4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
86 #include "../smpboot.h"
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/sched.h>
91 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
94 ktime_t soft, hard, now;
97 if (hrtimer_active(period_timer))
100 now = hrtimer_cb_get_time(period_timer);
101 hrtimer_forward(period_timer, now, period);
103 soft = hrtimer_get_softexpires(period_timer);
104 hard = hrtimer_get_expires(period_timer);
105 delta = ktime_to_ns(ktime_sub(hard, soft));
106 __hrtimer_start_range_ns(period_timer, soft, delta,
107 HRTIMER_MODE_ABS_PINNED, 0);
111 DEFINE_MUTEX(sched_domains_mutex);
112 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
114 static void update_rq_clock_task(struct rq *rq, s64 delta);
116 void update_rq_clock(struct rq *rq)
120 if (rq->skip_clock_update > 0)
123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
125 update_rq_clock_task(rq, delta);
129 * Debugging: various feature bits
132 #define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
135 const_debug unsigned int sysctl_sched_features =
136 #include "features.h"
141 #ifdef CONFIG_SCHED_DEBUG
142 #define SCHED_FEAT(name, enabled) \
145 static const char * const sched_feat_names[] = {
146 #include "features.h"
151 static int sched_feat_show(struct seq_file *m, void *v)
155 for (i = 0; i < __SCHED_FEAT_NR; i++) {
156 if (!(sysctl_sched_features & (1UL << i)))
158 seq_printf(m, "%s ", sched_feat_names[i]);
165 #ifdef HAVE_JUMP_LABEL
167 #define jump_label_key__true STATIC_KEY_INIT_TRUE
168 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 #define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
173 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174 #include "features.h"
179 static void sched_feat_disable(int i)
181 if (static_key_enabled(&sched_feat_keys[i]))
182 static_key_slow_dec(&sched_feat_keys[i]);
185 static void sched_feat_enable(int i)
187 if (!static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_inc(&sched_feat_keys[i]);
191 static void sched_feat_disable(int i) { };
192 static void sched_feat_enable(int i) { };
193 #endif /* HAVE_JUMP_LABEL */
196 sched_feat_write(struct file *filp, const char __user *ubuf,
197 size_t cnt, loff_t *ppos)
207 if (copy_from_user(&buf, ubuf, cnt))
213 if (strncmp(cmp, "NO_", 3) == 0) {
218 for (i = 0; i < __SCHED_FEAT_NR; i++) {
219 if (strcmp(cmp, sched_feat_names[i]) == 0) {
221 sysctl_sched_features &= ~(1UL << i);
222 sched_feat_disable(i);
224 sysctl_sched_features |= (1UL << i);
225 sched_feat_enable(i);
231 if (i == __SCHED_FEAT_NR)
239 static int sched_feat_open(struct inode *inode, struct file *filp)
241 return single_open(filp, sched_feat_show, NULL);
244 static const struct file_operations sched_feat_fops = {
245 .open = sched_feat_open,
246 .write = sched_feat_write,
249 .release = single_release,
252 static __init int sched_init_debug(void)
254 debugfs_create_file("sched_features", 0644, NULL, NULL,
259 late_initcall(sched_init_debug);
260 #endif /* CONFIG_SCHED_DEBUG */
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
266 const_debug unsigned int sysctl_sched_nr_migrate = 32;
269 * period over which we average the RT time consumption, measured
274 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
277 * period over which we measure -rt task cpu usage in us.
280 unsigned int sysctl_sched_rt_period = 1000000;
282 __read_mostly int scheduler_running;
285 * part of the period that we allow rt tasks to run in us.
288 int sysctl_sched_rt_runtime = 950000;
293 * __task_rq_lock - lock the rq @p resides on.
295 static inline struct rq *__task_rq_lock(struct task_struct *p)
300 lockdep_assert_held(&p->pi_lock);
304 raw_spin_lock(&rq->lock);
305 if (likely(rq == task_rq(p)))
307 raw_spin_unlock(&rq->lock);
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
314 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
315 __acquires(p->pi_lock)
321 raw_spin_lock_irqsave(&p->pi_lock, *flags);
323 raw_spin_lock(&rq->lock);
324 if (likely(rq == task_rq(p)))
326 raw_spin_unlock(&rq->lock);
327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
331 static void __task_rq_unlock(struct rq *rq)
334 raw_spin_unlock(&rq->lock);
338 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
340 __releases(p->pi_lock)
342 raw_spin_unlock(&rq->lock);
343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
347 * this_rq_lock - lock this runqueue and disable interrupts.
349 static struct rq *this_rq_lock(void)
356 raw_spin_lock(&rq->lock);
361 #ifdef CONFIG_SCHED_HRTICK
363 * Use HR-timers to deliver accurate preemption points.
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
373 static void hrtick_clear(struct rq *rq)
375 if (hrtimer_active(&rq->hrtick_timer))
376 hrtimer_cancel(&rq->hrtick_timer);
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
383 static enum hrtimer_restart hrtick(struct hrtimer *timer)
385 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
389 raw_spin_lock(&rq->lock);
391 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392 raw_spin_unlock(&rq->lock);
394 return HRTIMER_NORESTART;
399 * called from hardirq (IPI) context
401 static void __hrtick_start(void *arg)
405 raw_spin_lock(&rq->lock);
406 hrtimer_restart(&rq->hrtick_timer);
407 rq->hrtick_csd_pending = 0;
408 raw_spin_unlock(&rq->lock);
412 * Called to set the hrtick timer state.
414 * called with rq->lock held and irqs disabled
416 void hrtick_start(struct rq *rq, u64 delay)
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
421 hrtimer_set_expires(timer, time);
423 if (rq == this_rq()) {
424 hrtimer_restart(timer);
425 } else if (!rq->hrtick_csd_pending) {
426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
427 rq->hrtick_csd_pending = 1;
432 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
434 int cpu = (int)(long)hcpu;
437 case CPU_UP_CANCELED:
438 case CPU_UP_CANCELED_FROZEN:
439 case CPU_DOWN_PREPARE:
440 case CPU_DOWN_PREPARE_FROZEN:
442 case CPU_DEAD_FROZEN:
443 hrtick_clear(cpu_rq(cpu));
450 static __init void init_hrtick(void)
452 hotcpu_notifier(hotplug_hrtick, 0);
456 * Called to set the hrtick timer state.
458 * called with rq->lock held and irqs disabled
460 void hrtick_start(struct rq *rq, u64 delay)
462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
463 HRTIMER_MODE_REL_PINNED, 0);
466 static inline void init_hrtick(void)
469 #endif /* CONFIG_SMP */
471 static void init_rq_hrtick(struct rq *rq)
474 rq->hrtick_csd_pending = 0;
476 rq->hrtick_csd.flags = 0;
477 rq->hrtick_csd.func = __hrtick_start;
478 rq->hrtick_csd.info = rq;
481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
482 rq->hrtick_timer.function = hrtick;
484 #else /* CONFIG_SCHED_HRTICK */
485 static inline void hrtick_clear(struct rq *rq)
489 static inline void init_rq_hrtick(struct rq *rq)
493 static inline void init_hrtick(void)
496 #endif /* CONFIG_SCHED_HRTICK */
499 * resched_task - mark a task 'to be rescheduled now'.
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
507 #ifndef tsk_is_polling
508 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
511 void resched_task(struct task_struct *p)
515 assert_raw_spin_locked(&task_rq(p)->lock);
517 if (test_tsk_need_resched(p))
520 set_tsk_need_resched(p);
523 if (cpu == smp_processor_id())
526 /* NEED_RESCHED must be visible before we test polling */
528 if (!tsk_is_polling(p))
529 smp_send_reschedule(cpu);
532 void resched_cpu(int cpu)
534 struct rq *rq = cpu_rq(cpu);
537 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
539 resched_task(cpu_curr(cpu));
540 raw_spin_unlock_irqrestore(&rq->lock, flags);
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
552 int get_nohz_timer_target(void)
554 int cpu = smp_processor_id();
556 struct sched_domain *sd;
559 for_each_domain(cpu, sd) {
560 for_each_cpu(i, sched_domain_span(sd)) {
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
581 void wake_up_idle_cpu(int cpu)
583 struct rq *rq = cpu_rq(cpu);
585 if (cpu == smp_processor_id())
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
595 if (rq->curr != rq->idle)
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
603 set_tsk_need_resched(rq->idle);
605 /* NEED_RESCHED must be visible before we test polling */
607 if (!tsk_is_polling(rq->idle))
608 smp_send_reschedule(cpu);
611 static inline bool got_nohz_idle_kick(void)
613 int cpu = smp_processor_id();
614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
617 #else /* CONFIG_NO_HZ */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ */
626 void sched_avg_update(struct rq *rq)
628 s64 period = sched_avg_period();
630 while ((s64)(rq->clock - rq->age_stamp) > period) {
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
636 asm("" : "+rm" (rq->age_stamp));
637 rq->age_stamp += period;
642 #else /* !CONFIG_SMP */
643 void resched_task(struct task_struct *p)
645 assert_raw_spin_locked(&task_rq(p)->lock);
646 set_tsk_need_resched(p);
648 #endif /* CONFIG_SMP */
650 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
656 * Caller must hold rcu_lock or sufficient equivalent.
658 int walk_tg_tree_from(struct task_group *from,
659 tg_visitor down, tg_visitor up, void *data)
661 struct task_group *parent, *child;
667 ret = (*down)(parent, data);
670 list_for_each_entry_rcu(child, &parent->children, siblings) {
677 ret = (*up)(parent, data);
678 if (ret || parent == from)
682 parent = parent->parent;
689 int tg_nop(struct task_group *tg, void *data)
695 static void set_load_weight(struct task_struct *p)
697 int prio = p->static_prio - MAX_RT_PRIO;
698 struct load_weight *load = &p->se.load;
701 * SCHED_IDLE tasks get minimal weight:
703 if (p->policy == SCHED_IDLE) {
704 load->weight = scale_load(WEIGHT_IDLEPRIO);
705 load->inv_weight = WMULT_IDLEPRIO;
709 load->weight = scale_load(prio_to_weight[prio]);
710 load->inv_weight = prio_to_wmult[prio];
713 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
716 sched_info_queued(p);
717 p->sched_class->enqueue_task(rq, p, flags);
720 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
723 sched_info_dequeued(p);
724 p->sched_class->dequeue_task(rq, p, flags);
727 void activate_task(struct rq *rq, struct task_struct *p, int flags)
729 if (task_contributes_to_load(p))
730 rq->nr_uninterruptible--;
732 enqueue_task(rq, p, flags);
735 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible++;
740 dequeue_task(rq, p, flags);
743 static void update_rq_clock_task(struct rq *rq, s64 delta)
746 * In theory, the compile should just see 0 here, and optimize out the call
747 * to sched_rt_avg_update. But I don't trust it...
749 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
750 s64 steal = 0, irq_delta = 0;
752 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
753 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
756 * Since irq_time is only updated on {soft,}irq_exit, we might run into
757 * this case when a previous update_rq_clock() happened inside a
760 * When this happens, we stop ->clock_task and only update the
761 * prev_irq_time stamp to account for the part that fit, so that a next
762 * update will consume the rest. This ensures ->clock_task is
765 * It does however cause some slight miss-attribution of {soft,}irq
766 * time, a more accurate solution would be to update the irq_time using
767 * the current rq->clock timestamp, except that would require using
770 if (irq_delta > delta)
773 rq->prev_irq_time += irq_delta;
776 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
777 if (static_key_false((¶virt_steal_rq_enabled))) {
780 steal = paravirt_steal_clock(cpu_of(rq));
781 steal -= rq->prev_steal_time_rq;
783 if (unlikely(steal > delta))
786 st = steal_ticks(steal);
787 steal = st * TICK_NSEC;
789 rq->prev_steal_time_rq += steal;
795 rq->clock_task += delta;
797 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
798 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
799 sched_rt_avg_update(rq, irq_delta + steal);
803 void sched_set_stop_task(int cpu, struct task_struct *stop)
805 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
806 struct task_struct *old_stop = cpu_rq(cpu)->stop;
810 * Make it appear like a SCHED_FIFO task, its something
811 * userspace knows about and won't get confused about.
813 * Also, it will make PI more or less work without too
814 * much confusion -- but then, stop work should not
815 * rely on PI working anyway.
817 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
819 stop->sched_class = &stop_sched_class;
822 cpu_rq(cpu)->stop = stop;
826 * Reset it back to a normal scheduling class so that
827 * it can die in pieces.
829 old_stop->sched_class = &rt_sched_class;
834 * __normal_prio - return the priority that is based on the static prio
836 static inline int __normal_prio(struct task_struct *p)
838 return p->static_prio;
842 * Calculate the expected normal priority: i.e. priority
843 * without taking RT-inheritance into account. Might be
844 * boosted by interactivity modifiers. Changes upon fork,
845 * setprio syscalls, and whenever the interactivity
846 * estimator recalculates.
848 static inline int normal_prio(struct task_struct *p)
852 if (task_has_rt_policy(p))
853 prio = MAX_RT_PRIO-1 - p->rt_priority;
855 prio = __normal_prio(p);
860 * Calculate the current priority, i.e. the priority
861 * taken into account by the scheduler. This value might
862 * be boosted by RT tasks, or might be boosted by
863 * interactivity modifiers. Will be RT if the task got
864 * RT-boosted. If not then it returns p->normal_prio.
866 static int effective_prio(struct task_struct *p)
868 p->normal_prio = normal_prio(p);
870 * If we are RT tasks or we were boosted to RT priority,
871 * keep the priority unchanged. Otherwise, update priority
872 * to the normal priority:
874 if (!rt_prio(p->prio))
875 return p->normal_prio;
880 * task_curr - is this task currently executing on a CPU?
881 * @p: the task in question.
883 inline int task_curr(const struct task_struct *p)
885 return cpu_curr(task_cpu(p)) == p;
888 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
889 const struct sched_class *prev_class,
892 if (prev_class != p->sched_class) {
893 if (prev_class->switched_from)
894 prev_class->switched_from(rq, p);
895 p->sched_class->switched_to(rq, p);
896 } else if (oldprio != p->prio)
897 p->sched_class->prio_changed(rq, p, oldprio);
900 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
902 const struct sched_class *class;
904 if (p->sched_class == rq->curr->sched_class) {
905 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
907 for_each_class(class) {
908 if (class == rq->curr->sched_class)
910 if (class == p->sched_class) {
911 resched_task(rq->curr);
918 * A queue event has occurred, and we're going to schedule. In
919 * this case, we can save a useless back to back clock update.
921 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
922 rq->skip_clock_update = 1;
926 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
928 #ifdef CONFIG_SCHED_DEBUG
930 * We should never call set_task_cpu() on a blocked task,
931 * ttwu() will sort out the placement.
933 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
934 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
936 #ifdef CONFIG_LOCKDEP
938 * The caller should hold either p->pi_lock or rq->lock, when changing
939 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
941 * sched_move_task() holds both and thus holding either pins the cgroup,
944 * Furthermore, all task_rq users should acquire both locks, see
947 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
948 lockdep_is_held(&task_rq(p)->lock)));
952 trace_sched_migrate_task(p, new_cpu);
954 if (task_cpu(p) != new_cpu) {
955 p->se.nr_migrations++;
956 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
959 __set_task_cpu(p, new_cpu);
962 struct migration_arg {
963 struct task_struct *task;
967 static int migration_cpu_stop(void *data);
970 * wait_task_inactive - wait for a thread to unschedule.
972 * If @match_state is nonzero, it's the @p->state value just checked and
973 * not expected to change. If it changes, i.e. @p might have woken up,
974 * then return zero. When we succeed in waiting for @p to be off its CPU,
975 * we return a positive number (its total switch count). If a second call
976 * a short while later returns the same number, the caller can be sure that
977 * @p has remained unscheduled the whole time.
979 * The caller must ensure that the task *will* unschedule sometime soon,
980 * else this function might spin for a *long* time. This function can't
981 * be called with interrupts off, or it may introduce deadlock with
982 * smp_call_function() if an IPI is sent by the same process we are
983 * waiting to become inactive.
985 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
994 * We do the initial early heuristics without holding
995 * any task-queue locks at all. We'll only try to get
996 * the runqueue lock when things look like they will
1002 * If the task is actively running on another CPU
1003 * still, just relax and busy-wait without holding
1006 * NOTE! Since we don't hold any locks, it's not
1007 * even sure that "rq" stays as the right runqueue!
1008 * But we don't care, since "task_running()" will
1009 * return false if the runqueue has changed and p
1010 * is actually now running somewhere else!
1012 while (task_running(rq, p)) {
1013 if (match_state && unlikely(p->state != match_state))
1019 * Ok, time to look more closely! We need the rq
1020 * lock now, to be *sure*. If we're wrong, we'll
1021 * just go back and repeat.
1023 rq = task_rq_lock(p, &flags);
1024 trace_sched_wait_task(p);
1025 running = task_running(rq, p);
1028 if (!match_state || p->state == match_state)
1029 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1030 task_rq_unlock(rq, p, &flags);
1033 * If it changed from the expected state, bail out now.
1035 if (unlikely(!ncsw))
1039 * Was it really running after all now that we
1040 * checked with the proper locks actually held?
1042 * Oops. Go back and try again..
1044 if (unlikely(running)) {
1050 * It's not enough that it's not actively running,
1051 * it must be off the runqueue _entirely_, and not
1054 * So if it was still runnable (but just not actively
1055 * running right now), it's preempted, and we should
1056 * yield - it could be a while.
1058 if (unlikely(on_rq)) {
1059 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1061 set_current_state(TASK_UNINTERRUPTIBLE);
1062 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1067 * Ahh, all good. It wasn't running, and it wasn't
1068 * runnable, which means that it will never become
1069 * running in the future either. We're all done!
1078 * kick_process - kick a running thread to enter/exit the kernel
1079 * @p: the to-be-kicked thread
1081 * Cause a process which is running on another CPU to enter
1082 * kernel-mode, without any delay. (to get signals handled.)
1084 * NOTE: this function doesn't have to take the runqueue lock,
1085 * because all it wants to ensure is that the remote task enters
1086 * the kernel. If the IPI races and the task has been migrated
1087 * to another CPU then no harm is done and the purpose has been
1090 void kick_process(struct task_struct *p)
1096 if ((cpu != smp_processor_id()) && task_curr(p))
1097 smp_send_reschedule(cpu);
1100 EXPORT_SYMBOL_GPL(kick_process);
1101 #endif /* CONFIG_SMP */
1105 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1107 static int select_fallback_rq(int cpu, struct task_struct *p)
1109 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1110 enum { cpuset, possible, fail } state = cpuset;
1113 /* Look for allowed, online CPU in same node. */
1114 for_each_cpu(dest_cpu, nodemask) {
1115 if (!cpu_online(dest_cpu))
1117 if (!cpu_active(dest_cpu))
1119 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1124 /* Any allowed, online CPU? */
1125 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1126 if (!cpu_online(dest_cpu))
1128 if (!cpu_active(dest_cpu))
1135 /* No more Mr. Nice Guy. */
1136 cpuset_cpus_allowed_fallback(p);
1141 do_set_cpus_allowed(p, cpu_possible_mask);
1152 if (state != cpuset) {
1154 * Don't tell them about moving exiting tasks or
1155 * kernel threads (both mm NULL), since they never
1158 if (p->mm && printk_ratelimit()) {
1159 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1160 task_pid_nr(p), p->comm, cpu);
1168 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1171 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1173 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1176 * In order not to call set_task_cpu() on a blocking task we need
1177 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1180 * Since this is common to all placement strategies, this lives here.
