4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
201 if (strncmp(cmp, "NO_", 3) == 0) {
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
233 if (copy_from_user(&buf, ubuf, cnt))
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
248 static int sched_feat_open(struct inode *inode, struct file *filp)
250 return single_open(filp, sched_feat_show, NULL);
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
258 .release = single_release,
261 static __init int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
278 * period over which we average the RT time consumption, measured
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime = 950000;
300 * __task_rq_lock - lock the rq @p resides on.
302 static inline struct rq *__task_rq_lock(struct task_struct *p)
307 lockdep_assert_held(&p->pi_lock);
311 raw_spin_lock(&rq->lock);
312 if (likely(rq == task_rq(p)))
314 raw_spin_unlock(&rq->lock);
319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
321 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
322 __acquires(p->pi_lock)
328 raw_spin_lock_irqsave(&p->pi_lock, *flags);
330 raw_spin_lock(&rq->lock);
331 if (likely(rq == task_rq(p)))
333 raw_spin_unlock(&rq->lock);
334 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
338 static void __task_rq_unlock(struct rq *rq)
341 raw_spin_unlock(&rq->lock);
345 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
347 __releases(p->pi_lock)
349 raw_spin_unlock(&rq->lock);
350 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
354 * this_rq_lock - lock this runqueue and disable interrupts.
356 static struct rq *this_rq_lock(void)
363 raw_spin_lock(&rq->lock);
368 #ifdef CONFIG_SCHED_HRTICK
370 * Use HR-timers to deliver accurate preemption points.
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 static int __hrtick_restart(struct rq *rq)
401 struct hrtimer *timer = &rq->hrtick_timer;
402 ktime_t time = hrtimer_get_softexpires(timer);
404 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg)
414 raw_spin_lock(&rq->lock);
415 __hrtick_restart(rq);
416 rq->hrtick_csd_pending = 0;
417 raw_spin_unlock(&rq->lock);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq *rq, u64 delay)
427 struct hrtimer *timer = &rq->hrtick_timer;
428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 hrtimer_set_expires(timer, time);
432 if (rq == this_rq()) {
433 __hrtick_restart(rq);
434 } else if (!rq->hrtick_csd_pending) {
435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436 rq->hrtick_csd_pending = 1;
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 int cpu = (int)(long)hcpu;
446 case CPU_UP_CANCELED:
447 case CPU_UP_CANCELED_FROZEN:
448 case CPU_DOWN_PREPARE:
449 case CPU_DOWN_PREPARE_FROZEN:
451 case CPU_DEAD_FROZEN:
452 hrtick_clear(cpu_rq(cpu));
459 static __init void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq *rq, u64 delay)
471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472 HRTIMER_MODE_REL_PINNED, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq *rq)
483 rq->hrtick_csd_pending = 0;
485 rq->hrtick_csd.flags = 0;
486 rq->hrtick_csd.func = __hrtick_start;
487 rq->hrtick_csd.info = rq;
490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491 rq->hrtick_timer.function = hrtick;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
498 static inline void init_rq_hrtick(struct rq *rq)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
514 void resched_task(struct task_struct *p)
518 lockdep_assert_held(&task_rq(p)->lock);
520 if (test_tsk_need_resched(p))
523 set_tsk_need_resched(p);
526 if (cpu == smp_processor_id()) {
527 set_preempt_need_resched();
531 /* NEED_RESCHED must be visible before we test polling */
533 if (!tsk_is_polling(p))
534 smp_send_reschedule(cpu);
537 void resched_cpu(int cpu)
539 struct rq *rq = cpu_rq(cpu);
542 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
544 resched_task(cpu_curr(cpu));
545 raw_spin_unlock_irqrestore(&rq->lock, flags);
549 #ifdef CONFIG_NO_HZ_COMMON
551 * In the semi idle case, use the nearest busy cpu for migrating timers
552 * from an idle cpu. This is good for power-savings.
554 * We don't do similar optimization for completely idle system, as
555 * selecting an idle cpu will add more delays to the timers than intended
556 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558 int get_nohz_timer_target(void)
560 int cpu = smp_processor_id();
562 struct sched_domain *sd;
565 for_each_domain(cpu, sd) {
566 for_each_cpu(i, sched_domain_span(sd)) {
578 * When add_timer_on() enqueues a timer into the timer wheel of an
579 * idle CPU then this timer might expire before the next timer event
580 * which is scheduled to wake up that CPU. In case of a completely
581 * idle system the next event might even be infinite time into the
582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
583 * leaves the inner idle loop so the newly added timer is taken into
584 * account when the CPU goes back to idle and evaluates the timer
585 * wheel for the next timer event.
587 static void wake_up_idle_cpu(int cpu)
589 struct rq *rq = cpu_rq(cpu);
591 if (cpu == smp_processor_id())
595 * This is safe, as this function is called with the timer
596 * wheel base lock of (cpu) held. When the CPU is on the way
597 * to idle and has not yet set rq->curr to idle then it will
598 * be serialized on the timer wheel base lock and take the new
599 * timer into account automatically.
601 if (rq->curr != rq->idle)
605 * We can set TIF_RESCHED on the idle task of the other CPU
606 * lockless. The worst case is that the other CPU runs the
607 * idle task through an additional NOOP schedule()
609 set_tsk_need_resched(rq->idle);
611 /* NEED_RESCHED must be visible before we test polling */
613 if (!tsk_is_polling(rq->idle))
614 smp_send_reschedule(cpu);
617 static bool wake_up_full_nohz_cpu(int cpu)
619 if (tick_nohz_full_cpu(cpu)) {
620 if (cpu != smp_processor_id() ||
621 tick_nohz_tick_stopped())
622 smp_send_reschedule(cpu);
629 void wake_up_nohz_cpu(int cpu)
631 if (!wake_up_full_nohz_cpu(cpu))
632 wake_up_idle_cpu(cpu);
635 static inline bool got_nohz_idle_kick(void)
637 int cpu = smp_processor_id();
639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
642 if (idle_cpu(cpu) && !need_resched())
646 * We can't run Idle Load Balance on this CPU for this time so we
647 * cancel it and clear NOHZ_BALANCE_KICK
649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
653 #else /* CONFIG_NO_HZ_COMMON */
655 static inline bool got_nohz_idle_kick(void)
660 #endif /* CONFIG_NO_HZ_COMMON */
662 #ifdef CONFIG_NO_HZ_FULL
663 bool sched_can_stop_tick(void)
669 /* Make sure rq->nr_running update is visible after the IPI */
672 /* More than one running task need preemption */
673 if (rq->nr_running > 1)
678 #endif /* CONFIG_NO_HZ_FULL */
680 void sched_avg_update(struct rq *rq)
682 s64 period = sched_avg_period();
684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
686 * Inline assembly required to prevent the compiler
687 * optimising this loop into a divmod call.
688 * See __iter_div_u64_rem() for another example of this.
690 asm("" : "+rm" (rq->age_stamp));
691 rq->age_stamp += period;
696 #endif /* CONFIG_SMP */
698 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
699 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
701 * Iterate task_group tree rooted at *from, calling @down when first entering a
702 * node and @up when leaving it for the final time.
704 * Caller must hold rcu_lock or sufficient equivalent.
706 int walk_tg_tree_from(struct task_group *from,
707 tg_visitor down, tg_visitor up, void *data)
709 struct task_group *parent, *child;
715 ret = (*down)(parent, data);
718 list_for_each_entry_rcu(child, &parent->children, siblings) {
725 ret = (*up)(parent, data);
726 if (ret || parent == from)
730 parent = parent->parent;
737 int tg_nop(struct task_group *tg, void *data)
743 static void set_load_weight(struct task_struct *p)
745 int prio = p->static_prio - MAX_RT_PRIO;
746 struct load_weight *load = &p->se.load;
749 * SCHED_IDLE tasks get minimal weight:
751 if (p->policy == SCHED_IDLE) {
752 load->weight = scale_load(WEIGHT_IDLEPRIO);
753 load->inv_weight = WMULT_IDLEPRIO;
757 load->weight = scale_load(prio_to_weight[prio]);
758 load->inv_weight = prio_to_wmult[prio];
761 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
764 sched_info_queued(rq, p);
765 p->sched_class->enqueue_task(rq, p, flags);
768 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
771 sched_info_dequeued(rq, p);
772 p->sched_class->dequeue_task(rq, p, flags);
775 void activate_task(struct rq *rq, struct task_struct *p, int flags)
777 if (task_contributes_to_load(p))
778 rq->nr_uninterruptible--;
780 enqueue_task(rq, p, flags);
783 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
785 if (task_contributes_to_load(p))
786 rq->nr_uninterruptible++;
788 dequeue_task(rq, p, flags);
791 static void update_rq_clock_task(struct rq *rq, s64 delta)
794 * In theory, the compile should just see 0 here, and optimize out the call
795 * to sched_rt_avg_update. But I don't trust it...
797 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
798 s64 steal = 0, irq_delta = 0;
800 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
801 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
804 * Since irq_time is only updated on {soft,}irq_exit, we might run into
805 * this case when a previous update_rq_clock() happened inside a
808 * When this happens, we stop ->clock_task and only update the
809 * prev_irq_time stamp to account for the part that fit, so that a next
810 * update will consume the rest. This ensures ->clock_task is
813 * It does however cause some slight miss-attribution of {soft,}irq
814 * time, a more accurate solution would be to update the irq_time using
815 * the current rq->clock timestamp, except that would require using
818 if (irq_delta > delta)
821 rq->prev_irq_time += irq_delta;
824 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
825 if (static_key_false((¶virt_steal_rq_enabled))) {
826 steal = paravirt_steal_clock(cpu_of(rq));
827 steal -= rq->prev_steal_time_rq;
829 if (unlikely(steal > delta))
832 rq->prev_steal_time_rq += steal;
837 rq->clock_task += delta;
839 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
840 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
841 sched_rt_avg_update(rq, irq_delta + steal);
845 void sched_set_stop_task(int cpu, struct task_struct *stop)
847 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
848 struct task_struct *old_stop = cpu_rq(cpu)->stop;
852 * Make it appear like a SCHED_FIFO task, its something
853 * userspace knows about and won't get confused about.
855 * Also, it will make PI more or less work without too
856 * much confusion -- but then, stop work should not
857 * rely on PI working anyway.
859 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
861 stop->sched_class = &stop_sched_class;
864 cpu_rq(cpu)->stop = stop;
868 * Reset it back to a normal scheduling class so that
869 * it can die in pieces.
871 old_stop->sched_class = &rt_sched_class;
876 * __normal_prio - return the priority that is based on the static prio
878 static inline int __normal_prio(struct task_struct *p)
880 return p->static_prio;
884 * Calculate the expected normal priority: i.e. priority
885 * without taking RT-inheritance into account. Might be
886 * boosted by interactivity modifiers. Changes upon fork,
887 * setprio syscalls, and whenever the interactivity
888 * estimator recalculates.
890 static inline int normal_prio(struct task_struct *p)
894 if (task_has_dl_policy(p))
895 prio = MAX_DL_PRIO-1;
896 else if (task_has_rt_policy(p))
897 prio = MAX_RT_PRIO-1 - p->rt_priority;
899 prio = __normal_prio(p);
904 * Calculate the current priority, i.e. the priority
905 * taken into account by the scheduler. This value might
906 * be boosted by RT tasks, or might be boosted by
907 * interactivity modifiers. Will be RT if the task got
908 * RT-boosted. If not then it returns p->normal_prio.
910 static int effective_prio(struct task_struct *p)
912 p->normal_prio = normal_prio(p);
914 * If we are RT tasks or we were boosted to RT priority,
915 * keep the priority unchanged. Otherwise, update priority
916 * to the normal priority:
918 if (!rt_prio(p->prio))
919 return p->normal_prio;
924 * task_curr - is this task currently executing on a CPU?
925 * @p: the task in question.
927 * Return: 1 if the task is currently executing. 0 otherwise.
929 inline int task_curr(const struct task_struct *p)
931 return cpu_curr(task_cpu(p)) == p;
934 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
935 const struct sched_class *prev_class,
938 if (prev_class != p->sched_class) {
939 if (prev_class->switched_from)
940 prev_class->switched_from(rq, p);
941 p->sched_class->switched_to(rq, p);
942 } else if (oldprio != p->prio || dl_task(p))
943 p->sched_class->prio_changed(rq, p, oldprio);
946 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
948 const struct sched_class *class;
950 if (p->sched_class == rq->curr->sched_class) {
951 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
953 for_each_class(class) {
954 if (class == rq->curr->sched_class)
956 if (class == p->sched_class) {
957 resched_task(rq->curr);
964 * A queue event has occurred, and we're going to schedule. In
965 * this case, we can save a useless back to back clock update.
967 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
968 rq->skip_clock_update = 1;
972 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
974 #ifdef CONFIG_SCHED_DEBUG
976 * We should never call set_task_cpu() on a blocked task,
977 * ttwu() will sort out the placement.
979 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
980 !(task_preempt_count(p) & PREEMPT_ACTIVE));
982 #ifdef CONFIG_LOCKDEP
984 * The caller should hold either p->pi_lock or rq->lock, when changing
985 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
987 * sched_move_task() holds both and thus holding either pins the cgroup,
990 * Furthermore, all task_rq users should acquire both locks, see
993 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
994 lockdep_is_held(&task_rq(p)->lock)));
998 trace_sched_migrate_task(p, new_cpu);
1000 if (task_cpu(p) != new_cpu) {
1001 if (p->sched_class->migrate_task_rq)
1002 p->sched_class->migrate_task_rq(p, new_cpu);
1003 p->se.nr_migrations++;
1004 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1007 __set_task_cpu(p, new_cpu);
1010 static void __migrate_swap_task(struct task_struct *p, int cpu)
1013 struct rq *src_rq, *dst_rq;
1015 src_rq = task_rq(p);
1016 dst_rq = cpu_rq(cpu);
1018 deactivate_task(src_rq, p, 0);
1019 set_task_cpu(p, cpu);
1020 activate_task(dst_rq, p, 0);
1021 check_preempt_curr(dst_rq, p, 0);
1024 * Task isn't running anymore; make it appear like we migrated
1025 * it before it went to sleep. This means on wakeup we make the
1026 * previous cpu our targer instead of where it really is.
1032 struct migration_swap_arg {
1033 struct task_struct *src_task, *dst_task;
1034 int src_cpu, dst_cpu;
1037 static int migrate_swap_stop(void *data)
1039 struct migration_swap_arg *arg = data;
1040 struct rq *src_rq, *dst_rq;
1043 src_rq = cpu_rq(arg->src_cpu);
1044 dst_rq = cpu_rq(arg->dst_cpu);
1046 double_raw_lock(&arg->src_task->pi_lock,
1047 &arg->dst_task->pi_lock);
1048 double_rq_lock(src_rq, dst_rq);
1049 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1052 if (task_cpu(arg->src_task) != arg->src_cpu)
1055 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1058 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1061 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1062 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1067 double_rq_unlock(src_rq, dst_rq);
1068 raw_spin_unlock(&arg->dst_task->pi_lock);
1069 raw_spin_unlock(&arg->src_task->pi_lock);
1075 * Cross migrate two tasks
1077 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1079 struct migration_swap_arg arg;
1082 arg = (struct migration_swap_arg){
1084 .src_cpu = task_cpu(cur),
1086 .dst_cpu = task_cpu(p),
1089 if (arg.src_cpu == arg.dst_cpu)
1093 * These three tests are all lockless; this is OK since all of them
1094 * will be re-checked with proper locks held further down the line.
1096 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1099 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1102 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1105 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1106 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1112 struct migration_arg {
1113 struct task_struct *task;
1117 static int migration_cpu_stop(void *data);
1120 * wait_task_inactive - wait for a thread to unschedule.
1122 * If @match_state is nonzero, it's the @p->state value just checked and
1123 * not expected to change. If it changes, i.e. @p might have woken up,
1124 * then return zero. When we succeed in waiting for @p to be off its CPU,
1125 * we return a positive number (its total switch count). If a second call
1126 * a short while later returns the same number, the caller can be sure that
1127 * @p has remained unscheduled the whole time.
1129 * The caller must ensure that the task *will* unschedule sometime soon,
1130 * else this function might spin for a *long* time. This function can't
1131 * be called with interrupts off, or it may introduce deadlock with
1132 * smp_call_function() if an IPI is sent by the same process we are
1133 * waiting to become inactive.
1135 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1137 unsigned long flags;
1144 * We do the initial early heuristics without holding
1145 * any task-queue locks at all. We'll only try to get
1146 * the runqueue lock when things look like they will
1152 * If the task is actively running on another CPU
1153 * still, just relax and busy-wait without holding
1156 * NOTE! Since we don't hold any locks, it's not
1157 * even sure that "rq" stays as the right runqueue!
1158 * But we don't care, since "task_running()" will
1159 * return false if the runqueue has changed and p
1160 * is actually now running somewhere else!
1162 while (task_running(rq, p)) {
1163 if (match_state && unlikely(p->state != match_state))
1169 * Ok, time to look more closely! We need the rq
1170 * lock now, to be *sure*. If we're wrong, we'll
1171 * just go back and repeat.
1173 rq = task_rq_lock(p, &flags);
1174 trace_sched_wait_task(p);
1175 running = task_running(rq, p);
1178 if (!match_state || p->state == match_state)
1179 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1180 task_rq_unlock(rq, p, &flags);
1183 * If it changed from the expected state, bail out now.
1185 if (unlikely(!ncsw))
1189 * Was it really running after all now that we
1190 * checked with the proper locks actually held?
1192 * Oops. Go back and try again..
1194 if (unlikely(running)) {
1200 * It's not enough that it's not actively running,
1201 * it must be off the runqueue _entirely_, and not
1204 * So if it was still runnable (but just not actively
1205 * running right now), it's preempted, and we should
1206 * yield - it could be a while.
1208 if (unlikely(on_rq)) {
1209 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1211 set_current_state(TASK_UNINTERRUPTIBLE);
1212 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1217 * Ahh, all good. It wasn't running, and it wasn't
1218 * runnable, which means that it will never become
1219 * running in the future either. We're all done!
1228 * kick_process - kick a running thread to enter/exit the kernel
1229 * @p: the to-be-kicked thread
1231 * Cause a process which is running on another CPU to enter
1232 * kernel-mode, without any delay. (to get signals handled.)
1234 * NOTE: this function doesn't have to take the runqueue lock,
1235 * because all it wants to ensure is that the remote task enters
1236 * the kernel. If the IPI races and the task has been migrated
1237 * to another CPU then no harm is done and the purpose has been
1240 void kick_process(struct task_struct *p)
1246 if ((cpu != smp_processor_id()) && task_curr(p))
1247 smp_send_reschedule(cpu);
1250 EXPORT_SYMBOL_GPL(kick_process);
1251 #endif /* CONFIG_SMP */
1255 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1257 static int select_fallback_rq(int cpu, struct task_struct *p)
1259 int nid = cpu_to_node(cpu);
1260 const struct cpumask *nodemask = NULL;
1261 enum { cpuset, possible, fail } state = cpuset;
1265 * If the node that the cpu is on has been offlined, cpu_to_node()
1266 * will return -1. There is no cpu on the node, and we should
1267 * select the cpu on the other node.
1270 nodemask = cpumask_of_node(nid);
1272 /* Look for allowed, online CPU in same node. */
1273 for_each_cpu(dest_cpu, nodemask) {
1274 if (!cpu_online(dest_cpu))
1276 if (!cpu_active(dest_cpu))
1278 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1284 /* Any allowed, online CPU? */
1285 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1286 if (!cpu_online(dest_cpu))
1288 if (!cpu_active(dest_cpu))
1295 /* No more Mr. Nice Guy. */
1296 cpuset_cpus_allowed_fallback(p);
1301 do_set_cpus_allowed(p, cpu_possible_mask);
1312 if (state != cpuset) {
1314 * Don't tell them about moving exiting tasks or
1315 * kernel threads (both mm NULL), since they never
1318 if (p->mm && printk_ratelimit()) {
1319 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1320 task_pid_nr(p), p->comm, cpu);
1328 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1331 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1333 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1336 * In order not to call set_task_cpu() on a blocking task we need
1337 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1340 * Since this is common to all placement strategies, this lives here.
