock, flags);
7145 list_del_rcu(&tg->list);
7146 list_del_rcu(&tg->siblings);
7147 spin_unlock_irqrestore(&task_group_lock, flags);
7149 /* wait for possible concurrent references to cfs_rqs complete */
7150 call_rcu(&tg->rcu, free_sched_group_rcu);
7153 /* change task's runqueue when it moves between groups.
7154 * The caller of this function should have put the task in its new group
7155 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7156 * reflect its new group.
7158 void sched_move_task(struct task_struct *tsk)
7160 struct task_group *tg;
7162 unsigned long flags;
7165 rq = task_rq_lock(tsk, &flags);
7167 running = task_current(rq, tsk);
7171 dequeue_task(rq, tsk, 0);
7172 if (unlikely(running))
7173 tsk->sched_class->put_prev_task(rq, tsk);
7175 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7176 lockdep_is_held(&tsk->sighand->siglock)),
7177 struct task_group, css);
7178 tg = autogroup_task_group(tsk, tg);
7179 tsk->sched_task_group = tg;
7181 #ifdef CONFIG_FAIR_GROUP_SCHED
7182 if (tsk->sched_class->task_move_group)
7183 tsk->sched_class->task_move_group(tsk, on_rq);
7186 set_task_rq(tsk, task_cpu(tsk));
7188 if (unlikely(running))
7189 tsk->sched_class->set_curr_task(rq);
7191 enqueue_task(rq, tsk, 0);
7193 task_rq_unlock(rq, tsk, &flags);
7195 #endif /* CONFIG_CGROUP_SCHED */
7197 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7198 static unsigned long to_ratio(u64 period, u64 runtime)
7200 if (runtime == RUNTIME_INF)
7203 return div64_u64(runtime << 20, period);
7207 #ifdef CONFIG_RT_GROUP_SCHED
7209 * Ensure that the real time constraints are schedulable.
7211 static DEFINE_MUTEX(rt_constraints_mutex);
7213 /* Must be called with tasklist_lock held */
7214 static inline int tg_has_rt_tasks(struct task_group *tg)
7216 struct task_struct *g, *p;
7218 do_each_thread(g, p) {
7219 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7221 } while_each_thread(g, p);
7226 struct rt_schedulable_data {
7227 struct task_group *tg;
7232 static int tg_rt_schedulable(struct task_group *tg, void *data)
7234 struct rt_schedulable_data *d = data;
7235 struct task_group *child;
7236 unsigned long total, sum = 0;
7237 u64 period, runtime;
7239 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7240 runtime = tg->rt_bandwidth.rt_runtime;
7243 period = d->rt_period;
7244 runtime = d->rt_runtime;
7248 * Cannot have more runtime than the period.
7250 if (runtime > period && runtime != RUNTIME_INF)
7254 * Ensure we don't starve existing RT tasks.
7256 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7259 total = to_ratio(period, runtime);
7262 * Nobody can have more than the global setting allows.
7264 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7268 * The sum of our children's runtime should not exceed our own.
7270 list_for_each_entry_rcu(child, &tg->children, siblings) {
7271 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7272 runtime = child->rt_bandwidth.rt_runtime;
7274 if (child == d->tg) {
7275 period = d->rt_period;
7276 runtime = d->rt_runtime;
7279 sum += to_ratio(period, runtime);
7288 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7292 struct rt_schedulable_data data = {
7294 .rt_period = period,
7295 .rt_runtime = runtime,
7299 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7305 static int tg_set_rt_bandwidth(struct task_group *tg,
7306 u64 rt_period, u64 rt_runtime)
7310 mutex_lock(&rt_constraints_mutex);
7311 read_lock(&tasklist_lock);
7312 err = __rt_schedulable(tg, rt_period, rt_runtime);
7316 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7317 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7318 tg->rt_bandwidth.rt_runtime = rt_runtime;
7320 for_each_possible_cpu(i) {
7321 struct rt_rq *rt_rq = tg->rt_rq[i];
7323 raw_spin_lock(&rt_rq->rt_runtime_lock);
7324 rt_rq->rt_runtime = rt_runtime;
7325 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7327 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7329 read_unlock(&tasklist_lock);
7330 mutex_unlock(&rt_constraints_mutex);
7335 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7337 u64 rt_runtime, rt_period;
7339 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7340 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7341 if (rt_runtime_us < 0)
7342 rt_runtime = RUNTIME_INF;
7344 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7347 long sched_group_rt_runtime(struct task_group *tg)
7351 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7354 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7355 do_div(rt_runtime_us, NSEC_PER_USEC);
7356 return rt_runtime_us;
7359 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7361 u64 rt_runtime, rt_period;
7363 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7364 rt_runtime = tg->rt_bandwidth.rt_runtime;
7369 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7372 long sched_group_rt_period(struct task_group *tg)
7376 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7377 do_div(rt_period_us, NSEC_PER_USEC);
7378 return rt_period_us;
7381 static int sched_rt_global_constraints(void)
7383 u64 runtime, period;
7386 if (sysctl_sched_rt_period <= 0)
7389 runtime = global_rt_runtime();
7390 period = global_rt_period();
7393 * Sanity check on the sysctl variables.