1182 * [ this allows ->select_task() to simply return task_cpu(p) and
1183 * not worry about this generic constraint ]
1185 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1187 cpu = select_fallback_rq(task_cpu(p), p);
1192 static void update_avg(u64 *avg, u64 sample)
1194 s64 diff = sample - *avg;
1200 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1202 #ifdef CONFIG_SCHEDSTATS
1203 struct rq *rq = this_rq();
1206 int this_cpu = smp_processor_id();
1208 if (cpu == this_cpu) {
1209 schedstat_inc(rq, ttwu_local);
1210 schedstat_inc(p, se.statistics.nr_wakeups_local);
1212 struct sched_domain *sd;
1214 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1216 for_each_domain(this_cpu, sd) {
1217 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1218 schedstat_inc(sd, ttwu_wake_remote);
1225 if (wake_flags & WF_MIGRATED)
1226 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1228 #endif /* CONFIG_SMP */
1230 schedstat_inc(rq, ttwu_count);
1231 schedstat_inc(p, se.statistics.nr_wakeups);
1233 if (wake_flags & WF_SYNC)
1234 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1236 #endif /* CONFIG_SCHEDSTATS */
1239 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1241 activate_task(rq, p, en_flags);
1244 /* if a worker is waking up, notify workqueue */
1245 if (p->flags & PF_WQ_WORKER)
1246 wq_worker_waking_up(p, cpu_of(rq));
1250 * Mark the task runnable and perform wakeup-preemption.
1253 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1255 trace_sched_wakeup(p, true);
1256 check_preempt_curr(rq, p, wake_flags);
1258 p->state = TASK_RUNNING;
1260 if (p->sched_class->task_woken)
1261 p->sched_class->task_woken(rq, p);
1263 if (rq->idle_stamp) {
1264 u64 delta = rq->clock - rq->idle_stamp;
1265 u64 max = 2*sysctl_sched_migration_cost;
1270 update_avg(&rq->avg_idle, delta);
1277 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1280 if (p->sched_contributes_to_load)
1281 rq->nr_uninterruptible--;
1284 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1285 ttwu_do_wakeup(rq, p, wake_flags);
1289 * Called in case the task @p isn't fully descheduled from its runqueue,
1290 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1291 * since all we need to do is flip p->state to TASK_RUNNING, since
1292 * the task is still ->on_rq.
1294 static int ttwu_remote(struct task_struct *p, int wake_flags)
1299 rq = __task_rq_lock(p);
1301 ttwu_do_wakeup(rq, p, wake_flags);
1304 __task_rq_unlock(rq);
1310 static void sched_ttwu_pending(void)
1312 struct rq *rq = this_rq();
1313 struct llist_node *llist = llist_del_all(&rq->wake_list);
1314 struct task_struct *p;
1316 raw_spin_lock(&rq->lock);
1319 p = llist_entry(llist, struct task_struct, wake_entry);
1320 llist = llist_next(llist);
1321 ttwu_do_activate(rq, p, 0);
1324 raw_spin_unlock(&rq->lock);
1327 void scheduler_ipi(void)
1329 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1333 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1334 * traditionally all their work was done from the interrupt return
1335 * path. Now that we actually do some work, we need to make sure
1338 * Some archs already do call them, luckily irq_enter/exit nest
1341 * Arguably we should visit all archs and update all handlers,
1342 * however a fair share of IPIs are still resched only so this would
1343 * somewhat pessimize the simple resched case.
1346 sched_ttwu_pending();
1349 * Check if someone kicked us for doing the nohz idle load balance.
1351 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1352 this_rq()->idle_balance = 1;
1353 raise_softirq_irqoff(SCHED_SOFTIRQ);
1358 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1360 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1361 smp_send_reschedule(cpu);
1364 bool cpus_share_cache(int this_cpu, int that_cpu)
1366 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1368 #endif /* CONFIG_SMP */
1370 static void ttwu_queue(struct task_struct *p, int cpu)
1372 struct rq *rq = cpu_rq(cpu);
1374 #if defined(CONFIG_SMP)
1375 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1376 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1377 ttwu_queue_remote(p, cpu);
1382 raw_spin_lock(&rq->lock);
1383 ttwu_do_activate(rq, p, 0);
1384 raw_spin_unlock(&rq->lock);
1388 * try_to_wake_up - wake up a thread
1389 * @p: the thread to be awakened
1390 * @state: the mask of task states that can be woken
1391 * @wake_flags: wake modifier flags (WF_*)
1393 * Put it on the run-queue if it's not already there. The "current"
1394 * thread is always on the run-queue (except when the actual
1395 * re-schedule is in progress), and as such you're allowed to do
1396 * the simpler "current->state = TASK_RUNNING" to mark yourself
1397 * runnable without the overhead of this.
1399 * Returns %true if @p was woken up, %false if it was already running
1400 * or @state didn't match @p's state.
1403 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1405 unsigned long flags;
1406 int cpu, success = 0;
1409 raw_spin_lock_irqsave(&p->pi_lock, flags);
1410 if (!(p->state & state))
1413 success = 1; /* we're going to change ->state */
1416 if (p->on_rq && ttwu_remote(p, wake_flags))
1421 * If the owning (remote) cpu is still in the middle of schedule() with
1422 * this task as prev, wait until its done referencing the task.
1427 * Pairs with the smp_wmb() in finish_lock_switch().
1431 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1432 p->state = TASK_WAKING;
1434 if (p->sched_class->task_waking)
1435 p->sched_class->task_waking(p);
1437 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1438 if (task_cpu(p) != cpu) {
1439 wake_flags |= WF_MIGRATED;
1440 set_task_cpu(p, cpu);
1442 #endif /* CONFIG_SMP */
1446 ttwu_stat(p, cpu, wake_flags);
1448 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1454 * try_to_wake_up_local - try to wake up a local task with rq lock held
1455 * @p: the thread to be awakened
1457 * Put @p on the run-queue if it's not already there. The caller must
1458 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1461 static void try_to_wake_up_local(struct task_struct *p)
1463 struct rq *rq = task_rq(p);
1465 BUG_ON(rq != this_rq());
1466 BUG_ON(p == current);
1467 lockdep_assert_held(&rq->lock);
1469 if (!raw_spin_trylock(&p->pi_lock)) {
1470 raw_spin_unlock(&rq->lock);
1471 raw_spin_lock(&p->pi_lock);
1472 raw_spin_lock(&rq->lock);
1475 if (!(p->state & TASK_NORMAL))
1479 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1481 ttwu_do_wakeup(rq, p, 0);
1482 ttwu_stat(p, smp_processor_id(), 0);
1484 raw_spin_unlock(&p->pi_lock);
1488 * wake_up_process - Wake up a specific process
1489 * @p: The process to be woken up.
1491 * Attempt to wake up the nominated process and move it to the set of runnable
1492 * processes. Returns 1 if the process was woken up, 0 if it was already
1495 * It may be assumed that this function implies a write memory barrier before
1496 * changing the task state if and only if any tasks are woken up.
1498 int wake_up_process(struct task_struct *p)
1500 return try_to_wake_up(p, TASK_ALL, 0);
1502 EXPORT_SYMBOL(wake_up_process);
1504 int wake_up_state(struct task_struct *p, unsigned int state)
1506 return try_to_wake_up(p, state, 0);
1510 * Perform scheduler related setup for a newly forked process p.
1511 * p is forked by current.
1513 * __sched_fork() is basic setup used by init_idle() too:
1515 static void __sched_fork(struct task_struct *p)
1520 p->se.exec_start = 0;
1521 p->se.sum_exec_runtime = 0;
1522 p->se.prev_sum_exec_runtime = 0;
1523 p->se.nr_migrations = 0;
1525 INIT_LIST_HEAD(&p->se.group_node);
1527 #ifdef CONFIG_SCHEDSTATS
1528 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1531 INIT_LIST_HEAD(&p->rt.run_list);
1533 #ifdef CONFIG_PREEMPT_NOTIFIERS
1534 INIT_HLIST_HEAD(&p->preempt_notifiers);
1537 #ifdef CONFIG_SCHED_NUMA
1538 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1539 p->mm->numa_next_scan = jiffies;
1540 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_faults = NULL;
1548 p->numa_task_period = sysctl_sched_numa_task_period_min;
1549 #endif /* CONFIG_SCHED_NUMA */
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)) {
1791 task_numa_free(prev);
1793 * Remove function-return probe instances associated with this
1794 * task and put them back on the free list.
1796 kprobe_flush_task(prev);
1797 put_task_struct(prev);
1803 /* assumes rq->lock is held */
1804 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1806 if (prev->sched_class->pre_schedule)
1807 prev->sched_class->pre_schedule(rq, prev);
1810 /* rq->lock is NOT held, but preemption is disabled */
1811 static inline void post_schedule(struct rq *rq)
1813 if (rq->post_schedule) {
1814 unsigned long flags;
1816 raw_spin_lock_irqsave(&rq->lock, flags);
1817 if (rq->curr->sched_class->post_schedule)
1818 rq->curr->sched_class->post_schedule(rq);
1819 raw_spin_unlock_irqrestore(&rq->lock, flags);
1821 rq->post_schedule = 0;
1827 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1831 static inline void post_schedule(struct rq *rq)
1838 * schedule_tail - first thing a freshly forked thread must call.
1839 * @prev: the thread we just switched away from.
1841 asmlinkage void schedule_tail(struct task_struct *prev)
1842 __releases(rq->lock)
1844 struct rq *rq = this_rq();
1846 finish_task_switch(rq, prev);
1849 * FIXME: do we need to worry about rq being invalidated by the
1854 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1855 /* In this case, finish_task_switch does not reenable preemption */
1858 if (current->set_child_tid)
1859 put_user(task_pid_vnr(current), current->set_child_tid);
1863 * context_switch - switch to the new MM and the new
1864 * thread's register state.
1867 context_switch(struct rq *rq, struct task_struct *prev,
1868 struct task_struct *next)
1870 struct mm_struct *mm, *oldmm;
1872 prepare_task_switch(rq, prev, next);
1875 oldmm = prev->active_mm;
1877 * For paravirt, this is coupled with an exit in switch_to to
1878 * combine the page table reload and the switch backend into
1881 arch_start_context_switch(prev);
1884 next->active_mm = oldmm;
1885 atomic_inc(&oldmm->mm_count);
1886 enter_lazy_tlb(oldmm, next);
1888 switch_mm(oldmm, mm, next);
1891 prev->active_mm = NULL;
1892 rq->prev_mm = oldmm;
1895 * Since the runqueue lock will be released by the next
1896 * task (which is an invalid locking op but in the case
1897 * of the scheduler it's an obvious special-case), so we
1898 * do an early lockdep release here:
1900 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1901 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1904 /* Here we just switch the register state and the stack. */
1905 rcu_switch(prev, next);
1906 switch_to(prev, next, prev);
1910 * this_rq must be evaluated again because prev may have moved
1911 * CPUs since it called schedule(), thus the 'rq' on its stack
1912 * frame will be invalid.
1914 finish_task_switch(this_rq(), prev);
1918 * nr_running, nr_uninterruptible and nr_context_switches:
1920 * externally visible scheduler statistics: current number of runnable
1921 * threads, current number of uninterruptible-sleeping threads, total
1922 * number of context switches performed since bootup.
1924 unsigned long nr_running(void)
1926 unsigned long i, sum = 0;
1928 for_each_online_cpu(i)
1929 sum += cpu_rq(i)->nr_running;
1934 unsigned long nr_uninterruptible(void)
1936 unsigned long i, sum = 0;
1938 for_each_possible_cpu(i)
1939 sum += cpu_rq(i)->nr_uninterruptible;
1942 * Since we read the counters lockless, it might be slightly
1943 * inaccurate. Do not allow it to go below zero though:
1945 if (unlikely((long)sum < 0))
1951 unsigned long long nr_context_switches(void)
1954 unsigned long long sum = 0;
1956 for_each_possible_cpu(i)
1957 sum += cpu_rq(i)->nr_switches;
1962 unsigned long nr_iowait(void)
1964 unsigned long i, sum = 0;
1966 for_each_possible_cpu(i)
1967 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1972 unsigned long nr_iowait_cpu(int cpu)
1974 struct rq *this = cpu_rq(cpu);
1975 return atomic_read(&this->nr_iowait);
1978 unsigned long this_cpu_load(void)
1980 struct rq *this = this_rq();
1981 return this->cpu_load[0];
1986 * Global load-average calculations
1988 * We take a distributed and async approach to calculating the global load-avg
1989 * in order to minimize overhead.
1991 * The global load average is an exponentially decaying average of nr_running +
1992 * nr_uninterruptible.
1994 * Once every LOAD_FREQ:
1997 * for_each_possible_cpu(cpu)
1998 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2000 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2002 * Due to a number of reasons the above turns in the mess below:
2004 * - for_each_possible_cpu() is prohibitively expensive on machines with
2005 * serious number of cpus, therefore we need to take a distributed approach
2006 * to calculating nr_active.
2008 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2009 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2011 * So assuming nr_active := 0 when we start out -- true per definition, we
2012 * can simply take per-cpu deltas and fold those into a global accumulate
2013 * to obtain the same result. See calc_load_fold_active().
2015 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2016 * across the machine, we assume 10 ticks is sufficient time for every
2017 * cpu to have completed this task.
2019 * This places an upper-bound on the IRQ-off latency of the machine. Then
2020 * again, being late doesn't loose the delta, just wrecks the sample.
2022 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2023 * this would add another cross-cpu cacheline miss and atomic operation
2024 * to the wakeup path. Instead we increment on whatever cpu the task ran
2025 * when it went into uninterruptible state and decrement on whatever cpu
2026 * did the wakeup. This means that only the sum of nr_uninterruptible over
2027 * all cpus yields the correct result.
2029 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2032 /* Variables and functions for calc_load */
2033 static atomic_long_t calc_load_tasks;
2034 static unsigned long calc_load_update;
2035 unsigned long avenrun[3];
2036 EXPORT_SYMBOL(avenrun); /* should be removed */
2039 * get_avenrun - get the load average array
2040 * @loads: pointer to dest load array
2041 * @offset: offset to add
2042 * @shift: shift count to shift the result left
2044 * These values are estimates at best, so no need for locking.
2046 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2048 loads[0] = (avenrun[0] + offset) << shift;
2049 loads[1] = (avenrun[1] + offset) << shift;
2050 loads[2] = (avenrun[2] + offset) << shift;
2053 static long calc_load_fold_active(struct rq *this_rq)
2055 long nr_active, delta = 0;
2057 nr_active = this_rq->nr_running;
2058 nr_active += (long) this_rq->nr_uninterruptible;
2060 if (nr_active != this_rq->calc_load_active) {
2061 delta = nr_active - this_rq->calc_load_active;
2062 this_rq->calc_load_active = nr_active;
2069 * a1 = a0 * e + a * (1 - e)
2071 static unsigned long
2072 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2075 load += active * (FIXED_1 - exp);
2076 load += 1UL << (FSHIFT - 1);
2077 return load >> FSHIFT;
2082 * Handle NO_HZ for the global load-average.
2084 * Since the above described distributed algorithm to compute the global
2085 * load-average relies on per-cpu sampling from the tick, it is affected by
2088 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2089 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2090 * when we read the global state.
2092 * Obviously reality has to ruin such a delightfully simple scheme:
2094 * - When we go NO_HZ idle during the window, we can negate our sample
2095 * contribution, causing under-accounting.
2097 * We avoid this by keeping two idle-delta counters and flipping them
2098 * when the window starts, thus separating old and new NO_HZ load.
2100 * The only trick is the slight shift in index flip for read vs write.
2104 * |-|-----------|-|-----------|-|-----------|-|
2105 * r:0 0 1 1 0 0 1 1 0
2106 * w:0 1 1 0 0 1 1 0 0
2108 * This ensures we'll fold the old idle contribution in this window while
2109 * accumlating the new one.
2111 * - When we wake up from NO_HZ idle during the window, we push up our
2112 * contribution, since we effectively move our sample point to a known
2115 * This is solved by pushing the window forward, and thus skipping the
2116 * sample, for this cpu (effectively using the idle-delta for this cpu which
2117 * was in effect at the time the window opened). This also solves the issue
2118 * of having to deal with a cpu having been in NOHZ idle for multiple
2119 * LOAD_FREQ intervals.
2121 * When making the ILB scale, we should try to pull this in as well.
2123 static atomic_long_t calc_load_idle[2];
2124 static int calc_load_idx;
2126 static inline int calc_load_write_idx(void)
2128 int idx = calc_load_idx;
2131 * See calc_global_nohz(), if we observe the new index, we also
2132 * need to observe the new update time.
2137 * If the folding window started, make sure we start writing in the
2140 if (!time_before(jiffies, calc_load_update))
2146 static inline int calc_load_read_idx(void)
2148 return calc_load_idx & 1;
2151 void calc_load_enter_idle(void)
2153 struct rq *this_rq = this_rq();
2157 * We're going into NOHZ mode, if there's any pending delta, fold it
2158 * into the pending idle delta.
2160 delta = calc_load_fold_active(this_rq);
2162 int idx = calc_load_write_idx();
2163 atomic_long_add(delta, &calc_load_idle[idx]);
2167 void calc_load_exit_idle(void)
2169 struct rq *this_rq = this_rq();
2172 * If we're still before the sample window, we're done.
2174 if (time_before(jiffies, this_rq->calc_load_update))
2178 * We woke inside or after the sample window, this means we're already
2179 * accounted through the nohz accounting, so skip the entire deal and
2180 * sync up for the next window.
2182 this_rq->calc_load_update = calc_load_update;
2183 if (time_before(jiffies, this_rq->calc_load_update + 10))
2184 this_rq->calc_load_update += LOAD_FREQ;
2187 static long calc_load_fold_idle(void)
2189 int idx = calc_load_read_idx();
2192 if (atomic_long_read(&calc_load_idle[idx]))
2193 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2199 * fixed_power_int - compute: x^n, in O(log n) time
2201 * @x: base of the power
2202 * @frac_bits: fractional bits of @x
2203 * @n: power to raise @x to.
2205 * By exploiting the relation between the definition of the natural power
2206 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2207 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2208 * (where: n_i \elem {0, 1}, the binary vector representing n),
2209 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2210 * of course trivially computable in O(log_2 n), the length of our binary
2213 static unsigned long
2214 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2216 unsigned long result = 1UL << frac_bits;
2221 result += 1UL << (frac_bits - 1);
2222 result >>= frac_bits;
2228 x += 1UL << (frac_bits - 1);
2236 * a1 = a0 * e + a * (1 - e)
2238 * a2 = a1 * e + a * (1 - e)
2239 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2240 * = a0 * e^2 + a * (1 - e) * (1 + e)
2242 * a3 = a2 * e + a * (1 - e)
2243 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2244 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2248 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2249 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2250 * = a0 * e^n + a * (1 - e^n)
2252 * [1] application of the geometric series:
2255 * S_n := \Sum x^i = -------------
2258 static unsigned long
2259 calc_load_n(unsigned long load, unsigned long exp,
2260 unsigned long active, unsigned int n)
2263 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2267 * NO_HZ can leave us missing all per-cpu ticks calling
2268 * calc_load_account_active(), but since an idle CPU folds its delta into
2269 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2270 * in the pending idle delta if our idle period crossed a load cycle boundary.
2272 * Once we've updated the global active value, we need to apply the exponential
2273 * weights adjusted to the number of cycles missed.
2275 static void calc_global_nohz(void)
2277 long delta, active, n;
2279 if (!time_before(jiffies, calc_load_update + 10)) {
2281 * Catch-up, fold however many we are behind still
2283 delta = jiffies - calc_load_update - 10;
2284 n = 1 + (delta / LOAD_FREQ);
2286 active = atomic_long_read(&calc_load_tasks);
2287 active = active > 0 ? active * FIXED_1 : 0;
2289 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2290 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2291 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2293 calc_load_update += n * LOAD_FREQ;
2297 * Flip the idle index...
2299 * Make sure we first write the new time then flip the index, so that
2300 * calc_load_write_idx() will see the new time when it reads the new
2301 * index, this avoids a double flip messing things up.
2306 #else /* !CONFIG_NO_HZ */
2308 static inline long calc_load_fold_idle(void) { return 0; }
2309 static inline void calc_global_nohz(void) { }
2311 #endif /* CONFIG_NO_HZ */
2314 * calc_load - update the avenrun load estimates 10 ticks after the
2315 * CPUs have updated calc_load_tasks.