1342 * [ this allows ->select_task() to simply return task_cpu(p) and
1343 * not worry about this generic constraint ]
1345 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1347 cpu = select_fallback_rq(task_cpu(p), p);
1352 static void update_avg(u64 *avg, u64 sample)
1354 s64 diff = sample - *avg;
1360 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1362 #ifdef CONFIG_SCHEDSTATS
1363 struct rq *rq = this_rq();
1366 int this_cpu = smp_processor_id();
1368 if (cpu == this_cpu) {
1369 schedstat_inc(rq, ttwu_local);
1370 schedstat_inc(p, se.statistics.nr_wakeups_local);
1372 struct sched_domain *sd;
1374 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1376 for_each_domain(this_cpu, sd) {
1377 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1378 schedstat_inc(sd, ttwu_wake_remote);
1385 if (wake_flags & WF_MIGRATED)
1386 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1388 #endif /* CONFIG_SMP */
1390 schedstat_inc(rq, ttwu_count);
1391 schedstat_inc(p, se.statistics.nr_wakeups);
1393 if (wake_flags & WF_SYNC)
1394 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1396 #endif /* CONFIG_SCHEDSTATS */
1399 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1401 activate_task(rq, p, en_flags);
1404 /* if a worker is waking up, notify workqueue */
1405 if (p->flags & PF_WQ_WORKER)
1406 wq_worker_waking_up(p, cpu_of(rq));
1410 * Mark the task runnable and perform wakeup-preemption.
1413 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1415 check_preempt_curr(rq, p, wake_flags);
1416 trace_sched_wakeup(p, true);
1418 p->state = TASK_RUNNING;
1420 if (p->sched_class->task_woken)
1421 p->sched_class->task_woken(rq, p);
1423 if (rq->idle_stamp) {
1424 u64 delta = rq_clock(rq) - rq->idle_stamp;
1425 u64 max = 2*rq->max_idle_balance_cost;
1427 update_avg(&rq->avg_idle, delta);
1429 if (rq->avg_idle > max)
1438 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1441 if (p->sched_contributes_to_load)
1442 rq->nr_uninterruptible--;
1445 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1446 ttwu_do_wakeup(rq, p, wake_flags);
1450 * Called in case the task @p isn't fully descheduled from its runqueue,
1451 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1452 * since all we need to do is flip p->state to TASK_RUNNING, since
1453 * the task is still ->on_rq.
1455 static int ttwu_remote(struct task_struct *p, int wake_flags)
1460 rq = __task_rq_lock(p);
1462 /* check_preempt_curr() may use rq clock */
1463 update_rq_clock(rq);
1464 ttwu_do_wakeup(rq, p, wake_flags);
1467 __task_rq_unlock(rq);
1473 static void sched_ttwu_pending(void)
1475 struct rq *rq = this_rq();
1476 struct llist_node *llist = llist_del_all(&rq->wake_list);
1477 struct task_struct *p;
1479 raw_spin_lock(&rq->lock);
1482 p = llist_entry(llist, struct task_struct, wake_entry);
1483 llist = llist_next(llist);
1484 ttwu_do_activate(rq, p, 0);
1487 raw_spin_unlock(&rq->lock);
1490 void scheduler_ipi(void)
1493 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1494 * TIF_NEED_RESCHED remotely (for the first time) will also send
1497 preempt_fold_need_resched();
1499 if (llist_empty(&this_rq()->wake_list)
1500 && !tick_nohz_full_cpu(smp_processor_id())
1501 && !got_nohz_idle_kick())
1505 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1506 * traditionally all their work was done from the interrupt return
1507 * path. Now that we actually do some work, we need to make sure
1510 * Some archs already do call them, luckily irq_enter/exit nest
1513 * Arguably we should visit all archs and update all handlers,
1514 * however a fair share of IPIs are still resched only so this would
1515 * somewhat pessimize the simple resched case.
1518 tick_nohz_full_check();
1519 sched_ttwu_pending();
1522 * Check if someone kicked us for doing the nohz idle load balance.
1524 if (unlikely(got_nohz_idle_kick())) {
1525 this_rq()->idle_balance = 1;
1526 raise_softirq_irqoff(SCHED_SOFTIRQ);
1531 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1533 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1534 smp_send_reschedule(cpu);
1537 bool cpus_share_cache(int this_cpu, int that_cpu)
1539 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1541 #endif /* CONFIG_SMP */
1543 static void ttwu_queue(struct task_struct *p, int cpu)
1545 struct rq *rq = cpu_rq(cpu);
1547 #if defined(CONFIG_SMP)
1548 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1549 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1550 ttwu_queue_remote(p, cpu);
1555 raw_spin_lock(&rq->lock);
1556 ttwu_do_activate(rq, p, 0);
1557 raw_spin_unlock(&rq->lock);
1561 * try_to_wake_up - wake up a thread
1562 * @p: the thread to be awakened
1563 * @state: the mask of task states that can be woken
1564 * @wake_flags: wake modifier flags (WF_*)
1566 * Put it on the run-queue if it's not already there. The "current"
1567 * thread is always on the run-queue (except when the actual
1568 * re-schedule is in progress), and as such you're allowed to do
1569 * the simpler "current->state = TASK_RUNNING" to mark yourself
1570 * runnable without the overhead of this.
1572 * Return: %true if @p was woken up, %false if it was already running.
1573 * or @state didn't match @p's state.
1576 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1578 unsigned long flags;
1579 int cpu, success = 0;
1582 * If we are going to wake up a thread waiting for CONDITION we
1583 * need to ensure that CONDITION=1 done by the caller can not be
1584 * reordered with p->state check below. This pairs with mb() in
1585 * set_current_state() the waiting thread does.
1587 smp_mb__before_spinlock();
1588 raw_spin_lock_irqsave(&p->pi_lock, flags);
1589 if (!(p->state & state))
1592 success = 1; /* we're going to change ->state */
1595 if (p->on_rq && ttwu_remote(p, wake_flags))
1600 * If the owning (remote) cpu is still in the middle of schedule() with
1601 * this task as prev, wait until its done referencing the task.
1606 * Pairs with the smp_wmb() in finish_lock_switch().
1610 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1611 p->state = TASK_WAKING;
1613 if (p->sched_class->task_waking)
1614 p->sched_class->task_waking(p);
1616 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1617 if (task_cpu(p) != cpu) {
1618 wake_flags |= WF_MIGRATED;
1619 set_task_cpu(p, cpu);
1621 #endif /* CONFIG_SMP */
1625 ttwu_stat(p, cpu, wake_flags);
1627 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1633 * try_to_wake_up_local - try to wake up a local task with rq lock held
1634 * @p: the thread to be awakened
1636 * Put @p on the run-queue if it's not already there. The caller must
1637 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1640 static void try_to_wake_up_local(struct task_struct *p)
1642 struct rq *rq = task_rq(p);
1644 if (WARN_ON_ONCE(rq != this_rq()) ||
1645 WARN_ON_ONCE(p == current))
1648 lockdep_assert_held(&rq->lock);
1650 if (!raw_spin_trylock(&p->pi_lock)) {
1651 raw_spin_unlock(&rq->lock);
1652 raw_spin_lock(&p->pi_lock);
1653 raw_spin_lock(&rq->lock);
1656 if (!(p->state & TASK_NORMAL))
1660 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1662 ttwu_do_wakeup(rq, p, 0);
1663 ttwu_stat(p, smp_processor_id(), 0);
1665 raw_spin_unlock(&p->pi_lock);
1669 * wake_up_process - Wake up a specific process
1670 * @p: The process to be woken up.
1672 * Attempt to wake up the nominated process and move it to the set of runnable
1675 * Return: 1 if the process was woken up, 0 if it was already running.
1677 * It may be assumed that this function implies a write memory barrier before
1678 * changing the task state if and only if any tasks are woken up.
1680 int wake_up_process(struct task_struct *p)
1682 WARN_ON(task_is_stopped_or_traced(p));
1683 return try_to_wake_up(p, TASK_NORMAL, 0);
1685 EXPORT_SYMBOL(wake_up_process);
1687 int wake_up_state(struct task_struct *p, unsigned int state)
1689 return try_to_wake_up(p, state, 0);
1693 * Perform scheduler related setup for a newly forked process p.
1694 * p is forked by current.
1696 * __sched_fork() is basic setup used by init_idle() too:
1698 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1703 p->se.exec_start = 0;
1704 p->se.sum_exec_runtime = 0;
1705 p->se.prev_sum_exec_runtime = 0;
1706 p->se.nr_migrations = 0;
1708 INIT_LIST_HEAD(&p->se.group_node);
1710 #ifdef CONFIG_SCHEDSTATS
1711 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1714 RB_CLEAR_NODE(&p->dl.rb_node);
1715 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1716 p->dl.dl_runtime = p->dl.runtime = 0;
1717 p->dl.dl_deadline = p->dl.deadline = 0;
1718 p->dl.dl_period = 0;
1721 INIT_LIST_HEAD(&p->rt.run_list);
1723 #ifdef CONFIG_PREEMPT_NOTIFIERS
1724 INIT_HLIST_HEAD(&p->preempt_notifiers);
1727 #ifdef CONFIG_NUMA_BALANCING
1728 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1729 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1730 p->mm->numa_scan_seq = 0;
1733 if (clone_flags & CLONE_VM)
1734 p->numa_preferred_nid = current->numa_preferred_nid;
1736 p->numa_preferred_nid = -1;
1738 p->node_stamp = 0ULL;
1739 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1740 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1741 p->numa_work.next = &p->numa_work;
1742 p->numa_faults_memory = NULL;
1743 p->numa_faults_buffer_memory = NULL;
1744 p->last_task_numa_placement = 0;
1745 p->last_sum_exec_runtime = 0;
1747 INIT_LIST_HEAD(&p->numa_entry);
1748 p->numa_group = NULL;
1749 #endif /* CONFIG_NUMA_BALANCING */
1752 #ifdef CONFIG_NUMA_BALANCING
1753 #ifdef CONFIG_SCHED_DEBUG
1754 void set_numabalancing_state(bool enabled)
1757 sched_feat_set("NUMA");
1759 sched_feat_set("NO_NUMA");
1762 __read_mostly bool numabalancing_enabled;
1764 void set_numabalancing_state(bool enabled)
1766 numabalancing_enabled = enabled;
1768 #endif /* CONFIG_SCHED_DEBUG */
1770 #ifdef CONFIG_PROC_SYSCTL
1771 int sysctl_numa_balancing(struct ctl_table *table, int write,
1772 void __user *buffer, size_t *lenp, loff_t *ppos)
1776 int state = numabalancing_enabled;
1778 if (write && !capable(CAP_SYS_ADMIN))
1783 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1787 set_numabalancing_state(state);
1794 * fork()/clone()-time setup:
1796 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1798 unsigned long flags;
1799 int cpu = get_cpu();
1801 __sched_fork(clone_flags, p);
1803 * We mark the process as running here. This guarantees that
1804 * nobody will actually run it, and a signal or other external
1805 * event cannot wake it up and insert it on the runqueue either.
1807 p->state = TASK_RUNNING;
1810 * Make sure we do not leak PI boosting priority to the child.
1812 p->prio = current->normal_prio;
1815 * Revert to default priority/policy on fork if requested.
1817 if (unlikely(p->sched_reset_on_fork)) {
1818 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1819 p->policy = SCHED_NORMAL;
1820 p->static_prio = NICE_TO_PRIO(0);
1822 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1823 p->static_prio = NICE_TO_PRIO(0);
1825 p->prio = p->normal_prio = __normal_prio(p);
1829 * We don't need the reset flag anymore after the fork. It has
1830 * fulfilled its duty:
1832 p->sched_reset_on_fork = 0;
1835 if (dl_prio(p->prio)) {
1838 } else if (rt_prio(p->prio)) {
1839 p->sched_class = &rt_sched_class;
1841 p->sched_class = &fair_sched_class;
1844 if (p->sched_class->task_fork)
1845 p->sched_class->task_fork(p);
1848 * The child is not yet in the pid-hash so no cgroup attach races,
1849 * and the cgroup is pinned to this child due to cgroup_fork()
1850 * is ran before sched_fork().
1852 * Silence PROVE_RCU.
1854 raw_spin_lock_irqsave(&p->pi_lock, flags);
1855 set_task_cpu(p, cpu);
1856 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1858 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1859 if (likely(sched_info_on()))
1860 memset(&p->sched_info, 0, sizeof(p->sched_info));
1862 #if defined(CONFIG_SMP)
1865 init_task_preempt_count(p);
1867 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1868 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1875 unsigned long to_ratio(u64 period, u64 runtime)
1877 if (runtime == RUNTIME_INF)
1881 * Doing this here saves a lot of checks in all
1882 * the calling paths, and returning zero seems
1883 * safe for them anyway.
1888 return div64_u64(runtime << 20, period);
1892 inline struct dl_bw *dl_bw_of(int i)
1894 return &cpu_rq(i)->rd->dl_bw;
1897 static inline int dl_bw_cpus(int i)
1899 struct root_domain *rd = cpu_rq(i)->rd;
1902 for_each_cpu_and(i, rd->span, cpu_active_mask)
1908 inline struct dl_bw *dl_bw_of(int i)
1910 return &cpu_rq(i)->dl.dl_bw;
1913 static inline int dl_bw_cpus(int i)
1920 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1922 dl_b->total_bw -= tsk_bw;
1926 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1928 dl_b->total_bw += tsk_bw;
1932 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1934 return dl_b->bw != -1 &&
1935 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1939 * We must be sure that accepting a new task (or allowing changing the
1940 * parameters of an existing one) is consistent with the bandwidth
1941 * constraints. If yes, this function also accordingly updates the currently
1942 * allocated bandwidth to reflect the new situation.
1944 * This function is called while holding p's rq->lock.
1946 static int dl_overflow(struct task_struct *p, int policy,
1947 const struct sched_attr *attr)
1950 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1951 u64 period = attr->sched_period ?: attr->sched_deadline;
1952 u64 runtime = attr->sched_runtime;
1953 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1956 if (new_bw == p->dl.dl_bw)
1960 * Either if a task, enters, leave, or stays -deadline but changes
1961 * its parameters, we may need to update accordingly the total
1962 * allocated bandwidth of the container.
1964 raw_spin_lock(&dl_b->lock);
1965 cpus = dl_bw_cpus(task_cpu(p));
1966 if (dl_policy(policy) && !task_has_dl_policy(p) &&
1967 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
1968 __dl_add(dl_b, new_bw);
1970 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
1971 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
1972 __dl_clear(dl_b, p->dl.dl_bw);
1973 __dl_add(dl_b, new_bw);
1975 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
1976 __dl_clear(dl_b, p->dl.dl_bw);
1979 raw_spin_unlock(&dl_b->lock);
1984 extern void init_dl_bw(struct dl_bw *dl_b);
1987 * wake_up_new_task - wake up a newly created task for the first time.
1989 * This function will do some initial scheduler statistics housekeeping
1990 * that must be done for every newly created context, then puts the task
1991 * on the runqueue and wakes it.
1993 void wake_up_new_task(struct task_struct *p)
1995 unsigned long flags;
1998 raw_spin_lock_irqsave(&p->pi_lock, flags);
2001 * Fork balancing, do it here and not earlier because:
2002 * - cpus_allowed can change in the fork path
2003 * - any previously selected cpu might disappear through hotplug
2005 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2008 /* Initialize new task's runnable average */
2009 init_task_runnable_average(p);
2010 rq = __task_rq_lock(p);
2011 activate_task(rq, p, 0);
2013 trace_sched_wakeup_new(p, true);
2014 check_preempt_curr(rq, p, WF_FORK);
2016 if (p->sched_class->task_woken)
2017 p->sched_class->task_woken(rq, p);
2019 task_rq_unlock(rq, p, &flags);
2022 #ifdef CONFIG_PREEMPT_NOTIFIERS
2025 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2026 * @notifier: notifier struct to register
2028 void preempt_notifier_register(struct preempt_notifier *notifier)
2030 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2032 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2035 * preempt_notifier_unregister - no longer interested in preemption notifications
2036 * @notifier: notifier struct to unregister
2038 * This is safe to call from within a preemption notifier.
2040 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2042 hlist_del(¬ifier->link);
2044 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2046 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2048 struct preempt_notifier *notifier;
2050 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2051 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2055 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2056 struct task_struct *next)
2058 struct preempt_notifier *notifier;
2060 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2061 notifier->ops->sched_out(notifier, next);
2064 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2066 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2071 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2072 struct task_struct *next)
2076 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2079 * prepare_task_switch - prepare to switch tasks
2080 * @rq: the runqueue preparing to switch
2081 * @prev: the current task that is being switched out
2082 * @next: the task we are going to switch to.
2084 * This is called with the rq lock held and interrupts off. It must
2085 * be paired with a subsequent finish_task_switch after the context
2088 * prepare_task_switch sets up locking and calls architecture specific
2092 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2093 struct task_struct *next)
2095 trace_sched_switch(prev, next);
2096 sched_info_switch(rq, prev, next);
2097 perf_event_task_sched_out(prev, next);
2098 fire_sched_out_preempt_notifiers(prev, next);
2099 prepare_lock_switch(rq, next);
2100 prepare_arch_switch(next);
2104 * finish_task_switch - clean up after a task-switch
2105 * @rq: runqueue associated with task-switch
2106 * @prev: the thread we just switched away from.
2108 * finish_task_switch must be called after the context switch, paired
2109 * with a prepare_task_switch call before the context switch.
2110 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2111 * and do any other architecture-specific cleanup actions.
2113 * Note that we may have delayed dropping an mm in context_switch(). If
2114 * so, we finish that here outside of the runqueue lock. (Doing it
2115 * with the lock held can cause deadlocks; see schedule() for
2118 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2119 __releases(rq->lock)
2121 struct mm_struct *mm = rq->prev_mm;
2127 * A task struct has one reference for the use as "current".
2128 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2129 * schedule one last time. The schedule call will never return, and
2130 * the scheduled task must drop that reference.
2131 * The test for TASK_DEAD must occur while the runqueue locks are
2132 * still held, otherwise prev could be scheduled on another cpu, die
2133 * there before we look at prev->state, and then the reference would
2135 * Manfred Spraul <manfred@colorfullife.com>
2137 prev_state = prev->state;
2138 vtime_task_switch(prev);
2139 finish_arch_switch(prev);
2140 perf_event_task_sched_in(prev, current);
2141 finish_lock_switch(rq, prev);
2142 finish_arch_post_lock_switch();
2144 fire_sched_in_preempt_notifiers(current);
2147 if (unlikely(prev_state == TASK_DEAD)) {
2148 if (prev->sched_class->task_dead)
2149 prev->sched_class->task_dead(prev);
2152 * Remove function-return probe instances associated with this
2153 * task and put them back on the free list.
2155 kprobe_flush_task(prev);
2156 put_task_struct(prev);
2159 tick_nohz_task_switch(current);
2164 /* rq->lock is NOT held, but preemption is disabled */
2165 static inline void post_schedule(struct rq *rq)
2167 if (rq->post_schedule) {
2168 unsigned long flags;
2170 raw_spin_lock_irqsave(&rq->lock, flags);
2171 if (rq->curr->sched_class->post_schedule)
2172 rq->curr->sched_class->post_schedule(rq);
2173 raw_spin_unlock_irqrestore(&rq->lock, flags);
2175 rq->post_schedule = 0;
2181 static inline void post_schedule(struct rq *rq)
2188 * schedule_tail - first thing a freshly forked thread must call.
2189 * @prev: the thread we just switched away from.
2191 asmlinkage void schedule_tail(struct task_struct *prev)
2192 __releases(rq->lock)
2194 struct rq *rq = this_rq();
2196 finish_task_switch(rq, prev);
2199 * FIXME: do we need to worry about rq being invalidated by the
2204 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2205 /* In this case, finish_task_switch does not reenable preemption */
2208 if (current->set_child_tid)
2209 put_user(task_pid_vnr(current), current->set_child_tid);
2213 * context_switch - switch to the new MM and the new
2214 * thread's register state.