7395 if (runtime > period && runtime != RUNTIME_INF)
7398 mutex_lock(&rt_constraints_mutex);
7399 read_lock(&tasklist_lock);
7400 ret = __rt_schedulable(NULL, 0, 0);
7401 read_unlock(&tasklist_lock);
7402 mutex_unlock(&rt_constraints_mutex);
7407 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7409 /* Don't accept realtime tasks when there is no way for them to run */
7410 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7416 #else /* !CONFIG_RT_GROUP_SCHED */
7417 static int sched_rt_global_constraints(void)
7419 unsigned long flags;
7422 if (sysctl_sched_rt_period <= 0)
7426 * There's always some RT tasks in the root group
7427 * -- migration, kstopmachine etc..
7429 if (sysctl_sched_rt_runtime == 0)
7432 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7433 for_each_possible_cpu(i) {
7434 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7436 raw_spin_lock(&rt_rq->rt_runtime_lock);
7437 rt_rq->rt_runtime = global_rt_runtime();
7438 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7440 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7444 #endif /* CONFIG_RT_GROUP_SCHED */
7446 int sched_rt_handler(struct ctl_table *table, int write,
7447 void __user *buffer, size_t *lenp,
7451 int old_period, old_runtime;
7452 static DEFINE_MUTEX(mutex);
7455 old_period = sysctl_sched_rt_period;
7456 old_runtime = sysctl_sched_rt_runtime;
7458 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7460 if (!ret && write) {
7461 ret = sched_rt_global_constraints();
7463 sysctl_sched_rt_period = old_period;
7464 sysctl_sched_rt_runtime = old_runtime;
7466 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7467 def_rt_bandwidth.rt_period =
7468 ns_to_ktime(global_rt_period());
7471 mutex_unlock(&mutex);
7476 #ifdef CONFIG_CGROUP_SCHED
7478 /* return corresponding task_group object of a cgroup */
7479 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7481 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7482 struct task_group, css);
7485 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7487 struct task_group *tg, *parent;
7489 if (!cgrp->parent) {
7490 /* This is early initialization for the top cgroup */
7491 return &root_task_group.css;
7494 parent = cgroup_tg(cgrp->parent);
7495 tg = sched_create_group(parent);
7497 return ERR_PTR(-ENOMEM);
7502 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7504 struct task_group *tg = cgroup_tg(cgrp);
7506 sched_destroy_group(tg);
7509 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7510 struct cgroup_taskset *tset)
7512 struct task_struct *task;
7514 cgroup_taskset_for_each(task, cgrp, tset) {
7515 #ifdef CONFIG_RT_GROUP_SCHED
7516 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7519 /* We don't support RT-tasks being in separate groups */
7520 if (task->sched_class != &fair_sched_class)
7527 static void cpu_cgroup_attach(struct cgroup *cgrp,
7528 struct cgroup_taskset *tset)
7530 struct task_struct *task;
7532 cgroup_taskset_for_each(task, cgrp, tset)
7533 sched_move_task(task);
7537 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7538 struct task_struct *task)
7541 * cgroup_exit() is called in the copy_process() failure path.
7542 * Ignore this case since the task hasn't ran yet, this avoids
7543 * trying to poke a half freed task state from generic code.
7545 if (!(task->flags & PF_EXITING))
7548 sched_move_task(task);
7551 #ifdef CONFIG_FAIR_GROUP_SCHED
7552 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7555 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7558 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7560 struct task_group *tg = cgroup_tg(cgrp);
7562 return (u64) scale_load_down(tg->shares);
7565 #ifdef CONFIG_CFS_BANDWIDTH
7566 static DEFINE_MUTEX(cfs_constraints_mutex);
7568 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7569 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7571 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7573 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7575 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7576 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7578 if (tg == &root_task_group)
7582 * Ensure we have at some amount of bandwidth every period. This is
7583 * to prevent reaching a state of large arrears when throttled via
7584 * entity_tick() resulting in prolonged exit starvation.
