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 <linux/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>
76 #include <linux/compiler.h>
77 #include <linux/nospec.h>
79 #include <asm/switch_to.h>
81 #include <asm/irq_regs.h>
82 #include <asm/mutex.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
94 #ifdef smp_mb__before_atomic
95 void __smp_mb__before_atomic(void)
97 smp_mb__before_atomic();
99 EXPORT_SYMBOL(__smp_mb__before_atomic);
102 #ifdef smp_mb__after_atomic
103 void __smp_mb__after_atomic(void)
105 smp_mb__after_atomic();
107 EXPORT_SYMBOL(__smp_mb__after_atomic);
110 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
113 ktime_t soft, hard, now;
116 if (hrtimer_active(period_timer))
119 now = hrtimer_cb_get_time(period_timer);
120 hrtimer_forward(period_timer, now, period);
122 soft = hrtimer_get_softexpires(period_timer);
123 hard = hrtimer_get_expires(period_timer);
124 delta = ktime_to_ns(ktime_sub(hard, soft));
125 __hrtimer_start_range_ns(period_timer, soft, delta,
126 HRTIMER_MODE_ABS_PINNED, 0);
130 DEFINE_MUTEX(sched_domains_mutex);
131 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
133 static void update_rq_clock_task(struct rq *rq, s64 delta);
135 void update_rq_clock(struct rq *rq)
139 if (rq->skip_clock_update > 0)
142 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
144 update_rq_clock_task(rq, delta);
148 * Debugging: various feature bits
151 #define SCHED_FEAT(name, enabled) \
152 (1UL << __SCHED_FEAT_##name) * enabled |
154 const_debug unsigned int sysctl_sched_features =
155 #include "features.h"
160 #ifdef CONFIG_SCHED_DEBUG
161 #define SCHED_FEAT(name, enabled) \
164 static const char * const sched_feat_names[] = {
165 #include "features.h"
170 static int sched_feat_show(struct seq_file *m, void *v)
174 for (i = 0; i < __SCHED_FEAT_NR; i++) {
175 if (!(sysctl_sched_features & (1UL << i)))
177 seq_printf(m, "%s ", sched_feat_names[i]);
184 #ifdef HAVE_JUMP_LABEL
186 #define jump_label_key__true STATIC_KEY_INIT_TRUE
187 #define jump_label_key__false STATIC_KEY_INIT_FALSE
189 #define SCHED_FEAT(name, enabled) \
190 jump_label_key__##enabled ,
192 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
193 #include "features.h"
198 static void sched_feat_disable(int i)
200 static_key_disable(&sched_feat_keys[i]);
203 static void sched_feat_enable(int i)
205 static_key_enable(&sched_feat_keys[i]);
208 static void sched_feat_disable(int i) { };
209 static void sched_feat_enable(int i) { };
210 #endif /* HAVE_JUMP_LABEL */
212 static int sched_feat_set(char *cmp)
217 if (strncmp(cmp, "NO_", 3) == 0) {
222 for (i = 0; i < __SCHED_FEAT_NR; i++) {
223 if (strcmp(cmp, sched_feat_names[i]) == 0) {
225 sysctl_sched_features &= ~(1UL << i);
226 sched_feat_disable(i);
228 sysctl_sched_features |= (1UL << i);
229 sched_feat_enable(i);
239 sched_feat_write(struct file *filp, const char __user *ubuf,
240 size_t cnt, loff_t *ppos)
249 if (copy_from_user(&buf, ubuf, cnt))
255 i = sched_feat_set(cmp);
256 if (i == __SCHED_FEAT_NR)
264 static int sched_feat_open(struct inode *inode, struct file *filp)
266 return single_open(filp, sched_feat_show, NULL);
269 static const struct file_operations sched_feat_fops = {
270 .open = sched_feat_open,
271 .write = sched_feat_write,
274 .release = single_release,
277 static __init int sched_init_debug(void)
279 debugfs_create_file("sched_features", 0644, NULL, NULL,
284 late_initcall(sched_init_debug);
285 #endif /* CONFIG_SCHED_DEBUG */
288 * Number of tasks to iterate in a single balance run.
289 * Limited because this is done with IRQs disabled.
291 const_debug unsigned int sysctl_sched_nr_migrate = 32;
294 * period over which we average the RT time consumption, measured
299 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
302 * period over which we measure -rt task cpu usage in us.
305 unsigned int sysctl_sched_rt_period = 1000000;
307 __read_mostly int scheduler_running;
310 * part of the period that we allow rt tasks to run in us.
313 int sysctl_sched_rt_runtime = 950000;
316 * __task_rq_lock - lock the rq @p resides on.
318 static inline struct rq *__task_rq_lock(struct task_struct *p)
323 lockdep_assert_held(&p->pi_lock);
327 raw_spin_lock(&rq->lock);
328 if (likely(rq == task_rq(p)))
330 raw_spin_unlock(&rq->lock);
335 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
337 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
338 __acquires(p->pi_lock)
344 raw_spin_lock_irqsave(&p->pi_lock, *flags);
346 raw_spin_lock(&rq->lock);
347 if (likely(rq == task_rq(p)))
349 raw_spin_unlock(&rq->lock);
350 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
354 static void __task_rq_unlock(struct rq *rq)
357 raw_spin_unlock(&rq->lock);
361 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
363 __releases(p->pi_lock)
365 raw_spin_unlock(&rq->lock);
366 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
370 * this_rq_lock - lock this runqueue and disable interrupts.
372 static struct rq *this_rq_lock(void)
379 raw_spin_lock(&rq->lock);
384 #ifdef CONFIG_SCHED_HRTICK
386 * Use HR-timers to deliver accurate preemption points.
389 static void hrtick_clear(struct rq *rq)
391 if (hrtimer_active(&rq->hrtick_timer))
392 hrtimer_cancel(&rq->hrtick_timer);
396 * High-resolution timer tick.
397 * Runs from hardirq context with interrupts disabled.
399 static enum hrtimer_restart hrtick(struct hrtimer *timer)
401 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
403 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
405 raw_spin_lock(&rq->lock);
407 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
408 raw_spin_unlock(&rq->lock);
410 return HRTIMER_NORESTART;
415 static int __hrtick_restart(struct rq *rq)
417 struct hrtimer *timer = &rq->hrtick_timer;
418 ktime_t time = hrtimer_get_softexpires(timer);
420 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
424 * called from hardirq (IPI) context
426 static void __hrtick_start(void *arg)
430 raw_spin_lock(&rq->lock);
431 __hrtick_restart(rq);
432 rq->hrtick_csd_pending = 0;
433 raw_spin_unlock(&rq->lock);
437 * Called to set the hrtick timer state.
439 * called with rq->lock held and irqs disabled
441 void hrtick_start(struct rq *rq, u64 delay)
443 struct hrtimer *timer = &rq->hrtick_timer;
444 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
446 hrtimer_set_expires(timer, time);
448 if (rq == this_rq()) {
449 __hrtick_restart(rq);
450 } else if (!rq->hrtick_csd_pending) {
451 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
452 rq->hrtick_csd_pending = 1;
457 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
459 int cpu = (int)(long)hcpu;
462 case CPU_UP_CANCELED:
463 case CPU_UP_CANCELED_FROZEN:
464 case CPU_DOWN_PREPARE:
465 case CPU_DOWN_PREPARE_FROZEN:
467 case CPU_DEAD_FROZEN:
468 hrtick_clear(cpu_rq(cpu));
475 static __init void init_hrtick(void)
477 hotcpu_notifier(hotplug_hrtick, 0);
481 * Called to set the hrtick timer state.
483 * called with rq->lock held and irqs disabled
485 void hrtick_start(struct rq *rq, u64 delay)
487 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
488 HRTIMER_MODE_REL_PINNED, 0);
491 static inline void init_hrtick(void)
494 #endif /* CONFIG_SMP */
496 static void init_rq_hrtick(struct rq *rq)
499 rq->hrtick_csd_pending = 0;
501 rq->hrtick_csd.flags = 0;
502 rq->hrtick_csd.func = __hrtick_start;
503 rq->hrtick_csd.info = rq;
506 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
507 rq->hrtick_timer.function = hrtick;
509 #else /* CONFIG_SCHED_HRTICK */
510 static inline void hrtick_clear(struct rq *rq)
514 static inline void init_rq_hrtick(struct rq *rq)
518 static inline void init_hrtick(void)
521 #endif /* CONFIG_SCHED_HRTICK */
524 * cmpxchg based fetch_or, macro so it works for different integer types
526 #define fetch_or(ptr, val) \
527 ({ typeof(*(ptr)) __old, __val = *(ptr); \
529 __old = cmpxchg((ptr), __val, __val | (val)); \
530 if (__old == __val) \
537 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
539 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
540 * this avoids any races wrt polling state changes and thereby avoids
543 static bool set_nr_and_not_polling(struct task_struct *p)
545 struct thread_info *ti = task_thread_info(p);
546 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
550 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
552 * If this returns true, then the idle task promises to call
553 * sched_ttwu_pending() and reschedule soon.
555 static bool set_nr_if_polling(struct task_struct *p)
557 struct thread_info *ti = task_thread_info(p);
558 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
561 if (!(val & _TIF_POLLING_NRFLAG))
563 if (val & _TIF_NEED_RESCHED)
565 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
574 static bool set_nr_and_not_polling(struct task_struct *p)
576 set_tsk_need_resched(p);
581 static bool set_nr_if_polling(struct task_struct *p)
589 * resched_task - mark a task 'to be rescheduled now'.
591 * On UP this means the setting of the need_resched flag, on SMP it
592 * might also involve a cross-CPU call to trigger the scheduler on
595 void resched_task(struct task_struct *p)
599 lockdep_assert_held(&task_rq(p)->lock);
601 if (test_tsk_need_resched(p))
606 if (cpu == smp_processor_id()) {
607 set_tsk_need_resched(p);
608 set_preempt_need_resched();
612 if (set_nr_and_not_polling(p))
613 smp_send_reschedule(cpu);
615 trace_sched_wake_idle_without_ipi(cpu);
618 void resched_cpu(int cpu)
620 struct rq *rq = cpu_rq(cpu);
623 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
625 resched_task(cpu_curr(cpu));
626 raw_spin_unlock_irqrestore(&rq->lock, flags);
630 #ifdef CONFIG_NO_HZ_COMMON
632 * In the semi idle case, use the nearest busy cpu for migrating timers
633 * from an idle cpu. This is good for power-savings.
635 * We don't do similar optimization for completely idle system, as
636 * selecting an idle cpu will add more delays to the timers than intended
637 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
639 int get_nohz_timer_target(int pinned)
641 int cpu = smp_processor_id();
643 struct sched_domain *sd;
645 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
649 for_each_domain(cpu, sd) {
650 for_each_cpu(i, sched_domain_span(sd)) {
662 * When add_timer_on() enqueues a timer into the timer wheel of an
663 * idle CPU then this timer might expire before the next timer event
664 * which is scheduled to wake up that CPU. In case of a completely
665 * idle system the next event might even be infinite time into the
666 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
667 * leaves the inner idle loop so the newly added timer is taken into
668 * account when the CPU goes back to idle and evaluates the timer
669 * wheel for the next timer event.