2317 void calc_global_load(unsigned long ticks)
2321 if (time_before(jiffies, calc_load_update + 10))
2325 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2327 delta = calc_load_fold_idle();
2329 atomic_long_add(delta, &calc_load_tasks);
2331 active = atomic_long_read(&calc_load_tasks);
2332 active = active > 0 ? active * FIXED_1 : 0;
2334 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2335 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2336 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2338 calc_load_update += LOAD_FREQ;
2341 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2347 * Called from update_cpu_load() to periodically update this CPU's
2350 static void calc_load_account_active(struct rq *this_rq)
2354 if (time_before(jiffies, this_rq->calc_load_update))
2357 delta = calc_load_fold_active(this_rq);
2359 atomic_long_add(delta, &calc_load_tasks);
2361 this_rq->calc_load_update += LOAD_FREQ;
2365 * End of global load-average stuff
2369 * The exact cpuload at various idx values, calculated at every tick would be
2370 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2372 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2373 * on nth tick when cpu may be busy, then we have:
2374 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2375 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2377 * decay_load_missed() below does efficient calculation of
2378 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2379 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2381 * The calculation is approximated on a 128 point scale.
2382 * degrade_zero_ticks is the number of ticks after which load at any
2383 * particular idx is approximated to be zero.
2384 * degrade_factor is a precomputed table, a row for each load idx.
2385 * Each column corresponds to degradation factor for a power of two ticks,
2386 * based on 128 point scale.
2388 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2389 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2391 * With this power of 2 load factors, we can degrade the load n times
2392 * by looking at 1 bits in n and doing as many mult/shift instead of
2393 * n mult/shifts needed by the exact degradation.
2395 #define DEGRADE_SHIFT 7
2396 static const unsigned char
2397 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2398 static const unsigned char
2399 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2400 {0, 0, 0, 0, 0, 0, 0, 0},
2401 {64, 32, 8, 0, 0, 0, 0, 0},
2402 {96, 72, 40, 12, 1, 0, 0},
2403 {112, 98, 75, 43, 15, 1, 0},
2404 {120, 112, 98, 76, 45, 16, 2} };
2407 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2408 * would be when CPU is idle and so we just decay the old load without
2409 * adding any new load.
2411 static unsigned long
2412 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2416 if (!missed_updates)
2419 if (missed_updates >= degrade_zero_ticks[idx])
2423 return load >> missed_updates;
2425 while (missed_updates) {
2426 if (missed_updates % 2)
2427 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2429 missed_updates >>= 1;
2436 * Update rq->cpu_load[] statistics. This function is usually called every
2437 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2438 * every tick. We fix it up based on jiffies.
2440 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2441 unsigned long pending_updates)
2445 this_rq->nr_load_updates++;
2447 /* Update our load: */
2448 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2449 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2450 unsigned long old_load, new_load;
2452 /* scale is effectively 1 << i now, and >> i divides by scale */
2454 old_load = this_rq->cpu_load[i];
2455 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2456 new_load = this_load;
2458 * Round up the averaging division if load is increasing. This
2459 * prevents us from getting stuck on 9 if the load is 10, for
2462 if (new_load > old_load)
2463 new_load += scale - 1;
2465 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2468 sched_avg_update(this_rq);
2473 * There is no sane way to deal with nohz on smp when using jiffies because the
2474 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2475 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2477 * Therefore we cannot use the delta approach from the regular tick since that
2478 * would seriously skew the load calculation. However we'll make do for those
2479 * updates happening while idle (nohz_idle_balance) or coming out of idle
2480 * (tick_nohz_idle_exit).
2482 * This means we might still be one tick off for nohz periods.
2486 * Called from nohz_idle_balance() to update the load ratings before doing the
2489 void update_idle_cpu_load(struct rq *this_rq)
2491 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2492 unsigned long load = this_rq->load.weight;
2493 unsigned long pending_updates;
2496 * bail if there's load or we're actually up-to-date.
2498 if (load || curr_jiffies == this_rq->last_load_update_tick)
2501 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2502 this_rq->last_load_update_tick = curr_jiffies;
2504 __update_cpu_load(this_rq, load, pending_updates);
2508 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2510 void update_cpu_load_nohz(void)
2512 struct rq *this_rq = this_rq();
2513 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2514 unsigned long pending_updates;
2516 if (curr_jiffies == this_rq->last_load_update_tick)
2519 raw_spin_lock(&this_rq->lock);
2520 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2521 if (pending_updates) {
2522 this_rq->last_load_update_tick = curr_jiffies;
2524 * We were idle, this means load 0, the current load might be
2525 * !0 due to remote wakeups and the sort.
2527 __update_cpu_load(this_rq, 0, pending_updates);
2529 raw_spin_unlock(&this_rq->lock);
2531 #endif /* CONFIG_NO_HZ */
2534 * Called from scheduler_tick()
2536 static void update_cpu_load_active(struct rq *this_rq)
2539 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2541 this_rq->last_load_update_tick = jiffies;
2542 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2544 calc_load_account_active(this_rq);
2550 * sched_exec - execve() is a valuable balancing opportunity, because at
2551 * this point the task has the smallest effective memory and cache footprint.
2553 void sched_exec(void)
2555 struct task_struct *p = current;
2556 unsigned long flags;
2559 raw_spin_lock_irqsave(&p->pi_lock, flags);
2560 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2561 if (dest_cpu == smp_processor_id())
2564 if (likely(cpu_active(dest_cpu))) {
2565 struct migration_arg arg = { p, dest_cpu };
2567 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2568 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2572 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2577 DEFINE_PER_CPU(struct kernel_stat, kstat);
2578 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2580 EXPORT_PER_CPU_SYMBOL(kstat);
2581 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2584 * Return any ns on the sched_clock that have not yet been accounted in
2585 * @p in case that task is currently running.
2587 * Called with task_rq_lock() held on @rq.
2589 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2593 if (task_current(rq, p)) {
2594 update_rq_clock(rq);
2595 ns = rq->clock_task - p->se.exec_start;
2603 unsigned long long task_delta_exec(struct task_struct *p)
2605 unsigned long flags;
2609 rq = task_rq_lock(p, &flags);
2610 ns = do_task_delta_exec(p, rq);
2611 task_rq_unlock(rq, p, &flags);
2617 * Return accounted runtime for the task.
2618 * In case the task is currently running, return the runtime plus current's
2619 * pending runtime that have not been accounted yet.
2621 unsigned long long task_sched_runtime(struct task_struct *p)
2623 unsigned long flags;
2627 rq = task_rq_lock(p, &flags);
2628 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2629 task_rq_unlock(rq, p, &flags);
2635 * This function gets called by the timer code, with HZ frequency.
2636 * We call it with interrupts disabled.
2638 void scheduler_tick(void)
2640 int cpu = smp_processor_id();
2641 struct rq *rq = cpu_rq(cpu);
2642 struct task_struct *curr = rq->curr;
2646 raw_spin_lock(&rq->lock);
2647 update_rq_clock(rq);
2648 update_cpu_load_active(rq);
2649 curr->sched_class->task_tick(rq, curr, 0);
2650 raw_spin_unlock(&rq->lock);
2652 perf_event_task_tick();
2655 rq->idle_balance = idle_cpu(cpu);
2656 trigger_load_balance(rq, cpu);
2660 notrace unsigned long get_parent_ip(unsigned long addr)
2662 if (in_lock_functions(addr)) {
2663 addr = CALLER_ADDR2;
2664 if (in_lock_functions(addr))
2665 addr = CALLER_ADDR3;
2670 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2671 defined(CONFIG_PREEMPT_TRACER))
2673 void __kprobes add_preempt_count(int val)
2675 #ifdef CONFIG_DEBUG_PREEMPT
2679 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2682 preempt_count() += val;
2683 #ifdef CONFIG_DEBUG_PREEMPT
2685 * Spinlock count overflowing soon?
2687 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2690 if (preempt_count() == val)
2691 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2693 EXPORT_SYMBOL(add_preempt_count);
2695 void __kprobes sub_preempt_count(int val)
2697 #ifdef CONFIG_DEBUG_PREEMPT
2701 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2704 * Is the spinlock portion underflowing?
2706 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2707 !(preempt_count() & PREEMPT_MASK)))
2711 if (preempt_count() == val)
2712 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2713 preempt_count() -= val;
2715 EXPORT_SYMBOL(sub_preempt_count);
2720 * Print scheduling while atomic bug:
2722 static noinline void __schedule_bug(struct task_struct *prev)
2724 if (oops_in_progress)
2727 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2728 prev->comm, prev->pid, preempt_count());
2730 debug_show_held_locks(prev);
2732 if (irqs_disabled())
2733 print_irqtrace_events(prev);
2735 add_taint(TAINT_WARN);
2739 * Various schedule()-time debugging checks and statistics:
2741 static inline void schedule_debug(struct task_struct *prev)
2744 * Test if we are atomic. Since do_exit() needs to call into
2745 * schedule() atomically, we ignore that path for now.
2746 * Otherwise, whine if we are scheduling when we should not be.
2748 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2749 __schedule_bug(prev);
2752 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2754 schedstat_inc(this_rq(), sched_count);
2757 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2759 if (prev->on_rq || rq->skip_clock_update < 0)
2760 update_rq_clock(rq);
2761 prev->sched_class->put_prev_task(rq, prev);
2765 * Pick up the highest-prio task:
2767 static inline struct task_struct *
2768 pick_next_task(struct rq *rq)
2770 const struct sched_class *class;
2771 struct task_struct *p;
2774 * Optimization: we know that if all tasks are in
2775 * the fair class we can call that function directly:
2777 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2778 p = fair_sched_class.pick_next_task(rq);
2783 for_each_class(class) {
2784 p = class->pick_next_task(rq);
2789 BUG(); /* the idle class will always have a runnable task */
2793 * __schedule() is the main scheduler function.
2795 * The main means of driving the scheduler and thus entering this function are:
2797 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2799 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2800 * paths. For example, see arch/x86/entry_64.S.
2802 * To drive preemption between tasks, the scheduler sets the flag in timer
2803 * interrupt handler scheduler_tick().
2805 * 3. Wakeups don't really cause entry into schedule(). They add a
2806 * task to the run-queue and that's it.
2808 * Now, if the new task added to the run-queue preempts the current
2809 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2810 * called on the nearest possible occasion:
2812 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2814 * - in syscall or exception context, at the next outmost
2815 * preempt_enable(). (this might be as soon as the wake_up()'s
2818 * - in IRQ context, return from interrupt-handler to
2819 * preemptible context
2821 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2824 * - cond_resched() call
2825 * - explicit schedule() call
2826 * - return from syscall or exception to user-space
2827 * - return from interrupt-handler to user-space
2829 static void __sched __schedule(void)
2831 struct task_struct *prev, *next;
2832 unsigned long *switch_count;
2838 cpu = smp_processor_id();
2840 rcu_note_context_switch(cpu);
2843 schedule_debug(prev);
2845 if (sched_feat(HRTICK))
2848 raw_spin_lock_irq(&rq->lock);
2850 switch_count = &prev->nivcsw;
2851 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2852 if (unlikely(signal_pending_state(prev->state, prev))) {
2853 prev->state = TASK_RUNNING;
2855 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2859 * If a worker went to sleep, notify and ask workqueue
2860 * whether it wants to wake up a task to maintain
2863 if (prev->flags & PF_WQ_WORKER) {
2864 struct task_struct *to_wakeup;
2866 to_wakeup = wq_worker_sleeping(prev, cpu);
2868 try_to_wake_up_local(to_wakeup);
2871 switch_count = &prev->nvcsw;
2874 pre_schedule(rq, prev);
2876 if (unlikely(!rq->nr_running))
2877 idle_balance(cpu, rq);
2879 put_prev_task(rq, prev);
2880 next = pick_next_task(rq);
2881 clear_tsk_need_resched(prev);
2882 rq->skip_clock_update = 0;
2884 if (likely(prev != next)) {
2889 context_switch(rq, prev, next); /* unlocks the rq */
2891 * The context switch have flipped the stack from under us
2892 * and restored the local variables which were saved when
2893 * this task called schedule() in the past. prev == current
2894 * is still correct, but it can be moved to another cpu/rq.
2896 cpu = smp_processor_id();
2899 raw_spin_unlock_irq(&rq->lock);
2903 sched_preempt_enable_no_resched();
2908 static inline void sched_submit_work(struct task_struct *tsk)
2910 if (!tsk->state || tsk_is_pi_blocked(tsk))
2913 * If we are going to sleep and we have plugged IO queued,
2914 * make sure to submit it to avoid deadlocks.
2916 if (blk_needs_flush_plug(tsk))
2917 blk_schedule_flush_plug(tsk);
2920 asmlinkage void __sched schedule(void)
2922 struct task_struct *tsk = current;
2924 sched_submit_work(tsk);
2927 EXPORT_SYMBOL(schedule);
2929 #ifdef CONFIG_RCU_USER_QS
2930 asmlinkage void __sched schedule_user(void)
2933 * If we come here after a random call to set_need_resched(),
2934 * or we have been woken up remotely but the IPI has not yet arrived,
2935 * we haven't yet exited the RCU idle mode. Do it here manually until
2936 * we find a better solution.
2945 * schedule_preempt_disabled - called with preemption disabled
2947 * Returns with preemption disabled. Note: preempt_count must be 1
2949 void __sched schedule_preempt_disabled(void)
2951 sched_preempt_enable_no_resched();
2956 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2958 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2960 if (lock->owner != owner)
2964 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2965 * lock->owner still matches owner, if that fails, owner might
2966 * point to free()d memory, if it still matches, the rcu_read_lock()
2967 * ensures the memory stays valid.
2971 return owner->on_cpu;
2975 * Look out! "owner" is an entirely speculative pointer
2976 * access and not reliable.
2978 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
2980 if (!sched_feat(OWNER_SPIN))
2984 while (owner_running(lock, owner)) {
2988 arch_mutex_cpu_relax();
2993 * We break out the loop above on need_resched() and when the
2994 * owner changed, which is a sign for heavy contention. Return
2995 * success only when lock->owner is NULL.
2997 return lock->owner == NULL;
3001 #ifdef CONFIG_PREEMPT
3003 * this is the entry point to schedule() from in-kernel preemption
3004 * off of preempt_enable. Kernel preemptions off return from interrupt
3005 * occur there and call schedule directly.
3007 asmlinkage void __sched notrace preempt_schedule(void)
3009 struct thread_info *ti = current_thread_info();
3012 * If there is a non-zero preempt_count or interrupts are disabled,
3013 * we do not want to preempt the current task. Just return..
3015 if (likely(ti->preempt_count || irqs_disabled()))
3019 add_preempt_count_notrace(PREEMPT_ACTIVE);
3021 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3024 * Check again in case we missed a preemption opportunity
3025 * between schedule and now.
3028 } while (need_resched());
3030 EXPORT_SYMBOL(preempt_schedule);
3033 * this is the entry point to schedule() from kernel preemption
3034 * off of irq context.
3035 * Note, that this is called and return with irqs disabled. This will
3036 * protect us against recursive calling from irq.
3038 asmlinkage void __sched preempt_schedule_irq(void)
3040 struct thread_info *ti = current_thread_info();
3042 /* Catch callers which need to be fixed */
3043 BUG_ON(ti->preempt_count || !irqs_disabled());
3047 add_preempt_count(PREEMPT_ACTIVE);
3050 local_irq_disable();
3051 sub_preempt_count(PREEMPT_ACTIVE);
3054 * Check again in case we missed a preemption opportunity
3055 * between schedule and now.
3058 } while (need_resched());
3061 #endif /* CONFIG_PREEMPT */
3063 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3066 return try_to_wake_up(curr->private, mode, wake_flags);
3068 EXPORT_SYMBOL(default_wake_function);
3071 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3072 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3073 * number) then we wake all the non-exclusive tasks and one exclusive task.
3075 * There are circumstances in which we can try to wake a task which has already
3076 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3077 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3079 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3080 int nr_exclusive, int wake_flags, void *key)
3082 wait_queue_t *curr, *next;
3084 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3085 unsigned flags = curr->flags;
3087 if (curr->func(curr, mode, wake_flags, key) &&
3088 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3094 * __wake_up - wake up threads blocked on a waitqueue.
3096 * @mode: which threads
3097 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3098 * @key: is directly passed to the wakeup function
3100 * It may be assumed that this function implies a write memory barrier before
3101 * changing the task state if and only if any tasks are woken up.
3103 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3104 int nr_exclusive, void *key)
3106 unsigned long flags;
3108 spin_lock_irqsave(&q->lock, flags);
3109 __wake_up_common(q, mode, nr_exclusive, 0, key);
3110 spin_unlock_irqrestore(&q->lock, flags);
3112 EXPORT_SYMBOL(__wake_up);
3115 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3117 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3119 __wake_up_common(q, mode, nr, 0, NULL);
3121 EXPORT_SYMBOL_GPL(__wake_up_locked);
3123 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3125 __wake_up_common(q, mode, 1, 0, key);
3127 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3130 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3132 * @mode: which threads
3133 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3134 * @key: opaque value to be passed to wakeup targets
3136 * The sync wakeup differs that the waker knows that it will schedule
3137 * away soon, so while the target thread will be woken up, it will not
3138 * be migrated to another CPU - ie. the two threads are 'synchronized'
3139 * with each other. This can prevent needless bouncing between CPUs.
3141 * On UP it can prevent extra preemption.
3143 * It may be assumed that this function implies a write memory barrier before
3144 * changing the task state if and only if any tasks are woken up.
3146 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3147 int nr_exclusive, void *key)
3149 unsigned long flags;
3150 int wake_flags = WF_SYNC;
3155 if (unlikely(!nr_exclusive))
3158 spin_lock_irqsave(&q->lock, flags);
3159 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3160 spin_unlock_irqrestore(&q->lock, flags);
3162 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3165 * __wake_up_sync - see __wake_up_sync_key()
3167 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3169 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3171 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3174 * complete: - signals a single thread waiting on this completion
3175 * @x: holds the state of this particular completion
3177 * This will wake up a single thread waiting on this completion. Threads will be
3178 * awakened in the same order in which they were queued.
3180 * See also complete_all(), wait_for_completion() and related routines.
3182 * It may be assumed that this function implies a write memory barrier before
3183 * changing the task state if and only if any tasks are woken up.
3185 void complete(struct completion *x)
3187 unsigned long flags;
3189 spin_lock_irqsave(&x->wait.lock, flags);
3191 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3192 spin_unlock_irqrestore(&x->wait.lock, flags);
3194 EXPORT_SYMBOL(complete);
3197 * complete_all: - signals all threads waiting on this completion
3198 * @x: holds the state of this particular completion
3200 * This will wake up all threads waiting on this particular completion event.
3202 * It may be assumed that this function implies a write memory barrier before
3203 * changing the task state if and only if any tasks are woken up.
3205 void complete_all(struct completion *x)
3207 unsigned long flags;
3209 spin_lock_irqsave(&x->wait.lock, flags);
3210 x->done += UINT_MAX/2;
3211 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3212 spin_unlock_irqrestore(&x->wait.lock, flags);
3214 EXPORT_SYMBOL(complete_all);
3216 static inline long __sched
3217 do_wait_for_common(struct completion *x, long timeout, int state)
3220 DECLARE_WAITQUEUE(wait, current);
3222 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3224 if (signal_pending_state(state, current)) {
3225 timeout = -ERESTARTSYS;
3228 __set_current_state(state);
3229 spin_unlock_irq(&x->wait.lock);
3230 timeout = schedule_timeout(timeout);
3231 spin_lock_irq(&x->wait.lock);
3232 } while (!x->done && timeout);
3233 __remove_wait_queue(&x->wait, &wait);
3238 return timeout ?: 1;
3242 wait_for_common(struct completion *x, long timeout, int state)
3246 spin_lock_irq(&x->wait.lock);
3247 timeout = do_wait_for_common(x, timeout, state);
3248 spin_unlock_irq(&x->wait.lock);
3253 * wait_for_completion: - waits for completion of a task
3254 * @x: holds the state of this particular completion
3256 * This waits to be signaled for completion of a specific task. It is NOT
3257 * interruptible and there is no timeout.
3259 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3260 * and interrupt capability. Also see complete().