2217 context_switch(struct rq *rq, struct task_struct *prev,
2218 struct task_struct *next)
2220 struct mm_struct *mm, *oldmm;
2222 prepare_task_switch(rq, prev, next);
2225 oldmm = prev->active_mm;
2227 * For paravirt, this is coupled with an exit in switch_to to
2228 * combine the page table reload and the switch backend into
2231 arch_start_context_switch(prev);
2234 next->active_mm = oldmm;
2235 atomic_inc(&oldmm->mm_count);
2236 enter_lazy_tlb(oldmm, next);
2238 switch_mm(oldmm, mm, next);
2241 prev->active_mm = NULL;
2242 rq->prev_mm = oldmm;
2245 * Since the runqueue lock will be released by the next
2246 * task (which is an invalid locking op but in the case
2247 * of the scheduler it's an obvious special-case), so we
2248 * do an early lockdep release here:
2250 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2251 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2254 context_tracking_task_switch(prev, next);
2255 /* Here we just switch the register state and the stack. */
2256 switch_to(prev, next, prev);
2260 * this_rq must be evaluated again because prev may have moved
2261 * CPUs since it called schedule(), thus the 'rq' on its stack
2262 * frame will be invalid.
2264 finish_task_switch(this_rq(), prev);
2268 * nr_running and nr_context_switches:
2270 * externally visible scheduler statistics: current number of runnable
2271 * threads, total number of context switches performed since bootup.
2273 unsigned long nr_running(void)
2275 unsigned long i, sum = 0;
2277 for_each_online_cpu(i)
2278 sum += cpu_rq(i)->nr_running;
2283 unsigned long long nr_context_switches(void)
2286 unsigned long long sum = 0;
2288 for_each_possible_cpu(i)
2289 sum += cpu_rq(i)->nr_switches;
2294 unsigned long nr_iowait(void)
2296 unsigned long i, sum = 0;
2298 for_each_possible_cpu(i)
2299 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2304 unsigned long nr_iowait_cpu(int cpu)
2306 struct rq *this = cpu_rq(cpu);
2307 return atomic_read(&this->nr_iowait);
2313 * sched_exec - execve() is a valuable balancing opportunity, because at
2314 * this point the task has the smallest effective memory and cache footprint.
2316 void sched_exec(void)
2318 struct task_struct *p = current;
2319 unsigned long flags;
2322 raw_spin_lock_irqsave(&p->pi_lock, flags);
2323 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2324 if (dest_cpu == smp_processor_id())
2327 if (likely(cpu_active(dest_cpu))) {
2328 struct migration_arg arg = { p, dest_cpu };
2330 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2331 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2335 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2340 DEFINE_PER_CPU(struct kernel_stat, kstat);
2341 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2343 EXPORT_PER_CPU_SYMBOL(kstat);
2344 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2347 * Return any ns on the sched_clock that have not yet been accounted in
2348 * @p in case that task is currently running.
2350 * Called with task_rq_lock() held on @rq.
2352 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2356 if (task_current(rq, p)) {
2357 update_rq_clock(rq);
2358 ns = rq_clock_task(rq) - p->se.exec_start;
2366 unsigned long long task_delta_exec(struct task_struct *p)
2368 unsigned long flags;
2372 rq = task_rq_lock(p, &flags);
2373 ns = do_task_delta_exec(p, rq);
2374 task_rq_unlock(rq, p, &flags);
2380 * Return accounted runtime for the task.
2381 * In case the task is currently running, return the runtime plus current's
2382 * pending runtime that have not been accounted yet.
2384 unsigned long long task_sched_runtime(struct task_struct *p)
2386 unsigned long flags;
2390 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2392 * 64-bit doesn't need locks to atomically read a 64bit value.
2393 * So we have a optimization chance when the task's delta_exec is 0.
2394 * Reading ->on_cpu is racy, but this is ok.
2396 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2397 * If we race with it entering cpu, unaccounted time is 0. This is
2398 * indistinguishable from the read occurring a few cycles earlier.
2401 return p->se.sum_exec_runtime;
2404 rq = task_rq_lock(p, &flags);
2405 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2406 task_rq_unlock(rq, p, &flags);
2412 * This function gets called by the timer code, with HZ frequency.
2413 * We call it with interrupts disabled.
2415 void scheduler_tick(void)
2417 int cpu = smp_processor_id();
2418 struct rq *rq = cpu_rq(cpu);
2419 struct task_struct *curr = rq->curr;
2423 raw_spin_lock(&rq->lock);
2424 update_rq_clock(rq);
2425 curr->sched_class->task_tick(rq, curr, 0);
2426 update_cpu_load_active(rq);
2427 raw_spin_unlock(&rq->lock);
2429 perf_event_task_tick();
2432 rq->idle_balance = idle_cpu(cpu);
2433 trigger_load_balance(rq);
2435 rq_last_tick_reset(rq);
2438 #ifdef CONFIG_NO_HZ_FULL
2440 * scheduler_tick_max_deferment
2442 * Keep at least one tick per second when a single
2443 * active task is running because the scheduler doesn't
2444 * yet completely support full dynticks environment.
2446 * This makes sure that uptime, CFS vruntime, load
2447 * balancing, etc... continue to move forward, even
2448 * with a very low granularity.
2450 * Return: Maximum deferment in nanoseconds.
2452 u64 scheduler_tick_max_deferment(void)
2454 struct rq *rq = this_rq();
2455 unsigned long next, now = ACCESS_ONCE(jiffies);
2457 next = rq->last_sched_tick + HZ;
2459 if (time_before_eq(next, now))
2462 return jiffies_to_nsecs(next - now);
2466 notrace unsigned long get_parent_ip(unsigned long addr)
2468 if (in_lock_functions(addr)) {
2469 addr = CALLER_ADDR2;
2470 if (in_lock_functions(addr))
2471 addr = CALLER_ADDR3;
2476 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2477 defined(CONFIG_PREEMPT_TRACER))
2479 void __kprobes preempt_count_add(int val)
2481 #ifdef CONFIG_DEBUG_PREEMPT
2485 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2488 __preempt_count_add(val);
2489 #ifdef CONFIG_DEBUG_PREEMPT
2491 * Spinlock count overflowing soon?
2493 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2496 if (preempt_count() == val) {
2497 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2498 #ifdef CONFIG_DEBUG_PREEMPT
2499 current->preempt_disable_ip = ip;
2501 trace_preempt_off(CALLER_ADDR0, ip);
2504 EXPORT_SYMBOL(preempt_count_add);
2506 void __kprobes preempt_count_sub(int val)
2508 #ifdef CONFIG_DEBUG_PREEMPT
2512 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2515 * Is the spinlock portion underflowing?
2517 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2518 !(preempt_count() & PREEMPT_MASK)))
2522 if (preempt_count() == val)
2523 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2524 __preempt_count_sub(val);
2526 EXPORT_SYMBOL(preempt_count_sub);
2531 * Print scheduling while atomic bug:
2533 static noinline void __schedule_bug(struct task_struct *prev)
2535 if (oops_in_progress)
2538 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2539 prev->comm, prev->pid, preempt_count());
2541 debug_show_held_locks(prev);
2543 if (irqs_disabled())
2544 print_irqtrace_events(prev);
2545 #ifdef CONFIG_DEBUG_PREEMPT
2546 if (in_atomic_preempt_off()) {
2547 pr_err("Preemption disabled at:");
2548 print_ip_sym(current->preempt_disable_ip);
2553 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2557 * Various schedule()-time debugging checks and statistics:
2559 static inline void schedule_debug(struct task_struct *prev)
2562 * Test if we are atomic. Since do_exit() needs to call into
2563 * schedule() atomically, we ignore that path. Otherwise whine
2564 * if we are scheduling when we should not.
2566 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2567 __schedule_bug(prev);
2570 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2572 schedstat_inc(this_rq(), sched_count);
2576 * Pick up the highest-prio task:
2578 static inline struct task_struct *
2579 pick_next_task(struct rq *rq, struct task_struct *prev)
2581 const struct sched_class *class = &fair_sched_class;
2582 struct task_struct *p;
2585 * Optimization: we know that if all tasks are in
2586 * the fair class we can call that function directly:
2588 if (likely(prev->sched_class == class &&
2589 rq->nr_running == rq->cfs.h_nr_running)) {
2590 p = fair_sched_class.pick_next_task(rq, prev);
2591 if (likely(p && p != RETRY_TASK))
2596 for_each_class(class) {
2597 p = class->pick_next_task(rq, prev);
2599 if (unlikely(p == RETRY_TASK))
2605 BUG(); /* the idle class will always have a runnable task */
2609 * __schedule() is the main scheduler function.
2611 * The main means of driving the scheduler and thus entering this function are:
2613 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2615 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2616 * paths. For example, see arch/x86/entry_64.S.
2618 * To drive preemption between tasks, the scheduler sets the flag in timer
2619 * interrupt handler scheduler_tick().
2621 * 3. Wakeups don't really cause entry into schedule(). They add a
2622 * task to the run-queue and that's it.
2624 * Now, if the new task added to the run-queue preempts the current
2625 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2626 * called on the nearest possible occasion:
2628 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2630 * - in syscall or exception context, at the next outmost
2631 * preempt_enable(). (this might be as soon as the wake_up()'s
2634 * - in IRQ context, return from interrupt-handler to
2635 * preemptible context
2637 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2640 * - cond_resched() call
2641 * - explicit schedule() call
2642 * - return from syscall or exception to user-space
2643 * - return from interrupt-handler to user-space
2645 static void __sched __schedule(void)
2647 struct task_struct *prev, *next;
2648 unsigned long *switch_count;
2654 cpu = smp_processor_id();
2656 rcu_note_context_switch(cpu);
2659 schedule_debug(prev);
2661 if (sched_feat(HRTICK))
2665 * Make sure that signal_pending_state()->signal_pending() below
2666 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2667 * done by the caller to avoid the race with signal_wake_up().
2669 smp_mb__before_spinlock();
2670 raw_spin_lock_irq(&rq->lock);
2672 switch_count = &prev->nivcsw;
2673 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2674 if (unlikely(signal_pending_state(prev->state, prev))) {
2675 prev->state = TASK_RUNNING;
2677 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2681 * If a worker went to sleep, notify and ask workqueue
2682 * whether it wants to wake up a task to maintain
2685 if (prev->flags & PF_WQ_WORKER) {
2686 struct task_struct *to_wakeup;
2688 to_wakeup = wq_worker_sleeping(prev, cpu);
2690 try_to_wake_up_local(to_wakeup);
2693 switch_count = &prev->nvcsw;
2696 if (prev->on_rq || rq->skip_clock_update < 0)
2697 update_rq_clock(rq);
2699 next = pick_next_task(rq, prev);
2700 clear_tsk_need_resched(prev);
2701 clear_preempt_need_resched();
2702 rq->skip_clock_update = 0;
2704 if (likely(prev != next)) {
2709 context_switch(rq, prev, next); /* unlocks the rq */
2711 * The context switch have flipped the stack from under us
2712 * and restored the local variables which were saved when
2713 * this task called schedule() in the past. prev == current
2714 * is still correct, but it can be moved to another cpu/rq.
2716 cpu = smp_processor_id();
2719 raw_spin_unlock_irq(&rq->lock);
2723 sched_preempt_enable_no_resched();
2728 static inline void sched_submit_work(struct task_struct *tsk)
2730 if (!tsk->state || tsk_is_pi_blocked(tsk))
2733 * If we are going to sleep and we have plugged IO queued,
2734 * make sure to submit it to avoid deadlocks.
2736 if (blk_needs_flush_plug(tsk))
2737 blk_schedule_flush_plug(tsk);
2740 asmlinkage void __sched schedule(void)
2742 struct task_struct *tsk = current;
2744 sched_submit_work(tsk);
2747 EXPORT_SYMBOL(schedule);
2749 #ifdef CONFIG_CONTEXT_TRACKING
2750 asmlinkage void __sched schedule_user(void)
2753 * If we come here after a random call to set_need_resched(),
2754 * or we have been woken up remotely but the IPI has not yet arrived,
2755 * we haven't yet exited the RCU idle mode. Do it here manually until
2756 * we find a better solution.
2765 * schedule_preempt_disabled - called with preemption disabled
2767 * Returns with preemption disabled. Note: preempt_count must be 1
2769 void __sched schedule_preempt_disabled(void)
2771 sched_preempt_enable_no_resched();
2776 #ifdef CONFIG_PREEMPT
2778 * this is the entry point to schedule() from in-kernel preemption
2779 * off of preempt_enable. Kernel preemptions off return from interrupt
2780 * occur there and call schedule directly.
2782 asmlinkage void __sched notrace preempt_schedule(void)
2785 * If there is a non-zero preempt_count or interrupts are disabled,
2786 * we do not want to preempt the current task. Just return..
2788 if (likely(!preemptible()))
2792 __preempt_count_add(PREEMPT_ACTIVE);
2794 __preempt_count_sub(PREEMPT_ACTIVE);
2797 * Check again in case we missed a preemption opportunity
2798 * between schedule and now.
2801 } while (need_resched());
2803 EXPORT_SYMBOL(preempt_schedule);
2804 #endif /* CONFIG_PREEMPT */
2807 * this is the entry point to schedule() from kernel preemption
2808 * off of irq context.
2809 * Note, that this is called and return with irqs disabled. This will
2810 * protect us against recursive calling from irq.
2812 asmlinkage void __sched preempt_schedule_irq(void)
2814 enum ctx_state prev_state;
2816 /* Catch callers which need to be fixed */
2817 BUG_ON(preempt_count() || !irqs_disabled());
2819 prev_state = exception_enter();
2822 __preempt_count_add(PREEMPT_ACTIVE);
2825 local_irq_disable();
2826 __preempt_count_sub(PREEMPT_ACTIVE);
2829 * Check again in case we missed a preemption opportunity
2830 * between schedule and now.
2833 } while (need_resched());
2835 exception_exit(prev_state);
2838 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2841 return try_to_wake_up(curr->private, mode, wake_flags);
2843 EXPORT_SYMBOL(default_wake_function);
2846 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2848 unsigned long flags;
2851 init_waitqueue_entry(&wait, current);
2853 __set_current_state(state);
2855 spin_lock_irqsave(&q->lock, flags);
2856 __add_wait_queue(q, &wait);
2857 spin_unlock(&q->lock);
2858 timeout = schedule_timeout(timeout);
2859 spin_lock_irq(&q->lock);
2860 __remove_wait_queue(q, &wait);
2861 spin_unlock_irqrestore(&q->lock, flags);
2866 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2868 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2870 EXPORT_SYMBOL(interruptible_sleep_on);
2873 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2875 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2877 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2879 void __sched sleep_on(wait_queue_head_t *q)
2881 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2883 EXPORT_SYMBOL(sleep_on);
2885 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2887 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2889 EXPORT_SYMBOL(sleep_on_timeout);
2891 #ifdef CONFIG_RT_MUTEXES
2894 * rt_mutex_setprio - set the current priority of a task
2896 * @prio: prio value (kernel-internal form)
2898 * This function changes the 'effective' priority of a task. It does
2899 * not touch ->normal_prio like __setscheduler().
2901 * Used by the rt_mutex code to implement priority inheritance
2902 * logic. Call site only calls if the priority of the task changed.
2904 void rt_mutex_setprio(struct task_struct *p, int prio)
2906 int oldprio, on_rq, running, enqueue_flag = 0;
2908 const struct sched_class *prev_class;
2910 BUG_ON(prio > MAX_PRIO);
2912 rq = __task_rq_lock(p);
2915 * Idle task boosting is a nono in general. There is one
2916 * exception, when PREEMPT_RT and NOHZ is active:
2918 * The idle task calls get_next_timer_interrupt() and holds
2919 * the timer wheel base->lock on the CPU and another CPU wants
2920 * to access the timer (probably to cancel it). We can safely
2921 * ignore the boosting request, as the idle CPU runs this code
2922 * with interrupts disabled and will complete the lock
2923 * protected section without being interrupted. So there is no
2924 * real need to boost.
2926 if (unlikely(p == rq->idle)) {
2927 WARN_ON(p != rq->curr);
2928 WARN_ON(p->pi_blocked_on);
2932 trace_sched_pi_setprio(p, prio);
2933 p->pi_top_task = rt_mutex_get_top_task(p);
2935 prev_class = p->sched_class;
2937 running = task_current(rq, p);
2939 dequeue_task(rq, p, 0);
2941 p->sched_class->put_prev_task(rq, p);
2944 * Boosting condition are:
2945 * 1. -rt task is running and holds mutex A
2946 * --> -dl task blocks on mutex A
2948 * 2. -dl task is running and holds mutex A
2949 * --> -dl task blocks on mutex A and could preempt the
2952 if (dl_prio(prio)) {
2953 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2954 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2955 p->dl.dl_boosted = 1;
2956 p->dl.dl_throttled = 0;
2957 enqueue_flag = ENQUEUE_REPLENISH;
2959 p->dl.dl_boosted = 0;
2960 p->sched_class = &dl_sched_class;
2961 } else if (rt_prio(prio)) {
2962 if (dl_prio(oldprio))
2963 p->dl.dl_boosted = 0;
2965 enqueue_flag = ENQUEUE_HEAD;
2966 p->sched_class = &rt_sched_class;
2968 if (dl_prio(oldprio))
2969 p->dl.dl_boosted = 0;
2970 p->sched_class = &fair_sched_class;
2976 p->sched_class->set_curr_task(rq);
2978 enqueue_task(rq, p, enqueue_flag);
2980 check_class_changed(rq, p, prev_class, oldprio);
2982 __task_rq_unlock(rq);
2986 void set_user_nice(struct task_struct *p, long nice)
2988 int old_prio, delta, on_rq;
2989 unsigned long flags;
2992 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
2995 * We have to be careful, if called from sys_setpriority(),
2996 * the task might be in the middle of scheduling on another CPU.
2998 rq = task_rq_lock(p, &flags);
3000 * The RT priorities are set via sched_setscheduler(), but we still
3001 * allow the 'normal' nice value to be set - but as expected
3002 * it wont have any effect on scheduling until the task is
3003 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3005 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3006 p->static_prio = NICE_TO_PRIO(nice);
3011 dequeue_task(rq, p, 0);
3013 p->static_prio = NICE_TO_PRIO(nice);
3016 p->prio = effective_prio(p);
3017 delta = p->prio - old_prio;
3020 enqueue_task(rq, p, 0);
3022 * If the task increased its priority or is running and
3023 * lowered its priority, then reschedule its CPU:
3025 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3026 resched_task(rq->curr);
3029 task_rq_unlock(rq, p, &flags);
3031 EXPORT_SYMBOL(set_user_nice);
3034 * can_nice - check if a task can reduce its nice value
3038 int can_nice(const struct task_struct *p, const int nice)
3040 /* convert nice value [19,-20] to rlimit style value [1,40] */
3041 int nice_rlim = 20 - nice;
3043 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3044 capable(CAP_SYS_NICE));
3047 #ifdef __ARCH_WANT_SYS_NICE
3050 * sys_nice - change the priority of the current process.
3051 * @increment: priority increment
3053 * sys_setpriority is a more generic, but much slower function that
3054 * does similar things.
3056 SYSCALL_DEFINE1(nice, int, increment)
3061 * Setpriority might change our priority at the same moment.
3062 * We don't have to worry. Conceptually one call occurs first
3063 * and we have a single winner.
3065 if (increment < -40)
3070 nice = task_nice(current) + increment;
3071 if (nice < MIN_NICE)
3073 if (nice > MAX_NICE)
3076 if (increment < 0 && !can_nice(current, nice))
3079 retval = security_task_setnice(current, nice);
3083 set_user_nice(current, nice);
3090 * task_prio - return the priority value of a given task.
3091 * @p: the task in question.
3093 * Return: The priority value as seen by users in /proc.
3094 * RT tasks are offset by -200. Normal tasks are centered
3095 * around 0, value goes from -16 to +15.
3097 int task_prio(const struct task_struct *p)
3099 return p->prio - MAX_RT_PRIO;
3103 * idle_cpu - is a given cpu idle currently?
3104 * @cpu: the processor in question.
3106 * Return: 1 if the CPU is currently idle. 0 otherwise.
3108 int idle_cpu(int cpu)
3110 struct rq *rq = cpu_rq(cpu);
3112 if (rq->curr != rq->idle)
3119 if (!llist_empty(&rq->wake_list))
3127 * idle_task - return the idle task for a given cpu.
3128 * @cpu: the processor in question.
3130 * Return: The idle task for the cpu @cpu.
3132 struct task_struct *idle_task(int cpu)
3134 return cpu_rq(cpu)->idle;
3138 * find_process_by_pid - find a process with a matching PID value.
3139 * @pid: the pid in question.
3141 * The task of @pid, if found. %NULL otherwise.