671 static void wake_up_idle_cpu(int cpu)
673 struct rq *rq = cpu_rq(cpu);
675 if (cpu == smp_processor_id())
678 if (set_nr_and_not_polling(rq->idle))
679 smp_send_reschedule(cpu);
681 trace_sched_wake_idle_without_ipi(cpu);
684 static bool wake_up_full_nohz_cpu(int cpu)
686 if (tick_nohz_full_cpu(cpu)) {
687 if (cpu != smp_processor_id() ||
688 tick_nohz_tick_stopped())
689 smp_send_reschedule(cpu);
696 void wake_up_nohz_cpu(int cpu)
698 if (!wake_up_full_nohz_cpu(cpu))
699 wake_up_idle_cpu(cpu);
702 static inline bool got_nohz_idle_kick(void)
704 int cpu = smp_processor_id();
706 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
709 if (idle_cpu(cpu) && !need_resched())
713 * We can't run Idle Load Balance on this CPU for this time so we
714 * cancel it and clear NOHZ_BALANCE_KICK
716 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
720 #else /* CONFIG_NO_HZ_COMMON */
722 static inline bool got_nohz_idle_kick(void)
727 #endif /* CONFIG_NO_HZ_COMMON */
729 #ifdef CONFIG_NO_HZ_FULL
730 bool sched_can_stop_tick(void)
736 /* Make sure rq->nr_running update is visible after the IPI */
739 /* More than one running task need preemption */
740 if (rq->nr_running > 1)
745 #endif /* CONFIG_NO_HZ_FULL */
747 void sched_avg_update(struct rq *rq)
749 s64 period = sched_avg_period();
751 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
753 * Inline assembly required to prevent the compiler
754 * optimising this loop into a divmod call.
755 * See __iter_div_u64_rem() for another example of this.
757 asm("" : "+rm" (rq->age_stamp));
758 rq->age_stamp += period;
763 #endif /* CONFIG_SMP */
765 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
766 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
768 * Iterate task_group tree rooted at *from, calling @down when first entering a
769 * node and @up when leaving it for the final time.
771 * Caller must hold rcu_lock or sufficient equivalent.
773 int walk_tg_tree_from(struct task_group *from,
774 tg_visitor down, tg_visitor up, void *data)
776 struct task_group *parent, *child;
782 ret = (*down)(parent, data);
785 list_for_each_entry_rcu(child, &parent->children, siblings) {
792 ret = (*up)(parent, data);
793 if (ret || parent == from)
797 parent = parent->parent;
804 int tg_nop(struct task_group *tg, void *data)
810 static void set_load_weight(struct task_struct *p)
812 int prio = p->static_prio - MAX_RT_PRIO;
813 struct load_weight *load = &p->se.load;
816 * SCHED_IDLE tasks get minimal weight:
818 if (p->policy == SCHED_IDLE) {
819 load->weight = scale_load(WEIGHT_IDLEPRIO);
820 load->inv_weight = WMULT_IDLEPRIO;
824 prio = array_index_nospec(prio, 40);
826 load->weight = scale_load(prio_to_weight[prio]);
827 load->inv_weight = prio_to_wmult[prio];
830 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
833 sched_info_queued(rq, p);
834 p->sched_class->enqueue_task(rq, p, flags);
837 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
840 sched_info_dequeued(rq, p);
841 p->sched_class->dequeue_task(rq, p, flags);
844 void activate_task(struct rq *rq, struct task_struct *p, int flags)
846 if (task_contributes_to_load(p))
847 rq->nr_uninterruptible--;
849 enqueue_task(rq, p, flags);
852 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
854 if (task_contributes_to_load(p))
855 rq->nr_uninterruptible++;
857 dequeue_task(rq, p, flags);
860 static void update_rq_clock_task(struct rq *rq, s64 delta)
863 * In theory, the compile should just see 0 here, and optimize out the call
864 * to sched_rt_avg_update. But I don't trust it...
866 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
867 s64 steal = 0, irq_delta = 0;
869 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
870 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
873 * Since irq_time is only updated on {soft,}irq_exit, we might run into
874 * this case when a previous update_rq_clock() happened inside a
877 * When this happens, we stop ->clock_task and only update the
878 * prev_irq_time stamp to account for the part that fit, so that a next
879 * update will consume the rest. This ensures ->clock_task is
882 * It does however cause some slight miss-attribution of {soft,}irq
883 * time, a more accurate solution would be to update the irq_time using
884 * the current rq->clock timestamp, except that would require using
887 if (irq_delta > delta)
890 rq->prev_irq_time += irq_delta;
893 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
894 if (static_key_false((¶virt_steal_rq_enabled))) {
895 steal = paravirt_steal_clock(cpu_of(rq));
896 steal -= rq->prev_steal_time_rq;
898 if (unlikely(steal > delta))
901 rq->prev_steal_time_rq += steal;
906 rq->clock_task += delta;
908 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
909 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
910 sched_rt_avg_update(rq, irq_delta + steal);
914 void sched_set_stop_task(int cpu, struct task_struct *stop)
916 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
917 struct task_struct *old_stop = cpu_rq(cpu)->stop;
921 * Make it appear like a SCHED_FIFO task, its something
922 * userspace knows about and won't get confused about.
924 * Also, it will make PI more or less work without too
925 * much confusion -- but then, stop work should not
926 * rely on PI working anyway.
928 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
930 stop->sched_class = &stop_sched_class;
933 cpu_rq(cpu)->stop = stop;
937 * Reset it back to a normal scheduling class so that
938 * it can die in pieces.
940 old_stop->sched_class = &rt_sched_class;
945 * __normal_prio - return the priority that is based on the static prio
947 static inline int __normal_prio(struct task_struct *p)
949 return p->static_prio;
953 * Calculate the expected normal priority: i.e. priority
954 * without taking RT-inheritance into account. Might be
955 * boosted by interactivity modifiers. Changes upon fork,
956 * setprio syscalls, and whenever the interactivity
957 * estimator recalculates.
959 static inline int normal_prio(struct task_struct *p)
963 if (task_has_dl_policy(p))
964 prio = MAX_DL_PRIO-1;
965 else if (task_has_rt_policy(p))
966 prio = MAX_RT_PRIO-1 - p->rt_priority;
968 prio = __normal_prio(p);
973 * Calculate the current priority, i.e. the priority
974 * taken into account by the scheduler. This value might
975 * be boosted by RT tasks, or might be boosted by
976 * interactivity modifiers. Will be RT if the task got
977 * RT-boosted. If not then it returns p->normal_prio.
979 static int effective_prio(struct task_struct *p)
981 p->normal_prio = normal_prio(p);
983 * If we are RT tasks or we were boosted to RT priority,
984 * keep the priority unchanged. Otherwise, update priority
985 * to the normal priority:
987 if (!rt_prio(p->prio))
988 return p->normal_prio;
993 * task_curr - is this task currently executing on a CPU?
994 * @p: the task in question.
996 * Return: 1 if the task is currently executing. 0 otherwise.
998 inline int task_curr(const struct task_struct *p)
1000 return cpu_curr(task_cpu(p)) == p;
1004 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1005 * use the balance_callback list if you want balancing.
1007 * this means any call to check_class_changed() must be followed by a call to
1008 * balance_callback().
1010 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1011 const struct sched_class *prev_class,
1014 if (prev_class != p->sched_class) {
1015 if (prev_class->switched_from)
1016 prev_class->switched_from(rq, p);
1017 p->sched_class->switched_to(rq, p);
1018 } else if (oldprio != p->prio || dl_task(p))
1019 p->sched_class->prio_changed(rq, p, oldprio);
1022 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1024 const struct sched_class *class;
1026 if (p->sched_class == rq->curr->sched_class) {
1027 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1029 for_each_class(class) {
1030 if (class == rq->curr->sched_class)
1032 if (class == p->sched_class) {
1033 resched_task(rq->curr);
1040 * A queue event has occurred, and we're going to schedule. In
1041 * this case, we can save a useless back to back clock update.
1043 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1044 rq->skip_clock_update = 1;
1048 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1050 #ifdef CONFIG_SCHED_DEBUG
1052 * We should never call set_task_cpu() on a blocked task,
1053 * ttwu() will sort out the placement.
1055 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1056 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1058 #ifdef CONFIG_LOCKDEP
1060 * The caller should hold either p->pi_lock or rq->lock, when changing
1061 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1063 * sched_move_task() holds both and thus holding either pins the cgroup,
1066 * Furthermore, all task_rq users should acquire both locks, see
1069 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1070 lockdep_is_held(&task_rq(p)->lock)));
1074 trace_sched_migrate_task(p, new_cpu);
1076 if (task_cpu(p) != new_cpu) {
1077 if (p->sched_class->migrate_task_rq)
1078 p->sched_class->migrate_task_rq(p, new_cpu);
1079 p->se.nr_migrations++;
1080 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1083 __set_task_cpu(p, new_cpu);
1086 static void __migrate_swap_task(struct task_struct *p, int cpu)
1089 struct rq *src_rq, *dst_rq;
1091 src_rq = task_rq(p);
1092 dst_rq = cpu_rq(cpu);
1094 deactivate_task(src_rq, p, 0);
1095 set_task_cpu(p, cpu);
1096 activate_task(dst_rq, p, 0);
1097 check_preempt_curr(dst_rq, p, 0);
1100 * Task isn't running anymore; make it appear like we migrated
1101 * it before it went to sleep. This means on wakeup we make the
1102 * previous cpu our targer instead of where it really is.
1108 struct migration_swap_arg {
1109 struct task_struct *src_task, *dst_task;
1110 int src_cpu, dst_cpu;
1113 static int migrate_swap_stop(void *data)
1115 struct migration_swap_arg *arg = data;
1116 struct rq *src_rq, *dst_rq;
1119 src_rq = cpu_rq(arg->src_cpu);
1120 dst_rq = cpu_rq(arg->dst_cpu);
1122 double_raw_lock(&arg->src_task->pi_lock,
1123 &arg->dst_task->pi_lock);
1124 double_rq_lock(src_rq, dst_rq);
1125 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1128 if (task_cpu(arg->src_task) != arg->src_cpu)
1131 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1134 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1137 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1138 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1143 double_rq_unlock(src_rq, dst_rq);
1144 raw_spin_unlock(&arg->dst_task->pi_lock);
1145 raw_spin_unlock(&arg->src_task->pi_lock);
1151 * Cross migrate two tasks
1153 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1155 struct migration_swap_arg arg;
1158 arg = (struct migration_swap_arg){
1160 .src_cpu = task_cpu(cur),
1162 .dst_cpu = task_cpu(p),
1165 if (arg.src_cpu == arg.dst_cpu)
1169 * These three tests are all lockless; this is OK since all of them
1170 * will be re-checked with proper locks held further down the line.
1172 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1175 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1178 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1181 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1182 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1188 struct migration_arg {
1189 struct task_struct *task;
1193 static int migration_cpu_stop(void *data);
1196 * wait_task_inactive - wait for a thread to unschedule.
1198 * If @match_state is nonzero, it's the @p->state value just checked and
1199 * not expected to change. If it changes, i.e. @p might have woken up,
1200 * then return zero. When we succeed in waiting for @p to be off its CPU,
1201 * we return a positive number (its total switch count). If a second call
1202 * a short while later returns the same number, the caller can be sure that
1203 * @p has remained unscheduled the whole time.