3262 void __sched wait_for_completion(struct completion *x)
3264 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3266 EXPORT_SYMBOL(wait_for_completion);
3269 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3270 * @x: holds the state of this particular completion
3271 * @timeout: timeout value in jiffies
3273 * This waits for either a completion of a specific task to be signaled or for a
3274 * specified timeout to expire. The timeout is in jiffies. It is not
3277 * The return value is 0 if timed out, and positive (at least 1, or number of
3278 * jiffies left till timeout) if completed.
3280 unsigned long __sched
3281 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3283 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3285 EXPORT_SYMBOL(wait_for_completion_timeout);
3288 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3289 * @x: holds the state of this particular completion
3291 * This waits for completion of a specific task to be signaled. It is
3294 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3296 int __sched wait_for_completion_interruptible(struct completion *x)
3298 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3299 if (t == -ERESTARTSYS)
3303 EXPORT_SYMBOL(wait_for_completion_interruptible);
3306 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3307 * @x: holds the state of this particular completion
3308 * @timeout: timeout value in jiffies
3310 * This waits for either a completion of a specific task to be signaled or for a
3311 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3313 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3314 * positive (at least 1, or number of jiffies left till timeout) if completed.
3317 wait_for_completion_interruptible_timeout(struct completion *x,
3318 unsigned long timeout)
3320 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3322 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3325 * wait_for_completion_killable: - waits for completion of a task (killable)
3326 * @x: holds the state of this particular completion
3328 * This waits to be signaled for completion of a specific task. It can be
3329 * interrupted by a kill signal.
3331 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3333 int __sched wait_for_completion_killable(struct completion *x)
3335 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3336 if (t == -ERESTARTSYS)
3340 EXPORT_SYMBOL(wait_for_completion_killable);
3343 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3344 * @x: holds the state of this particular completion
3345 * @timeout: timeout value in jiffies
3347 * This waits for either a completion of a specific task to be
3348 * signaled or for a specified timeout to expire. It can be
3349 * interrupted by a kill signal. The timeout is in jiffies.
3351 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3352 * positive (at least 1, or number of jiffies left till timeout) if completed.
3355 wait_for_completion_killable_timeout(struct completion *x,
3356 unsigned long timeout)
3358 return wait_for_common(x, timeout, TASK_KILLABLE);
3360 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3363 * try_wait_for_completion - try to decrement a completion without blocking
3364 * @x: completion structure
3366 * Returns: 0 if a decrement cannot be done without blocking
3367 * 1 if a decrement succeeded.
3369 * If a completion is being used as a counting completion,
3370 * attempt to decrement the counter without blocking. This
3371 * enables us to avoid waiting if the resource the completion
3372 * is protecting is not available.
3374 bool try_wait_for_completion(struct completion *x)
3376 unsigned long flags;
3379 spin_lock_irqsave(&x->wait.lock, flags);
3384 spin_unlock_irqrestore(&x->wait.lock, flags);
3387 EXPORT_SYMBOL(try_wait_for_completion);
3390 * completion_done - Test to see if a completion has any waiters
3391 * @x: completion structure
3393 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3394 * 1 if there are no waiters.
3397 bool completion_done(struct completion *x)
3399 unsigned long flags;
3402 spin_lock_irqsave(&x->wait.lock, flags);
3405 spin_unlock_irqrestore(&x->wait.lock, flags);
3408 EXPORT_SYMBOL(completion_done);
3411 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3413 unsigned long flags;
3416 init_waitqueue_entry(&wait, current);
3418 __set_current_state(state);
3420 spin_lock_irqsave(&q->lock, flags);
3421 __add_wait_queue(q, &wait);
3422 spin_unlock(&q->lock);
3423 timeout = schedule_timeout(timeout);
3424 spin_lock_irq(&q->lock);
3425 __remove_wait_queue(q, &wait);
3426 spin_unlock_irqrestore(&q->lock, flags);
3431 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3433 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3435 EXPORT_SYMBOL(interruptible_sleep_on);
3438 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3440 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3442 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3444 void __sched sleep_on(wait_queue_head_t *q)
3446 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3448 EXPORT_SYMBOL(sleep_on);
3450 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3452 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3454 EXPORT_SYMBOL(sleep_on_timeout);
3456 #ifdef CONFIG_RT_MUTEXES
3459 * rt_mutex_setprio - set the current priority of a task
3461 * @prio: prio value (kernel-internal form)
3463 * This function changes the 'effective' priority of a task. It does
3464 * not touch ->normal_prio like __setscheduler().
3466 * Used by the rt_mutex code to implement priority inheritance logic.
3468 void rt_mutex_setprio(struct task_struct *p, int prio)
3470 int oldprio, on_rq, running;
3472 const struct sched_class *prev_class;
3474 BUG_ON(prio < 0 || prio > MAX_PRIO);
3476 rq = __task_rq_lock(p);
3479 * Idle task boosting is a nono in general. There is one
3480 * exception, when PREEMPT_RT and NOHZ is active:
3482 * The idle task calls get_next_timer_interrupt() and holds
3483 * the timer wheel base->lock on the CPU and another CPU wants
3484 * to access the timer (probably to cancel it). We can safely
3485 * ignore the boosting request, as the idle CPU runs this code
3486 * with interrupts disabled and will complete the lock
3487 * protected section without being interrupted. So there is no
3488 * real need to boost.
3490 if (unlikely(p == rq->idle)) {
3491 WARN_ON(p != rq->curr);
3492 WARN_ON(p->pi_blocked_on);
3496 trace_sched_pi_setprio(p, prio);
3498 prev_class = p->sched_class;
3500 running = task_current(rq, p);
3502 dequeue_task(rq, p, 0);
3504 p->sched_class->put_prev_task(rq, p);
3507 p->sched_class = &rt_sched_class;
3509 p->sched_class = &fair_sched_class;
3514 p->sched_class->set_curr_task(rq);
3516 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3518 check_class_changed(rq, p, prev_class, oldprio);
3520 __task_rq_unlock(rq);
3523 void set_user_nice(struct task_struct *p, long nice)
3525 int old_prio, delta, on_rq;
3526 unsigned long flags;
3529 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3532 * We have to be careful, if called from sys_setpriority(),
3533 * the task might be in the middle of scheduling on another CPU.
3535 rq = task_rq_lock(p, &flags);
3537 * The RT priorities are set via sched_setscheduler(), but we still
3538 * allow the 'normal' nice value to be set - but as expected
3539 * it wont have any effect on scheduling until the task is
3540 * SCHED_FIFO/SCHED_RR:
3542 if (task_has_rt_policy(p)) {
3543 p->static_prio = NICE_TO_PRIO(nice);
3548 dequeue_task(rq, p, 0);
3550 p->static_prio = NICE_TO_PRIO(nice);
3553 p->prio = effective_prio(p);
3554 delta = p->prio - old_prio;
3557 enqueue_task(rq, p, 0);
3559 * If the task increased its priority or is running and
3560 * lowered its priority, then reschedule its CPU:
3562 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3563 resched_task(rq->curr);
3566 task_rq_unlock(rq, p, &flags);
3568 EXPORT_SYMBOL(set_user_nice);
3571 * can_nice - check if a task can reduce its nice value
3575 int can_nice(const struct task_struct *p, const int nice)
3577 /* convert nice value [19,-20] to rlimit style value [1,40] */
3578 int nice_rlim = 20 - nice;
3580 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3581 capable(CAP_SYS_NICE));
3584 #ifdef __ARCH_WANT_SYS_NICE
3587 * sys_nice - change the priority of the current process.
3588 * @increment: priority increment
3590 * sys_setpriority is a more generic, but much slower function that
3591 * does similar things.
3593 SYSCALL_DEFINE1(nice, int, increment)
3598 * Setpriority might change our priority at the same moment.
3599 * We don't have to worry. Conceptually one call occurs first
3600 * and we have a single winner.
3602 if (increment < -40)
3607 nice = TASK_NICE(current) + increment;
3613 if (increment < 0 && !can_nice(current, nice))
3616 retval = security_task_setnice(current, nice);
3620 set_user_nice(current, nice);
3627 * task_prio - return the priority value of a given task.
3628 * @p: the task in question.
3630 * This is the priority value as seen by users in /proc.
3631 * RT tasks are offset by -200. Normal tasks are centered
3632 * around 0, value goes from -16 to +15.
3634 int task_prio(const struct task_struct *p)
3636 return p->prio - MAX_RT_PRIO;
3640 * task_nice - return the nice value of a given task.
3641 * @p: the task in question.
3643 int task_nice(const struct task_struct *p)
3645 return TASK_NICE(p);
3647 EXPORT_SYMBOL(task_nice);
3650 * idle_cpu - is a given cpu idle currently?
3651 * @cpu: the processor in question.
3653 int idle_cpu(int cpu)
3655 struct rq *rq = cpu_rq(cpu);
3657 if (rq->curr != rq->idle)
3664 if (!llist_empty(&rq->wake_list))
3672 * idle_task - return the idle task for a given cpu.
3673 * @cpu: the processor in question.
3675 struct task_struct *idle_task(int cpu)
3677 return cpu_rq(cpu)->idle;
3681 * find_process_by_pid - find a process with a matching PID value.
3682 * @pid: the pid in question.
3684 static struct task_struct *find_process_by_pid(pid_t pid)
3686 return pid ? find_task_by_vpid(pid) : current;
3689 /* Actually do priority change: must hold rq lock. */
3691 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3694 p->rt_priority = prio;
3695 p->normal_prio = normal_prio(p);
3696 /* we are holding p->pi_lock already */
3697 p->prio = rt_mutex_getprio(p);
3698 if (rt_prio(p->prio))
3699 p->sched_class = &rt_sched_class;
3701 p->sched_class = &fair_sched_class;
3706 * check the target process has a UID that matches the current process's
3708 static bool check_same_owner(struct task_struct *p)
3710 const struct cred *cred = current_cred(), *pcred;
3714 pcred = __task_cred(p);
3715 match = (uid_eq(cred->euid, pcred->euid) ||
3716 uid_eq(cred->euid, pcred->uid));
3721 static int __sched_setscheduler(struct task_struct *p, int policy,
3722 const struct sched_param *param, bool user)
3724 int retval, oldprio, oldpolicy = -1, on_rq, running;
3725 unsigned long flags;
3726 const struct sched_class *prev_class;
3730 /* may grab non-irq protected spin_locks */
3731 BUG_ON(in_interrupt());
3733 /* double check policy once rq lock held */
3735 reset_on_fork = p->sched_reset_on_fork;
3736 policy = oldpolicy = p->policy;
3738 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3739 policy &= ~SCHED_RESET_ON_FORK;
3741 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3742 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3743 policy != SCHED_IDLE)
3748 * Valid priorities for SCHED_FIFO and SCHED_RR are
3749 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3750 * SCHED_BATCH and SCHED_IDLE is 0.
3752 if (param->sched_priority < 0 ||
3753 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3754 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3756 if (rt_policy(policy) != (param->sched_priority != 0))
3760 * Allow unprivileged RT tasks to decrease priority:
3762 if (user && !capable(CAP_SYS_NICE)) {
3763 if (rt_policy(policy)) {
3764 unsigned long rlim_rtprio =
3765 task_rlimit(p, RLIMIT_RTPRIO);
3767 /* can't set/change the rt policy */
3768 if (policy != p->policy && !rlim_rtprio)
3771 /* can't increase priority */
3772 if (param->sched_priority > p->rt_priority &&
3773 param->sched_priority > rlim_rtprio)
3778 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3779 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3781 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3782 if (!can_nice(p, TASK_NICE(p)))
3786 /* can't change other user's priorities */
3787 if (!check_same_owner(p))
3790 /* Normal users shall not reset the sched_reset_on_fork flag */
3791 if (p->sched_reset_on_fork && !reset_on_fork)
3796 retval = security_task_setscheduler(p);
3802 * make sure no PI-waiters arrive (or leave) while we are
3803 * changing the priority of the task:
3805 * To be able to change p->policy safely, the appropriate
3806 * runqueue lock must be held.
3808 rq = task_rq_lock(p, &flags);
3811 * Changing the policy of the stop threads its a very bad idea
3813 if (p == rq->stop) {
3814 task_rq_unlock(rq, p, &flags);
3819 * If not changing anything there's no need to proceed further:
3821 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3822 param->sched_priority == p->rt_priority))) {
3823 task_rq_unlock(rq, p, &flags);
3827 #ifdef CONFIG_RT_GROUP_SCHED
3830 * Do not allow realtime tasks into groups that have no runtime
3833 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3834 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3835 !task_group_is_autogroup(task_group(p))) {
3836 task_rq_unlock(rq, p, &flags);
3842 /* recheck policy now with rq lock held */
3843 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3844 policy = oldpolicy = -1;
3845 task_rq_unlock(rq, p, &flags);
3849 running = task_current(rq, p);
3851 dequeue_task(rq, p, 0);
3853 p->sched_class->put_prev_task(rq, p);
3855 p->sched_reset_on_fork = reset_on_fork;
3858 prev_class = p->sched_class;
3859 __setscheduler(rq, p, policy, param->sched_priority);
3862 p->sched_class->set_curr_task(rq);
3864 enqueue_task(rq, p, 0);
3866 check_class_changed(rq, p, prev_class, oldprio);
3867 task_rq_unlock(rq, p, &flags);
3869 rt_mutex_adjust_pi(p);
3875 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3876 * @p: the task in question.
3877 * @policy: new policy.
3878 * @param: structure containing the new RT priority.
3880 * NOTE that the task may be already dead.
3882 int sched_setscheduler(struct task_struct *p, int policy,
3883 const struct sched_param *param)
3885 return __sched_setscheduler(p, policy, param, true);
3887 EXPORT_SYMBOL_GPL(sched_setscheduler);
3890 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3891 * @p: the task in question.
3892 * @policy: new policy.
3893 * @param: structure containing the new RT priority.
3895 * Just like sched_setscheduler, only don't bother checking if the
3896 * current context has permission. For example, this is needed in
3897 * stop_machine(): we create temporary high priority worker threads,
3898 * but our caller might not have that capability.
3900 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3901 const struct sched_param *param)
3903 return __sched_setscheduler(p, policy, param, false);
3907 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3909 struct sched_param lparam;
3910 struct task_struct *p;
3913 if (!param || pid < 0)
3915 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3920 p = find_process_by_pid(pid);
3922 retval = sched_setscheduler(p, policy, &lparam);
3929 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3930 * @pid: the pid in question.
3931 * @policy: new policy.
3932 * @param: structure containing the new RT priority.
3934 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3935 struct sched_param __user *, param)
3937 /* negative values for policy are not valid */
3941 return do_sched_setscheduler(pid, policy, param);
3945 * sys_sched_setparam - set/change the RT priority of a thread
3946 * @pid: the pid in question.
3947 * @param: structure containing the new RT priority.
3949 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3951 return do_sched_setscheduler(pid, -1, param);
3955 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3956 * @pid: the pid in question.
3958 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3960 struct task_struct *p;
3968 p = find_process_by_pid(pid);
3970 retval = security_task_getscheduler(p);
3973 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3980 * sys_sched_getparam - get the RT priority of a thread
3981 * @pid: the pid in question.
3982 * @param: structure containing the RT priority.
3984 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3986 struct sched_param lp;
3987 struct task_struct *p;
3990 if (!param || pid < 0)
3994 p = find_process_by_pid(pid);
3999 retval = security_task_getscheduler(p);
4003 lp.sched_priority = p->rt_priority;
4007 * This one might sleep, we cannot do it with a spinlock held ...
4009 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4018 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4020 cpumask_var_t cpus_allowed, new_mask;
4021 struct task_struct *p;
4027 p = find_process_by_pid(pid);
4034 /* Prevent p going away */
4038 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4042 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4044 goto out_free_cpus_allowed;
4047 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4050 retval = security_task_setscheduler(p);
4054 cpuset_cpus_allowed(p, cpus_allowed);
4055 cpumask_and(new_mask, in_mask, cpus_allowed);
4057 retval = set_cpus_allowed_ptr(p, new_mask);
4060 cpuset_cpus_allowed(p, cpus_allowed);
4061 if (!cpumask_subset(new_mask, cpus_allowed)) {
4063 * We must have raced with a concurrent cpuset
4064 * update. Just reset the cpus_allowed to the
4065 * cpuset's cpus_allowed
4067 cpumask_copy(new_mask, cpus_allowed);
4072 free_cpumask_var(new_mask);
4073 out_free_cpus_allowed:
4074 free_cpumask_var(cpus_allowed);
4081 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4082 struct cpumask *new_mask)
4084 if (len < cpumask_size())
4085 cpumask_clear(new_mask);
4086 else if (len > cpumask_size())
4087 len = cpumask_size();
4089 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4093 * sys_sched_setaffinity - set the cpu affinity of a process
4094 * @pid: pid of the process
4095 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4096 * @user_mask_ptr: user-space pointer to the new cpu mask
4098 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4099 unsigned long __user *, user_mask_ptr)
4101 cpumask_var_t new_mask;
4104 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4107 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4109 retval = sched_setaffinity(pid, new_mask);
4110 free_cpumask_var(new_mask);
4114 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4116 struct task_struct *p;
4117 unsigned long flags;
4124 p = find_process_by_pid(pid);
4128 retval = security_task_getscheduler(p);
4132 raw_spin_lock_irqsave(&p->pi_lock, flags);
4133 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4134 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4144 * sys_sched_getaffinity - get the cpu affinity of a process
4145 * @pid: pid of the process
4146 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4147 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4149 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4150 unsigned long __user *, user_mask_ptr)
4155 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4157 if (len & (sizeof(unsigned long)-1))
4160 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4163 ret = sched_getaffinity(pid, mask);
4165 size_t retlen = min_t(size_t, len, cpumask_size());
4167 if (copy_to_user(user_mask_ptr, mask, retlen))
4172 free_cpumask_var(mask);
4178 * sys_sched_yield - yield the current processor to other threads.
4180 * This function yields the current CPU to other tasks. If there are no
4181 * other threads running on this CPU then this function will return.
4183 SYSCALL_DEFINE0(sched_yield)
4185 struct rq *rq = this_rq_lock();
4187 schedstat_inc(rq, yld_count);
4188 current->sched_class->yield_task(rq);
4191 * Since we are going to call schedule() anyway, there's
4192 * no need to preempt or enable interrupts:
4194 __release(rq->lock);
4195 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4196 do_raw_spin_unlock(&rq->lock);
4197 sched_preempt_enable_no_resched();
4204 static inline int should_resched(void)
4206 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4209 static void __cond_resched(void)
4211 add_preempt_count(PREEMPT_ACTIVE);
4213 sub_preempt_count(PREEMPT_ACTIVE);
4216 int __sched _cond_resched(void)
4218 if (should_resched()) {
4224 EXPORT_SYMBOL(_cond_resched);
4227 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4228 * call schedule, and on return reacquire the lock.
4230 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4231 * operations here to prevent schedule() from being called twice (once via
4232 * spin_unlock(), once by hand).
4234 int __cond_resched_lock(spinlock_t *lock)
4236 int resched = should_resched();
4239 lockdep_assert_held(lock);
4241 if (spin_needbreak(lock) || resched) {
4252 EXPORT_SYMBOL(__cond_resched_lock);
4254 int __sched __cond_resched_softirq(void)
4256 BUG_ON(!in_softirq());
4258 if (should_resched()) {
4266 EXPORT_SYMBOL(__cond_resched_softirq);
4269 * yield - yield the current processor to other threads.
4271 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4273 * The scheduler is at all times free to pick the calling task as the most
4274 * eligible task to run, if removing the yield() call from your code breaks
4275 * it, its already broken.
4277 * Typical broken usage is:
4282 * where one assumes that yield() will let 'the other' process run that will
4283 * make event true. If the current task is a SCHED_FIFO task that will never
4284 * happen. Never use yield() as a progress guarantee!!
4286 * If you want to use yield() to wait for something, use wait_event().
4287 * If you want to use yield() to be 'nice' for others, use cond_resched().
4288 * If you still want to use yield(), do not!
4290 void __sched yield(void)
4292 set_current_state(TASK_RUNNING);
4295 EXPORT_SYMBOL(yield);
4298 * yield_to - yield the current processor to another thread in
4299 * your thread group, or accelerate that thread toward the
4300 * processor it's on.