3143 static struct task_struct *find_process_by_pid(pid_t pid)
3145 return pid ? find_task_by_vpid(pid) : current;
3149 * This function initializes the sched_dl_entity of a newly becoming
3150 * SCHED_DEADLINE task.
3152 * Only the static values are considered here, the actual runtime and the
3153 * absolute deadline will be properly calculated when the task is enqueued
3154 * for the first time with its new policy.
3157 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3159 struct sched_dl_entity *dl_se = &p->dl;
3161 init_dl_task_timer(dl_se);
3162 dl_se->dl_runtime = attr->sched_runtime;
3163 dl_se->dl_deadline = attr->sched_deadline;
3164 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3165 dl_se->flags = attr->sched_flags;
3166 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3167 dl_se->dl_throttled = 0;
3171 static void __setscheduler_params(struct task_struct *p,
3172 const struct sched_attr *attr)
3174 int policy = attr->sched_policy;
3176 if (policy == -1) /* setparam */
3181 if (dl_policy(policy))
3182 __setparam_dl(p, attr);
3183 else if (fair_policy(policy))
3184 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3187 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3188 * !rt_policy. Always setting this ensures that things like
3189 * getparam()/getattr() don't report silly values for !rt tasks.
3191 p->rt_priority = attr->sched_priority;
3192 p->normal_prio = normal_prio(p);
3196 /* Actually do priority change: must hold pi & rq lock. */
3197 static void __setscheduler(struct rq *rq, struct task_struct *p,
3198 const struct sched_attr *attr)
3200 __setscheduler_params(p, attr);
3203 * If we get here, there was no pi waiters boosting the
3204 * task. It is safe to use the normal prio.
3206 p->prio = normal_prio(p);
3208 if (dl_prio(p->prio))
3209 p->sched_class = &dl_sched_class;
3210 else if (rt_prio(p->prio))
3211 p->sched_class = &rt_sched_class;
3213 p->sched_class = &fair_sched_class;
3217 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3219 struct sched_dl_entity *dl_se = &p->dl;
3221 attr->sched_priority = p->rt_priority;
3222 attr->sched_runtime = dl_se->dl_runtime;
3223 attr->sched_deadline = dl_se->dl_deadline;
3224 attr->sched_period = dl_se->dl_period;
3225 attr->sched_flags = dl_se->flags;
3229 * This function validates the new parameters of a -deadline task.
3230 * We ask for the deadline not being zero, and greater or equal
3231 * than the runtime, as well as the period of being zero or
3232 * greater than deadline. Furthermore, we have to be sure that
3233 * user parameters are above the internal resolution (1us); we
3234 * check sched_runtime only since it is always the smaller one.
3237 __checkparam_dl(const struct sched_attr *attr)
3239 return attr && attr->sched_deadline != 0 &&
3240 (attr->sched_period == 0 ||
3241 (s64)(attr->sched_period - attr->sched_deadline) >= 0) &&
3242 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 &&
3243 attr->sched_runtime >= (2 << (DL_SCALE - 1));
3247 * check the target process has a UID that matches the current process's
3249 static bool check_same_owner(struct task_struct *p)
3251 const struct cred *cred = current_cred(), *pcred;
3255 pcred = __task_cred(p);
3256 match = (uid_eq(cred->euid, pcred->euid) ||
3257 uid_eq(cred->euid, pcred->uid));
3262 static int __sched_setscheduler(struct task_struct *p,
3263 const struct sched_attr *attr,
3266 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3267 MAX_RT_PRIO - 1 - attr->sched_priority;
3268 int retval, oldprio, oldpolicy = -1, on_rq, running;
3269 int policy = attr->sched_policy;
3270 unsigned long flags;
3271 const struct sched_class *prev_class;
3275 /* may grab non-irq protected spin_locks */
3276 BUG_ON(in_interrupt());
3278 /* double check policy once rq lock held */
3280 reset_on_fork = p->sched_reset_on_fork;
3281 policy = oldpolicy = p->policy;
3283 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3285 if (policy != SCHED_DEADLINE &&
3286 policy != SCHED_FIFO && policy != SCHED_RR &&
3287 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3288 policy != SCHED_IDLE)
3292 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3296 * Valid priorities for SCHED_FIFO and SCHED_RR are
3297 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3298 * SCHED_BATCH and SCHED_IDLE is 0.
3300 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3301 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3303 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3304 (rt_policy(policy) != (attr->sched_priority != 0)))
3308 * Allow unprivileged RT tasks to decrease priority:
3310 if (user && !capable(CAP_SYS_NICE)) {
3311 if (fair_policy(policy)) {
3312 if (attr->sched_nice < task_nice(p) &&
3313 !can_nice(p, attr->sched_nice))
3317 if (rt_policy(policy)) {
3318 unsigned long rlim_rtprio =
3319 task_rlimit(p, RLIMIT_RTPRIO);
3321 /* can't set/change the rt policy */
3322 if (policy != p->policy && !rlim_rtprio)
3325 /* can't increase priority */
3326 if (attr->sched_priority > p->rt_priority &&
3327 attr->sched_priority > rlim_rtprio)
3332 * Can't set/change SCHED_DEADLINE policy at all for now
3333 * (safest behavior); in the future we would like to allow
3334 * unprivileged DL tasks to increase their relative deadline
3335 * or reduce their runtime (both ways reducing utilization)
3337 if (dl_policy(policy))
3341 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3342 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3344 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3345 if (!can_nice(p, task_nice(p)))
3349 /* can't change other user's priorities */
3350 if (!check_same_owner(p))
3353 /* Normal users shall not reset the sched_reset_on_fork flag */
3354 if (p->sched_reset_on_fork && !reset_on_fork)
3359 retval = security_task_setscheduler(p);
3365 * make sure no PI-waiters arrive (or leave) while we are
3366 * changing the priority of the task:
3368 * To be able to change p->policy safely, the appropriate
3369 * runqueue lock must be held.
3371 rq = task_rq_lock(p, &flags);
3374 * Changing the policy of the stop threads its a very bad idea
3376 if (p == rq->stop) {
3377 task_rq_unlock(rq, p, &flags);
3382 * If not changing anything there's no need to proceed further,
3383 * but store a possible modification of reset_on_fork.
3385 if (unlikely(policy == p->policy)) {
3386 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3388 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3390 if (dl_policy(policy))
3393 p->sched_reset_on_fork = reset_on_fork;
3394 task_rq_unlock(rq, p, &flags);
3400 #ifdef CONFIG_RT_GROUP_SCHED
3402 * Do not allow realtime tasks into groups that have no runtime
3405 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3406 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3407 !task_group_is_autogroup(task_group(p))) {
3408 task_rq_unlock(rq, p, &flags);
3413 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3414 cpumask_t *span = rq->rd->span;
3417 * Don't allow tasks with an affinity mask smaller than
3418 * the entire root_domain to become SCHED_DEADLINE. We
3419 * will also fail if there's no bandwidth available.
3421 if (!cpumask_subset(span, &p->cpus_allowed) ||
3422 rq->rd->dl_bw.bw == 0) {
3423 task_rq_unlock(rq, p, &flags);
3430 /* recheck policy now with rq lock held */
3431 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3432 policy = oldpolicy = -1;
3433 task_rq_unlock(rq, p, &flags);
3438 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3439 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3442 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3443 task_rq_unlock(rq, p, &flags);
3447 p->sched_reset_on_fork = reset_on_fork;
3451 * Special case for priority boosted tasks.
3453 * If the new priority is lower or equal (user space view)
3454 * than the current (boosted) priority, we just store the new
3455 * normal parameters and do not touch the scheduler class and
3456 * the runqueue. This will be done when the task deboost
3459 if (rt_mutex_check_prio(p, newprio)) {
3460 __setscheduler_params(p, attr);
3461 task_rq_unlock(rq, p, &flags);
3466 running = task_current(rq, p);
3468 dequeue_task(rq, p, 0);
3470 p->sched_class->put_prev_task(rq, p);
3472 prev_class = p->sched_class;
3473 __setscheduler(rq, p, attr);
3476 p->sched_class->set_curr_task(rq);
3479 * We enqueue to tail when the priority of a task is
3480 * increased (user space view).
3482 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3485 check_class_changed(rq, p, prev_class, oldprio);
3486 task_rq_unlock(rq, p, &flags);
3488 rt_mutex_adjust_pi(p);
3493 static int _sched_setscheduler(struct task_struct *p, int policy,
3494 const struct sched_param *param, bool check)
3496 struct sched_attr attr = {
3497 .sched_policy = policy,
3498 .sched_priority = param->sched_priority,
3499 .sched_nice = PRIO_TO_NICE(p->static_prio),
3503 * Fixup the legacy SCHED_RESET_ON_FORK hack
3505 if (policy & SCHED_RESET_ON_FORK) {
3506 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3507 policy &= ~SCHED_RESET_ON_FORK;
3508 attr.sched_policy = policy;
3511 return __sched_setscheduler(p, &attr, check);
3514 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3515 * @p: the task in question.
3516 * @policy: new policy.
3517 * @param: structure containing the new RT priority.
3519 * Return: 0 on success. An error code otherwise.
3521 * NOTE that the task may be already dead.
3523 int sched_setscheduler(struct task_struct *p, int policy,
3524 const struct sched_param *param)
3526 return _sched_setscheduler(p, policy, param, true);
3528 EXPORT_SYMBOL_GPL(sched_setscheduler);
3530 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3532 return __sched_setscheduler(p, attr, true);
3534 EXPORT_SYMBOL_GPL(sched_setattr);
3537 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3538 * @p: the task in question.
3539 * @policy: new policy.
3540 * @param: structure containing the new RT priority.
3542 * Just like sched_setscheduler, only don't bother checking if the
3543 * current context has permission. For example, this is needed in
3544 * stop_machine(): we create temporary high priority worker threads,
3545 * but our caller might not have that capability.
3547 * Return: 0 on success. An error code otherwise.
3549 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3550 const struct sched_param *param)
3552 return _sched_setscheduler(p, policy, param, false);
3556 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3558 struct sched_param lparam;
3559 struct task_struct *p;
3562 if (!param || pid < 0)
3564 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3569 p = find_process_by_pid(pid);
3571 retval = sched_setscheduler(p, policy, &lparam);
3578 * Mimics kernel/events/core.c perf_copy_attr().
3580 static int sched_copy_attr(struct sched_attr __user *uattr,
3581 struct sched_attr *attr)
3586 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3590 * zero the full structure, so that a short copy will be nice.
3592 memset(attr, 0, sizeof(*attr));
3594 ret = get_user(size, &uattr->size);
3598 if (size > PAGE_SIZE) /* silly large */
3601 if (!size) /* abi compat */
3602 size = SCHED_ATTR_SIZE_VER0;
3604 if (size < SCHED_ATTR_SIZE_VER0)
3608 * If we're handed a bigger struct than we know of,
3609 * ensure all the unknown bits are 0 - i.e. new
3610 * user-space does not rely on any kernel feature
3611 * extensions we dont know about yet.
3613 if (size > sizeof(*attr)) {
3614 unsigned char __user *addr;
3615 unsigned char __user *end;
3618 addr = (void __user *)uattr + sizeof(*attr);
3619 end = (void __user *)uattr + size;
3621 for (; addr < end; addr++) {
3622 ret = get_user(val, addr);
3628 size = sizeof(*attr);
3631 ret = copy_from_user(attr, uattr, size);
3636 * XXX: do we want to be lenient like existing syscalls; or do we want
3637 * to be strict and return an error on out-of-bounds values?
3639 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3645 put_user(sizeof(*attr), &uattr->size);
3651 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3652 * @pid: the pid in question.
3653 * @policy: new policy.
3654 * @param: structure containing the new RT priority.
3656 * Return: 0 on success. An error code otherwise.
3658 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3659 struct sched_param __user *, param)
3661 /* negative values for policy are not valid */
3665 return do_sched_setscheduler(pid, policy, param);
3669 * sys_sched_setparam - set/change the RT priority of a thread
3670 * @pid: the pid in question.
3671 * @param: structure containing the new RT priority.
3673 * Return: 0 on success. An error code otherwise.
3675 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3677 return do_sched_setscheduler(pid, -1, param);
3681 * sys_sched_setattr - same as above, but with extended sched_attr
3682 * @pid: the pid in question.
3683 * @uattr: structure containing the extended parameters.
3685 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3686 unsigned int, flags)
3688 struct sched_attr attr;
3689 struct task_struct *p;
3692 if (!uattr || pid < 0 || flags)
3695 if (sched_copy_attr(uattr, &attr))
3700 p = find_process_by_pid(pid);
3702 retval = sched_setattr(p, &attr);
3709 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3710 * @pid: the pid in question.
3712 * Return: On success, the policy of the thread. Otherwise, a negative error
3715 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3717 struct task_struct *p;
3725 p = find_process_by_pid(pid);
3727 retval = security_task_getscheduler(p);
3730 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3737 * sys_sched_getparam - get the RT priority of a thread
3738 * @pid: the pid in question.
3739 * @param: structure containing the RT priority.
3741 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3744 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3746 struct sched_param lp;
3747 struct task_struct *p;
3750 if (!param || pid < 0)
3754 p = find_process_by_pid(pid);
3759 retval = security_task_getscheduler(p);
3763 if (task_has_dl_policy(p)) {
3767 lp.sched_priority = p->rt_priority;
3771 * This one might sleep, we cannot do it with a spinlock held ...
3773 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3782 static int sched_read_attr(struct sched_attr __user *uattr,
3783 struct sched_attr *attr,
3788 if (!access_ok(VERIFY_WRITE, uattr, usize))
3792 * If we're handed a smaller struct than we know of,
3793 * ensure all the unknown bits are 0 - i.e. old
3794 * user-space does not get uncomplete information.
3796 if (usize < sizeof(*attr)) {
3797 unsigned char *addr;
3800 addr = (void *)attr + usize;
3801 end = (void *)attr + sizeof(*attr);
3803 for (; addr < end; addr++) {
3811 ret = copy_to_user(uattr, attr, attr->size);
3824 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3825 * @pid: the pid in question.
3826 * @uattr: structure containing the extended parameters.
3827 * @size: sizeof(attr) for fwd/bwd comp.
3829 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3830 unsigned int, size, unsigned int, flags)
3832 struct sched_attr attr = {
3833 .size = sizeof(struct sched_attr),
3835 struct task_struct *p;
3838 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3839 size < SCHED_ATTR_SIZE_VER0 || flags)
3843 p = find_process_by_pid(pid);
3848 retval = security_task_getscheduler(p);
3852 attr.sched_policy = p->policy;
3853 if (p->sched_reset_on_fork)
3854 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3855 if (task_has_dl_policy(p))
3856 __getparam_dl(p, &attr);
3857 else if (task_has_rt_policy(p))
3858 attr.sched_priority = p->rt_priority;
3860 attr.sched_nice = task_nice(p);
3864 retval = sched_read_attr(uattr, &attr, size);
3872 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3874 cpumask_var_t cpus_allowed, new_mask;
3875 struct task_struct *p;
3880 p = find_process_by_pid(pid);
3886 /* Prevent p going away */
3890 if (p->flags & PF_NO_SETAFFINITY) {
3894 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3898 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3900 goto out_free_cpus_allowed;
3903 if (!check_same_owner(p)) {
3905 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3912 retval = security_task_setscheduler(p);
3917 cpuset_cpus_allowed(p, cpus_allowed);
3918 cpumask_and(new_mask, in_mask, cpus_allowed);
3921 * Since bandwidth control happens on root_domain basis,
3922 * if admission test is enabled, we only admit -deadline
3923 * tasks allowed to run on all the CPUs in the task's
3927 if (task_has_dl_policy(p)) {
3928 const struct cpumask *span = task_rq(p)->rd->span;
3930 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3937 retval = set_cpus_allowed_ptr(p, new_mask);
3940 cpuset_cpus_allowed(p, cpus_allowed);
3941 if (!cpumask_subset(new_mask, cpus_allowed)) {
3943 * We must have raced with a concurrent cpuset
3944 * update. Just reset the cpus_allowed to the
3945 * cpuset's cpus_allowed
3947 cpumask_copy(new_mask, cpus_allowed);
3952 free_cpumask_var(new_mask);
3953 out_free_cpus_allowed:
3954 free_cpumask_var(cpus_allowed);
3960 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3961 struct cpumask *new_mask)
3963 if (len < cpumask_size())
3964 cpumask_clear(new_mask);
3965 else if (len > cpumask_size())
3966 len = cpumask_size();
3968 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3972 * sys_sched_setaffinity - set the cpu affinity of a process
3973 * @pid: pid of the process
3974 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3975 * @user_mask_ptr: user-space pointer to the new cpu mask
3977 * Return: 0 on success. An error code otherwise.
3979 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3980 unsigned long __user *, user_mask_ptr)
3982 cpumask_var_t new_mask;
3985 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3988 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3990 retval = sched_setaffinity(pid, new_mask);
3991 free_cpumask_var(new_mask);
3995 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3997 struct task_struct *p;
3998 unsigned long flags;
4004 p = find_process_by_pid(pid);
4008 retval = security_task_getscheduler(p);
4012 raw_spin_lock_irqsave(&p->pi_lock, flags);
4013 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4014 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4023 * sys_sched_getaffinity - get the cpu affinity of a process
4024 * @pid: pid of the process
4025 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4026 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4028 * Return: 0 on success. An error code otherwise.
4030 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4031 unsigned long __user *, user_mask_ptr)
4036 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4038 if (len & (sizeof(unsigned long)-1))
4041 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4044 ret = sched_getaffinity(pid, mask);
4046 size_t retlen = min_t(size_t, len, cpumask_size());
4048 if (copy_to_user(user_mask_ptr, mask, retlen))
4053 free_cpumask_var(mask);
4059 * sys_sched_yield - yield the current processor to other threads.
4061 * This function yields the current CPU to other tasks. If there are no
4062 * other threads running on this CPU then this function will return.
4066 SYSCALL_DEFINE0(sched_yield)
4068 struct rq *rq = this_rq_lock();
4070 schedstat_inc(rq, yld_count);
4071 current->sched_class->yield_task(rq);
4074 * Since we are going to call schedule() anyway, there's
4075 * no need to preempt or enable interrupts:
4077 __release(rq->lock);
4078 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4079 do_raw_spin_unlock(&rq->lock);
4080 sched_preempt_enable_no_resched();
4087 static void __cond_resched(void)
4089 __preempt_count_add(PREEMPT_ACTIVE);
4091 __preempt_count_sub(PREEMPT_ACTIVE);
4094 int __sched _cond_resched(void)
4096 if (should_resched()) {
4102 EXPORT_SYMBOL(_cond_resched);
4105 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4106 * call schedule, and on return reacquire the lock.
4108 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4109 * operations here to prevent schedule() from being called twice (once via
4110 * spin_unlock(), once by hand).
4112 int __cond_resched_lock(spinlock_t *lock)
4114 int resched = should_resched();
4117 lockdep_assert_held(lock);
4119 if (spin_needbreak(lock) || resched) {
4130 EXPORT_SYMBOL(__cond_resched_lock);
4132 int __sched __cond_resched_softirq(void)
4134 BUG_ON(!in_softirq());
4136 if (should_resched()) {
4144 EXPORT_SYMBOL(__cond_resched_softirq);
4147 * yield - yield the current processor to other threads.
4149 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4151 * The scheduler is at all times free to pick the calling task as the most
4152 * eligible task to run, if removing the yield() call from your code breaks
4153 * it, its already broken.
4155 * Typical broken usage is:
4160 * where one assumes that yield() will let 'the other' process run that will
4161 * make event true. If the current task is a SCHED_FIFO task that will never
4162 * happen. Never use yield() as a progress guarantee!!
4164 * If you want to use yield() to wait for something, use wait_event().
4165 * If you want to use yield() to be 'nice' for others, use cond_resched().
4166 * If you still want to use yield(), do not!
4168 void __sched yield(void)
4170 set_current_state(TASK_RUNNING);
4173 EXPORT_SYMBOL(yield);
4176 * yield_to - yield the current processor to another thread in
4177 * your thread group, or accelerate that thread toward the
4178 * processor it's on.
4180 * @preempt: whether task preemption is allowed or not
4182 * It's the caller's job to ensure that the target task struct
4183 * can't go away on us before we can do any checks.
4186 * true (>0) if we indeed boosted the target task.
4187 * false (0) if we failed to boost the target.
4188 * -ESRCH if there's no task to yield to.