1205 * The caller must ensure that the task *will* unschedule sometime soon,
1206 * else this function might spin for a *long* time. This function can't
1207 * be called with interrupts off, or it may introduce deadlock with
1208 * smp_call_function() if an IPI is sent by the same process we are
1209 * waiting to become inactive.
1211 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1213 unsigned long flags;
1220 * We do the initial early heuristics without holding
1221 * any task-queue locks at all. We'll only try to get
1222 * the runqueue lock when things look like they will
1228 * If the task is actively running on another CPU
1229 * still, just relax and busy-wait without holding
1232 * NOTE! Since we don't hold any locks, it's not
1233 * even sure that "rq" stays as the right runqueue!
1234 * But we don't care, since "task_running()" will
1235 * return false if the runqueue has changed and p
1236 * is actually now running somewhere else!
1238 while (task_running(rq, p)) {
1239 if (match_state && unlikely(p->state != match_state))
1245 * Ok, time to look more closely! We need the rq
1246 * lock now, to be *sure*. If we're wrong, we'll
1247 * just go back and repeat.
1249 rq = task_rq_lock(p, &flags);
1250 trace_sched_wait_task(p);
1251 running = task_running(rq, p);
1254 if (!match_state || p->state == match_state)
1255 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1256 task_rq_unlock(rq, p, &flags);
1259 * If it changed from the expected state, bail out now.
1261 if (unlikely(!ncsw))
1265 * Was it really running after all now that we
1266 * checked with the proper locks actually held?
1268 * Oops. Go back and try again..
1270 if (unlikely(running)) {
1276 * It's not enough that it's not actively running,
1277 * it must be off the runqueue _entirely_, and not
1280 * So if it was still runnable (but just not actively
1281 * running right now), it's preempted, and we should
1282 * yield - it could be a while.
1284 if (unlikely(on_rq)) {
1285 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1287 set_current_state(TASK_UNINTERRUPTIBLE);
1288 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1293 * Ahh, all good. It wasn't running, and it wasn't
1294 * runnable, which means that it will never become
1295 * running in the future either. We're all done!
1304 * kick_process - kick a running thread to enter/exit the kernel
1305 * @p: the to-be-kicked thread
1307 * Cause a process which is running on another CPU to enter
1308 * kernel-mode, without any delay. (to get signals handled.)
1310 * NOTE: this function doesn't have to take the runqueue lock,
1311 * because all it wants to ensure is that the remote task enters
1312 * the kernel. If the IPI races and the task has been migrated
1313 * to another CPU then no harm is done and the purpose has been
1316 void kick_process(struct task_struct *p)
1322 if ((cpu != smp_processor_id()) && task_curr(p))
1323 smp_send_reschedule(cpu);
1326 EXPORT_SYMBOL_GPL(kick_process);
1327 #endif /* CONFIG_SMP */
1331 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1333 static int select_fallback_rq(int cpu, struct task_struct *p)
1335 int nid = cpu_to_node(cpu);
1336 const struct cpumask *nodemask = NULL;
1337 enum { cpuset, possible, fail } state = cpuset;
1341 * If the node that the cpu is on has been offlined, cpu_to_node()
1342 * will return -1. There is no cpu on the node, and we should
1343 * select the cpu on the other node.
1346 nodemask = cpumask_of_node(nid);
1348 /* Look for allowed, online CPU in same node. */
1349 for_each_cpu(dest_cpu, nodemask) {
1350 if (!cpu_online(dest_cpu))
1352 if (!cpu_active(dest_cpu))
1354 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1360 /* Any allowed, online CPU? */
1361 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1362 if (!cpu_online(dest_cpu))
1364 if (!cpu_active(dest_cpu))
1371 /* No more Mr. Nice Guy. */
1372 cpuset_cpus_allowed_fallback(p);
1377 do_set_cpus_allowed(p, cpu_possible_mask);
1388 if (state != cpuset) {
1390 * Don't tell them about moving exiting tasks or
1391 * kernel threads (both mm NULL), since they never
1394 if (p->mm && printk_ratelimit()) {
1395 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1396 task_pid_nr(p), p->comm, cpu);
1404 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1407 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1409 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1412 * In order not to call set_task_cpu() on a blocking task we need
1413 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1416 * Since this is common to all placement strategies, this lives here.
1418 * [ this allows ->select_task() to simply return task_cpu(p) and
1419 * not worry about this generic constraint ]
1421 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1423 cpu = select_fallback_rq(task_cpu(p), p);
1428 static void update_avg(u64 *avg, u64 sample)
1430 s64 diff = sample - *avg;
1436 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1438 #ifdef CONFIG_SCHEDSTATS
1439 struct rq *rq = this_rq();
1442 int this_cpu = smp_processor_id();
1444 if (cpu == this_cpu) {
1445 schedstat_inc(rq, ttwu_local);
1446 schedstat_inc(p, se.statistics.nr_wakeups_local);
1448 struct sched_domain *sd;
1450 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1452 for_each_domain(this_cpu, sd) {
1453 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1454 schedstat_inc(sd, ttwu_wake_remote);
1461 if (wake_flags & WF_MIGRATED)
1462 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1464 #endif /* CONFIG_SMP */
1466 schedstat_inc(rq, ttwu_count);
1467 schedstat_inc(p, se.statistics.nr_wakeups);
1469 if (wake_flags & WF_SYNC)
1470 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1472 #endif /* CONFIG_SCHEDSTATS */
1475 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1477 activate_task(rq, p, en_flags);
1480 /* if a worker is waking up, notify workqueue */
1481 if (p->flags & PF_WQ_WORKER)
1482 wq_worker_waking_up(p, cpu_of(rq));
1486 * Mark the task runnable and perform wakeup-preemption.
1489 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1491 check_preempt_curr(rq, p, wake_flags);
1492 trace_sched_wakeup(p, true);
1494 p->state = TASK_RUNNING;
1496 if (p->sched_class->task_woken) {
1498 * XXX can drop rq->lock; most likely ok.
1500 p->sched_class->task_woken(rq, p);
1503 if (rq->idle_stamp) {
1504 u64 delta = rq_clock(rq) - rq->idle_stamp;
1505 u64 max = 2*rq->max_idle_balance_cost;
1507 update_avg(&rq->avg_idle, delta);
1509 if (rq->avg_idle > max)
1518 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1521 if (p->sched_contributes_to_load)
1522 rq->nr_uninterruptible--;
1525 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1526 ttwu_do_wakeup(rq, p, wake_flags);
1530 * Called in case the task @p isn't fully descheduled from its runqueue,
1531 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1532 * since all we need to do is flip p->state to TASK_RUNNING, since
1533 * the task is still ->on_rq.
1535 static int ttwu_remote(struct task_struct *p, int wake_flags)
1540 rq = __task_rq_lock(p);
1542 /* check_preempt_curr() may use rq clock */
1543 update_rq_clock(rq);
1544 ttwu_do_wakeup(rq, p, wake_flags);
1547 __task_rq_unlock(rq);
1553 void sched_ttwu_pending(void)
1555 struct rq *rq = this_rq();
1556 struct llist_node *llist = llist_del_all(&rq->wake_list);
1557 struct task_struct *p;
1558 unsigned long flags;
1563 raw_spin_lock_irqsave(&rq->lock, flags);
1566 p = llist_entry(llist, struct task_struct, wake_entry);
1567 llist = llist_next(llist);
1568 ttwu_do_activate(rq, p, 0);
1571 raw_spin_unlock_irqrestore(&rq->lock, flags);
1574 void scheduler_ipi(void)
1577 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1578 * TIF_NEED_RESCHED remotely (for the first time) will also send
1581 preempt_fold_need_resched();
1583 if (llist_empty(&this_rq()->wake_list)
1584 && !tick_nohz_full_cpu(smp_processor_id())
1585 && !got_nohz_idle_kick())
1589 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1590 * traditionally all their work was done from the interrupt return
1591 * path. Now that we actually do some work, we need to make sure
1594 * Some archs already do call them, luckily irq_enter/exit nest
1597 * Arguably we should visit all archs and update all handlers,
1598 * however a fair share of IPIs are still resched only so this would
1599 * somewhat pessimize the simple resched case.
1602 tick_nohz_full_check();
1603 sched_ttwu_pending();
1606 * Check if someone kicked us for doing the nohz idle load balance.
1608 if (unlikely(got_nohz_idle_kick())) {
1609 this_rq()->idle_balance = 1;
1610 raise_softirq_irqoff(SCHED_SOFTIRQ);
1615 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1617 struct rq *rq = cpu_rq(cpu);
1619 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1620 if (!set_nr_if_polling(rq->idle))
1621 smp_send_reschedule(cpu);
1623 trace_sched_wake_idle_without_ipi(cpu);
1627 bool cpus_share_cache(int this_cpu, int that_cpu)
1629 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1631 #endif /* CONFIG_SMP */
1633 static void ttwu_queue(struct task_struct *p, int cpu)
1635 struct rq *rq = cpu_rq(cpu);
1637 #if defined(CONFIG_SMP)
1638 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1639 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1640 ttwu_queue_remote(p, cpu);
1645 raw_spin_lock(&rq->lock);
1646 ttwu_do_activate(rq, p, 0);
1647 raw_spin_unlock(&rq->lock);
1651 * try_to_wake_up - wake up a thread
1652 * @p: the thread to be awakened
1653 * @state: the mask of task states that can be woken
1654 * @wake_flags: wake modifier flags (WF_*)
1656 * Put it on the run-queue if it's not already there. The "current"
1657 * thread is always on the run-queue (except when the actual
1658 * re-schedule is in progress), and as such you're allowed to do
1659 * the simpler "current->state = TASK_RUNNING" to mark yourself
1660 * runnable without the overhead of this.
1662 * Return: %true if @p was woken up, %false if it was already running.
1663 * or @state didn't match @p's state.
1666 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1668 unsigned long flags;
1669 int cpu, success = 0;
1672 * If we are going to wake up a thread waiting for CONDITION we
1673 * need to ensure that CONDITION=1 done by the caller can not be
1674 * reordered with p->state check below. This pairs with mb() in
1675 * set_current_state() the waiting thread does.
1677 smp_mb__before_spinlock();
1678 raw_spin_lock_irqsave(&p->pi_lock, flags);
1679 if (!(p->state & state))
1682 success = 1; /* we're going to change ->state */
1686 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1687 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1688 * in smp_cond_load_acquire() below.
1690 * sched_ttwu_pending() try_to_wake_up()
1691 * [S] p->on_rq = 1; [L] P->state
1692 * UNLOCK rq->lock -----.
1696 * LOCK rq->lock -----'
1700 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1702 * Pairs with the UNLOCK+LOCK on rq->lock from the
1703 * last wakeup of our task and the schedule that got our task
1707 if (p->on_rq && ttwu_remote(p, wake_flags))
1712 * If the owning (remote) cpu is still in the middle of schedule() with
1713 * this task as prev, wait until its done referencing the task.
1718 * Pairs with the smp_wmb() in finish_lock_switch().