4302 * @preempt: whether task preemption is allowed or not
4304 * It's the caller's job to ensure that the target task struct
4305 * can't go away on us before we can do any checks.
4307 * Returns true if we indeed boosted the target task.
4309 bool __sched yield_to(struct task_struct *p, bool preempt)
4311 struct task_struct *curr = current;
4312 struct rq *rq, *p_rq;
4313 unsigned long flags;
4316 local_irq_save(flags);
4321 double_rq_lock(rq, p_rq);
4322 while (task_rq(p) != p_rq) {
4323 double_rq_unlock(rq, p_rq);
4327 if (!curr->sched_class->yield_to_task)
4330 if (curr->sched_class != p->sched_class)
4333 if (task_running(p_rq, p) || p->state)
4336 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4338 schedstat_inc(rq, yld_count);
4340 * Make p's CPU reschedule; pick_next_entity takes care of
4343 if (preempt && rq != p_rq)
4344 resched_task(p_rq->curr);
4348 double_rq_unlock(rq, p_rq);
4349 local_irq_restore(flags);
4356 EXPORT_SYMBOL_GPL(yield_to);
4359 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4360 * that process accounting knows that this is a task in IO wait state.
4362 void __sched io_schedule(void)
4364 struct rq *rq = raw_rq();
4366 delayacct_blkio_start();
4367 atomic_inc(&rq->nr_iowait);
4368 blk_flush_plug(current);
4369 current->in_iowait = 1;
4371 current->in_iowait = 0;
4372 atomic_dec(&rq->nr_iowait);
4373 delayacct_blkio_end();
4375 EXPORT_SYMBOL(io_schedule);
4377 long __sched io_schedule_timeout(long timeout)
4379 struct rq *rq = raw_rq();
4382 delayacct_blkio_start();
4383 atomic_inc(&rq->nr_iowait);
4384 blk_flush_plug(current);
4385 current->in_iowait = 1;
4386 ret = schedule_timeout(timeout);
4387 current->in_iowait = 0;
4388 atomic_dec(&rq->nr_iowait);
4389 delayacct_blkio_end();
4394 * sys_sched_get_priority_max - return maximum RT priority.
4395 * @policy: scheduling class.
4397 * this syscall returns the maximum rt_priority that can be used
4398 * by a given scheduling class.
4400 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4407 ret = MAX_USER_RT_PRIO-1;
4419 * sys_sched_get_priority_min - return minimum RT priority.
4420 * @policy: scheduling class.
4422 * this syscall returns the minimum rt_priority that can be used
4423 * by a given scheduling class.
4425 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4443 * sys_sched_rr_get_interval - return the default timeslice of a process.
4444 * @pid: pid of the process.
4445 * @interval: userspace pointer to the timeslice value.
4447 * this syscall writes the default timeslice value of a given process
4448 * into the user-space timespec buffer. A value of '0' means infinity.
4450 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4451 struct timespec __user *, interval)
4453 struct task_struct *p;
4454 unsigned int time_slice;
4455 unsigned long flags;
4465 p = find_process_by_pid(pid);
4469 retval = security_task_getscheduler(p);
4473 rq = task_rq_lock(p, &flags);
4474 time_slice = p->sched_class->get_rr_interval(rq, p);
4475 task_rq_unlock(rq, p, &flags);
4478 jiffies_to_timespec(time_slice, &t);
4479 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4487 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4489 void sched_show_task(struct task_struct *p)
4491 unsigned long free = 0;
4494 state = p->state ? __ffs(p->state) + 1 : 0;
4495 printk(KERN_INFO "%-15.15s %c", p->comm,
4496 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4497 #if BITS_PER_LONG == 32
4498 if (state == TASK_RUNNING)
4499 printk(KERN_CONT " running ");
4501 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4503 if (state == TASK_RUNNING)
4504 printk(KERN_CONT " running task ");
4506 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4508 #ifdef CONFIG_DEBUG_STACK_USAGE
4509 free = stack_not_used(p);
4511 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4512 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4513 (unsigned long)task_thread_info(p)->flags);
4515 show_stack(p, NULL);
4518 void show_state_filter(unsigned long state_filter)
4520 struct task_struct *g, *p;
4522 #if BITS_PER_LONG == 32
4524 " task PC stack pid father\n");
4527 " task PC stack pid father\n");
4530 do_each_thread(g, p) {
4532 * reset the NMI-timeout, listing all files on a slow
4533 * console might take a lot of time:
4535 touch_nmi_watchdog();
4536 if (!state_filter || (p->state & state_filter))
4538 } while_each_thread(g, p);
4540 touch_all_softlockup_watchdogs();
4542 #ifdef CONFIG_SCHED_DEBUG
4543 sysrq_sched_debug_show();
4547 * Only show locks if all tasks are dumped:
4550 debug_show_all_locks();
4553 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4555 idle->sched_class = &idle_sched_class;
4559 * init_idle - set up an idle thread for a given CPU
4560 * @idle: task in question
4561 * @cpu: cpu the idle task belongs to
4563 * NOTE: this function does not set the idle thread's NEED_RESCHED
4564 * flag, to make booting more robust.
4566 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4568 struct rq *rq = cpu_rq(cpu);
4569 unsigned long flags;
4571 raw_spin_lock_irqsave(&rq->lock, flags);
4574 idle->state = TASK_RUNNING;
4575 idle->se.exec_start = sched_clock();
4577 do_set_cpus_allowed(idle, cpumask_of(cpu));
4579 * We're having a chicken and egg problem, even though we are
4580 * holding rq->lock, the cpu isn't yet set to this cpu so the
4581 * lockdep check in task_group() will fail.
4583 * Similar case to sched_fork(). / Alternatively we could
4584 * use task_rq_lock() here and obtain the other rq->lock.
4589 __set_task_cpu(idle, cpu);
4592 rq->curr = rq->idle = idle;
4593 #if defined(CONFIG_SMP)
4596 raw_spin_unlock_irqrestore(&rq->lock, flags);
4598 /* Set the preempt count _outside_ the spinlocks! */
4599 task_thread_info(idle)->preempt_count = 0;
4602 * The idle tasks have their own, simple scheduling class:
4604 idle->sched_class = &idle_sched_class;
4605 ftrace_graph_init_idle_task(idle, cpu);
4606 #if defined(CONFIG_SMP)
4607 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4612 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4614 if (p->sched_class && p->sched_class->set_cpus_allowed)
4615 p->sched_class->set_cpus_allowed(p, new_mask);
4617 cpumask_copy(&p->cpus_allowed, new_mask);
4618 p->nr_cpus_allowed = cpumask_weight(new_mask);
4622 * This is how migration works:
4624 * 1) we invoke migration_cpu_stop() on the target CPU using
4626 * 2) stopper starts to run (implicitly forcing the migrated thread
4628 * 3) it checks whether the migrated task is still in the wrong runqueue.
4629 * 4) if it's in the wrong runqueue then the migration thread removes
4630 * it and puts it into the right queue.
4631 * 5) stopper completes and stop_one_cpu() returns and the migration
4636 * Change a given task's CPU affinity. Migrate the thread to a
4637 * proper CPU and schedule it away if the CPU it's executing on
4638 * is removed from the allowed bitmask.
4640 * NOTE: the caller must have a valid reference to the task, the
4641 * task must not exit() & deallocate itself prematurely. The
4642 * call is not atomic; no spinlocks may be held.
4644 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4646 unsigned long flags;
4648 unsigned int dest_cpu;
4651 rq = task_rq_lock(p, &flags);
4653 if (cpumask_equal(&p->cpus_allowed, new_mask))
4656 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4661 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4666 do_set_cpus_allowed(p, new_mask);
4668 /* Can the task run on the task's current CPU? If so, we're done */
4669 if (cpumask_test_cpu(task_cpu(p), new_mask))
4672 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4674 struct migration_arg arg = { p, dest_cpu };
4675 /* Need help from migration thread: drop lock and wait. */
4676 task_rq_unlock(rq, p, &flags);
4677 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4678 tlb_migrate_finish(p->mm);
4682 task_rq_unlock(rq, p, &flags);
4686 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4689 * Move (not current) task off this cpu, onto dest cpu. We're doing
4690 * this because either it can't run here any more (set_cpus_allowed()
4691 * away from this CPU, or CPU going down), or because we're
4692 * attempting to rebalance this task on exec (sched_exec).
4694 * So we race with normal scheduler movements, but that's OK, as long
4695 * as the task is no longer on this CPU.
4697 * Returns non-zero if task was successfully migrated.
4699 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4701 struct rq *rq_dest, *rq_src;
4704 if (unlikely(!cpu_active(dest_cpu)))
4707 rq_src = cpu_rq(src_cpu);
4708 rq_dest = cpu_rq(dest_cpu);
4710 raw_spin_lock(&p->pi_lock);
4711 double_rq_lock(rq_src, rq_dest);
4712 /* Already moved. */
4713 if (task_cpu(p) != src_cpu)
4715 /* Affinity changed (again). */
4716 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4720 * If we're not on a rq, the next wake-up will ensure we're
4724 dequeue_task(rq_src, p, 0);
4725 set_task_cpu(p, dest_cpu);
4726 enqueue_task(rq_dest, p, 0);
4727 check_preempt_curr(rq_dest, p, 0);
4732 double_rq_unlock(rq_src, rq_dest);
4733 raw_spin_unlock(&p->pi_lock);
4738 * migration_cpu_stop - this will be executed by a highprio stopper thread
4739 * and performs thread migration by bumping thread off CPU then
4740 * 'pushing' onto another runqueue.
4742 static int migration_cpu_stop(void *data)
4744 struct migration_arg *arg = data;
4747 * The original target cpu might have gone down and we might
4748 * be on another cpu but it doesn't matter.
4750 local_irq_disable();
4751 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4756 #ifdef CONFIG_HOTPLUG_CPU
4759 * Ensures that the idle task is using init_mm right before its cpu goes
4762 void idle_task_exit(void)
4764 struct mm_struct *mm = current->active_mm;
4766 BUG_ON(cpu_online(smp_processor_id()));
4769 switch_mm(mm, &init_mm, current);
4774 * Since this CPU is going 'away' for a while, fold any nr_active delta
4775 * we might have. Assumes we're called after migrate_tasks() so that the
4776 * nr_active count is stable.
4778 * Also see the comment "Global load-average calculations".
4780 static void calc_load_migrate(struct rq *rq)
4782 long delta = calc_load_fold_active(rq);
4784 atomic_long_add(delta, &calc_load_tasks);
4788 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4789 * try_to_wake_up()->select_task_rq().
4791 * Called with rq->lock held even though we'er in stop_machine() and
4792 * there's no concurrency possible, we hold the required locks anyway
4793 * because of lock validation efforts.
4795 static void migrate_tasks(unsigned int dead_cpu)
4797 struct rq *rq = cpu_rq(dead_cpu);
4798 struct task_struct *next, *stop = rq->stop;
4802 * Fudge the rq selection such that the below task selection loop
4803 * doesn't get stuck on the currently eligible stop task.
4805 * We're currently inside stop_machine() and the rq is either stuck
4806 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4807 * either way we should never end up calling schedule() until we're
4814 * There's this thread running, bail when that's the only
4817 if (rq->nr_running == 1)
4820 next = pick_next_task(rq);
4822 next->sched_class->put_prev_task(rq, next);
4824 /* Find suitable destination for @next, with force if needed. */
4825 dest_cpu = select_fallback_rq(dead_cpu, next);
4826 raw_spin_unlock(&rq->lock);
4828 __migrate_task(next, dead_cpu, dest_cpu);
4830 raw_spin_lock(&rq->lock);
4836 #endif /* CONFIG_HOTPLUG_CPU */
4838 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4840 static struct ctl_table sd_ctl_dir[] = {
4842 .procname = "sched_domain",
4848 static struct ctl_table sd_ctl_root[] = {
4850 .procname = "kernel",
4852 .child = sd_ctl_dir,
4857 static struct ctl_table *sd_alloc_ctl_entry(int n)
4859 struct ctl_table *entry =
4860 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4865 static void sd_free_ctl_entry(struct ctl_table **tablep)
4867 struct ctl_table *entry;
4870 * In the intermediate directories, both the child directory and
4871 * procname are dynamically allocated and could fail but the mode
4872 * will always be set. In the lowest directory the names are
4873 * static strings and all have proc handlers.
4875 for (entry = *tablep; entry->mode; entry++) {
4877 sd_free_ctl_entry(&entry->child);
4878 if (entry->proc_handler == NULL)
4879 kfree(entry->procname);
4886 static int min_load_idx = 0;
4887 static int max_load_idx = CPU_LOAD_IDX_MAX;
4890 set_table_entry(struct ctl_table *entry,
4891 const char *procname, void *data, int maxlen,
4892 umode_t mode, proc_handler *proc_handler,
4895 entry->procname = procname;
4897 entry->maxlen = maxlen;
4899 entry->proc_handler = proc_handler;
4902 entry->extra1 = &min_load_idx;
4903 entry->extra2 = &max_load_idx;
4907 static struct ctl_table *
4908 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4910 struct ctl_table *table = sd_alloc_ctl_entry(13);
4915 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4916 sizeof(long), 0644, proc_doulongvec_minmax, false);
4917 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4918 sizeof(long), 0644, proc_doulongvec_minmax, false);
4919 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4920 sizeof(int), 0644, proc_dointvec_minmax, true);
4921 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4922 sizeof(int), 0644, proc_dointvec_minmax, true);
4923 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4924 sizeof(int), 0644, proc_dointvec_minmax, true);
4925 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4926 sizeof(int), 0644, proc_dointvec_minmax, true);
4927 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4928 sizeof(int), 0644, proc_dointvec_minmax, true);
4929 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4930 sizeof(int), 0644, proc_dointvec_minmax, false);
4931 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4932 sizeof(int), 0644, proc_dointvec_minmax, false);
4933 set_table_entry(&table[9], "cache_nice_tries",
4934 &sd->cache_nice_tries,
4935 sizeof(int), 0644, proc_dointvec_minmax, false);
4936 set_table_entry(&table[10], "flags", &sd->flags,
4937 sizeof(int), 0644, proc_dointvec_minmax, false);
4938 set_table_entry(&table[11], "name", sd->name,
4939 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4940 /* &table[12] is terminator */
4945 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4947 struct ctl_table *entry, *table;
4948 struct sched_domain *sd;
4949 int domain_num = 0, i;
4952 for_each_domain(cpu, sd)
4954 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4959 for_each_domain(cpu, sd) {
4960 snprintf(buf, 32, "domain%d", i);
4961 entry->procname = kstrdup(buf, GFP_KERNEL);
4963 entry->child = sd_alloc_ctl_domain_table(sd);
4970 static struct ctl_table_header *sd_sysctl_header;
4971 static void register_sched_domain_sysctl(void)
4973 int i, cpu_num = num_possible_cpus();
4974 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4977 WARN_ON(sd_ctl_dir[0].child);
4978 sd_ctl_dir[0].child = entry;
4983 for_each_possible_cpu(i) {
4984 snprintf(buf, 32, "cpu%d", i);
4985 entry->procname = kstrdup(buf, GFP_KERNEL);
4987 entry->child = sd_alloc_ctl_cpu_table(i);
4991 WARN_ON(sd_sysctl_header);
4992 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4995 /* may be called multiple times per register */
4996 static void unregister_sched_domain_sysctl(void)
4998 if (sd_sysctl_header)
4999 unregister_sysctl_table(sd_sysctl_header);
5000 sd_sysctl_header = NULL;
5001 if (sd_ctl_dir[0].child)
5002 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5005 static void register_sched_domain_sysctl(void)
5008 static void unregister_sched_domain_sysctl(void)
5013 static void set_rq_online(struct rq *rq)
5016 const struct sched_class *class;
5018 cpumask_set_cpu(rq->cpu, rq->rd->online);
5021 for_each_class(class) {
5022 if (class->rq_online)
5023 class->rq_online(rq);
5028 static void set_rq_offline(struct rq *rq)
5031 const struct sched_class *class;
5033 for_each_class(class) {
5034 if (class->rq_offline)
5035 class->rq_offline(rq);
5038 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5044 * migration_call - callback that gets triggered when a CPU is added.
5045 * Here we can start up the necessary migration thread for the new CPU.
5047 static int __cpuinit
5048 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5050 int cpu = (long)hcpu;
5051 unsigned long flags;
5052 struct rq *rq = cpu_rq(cpu);
5054 switch (action & ~CPU_TASKS_FROZEN) {
5056 case CPU_UP_PREPARE:
5057 rq->calc_load_update = calc_load_update;
5061 /* Update our root-domain */
5062 raw_spin_lock_irqsave(&rq->lock, flags);
5064 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5068 raw_spin_unlock_irqrestore(&rq->lock, flags);
5071 #ifdef CONFIG_HOTPLUG_CPU
5073 sched_ttwu_pending();
5074 /* Update our root-domain */
5075 raw_spin_lock_irqsave(&rq->lock, flags);
5077 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5081 BUG_ON(rq->nr_running != 1); /* the migration thread */
5082 raw_spin_unlock_irqrestore(&rq->lock, flags);
5086 calc_load_migrate(rq);
5091 update_max_interval();
5097 * Register at high priority so that task migration (migrate_all_tasks)
5098 * happens before everything else. This has to be lower priority than
5099 * the notifier in the perf_event subsystem, though.
5101 static struct notifier_block __cpuinitdata migration_notifier = {
5102 .notifier_call = migration_call,
5103 .priority = CPU_PRI_MIGRATION,
5106 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5107 unsigned long action, void *hcpu)
5109 switch (action & ~CPU_TASKS_FROZEN) {
5111 case CPU_DOWN_FAILED:
5112 set_cpu_active((long)hcpu, true);
5119 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5120 unsigned long action, void *hcpu)
5122 switch (action & ~CPU_TASKS_FROZEN) {
5123 case CPU_DOWN_PREPARE:
5124 set_cpu_active((long)hcpu, false);
5131 static int __init migration_init(void)
5133 void *cpu = (void *)(long)smp_processor_id();
5136 /* Initialize migration for the boot CPU */
5137 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5138 BUG_ON(err == NOTIFY_BAD);
5139 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5140 register_cpu_notifier(&migration_notifier);
5142 /* Register cpu active notifiers */
5143 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5144 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5148 early_initcall(migration_init);
5153 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5155 #ifdef CONFIG_SCHED_DEBUG
5157 static __read_mostly int sched_debug_enabled;
5159 static int __init sched_debug_setup(char *str)
5161 sched_debug_enabled = 1;
5165 early_param("sched_debug", sched_debug_setup);
5167 static inline bool sched_debug(void)
5169 return sched_debug_enabled;
5172 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5173 struct cpumask *groupmask)
5175 struct sched_group *group = sd->groups;
5178 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5179 cpumask_clear(groupmask);
5181 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5183 if (!(sd->flags & SD_LOAD_BALANCE)) {
5184 printk("does not load-balance\n");
5186 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5191 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5193 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5194 printk(KERN_ERR "ERROR: domain->span does not contain "
5197 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5198 printk(KERN_ERR "ERROR: domain->groups does not contain"
5202 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5206 printk(KERN_ERR "ERROR: group is NULL\n");
5211 * Even though we initialize ->power to something semi-sane,
5212 * we leave power_orig unset. This allows us to detect if
5213 * domain iteration is still funny without causing /0 traps.