4190 bool __sched yield_to(struct task_struct *p, bool preempt)
4192 struct task_struct *curr = current;
4193 struct rq *rq, *p_rq;
4194 unsigned long flags;
4197 local_irq_save(flags);
4203 * If we're the only runnable task on the rq and target rq also
4204 * has only one task, there's absolutely no point in yielding.
4206 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4211 double_rq_lock(rq, p_rq);
4212 if (task_rq(p) != p_rq) {
4213 double_rq_unlock(rq, p_rq);
4217 if (!curr->sched_class->yield_to_task)
4220 if (curr->sched_class != p->sched_class)
4223 if (task_running(p_rq, p) || p->state)
4226 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4228 schedstat_inc(rq, yld_count);
4230 * Make p's CPU reschedule; pick_next_entity takes care of
4233 if (preempt && rq != p_rq)
4234 resched_task(p_rq->curr);
4238 double_rq_unlock(rq, p_rq);
4240 local_irq_restore(flags);
4247 EXPORT_SYMBOL_GPL(yield_to);
4250 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4251 * that process accounting knows that this is a task in IO wait state.
4253 void __sched io_schedule(void)
4255 struct rq *rq = raw_rq();
4257 delayacct_blkio_start();
4258 atomic_inc(&rq->nr_iowait);
4259 blk_flush_plug(current);
4260 current->in_iowait = 1;
4262 current->in_iowait = 0;
4263 atomic_dec(&rq->nr_iowait);
4264 delayacct_blkio_end();
4266 EXPORT_SYMBOL(io_schedule);
4268 long __sched io_schedule_timeout(long timeout)
4270 struct rq *rq = raw_rq();
4273 delayacct_blkio_start();
4274 atomic_inc(&rq->nr_iowait);
4275 blk_flush_plug(current);
4276 current->in_iowait = 1;
4277 ret = schedule_timeout(timeout);
4278 current->in_iowait = 0;
4279 atomic_dec(&rq->nr_iowait);
4280 delayacct_blkio_end();
4285 * sys_sched_get_priority_max - return maximum RT priority.
4286 * @policy: scheduling class.
4288 * Return: On success, this syscall returns the maximum
4289 * rt_priority that can be used by a given scheduling class.
4290 * On failure, a negative error code is returned.
4292 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4299 ret = MAX_USER_RT_PRIO-1;
4301 case SCHED_DEADLINE:
4312 * sys_sched_get_priority_min - return minimum RT priority.
4313 * @policy: scheduling class.
4315 * Return: On success, this syscall returns the minimum
4316 * rt_priority that can be used by a given scheduling class.
4317 * On failure, a negative error code is returned.
4319 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4328 case SCHED_DEADLINE:
4338 * sys_sched_rr_get_interval - return the default timeslice of a process.
4339 * @pid: pid of the process.
4340 * @interval: userspace pointer to the timeslice value.
4342 * this syscall writes the default timeslice value of a given process
4343 * into the user-space timespec buffer. A value of '0' means infinity.
4345 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4348 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4349 struct timespec __user *, interval)
4351 struct task_struct *p;
4352 unsigned int time_slice;
4353 unsigned long flags;
4363 p = find_process_by_pid(pid);
4367 retval = security_task_getscheduler(p);
4371 rq = task_rq_lock(p, &flags);
4373 if (p->sched_class->get_rr_interval)
4374 time_slice = p->sched_class->get_rr_interval(rq, p);
4375 task_rq_unlock(rq, p, &flags);
4378 jiffies_to_timespec(time_slice, &t);
4379 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4387 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4389 void sched_show_task(struct task_struct *p)
4391 unsigned long free = 0;
4395 state = p->state ? __ffs(p->state) + 1 : 0;
4396 printk(KERN_INFO "%-15.15s %c", p->comm,
4397 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4398 #if BITS_PER_LONG == 32
4399 if (state == TASK_RUNNING)
4400 printk(KERN_CONT " running ");
4402 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4404 if (state == TASK_RUNNING)
4405 printk(KERN_CONT " running task ");
4407 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4409 #ifdef CONFIG_DEBUG_STACK_USAGE
4410 free = stack_not_used(p);
4413 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4415 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4416 task_pid_nr(p), ppid,
4417 (unsigned long)task_thread_info(p)->flags);
4419 print_worker_info(KERN_INFO, p);
4420 show_stack(p, NULL);
4423 void show_state_filter(unsigned long state_filter)
4425 struct task_struct *g, *p;
4427 #if BITS_PER_LONG == 32
4429 " task PC stack pid father\n");
4432 " task PC stack pid father\n");
4435 do_each_thread(g, p) {
4437 * reset the NMI-timeout, listing all files on a slow
4438 * console might take a lot of time:
4440 touch_nmi_watchdog();
4441 if (!state_filter || (p->state & state_filter))
4443 } while_each_thread(g, p);
4445 touch_all_softlockup_watchdogs();
4447 #ifdef CONFIG_SCHED_DEBUG
4448 sysrq_sched_debug_show();
4452 * Only show locks if all tasks are dumped:
4455 debug_show_all_locks();
4458 void init_idle_bootup_task(struct task_struct *idle)
4460 idle->sched_class = &idle_sched_class;
4464 * init_idle - set up an idle thread for a given CPU
4465 * @idle: task in question
4466 * @cpu: cpu the idle task belongs to
4468 * NOTE: this function does not set the idle thread's NEED_RESCHED
4469 * flag, to make booting more robust.
4471 void init_idle(struct task_struct *idle, int cpu)
4473 struct rq *rq = cpu_rq(cpu);
4474 unsigned long flags;
4476 raw_spin_lock_irqsave(&rq->lock, flags);
4478 __sched_fork(0, idle);
4479 idle->state = TASK_RUNNING;
4480 idle->se.exec_start = sched_clock();
4482 do_set_cpus_allowed(idle, cpumask_of(cpu));
4484 * We're having a chicken and egg problem, even though we are
4485 * holding rq->lock, the cpu isn't yet set to this cpu so the
4486 * lockdep check in task_group() will fail.
4488 * Similar case to sched_fork(). / Alternatively we could
4489 * use task_rq_lock() here and obtain the other rq->lock.
4494 __set_task_cpu(idle, cpu);
4497 rq->curr = rq->idle = idle;
4499 #if defined(CONFIG_SMP)
4502 raw_spin_unlock_irqrestore(&rq->lock, flags);
4504 /* Set the preempt count _outside_ the spinlocks! */
4505 init_idle_preempt_count(idle, cpu);
4508 * The idle tasks have their own, simple scheduling class:
4510 idle->sched_class = &idle_sched_class;
4511 ftrace_graph_init_idle_task(idle, cpu);
4512 vtime_init_idle(idle, cpu);
4513 #if defined(CONFIG_SMP)
4514 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4519 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4521 if (p->sched_class && p->sched_class->set_cpus_allowed)
4522 p->sched_class->set_cpus_allowed(p, new_mask);
4524 cpumask_copy(&p->cpus_allowed, new_mask);
4525 p->nr_cpus_allowed = cpumask_weight(new_mask);
4529 * This is how migration works:
4531 * 1) we invoke migration_cpu_stop() on the target CPU using
4533 * 2) stopper starts to run (implicitly forcing the migrated thread
4535 * 3) it checks whether the migrated task is still in the wrong runqueue.
4536 * 4) if it's in the wrong runqueue then the migration thread removes
4537 * it and puts it into the right queue.
4538 * 5) stopper completes and stop_one_cpu() returns and the migration
4543 * Change a given task's CPU affinity. Migrate the thread to a
4544 * proper CPU and schedule it away if the CPU it's executing on
4545 * is removed from the allowed bitmask.
4547 * NOTE: the caller must have a valid reference to the task, the
4548 * task must not exit() & deallocate itself prematurely. The
4549 * call is not atomic; no spinlocks may be held.
4551 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4553 unsigned long flags;
4555 unsigned int dest_cpu;
4558 rq = task_rq_lock(p, &flags);
4560 if (cpumask_equal(&p->cpus_allowed, new_mask))
4563 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4568 do_set_cpus_allowed(p, new_mask);
4570 /* Can the task run on the task's current CPU? If so, we're done */
4571 if (cpumask_test_cpu(task_cpu(p), new_mask))
4574 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4576 struct migration_arg arg = { p, dest_cpu };
4577 /* Need help from migration thread: drop lock and wait. */
4578 task_rq_unlock(rq, p, &flags);
4579 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4580 tlb_migrate_finish(p->mm);
4584 task_rq_unlock(rq, p, &flags);
4588 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4591 * Move (not current) task off this cpu, onto dest cpu. We're doing
4592 * this because either it can't run here any more (set_cpus_allowed()
4593 * away from this CPU, or CPU going down), or because we're
4594 * attempting to rebalance this task on exec (sched_exec).
4596 * So we race with normal scheduler movements, but that's OK, as long
4597 * as the task is no longer on this CPU.
4599 * Returns non-zero if task was successfully migrated.
4601 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4603 struct rq *rq_dest, *rq_src;
4606 if (unlikely(!cpu_active(dest_cpu)))
4609 rq_src = cpu_rq(src_cpu);
4610 rq_dest = cpu_rq(dest_cpu);
4612 raw_spin_lock(&p->pi_lock);
4613 double_rq_lock(rq_src, rq_dest);
4614 /* Already moved. */
4615 if (task_cpu(p) != src_cpu)
4617 /* Affinity changed (again). */
4618 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4622 * If we're not on a rq, the next wake-up will ensure we're
4626 dequeue_task(rq_src, p, 0);
4627 set_task_cpu(p, dest_cpu);
4628 enqueue_task(rq_dest, p, 0);
4629 check_preempt_curr(rq_dest, p, 0);
4634 double_rq_unlock(rq_src, rq_dest);
4635 raw_spin_unlock(&p->pi_lock);
4639 #ifdef CONFIG_NUMA_BALANCING
4640 /* Migrate current task p to target_cpu */
4641 int migrate_task_to(struct task_struct *p, int target_cpu)
4643 struct migration_arg arg = { p, target_cpu };
4644 int curr_cpu = task_cpu(p);
4646 if (curr_cpu == target_cpu)
4649 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4652 /* TODO: This is not properly updating schedstats */
4654 trace_sched_move_numa(p, curr_cpu, target_cpu);
4655 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4659 * Requeue a task on a given node and accurately track the number of NUMA
4660 * tasks on the runqueues
4662 void sched_setnuma(struct task_struct *p, int nid)
4665 unsigned long flags;
4666 bool on_rq, running;
4668 rq = task_rq_lock(p, &flags);
4670 running = task_current(rq, p);
4673 dequeue_task(rq, p, 0);
4675 p->sched_class->put_prev_task(rq, p);
4677 p->numa_preferred_nid = nid;
4680 p->sched_class->set_curr_task(rq);
4682 enqueue_task(rq, p, 0);
4683 task_rq_unlock(rq, p, &flags);
4688 * migration_cpu_stop - this will be executed by a highprio stopper thread
4689 * and performs thread migration by bumping thread off CPU then
4690 * 'pushing' onto another runqueue.
4692 static int migration_cpu_stop(void *data)
4694 struct migration_arg *arg = data;
4697 * The original target cpu might have gone down and we might
4698 * be on another cpu but it doesn't matter.
4700 local_irq_disable();
4701 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4706 #ifdef CONFIG_HOTPLUG_CPU
4709 * Ensures that the idle task is using init_mm right before its cpu goes
4712 void idle_task_exit(void)
4714 struct mm_struct *mm = current->active_mm;
4716 BUG_ON(cpu_online(smp_processor_id()));
4718 if (mm != &init_mm) {
4719 switch_mm(mm, &init_mm, current);
4720 finish_arch_post_lock_switch();
4726 * Since this CPU is going 'away' for a while, fold any nr_active delta
4727 * we might have. Assumes we're called after migrate_tasks() so that the
4728 * nr_active count is stable.
4730 * Also see the comment "Global load-average calculations".
4732 static void calc_load_migrate(struct rq *rq)
4734 long delta = calc_load_fold_active(rq);
4736 atomic_long_add(delta, &calc_load_tasks);
4739 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4743 static const struct sched_class fake_sched_class = {
4744 .put_prev_task = put_prev_task_fake,
4747 static struct task_struct fake_task = {
4749 * Avoid pull_{rt,dl}_task()
4751 .prio = MAX_PRIO + 1,
4752 .sched_class = &fake_sched_class,
4756 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4757 * try_to_wake_up()->select_task_rq().
4759 * Called with rq->lock held even though we'er in stop_machine() and
4760 * there's no concurrency possible, we hold the required locks anyway
4761 * because of lock validation efforts.
4763 static void migrate_tasks(unsigned int dead_cpu)
4765 struct rq *rq = cpu_rq(dead_cpu);
4766 struct task_struct *next, *stop = rq->stop;
4770 * Fudge the rq selection such that the below task selection loop
4771 * doesn't get stuck on the currently eligible stop task.
4773 * We're currently inside stop_machine() and the rq is either stuck
4774 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4775 * either way we should never end up calling schedule() until we're
4781 * put_prev_task() and pick_next_task() sched
4782 * class method both need to have an up-to-date
4783 * value of rq->clock[_task]
4785 update_rq_clock(rq);
4789 * There's this thread running, bail when that's the only
4792 if (rq->nr_running == 1)
4795 next = pick_next_task(rq, &fake_task);
4797 next->sched_class->put_prev_task(rq, next);
4799 /* Find suitable destination for @next, with force if needed. */
4800 dest_cpu = select_fallback_rq(dead_cpu, next);
4801 raw_spin_unlock(&rq->lock);
4803 __migrate_task(next, dead_cpu, dest_cpu);
4805 raw_spin_lock(&rq->lock);
4811 #endif /* CONFIG_HOTPLUG_CPU */
4813 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4815 static struct ctl_table sd_ctl_dir[] = {
4817 .procname = "sched_domain",
4823 static struct ctl_table sd_ctl_root[] = {
4825 .procname = "kernel",
4827 .child = sd_ctl_dir,
4832 static struct ctl_table *sd_alloc_ctl_entry(int n)
4834 struct ctl_table *entry =
4835 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4840 static void sd_free_ctl_entry(struct ctl_table **tablep)
4842 struct ctl_table *entry;
4845 * In the intermediate directories, both the child directory and
4846 * procname are dynamically allocated and could fail but the mode
4847 * will always be set. In the lowest directory the names are
4848 * static strings and all have proc handlers.
4850 for (entry = *tablep; entry->mode; entry++) {
4852 sd_free_ctl_entry(&entry->child);
4853 if (entry->proc_handler == NULL)
4854 kfree(entry->procname);
4861 static int min_load_idx = 0;
4862 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4865 set_table_entry(struct ctl_table *entry,
4866 const char *procname, void *data, int maxlen,
4867 umode_t mode, proc_handler *proc_handler,
4870 entry->procname = procname;
4872 entry->maxlen = maxlen;
4874 entry->proc_handler = proc_handler;
4877 entry->extra1 = &min_load_idx;
4878 entry->extra2 = &max_load_idx;
4882 static struct ctl_table *
4883 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4885 struct ctl_table *table = sd_alloc_ctl_entry(14);
4890 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4891 sizeof(long), 0644, proc_doulongvec_minmax, false);
4892 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4893 sizeof(long), 0644, proc_doulongvec_minmax, false);
4894 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4895 sizeof(int), 0644, proc_dointvec_minmax, true);
4896 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4897 sizeof(int), 0644, proc_dointvec_minmax, true);
4898 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4899 sizeof(int), 0644, proc_dointvec_minmax, true);
4900 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4901 sizeof(int), 0644, proc_dointvec_minmax, true);
4902 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4903 sizeof(int), 0644, proc_dointvec_minmax, true);
4904 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4905 sizeof(int), 0644, proc_dointvec_minmax, false);
4906 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4907 sizeof(int), 0644, proc_dointvec_minmax, false);
4908 set_table_entry(&table[9], "cache_nice_tries",
4909 &sd->cache_nice_tries,
4910 sizeof(int), 0644, proc_dointvec_minmax, false);
4911 set_table_entry(&table[10], "flags", &sd->flags,
4912 sizeof(int), 0644, proc_dointvec_minmax, false);
4913 set_table_entry(&table[11], "max_newidle_lb_cost",
4914 &sd->max_newidle_lb_cost,
4915 sizeof(long), 0644, proc_doulongvec_minmax, false);
4916 set_table_entry(&table[12], "name", sd->name,
4917 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4918 /* &table[13] is terminator */
4923 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4925 struct ctl_table *entry, *table;
4926 struct sched_domain *sd;
4927 int domain_num = 0, i;
4930 for_each_domain(cpu, sd)
4932 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4937 for_each_domain(cpu, sd) {
4938 snprintf(buf, 32, "domain%d", i);
4939 entry->procname = kstrdup(buf, GFP_KERNEL);
4941 entry->child = sd_alloc_ctl_domain_table(sd);
4948 static struct ctl_table_header *sd_sysctl_header;
4949 static void register_sched_domain_sysctl(void)
4951 int i, cpu_num = num_possible_cpus();
4952 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4955 WARN_ON(sd_ctl_dir[0].child);
4956 sd_ctl_dir[0].child = entry;
4961 for_each_possible_cpu(i) {
4962 snprintf(buf, 32, "cpu%d", i);
4963 entry->procname = kstrdup(buf, GFP_KERNEL);
4965 entry->child = sd_alloc_ctl_cpu_table(i);
4969 WARN_ON(sd_sysctl_header);
4970 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4973 /* may be called multiple times per register */
4974 static void unregister_sched_domain_sysctl(void)
4976 if (sd_sysctl_header)
4977 unregister_sysctl_table(sd_sysctl_header);
4978 sd_sysctl_header = NULL;
4979 if (sd_ctl_dir[0].child)
4980 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4983 static void register_sched_domain_sysctl(void)
4986 static void unregister_sched_domain_sysctl(void)
4991 static void set_rq_online(struct rq *rq)
4994 const struct sched_class *class;
4996 cpumask_set_cpu(rq->cpu, rq->rd->online);
4999 for_each_class(class) {
5000 if (class->rq_online)
5001 class->rq_online(rq);
5006 static void set_rq_offline(struct rq *rq)
5009 const struct sched_class *class;
5011 for_each_class(class) {
5012 if (class->rq_offline)
5013 class->rq_offline(rq);
5016 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5022 * migration_call - callback that gets triggered when a CPU is added.
5023 * Here we can start up the necessary migration thread for the new CPU.
5026 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5028 int cpu = (long)hcpu;
5029 unsigned long flags;
5030 struct rq *rq = cpu_rq(cpu);
5032 switch (action & ~CPU_TASKS_FROZEN) {
5034 case CPU_UP_PREPARE:
5035 rq->calc_load_update = calc_load_update;
5039 /* Update our root-domain */
5040 raw_spin_lock_irqsave(&rq->lock, flags);
5042 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5046 raw_spin_unlock_irqrestore(&rq->lock, flags);
5049 #ifdef CONFIG_HOTPLUG_CPU
5051 sched_ttwu_pending();
5052 /* Update our root-domain */
5053 raw_spin_lock_irqsave(&rq->lock, flags);
5055 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5059 BUG_ON(rq->nr_running != 1); /* the migration thread */
5060 raw_spin_unlock_irqrestore(&rq->lock, flags);
5064 calc_load_migrate(rq);
5069 update_max_interval();
5075 * Register at high priority so that task migration (migrate_all_tasks)
5076 * happens before everything else. This has to be lower priority than
5077 * the notifier in the perf_event subsystem, though.