1722 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1723 p->state = TASK_WAKING;
1725 if (p->sched_class->task_waking)
1726 p->sched_class->task_waking(p);
1728 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1729 if (task_cpu(p) != cpu) {
1730 wake_flags |= WF_MIGRATED;
1731 set_task_cpu(p, cpu);
1733 #endif /* CONFIG_SMP */
1737 ttwu_stat(p, cpu, wake_flags);
1739 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1745 * try_to_wake_up_local - try to wake up a local task with rq lock held
1746 * @p: the thread to be awakened
1748 * Put @p on the run-queue if it's not already there. The caller must
1749 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1752 static void try_to_wake_up_local(struct task_struct *p)
1754 struct rq *rq = task_rq(p);
1756 if (WARN_ON_ONCE(rq != this_rq()) ||
1757 WARN_ON_ONCE(p == current))
1760 lockdep_assert_held(&rq->lock);
1762 if (!raw_spin_trylock(&p->pi_lock)) {
1763 raw_spin_unlock(&rq->lock);
1764 raw_spin_lock(&p->pi_lock);
1765 raw_spin_lock(&rq->lock);
1768 if (!(p->state & TASK_NORMAL))
1772 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1774 ttwu_do_wakeup(rq, p, 0);
1775 ttwu_stat(p, smp_processor_id(), 0);
1777 raw_spin_unlock(&p->pi_lock);
1781 * wake_up_process - Wake up a specific process
1782 * @p: The process to be woken up.
1784 * Attempt to wake up the nominated process and move it to the set of runnable
1787 * Return: 1 if the process was woken up, 0 if it was already running.
1789 * It may be assumed that this function implies a write memory barrier before
1790 * changing the task state if and only if any tasks are woken up.
1792 int wake_up_process(struct task_struct *p)
1794 return try_to_wake_up(p, TASK_NORMAL, 0);
1796 EXPORT_SYMBOL(wake_up_process);
1798 int wake_up_state(struct task_struct *p, unsigned int state)
1800 return try_to_wake_up(p, state, 0);
1804 * Perform scheduler related setup for a newly forked process p.
1805 * p is forked by current.
1807 * __sched_fork() is basic setup used by init_idle() too:
1809 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1814 p->se.exec_start = 0;
1815 p->se.sum_exec_runtime = 0;
1816 p->se.prev_sum_exec_runtime = 0;
1817 p->se.nr_migrations = 0;
1819 INIT_LIST_HEAD(&p->se.group_node);
1821 #ifdef CONFIG_SCHEDSTATS
1822 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1825 RB_CLEAR_NODE(&p->dl.rb_node);
1826 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1827 p->dl.dl_runtime = p->dl.runtime = 0;
1828 p->dl.dl_deadline = p->dl.deadline = 0;
1829 p->dl.dl_period = 0;
1832 INIT_LIST_HEAD(&p->rt.run_list);
1834 #ifdef CONFIG_PREEMPT_NOTIFIERS
1835 INIT_HLIST_HEAD(&p->preempt_notifiers);
1838 #ifdef CONFIG_NUMA_BALANCING
1839 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1840 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1841 p->mm->numa_scan_seq = 0;
1844 if (clone_flags & CLONE_VM)
1845 p->numa_preferred_nid = current->numa_preferred_nid;
1847 p->numa_preferred_nid = -1;
1849 p->node_stamp = 0ULL;
1850 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1851 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1852 p->numa_work.next = &p->numa_work;
1853 p->numa_faults_memory = NULL;
1854 p->numa_faults_buffer_memory = NULL;
1855 p->last_task_numa_placement = 0;
1856 p->last_sum_exec_runtime = 0;
1858 INIT_LIST_HEAD(&p->numa_entry);
1859 p->numa_group = NULL;
1860 #endif /* CONFIG_NUMA_BALANCING */
1863 #ifdef CONFIG_NUMA_BALANCING
1864 #ifdef CONFIG_SCHED_DEBUG
1865 void set_numabalancing_state(bool enabled)
1868 sched_feat_set("NUMA");
1870 sched_feat_set("NO_NUMA");
1873 __read_mostly bool numabalancing_enabled;
1875 void set_numabalancing_state(bool enabled)
1877 numabalancing_enabled = enabled;
1879 #endif /* CONFIG_SCHED_DEBUG */
1881 #ifdef CONFIG_PROC_SYSCTL
1882 int sysctl_numa_balancing(struct ctl_table *table, int write,
1883 void __user *buffer, size_t *lenp, loff_t *ppos)
1887 int state = numabalancing_enabled;
1889 if (write && !capable(CAP_SYS_ADMIN))
1894 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1898 set_numabalancing_state(state);
1905 * fork()/clone()-time setup:
1907 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1909 unsigned long flags;
1910 int cpu = get_cpu();
1912 __sched_fork(clone_flags, p);
1914 * We mark the process as running here. This guarantees that
1915 * nobody will actually run it, and a signal or other external
1916 * event cannot wake it up and insert it on the runqueue either.
1918 p->state = TASK_RUNNING;
1921 * Make sure we do not leak PI boosting priority to the child.
1923 p->prio = current->normal_prio;
1926 * Revert to default priority/policy on fork if requested.
1928 if (unlikely(p->sched_reset_on_fork)) {
1929 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1930 p->policy = SCHED_NORMAL;
1931 p->static_prio = NICE_TO_PRIO(0);
1933 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1934 p->static_prio = NICE_TO_PRIO(0);
1936 p->prio = p->normal_prio = __normal_prio(p);
1940 * We don't need the reset flag anymore after the fork. It has
1941 * fulfilled its duty:
1943 p->sched_reset_on_fork = 0;
1946 if (dl_prio(p->prio)) {
1949 } else if (rt_prio(p->prio)) {
1950 p->sched_class = &rt_sched_class;
1952 p->sched_class = &fair_sched_class;
1955 if (p->sched_class->task_fork)
1956 p->sched_class->task_fork(p);
1959 * The child is not yet in the pid-hash so no cgroup attach races,
1960 * and the cgroup is pinned to this child due to cgroup_fork()
1961 * is ran before sched_fork().
1963 * Silence PROVE_RCU.
1965 raw_spin_lock_irqsave(&p->pi_lock, flags);
1966 set_task_cpu(p, cpu);
1967 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1969 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1970 if (likely(sched_info_on()))
1971 memset(&p->sched_info, 0, sizeof(p->sched_info));
1973 #if defined(CONFIG_SMP)
1976 init_task_preempt_count(p);
1978 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1979 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1986 unsigned long to_ratio(u64 period, u64 runtime)
1988 if (runtime == RUNTIME_INF)
1992 * Doing this here saves a lot of checks in all
1993 * the calling paths, and returning zero seems
1994 * safe for them anyway.
1999 return div64_u64(runtime << 20, period);
2003 inline struct dl_bw *dl_bw_of(int i)
2005 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2006 "sched RCU must be held");
2007 return &cpu_rq(i)->rd->dl_bw;
2010 static inline int dl_bw_cpus(int i)
2012 struct root_domain *rd = cpu_rq(i)->rd;
2015 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2016 "sched RCU must be held");
2017 for_each_cpu_and(i, rd->span, cpu_active_mask)
2023 inline struct dl_bw *dl_bw_of(int i)
2025 return &cpu_rq(i)->dl.dl_bw;
2028 static inline int dl_bw_cpus(int i)
2035 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2037 dl_b->total_bw -= tsk_bw;
2041 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2043 dl_b->total_bw += tsk_bw;
2047 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2049 return dl_b->bw != -1 &&
2050 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2054 * We must be sure that accepting a new task (or allowing changing the
2055 * parameters of an existing one) is consistent with the bandwidth
2056 * constraints. If yes, this function also accordingly updates the currently
2057 * allocated bandwidth to reflect the new situation.
2059 * This function is called while holding p's rq->lock.
2061 static int dl_overflow(struct task_struct *p, int policy,
2062 const struct sched_attr *attr)
2065 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2066 u64 period = attr->sched_period ?: attr->sched_deadline;
2067 u64 runtime = attr->sched_runtime;
2068 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2071 if (new_bw == p->dl.dl_bw)
2075 * Either if a task, enters, leave, or stays -deadline but changes
2076 * its parameters, we may need to update accordingly the total
2077 * allocated bandwidth of the container.
2079 raw_spin_lock(&dl_b->lock);
2080 cpus = dl_bw_cpus(task_cpu(p));
2081 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2082 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2083 __dl_add(dl_b, new_bw);
2085 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2086 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2087 __dl_clear(dl_b, p->dl.dl_bw);
2088 __dl_add(dl_b, new_bw);
2090 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2091 __dl_clear(dl_b, p->dl.dl_bw);
2094 raw_spin_unlock(&dl_b->lock);
2099 extern void init_dl_bw(struct dl_bw *dl_b);
2102 * wake_up_new_task - wake up a newly created task for the first time.
2104 * This function will do some initial scheduler statistics housekeeping
2105 * that must be done for every newly created context, then puts the task
2106 * on the runqueue and wakes it.
2108 void wake_up_new_task(struct task_struct *p)
2110 unsigned long flags;
2113 raw_spin_lock_irqsave(&p->pi_lock, flags);
2116 * Fork balancing, do it here and not earlier because:
2117 * - cpus_allowed can change in the fork path
2118 * - any previously selected cpu might disappear through hotplug
2120 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2123 /* Initialize new task's runnable average */
2124 init_task_runnable_average(p);
2125 rq = __task_rq_lock(p);
2126 activate_task(rq, p, 0);
2128 trace_sched_wakeup_new(p, true);
2129 check_preempt_curr(rq, p, WF_FORK);
2131 if (p->sched_class->task_woken)
2132 p->sched_class->task_woken(rq, p);
2134 task_rq_unlock(rq, p, &flags);
2137 #ifdef CONFIG_PREEMPT_NOTIFIERS
2140 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2141 * @notifier: notifier struct to register
2143 void preempt_notifier_register(struct preempt_notifier *notifier)
2145 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2147 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2150 * preempt_notifier_unregister - no longer interested in preemption notifications
2151 * @notifier: notifier struct to unregister
2153 * This is safe to call from within a preemption notifier.
2155 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2157 hlist_del(¬ifier->link);
2159 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2161 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2163 struct preempt_notifier *notifier;
2165 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2166 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2170 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2171 struct task_struct *next)
2173 struct preempt_notifier *notifier;
2175 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2176 notifier->ops->sched_out(notifier, next);
2179 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2181 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2186 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2187 struct task_struct *next)
2191 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2194 * prepare_task_switch - prepare to switch tasks
2195 * @rq: the runqueue preparing to switch
2196 * @prev: the current task that is being switched out
2197 * @next: the task we are going to switch to.
2199 * This is called with the rq lock held and interrupts off. It must
2200 * be paired with a subsequent finish_task_switch after the context
2203 * prepare_task_switch sets up locking and calls architecture specific
2207 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2208 struct task_struct *next)
2210 trace_sched_switch(prev, next);
2211 sched_info_switch(rq, prev, next);
2212 perf_event_task_sched_out(prev, next);
2213 fire_sched_out_preempt_notifiers(prev, next);
2214 prepare_lock_switch(rq, next);
2215 prepare_arch_switch(next);
2219 * finish_task_switch - clean up after a task-switch
2220 * @rq: runqueue associated with task-switch
2221 * @prev: the thread we just switched away from.
2223 * finish_task_switch must be called after the context switch, paired
2224 * with a prepare_task_switch call before the context switch.
2225 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2226 * and do any other architecture-specific cleanup actions.