5215 if (!group->sgp->power_orig) {
5216 printk(KERN_CONT "\n");
5217 printk(KERN_ERR "ERROR: domain->cpu_power not "
5222 if (!cpumask_weight(sched_group_cpus(group))) {
5223 printk(KERN_CONT "\n");
5224 printk(KERN_ERR "ERROR: empty group\n");
5228 if (!(sd->flags & SD_OVERLAP) &&
5229 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5230 printk(KERN_CONT "\n");
5231 printk(KERN_ERR "ERROR: repeated CPUs\n");
5235 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5237 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5239 printk(KERN_CONT " %s", str);
5240 if (group->sgp->power != SCHED_POWER_SCALE) {
5241 printk(KERN_CONT " (cpu_power = %d)",
5245 group = group->next;
5246 } while (group != sd->groups);
5247 printk(KERN_CONT "\n");
5249 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5250 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5253 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5254 printk(KERN_ERR "ERROR: parent span is not a superset "
5255 "of domain->span\n");
5259 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5263 if (!sched_debug_enabled)
5267 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5271 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5274 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5282 #else /* !CONFIG_SCHED_DEBUG */
5283 # define sched_domain_debug(sd, cpu) do { } while (0)
5284 static inline bool sched_debug(void)
5288 #endif /* CONFIG_SCHED_DEBUG */
5290 static int sd_degenerate(struct sched_domain *sd)
5292 if (cpumask_weight(sched_domain_span(sd)) == 1)
5295 /* Following flags need at least 2 groups */
5296 if (sd->flags & (SD_LOAD_BALANCE |
5297 SD_BALANCE_NEWIDLE |
5301 SD_SHARE_PKG_RESOURCES)) {
5302 if (sd->groups != sd->groups->next)
5306 /* Following flags don't use groups */
5307 if (sd->flags & (SD_WAKE_AFFINE))
5314 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5316 unsigned long cflags = sd->flags, pflags = parent->flags;
5318 if (sd_degenerate(parent))
5321 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5324 /* Flags needing groups don't count if only 1 group in parent */
5325 if (parent->groups == parent->groups->next) {
5326 pflags &= ~(SD_LOAD_BALANCE |
5327 SD_BALANCE_NEWIDLE |
5331 SD_SHARE_PKG_RESOURCES);
5332 if (nr_node_ids == 1)
5333 pflags &= ~SD_SERIALIZE;
5335 if (~cflags & pflags)
5341 static void free_rootdomain(struct rcu_head *rcu)
5343 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5345 cpupri_cleanup(&rd->cpupri);
5346 free_cpumask_var(rd->rto_mask);
5347 free_cpumask_var(rd->online);
5348 free_cpumask_var(rd->span);
5352 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5354 struct root_domain *old_rd = NULL;
5355 unsigned long flags;
5357 raw_spin_lock_irqsave(&rq->lock, flags);
5362 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5365 cpumask_clear_cpu(rq->cpu, old_rd->span);
5368 * If we dont want to free the old_rt yet then
5369 * set old_rd to NULL to skip the freeing later
5372 if (!atomic_dec_and_test(&old_rd->refcount))
5376 atomic_inc(&rd->refcount);
5379 cpumask_set_cpu(rq->cpu, rd->span);
5380 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5383 raw_spin_unlock_irqrestore(&rq->lock, flags);
5386 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5389 static int init_rootdomain(struct root_domain *rd)
5391 memset(rd, 0, sizeof(*rd));
5393 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5395 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5397 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5400 if (cpupri_init(&rd->cpupri) != 0)
5405 free_cpumask_var(rd->rto_mask);
5407 free_cpumask_var(rd->online);
5409 free_cpumask_var(rd->span);
5415 * By default the system creates a single root-domain with all cpus as
5416 * members (mimicking the global state we have today).
5418 struct root_domain def_root_domain;
5420 static void init_defrootdomain(void)
5422 init_rootdomain(&def_root_domain);
5424 atomic_set(&def_root_domain.refcount, 1);
5427 static struct root_domain *alloc_rootdomain(void)
5429 struct root_domain *rd;
5431 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5435 if (init_rootdomain(rd) != 0) {
5443 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5445 struct sched_group *tmp, *first;
5454 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5459 } while (sg != first);
5462 static void free_sched_domain(struct rcu_head *rcu)
5464 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5467 * If its an overlapping domain it has private groups, iterate and
5470 if (sd->flags & SD_OVERLAP) {
5471 free_sched_groups(sd->groups, 1);
5472 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5473 kfree(sd->groups->sgp);
5479 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5481 call_rcu(&sd->rcu, free_sched_domain);
5484 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5486 for (; sd; sd = sd->parent)
5487 destroy_sched_domain(sd, cpu);
5491 * Keep a special pointer to the highest sched_domain that has
5492 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5493 * allows us to avoid some pointer chasing select_idle_sibling().
5495 * Also keep a unique ID per domain (we use the first cpu number in
5496 * the cpumask of the domain), this allows us to quickly tell if
5497 * two cpus are in the same cache domain, see cpus_share_cache().
5499 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5500 DEFINE_PER_CPU(int, sd_llc_id);
5502 DEFINE_PER_CPU(struct sched_domain *, sd_node);
5504 static void update_domain_cache(int cpu)
5506 struct sched_domain *sd;
5509 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5511 id = cpumask_first(sched_domain_span(sd));
5513 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5514 per_cpu(sd_llc_id, cpu) = id;
5516 for_each_domain(cpu, sd) {
5517 if (cpumask_equal(sched_domain_span(sd),
5518 cpumask_of_node(cpu_to_node(cpu))))
5523 rcu_assign_pointer(per_cpu(sd_node, cpu), sd);
5527 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5528 * hold the hotplug lock.
5531 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5533 struct rq *rq = cpu_rq(cpu);
5534 struct sched_domain *tmp;
5536 /* Remove the sched domains which do not contribute to scheduling. */
5537 for (tmp = sd; tmp; ) {
5538 struct sched_domain *parent = tmp->parent;
5542 if (sd_parent_degenerate(tmp, parent)) {
5543 tmp->parent = parent->parent;
5545 parent->parent->child = tmp;
5546 destroy_sched_domain(parent, cpu);
5551 if (sd && sd_degenerate(sd)) {
5554 destroy_sched_domain(tmp, cpu);
5559 sched_domain_debug(sd, cpu);
5561 rq_attach_root(rq, rd);
5563 rcu_assign_pointer(rq->sd, sd);
5564 destroy_sched_domains(tmp, cpu);
5566 update_domain_cache(cpu);
5569 /* cpus with isolated domains */
5570 static cpumask_var_t cpu_isolated_map;
5572 /* Setup the mask of cpus configured for isolated domains */
5573 static int __init isolated_cpu_setup(char *str)
5575 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5576 cpulist_parse(str, cpu_isolated_map);
5580 __setup("isolcpus=", isolated_cpu_setup);
5582 static const struct cpumask *cpu_cpu_mask(int cpu)
5584 return cpumask_of_node(cpu_to_node(cpu));
5588 struct sched_domain **__percpu sd;
5589 struct sched_group **__percpu sg;
5590 struct sched_group_power **__percpu sgp;
5594 struct sched_domain ** __percpu sd;
5595 struct root_domain *rd;
5605 struct sched_domain_topology_level;
5607 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5608 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5610 #define SDTL_OVERLAP 0x01
5612 struct sched_domain_topology_level {
5613 sched_domain_init_f init;
5614 sched_domain_mask_f mask;
5617 struct sd_data data;
5621 * Build an iteration mask that can exclude certain CPUs from the upwards
5624 * Asymmetric node setups can result in situations where the domain tree is of
5625 * unequal depth, make sure to skip domains that already cover the entire
5628 * In that case build_sched_domains() will have terminated the iteration early
5629 * and our sibling sd spans will be empty. Domains should always include the
5630 * cpu they're built on, so check that.
5633 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5635 const struct cpumask *span = sched_domain_span(sd);
5636 struct sd_data *sdd = sd->private;
5637 struct sched_domain *sibling;
5640 for_each_cpu(i, span) {
5641 sibling = *per_cpu_ptr(sdd->sd, i);
5642 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5645 cpumask_set_cpu(i, sched_group_mask(sg));
5650 * Return the canonical balance cpu for this group, this is the first cpu
5651 * of this group that's also in the iteration mask.
5653 int group_balance_cpu(struct sched_group *sg)
5655 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5659 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5661 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5662 const struct cpumask *span = sched_domain_span(sd);
5663 struct cpumask *covered = sched_domains_tmpmask;
5664 struct sd_data *sdd = sd->private;
5665 struct sched_domain *child;
5668 cpumask_clear(covered);
5670 for_each_cpu(i, span) {
5671 struct cpumask *sg_span;
5673 if (cpumask_test_cpu(i, covered))
5676 child = *per_cpu_ptr(sdd->sd, i);
5678 /* See the comment near build_group_mask(). */
5679 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5682 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5683 GFP_KERNEL, cpu_to_node(cpu));
5688 sg_span = sched_group_cpus(sg);
5690 child = child->child;
5691 cpumask_copy(sg_span, sched_domain_span(child));
5693 cpumask_set_cpu(i, sg_span);
5695 cpumask_or(covered, covered, sg_span);
5697 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5698 if (atomic_inc_return(&sg->sgp->ref) == 1)
5699 build_group_mask(sd, sg);
5702 * Initialize sgp->power such that even if we mess up the
5703 * domains and no possible iteration will get us here, we won't
5706 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5709 * Make sure the first group of this domain contains the
5710 * canonical balance cpu. Otherwise the sched_domain iteration
5711 * breaks. See update_sg_lb_stats().
5713 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5714 group_balance_cpu(sg) == cpu)
5724 sd->groups = groups;
5729 free_sched_groups(first, 0);
5734 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5736 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5737 struct sched_domain *child = sd->child;
5740 cpu = cpumask_first(sched_domain_span(child));
5743 *sg = *per_cpu_ptr(sdd->sg, cpu);
5744 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5745 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5752 * build_sched_groups will build a circular linked list of the groups
5753 * covered by the given span, and will set each group's ->cpumask correctly,
5754 * and ->cpu_power to 0.
5756 * Assumes the sched_domain tree is fully constructed
5759 build_sched_groups(struct sched_domain *sd, int cpu)
5761 struct sched_group *first = NULL, *last = NULL;
5762 struct sd_data *sdd = sd->private;
5763 const struct cpumask *span = sched_domain_span(sd);
5764 struct cpumask *covered;
5767 get_group(cpu, sdd, &sd->groups);
5768 atomic_inc(&sd->groups->ref);
5770 if (cpu != cpumask_first(sched_domain_span(sd)))
5773 lockdep_assert_held(&sched_domains_mutex);
5774 covered = sched_domains_tmpmask;
5776 cpumask_clear(covered);
5778 for_each_cpu(i, span) {
5779 struct sched_group *sg;
5780 int group = get_group(i, sdd, &sg);
5783 if (cpumask_test_cpu(i, covered))
5786 cpumask_clear(sched_group_cpus(sg));
5788 cpumask_setall(sched_group_mask(sg));
5790 for_each_cpu(j, span) {
5791 if (get_group(j, sdd, NULL) != group)
5794 cpumask_set_cpu(j, covered);
5795 cpumask_set_cpu(j, sched_group_cpus(sg));
5810 * Initialize sched groups cpu_power.
5812 * cpu_power indicates the capacity of sched group, which is used while
5813 * distributing the load between different sched groups in a sched domain.
5814 * Typically cpu_power for all the groups in a sched domain will be same unless
5815 * there are asymmetries in the topology. If there are asymmetries, group
5816 * having more cpu_power will pickup more load compared to the group having
5819 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5821 struct sched_group *sg = sd->groups;
5823 WARN_ON(!sd || !sg);
5826 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5828 } while (sg != sd->groups);
5830 if (cpu != group_balance_cpu(sg))
5833 update_group_power(sd, cpu);
5834 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5837 int __weak arch_sd_sibling_asym_packing(void)
5839 return 0*SD_ASYM_PACKING;
5843 * Initializers for schedule domains
5844 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5847 #ifdef CONFIG_SCHED_DEBUG
5848 # define SD_INIT_NAME(sd, type) sd->name = #type
5850 # define SD_INIT_NAME(sd, type) do { } while (0)
5853 #define SD_INIT_FUNC(type) \
5854 static noinline struct sched_domain * \
5855 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5857 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5858 *sd = SD_##type##_INIT; \
5859 SD_INIT_NAME(sd, type); \
5860 sd->private = &tl->data; \
5865 #ifdef CONFIG_SCHED_SMT
5866 SD_INIT_FUNC(SIBLING)
5868 #ifdef CONFIG_SCHED_MC
5871 #ifdef CONFIG_SCHED_BOOK
5875 static int default_relax_domain_level = -1;
5876 int sched_domain_level_max;
5878 static int __init setup_relax_domain_level(char *str)
5880 if (kstrtoint(str, 0, &default_relax_domain_level))
5881 pr_warn("Unable to set relax_domain_level\n");
5885 __setup("relax_domain_level=", setup_relax_domain_level);
5887 static void set_domain_attribute(struct sched_domain *sd,
5888 struct sched_domain_attr *attr)
5892 if (!attr || attr->relax_domain_level < 0) {
5893 if (default_relax_domain_level < 0)
5896 request = default_relax_domain_level;
5898 request = attr->relax_domain_level;
5899 if (request < sd->level) {
5900 /* turn off idle balance on this domain */
5901 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5903 /* turn on idle balance on this domain */
5904 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5908 static void __sdt_free(const struct cpumask *cpu_map);
5909 static int __sdt_alloc(const struct cpumask *cpu_map);
5911 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5912 const struct cpumask *cpu_map)
5916 if (!atomic_read(&d->rd->refcount))
5917 free_rootdomain(&d->rd->rcu); /* fall through */
5919 free_percpu(d->sd); /* fall through */
5921 __sdt_free(cpu_map); /* fall through */
5927 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5928 const struct cpumask *cpu_map)
5930 memset(d, 0, sizeof(*d));
5932 if (__sdt_alloc(cpu_map))
5933 return sa_sd_storage;
5934 d->sd = alloc_percpu(struct sched_domain *);
5936 return sa_sd_storage;
5937 d->rd = alloc_rootdomain();
5940 return sa_rootdomain;
5944 * NULL the sd_data elements we've used to build the sched_domain and
5945 * sched_group structure so that the subsequent __free_domain_allocs()
5946 * will not free the data we're using.
5948 static void claim_allocations(int cpu, struct sched_domain *sd)
5950 struct sd_data *sdd = sd->private;
5952 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5953 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5955 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5956 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5958 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5959 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5962 #ifdef CONFIG_SCHED_SMT
5963 static const struct cpumask *cpu_smt_mask(int cpu)
5965 return topology_thread_cpumask(cpu);
5970 * Topology list, bottom-up.
5972 static struct sched_domain_topology_level default_topology[] = {
5973 #ifdef CONFIG_SCHED_SMT
5974 { sd_init_SIBLING, cpu_smt_mask, },
5976 #ifdef CONFIG_SCHED_MC
5977 { sd_init_MC, cpu_coregroup_mask, },
5979 #ifdef CONFIG_SCHED_BOOK
5980 { sd_init_BOOK, cpu_book_mask, },
5982 { sd_init_CPU, cpu_cpu_mask, },
5986 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5988 #ifdef CONFIG_SCHED_NUMA
5991 * Requeues a task ensuring its on the right load-balance list so
5992 * that it might get migrated to its new home.
5994 * Since home-node is pure preference there's no hard migrate to force
5995 * us anywhere, this also allows us to call this from atomic context if
5998 void sched_setnode(struct task_struct *p, int node)
6000 unsigned long flags;
6004 rq = task_rq_lock(p, &flags);
6006 running = task_current(rq, p);
6009 dequeue_task(rq, p, 0);
6011 p->sched_class->put_prev_task(rq, p);
6016 p->sched_class->set_curr_task(rq);
6018 enqueue_task(rq, p, 0);
6019 task_rq_unlock(rq, p, &flags);
6022 #endif /* CONFIG_SCHED_NUMA */
6026 static int sched_domains_numa_levels;
6027 static int *sched_domains_numa_distance;
6028 static struct cpumask ***sched_domains_numa_masks;
6029 static int sched_domains_curr_level;
6031 static inline int sd_local_flags(int level)
6033 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6036 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6039 static struct sched_domain *
6040 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6042 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6043 int level = tl->numa_level;
6044 int sd_weight = cpumask_weight(
6045 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6047 *sd = (struct sched_domain){
6048 .min_interval = sd_weight,
6049 .max_interval = 2*sd_weight,
6051 .imbalance_pct = 125,
6052 .cache_nice_tries = 2,
6059 .flags = 1*SD_LOAD_BALANCE
6060 | 1*SD_BALANCE_NEWIDLE
6065 | 0*SD_SHARE_CPUPOWER
6066 | 0*SD_SHARE_PKG_RESOURCES
6068 | 0*SD_PREFER_SIBLING
6070 | sd_local_flags(level)
6072 .last_balance = jiffies,
6073 .balance_interval = sd_weight,
6075 SD_INIT_NAME(sd, NUMA);
6076 sd->private = &tl->data;
6079 * Ugly hack to pass state to sd_numa_mask()...
6081 sched_domains_curr_level = tl->numa_level;
6086 static const struct cpumask *sd_numa_mask(int cpu)
6088 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6091 static void sched_numa_warn(const char *str)
6093 static int done = false;
6101 printk(KERN_WARNING "ERROR: %s\n\n", str);
6103 for (i = 0; i < nr_node_ids; i++) {
6104 printk(KERN_WARNING " ");
6105 for (j = 0; j < nr_node_ids; j++)
6106 printk(KERN_CONT "%02d ", node_distance(i,j));
6107 printk(KERN_CONT "\n");
6109 printk(KERN_WARNING "\n");
6112 static bool find_numa_distance(int distance)
6116 if (distance == node_distance(0, 0))
6119 for (i = 0; i < sched_domains_numa_levels; i++) {
6120 if (sched_domains_numa_distance[i] == distance)
6127 static void sched_init_numa(void)
6129 int next_distance, curr_distance = node_distance(0, 0);
6130 struct sched_domain_topology_level *tl;
6134 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6135 if (!sched_domains_numa_distance)
6139 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6140 * unique distances in the node_distance() table.
6142 * Assumes node_distance(0,j) includes all distances in
6143 * node_distance(i,j) in order to avoid cubic time.
6145 next_distance = curr_distance;
6146 for (i = 0; i < nr_node_ids; i++) {
6147 for (j = 0; j < nr_node_ids; j++) {
6148 for (k = 0; k < nr_node_ids; k++) {
6149 int distance = node_distance(i, k);
6151 if (distance > curr_distance &&
6152 (distance < next_distance ||
6153 next_distance == curr_distance))
6154 next_distance = distance;
6157 * While not a strong assumption it would be nice to know
6158 * about cases where if node A is connected to B, B is not
6159 * equally connected to A.
6161 if (sched_debug() && node_distance(k, i) != distance)
6162 sched_numa_warn("Node-distance not symmetric");
6164 if (sched_debug() && i && !find_numa_distance(distance))
6165 sched_numa_warn("Node-0 not representative");
6167 if (next_distance != curr_distance) {
6168 sched_domains_numa_distance[level++] = next_distance;
6169 sched_domains_numa_levels = level;
6170 curr_distance = next_distance;
6175 * In case of sched_debug() we verify the above assumption.
6181 * 'level' contains the number of unique distances, excluding the
6182 * identity distance node_distance(i,i).
6184 * The sched_domains_nume_distance[] array includes the actual distance
6189 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6190 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6191 * the array will contain less then 'level' members. This could be
6192 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6193 * in other functions.
6195 * We reset it to 'level' at the end of this function.
6197 sched_domains_numa_levels = 0;
6199 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6200 if (!sched_domains_numa_masks)
6204 * Now for each level, construct a mask per node which contains all
6205 * cpus of nodes that are that many hops away from us.
6207 for (i = 0; i < level; i++) {
6208 sched_domains_numa_masks[i] =
6209 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6210 if (!sched_domains_numa_masks[i])
6213 for (j = 0; j < nr_node_ids; j++) {
6214 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6218 sched_domains_numa_masks[i][j] = mask;
6220 for (k = 0; k < nr_node_ids; k++) {
6221 if (node_distance(j, k) > sched_domains_numa_distance[i])
6224 cpumask_or(mask, mask, cpumask_of_node(k));
6229 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6230 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6235 * Copy the default topology bits..
6237 for (i = 0; default_topology[i].init; i++)
6238 tl[i] = default_topology[i];
6241 * .. and append 'j' levels of NUMA goodness.