5079 static struct notifier_block migration_notifier = {
5080 .notifier_call = migration_call,
5081 .priority = CPU_PRI_MIGRATION,
5084 static int sched_cpu_active(struct notifier_block *nfb,
5085 unsigned long action, void *hcpu)
5087 switch (action & ~CPU_TASKS_FROZEN) {
5089 case CPU_DOWN_FAILED:
5090 set_cpu_active((long)hcpu, true);
5097 static int sched_cpu_inactive(struct notifier_block *nfb,
5098 unsigned long action, void *hcpu)
5100 unsigned long flags;
5101 long cpu = (long)hcpu;
5103 switch (action & ~CPU_TASKS_FROZEN) {
5104 case CPU_DOWN_PREPARE:
5105 set_cpu_active(cpu, false);
5107 /* explicitly allow suspend */
5108 if (!(action & CPU_TASKS_FROZEN)) {
5109 struct dl_bw *dl_b = dl_bw_of(cpu);
5113 raw_spin_lock_irqsave(&dl_b->lock, flags);
5114 cpus = dl_bw_cpus(cpu);
5115 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5116 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5119 return notifier_from_errno(-EBUSY);
5127 static int __init migration_init(void)
5129 void *cpu = (void *)(long)smp_processor_id();
5132 /* Initialize migration for the boot CPU */
5133 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5134 BUG_ON(err == NOTIFY_BAD);
5135 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5136 register_cpu_notifier(&migration_notifier);
5138 /* Register cpu active notifiers */
5139 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5140 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5144 early_initcall(migration_init);
5149 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5151 #ifdef CONFIG_SCHED_DEBUG
5153 static __read_mostly int sched_debug_enabled;
5155 static int __init sched_debug_setup(char *str)
5157 sched_debug_enabled = 1;
5161 early_param("sched_debug", sched_debug_setup);
5163 static inline bool sched_debug(void)
5165 return sched_debug_enabled;
5168 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5169 struct cpumask *groupmask)
5171 struct sched_group *group = sd->groups;
5174 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5175 cpumask_clear(groupmask);
5177 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5179 if (!(sd->flags & SD_LOAD_BALANCE)) {
5180 printk("does not load-balance\n");
5182 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5187 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5189 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5190 printk(KERN_ERR "ERROR: domain->span does not contain "
5193 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5194 printk(KERN_ERR "ERROR: domain->groups does not contain"
5198 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5202 printk(KERN_ERR "ERROR: group is NULL\n");
5207 * Even though we initialize ->power to something semi-sane,
5208 * we leave power_orig unset. This allows us to detect if
5209 * domain iteration is still funny without causing /0 traps.
5211 if (!group->sgp->power_orig) {
5212 printk(KERN_CONT "\n");
5213 printk(KERN_ERR "ERROR: domain->cpu_power not "
5218 if (!cpumask_weight(sched_group_cpus(group))) {
5219 printk(KERN_CONT "\n");
5220 printk(KERN_ERR "ERROR: empty group\n");
5224 if (!(sd->flags & SD_OVERLAP) &&
5225 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5226 printk(KERN_CONT "\n");
5227 printk(KERN_ERR "ERROR: repeated CPUs\n");
5231 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5233 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5235 printk(KERN_CONT " %s", str);
5236 if (group->sgp->power != SCHED_POWER_SCALE) {
5237 printk(KERN_CONT " (cpu_power = %d)",
5241 group = group->next;
5242 } while (group != sd->groups);
5243 printk(KERN_CONT "\n");
5245 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5246 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5249 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5250 printk(KERN_ERR "ERROR: parent span is not a superset "
5251 "of domain->span\n");
5255 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5259 if (!sched_debug_enabled)
5263 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5267 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5270 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5278 #else /* !CONFIG_SCHED_DEBUG */
5279 # define sched_domain_debug(sd, cpu) do { } while (0)
5280 static inline bool sched_debug(void)
5284 #endif /* CONFIG_SCHED_DEBUG */
5286 static int sd_degenerate(struct sched_domain *sd)
5288 if (cpumask_weight(sched_domain_span(sd)) == 1)
5291 /* Following flags need at least 2 groups */
5292 if (sd->flags & (SD_LOAD_BALANCE |
5293 SD_BALANCE_NEWIDLE |
5297 SD_SHARE_PKG_RESOURCES)) {
5298 if (sd->groups != sd->groups->next)
5302 /* Following flags don't use groups */
5303 if (sd->flags & (SD_WAKE_AFFINE))
5310 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5312 unsigned long cflags = sd->flags, pflags = parent->flags;
5314 if (sd_degenerate(parent))
5317 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5320 /* Flags needing groups don't count if only 1 group in parent */
5321 if (parent->groups == parent->groups->next) {
5322 pflags &= ~(SD_LOAD_BALANCE |
5323 SD_BALANCE_NEWIDLE |
5327 SD_SHARE_PKG_RESOURCES |
5329 if (nr_node_ids == 1)
5330 pflags &= ~SD_SERIALIZE;
5332 if (~cflags & pflags)
5338 static void free_rootdomain(struct rcu_head *rcu)
5340 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5342 cpupri_cleanup(&rd->cpupri);
5343 cpudl_cleanup(&rd->cpudl);
5344 free_cpumask_var(rd->dlo_mask);
5345 free_cpumask_var(rd->rto_mask);
5346 free_cpumask_var(rd->online);
5347 free_cpumask_var(rd->span);
5351 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5353 struct root_domain *old_rd = NULL;
5354 unsigned long flags;
5356 raw_spin_lock_irqsave(&rq->lock, flags);
5361 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5364 cpumask_clear_cpu(rq->cpu, old_rd->span);
5367 * If we dont want to free the old_rd yet then
5368 * set old_rd to NULL to skip the freeing later
5371 if (!atomic_dec_and_test(&old_rd->refcount))
5375 atomic_inc(&rd->refcount);
5378 cpumask_set_cpu(rq->cpu, rd->span);
5379 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5382 raw_spin_unlock_irqrestore(&rq->lock, flags);
5385 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5388 static int init_rootdomain(struct root_domain *rd)
5390 memset(rd, 0, sizeof(*rd));
5392 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5394 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5396 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5398 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5401 init_dl_bw(&rd->dl_bw);
5402 if (cpudl_init(&rd->cpudl) != 0)
5405 if (cpupri_init(&rd->cpupri) != 0)
5410 free_cpumask_var(rd->rto_mask);
5412 free_cpumask_var(rd->dlo_mask);
5414 free_cpumask_var(rd->online);
5416 free_cpumask_var(rd->span);
5422 * By default the system creates a single root-domain with all cpus as
5423 * members (mimicking the global state we have today).
5425 struct root_domain def_root_domain;
5427 static void init_defrootdomain(void)
5429 init_rootdomain(&def_root_domain);
5431 atomic_set(&def_root_domain.refcount, 1);
5434 static struct root_domain *alloc_rootdomain(void)
5436 struct root_domain *rd;
5438 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5442 if (init_rootdomain(rd) != 0) {
5450 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5452 struct sched_group *tmp, *first;
5461 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5466 } while (sg != first);
5469 static void free_sched_domain(struct rcu_head *rcu)
5471 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5474 * If its an overlapping domain it has private groups, iterate and
5477 if (sd->flags & SD_OVERLAP) {
5478 free_sched_groups(sd->groups, 1);
5479 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5480 kfree(sd->groups->sgp);
5486 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5488 call_rcu(&sd->rcu, free_sched_domain);
5491 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5493 for (; sd; sd = sd->parent)
5494 destroy_sched_domain(sd, cpu);
5498 * Keep a special pointer to the highest sched_domain that has
5499 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5500 * allows us to avoid some pointer chasing select_idle_sibling().
5502 * Also keep a unique ID per domain (we use the first cpu number in
5503 * the cpumask of the domain), this allows us to quickly tell if
5504 * two cpus are in the same cache domain, see cpus_share_cache().
5506 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5507 DEFINE_PER_CPU(int, sd_llc_size);
5508 DEFINE_PER_CPU(int, sd_llc_id);
5509 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5510 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5511 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5513 static void update_top_cache_domain(int cpu)
5515 struct sched_domain *sd;
5516 struct sched_domain *busy_sd = NULL;
5520 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5522 id = cpumask_first(sched_domain_span(sd));
5523 size = cpumask_weight(sched_domain_span(sd));
5524 busy_sd = sd->parent; /* sd_busy */
5526 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5528 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5529 per_cpu(sd_llc_size, cpu) = size;
5530 per_cpu(sd_llc_id, cpu) = id;
5532 sd = lowest_flag_domain(cpu, SD_NUMA);
5533 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5535 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5536 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5540 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5541 * hold the hotplug lock.
5544 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5546 struct rq *rq = cpu_rq(cpu);
5547 struct sched_domain *tmp;
5549 /* Remove the sched domains which do not contribute to scheduling. */
5550 for (tmp = sd; tmp; ) {
5551 struct sched_domain *parent = tmp->parent;
5555 if (sd_parent_degenerate(tmp, parent)) {
5556 tmp->parent = parent->parent;
5558 parent->parent->child = tmp;
5560 * Transfer SD_PREFER_SIBLING down in case of a
5561 * degenerate parent; the spans match for this
5562 * so the property transfers.
5564 if (parent->flags & SD_PREFER_SIBLING)
5565 tmp->flags |= SD_PREFER_SIBLING;
5566 destroy_sched_domain(parent, cpu);
5571 if (sd && sd_degenerate(sd)) {
5574 destroy_sched_domain(tmp, cpu);
5579 sched_domain_debug(sd, cpu);
5581 rq_attach_root(rq, rd);
5583 rcu_assign_pointer(rq->sd, sd);
5584 destroy_sched_domains(tmp, cpu);
5586 update_top_cache_domain(cpu);
5589 /* cpus with isolated domains */
5590 static cpumask_var_t cpu_isolated_map;
5592 /* Setup the mask of cpus configured for isolated domains */
5593 static int __init isolated_cpu_setup(char *str)
5595 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5596 cpulist_parse(str, cpu_isolated_map);
5600 __setup("isolcpus=", isolated_cpu_setup);
5602 static const struct cpumask *cpu_cpu_mask(int cpu)
5604 return cpumask_of_node(cpu_to_node(cpu));
5608 struct sched_domain **__percpu sd;
5609 struct sched_group **__percpu sg;
5610 struct sched_group_power **__percpu sgp;
5614 struct sched_domain ** __percpu sd;
5615 struct root_domain *rd;
5625 struct sched_domain_topology_level;
5627 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5628 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5630 #define SDTL_OVERLAP 0x01
5632 struct sched_domain_topology_level {
5633 sched_domain_init_f init;
5634 sched_domain_mask_f mask;
5637 struct sd_data data;
5641 * Build an iteration mask that can exclude certain CPUs from the upwards
5644 * Asymmetric node setups can result in situations where the domain tree is of
5645 * unequal depth, make sure to skip domains that already cover the entire
5648 * In that case build_sched_domains() will have terminated the iteration early
5649 * and our sibling sd spans will be empty. Domains should always include the
5650 * cpu they're built on, so check that.
5653 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5655 const struct cpumask *span = sched_domain_span(sd);
5656 struct sd_data *sdd = sd->private;
5657 struct sched_domain *sibling;
5660 for_each_cpu(i, span) {
5661 sibling = *per_cpu_ptr(sdd->sd, i);
5662 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5665 cpumask_set_cpu(i, sched_group_mask(sg));
5670 * Return the canonical balance cpu for this group, this is the first cpu
5671 * of this group that's also in the iteration mask.
5673 int group_balance_cpu(struct sched_group *sg)
5675 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5679 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5681 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5682 const struct cpumask *span = sched_domain_span(sd);
5683 struct cpumask *covered = sched_domains_tmpmask;
5684 struct sd_data *sdd = sd->private;
5685 struct sched_domain *child;
5688 cpumask_clear(covered);
5690 for_each_cpu(i, span) {
5691 struct cpumask *sg_span;
5693 if (cpumask_test_cpu(i, covered))
5696 child = *per_cpu_ptr(sdd->sd, i);
5698 /* See the comment near build_group_mask(). */
5699 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5702 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5703 GFP_KERNEL, cpu_to_node(cpu));
5708 sg_span = sched_group_cpus(sg);
5710 child = child->child;
5711 cpumask_copy(sg_span, sched_domain_span(child));
5713 cpumask_set_cpu(i, sg_span);
5715 cpumask_or(covered, covered, sg_span);
5717 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5718 if (atomic_inc_return(&sg->sgp->ref) == 1)
5719 build_group_mask(sd, sg);
5722 * Initialize sgp->power such that even if we mess up the
5723 * domains and no possible iteration will get us here, we won't
5726 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5727 sg->sgp->power_orig = sg->sgp->power;
5730 * Make sure the first group of this domain contains the
5731 * canonical balance cpu. Otherwise the sched_domain iteration
5732 * breaks. See update_sg_lb_stats().
5734 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5735 group_balance_cpu(sg) == cpu)
5745 sd->groups = groups;
5750 free_sched_groups(first, 0);
5755 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5757 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5758 struct sched_domain *child = sd->child;
5761 cpu = cpumask_first(sched_domain_span(child));
5764 *sg = *per_cpu_ptr(sdd->sg, cpu);
5765 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5766 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5773 * build_sched_groups will build a circular linked list of the groups
5774 * covered by the given span, and will set each group's ->cpumask correctly,
5775 * and ->cpu_power to 0.
5777 * Assumes the sched_domain tree is fully constructed
5780 build_sched_groups(struct sched_domain *sd, int cpu)
5782 struct sched_group *first = NULL, *last = NULL;
5783 struct sd_data *sdd = sd->private;
5784 const struct cpumask *span = sched_domain_span(sd);
5785 struct cpumask *covered;
5788 get_group(cpu, sdd, &sd->groups);
5789 atomic_inc(&sd->groups->ref);
5791 if (cpu != cpumask_first(span))
5794 lockdep_assert_held(&sched_domains_mutex);
5795 covered = sched_domains_tmpmask;
5797 cpumask_clear(covered);
5799 for_each_cpu(i, span) {
5800 struct sched_group *sg;
5803 if (cpumask_test_cpu(i, covered))
5806 group = get_group(i, sdd, &sg);
5807 cpumask_clear(sched_group_cpus(sg));
5809 cpumask_setall(sched_group_mask(sg));
5811 for_each_cpu(j, span) {
5812 if (get_group(j, sdd, NULL) != group)
5815 cpumask_set_cpu(j, covered);
5816 cpumask_set_cpu(j, sched_group_cpus(sg));
5831 * Initialize sched groups cpu_power.
5833 * cpu_power indicates the capacity of sched group, which is used while
5834 * distributing the load between different sched groups in a sched domain.
5835 * Typically cpu_power for all the groups in a sched domain will be same unless
5836 * there are asymmetries in the topology. If there are asymmetries, group
5837 * having more cpu_power will pickup more load compared to the group having
5840 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5842 struct sched_group *sg = sd->groups;
5847 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5849 } while (sg != sd->groups);
5851 if (cpu != group_balance_cpu(sg))
5854 update_group_power(sd, cpu);
5855 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5858 int __weak arch_sd_sibling_asym_packing(void)
5860 return 0*SD_ASYM_PACKING;
5864 * Initializers for schedule domains
5865 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5868 #ifdef CONFIG_SCHED_DEBUG
5869 # define SD_INIT_NAME(sd, type) sd->name = #type
5871 # define SD_INIT_NAME(sd, type) do { } while (0)
5874 #define SD_INIT_FUNC(type) \
5875 static noinline struct sched_domain * \
5876 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5878 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5879 *sd = SD_##type##_INIT; \
5880 SD_INIT_NAME(sd, type); \
5881 sd->private = &tl->data; \
5886 #ifdef CONFIG_SCHED_SMT
5887 SD_INIT_FUNC(SIBLING)
5889 #ifdef CONFIG_SCHED_MC
5892 #ifdef CONFIG_SCHED_BOOK
5896 static int default_relax_domain_level = -1;
5897 int sched_domain_level_max;
5899 static int __init setup_relax_domain_level(char *str)
5901 if (kstrtoint(str, 0, &default_relax_domain_level))
5902 pr_warn("Unable to set relax_domain_level\n");
5906 __setup("relax_domain_level=", setup_relax_domain_level);
5908 static void set_domain_attribute(struct sched_domain *sd,
5909 struct sched_domain_attr *attr)
5913 if (!attr || attr->relax_domain_level < 0) {
5914 if (default_relax_domain_level < 0)
5917 request = default_relax_domain_level;
5919 request = attr->relax_domain_level;
5920 if (request < sd->level) {
5921 /* turn off idle balance on this domain */
5922 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5924 /* turn on idle balance on this domain */
5925 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5929 static void __sdt_free(const struct cpumask *cpu_map);
5930 static int __sdt_alloc(const struct cpumask *cpu_map);
5932 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5933 const struct cpumask *cpu_map)
5937 if (!atomic_read(&d->rd->refcount))
5938 free_rootdomain(&d->rd->rcu); /* fall through */
5940 free_percpu(d->sd); /* fall through */
5942 __sdt_free(cpu_map); /* fall through */
5948 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5949 const struct cpumask *cpu_map)
5951 memset(d, 0, sizeof(*d));
5953 if (__sdt_alloc(cpu_map))
5954 return sa_sd_storage;
5955 d->sd = alloc_percpu(struct sched_domain *);
5957 return sa_sd_storage;
5958 d->rd = alloc_rootdomain();
5961 return sa_rootdomain;
5965 * NULL the sd_data elements we've used to build the sched_domain and
5966 * sched_group structure so that the subsequent __free_domain_allocs()
5967 * will not free the data we're using.
5969 static void claim_allocations(int cpu, struct sched_domain *sd)
5971 struct sd_data *sdd = sd->private;
5973 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5974 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5976 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5977 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5979 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5980 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5983 #ifdef CONFIG_SCHED_SMT
5984 static const struct cpumask *cpu_smt_mask(int cpu)
5986 return topology_thread_cpumask(cpu);
5991 * Topology list, bottom-up.
5993 static struct sched_domain_topology_level default_topology[] = {
5994 #ifdef CONFIG_SCHED_SMT
5995 { sd_init_SIBLING, cpu_smt_mask, },
5997 #ifdef CONFIG_SCHED_MC
5998 { sd_init_MC, cpu_coregroup_mask, },
6000 #ifdef CONFIG_SCHED_BOOK
6001 { sd_init_BOOK, cpu_book_mask, },
6003 { sd_init_CPU, cpu_cpu_mask, },
6007 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6009 #define for_each_sd_topology(tl) \
6010 for (tl = sched_domain_topology; tl->init; tl++)
6014 static int sched_domains_numa_levels;
6015 static int *sched_domains_numa_distance;
6016 static struct cpumask ***sched_domains_numa_masks;
6017 static int sched_domains_curr_level;
6019 static inline int sd_local_flags(int level)
6021 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6024 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6027 static struct sched_domain *
6028 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6030 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6031 int level = tl->numa_level;
6032 int sd_weight = cpumask_weight(
6033 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6035 *sd = (struct sched_domain){
6036 .min_interval = sd_weight,
6037 .max_interval = 2*sd_weight,
6039 .imbalance_pct = 125,
6040 .cache_nice_tries = 2,
6047 .flags = 1*SD_LOAD_BALANCE
6048 | 1*SD_BALANCE_NEWIDLE
6053 | 0*SD_SHARE_CPUPOWER
6054 | 0*SD_SHARE_PKG_RESOURCES
6056 | 0*SD_PREFER_SIBLING
6058 | sd_local_flags(level)
6060 .last_balance = jiffies,
6061 .balance_interval = sd_weight,
6063 SD_INIT_NAME(sd, NUMA);
6064 sd->private = &tl->data;
6067 * Ugly hack to pass state to sd_numa_mask()...
6069 sched_domains_curr_level = tl->numa_level;
6074 static const struct cpumask *sd_numa_mask(int cpu)
6076 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6079 static void sched_numa_warn(const char *str)
6081 static int done = false;
6089 printk(KERN_WARNING "ERROR: %s\n\n", str);
6091 for (i = 0; i < nr_node_ids; i++) {
6092 printk(KERN_WARNING " ");
6093 for (j = 0; j < nr_node_ids; j++)
6094 printk(KERN_CONT "%02d ", node_distance(i,j));
6095 printk(KERN_CONT "\n");
6097 printk(KERN_WARNING "\n");
6100 static bool find_numa_distance(int distance)
6104 if (distance == node_distance(0, 0))
6107 for (i = 0; i < sched_domains_numa_levels; i++) {
6108 if (sched_domains_numa_distance[i] == distance)
6115 static void sched_init_numa(void)
6117 int next_distance, curr_distance = node_distance(0, 0);
6118 struct sched_domain_topology_level *tl;
6122 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6123 if (!sched_domains_numa_distance)
6127 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6128 * unique distances in the node_distance() table.
6130 * Assumes node_distance(0,j) includes all distances in
6131 * node_distance(i,j) in order to avoid cubic time.
6133 next_distance = curr_distance;
6134 for (i = 0; i < nr_node_ids; i++) {
6135 for (j = 0; j < nr_node_ids; j++) {
6136 for (k = 0; k < nr_node_ids; k++) {
6137 int distance = node_distance(i, k);
6139 if (distance > curr_distance &&
6140 (distance < next_distance ||
6141 next_distance == curr_distance))
6142 next_distance = distance;
6145 * While not a strong assumption it would be nice to know
6146 * about cases where if node A is connected to B, B is not
6147 * equally connected to A.