2228 * Note that we may have delayed dropping an mm in context_switch(). If
2229 * so, we finish that here outside of the runqueue lock. (Doing it
2230 * with the lock held can cause deadlocks; see schedule() for
2233 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2234 __releases(rq->lock)
2236 struct mm_struct *mm = rq->prev_mm;
2242 * A task struct has one reference for the use as "current".
2243 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2244 * schedule one last time. The schedule call will never return, and
2245 * the scheduled task must drop that reference.
2247 * We must observe prev->state before clearing prev->on_cpu (in
2248 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2249 * running on another CPU and we could rave with its RUNNING -> DEAD
2250 * transition, resulting in a double drop.
2252 prev_state = prev->state;
2253 vtime_task_switch(prev);
2254 finish_arch_switch(prev);
2255 perf_event_task_sched_in(prev, current);
2256 finish_lock_switch(rq, prev);
2257 finish_arch_post_lock_switch();
2259 fire_sched_in_preempt_notifiers(current);
2262 if (unlikely(prev_state == TASK_DEAD)) {
2263 if (prev->sched_class->task_dead)
2264 prev->sched_class->task_dead(prev);
2267 * Remove function-return probe instances associated with this
2268 * task and put them back on the free list.
2270 kprobe_flush_task(prev);
2271 put_task_struct(prev);
2274 tick_nohz_task_switch(current);
2279 /* rq->lock is NOT held, but preemption is disabled */
2280 static void __balance_callback(struct rq *rq)
2282 struct callback_head *head, *next;
2283 void (*func)(struct rq *rq);
2284 unsigned long flags;
2286 raw_spin_lock_irqsave(&rq->lock, flags);
2287 head = rq->balance_callback;
2288 rq->balance_callback = NULL;
2290 func = (void (*)(struct rq *))head->func;
2297 raw_spin_unlock_irqrestore(&rq->lock, flags);
2300 static inline void balance_callback(struct rq *rq)
2302 if (unlikely(rq->balance_callback))
2303 __balance_callback(rq);
2308 static inline void balance_callback(struct rq *rq)
2315 * schedule_tail - first thing a freshly forked thread must call.
2316 * @prev: the thread we just switched away from.
2318 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2319 __releases(rq->lock)
2321 struct rq *rq = this_rq();
2323 finish_task_switch(rq, prev);
2326 * FIXME: do we need to worry about rq being invalidated by the
2329 balance_callback(rq);
2331 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2332 /* In this case, finish_task_switch does not reenable preemption */
2335 if (current->set_child_tid)
2336 put_user(task_pid_vnr(current), current->set_child_tid);
2340 * context_switch - switch to the new MM and the new
2341 * thread's register state.
2344 context_switch(struct rq *rq, struct task_struct *prev,
2345 struct task_struct *next)
2347 struct mm_struct *mm, *oldmm;
2349 prepare_task_switch(rq, prev, next);
2352 oldmm = prev->active_mm;
2354 * For paravirt, this is coupled with an exit in switch_to to
2355 * combine the page table reload and the switch backend into
2358 arch_start_context_switch(prev);
2361 next->active_mm = oldmm;
2362 atomic_inc(&oldmm->mm_count);
2363 enter_lazy_tlb(oldmm, next);
2365 switch_mm_irqs_off(oldmm, mm, next);
2368 prev->active_mm = NULL;
2369 rq->prev_mm = oldmm;
2372 * Since the runqueue lock will be released by the next
2373 * task (which is an invalid locking op but in the case
2374 * of the scheduler it's an obvious special-case), so we
2375 * do an early lockdep release here:
2377 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2378 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2381 context_tracking_task_switch(prev, next);
2382 /* Here we just switch the register state and the stack. */
2383 switch_to(prev, next, prev);
2387 * this_rq must be evaluated again because prev may have moved
2388 * CPUs since it called schedule(), thus the 'rq' on its stack
2389 * frame will be invalid.
2391 finish_task_switch(this_rq(), prev);
2395 * nr_running and nr_context_switches:
2397 * externally visible scheduler statistics: current number of runnable
2398 * threads, total number of context switches performed since bootup.
2400 unsigned long nr_running(void)
2402 unsigned long i, sum = 0;
2404 for_each_online_cpu(i)
2405 sum += cpu_rq(i)->nr_running;
2410 unsigned long long nr_context_switches(void)
2413 unsigned long long sum = 0;
2415 for_each_possible_cpu(i)
2416 sum += cpu_rq(i)->nr_switches;
2421 unsigned long nr_iowait(void)
2423 unsigned long i, sum = 0;
2425 for_each_possible_cpu(i)
2426 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2431 unsigned long nr_iowait_cpu(int cpu)
2433 struct rq *this = cpu_rq(cpu);
2434 return atomic_read(&this->nr_iowait);
2440 * sched_exec - execve() is a valuable balancing opportunity, because at
2441 * this point the task has the smallest effective memory and cache footprint.
2443 void sched_exec(void)
2445 struct task_struct *p = current;
2446 unsigned long flags;
2449 raw_spin_lock_irqsave(&p->pi_lock, flags);
2450 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2451 if (dest_cpu == smp_processor_id())
2454 if (likely(cpu_active(dest_cpu))) {
2455 struct migration_arg arg = { p, dest_cpu };
2457 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2458 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2462 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2467 DEFINE_PER_CPU(struct kernel_stat, kstat);
2468 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2470 EXPORT_PER_CPU_SYMBOL(kstat);
2471 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2474 * Return any ns on the sched_clock that have not yet been accounted in
2475 * @p in case that task is currently running.
2477 * Called with task_rq_lock() held on @rq.
2479 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2483 if (task_current(rq, p)) {
2484 update_rq_clock(rq);
2485 ns = rq_clock_task(rq) - p->se.exec_start;
2493 unsigned long long task_delta_exec(struct task_struct *p)
2495 unsigned long flags;
2499 rq = task_rq_lock(p, &flags);
2500 ns = do_task_delta_exec(p, rq);
2501 task_rq_unlock(rq, p, &flags);
2507 * Return accounted runtime for the task.
2508 * In case the task is currently running, return the runtime plus current's
2509 * pending runtime that have not been accounted yet.
2511 unsigned long long task_sched_runtime(struct task_struct *p)
2513 unsigned long flags;
2517 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2519 * 64-bit doesn't need locks to atomically read a 64bit value.
2520 * So we have a optimization chance when the task's delta_exec is 0.
2521 * Reading ->on_cpu is racy, but this is ok.
2523 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2524 * If we race with it entering cpu, unaccounted time is 0. This is
2525 * indistinguishable from the read occurring a few cycles earlier.
2528 return p->se.sum_exec_runtime;
2531 rq = task_rq_lock(p, &flags);
2532 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2533 task_rq_unlock(rq, p, &flags);
2539 * This function gets called by the timer code, with HZ frequency.
2540 * We call it with interrupts disabled.
2542 void scheduler_tick(void)
2544 int cpu = smp_processor_id();
2545 struct rq *rq = cpu_rq(cpu);
2546 struct task_struct *curr = rq->curr;
2550 raw_spin_lock(&rq->lock);
2551 update_rq_clock(rq);
2552 curr->sched_class->task_tick(rq, curr, 0);
2553 update_cpu_load_active(rq);
2554 raw_spin_unlock(&rq->lock);
2556 perf_event_task_tick();
2559 rq->idle_balance = idle_cpu(cpu);
2560 trigger_load_balance(rq);
2562 rq_last_tick_reset(rq);
2565 #ifdef CONFIG_NO_HZ_FULL
2567 * scheduler_tick_max_deferment
2569 * Keep at least one tick per second when a single
2570 * active task is running because the scheduler doesn't
2571 * yet completely support full dynticks environment.
2573 * This makes sure that uptime, CFS vruntime, load
2574 * balancing, etc... continue to move forward, even
2575 * with a very low granularity.
2577 * Return: Maximum deferment in nanoseconds.
2579 u64 scheduler_tick_max_deferment(void)
2581 struct rq *rq = this_rq();
2582 unsigned long next, now = ACCESS_ONCE(jiffies);
2584 next = rq->last_sched_tick + HZ;
2586 if (time_before_eq(next, now))
2589 return jiffies_to_nsecs(next - now);
2593 notrace unsigned long get_parent_ip(unsigned long addr)
2595 if (in_lock_functions(addr)) {
2596 addr = CALLER_ADDR2;
2597 if (in_lock_functions(addr))
2598 addr = CALLER_ADDR3;
2603 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2604 defined(CONFIG_PREEMPT_TRACER))
2606 void preempt_count_add(int val)
2608 #ifdef CONFIG_DEBUG_PREEMPT
2612 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2615 __preempt_count_add(val);
2616 #ifdef CONFIG_DEBUG_PREEMPT
2618 * Spinlock count overflowing soon?
2620 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2623 if (preempt_count() == val) {
2624 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2625 #ifdef CONFIG_DEBUG_PREEMPT
2626 current->preempt_disable_ip = ip;
2628 trace_preempt_off(CALLER_ADDR0, ip);
2631 EXPORT_SYMBOL(preempt_count_add);
2632 NOKPROBE_SYMBOL(preempt_count_add);
2634 void preempt_count_sub(int val)
2636 #ifdef CONFIG_DEBUG_PREEMPT
2640 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2643 * Is the spinlock portion underflowing?
2645 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2646 !(preempt_count() & PREEMPT_MASK)))
2650 if (preempt_count() == val)
2651 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2652 __preempt_count_sub(val);
2654 EXPORT_SYMBOL(preempt_count_sub);
2655 NOKPROBE_SYMBOL(preempt_count_sub);
2660 * Print scheduling while atomic bug:
2662 static noinline void __schedule_bug(struct task_struct *prev)
2664 if (oops_in_progress)
2667 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2668 prev->comm, prev->pid, preempt_count());
2670 debug_show_held_locks(prev);
2672 if (irqs_disabled())
2673 print_irqtrace_events(prev);
2674 #ifdef CONFIG_DEBUG_PREEMPT
2675 if (in_atomic_preempt_off()) {
2676 pr_err("Preemption disabled at:");
2677 print_ip_sym(current->preempt_disable_ip);
2682 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2686 * Various schedule()-time debugging checks and statistics:
2688 static inline void schedule_debug(struct task_struct *prev)
2691 * Test if we are atomic. Since do_exit() needs to call into
2692 * schedule() atomically, we ignore that path. Otherwise whine
2693 * if we are scheduling when we should not.
2695 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2696 __schedule_bug(prev);
2699 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2701 schedstat_inc(this_rq(), sched_count);
2705 * Pick up the highest-prio task:
2707 static inline struct task_struct *
2708 pick_next_task(struct rq *rq, struct task_struct *prev)
2710 const struct sched_class *class = &fair_sched_class;
2711 struct task_struct *p;
2714 * Optimization: we know that if all tasks are in
2715 * the fair class we can call that function directly:
2717 if (likely(prev->sched_class == class &&
2718 rq->nr_running == rq->cfs.h_nr_running)) {
2719 p = fair_sched_class.pick_next_task(rq, prev);
2720 if (unlikely(p == RETRY_TASK))
2723 /* assumes fair_sched_class->next == idle_sched_class */
2725 p = idle_sched_class.pick_next_task(rq, prev);
2731 for_each_class(class) {
2732 p = class->pick_next_task(rq, prev);
2734 if (unlikely(p == RETRY_TASK))
2740 BUG(); /* the idle class will always have a runnable task */
2744 * __schedule() is the main scheduler function.