6243 for (j = 0; j < level; i++, j++) {
6244 tl[i] = (struct sched_domain_topology_level){
6245 .init = sd_numa_init,
6246 .mask = sd_numa_mask,
6247 .flags = SDTL_OVERLAP,
6252 sched_domain_topology = tl;
6254 sched_domains_numa_levels = level;
6257 static void sched_domains_numa_masks_set(int cpu)
6260 int node = cpu_to_node(cpu);
6262 for (i = 0; i < sched_domains_numa_levels; i++) {
6263 for (j = 0; j < nr_node_ids; j++) {
6264 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6265 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6270 static void sched_domains_numa_masks_clear(int cpu)
6273 for (i = 0; i < sched_domains_numa_levels; i++) {
6274 for (j = 0; j < nr_node_ids; j++)
6275 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6280 * Update sched_domains_numa_masks[level][node] array when new cpus
6283 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6284 unsigned long action,
6287 int cpu = (long)hcpu;
6289 switch (action & ~CPU_TASKS_FROZEN) {
6291 sched_domains_numa_masks_set(cpu);
6295 sched_domains_numa_masks_clear(cpu);
6305 static inline void sched_init_numa(void)
6309 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6310 unsigned long action,
6315 #endif /* CONFIG_NUMA */
6317 static int __sdt_alloc(const struct cpumask *cpu_map)
6319 struct sched_domain_topology_level *tl;
6322 for (tl = sched_domain_topology; tl->init; tl++) {
6323 struct sd_data *sdd = &tl->data;
6325 sdd->sd = alloc_percpu(struct sched_domain *);
6329 sdd->sg = alloc_percpu(struct sched_group *);
6333 sdd->sgp = alloc_percpu(struct sched_group_power *);
6337 for_each_cpu(j, cpu_map) {
6338 struct sched_domain *sd;
6339 struct sched_group *sg;
6340 struct sched_group_power *sgp;
6342 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6343 GFP_KERNEL, cpu_to_node(j));
6347 *per_cpu_ptr(sdd->sd, j) = sd;
6349 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6350 GFP_KERNEL, cpu_to_node(j));
6356 *per_cpu_ptr(sdd->sg, j) = sg;
6358 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6359 GFP_KERNEL, cpu_to_node(j));
6363 *per_cpu_ptr(sdd->sgp, j) = sgp;
6370 static void __sdt_free(const struct cpumask *cpu_map)
6372 struct sched_domain_topology_level *tl;
6375 for (tl = sched_domain_topology; tl->init; tl++) {
6376 struct sd_data *sdd = &tl->data;
6378 for_each_cpu(j, cpu_map) {
6379 struct sched_domain *sd;
6382 sd = *per_cpu_ptr(sdd->sd, j);
6383 if (sd && (sd->flags & SD_OVERLAP))
6384 free_sched_groups(sd->groups, 0);
6385 kfree(*per_cpu_ptr(sdd->sd, j));
6389 kfree(*per_cpu_ptr(sdd->sg, j));
6391 kfree(*per_cpu_ptr(sdd->sgp, j));
6393 free_percpu(sdd->sd);
6395 free_percpu(sdd->sg);
6397 free_percpu(sdd->sgp);
6402 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6403 struct s_data *d, const struct cpumask *cpu_map,
6404 struct sched_domain_attr *attr, struct sched_domain *child,
6407 struct sched_domain *sd = tl->init(tl, cpu);
6411 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6413 sd->level = child->level + 1;
6414 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6418 set_domain_attribute(sd, attr);
6424 * Build sched domains for a given set of cpus and attach the sched domains
6425 * to the individual cpus
6427 static int build_sched_domains(const struct cpumask *cpu_map,
6428 struct sched_domain_attr *attr)
6430 enum s_alloc alloc_state = sa_none;
6431 struct sched_domain *sd;
6433 int i, ret = -ENOMEM;
6435 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6436 if (alloc_state != sa_rootdomain)
6439 /* Set up domains for cpus specified by the cpu_map. */
6440 for_each_cpu(i, cpu_map) {
6441 struct sched_domain_topology_level *tl;
6444 for (tl = sched_domain_topology; tl->init; tl++) {
6445 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6446 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6447 sd->flags |= SD_OVERLAP;
6448 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6455 *per_cpu_ptr(d.sd, i) = sd;
6458 /* Build the groups for the domains */
6459 for_each_cpu(i, cpu_map) {
6460 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6461 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6462 if (sd->flags & SD_OVERLAP) {
6463 if (build_overlap_sched_groups(sd, i))
6466 if (build_sched_groups(sd, i))
6472 /* Calculate CPU power for physical packages and nodes */
6473 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6474 if (!cpumask_test_cpu(i, cpu_map))
6477 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6478 claim_allocations(i, sd);
6479 init_sched_groups_power(i, sd);
6483 /* Attach the domains */
6485 for_each_cpu(i, cpu_map) {
6486 sd = *per_cpu_ptr(d.sd, i);
6487 cpu_attach_domain(sd, d.rd, i);
6493 __free_domain_allocs(&d, alloc_state, cpu_map);
6497 static cpumask_var_t *doms_cur; /* current sched domains */
6498 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6499 static struct sched_domain_attr *dattr_cur;
6500 /* attribues of custom domains in 'doms_cur' */
6503 * Special case: If a kmalloc of a doms_cur partition (array of
6504 * cpumask) fails, then fallback to a single sched domain,
6505 * as determined by the single cpumask fallback_doms.
6507 static cpumask_var_t fallback_doms;
6510 * arch_update_cpu_topology lets virtualized architectures update the
6511 * cpu core maps. It is supposed to return 1 if the topology changed
6512 * or 0 if it stayed the same.
6514 int __attribute__((weak)) arch_update_cpu_topology(void)
6519 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6522 cpumask_var_t *doms;
6524 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6527 for (i = 0; i < ndoms; i++) {
6528 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6529 free_sched_domains(doms, i);
6536 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6539 for (i = 0; i < ndoms; i++)
6540 free_cpumask_var(doms[i]);
6545 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6546 * For now this just excludes isolated cpus, but could be used to
6547 * exclude other special cases in the future.
6549 static int init_sched_domains(const struct cpumask *cpu_map)
6553 arch_update_cpu_topology();
6555 doms_cur = alloc_sched_domains(ndoms_cur);
6557 doms_cur = &fallback_doms;
6558 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6559 err = build_sched_domains(doms_cur[0], NULL);
6560 register_sched_domain_sysctl();
6566 * Detach sched domains from a group of cpus specified in cpu_map
6567 * These cpus will now be attached to the NULL domain
6569 static void detach_destroy_domains(const struct cpumask *cpu_map)
6574 for_each_cpu(i, cpu_map)
6575 cpu_attach_domain(NULL, &def_root_domain, i);
6579 /* handle null as "default" */
6580 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6581 struct sched_domain_attr *new, int idx_new)
6583 struct sched_domain_attr tmp;
6590 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6591 new ? (new + idx_new) : &tmp,
6592 sizeof(struct sched_domain_attr));
6596 * Partition sched domains as specified by the 'ndoms_new'
6597 * cpumasks in the array doms_new[] of cpumasks. This compares
6598 * doms_new[] to the current sched domain partitioning, doms_cur[].
6599 * It destroys each deleted domain and builds each new domain.
6601 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6602 * The masks don't intersect (don't overlap.) We should setup one
6603 * sched domain for each mask. CPUs not in any of the cpumasks will
6604 * not be load balanced. If the same cpumask appears both in the
6605 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6608 * The passed in 'doms_new' should be allocated using
6609 * alloc_sched_domains. This routine takes ownership of it and will
6610 * free_sched_domains it when done with it. If the caller failed the
6611 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6612 * and partition_sched_domains() will fallback to the single partition
6613 * 'fallback_doms', it also forces the domains to be rebuilt.
6615 * If doms_new == NULL it will be replaced with cpu_online_mask.
6616 * ndoms_new == 0 is a special case for destroying existing domains,
6617 * and it will not create the default domain.
6619 * Call with hotplug lock held
6621 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6622 struct sched_domain_attr *dattr_new)
6627 mutex_lock(&sched_domains_mutex);
6629 /* always unregister in case we don't destroy any domains */
6630 unregister_sched_domain_sysctl();
6632 /* Let architecture update cpu core mappings. */
6633 new_topology = arch_update_cpu_topology();
6635 n = doms_new ? ndoms_new : 0;
6637 /* Destroy deleted domains */
6638 for (i = 0; i < ndoms_cur; i++) {
6639 for (j = 0; j < n && !new_topology; j++) {
6640 if (cpumask_equal(doms_cur[i], doms_new[j])
6641 && dattrs_equal(dattr_cur, i, dattr_new, j))
6644 /* no match - a current sched domain not in new doms_new[] */
6645 detach_destroy_domains(doms_cur[i]);
6650 if (doms_new == NULL) {
6652 doms_new = &fallback_doms;
6653 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6654 WARN_ON_ONCE(dattr_new);
6657 /* Build new domains */
6658 for (i = 0; i < ndoms_new; i++) {
6659 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6660 if (cpumask_equal(doms_new[i], doms_cur[j])
6661 && dattrs_equal(dattr_new, i, dattr_cur, j))
6664 /* no match - add a new doms_new */
6665 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6670 /* Remember the new sched domains */
6671 if (doms_cur != &fallback_doms)
6672 free_sched_domains(doms_cur, ndoms_cur);
6673 kfree(dattr_cur); /* kfree(NULL) is safe */
6674 doms_cur = doms_new;
6675 dattr_cur = dattr_new;
6676 ndoms_cur = ndoms_new;
6678 register_sched_domain_sysctl();
6680 mutex_unlock(&sched_domains_mutex);
6683 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6686 * Update cpusets according to cpu_active mask. If cpusets are
6687 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6688 * around partition_sched_domains().
6690 * If we come here as part of a suspend/resume, don't touch cpusets because we
6691 * want to restore it back to its original state upon resume anyway.
6693 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6697 case CPU_ONLINE_FROZEN:
6698 case CPU_DOWN_FAILED_FROZEN:
6701 * num_cpus_frozen tracks how many CPUs are involved in suspend
6702 * resume sequence. As long as this is not the last online
6703 * operation in the resume sequence, just build a single sched
6704 * domain, ignoring cpusets.
6707 if (likely(num_cpus_frozen)) {
6708 partition_sched_domains(1, NULL, NULL);
6713 * This is the last CPU online operation. So fall through and
6714 * restore the original sched domains by considering the
6715 * cpuset configurations.
6719 case CPU_DOWN_FAILED:
6720 cpuset_update_active_cpus(true);
6728 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6732 case CPU_DOWN_PREPARE:
6733 cpuset_update_active_cpus(false);
6735 case CPU_DOWN_PREPARE_FROZEN:
6737 partition_sched_domains(1, NULL, NULL);
6745 void __init sched_init_smp(void)
6747 cpumask_var_t non_isolated_cpus;
6749 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6750 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6755 mutex_lock(&sched_domains_mutex);
6756 init_sched_domains(cpu_active_mask);
6757 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6758 if (cpumask_empty(non_isolated_cpus))
6759 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6760 mutex_unlock(&sched_domains_mutex);
6763 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6764 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6765 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6767 /* RT runtime code needs to handle some hotplug events */
6768 hotcpu_notifier(update_runtime, 0);
6772 /* Move init over to a non-isolated CPU */
6773 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6775 sched_init_granularity();
6776 free_cpumask_var(non_isolated_cpus);
6778 init_sched_rt_class();
6781 void __init sched_init_smp(void)
6783 sched_init_granularity();
6785 #endif /* CONFIG_SMP */
6787 const_debug unsigned int sysctl_timer_migration = 1;
6789 int in_sched_functions(unsigned long addr)
6791 return in_lock_functions(addr) ||
6792 (addr >= (unsigned long)__sched_text_start
6793 && addr < (unsigned long)__sched_text_end);
6796 #ifdef CONFIG_CGROUP_SCHED
6797 struct task_group root_task_group;
6798 LIST_HEAD(task_groups);
6801 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6803 void __init sched_init(void)
6806 unsigned long alloc_size = 0, ptr;
6808 #ifdef CONFIG_FAIR_GROUP_SCHED
6809 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6811 #ifdef CONFIG_RT_GROUP_SCHED
6812 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6814 #ifdef CONFIG_CPUMASK_OFFSTACK
6815 alloc_size += num_possible_cpus() * cpumask_size();
6818 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6820 #ifdef CONFIG_FAIR_GROUP_SCHED
6821 root_task_group.se = (struct sched_entity **)ptr;
6822 ptr += nr_cpu_ids * sizeof(void **);
6824 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6825 ptr += nr_cpu_ids * sizeof(void **);
6827 #endif /* CONFIG_FAIR_GROUP_SCHED */
6828 #ifdef CONFIG_RT_GROUP_SCHED
6829 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6830 ptr += nr_cpu_ids * sizeof(void **);
6832 root_task_group.rt_rq = (struct rt_rq **)ptr;
6833 ptr += nr_cpu_ids * sizeof(void **);
6835 #endif /* CONFIG_RT_GROUP_SCHED */
6836 #ifdef CONFIG_CPUMASK_OFFSTACK
6837 for_each_possible_cpu(i) {
6838 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6839 ptr += cpumask_size();
6841 #endif /* CONFIG_CPUMASK_OFFSTACK */
6845 init_defrootdomain();
6848 init_rt_bandwidth(&def_rt_bandwidth,
6849 global_rt_period(), global_rt_runtime());
6851 #ifdef CONFIG_RT_GROUP_SCHED
6852 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6853 global_rt_period(), global_rt_runtime());
6854 #endif /* CONFIG_RT_GROUP_SCHED */
6856 #ifdef CONFIG_CGROUP_SCHED
6857 list_add(&root_task_group.list, &task_groups);
6858 INIT_LIST_HEAD(&root_task_group.children);
6859 INIT_LIST_HEAD(&root_task_group.siblings);
6860 autogroup_init(&init_task);
6862 #endif /* CONFIG_CGROUP_SCHED */
6864 #ifdef CONFIG_CGROUP_CPUACCT
6865 root_cpuacct.cpustat = &kernel_cpustat;
6866 root_cpuacct.cpuusage = alloc_percpu(u64);
6867 /* Too early, not expected to fail */
6868 BUG_ON(!root_cpuacct.cpuusage);
6870 for_each_possible_cpu(i) {
6874 raw_spin_lock_init(&rq->lock);
6876 rq->calc_load_active = 0;
6877 rq->calc_load_update = jiffies + LOAD_FREQ;
6878 init_cfs_rq(&rq->cfs);
6879 init_rt_rq(&rq->rt, rq);
6880 #ifdef CONFIG_FAIR_GROUP_SCHED
6881 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6882 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6884 * How much cpu bandwidth does root_task_group get?
6886 * In case of task-groups formed thr' the cgroup filesystem, it
6887 * gets 100% of the cpu resources in the system. This overall
6888 * system cpu resource is divided among the tasks of
6889 * root_task_group and its child task-groups in a fair manner,
6890 * based on each entity's (task or task-group's) weight
6891 * (se->load.weight).
6893 * In other words, if root_task_group has 10 tasks of weight
6894 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6895 * then A0's share of the cpu resource is:
6897 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6899 * We achieve this by letting root_task_group's tasks sit
6900 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6902 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6903 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6904 #endif /* CONFIG_FAIR_GROUP_SCHED */
6906 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6907 #ifdef CONFIG_RT_GROUP_SCHED
6908 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6909 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6912 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6913 rq->cpu_load[j] = 0;
6915 rq->last_load_update_tick = jiffies;
6920 rq->cpu_power = SCHED_POWER_SCALE;
6921 rq->post_schedule = 0;
6922 rq->active_balance = 0;
6923 rq->next_balance = jiffies;
6928 rq->avg_idle = 2*sysctl_sched_migration_cost;
6930 INIT_LIST_HEAD(&rq->cfs_tasks);
6931 #ifdef CONFIG_SCHED_NUMA
6932 INIT_LIST_HEAD(&rq->offnode_tasks);
6933 rq->onnode_running = 0;
6934 rq->offnode_running = 0;
6935 rq->offnode_weight = 0;
6938 rq_attach_root(rq, &def_root_domain);
6944 atomic_set(&rq->nr_iowait, 0);
6947 set_load_weight(&init_task);
6949 #ifdef CONFIG_PREEMPT_NOTIFIERS
6950 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6953 #ifdef CONFIG_RT_MUTEXES
6954 plist_head_init(&init_task.pi_waiters);
6958 * The boot idle thread does lazy MMU switching as well:
6960 atomic_inc(&init_mm.mm_count);
6961 enter_lazy_tlb(&init_mm, current);
6964 * Make us the idle thread. Technically, schedule() should not be
6965 * called from this thread, however somewhere below it might be,
6966 * but because we are the idle thread, we just pick up running again
6967 * when this runqueue becomes "idle".
6969 init_idle(current, smp_processor_id());
6971 calc_load_update = jiffies + LOAD_FREQ;
6974 * During early bootup we pretend to be a normal task:
6976 current->sched_class = &fair_sched_class;
6979 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6980 /* May be allocated at isolcpus cmdline parse time */
6981 if (cpu_isolated_map == NULL)
6982 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6983 idle_thread_set_boot_cpu();
6985 init_sched_fair_class();
6987 scheduler_running = 1;
6990 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6991 static inline int preempt_count_equals(int preempt_offset)
6993 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6995 return (nested == preempt_offset);
6998 void __might_sleep(const char *file, int line, int preempt_offset)
7000 static unsigned long prev_jiffy; /* ratelimiting */
7002 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7003 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7004 system_state != SYSTEM_RUNNING || oops_in_progress)
7006 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7008 prev_jiffy = jiffies;
7011 "BUG: sleeping function called from invalid context at %s:%d\n",
7014 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7015 in_atomic(), irqs_disabled(),
7016 current->pid, current->comm);
7018 debug_show_held_locks(current);
7019 if (irqs_disabled())
7020 print_irqtrace_events(current);
7023 EXPORT_SYMBOL(__might_sleep);
7026 #ifdef CONFIG_MAGIC_SYSRQ
7027 static void normalize_task(struct rq *rq, struct task_struct *p)
7029 const struct sched_class *prev_class = p->sched_class;
7030 int old_prio = p->prio;
7035 dequeue_task(rq, p, 0);
7036 __setscheduler(rq, p, SCHED_NORMAL, 0);
7038 enqueue_task(rq, p, 0);
7039 resched_task(rq->curr);
7042 check_class_changed(rq, p, prev_class, old_prio);
7045 void normalize_rt_tasks(void)
7047 struct task_struct *g, *p;
7048 unsigned long flags;
7051 read_lock_irqsave(&tasklist_lock, flags);
7052 do_each_thread(g, p) {
7054 * Only normalize user tasks:
7059 p->se.exec_start = 0;
7060 #ifdef CONFIG_SCHEDSTATS
7061 p->se.statistics.wait_start = 0;
7062 p->se.statistics.sleep_start = 0;
7063 p->se.statistics.block_start = 0;
7068 * Renice negative nice level userspace
7071 if (TASK_NICE(p) < 0 && p->mm)
7072 set_user_nice(p, 0);
7076 raw_spin_lock(&p->pi_lock);
7077 rq = __task_rq_lock(p);
7079 normalize_task(rq, p);
7081 __task_rq_unlock(rq);
7082 raw_spin_unlock(&p->pi_lock);
7083 } while_each_thread(g, p);
7085 read_unlock_irqrestore(&tasklist_lock, flags);
7088 #endif /* CONFIG_MAGIC_SYSRQ */
7090 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7092 * These functions are only useful for the IA64 MCA handling, or kdb.
7094 * They can only be called when the whole system has been
7095 * stopped - every CPU needs to be quiescent, and no scheduling
7096 * activity can take place. Using them for anything else would
7097 * be a serious bug, and as a result, they aren't even visible
7098 * under any other configuration.
7102 * curr_task - return the current task for a given cpu.
7103 * @cpu: the processor in question.
7105 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7107 struct task_struct *curr_task(int cpu)
7109 return cpu_curr(cpu);
7112 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7116 * set_curr_task - set the current task for a given cpu.
7117 * @cpu: the processor in question.
7118 * @p: the task pointer to set.
7120 * Description: This function must only be used when non-maskable interrupts
7121 * are serviced on a separate stack. It allows the architecture to switch the
7122 * notion of the current task on a cpu in a non-blocking manner. This function
7123 * must be called with all CPU's synchronized, and interrupts disabled, the
7124 * and caller must save the original value of the current task (see
7125 * curr_task() above) and restore that value before reenabling interrupts and
7126 * re-starting the system.