7586 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7590 * Likewise, bound things on the otherside by preventing insane quota
7591 * periods. This also allows us to normalize in computing quota
7594 if (period > max_cfs_quota_period)
7597 mutex_lock(&cfs_constraints_mutex);
7598 ret = __cfs_schedulable(tg, period, quota);
7602 runtime_enabled = quota != RUNTIME_INF;
7603 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7604 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7605 raw_spin_lock_irq(&cfs_b->lock);
7606 cfs_b->period = ns_to_ktime(period);
7607 cfs_b->quota = quota;
7609 __refill_cfs_bandwidth_runtime(cfs_b);
7610 /* restart the period timer (if active) to handle new period expiry */
7611 if (runtime_enabled && cfs_b->timer_active) {
7612 /* force a reprogram */
7613 cfs_b->timer_active = 0;
7614 __start_cfs_bandwidth(cfs_b);
7616 raw_spin_unlock_irq(&cfs_b->lock);
7618 for_each_possible_cpu(i) {
7619 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7620 struct rq *rq = cfs_rq->rq;
7622 raw_spin_lock_irq(&rq->lock);
7623 cfs_rq->runtime_enabled = runtime_enabled;
7624 cfs_rq->runtime_remaining = 0;
7626 if (cfs_rq->throttled)
7627 unthrottle_cfs_rq(cfs_rq);
7628 raw_spin_unlock_irq(&rq->lock);
7631 mutex_unlock(&cfs_constraints_mutex);
7636 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7640 period = ktime_to_ns(tg->cfs_bandwidth.period);
7641 if (cfs_quota_us < 0)
7642 quota = RUNTIME_INF;
7644 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7646 return tg_set_cfs_bandwidth(tg, period, quota);
7649 long tg_get_cfs_quota(struct task_group *tg)
7653 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7656 quota_us = tg->cfs_bandwidth.quota;
7657 do_div(quota_us, NSEC_PER_USEC);
7662 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7666 period = (u64)cfs_period_us * NSEC_PER_USEC;
7667 quota = tg->cfs_bandwidth.quota;
7669 return tg_set_cfs_bandwidth(tg, period, quota);
7672 long tg_get_cfs_period(struct task_group *tg)
7676 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7677 do_div(cfs_period_us, NSEC_PER_USEC);
7679 return cfs_period_us;
7682 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7684 return tg_get_cfs_quota(cgroup_tg(cgrp));
7687 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7690 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7693 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7695 return tg_get_cfs_period(cgroup_tg(cgrp));
7698 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7701 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7704 struct cfs_schedulable_data {
7705 struct task_group *tg;
7710 * normalize group quota/period to be quota/max_period
7711 * note: units are usecs
7713 static u64 normalize_cfs_quota(struct task_group *tg,
7714 struct cfs_schedulable_data *d)
7722 period = tg_get_cfs_period(tg);
7723 quota = tg_get_cfs_quota(tg);
7726 /* note: these should typically be equivalent */
7727 if (quota == RUNTIME_INF || quota == -1)
7730 return to_ratio(period, quota);
7733 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7735 struct cfs_schedulable_data *d = data;
7736 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7737 s64 quota = 0, parent_quota = -1;
7740 quota = RUNTIME_INF;
7742 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7744 quota = normalize_cfs_quota(tg, d);
7745 parent_quota = parent_b->hierarchal_quota;
7748 * ensure max(child_quota) <= parent_quota, inherit when no
7751 if (quota == RUNTIME_INF)
7752 quota = parent_quota;
7753 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7756 cfs_b->hierarchal_quota = quota;
7761 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7764 struct cfs_schedulable_data data = {
7770 if (quota != RUNTIME_INF) {
7771 do_div(data.period, NSEC_PER_USEC);
7772 do_div(data.quota, NSEC_PER_USEC);
7776 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7782 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7783 struct cgroup_map_cb *cb)
7785 struct task_group *tg = cgroup_tg(cgrp);
7786 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7788 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7789 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7790 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7794 #endif /* CONFIG_CFS_BANDWIDTH */
7795 #endif /* CONFIG_FAIR_GROUP_SCHED */
7797 #ifdef CONFIG_RT_GROUP_SCHED
7798 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7801 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7804 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7806 return sched_group_rt_runtime(cgroup_tg(cgrp));
7809 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7812 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7815 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7817 return sched_group_rt_period(cgroup_tg(cgrp));
7819 #endif /* CONFIG_RT_GROUP_SCHED */
7821 static struct cftype cpu_files[] = {
7822 #ifdef CONFIG_FAIR_GROUP_SCHED
7825 .read_u64 = cpu_shares_read_u64,
7826 .write_u64 = cpu_shares_write_u64,
7829 #ifdef CONFIG_CFS_BANDWIDTH
7831 .name = "cfs_quota_us",
7832 .read_s64 = cpu_cfs_quota_read_s64,
7833 .write_s64 = cpu_cfs_quota_write_s64,
7836 .name = "cfs_period_us",
7837 .read_u64 = cpu_cfs_period_read_u64,
7838 .write_u64 = cpu_cfs_period_write_u64,
7842 .read_map = cpu_stats_show,
7845 #ifdef CONFIG_RT_GROUP_SCHED
7847 .name = "rt_runtime_us",
7848 .read_s64 = cpu_rt_runtime_read,
7849 .write_s64 = cpu_rt_runtime_write,
7852 .name = "rt_period_us",
7853 .read_u64 = cpu_rt_period_read_uint,
7854 .write_u64 = cpu_rt_period_write_uint,
7860 struct cgroup_subsys cpu_cgroup_subsys = {
7862 .create = cpu_cgroup_create,
7863 .destroy = cpu_cgroup_destroy,
7864 .can_attach = cpu_cgroup_can_attach,
7865 .attach = cpu_cgroup_attach,
7866 .exit = cpu_cgroup_exit,
7867 .subsys_id = cpu_cgroup_subsys_id,
7868 .base_cftypes = cpu_files,
7872 #endif /* CONFIG_CGROUP_SCHED */
7874 #ifdef CONFIG_CGROUP_CPUACCT
7877 * CPU accounting code for task groups.