6149 if (sched_debug() && node_distance(k, i) != distance)
6150 sched_numa_warn("Node-distance not symmetric");
6152 if (sched_debug() && i && !find_numa_distance(distance))
6153 sched_numa_warn("Node-0 not representative");
6155 if (next_distance != curr_distance) {
6156 sched_domains_numa_distance[level++] = next_distance;
6157 sched_domains_numa_levels = level;
6158 curr_distance = next_distance;
6163 * In case of sched_debug() we verify the above assumption.
6169 * 'level' contains the number of unique distances, excluding the
6170 * identity distance node_distance(i,i).
6172 * The sched_domains_numa_distance[] array includes the actual distance
6177 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6178 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6179 * the array will contain less then 'level' members. This could be
6180 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6181 * in other functions.
6183 * We reset it to 'level' at the end of this function.
6185 sched_domains_numa_levels = 0;
6187 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6188 if (!sched_domains_numa_masks)
6192 * Now for each level, construct a mask per node which contains all
6193 * cpus of nodes that are that many hops away from us.
6195 for (i = 0; i < level; i++) {
6196 sched_domains_numa_masks[i] =
6197 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6198 if (!sched_domains_numa_masks[i])
6201 for (j = 0; j < nr_node_ids; j++) {
6202 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6206 sched_domains_numa_masks[i][j] = mask;
6208 for (k = 0; k < nr_node_ids; k++) {
6209 if (node_distance(j, k) > sched_domains_numa_distance[i])
6212 cpumask_or(mask, mask, cpumask_of_node(k));
6217 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6218 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6223 * Copy the default topology bits..
6225 for (i = 0; default_topology[i].init; i++)
6226 tl[i] = default_topology[i];
6229 * .. and append 'j' levels of NUMA goodness.
6231 for (j = 0; j < level; i++, j++) {
6232 tl[i] = (struct sched_domain_topology_level){
6233 .init = sd_numa_init,
6234 .mask = sd_numa_mask,
6235 .flags = SDTL_OVERLAP,
6240 sched_domain_topology = tl;
6242 sched_domains_numa_levels = level;
6245 static void sched_domains_numa_masks_set(int cpu)
6248 int node = cpu_to_node(cpu);
6250 for (i = 0; i < sched_domains_numa_levels; i++) {
6251 for (j = 0; j < nr_node_ids; j++) {
6252 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6253 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6258 static void sched_domains_numa_masks_clear(int cpu)
6261 for (i = 0; i < sched_domains_numa_levels; i++) {
6262 for (j = 0; j < nr_node_ids; j++)
6263 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6268 * Update sched_domains_numa_masks[level][node] array when new cpus
6271 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6272 unsigned long action,
6275 int cpu = (long)hcpu;
6277 switch (action & ~CPU_TASKS_FROZEN) {
6279 sched_domains_numa_masks_set(cpu);
6283 sched_domains_numa_masks_clear(cpu);
6293 static inline void sched_init_numa(void)
6297 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6298 unsigned long action,
6303 #endif /* CONFIG_NUMA */
6305 static int __sdt_alloc(const struct cpumask *cpu_map)
6307 struct sched_domain_topology_level *tl;
6310 for_each_sd_topology(tl) {
6311 struct sd_data *sdd = &tl->data;
6313 sdd->sd = alloc_percpu(struct sched_domain *);
6317 sdd->sg = alloc_percpu(struct sched_group *);
6321 sdd->sgp = alloc_percpu(struct sched_group_power *);
6325 for_each_cpu(j, cpu_map) {
6326 struct sched_domain *sd;
6327 struct sched_group *sg;
6328 struct sched_group_power *sgp;
6330 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6331 GFP_KERNEL, cpu_to_node(j));
6335 *per_cpu_ptr(sdd->sd, j) = sd;
6337 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6338 GFP_KERNEL, cpu_to_node(j));
6344 *per_cpu_ptr(sdd->sg, j) = sg;
6346 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6347 GFP_KERNEL, cpu_to_node(j));
6351 *per_cpu_ptr(sdd->sgp, j) = sgp;
6358 static void __sdt_free(const struct cpumask *cpu_map)
6360 struct sched_domain_topology_level *tl;
6363 for_each_sd_topology(tl) {
6364 struct sd_data *sdd = &tl->data;
6366 for_each_cpu(j, cpu_map) {
6367 struct sched_domain *sd;
6370 sd = *per_cpu_ptr(sdd->sd, j);
6371 if (sd && (sd->flags & SD_OVERLAP))
6372 free_sched_groups(sd->groups, 0);
6373 kfree(*per_cpu_ptr(sdd->sd, j));
6377 kfree(*per_cpu_ptr(sdd->sg, j));
6379 kfree(*per_cpu_ptr(sdd->sgp, j));
6381 free_percpu(sdd->sd);
6383 free_percpu(sdd->sg);
6385 free_percpu(sdd->sgp);
6390 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6391 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6392 struct sched_domain *child, int cpu)
6394 struct sched_domain *sd = tl->init(tl, cpu);
6398 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6400 sd->level = child->level + 1;
6401 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6405 set_domain_attribute(sd, attr);
6411 * Build sched domains for a given set of cpus and attach the sched domains
6412 * to the individual cpus
6414 static int build_sched_domains(const struct cpumask *cpu_map,
6415 struct sched_domain_attr *attr)
6417 enum s_alloc alloc_state;
6418 struct sched_domain *sd;
6420 int i, ret = -ENOMEM;
6422 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6423 if (alloc_state != sa_rootdomain)
6426 /* Set up domains for cpus specified by the cpu_map. */
6427 for_each_cpu(i, cpu_map) {
6428 struct sched_domain_topology_level *tl;
6431 for_each_sd_topology(tl) {
6432 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6433 if (tl == sched_domain_topology)
6434 *per_cpu_ptr(d.sd, i) = sd;
6435 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6436 sd->flags |= SD_OVERLAP;
6437 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6442 /* Build the groups for the domains */
6443 for_each_cpu(i, cpu_map) {
6444 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6445 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6446 if (sd->flags & SD_OVERLAP) {
6447 if (build_overlap_sched_groups(sd, i))
6450 if (build_sched_groups(sd, i))
6456 /* Calculate CPU power for physical packages and nodes */
6457 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6458 if (!cpumask_test_cpu(i, cpu_map))
6461 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6462 claim_allocations(i, sd);
6463 init_sched_groups_power(i, sd);
6467 /* Attach the domains */
6469 for_each_cpu(i, cpu_map) {
6470 sd = *per_cpu_ptr(d.sd, i);
6471 cpu_attach_domain(sd, d.rd, i);
6477 __free_domain_allocs(&d, alloc_state, cpu_map);
6481 static cpumask_var_t *doms_cur; /* current sched domains */
6482 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6483 static struct sched_domain_attr *dattr_cur;
6484 /* attribues of custom domains in 'doms_cur' */
6487 * Special case: If a kmalloc of a doms_cur partition (array of
6488 * cpumask) fails, then fallback to a single sched domain,
6489 * as determined by the single cpumask fallback_doms.
6491 static cpumask_var_t fallback_doms;
6494 * arch_update_cpu_topology lets virtualized architectures update the
6495 * cpu core maps. It is supposed to return 1 if the topology changed
6496 * or 0 if it stayed the same.
6498 int __attribute__((weak)) arch_update_cpu_topology(void)
6503 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6506 cpumask_var_t *doms;
6508 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6511 for (i = 0; i < ndoms; i++) {
6512 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6513 free_sched_domains(doms, i);
6520 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6523 for (i = 0; i < ndoms; i++)
6524 free_cpumask_var(doms[i]);
6529 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6530 * For now this just excludes isolated cpus, but could be used to
6531 * exclude other special cases in the future.
6533 static int init_sched_domains(const struct cpumask *cpu_map)
6537 arch_update_cpu_topology();
6539 doms_cur = alloc_sched_domains(ndoms_cur);
6541 doms_cur = &fallback_doms;
6542 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6543 err = build_sched_domains(doms_cur[0], NULL);
6544 register_sched_domain_sysctl();
6550 * Detach sched domains from a group of cpus specified in cpu_map
6551 * These cpus will now be attached to the NULL domain
6553 static void detach_destroy_domains(const struct cpumask *cpu_map)
6558 for_each_cpu(i, cpu_map)
6559 cpu_attach_domain(NULL, &def_root_domain, i);
6563 /* handle null as "default" */
6564 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6565 struct sched_domain_attr *new, int idx_new)
6567 struct sched_domain_attr tmp;
6574 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6575 new ? (new + idx_new) : &tmp,
6576 sizeof(struct sched_domain_attr));
6580 * Partition sched domains as specified by the 'ndoms_new'
6581 * cpumasks in the array doms_new[] of cpumasks. This compares
6582 * doms_new[] to the current sched domain partitioning, doms_cur[].
6583 * It destroys each deleted domain and builds each new domain.
6585 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6586 * The masks don't intersect (don't overlap.) We should setup one
6587 * sched domain for each mask. CPUs not in any of the cpumasks will
6588 * not be load balanced. If the same cpumask appears both in the
6589 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6592 * The passed in 'doms_new' should be allocated using
6593 * alloc_sched_domains. This routine takes ownership of it and will
6594 * free_sched_domains it when done with it. If the caller failed the
6595 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6596 * and partition_sched_domains() will fallback to the single partition
6597 * 'fallback_doms', it also forces the domains to be rebuilt.
6599 * If doms_new == NULL it will be replaced with cpu_online_mask.
6600 * ndoms_new == 0 is a special case for destroying existing domains,
6601 * and it will not create the default domain.
6603 * Call with hotplug lock held
6605 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6606 struct sched_domain_attr *dattr_new)
6611 mutex_lock(&sched_domains_mutex);
6613 /* always unregister in case we don't destroy any domains */
6614 unregister_sched_domain_sysctl();
6616 /* Let architecture update cpu core mappings. */
6617 new_topology = arch_update_cpu_topology();
6619 n = doms_new ? ndoms_new : 0;
6621 /* Destroy deleted domains */
6622 for (i = 0; i < ndoms_cur; i++) {
6623 for (j = 0; j < n && !new_topology; j++) {
6624 if (cpumask_equal(doms_cur[i], doms_new[j])
6625 && dattrs_equal(dattr_cur, i, dattr_new, j))
6628 /* no match - a current sched domain not in new doms_new[] */
6629 detach_destroy_domains(doms_cur[i]);
6635 if (doms_new == NULL) {
6637 doms_new = &fallback_doms;
6638 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6639 WARN_ON_ONCE(dattr_new);
6642 /* Build new domains */
6643 for (i = 0; i < ndoms_new; i++) {
6644 for (j = 0; j < n && !new_topology; j++) {
6645 if (cpumask_equal(doms_new[i], doms_cur[j])
6646 && dattrs_equal(dattr_new, i, dattr_cur, j))
6649 /* no match - add a new doms_new */
6650 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6655 /* Remember the new sched domains */
6656 if (doms_cur != &fallback_doms)
6657 free_sched_domains(doms_cur, ndoms_cur);
6658 kfree(dattr_cur); /* kfree(NULL) is safe */
6659 doms_cur = doms_new;
6660 dattr_cur = dattr_new;
6661 ndoms_cur = ndoms_new;
6663 register_sched_domain_sysctl();
6665 mutex_unlock(&sched_domains_mutex);
6668 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6671 * Update cpusets according to cpu_active mask. If cpusets are
6672 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6673 * around partition_sched_domains().
6675 * If we come here as part of a suspend/resume, don't touch cpusets because we
6676 * want to restore it back to its original state upon resume anyway.
6678 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6682 case CPU_ONLINE_FROZEN:
6683 case CPU_DOWN_FAILED_FROZEN:
6686 * num_cpus_frozen tracks how many CPUs are involved in suspend
6687 * resume sequence. As long as this is not the last online
6688 * operation in the resume sequence, just build a single sched
6689 * domain, ignoring cpusets.
6692 if (likely(num_cpus_frozen)) {
6693 partition_sched_domains(1, NULL, NULL);
6698 * This is the last CPU online operation. So fall through and
6699 * restore the original sched domains by considering the
6700 * cpuset configurations.
6704 case CPU_DOWN_FAILED:
6705 cpuset_update_active_cpus(true);
6713 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6717 case CPU_DOWN_PREPARE:
6718 cpuset_update_active_cpus(false);
6720 case CPU_DOWN_PREPARE_FROZEN:
6722 partition_sched_domains(1, NULL, NULL);
6730 void __init sched_init_smp(void)
6732 cpumask_var_t non_isolated_cpus;
6734 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6735 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6740 * There's no userspace yet to cause hotplug operations; hence all the
6741 * cpu masks are stable and all blatant races in the below code cannot
6744 mutex_lock(&sched_domains_mutex);
6745 init_sched_domains(cpu_active_mask);
6746 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6747 if (cpumask_empty(non_isolated_cpus))
6748 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6749 mutex_unlock(&sched_domains_mutex);
6751 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6752 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6753 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6757 /* Move init over to a non-isolated CPU */
6758 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6760 sched_init_granularity();
6761 free_cpumask_var(non_isolated_cpus);
6763 init_sched_rt_class();
6764 init_sched_dl_class();
6767 void __init sched_init_smp(void)
6769 sched_init_granularity();
6771 #endif /* CONFIG_SMP */
6773 const_debug unsigned int sysctl_timer_migration = 1;
6775 int in_sched_functions(unsigned long addr)
6777 return in_lock_functions(addr) ||
6778 (addr >= (unsigned long)__sched_text_start
6779 && addr < (unsigned long)__sched_text_end);
6782 #ifdef CONFIG_CGROUP_SCHED
6784 * Default task group.
6785 * Every task in system belongs to this group at bootup.
6787 struct task_group root_task_group;
6788 LIST_HEAD(task_groups);
6791 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6793 void __init sched_init(void)
6796 unsigned long alloc_size = 0, ptr;
6798 #ifdef CONFIG_FAIR_GROUP_SCHED
6799 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6801 #ifdef CONFIG_RT_GROUP_SCHED
6802 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6804 #ifdef CONFIG_CPUMASK_OFFSTACK
6805 alloc_size += num_possible_cpus() * cpumask_size();
6808 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6810 #ifdef CONFIG_FAIR_GROUP_SCHED
6811 root_task_group.se = (struct sched_entity **)ptr;
6812 ptr += nr_cpu_ids * sizeof(void **);
6814 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6815 ptr += nr_cpu_ids * sizeof(void **);
6817 #endif /* CONFIG_FAIR_GROUP_SCHED */
6818 #ifdef CONFIG_RT_GROUP_SCHED
6819 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6820 ptr += nr_cpu_ids * sizeof(void **);
6822 root_task_group.rt_rq = (struct rt_rq **)ptr;
6823 ptr += nr_cpu_ids * sizeof(void **);
6825 #endif /* CONFIG_RT_GROUP_SCHED */
6826 #ifdef CONFIG_CPUMASK_OFFSTACK
6827 for_each_possible_cpu(i) {
6828 per_cpu(load_balance_mask, i) = (void *)ptr;
6829 ptr += cpumask_size();
6831 #endif /* CONFIG_CPUMASK_OFFSTACK */
6834 init_rt_bandwidth(&def_rt_bandwidth,
6835 global_rt_period(), global_rt_runtime());
6836 init_dl_bandwidth(&def_dl_bandwidth,
6837 global_rt_period(), global_rt_runtime());
6840 init_defrootdomain();
6843 #ifdef CONFIG_RT_GROUP_SCHED
6844 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6845 global_rt_period(), global_rt_runtime());
6846 #endif /* CONFIG_RT_GROUP_SCHED */
6848 #ifdef CONFIG_CGROUP_SCHED
6849 list_add(&root_task_group.list, &task_groups);
6850 INIT_LIST_HEAD(&root_task_group.children);
6851 INIT_LIST_HEAD(&root_task_group.siblings);
6852 autogroup_init(&init_task);
6854 #endif /* CONFIG_CGROUP_SCHED */
6856 for_each_possible_cpu(i) {
6860 raw_spin_lock_init(&rq->lock);
6862 rq->calc_load_active = 0;
6863 rq->calc_load_update = jiffies + LOAD_FREQ;
6864 init_cfs_rq(&rq->cfs);
6865 init_rt_rq(&rq->rt, rq);
6866 init_dl_rq(&rq->dl, rq);
6867 #ifdef CONFIG_FAIR_GROUP_SCHED
6868 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6869 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6871 * How much cpu bandwidth does root_task_group get?
6873 * In case of task-groups formed thr' the cgroup filesystem, it
6874 * gets 100% of the cpu resources in the system. This overall
6875 * system cpu resource is divided among the tasks of
6876 * root_task_group and its child task-groups in a fair manner,
6877 * based on each entity's (task or task-group's) weight
6878 * (se->load.weight).
6880 * In other words, if root_task_group has 10 tasks of weight
6881 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6882 * then A0's share of the cpu resource is:
6884 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6886 * We achieve this by letting root_task_group's tasks sit
6887 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6889 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6890 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6891 #endif /* CONFIG_FAIR_GROUP_SCHED */
6893 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6894 #ifdef CONFIG_RT_GROUP_SCHED
6895 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6898 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6899 rq->cpu_load[j] = 0;
6901 rq->last_load_update_tick = jiffies;
6906 rq->cpu_power = SCHED_POWER_SCALE;
6907 rq->post_schedule = 0;
6908 rq->active_balance = 0;
6909 rq->next_balance = jiffies;
6914 rq->avg_idle = 2*sysctl_sched_migration_cost;
6915 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6917 INIT_LIST_HEAD(&rq->cfs_tasks);
6919 rq_attach_root(rq, &def_root_domain);
6920 #ifdef CONFIG_NO_HZ_COMMON
6923 #ifdef CONFIG_NO_HZ_FULL
6924 rq->last_sched_tick = 0;
6928 atomic_set(&rq->nr_iowait, 0);
6931 set_load_weight(&init_task);
6933 #ifdef CONFIG_PREEMPT_NOTIFIERS
6934 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6938 * The boot idle thread does lazy MMU switching as well:
6940 atomic_inc(&init_mm.mm_count);
6941 enter_lazy_tlb(&init_mm, current);
6944 * Make us the idle thread. Technically, schedule() should not be
6945 * called from this thread, however somewhere below it might be,
6946 * but because we are the idle thread, we just pick up running again
6947 * when this runqueue becomes "idle".
6949 init_idle(current, smp_processor_id());
6951 calc_load_update = jiffies + LOAD_FREQ;
6954 * During early bootup we pretend to be a normal task:
6956 current->sched_class = &fair_sched_class;
6959 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6960 /* May be allocated at isolcpus cmdline parse time */
6961 if (cpu_isolated_map == NULL)
6962 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6963 idle_thread_set_boot_cpu();
6965 init_sched_fair_class();
6967 scheduler_running = 1;
6970 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6971 static inline int preempt_count_equals(int preempt_offset)
6973 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6975 return (nested == preempt_offset);
6978 void __might_sleep(const char *file, int line, int preempt_offset)
6980 static unsigned long prev_jiffy; /* ratelimiting */
6982 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6983 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6984 !is_idle_task(current)) ||
6985 system_state != SYSTEM_RUNNING || oops_in_progress)
6987 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6989 prev_jiffy = jiffies;
6992 "BUG: sleeping function called from invalid context at %s:%d\n",
6995 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6996 in_atomic(), irqs_disabled(),
6997 current->pid, current->comm);
6999 debug_show_held_locks(current);
7000 if (irqs_disabled())
7001 print_irqtrace_events(current);
7002 #ifdef CONFIG_DEBUG_PREEMPT
7003 if (!preempt_count_equals(preempt_offset)) {
7004 pr_err("Preemption disabled at:");
7005 print_ip_sym(current->preempt_disable_ip);
7011 EXPORT_SYMBOL(__might_sleep);
7014 #ifdef CONFIG_MAGIC_SYSRQ
7015 static void normalize_task(struct rq *rq, struct task_struct *p)
7017 const struct sched_class *prev_class = p->sched_class;
7018 struct sched_attr attr = {
7019 .sched_policy = SCHED_NORMAL,
7021 int old_prio = p->prio;
7026 dequeue_task(rq, p, 0);
7027 __setscheduler(rq, p, &attr);
7029 enqueue_task(rq, p, 0);
7030 resched_task(rq->curr);
7033 check_class_changed(rq, p, prev_class, old_prio);
7036 void normalize_rt_tasks(void)
7038 struct task_struct *g, *p;
7039 unsigned long flags;
7042 read_lock_irqsave(&tasklist_lock, flags);
7043 do_each_thread(g, p) {
7045 * Only normalize user tasks:
7050 p->se.exec_start = 0;
7051 #ifdef CONFIG_SCHEDSTATS
7052 p->se.statistics.wait_start = 0;
7053 p->se.statistics.sleep_start = 0;
7054 p->se.statistics.block_start = 0;
7057 if (!dl_task(p) && !rt_task(p)) {
7059 * Renice negative nice level userspace
7062 if (task_nice(p) < 0 && p->mm)
7063 set_user_nice(p, 0);
7067 raw_spin_lock(&p->pi_lock);
7068 rq = __task_rq_lock(p);
7070 normalize_task(rq, p);
7072 __task_rq_unlock(rq);
7073 raw_spin_unlock(&p->pi_lock);
7074 } while_each_thread(g, p);
7076 read_unlock_irqrestore(&tasklist_lock, flags);
7079 #endif /* CONFIG_MAGIC_SYSRQ */
7081 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7083 * These functions are only useful for the IA64 MCA handling, or kdb.