7128 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7130 void set_curr_task(int cpu, struct task_struct *p)
7137 #ifdef CONFIG_CGROUP_SCHED
7138 /* task_group_lock serializes the addition/removal of task groups */
7139 static DEFINE_SPINLOCK(task_group_lock);
7141 static void free_sched_group(struct task_group *tg)
7143 free_fair_sched_group(tg);
7144 free_rt_sched_group(tg);
7149 /* allocate runqueue etc for a new task group */
7150 struct task_group *sched_create_group(struct task_group *parent)
7152 struct task_group *tg;
7153 unsigned long flags;
7155 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7157 return ERR_PTR(-ENOMEM);
7159 if (!alloc_fair_sched_group(tg, parent))
7162 if (!alloc_rt_sched_group(tg, parent))
7165 spin_lock_irqsave(&task_group_lock, flags);
7166 list_add_rcu(&tg->list, &task_groups);
7168 WARN_ON(!parent); /* root should already exist */
7170 tg->parent = parent;
7171 INIT_LIST_HEAD(&tg->children);
7172 list_add_rcu(&tg->siblings, &parent->children);
7173 spin_unlock_irqrestore(&task_group_lock, flags);
7178 free_sched_group(tg);
7179 return ERR_PTR(-ENOMEM);
7182 /* rcu callback to free various structures associated with a task group */
7183 static void free_sched_group_rcu(struct rcu_head *rhp)
7185 /* now it should be safe to free those cfs_rqs */
7186 free_sched_group(container_of(rhp, struct task_group, rcu));
7189 /* Destroy runqueue etc associated with a task group */
7190 void sched_destroy_group(struct task_group *tg)
7192 unsigned long flags;
7195 /* end participation in shares distribution */
7196 for_each_possible_cpu(i)
7197 unregister_fair_sched_group(tg, i);
7199 spin_lock_irqsave(&task_group_lock, flags);
7200 list_del_rcu(&tg->list);
7201 list_del_rcu(&tg->siblings);
7202 spin_unlock_irqrestore(&task_group_lock, flags);
7204 /* wait for possible concurrent references to cfs_rqs complete */
7205 call_rcu(&tg->rcu, free_sched_group_rcu);
7208 /* change task's runqueue when it moves between groups.
7209 * The caller of this function should have put the task in its new group
7210 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7211 * reflect its new group.
7213 void sched_move_task(struct task_struct *tsk)
7215 struct task_group *tg;
7217 unsigned long flags;
7220 rq = task_rq_lock(tsk, &flags);
7222 running = task_current(rq, tsk);
7226 dequeue_task(rq, tsk, 0);
7227 if (unlikely(running))
7228 tsk->sched_class->put_prev_task(rq, tsk);
7230 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7231 lockdep_is_held(&tsk->sighand->siglock)),
7232 struct task_group, css);
7233 tg = autogroup_task_group(tsk, tg);
7234 tsk->sched_task_group = tg;
7236 #ifdef CONFIG_FAIR_GROUP_SCHED
7237 if (tsk->sched_class->task_move_group)
7238 tsk->sched_class->task_move_group(tsk, on_rq);
7241 set_task_rq(tsk, task_cpu(tsk));
7243 if (unlikely(running))
7244 tsk->sched_class->set_curr_task(rq);
7246 enqueue_task(rq, tsk, 0);
7248 task_rq_unlock(rq, tsk, &flags);
7250 #endif /* CONFIG_CGROUP_SCHED */
7252 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7253 static unsigned long to_ratio(u64 period, u64 runtime)
7255 if (runtime == RUNTIME_INF)
7258 return div64_u64(runtime << 20, period);
7262 #ifdef CONFIG_RT_GROUP_SCHED
7264 * Ensure that the real time constraints are schedulable.
7266 static DEFINE_MUTEX(rt_constraints_mutex);
7268 /* Must be called with tasklist_lock held */
7269 static inline int tg_has_rt_tasks(struct task_group *tg)
7271 struct task_struct *g, *p;
7273 do_each_thread(g, p) {
7274 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7276 } while_each_thread(g, p);
7281 struct rt_schedulable_data {
7282 struct task_group *tg;
7287 static int tg_rt_schedulable(struct task_group *tg, void *data)
7289 struct rt_schedulable_data *d = data;
7290 struct task_group *child;
7291 unsigned long total, sum = 0;
7292 u64 period, runtime;
7294 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7295 runtime = tg->rt_bandwidth.rt_runtime;
7298 period = d->rt_period;
7299 runtime = d->rt_runtime;
7303 * Cannot have more runtime than the period.
7305 if (runtime > period && runtime != RUNTIME_INF)
7309 * Ensure we don't starve existing RT tasks.
7311 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7314 total = to_ratio(period, runtime);
7317 * Nobody can have more than the global setting allows.
7319 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7323 * The sum of our children's runtime should not exceed our own.
7325 list_for_each_entry_rcu(child, &tg->children, siblings) {
7326 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7327 runtime = child->rt_bandwidth.rt_runtime;
7329 if (child == d->tg) {
7330 period = d->rt_period;
7331 runtime = d->rt_runtime;
7334 sum += to_ratio(period, runtime);
7343 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7347 struct rt_schedulable_data data = {
7349 .rt_period = period,
7350 .rt_runtime = runtime,
7354 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7360 static int tg_set_rt_bandwidth(struct task_group *tg,
7361 u64 rt_period, u64 rt_runtime)
7365 mutex_lock(&rt_constraints_mutex);
7366 read_lock(&tasklist_lock);
7367 err = __rt_schedulable(tg, rt_period, rt_runtime);
7371 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7372 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7373 tg->rt_bandwidth.rt_runtime = rt_runtime;
7375 for_each_possible_cpu(i) {
7376 struct rt_rq *rt_rq = tg->rt_rq[i];
7378 raw_spin_lock(&rt_rq->rt_runtime_lock);
7379 rt_rq->rt_runtime = rt_runtime;
7380 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7382 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7384 read_unlock(&tasklist_lock);
7385 mutex_unlock(&rt_constraints_mutex);
7390 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7392 u64 rt_runtime, rt_period;
7394 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7395 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7396 if (rt_runtime_us < 0)
7397 rt_runtime = RUNTIME_INF;
7399 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7402 long sched_group_rt_runtime(struct task_group *tg)
7406 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7409 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7410 do_div(rt_runtime_us, NSEC_PER_USEC);
7411 return rt_runtime_us;
7414 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7416 u64 rt_runtime, rt_period;
7418 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7419 rt_runtime = tg->rt_bandwidth.rt_runtime;
7424 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7427 long sched_group_rt_period(struct task_group *tg)
7431 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7432 do_div(rt_period_us, NSEC_PER_USEC);
7433 return rt_period_us;
7436 static int sched_rt_global_constraints(void)
7438 u64 runtime, period;
7441 if (sysctl_sched_rt_period <= 0)
7444 runtime = global_rt_runtime();
7445 period = global_rt_period();
7448 * Sanity check on the sysctl variables.
7450 if (runtime > period && runtime != RUNTIME_INF)
7453 mutex_lock(&rt_constraints_mutex);
7454 read_lock(&tasklist_lock);
7455 ret = __rt_schedulable(NULL, 0, 0);
7456 read_unlock(&tasklist_lock);
7457 mutex_unlock(&rt_constraints_mutex);
7462 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7464 /* Don't accept realtime tasks when there is no way for them to run */
7465 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7471 #else /* !CONFIG_RT_GROUP_SCHED */
7472 static int sched_rt_global_constraints(void)
7474 unsigned long flags;
7477 if (sysctl_sched_rt_period <= 0)
7481 * There's always some RT tasks in the root group
7482 * -- migration, kstopmachine etc..
7484 if (sysctl_sched_rt_runtime == 0)
7487 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7488 for_each_possible_cpu(i) {
7489 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7491 raw_spin_lock(&rt_rq->rt_runtime_lock);
7492 rt_rq->rt_runtime = global_rt_runtime();
7493 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7495 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7499 #endif /* CONFIG_RT_GROUP_SCHED */
7501 int sched_rt_handler(struct ctl_table *table, int write,
7502 void __user *buffer, size_t *lenp,
7506 int old_period, old_runtime;
7507 static DEFINE_MUTEX(mutex);
7510 old_period = sysctl_sched_rt_period;
7511 old_runtime = sysctl_sched_rt_runtime;
7513 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7515 if (!ret && write) {
7516 ret = sched_rt_global_constraints();
7518 sysctl_sched_rt_period = old_period;
7519 sysctl_sched_rt_runtime = old_runtime;
7521 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7522 def_rt_bandwidth.rt_period =
7523 ns_to_ktime(global_rt_period());
7526 mutex_unlock(&mutex);
7531 #ifdef CONFIG_CGROUP_SCHED
7533 /* return corresponding task_group object of a cgroup */
7534 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7536 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7537 struct task_group, css);
7540 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7542 struct task_group *tg, *parent;
7544 if (!cgrp->parent) {
7545 /* This is early initialization for the top cgroup */
7546 return &root_task_group.css;
7549 parent = cgroup_tg(cgrp->parent);
7550 tg = sched_create_group(parent);
7552 return ERR_PTR(-ENOMEM);
7557 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7559 struct task_group *tg = cgroup_tg(cgrp);
7561 sched_destroy_group(tg);
7564 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7565 struct cgroup_taskset *tset)
7567 struct task_struct *task;
7569 cgroup_taskset_for_each(task, cgrp, tset) {
7570 #ifdef CONFIG_RT_GROUP_SCHED
7571 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7574 /* We don't support RT-tasks being in separate groups */
7575 if (task->sched_class != &fair_sched_class)
7582 static void cpu_cgroup_attach(struct cgroup *cgrp,
7583 struct cgroup_taskset *tset)
7585 struct task_struct *task;
7587 cgroup_taskset_for_each(task, cgrp, tset)
7588 sched_move_task(task);
7592 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7593 struct task_struct *task)
7596 * cgroup_exit() is called in the copy_process() failure path.
7597 * Ignore this case since the task hasn't ran yet, this avoids
7598 * trying to poke a half freed task state from generic code.
7600 if (!(task->flags & PF_EXITING))
7603 sched_move_task(task);
7606 #ifdef CONFIG_FAIR_GROUP_SCHED
7607 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7610 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7613 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7615 struct task_group *tg = cgroup_tg(cgrp);
7617 return (u64) scale_load_down(tg->shares);
7620 #ifdef CONFIG_CFS_BANDWIDTH
7621 static DEFINE_MUTEX(cfs_constraints_mutex);
7623 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7624 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7626 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7628 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7630 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7631 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7633 if (tg == &root_task_group)
7637 * Ensure we have at some amount of bandwidth every period. This is
7638 * to prevent reaching a state of large arrears when throttled via
7639 * entity_tick() resulting in prolonged exit starvation.
7641 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7645 * Likewise, bound things on the otherside by preventing insane quota
7646 * periods. This also allows us to normalize in computing quota
7649 if (period > max_cfs_quota_period)
7652 mutex_lock(&cfs_constraints_mutex);
7653 ret = __cfs_schedulable(tg, period, quota);
7657 runtime_enabled = quota != RUNTIME_INF;
7658 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7659 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7660 raw_spin_lock_irq(&cfs_b->lock);
7661 cfs_b->period = ns_to_ktime(period);
7662 cfs_b->quota = quota;
7664 __refill_cfs_bandwidth_runtime(cfs_b);
7665 /* restart the period timer (if active) to handle new period expiry */
7666 if (runtime_enabled && cfs_b->timer_active) {
7667 /* force a reprogram */
7668 cfs_b->timer_active = 0;
7669 __start_cfs_bandwidth(cfs_b);
7671 raw_spin_unlock_irq(&cfs_b->lock);
7673 for_each_possible_cpu(i) {
7674 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7675 struct rq *rq = cfs_rq->rq;
7677 raw_spin_lock_irq(&rq->lock);
7678 cfs_rq->runtime_enabled = runtime_enabled;
7679 cfs_rq->runtime_remaining = 0;
7681 if (cfs_rq->throttled)
7682 unthrottle_cfs_rq(cfs_rq);
7683 raw_spin_unlock_irq(&rq->lock);
7686 mutex_unlock(&cfs_constraints_mutex);
7691 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7695 period = ktime_to_ns(tg->cfs_bandwidth.period);
7696 if (cfs_quota_us < 0)
7697 quota = RUNTIME_INF;
7699 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7701 return tg_set_cfs_bandwidth(tg, period, quota);
7704 long tg_get_cfs_quota(struct task_group *tg)
7708 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7711 quota_us = tg->cfs_bandwidth.quota;
7712 do_div(quota_us, NSEC_PER_USEC);
7717 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7721 period = (u64)cfs_period_us * NSEC_PER_USEC;
7722 quota = tg->cfs_bandwidth.quota;
7724 return tg_set_cfs_bandwidth(tg, period, quota);
7727 long tg_get_cfs_period(struct task_group *tg)
7731 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7732 do_div(cfs_period_us, NSEC_PER_USEC);
7734 return cfs_period_us;
7737 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7739 return tg_get_cfs_quota(cgroup_tg(cgrp));
7742 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7745 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7748 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7750 return tg_get_cfs_period(cgroup_tg(cgrp));
7753 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7756 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7759 struct cfs_schedulable_data {
7760 struct task_group *tg;
7765 * normalize group quota/period to be quota/max_period
7766 * note: units are usecs
7768 static u64 normalize_cfs_quota(struct task_group *tg,
7769 struct cfs_schedulable_data *d)
7777 period = tg_get_cfs_period(tg);
7778 quota = tg_get_cfs_quota(tg);
7781 /* note: these should typically be equivalent */
7782 if (quota == RUNTIME_INF || quota == -1)
7785 return to_ratio(period, quota);
7788 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7790 struct cfs_schedulable_data *d = data;
7791 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7792 s64 quota = 0, parent_quota = -1;
7795 quota = RUNTIME_INF;
7797 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7799 quota = normalize_cfs_quota(tg, d);
7800 parent_quota = parent_b->hierarchal_quota;
7803 * ensure max(child_quota) <= parent_quota, inherit when no
7806 if (quota == RUNTIME_INF)
7807 quota = parent_quota;
7808 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7811 cfs_b->hierarchal_quota = quota;
7816 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7819 struct cfs_schedulable_data data = {
7825 if (quota != RUNTIME_INF) {
7826 do_div(data.period, NSEC_PER_USEC);
7827 do_div(data.quota, NSEC_PER_USEC);
7831 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7837 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7838 struct cgroup_map_cb *cb)
7840 struct task_group *tg = cgroup_tg(cgrp);
7841 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7843 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7844 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7845 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7849 #endif /* CONFIG_CFS_BANDWIDTH */
7850 #endif /* CONFIG_FAIR_GROUP_SCHED */
7852 #ifdef CONFIG_RT_GROUP_SCHED
7853 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7856 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7859 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7861 return sched_group_rt_runtime(cgroup_tg(cgrp));
7864 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7867 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7870 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7872 return sched_group_rt_period(cgroup_tg(cgrp));
7874 #endif /* CONFIG_RT_GROUP_SCHED */
7876 static struct cftype cpu_files[] = {
7877 #ifdef CONFIG_FAIR_GROUP_SCHED
7880 .read_u64 = cpu_shares_read_u64,
7881 .write_u64 = cpu_shares_write_u64,
7884 #ifdef CONFIG_CFS_BANDWIDTH
7886 .name = "cfs_quota_us",
7887 .read_s64 = cpu_cfs_quota_read_s64,
7888 .write_s64 = cpu_cfs_quota_write_s64,
7891 .name = "cfs_period_us",
7892 .read_u64 = cpu_cfs_period_read_u64,
7893 .write_u64 = cpu_cfs_period_write_u64,
7897 .read_map = cpu_stats_show,
7900 #ifdef CONFIG_RT_GROUP_SCHED
7902 .name = "rt_runtime_us",
7903 .read_s64 = cpu_rt_runtime_read,
7904 .write_s64 = cpu_rt_runtime_write,
7907 .name = "rt_period_us",
7908 .read_u64 = cpu_rt_period_read_uint,
7909 .write_u64 = cpu_rt_period_write_uint,
7915 struct cgroup_subsys cpu_cgroup_subsys = {
7917 .create = cpu_cgroup_create,
7918 .destroy = cpu_cgroup_destroy,
7919 .can_attach = cpu_cgroup_can_attach,
7920 .attach = cpu_cgroup_attach,
7921 .exit = cpu_cgroup_exit,
7922 .subsys_id = cpu_cgroup_subsys_id,
7923 .base_cftypes = cpu_files,
7927 #endif /* CONFIG_CGROUP_SCHED */
7929 #ifdef CONFIG_CGROUP_CPUACCT
7932 * CPU accounting code for task groups.
7934 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7935 * (balbir@in.ibm.com).
7938 struct cpuacct root_cpuacct;
7940 /* create a new cpu accounting group */
7941 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
7946 return &root_cpuacct.css;
7948 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7952 ca->cpuusage = alloc_percpu(u64);
7956 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7958 goto out_free_cpuusage;
7963 free_percpu(ca->cpuusage);
7967 return ERR_PTR(-ENOMEM);
7970 /* destroy an existing cpu accounting group */
7971 static void cpuacct_destroy(struct cgroup *cgrp)
7973 struct cpuacct *ca = cgroup_ca(cgrp);
7975 free_percpu(ca->cpustat);
7976 free_percpu(ca->cpuusage);
7980 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7982 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7985 #ifndef CONFIG_64BIT
7987 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7989 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7991 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7999 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8001 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8003 #ifndef CONFIG_64BIT
8005 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8007 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8009 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8015 /* return total cpu usage (in nanoseconds) of a group */
8016 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8018 struct cpuacct *ca = cgroup_ca(cgrp);
8019 u64 totalcpuusage = 0;
8022 for_each_present_cpu(i)
8023 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8025 return totalcpuusage;
8028 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8031 struct cpuacct *ca = cgroup_ca(cgrp);
8040 for_each_present_cpu(i)
8041 cpuacct_cpuusage_write(ca, i, 0);
8047 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8050 struct cpuacct *ca = cgroup_ca(cgroup);
8054 for_each_present_cpu(i) {
8055 percpu = cpuacct_cpuusage_read(ca, i);
8056 seq_printf(m, "%llu ", (unsigned long long) percpu);
8058 seq_printf(m, "\n");
8062 static const char *cpuacct_stat_desc[] = {
8063 [CPUACCT_STAT_USER] = "user",
8064 [CPUACCT_STAT_SYSTEM] = "system",
8067 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8068 struct cgroup_map_cb *cb)
8070 struct cpuacct *ca = cgroup_ca(cgrp);
8074 for_each_online_cpu(cpu) {
8075 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8076 val += kcpustat->cpustat[CPUTIME_USER];
8077 val += kcpustat->cpustat[CPUTIME_NICE];
8079 val = cputime64_to_clock_t(val);
8080 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8083 for_each_online_cpu(cpu) {
8084 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8085 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8086 val += kcpustat->cpustat[CPUTIME_IRQ];
8087 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8090 val = cputime64_to_clock_t(val);
8091 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8096 static struct cftype files[] = {
8099 .read_u64 = cpuusage_read,
8100 .write_u64 = cpuusage_write,
8103 .name = "usage_percpu",
8104 .read_seq_string = cpuacct_percpu_seq_read,
8108 .read_map = cpuacct_stats_show,
8114 * charge this task's execution time to its accounting group.
8116 * called with rq->lock held.
8118 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8123 if (unlikely(!cpuacct_subsys.active))
8126 cpu = task_cpu(tsk);
8132 for (; ca; ca = parent_ca(ca)) {
8133 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8134 *cpuusage += cputime;
8140 struct cgroup_subsys cpuacct_subsys = {
8142 .create = cpuacct_create,
8143 .destroy = cpuacct_destroy,
8144 .subsys_id = cpuacct_subsys_id,
8145 .base_cftypes = files,
8147 #endif /* CONFIG_CGROUP_CPUACCT */