2746 * The main means of driving the scheduler and thus entering this function are:
2748 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2750 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2751 * paths. For example, see arch/x86/entry_64.S.
2753 * To drive preemption between tasks, the scheduler sets the flag in timer
2754 * interrupt handler scheduler_tick().
2756 * 3. Wakeups don't really cause entry into schedule(). They add a
2757 * task to the run-queue and that's it.
2759 * Now, if the new task added to the run-queue preempts the current
2760 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2761 * called on the nearest possible occasion:
2763 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2765 * - in syscall or exception context, at the next outmost
2766 * preempt_enable(). (this might be as soon as the wake_up()'s
2769 * - in IRQ context, return from interrupt-handler to
2770 * preemptible context
2772 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2775 * - cond_resched() call
2776 * - explicit schedule() call
2777 * - return from syscall or exception to user-space
2778 * - return from interrupt-handler to user-space
2780 static void __sched __schedule(void)
2782 struct task_struct *prev, *next;
2783 unsigned long *switch_count;
2789 cpu = smp_processor_id();
2791 rcu_note_context_switch(cpu);
2794 schedule_debug(prev);
2796 if (sched_feat(HRTICK))
2800 * Make sure that signal_pending_state()->signal_pending() below
2801 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2802 * done by the caller to avoid the race with signal_wake_up().
2804 smp_mb__before_spinlock();
2805 raw_spin_lock_irq(&rq->lock);
2807 switch_count = &prev->nivcsw;
2808 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2809 if (unlikely(signal_pending_state(prev->state, prev))) {
2810 prev->state = TASK_RUNNING;
2812 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2816 * If a worker went to sleep, notify and ask workqueue
2817 * whether it wants to wake up a task to maintain
2820 if (prev->flags & PF_WQ_WORKER) {
2821 struct task_struct *to_wakeup;
2823 to_wakeup = wq_worker_sleeping(prev, cpu);
2825 try_to_wake_up_local(to_wakeup);
2828 switch_count = &prev->nvcsw;
2831 if (prev->on_rq || rq->skip_clock_update < 0)
2832 update_rq_clock(rq);
2834 next = pick_next_task(rq, prev);
2835 clear_tsk_need_resched(prev);
2836 clear_preempt_need_resched();
2837 rq->skip_clock_update = 0;
2839 if (likely(prev != next)) {
2844 context_switch(rq, prev, next); /* unlocks the rq */
2846 * The context switch have flipped the stack from under us
2847 * and restored the local variables which were saved when
2848 * this task called schedule() in the past. prev == current
2849 * is still correct, but it can be moved to another cpu/rq.
2851 cpu = smp_processor_id();
2854 raw_spin_unlock_irq(&rq->lock);
2856 balance_callback(rq);
2858 sched_preempt_enable_no_resched();
2863 static inline void sched_submit_work(struct task_struct *tsk)
2865 if (!tsk->state || tsk_is_pi_blocked(tsk))
2868 * If we are going to sleep and we have plugged IO queued,
2869 * make sure to submit it to avoid deadlocks.
2871 if (blk_needs_flush_plug(tsk))
2872 blk_schedule_flush_plug(tsk);
2875 asmlinkage __visible void __sched schedule(void)
2877 struct task_struct *tsk = current;
2879 sched_submit_work(tsk);
2882 EXPORT_SYMBOL(schedule);
2884 #ifdef CONFIG_CONTEXT_TRACKING
2885 asmlinkage __visible void __sched schedule_user(void)
2888 * If we come here after a random call to set_need_resched(),
2889 * or we have been woken up remotely but the IPI has not yet arrived,
2890 * we haven't yet exited the RCU idle mode. Do it here manually until
2891 * we find a better solution.
2900 * schedule_preempt_disabled - called with preemption disabled
2902 * Returns with preemption disabled. Note: preempt_count must be 1
2904 void __sched schedule_preempt_disabled(void)
2906 sched_preempt_enable_no_resched();
2911 #ifdef CONFIG_PREEMPT
2913 * this is the entry point to schedule() from in-kernel preemption
2914 * off of preempt_enable. Kernel preemptions off return from interrupt
2915 * occur there and call schedule directly.
2917 asmlinkage __visible void __sched notrace preempt_schedule(void)
2920 * If there is a non-zero preempt_count or interrupts are disabled,
2921 * we do not want to preempt the current task. Just return..
2923 if (likely(!preemptible()))
2927 __preempt_count_add(PREEMPT_ACTIVE);
2929 __preempt_count_sub(PREEMPT_ACTIVE);
2932 * Check again in case we missed a preemption opportunity
2933 * between schedule and now.
2936 } while (need_resched());
2938 NOKPROBE_SYMBOL(preempt_schedule);
2939 EXPORT_SYMBOL(preempt_schedule);
2940 #endif /* CONFIG_PREEMPT */
2943 * this is the entry point to schedule() from kernel preemption
2944 * off of irq context.
2945 * Note, that this is called and return with irqs disabled. This will
2946 * protect us against recursive calling from irq.
2948 asmlinkage __visible void __sched preempt_schedule_irq(void)
2950 enum ctx_state prev_state;
2952 /* Catch callers which need to be fixed */
2953 BUG_ON(preempt_count() || !irqs_disabled());
2955 prev_state = exception_enter();
2958 __preempt_count_add(PREEMPT_ACTIVE);
2961 local_irq_disable();
2962 __preempt_count_sub(PREEMPT_ACTIVE);
2965 * Check again in case we missed a preemption opportunity
2966 * between schedule and now.
2969 } while (need_resched());
2971 exception_exit(prev_state);
2974 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2977 return try_to_wake_up(curr->private, mode, wake_flags);
2979 EXPORT_SYMBOL(default_wake_function);
2981 #ifdef CONFIG_RT_MUTEXES
2984 * rt_mutex_setprio - set the current priority of a task
2986 * @prio: prio value (kernel-internal form)
2988 * This function changes the 'effective' priority of a task. It does
2989 * not touch ->normal_prio like __setscheduler().
2991 * Used by the rt_mutex code to implement priority inheritance
2992 * logic. Call site only calls if the priority of the task changed.
2994 void rt_mutex_setprio(struct task_struct *p, int prio)
2996 int oldprio, on_rq, running, enqueue_flag = 0;
2998 const struct sched_class *prev_class;
3000 BUG_ON(prio > MAX_PRIO);
3002 rq = __task_rq_lock(p);
3005 * Idle task boosting is a nono in general. There is one
3006 * exception, when PREEMPT_RT and NOHZ is active:
3008 * The idle task calls get_next_timer_interrupt() and holds
3009 * the timer wheel base->lock on the CPU and another CPU wants
3010 * to access the timer (probably to cancel it). We can safely
3011 * ignore the boosting request, as the idle CPU runs this code
3012 * with interrupts disabled and will complete the lock
3013 * protected section without being interrupted. So there is no
3014 * real need to boost.
3016 if (unlikely(p == rq->idle)) {
3017 WARN_ON(p != rq->curr);
3018 WARN_ON(p->pi_blocked_on);
3022 trace_sched_pi_setprio(p, prio);
3023 p->pi_top_task = rt_mutex_get_top_task(p);
3025 prev_class = p->sched_class;
3027 running = task_current(rq, p);
3029 dequeue_task(rq, p, 0);
3031 p->sched_class->put_prev_task(rq, p);
3034 * Boosting condition are:
3035 * 1. -rt task is running and holds mutex A
3036 * --> -dl task blocks on mutex A
3038 * 2. -dl task is running and holds mutex A
3039 * --> -dl task blocks on mutex A and could preempt the
3042 if (dl_prio(prio)) {
3043 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
3044 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
3045 p->dl.dl_boosted = 1;
3046 p->dl.dl_throttled = 0;
3047 enqueue_flag = ENQUEUE_REPLENISH;
3049 p->dl.dl_boosted = 0;
3050 p->sched_class = &dl_sched_class;
3051 } else if (rt_prio(prio)) {
3052 if (dl_prio(oldprio))
3053 p->dl.dl_boosted = 0;
3055 enqueue_flag = ENQUEUE_HEAD;
3056 p->sched_class = &rt_sched_class;
3058 if (dl_prio(oldprio))
3059 p->dl.dl_boosted = 0;
3060 if (rt_prio(oldprio))
3062 p->sched_class = &fair_sched_class;
3068 p->sched_class->set_curr_task(rq);
3070 enqueue_task(rq, p, enqueue_flag);
3072 check_class_changed(rq, p, prev_class, oldprio);
3074 preempt_disable(); /* avoid rq from going away on us */
3075 __task_rq_unlock(rq);
3077 balance_callback(rq);
3082 void set_user_nice(struct task_struct *p, long nice)
3084 int old_prio, delta, on_rq;
3085 unsigned long flags;
3088 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3091 * We have to be careful, if called from sys_setpriority(),
3092 * the task might be in the middle of scheduling on another CPU.
3094 rq = task_rq_lock(p, &flags);
3096 * The RT priorities are set via sched_setscheduler(), but we still
3097 * allow the 'normal' nice value to be set - but as expected
3098 * it wont have any effect on scheduling until the task is
3099 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3101 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3102 p->static_prio = NICE_TO_PRIO(nice);
3107 dequeue_task(rq, p, 0);
3109 p->static_prio = NICE_TO_PRIO(nice);
3112 p->prio = effective_prio(p);
3113 delta = p->prio - old_prio;
3116 enqueue_task(rq, p, 0);
3118 * If the task increased its priority or is running and
3119 * lowered its priority, then reschedule its CPU:
3121 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3122 resched_task(rq->curr);
3125 task_rq_unlock(rq, p, &flags);
3127 EXPORT_SYMBOL(set_user_nice);
3130 * can_nice - check if a task can reduce its nice value
3134 int can_nice(const struct task_struct *p, const int nice)
3136 /* convert nice value [19,-20] to rlimit style value [1,40] */
3137 int nice_rlim = nice_to_rlimit(nice);
3139 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3140 capable(CAP_SYS_NICE));
3143 #ifdef __ARCH_WANT_SYS_NICE
3146 * sys_nice - change the priority of the current process.
3147 * @increment: priority increment
3149 * sys_setpriority is a more generic, but much slower function that
3150 * does similar things.
3152 SYSCALL_DEFINE1(nice, int, increment)
3157 * Setpriority might change our priority at the same moment.
3158 * We don't have to worry. Conceptually one call occurs first
3159 * and we have a single winner.
3161 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3162 nice = task_nice(current) + increment;
3164 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3165 if (increment < 0 && !can_nice(current, nice))
3168 retval = security_task_setnice(current, nice);
3172 set_user_nice(current, nice);
3179 * task_prio - return the priority value of a given task.
3180 * @p: the task in question.
3182 * Return: The priority value as seen by users in /proc.
3183 * RT tasks are offset by -200. Normal tasks are centered
3184 * around 0, value goes from -16 to +15.
3186 int task_prio(const struct task_struct *p)
3188 return p->prio - MAX_RT_PRIO;
3192 * idle_cpu - is a given cpu idle currently?
3193 * @cpu: the processor in question.
3195 * Return: 1 if the CPU is currently idle. 0 otherwise.