7085 * They can only be called when the whole system has been
7086 * stopped - every CPU needs to be quiescent, and no scheduling
7087 * activity can take place. Using them for anything else would
7088 * be a serious bug, and as a result, they aren't even visible
7089 * under any other configuration.
7093 * curr_task - return the current task for a given cpu.
7094 * @cpu: the processor in question.
7096 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7098 * Return: The current task for @cpu.
7100 struct task_struct *curr_task(int cpu)
7102 return cpu_curr(cpu);
7105 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7109 * set_curr_task - set the current task for a given cpu.
7110 * @cpu: the processor in question.
7111 * @p: the task pointer to set.
7113 * Description: This function must only be used when non-maskable interrupts
7114 * are serviced on a separate stack. It allows the architecture to switch the
7115 * notion of the current task on a cpu in a non-blocking manner. This function
7116 * must be called with all CPU's synchronized, and interrupts disabled, the
7117 * and caller must save the original value of the current task (see
7118 * curr_task() above) and restore that value before reenabling interrupts and
7119 * re-starting the system.
7121 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7123 void set_curr_task(int cpu, struct task_struct *p)
7130 #ifdef CONFIG_CGROUP_SCHED
7131 /* task_group_lock serializes the addition/removal of task groups */
7132 static DEFINE_SPINLOCK(task_group_lock);
7134 static void free_sched_group(struct task_group *tg)
7136 free_fair_sched_group(tg);
7137 free_rt_sched_group(tg);
7142 /* allocate runqueue etc for a new task group */
7143 struct task_group *sched_create_group(struct task_group *parent)
7145 struct task_group *tg;
7147 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7149 return ERR_PTR(-ENOMEM);
7151 if (!alloc_fair_sched_group(tg, parent))
7154 if (!alloc_rt_sched_group(tg, parent))
7160 free_sched_group(tg);
7161 return ERR_PTR(-ENOMEM);
7164 void sched_online_group(struct task_group *tg, struct task_group *parent)
7166 unsigned long flags;
7168 spin_lock_irqsave(&task_group_lock, flags);
7169 list_add_rcu(&tg->list, &task_groups);
7171 WARN_ON(!parent); /* root should already exist */
7173 tg->parent = parent;
7174 INIT_LIST_HEAD(&tg->children);
7175 list_add_rcu(&tg->siblings, &parent->children);
7176 spin_unlock_irqrestore(&task_group_lock, flags);
7179 /* rcu callback to free various structures associated with a task group */
7180 static void free_sched_group_rcu(struct rcu_head *rhp)
7182 /* now it should be safe to free those cfs_rqs */
7183 free_sched_group(container_of(rhp, struct task_group, rcu));
7186 /* Destroy runqueue etc associated with a task group */
7187 void sched_destroy_group(struct task_group *tg)
7189 /* wait for possible concurrent references to cfs_rqs complete */
7190 call_rcu(&tg->rcu, free_sched_group_rcu);
7193 void sched_offline_group(struct task_group *tg)
7195 unsigned long flags;
7198 /* end participation in shares distribution */
7199 for_each_possible_cpu(i)
7200 unregister_fair_sched_group(tg, i);
7202 spin_lock_irqsave(&task_group_lock, flags);
7203 list_del_rcu(&tg->list);
7204 list_del_rcu(&tg->siblings);
7205 spin_unlock_irqrestore(&task_group_lock, flags);
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_css_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 #ifdef CONFIG_RT_GROUP_SCHED
7254 * Ensure that the real time constraints are schedulable.
7256 static DEFINE_MUTEX(rt_constraints_mutex);
7258 /* Must be called with tasklist_lock held */
7259 static inline int tg_has_rt_tasks(struct task_group *tg)
7261 struct task_struct *g, *p;
7263 do_each_thread(g, p) {
7264 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7266 } while_each_thread(g, p);
7271 struct rt_schedulable_data {
7272 struct task_group *tg;
7277 static int tg_rt_schedulable(struct task_group *tg, void *data)
7279 struct rt_schedulable_data *d = data;
7280 struct task_group *child;
7281 unsigned long total, sum = 0;
7282 u64 period, runtime;
7284 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7285 runtime = tg->rt_bandwidth.rt_runtime;
7288 period = d->rt_period;
7289 runtime = d->rt_runtime;
7293 * Cannot have more runtime than the period.
7295 if (runtime > period && runtime != RUNTIME_INF)
7299 * Ensure we don't starve existing RT tasks.
7301 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7304 total = to_ratio(period, runtime);
7307 * Nobody can have more than the global setting allows.
7309 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7313 * The sum of our children's runtime should not exceed our own.
7315 list_for_each_entry_rcu(child, &tg->children, siblings) {
7316 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7317 runtime = child->rt_bandwidth.rt_runtime;
7319 if (child == d->tg) {
7320 period = d->rt_period;
7321 runtime = d->rt_runtime;
7324 sum += to_ratio(period, runtime);
7333 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7337 struct rt_schedulable_data data = {
7339 .rt_period = period,
7340 .rt_runtime = runtime,
7344 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7350 static int tg_set_rt_bandwidth(struct task_group *tg,
7351 u64 rt_period, u64 rt_runtime)
7355 mutex_lock(&rt_constraints_mutex);
7356 read_lock(&tasklist_lock);
7357 err = __rt_schedulable(tg, rt_period, rt_runtime);
7361 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7362 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7363 tg->rt_bandwidth.rt_runtime = rt_runtime;
7365 for_each_possible_cpu(i) {
7366 struct rt_rq *rt_rq = tg->rt_rq[i];
7368 raw_spin_lock(&rt_rq->rt_runtime_lock);
7369 rt_rq->rt_runtime = rt_runtime;
7370 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7372 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7374 read_unlock(&tasklist_lock);
7375 mutex_unlock(&rt_constraints_mutex);
7380 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7382 u64 rt_runtime, rt_period;
7384 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7385 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7386 if (rt_runtime_us < 0)
7387 rt_runtime = RUNTIME_INF;
7389 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7392 static long sched_group_rt_runtime(struct task_group *tg)
7396 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7399 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7400 do_div(rt_runtime_us, NSEC_PER_USEC);
7401 return rt_runtime_us;
7404 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7406 u64 rt_runtime, rt_period;
7408 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7409 rt_runtime = tg->rt_bandwidth.rt_runtime;
7414 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7417 static long sched_group_rt_period(struct task_group *tg)
7421 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7422 do_div(rt_period_us, NSEC_PER_USEC);
7423 return rt_period_us;
7425 #endif /* CONFIG_RT_GROUP_SCHED */
7427 #ifdef CONFIG_RT_GROUP_SCHED
7428 static int sched_rt_global_constraints(void)
7432 mutex_lock(&rt_constraints_mutex);
7433 read_lock(&tasklist_lock);
7434 ret = __rt_schedulable(NULL, 0, 0);
7435 read_unlock(&tasklist_lock);
7436 mutex_unlock(&rt_constraints_mutex);
7441 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7443 /* Don't accept realtime tasks when there is no way for them to run */
7444 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7450 #else /* !CONFIG_RT_GROUP_SCHED */
7451 static int sched_rt_global_constraints(void)
7453 unsigned long flags;
7456 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7457 for_each_possible_cpu(i) {
7458 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7460 raw_spin_lock(&rt_rq->rt_runtime_lock);
7461 rt_rq->rt_runtime = global_rt_runtime();
7462 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7464 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7468 #endif /* CONFIG_RT_GROUP_SCHED */
7470 static int sched_dl_global_constraints(void)
7472 u64 runtime = global_rt_runtime();
7473 u64 period = global_rt_period();
7474 u64 new_bw = to_ratio(period, runtime);
7476 unsigned long flags;
7479 * Here we want to check the bandwidth not being set to some
7480 * value smaller than the currently allocated bandwidth in
7481 * any of the root_domains.
7483 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7484 * cycling on root_domains... Discussion on different/better
7485 * solutions is welcome!
7487 for_each_possible_cpu(cpu) {
7488 struct dl_bw *dl_b = dl_bw_of(cpu);
7490 raw_spin_lock_irqsave(&dl_b->lock, flags);
7491 if (new_bw < dl_b->total_bw)
7493 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7502 static void sched_dl_do_global(void)
7506 unsigned long flags;
7508 def_dl_bandwidth.dl_period = global_rt_period();
7509 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7511 if (global_rt_runtime() != RUNTIME_INF)
7512 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7515 * FIXME: As above...
7517 for_each_possible_cpu(cpu) {
7518 struct dl_bw *dl_b = dl_bw_of(cpu);
7520 raw_spin_lock_irqsave(&dl_b->lock, flags);
7522 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7526 static int sched_rt_global_validate(void)
7528 if (sysctl_sched_rt_period <= 0)
7531 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7532 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7538 static void sched_rt_do_global(void)
7540 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7541 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7544 int sched_rt_handler(struct ctl_table *table, int write,
7545 void __user *buffer, size_t *lenp,
7548 int old_period, old_runtime;
7549 static DEFINE_MUTEX(mutex);
7553 old_period = sysctl_sched_rt_period;
7554 old_runtime = sysctl_sched_rt_runtime;
7556 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7558 if (!ret && write) {
7559 ret = sched_rt_global_validate();
7563 ret = sched_rt_global_constraints();
7567 ret = sched_dl_global_constraints();
7571 sched_rt_do_global();
7572 sched_dl_do_global();
7576 sysctl_sched_rt_period = old_period;
7577 sysctl_sched_rt_runtime = old_runtime;
7579 mutex_unlock(&mutex);
7584 int sched_rr_handler(struct ctl_table *table, int write,
7585 void __user *buffer, size_t *lenp,
7589 static DEFINE_MUTEX(mutex);
7592 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7593 /* make sure that internally we keep jiffies */
7594 /* also, writing zero resets timeslice to default */
7595 if (!ret && write) {
7596 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7597 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7599 mutex_unlock(&mutex);
7603 #ifdef CONFIG_CGROUP_SCHED
7605 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7607 return css ? container_of(css, struct task_group, css) : NULL;
7610 static struct cgroup_subsys_state *
7611 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7613 struct task_group *parent = css_tg(parent_css);
7614 struct task_group *tg;
7617 /* This is early initialization for the top cgroup */
7618 return &root_task_group.css;
7621 tg = sched_create_group(parent);
7623 return ERR_PTR(-ENOMEM);
7628 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7630 struct task_group *tg = css_tg(css);
7631 struct task_group *parent = css_tg(css_parent(css));
7634 sched_online_group(tg, parent);
7638 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7640 struct task_group *tg = css_tg(css);
7642 sched_destroy_group(tg);
7645 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7647 struct task_group *tg = css_tg(css);
7649 sched_offline_group(tg);
7652 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7653 struct cgroup_taskset *tset)
7655 struct task_struct *task;
7657 cgroup_taskset_for_each(task, css, tset) {
7658 #ifdef CONFIG_RT_GROUP_SCHED
7659 if (!sched_rt_can_attach(css_tg(css), task))
7662 /* We don't support RT-tasks being in separate groups */
7663 if (task->sched_class != &fair_sched_class)
7670 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7671 struct cgroup_taskset *tset)
7673 struct task_struct *task;
7675 cgroup_taskset_for_each(task, css, tset)
7676 sched_move_task(task);
7679 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7680 struct cgroup_subsys_state *old_css,
7681 struct task_struct *task)
7684 * cgroup_exit() is called in the copy_process() failure path.
7685 * Ignore this case since the task hasn't ran yet, this avoids
7686 * trying to poke a half freed task state from generic code.
7688 if (!(task->flags & PF_EXITING))
7691 sched_move_task(task);
7694 #ifdef CONFIG_FAIR_GROUP_SCHED
7695 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7696 struct cftype *cftype, u64 shareval)
7698 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7701 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7704 struct task_group *tg = css_tg(css);
7706 return (u64) scale_load_down(tg->shares);
7709 #ifdef CONFIG_CFS_BANDWIDTH
7710 static DEFINE_MUTEX(cfs_constraints_mutex);
7712 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7713 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7715 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7717 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7719 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7720 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7722 if (tg == &root_task_group)
7726 * Ensure we have at some amount of bandwidth every period. This is
7727 * to prevent reaching a state of large arrears when throttled via
7728 * entity_tick() resulting in prolonged exit starvation.
7730 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7734 * Likewise, bound things on the otherside by preventing insane quota
7735 * periods. This also allows us to normalize in computing quota
7738 if (period > max_cfs_quota_period)
7741 mutex_lock(&cfs_constraints_mutex);
7742 ret = __cfs_schedulable(tg, period, quota);
7746 runtime_enabled = quota != RUNTIME_INF;
7747 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7749 * If we need to toggle cfs_bandwidth_used, off->on must occur
7750 * before making related changes, and on->off must occur afterwards
7752 if (runtime_enabled && !runtime_was_enabled)
7753 cfs_bandwidth_usage_inc();
7754 raw_spin_lock_irq(&cfs_b->lock);
7755 cfs_b->period = ns_to_ktime(period);
7756 cfs_b->quota = quota;
7758 __refill_cfs_bandwidth_runtime(cfs_b);
7759 /* restart the period timer (if active) to handle new period expiry */
7760 if (runtime_enabled && cfs_b->timer_active) {
7761 /* force a reprogram */
7762 cfs_b->timer_active = 0;
7763 __start_cfs_bandwidth(cfs_b);
7765 raw_spin_unlock_irq(&cfs_b->lock);
7767 for_each_possible_cpu(i) {
7768 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7769 struct rq *rq = cfs_rq->rq;
7771 raw_spin_lock_irq(&rq->lock);
7772 cfs_rq->runtime_enabled = runtime_enabled;
7773 cfs_rq->runtime_remaining = 0;
7775 if (cfs_rq->throttled)
7776 unthrottle_cfs_rq(cfs_rq);
7777 raw_spin_unlock_irq(&rq->lock);
7779 if (runtime_was_enabled && !runtime_enabled)
7780 cfs_bandwidth_usage_dec();
7782 mutex_unlock(&cfs_constraints_mutex);
7787 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7791 period = ktime_to_ns(tg->cfs_bandwidth.period);
7792 if (cfs_quota_us < 0)
7793 quota = RUNTIME_INF;
7795 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7797 return tg_set_cfs_bandwidth(tg, period, quota);
7800 long tg_get_cfs_quota(struct task_group *tg)
7804 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7807 quota_us = tg->cfs_bandwidth.quota;
7808 do_div(quota_us, NSEC_PER_USEC);
7813 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7817 period = (u64)cfs_period_us * NSEC_PER_USEC;
7818 quota = tg->cfs_bandwidth.quota;
7820 return tg_set_cfs_bandwidth(tg, period, quota);
7823 long tg_get_cfs_period(struct task_group *tg)
7827 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7828 do_div(cfs_period_us, NSEC_PER_USEC);
7830 return cfs_period_us;
7833 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7836 return tg_get_cfs_quota(css_tg(css));
7839 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7840 struct cftype *cftype, s64 cfs_quota_us)
7842 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7845 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7848 return tg_get_cfs_period(css_tg(css));
7851 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7852 struct cftype *cftype, u64 cfs_period_us)
7854 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7857 struct cfs_schedulable_data {
7858 struct task_group *tg;
7863 * normalize group quota/period to be quota/max_period
7864 * note: units are usecs
7866 static u64 normalize_cfs_quota(struct task_group *tg,
7867 struct cfs_schedulable_data *d)
7875 period = tg_get_cfs_period(tg);
7876 quota = tg_get_cfs_quota(tg);
7879 /* note: these should typically be equivalent */
7880 if (quota == RUNTIME_INF || quota == -1)
7883 return to_ratio(period, quota);
7886 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7888 struct cfs_schedulable_data *d = data;
7889 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7890 s64 quota = 0, parent_quota = -1;
7893 quota = RUNTIME_INF;
7895 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7897 quota = normalize_cfs_quota(tg, d);
7898 parent_quota = parent_b->hierarchal_quota;
7901 * ensure max(child_quota) <= parent_quota, inherit when no
7904 if (quota == RUNTIME_INF)
7905 quota = parent_quota;
7906 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7909 cfs_b->hierarchal_quota = quota;
7914 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7917 struct cfs_schedulable_data data = {
7923 if (quota != RUNTIME_INF) {
7924 do_div(data.period, NSEC_PER_USEC);
7925 do_div(data.quota, NSEC_PER_USEC);
7929 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7935 static int cpu_stats_show(struct seq_file *sf, void *v)
7937 struct task_group *tg = css_tg(seq_css(sf));
7938 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7940 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7941 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7942 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7946 #endif /* CONFIG_CFS_BANDWIDTH */
7947 #endif /* CONFIG_FAIR_GROUP_SCHED */
7949 #ifdef CONFIG_RT_GROUP_SCHED
7950 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7951 struct cftype *cft, s64 val)
7953 return sched_group_set_rt_runtime(css_tg(css), val);
7956 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7959 return sched_group_rt_runtime(css_tg(css));
7962 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7963 struct cftype *cftype, u64 rt_period_us)
7965 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7968 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7971 return sched_group_rt_period(css_tg(css));
7973 #endif /* CONFIG_RT_GROUP_SCHED */
7975 static struct cftype cpu_files[] = {
7976 #ifdef CONFIG_FAIR_GROUP_SCHED
7979 .read_u64 = cpu_shares_read_u64,
7980 .write_u64 = cpu_shares_write_u64,
7983 #ifdef CONFIG_CFS_BANDWIDTH
7985 .name = "cfs_quota_us",
7986 .read_s64 = cpu_cfs_quota_read_s64,
7987 .write_s64 = cpu_cfs_quota_write_s64,
7990 .name = "cfs_period_us",
7991 .read_u64 = cpu_cfs_period_read_u64,
7992 .write_u64 = cpu_cfs_period_write_u64,
7996 .seq_show = cpu_stats_show,
7999 #ifdef CONFIG_RT_GROUP_SCHED
8001 .name = "rt_runtime_us",
8002 .read_s64 = cpu_rt_runtime_read,
8003 .write_s64 = cpu_rt_runtime_write,
8006 .name = "rt_period_us",
8007 .read_u64 = cpu_rt_period_read_uint,
8008 .write_u64 = cpu_rt_period_write_uint,
8014 struct cgroup_subsys cpu_cgroup_subsys = {
8016 .css_alloc = cpu_cgroup_css_alloc,
8017 .css_free = cpu_cgroup_css_free,
8018 .css_online = cpu_cgroup_css_online,
8019 .css_offline = cpu_cgroup_css_offline,
8020 .can_attach = cpu_cgroup_can_attach,
8021 .attach = cpu_cgroup_attach,
8022 .exit = cpu_cgroup_exit,
8023 .subsys_id = cpu_cgroup_subsys_id,
8024 .base_cftypes = cpu_files,
8028 #endif /* CONFIG_CGROUP_SCHED */
8030 void dump_cpu_task(int cpu)
8032 pr_info("Task dump for CPU %d:\n", cpu);
8033 sched_show_task(cpu_curr(cpu));