7879 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7880 * (balbir@in.ibm.com).
7883 struct cpuacct root_cpuacct;
7885 /* create a new cpu accounting group */
7886 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
7891 return &root_cpuacct.css;
7893 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7897 ca->cpuusage = alloc_percpu(u64);
7901 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7903 goto out_free_cpuusage;
7908 free_percpu(ca->cpuusage);
7912 return ERR_PTR(-ENOMEM);
7915 /* destroy an existing cpu accounting group */
7916 static void cpuacct_destroy(struct cgroup *cgrp)
7918 struct cpuacct *ca = cgroup_ca(cgrp);
7920 free_percpu(ca->cpustat);
7921 free_percpu(ca->cpuusage);
7925 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7927 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7930 #ifndef CONFIG_64BIT
7932 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7934 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7936 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7944 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7946 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7948 #ifndef CONFIG_64BIT
7950 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7952 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7954 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7960 /* return total cpu usage (in nanoseconds) of a group */
7961 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7963 struct cpuacct *ca = cgroup_ca(cgrp);
7964 u64 totalcpuusage = 0;
7967 for_each_present_cpu(i)
7968 totalcpuusage += cpuacct_cpuusage_read(ca, i);
7970 return totalcpuusage;
7973 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
7976 struct cpuacct *ca = cgroup_ca(cgrp);
7985 for_each_present_cpu(i)
7986 cpuacct_cpuusage_write(ca, i, 0);
7992 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
7995 struct cpuacct *ca = cgroup_ca(cgroup);
7999 for_each_present_cpu(i) {
8000 percpu = cpuacct_cpuusage_read(ca, i);
8001 seq_printf(m, "%llu ", (unsigned long long) percpu);
8003 seq_printf(m, "\n");
8007 static const char *cpuacct_stat_desc[] = {
8008 [CPUACCT_STAT_USER] = "user",
8009 [CPUACCT_STAT_SYSTEM] = "system",
8012 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8013 struct cgroup_map_cb *cb)
8015 struct cpuacct *ca = cgroup_ca(cgrp);
8019 for_each_online_cpu(cpu) {
8020 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8021 val += kcpustat->cpustat[CPUTIME_USER];
8022 val += kcpustat->cpustat[CPUTIME_NICE];
8024 val = cputime64_to_clock_t(val);
8025 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8028 for_each_online_cpu(cpu) {
8029 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8030 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8031 val += kcpustat->cpustat[CPUTIME_IRQ];
8032 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8035 val = cputime64_to_clock_t(val);
8036 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8041 static struct cftype files[] = {
8044 .read_u64 = cpuusage_read,
8045 .write_u64 = cpuusage_write,
8048 .name = "usage_percpu",
8049 .read_seq_string = cpuacct_percpu_seq_read,
8053 .read_map = cpuacct_stats_show,
8059 * charge this task's execution time to its accounting group.
8061 * called with rq->lock held.
8063 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8068 if (unlikely(!cpuacct_subsys.active))
8071 cpu = task_cpu(tsk);
8077 for (; ca; ca = parent_ca(ca)) {
8078 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8079 *cpuusage += cputime;
8085 struct cgroup_subsys cpuacct_subsys = {
8087 .create = cpuacct_create,
8088 .destroy = cpuacct_destroy,
8089 .subsys_id = cpuacct_subsys_id,
8090 .base_cftypes = files,
8092 #endif /* CONFIG_CGROUP_CPUACCT */