3197 int idle_cpu(int cpu)
3199 struct rq *rq = cpu_rq(cpu);
3201 if (rq->curr != rq->idle)
3208 if (!llist_empty(&rq->wake_list))
3216 * idle_task - return the idle task for a given cpu.
3217 * @cpu: the processor in question.
3219 * Return: The idle task for the cpu @cpu.
3221 struct task_struct *idle_task(int cpu)
3223 return cpu_rq(cpu)->idle;
3227 * find_process_by_pid - find a process with a matching PID value.
3228 * @pid: the pid in question.
3230 * The task of @pid, if found. %NULL otherwise.
3232 static struct task_struct *find_process_by_pid(pid_t pid)
3234 return pid ? find_task_by_vpid(pid) : current;
3238 * This function initializes the sched_dl_entity of a newly becoming
3239 * SCHED_DEADLINE task.
3241 * Only the static values are considered here, the actual runtime and the
3242 * absolute deadline will be properly calculated when the task is enqueued
3243 * for the first time with its new policy.
3246 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3248 struct sched_dl_entity *dl_se = &p->dl;
3250 init_dl_task_timer(dl_se);
3251 dl_se->dl_runtime = attr->sched_runtime;
3252 dl_se->dl_deadline = attr->sched_deadline;
3253 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3254 dl_se->flags = attr->sched_flags;
3255 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3256 dl_se->dl_throttled = 0;
3258 dl_se->dl_yielded = 0;
3261 static void __setscheduler_params(struct task_struct *p,
3262 const struct sched_attr *attr)
3264 int policy = attr->sched_policy;
3266 if (policy == -1) /* setparam */
3271 if (dl_policy(policy))
3272 __setparam_dl(p, attr);
3273 else if (fair_policy(policy))
3274 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3277 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3278 * !rt_policy. Always setting this ensures that things like
3279 * getparam()/getattr() don't report silly values for !rt tasks.
3281 p->rt_priority = attr->sched_priority;
3282 p->normal_prio = normal_prio(p);
3286 /* Actually do priority change: must hold pi & rq lock. */
3287 static void __setscheduler(struct rq *rq, struct task_struct *p,
3288 const struct sched_attr *attr, bool keep_boost)
3290 __setscheduler_params(p, attr);
3293 * Keep a potential priority boosting if called from
3294 * sched_setscheduler().
3297 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3299 p->prio = normal_prio(p);
3301 if (dl_prio(p->prio))
3302 p->sched_class = &dl_sched_class;
3303 else if (rt_prio(p->prio))
3304 p->sched_class = &rt_sched_class;
3306 p->sched_class = &fair_sched_class;
3310 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3312 struct sched_dl_entity *dl_se = &p->dl;
3314 attr->sched_priority = p->rt_priority;
3315 attr->sched_runtime = dl_se->dl_runtime;
3316 attr->sched_deadline = dl_se->dl_deadline;
3317 attr->sched_period = dl_se->dl_period;
3318 attr->sched_flags = dl_se->flags;
3322 * This function validates the new parameters of a -deadline task.
3323 * We ask for the deadline not being zero, and greater or equal
3324 * than the runtime, as well as the period of being zero or
3325 * greater than deadline. Furthermore, we have to be sure that
3326 * user parameters are above the internal resolution of 1us (we
3327 * check sched_runtime only since it is always the smaller one) and
3328 * below 2^63 ns (we have to check both sched_deadline and
3329 * sched_period, as the latter can be zero).
3332 __checkparam_dl(const struct sched_attr *attr)
3335 if (attr->sched_deadline == 0)
3339 * Since we truncate DL_SCALE bits, make sure we're at least
3342 if (attr->sched_runtime < (1ULL << DL_SCALE))
3346 * Since we use the MSB for wrap-around and sign issues, make
3347 * sure it's not set (mind that period can be equal to zero).
3349 if (attr->sched_deadline & (1ULL << 63) ||
3350 attr->sched_period & (1ULL << 63))
3353 /* runtime <= deadline <= period (if period != 0) */
3354 if ((attr->sched_period != 0 &&
3355 attr->sched_period < attr->sched_deadline) ||
3356 attr->sched_deadline < attr->sched_runtime)
3363 * check the target process has a UID that matches the current process's
3365 static bool check_same_owner(struct task_struct *p)
3367 const struct cred *cred = current_cred(), *pcred;
3371 pcred = __task_cred(p);
3372 match = (uid_eq(cred->euid, pcred->euid) ||
3373 uid_eq(cred->euid, pcred->uid));
3378 static int __sched_setscheduler(struct task_struct *p,
3379 const struct sched_attr *attr,
3382 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3383 MAX_RT_PRIO - 1 - attr->sched_priority;
3384 int retval, oldprio, oldpolicy = -1, on_rq, running;
3385 int new_effective_prio, policy = attr->sched_policy;
3386 unsigned long flags;
3387 const struct sched_class *prev_class;
3391 /* may grab non-irq protected spin_locks */
3392 BUG_ON(in_interrupt());
3394 /* double check policy once rq lock held */
3396 reset_on_fork = p->sched_reset_on_fork;
3397 policy = oldpolicy = p->policy;
3399 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3401 if (policy != SCHED_DEADLINE &&
3402 policy != SCHED_FIFO && policy != SCHED_RR &&
3403 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3404 policy != SCHED_IDLE)
3408 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3412 * Valid priorities for SCHED_FIFO and SCHED_RR are
3413 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3414 * SCHED_BATCH and SCHED_IDLE is 0.
3416 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3417 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3419 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3420 (rt_policy(policy) != (attr->sched_priority != 0)))
3424 * Allow unprivileged RT tasks to decrease priority:
3426 if (user && !capable(CAP_SYS_NICE)) {
3427 if (fair_policy(policy)) {
3428 if (attr->sched_nice < task_nice(p) &&
3429 !can_nice(p, attr->sched_nice))
3433 if (rt_policy(policy)) {
3434 unsigned long rlim_rtprio =
3435 task_rlimit(p, RLIMIT_RTPRIO);
3437 /* can't set/change the rt policy */
3438 if (policy != p->policy && !rlim_rtprio)
3441 /* can't increase priority */
3442 if (attr->sched_priority > p->rt_priority &&
3443 attr->sched_priority > rlim_rtprio)
3448 * Can't set/change SCHED_DEADLINE policy at all for now
3449 * (safest behavior); in the future we would like to allow
3450 * unprivileged DL tasks to increase their relative deadline
3451 * or reduce their runtime (both ways reducing utilization)
3453 if (dl_policy(policy))
3457 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3458 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3460 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3461 if (!can_nice(p, task_nice(p)))
3465 /* can't change other user's priorities */
3466 if (!check_same_owner(p))
3469 /* Normal users shall not reset the sched_reset_on_fork flag */
3470 if (p->sched_reset_on_fork && !reset_on_fork)
3475 retval = security_task_setscheduler(p);
3481 * make sure no PI-waiters arrive (or leave) while we are
3482 * changing the priority of the task:
3484 * To be able to change p->policy safely, the appropriate
3485 * runqueue lock must be held.
3487 rq = task_rq_lock(p, &flags);
3490 * Changing the policy of the stop threads its a very bad idea
3492 if (p == rq->stop) {
3493 task_rq_unlock(rq, p, &flags);
3498 * If not changing anything there's no need to proceed further,
3499 * but store a possible modification of reset_on_fork.
3501 if (unlikely(policy == p->policy)) {
3502 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3504 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3506 if (dl_policy(policy))
3509 p->sched_reset_on_fork = reset_on_fork;
3510 task_rq_unlock(rq, p, &flags);
3516 #ifdef CONFIG_RT_GROUP_SCHED
3518 * Do not allow realtime tasks into groups that have no runtime
3521 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3522 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3523 !task_group_is_autogroup(task_group(p))) {
3524 task_rq_unlock(rq, p, &flags);
3529 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3530 cpumask_t *span = rq->rd->span;
3533 * Don't allow tasks with an affinity mask smaller than
3534 * the entire root_domain to become SCHED_DEADLINE. We
3535 * will also fail if there's no bandwidth available.
3537 if (!cpumask_subset(span, &p->cpus_allowed) ||
3538 rq->rd->dl_bw.bw == 0) {
3539 task_rq_unlock(rq, p, &flags);
3546 /* recheck policy now with rq lock held */
3547 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3548 policy = oldpolicy = -1;
3549 task_rq_unlock(rq, p, &flags);
3554 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3555 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3558 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3559 task_rq_unlock(rq, p, &flags);
3563 p->sched_reset_on_fork = reset_on_fork;
3567 * Take priority boosted tasks into account. If the new
3568 * effective priority is unchanged, we just store the new
3569 * normal parameters and do not touch the scheduler class and
3570 * the runqueue. This will be done when the task deboost
3573 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3574 if (new_effective_prio == oldprio) {
3575 __setscheduler_params(p, attr);
3576 task_rq_unlock(rq, p, &flags);
3581 running = task_current(rq, p);
3583 dequeue_task(rq, p, 0);
3585 p->sched_class->put_prev_task(rq, p);
3587 prev_class = p->sched_class;
3588 __setscheduler(rq, p, attr, true);
3591 p->sched_class->set_curr_task(rq);
3594 * We enqueue to tail when the priority of a task is
3595 * increased (user space view).
3597 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3600 check_class_changed(rq, p, prev_class, oldprio);
3601 preempt_disable(); /* avoid rq from going away on us */
3602 task_rq_unlock(rq, p, &flags);
3604 rt_mutex_adjust_pi(p);
3607 * Run balance callbacks after we've adjusted the PI chain.
3609 balance_callback(rq);
3615 static int _sched_setscheduler(struct task_struct *p, int policy,
3616 const struct sched_param *param, bool check)
3618 struct sched_attr attr = {
3619 .sched_policy = policy,
3620 .sched_priority = param->sched_priority,
3621 .sched_nice = PRIO_TO_NICE(p->static_prio),
3625 * Fixup the legacy SCHED_RESET_ON_FORK hack, except if
3626 * the policy=-1 was passed by sched_setparam().
3628 if ((policy != -1) && (policy & SCHED_RESET_ON_FORK)) {
3629 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3630 policy &= ~SCHED_RESET_ON_FORK;
3631 attr.sched_policy = policy;
3634 return __sched_setscheduler(p, &attr, check);
3637 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3638 * @p: the task in question.
3639 * @policy: new policy.
3640 * @param: structure containing the new RT priority.
3642 * Return: 0 on success. An error code otherwise.
3644 * NOTE that the task may be already dead.
3646 int sched_setscheduler(struct task_struct *p, int policy,
3647 const struct sched_param *param)
3649 return _sched_setscheduler(p, policy, param, true);
3651 EXPORT_SYMBOL_GPL(sched_setscheduler);
3653 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3655 return __sched_setscheduler(p, attr, true);
3657 EXPORT_SYMBOL_GPL(sched_setattr);
3660 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3661 * @p: the task in question.
3662 * @policy: new policy.
3663 * @param: structure containing the new RT priority.
3665 * Just like sched_setscheduler, only don't bother checking if the
3666 * current context has permission. For example, this is needed in
3667 * stop_machine(): we create temporary high priority worker threads,
3668 * but our caller might not have that capability.
3670 * Return: 0 on success. An error code otherwise.
3672 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3673 const struct sched_param *param)
3675 return _sched_setscheduler(p, policy, param, false);
3679 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3681 struct sched_param lparam;
3682 struct task_struct *p;
3685 if (!param || pid < 0)
3687 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3692 p = find_process_by_pid(pid);
3694 retval = sched_setscheduler(p, policy, &lparam);
3701 * Mimics kernel/events/core.c perf_copy_attr().
3703 static int sched_copy_attr(struct sched_attr __user *uattr,
3704 struct sched_attr *attr)
3709 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3713 * zero the full structure, so that a short copy will be nice.
3715 memset(attr, 0, sizeof(*attr));
3717 ret = get_user(size, &uattr->size);
3721 if (size > PAGE_SIZE) /* silly large */
3724 if (!size) /* abi compat */
3725 size = SCHED_ATTR_SIZE_VER0;
3727 if (size < SCHED_ATTR_SIZE_VER0)
3731 * If we're handed a bigger struct than we know of,
3732 * ensure all the unknown bits are 0 - i.e. new
3733 * user-space does not rely on any kernel feature
3734 * extensions we dont know about yet.
3736 if (size > sizeof(*attr)) {
3737 unsigned char __user *addr;
3738 unsigned char __user *end;
3741 addr = (void __user *)uattr + sizeof(*attr);
3742 end = (void __user *)uattr + size;
3744 for (; addr < end; addr++) {
3745 ret = get_user(val, addr);
3751 size = sizeof(*attr);
3754 ret = copy_from_user(attr, uattr, size);
3759 * XXX: do we want to be lenient like existing syscalls; or do we want
3760 * to be strict and return an error on out-of-bounds values?
3762 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3767 put_user(sizeof(*attr), &uattr->size);
3772 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3773 * @pid: the pid in question.
3774 * @policy: new policy.
3775 * @param: structure containing the new RT priority.
3777 * Return: 0 on success. An error code otherwise.
3779 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3780 struct sched_param __user *, param)
3782 /* negative values for policy are not valid */
3786 return do_sched_setscheduler(pid, policy, param);
3790 * sys_sched_setparam - set/change the RT priority of a thread
3791 * @pid: the pid in question.
3792 * @param: structure containing the new RT priority.
3794 * Return: 0 on success. An error code otherwise.
3796 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3798 return do_sched_setscheduler(pid, -1, param);
3802 * sys_sched_setattr - same as above, but with extended sched_attr
3803 * @pid: the pid in question.
3804 * @uattr: structure containing the extended parameters.
3805 * @flags: for future extension.
3807 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3808 unsigned int, flags)
3810 struct sched_attr attr;
3811 struct task_struct *p;
3814 if (!uattr || pid < 0 || flags)
3817 retval = sched_copy_attr(uattr, &attr);
3821 if ((int)attr.sched_policy < 0)
3826 p = find_process_by_pid(pid);
3828 retval = sched_setattr(p, &attr);
3835 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3836 * @pid: the pid in question.
3838 * Return: On success, the policy of the thread. Otherwise, a negative error
3841 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3843 struct task_struct *p;
3851 p = find_process_by_pid(pid);
3853 retval = security_task_getscheduler(p);
3856 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3863 * sys_sched_getparam - get the RT priority of a thread
3864 * @pid: the pid in question.
3865 * @param: structure containing the RT priority.
3867 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3870 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3872 struct sched_param lp = { .sched_priority = 0 };
3873 struct task_struct *p;
3876 if (!param || pid < 0)
3880 p = find_process_by_pid(pid);
3885 retval = security_task_getscheduler(p);
3889 if (task_has_rt_policy(p))
3890 lp.sched_priority = p->rt_priority;
3894 * This one might sleep, we cannot do it with a spinlock held ...
3896 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3905 static int sched_read_attr(struct sched_attr __user *uattr,
3906 struct sched_attr *attr,
3911 if (!access_ok(VERIFY_WRITE, uattr, usize))
3915 * If we're handed a smaller struct than we know of,
3916 * ensure all the unknown bits are 0 - i.e. old
3917 * user-space does not get uncomplete information.
3919 if (usize < sizeof(*attr)) {
3920 unsigned char *addr;
3923 addr = (void *)attr + usize;
3924 end = (void *)attr + sizeof(*attr);
3926 for (; addr < end; addr++) {
3934 ret = copy_to_user(uattr, attr, attr->size);
3942 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3943 * @pid: the pid in question.
3944 * @uattr: structure containing the extended parameters.
3945 * @size: sizeof(attr) for fwd/bwd comp.
3946 * @flags: for future extension.
3948 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3949 unsigned int, size, unsigned int, flags)
3951 struct sched_attr attr = {
3952 .size = sizeof(struct sched_attr),
3954 struct task_struct *p;
3957 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3958 size < SCHED_ATTR_SIZE_VER0 || flags)
3962 p = find_process_by_pid(pid);
3967 retval = security_task_getscheduler(p);
3971 attr.sched_policy = p->policy;
3972 if (p->sched_reset_on_fork)
3973 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3974 if (task_has_dl_policy(p))
3975 __getparam_dl(p, &attr);
3976 else if (task_has_rt_policy(p))
3977 attr.sched_priority = p->rt_priority;
3979 attr.sched_nice = task_nice(p);
3983 retval = sched_read_attr(uattr, &attr, size);
3991 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3993 cpumask_var_t cpus_allowed, new_mask;
3994 struct task_struct *p;
3999 p = find_process_by_pid(pid);
4005 /* Prevent p going away */
4009 if (p->flags & PF_NO_SETAFFINITY) {
4013 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4017 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4019 goto out_free_cpus_allowed;
4022 if (!check_same_owner(p)) {
4024 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4031 retval = security_task_setscheduler(p);
4036 cpuset_cpus_allowed(p, cpus_allowed);
4037 cpumask_and(new_mask, in_mask, cpus_allowed);
4040 * Since bandwidth control happens on root_domain basis,
4041 * if admission test is enabled, we only admit -deadline
4042 * tasks allowed to run on all the CPUs in the task's
4046 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4048 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4057 retval = set_cpus_allowed_ptr(p, new_mask);
4060 cpuset_cpus_allowed(p, cpus_allowed);
4061 if (!cpumask_subset(new_mask, cpus_allowed)) {
4063 * We must have raced with a concurrent cpuset
4064 * update. Just reset the cpus_allowed to the
4065 * cpuset's cpus_allowed
4067 cpumask_copy(new_mask, cpus_allowed);
4072 free_cpumask_var(new_mask);
4073 out_free_cpus_allowed:
4074 free_cpumask_var(cpus_allowed);
4080 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4081 struct cpumask *new_mask)
4083 if (len < cpumask_size())
4084 cpumask_clear(new_mask);
4085 else if (len > cpumask_size())
4086 len = cpumask_size();
4088 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4092 * sys_sched_setaffinity - set the cpu affinity of a process
4093 * @pid: pid of the process
4094 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4095 * @user_mask_ptr: user-space pointer to the new cpu mask
4097 * Return: 0 on success. An error code otherwise.
4099 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4100 unsigned long __user *, user_mask_ptr)
4102 cpumask_var_t new_mask;
4105 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4108 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4110 retval = sched_setaffinity(pid, new_mask);
4111 free_cpumask_var(new_mask);
4115 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4117 struct task_struct *p;
4118 unsigned long flags;
4124 p = find_process_by_pid(pid);
4128 retval = security_task_getscheduler(p);
4132 raw_spin_lock_irqsave(&p->pi_lock, flags);
4133 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4134 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4143 * sys_sched_getaffinity - get the cpu affinity of a process
4144 * @pid: pid of the process
4145 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4146 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4148 * Return: 0 on success. An error code otherwise.
4150 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4151 unsigned long __user *, user_mask_ptr)
4156 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4158 if (len & (sizeof(unsigned long)-1))
4161 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4164 ret = sched_getaffinity(pid, mask);
4166 size_t retlen = min_t(size_t, len, cpumask_size());
4168 if (copy_to_user(user_mask_ptr, mask, retlen))
4173 free_cpumask_var(mask);
4179 * sys_sched_yield - yield the current processor to other threads.
4181 * This function yields the current CPU to other tasks. If there are no
4182 * other threads running on this CPU then this function will return.
4186 SYSCALL_DEFINE0(sched_yield)
4188 struct rq *rq = this_rq_lock();
4190 schedstat_inc(rq, yld_count);
4191 current->sched_class->yield_task(rq);
4194 * Since we are going to call schedule() anyway, there's
4195 * no need to preempt or enable interrupts:
4197 __release(rq->lock);
4198 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4199 do_raw_spin_unlock(&rq->lock);
4200 sched_preempt_enable_no_resched();
4207 static void __cond_resched(void)
4209 __preempt_count_add(PREEMPT_ACTIVE);
4211 __preempt_count_sub(PREEMPT_ACTIVE);
4214 int __sched _cond_resched(void)
4216 if (should_resched(0)) {
4222 EXPORT_SYMBOL(_cond_resched);
4225 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4226 * call schedule, and on return reacquire the lock.
4228 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4229 * operations here to prevent schedule() from being called twice (once via
4230 * spin_unlock(), once by hand).
4232 int __cond_resched_lock(spinlock_t *lock)
4234 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4237 lockdep_assert_held(lock);
4239 if (spin_needbreak(lock) || resched) {
4250 EXPORT_SYMBOL(__cond_resched_lock);
4252 int __sched __cond_resched_softirq(void)
4254 BUG_ON(!in_softirq());
4256 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4264 EXPORT_SYMBOL(__cond_resched_softirq);
4267 * yield - yield the current processor to other threads.
4269 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4271 * The scheduler is at all times free to pick the calling task as the most
4272 * eligible task to run, if removing the yield() call from your code breaks
4273 * it, its already broken.
4275 * Typical broken usage is:
4280 * where one assumes that yield() will let 'the other' process run that will
4281 * make event true. If the current task is a SCHED_FIFO task that will never
4282 * happen. Never use yield() as a progress guarantee!!
4284 * If you want to use yield() to wait for something, use wait_event().
4285 * If you want to use yield() to be 'nice' for others, use cond_resched().
4286 * If you still want to use yield(), do not!
4288 void __sched yield(void)
4290 set_current_state(TASK_RUNNING);
4293 EXPORT_SYMBOL(yield);
4296 * yield_to - yield the current processor to another thread in
4297 * your thread group, or accelerate that thread toward the
4298 * processor it's on.
4300 * @preempt: whether task preemption is allowed or not
4302 * It's the caller's job to ensure that the target task struct
4303 * can't go away on us before we can do any checks.
4306 * true (>0) if we indeed boosted the target task.
4307 * false (0) if we failed to boost the target.
4308 * -ESRCH if there's no task to yield to.
4310 int __sched yield_to(struct task_struct *p, bool preempt)
4312 struct task_struct *curr = current;
4313 struct rq *rq, *p_rq;
4314 unsigned long flags;
4317 local_irq_save(flags);
4323 * If we're the only runnable task on the rq and target rq also
4324 * has only one task, there's absolutely no point in yielding.
4326 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4331 double_rq_lock(rq, p_rq);
4332 if (task_rq(p) != p_rq) {
4333 double_rq_unlock(rq, p_rq);
4337 if (!curr->sched_class->yield_to_task)
4340 if (curr->sched_class != p->sched_class)
4343 if (task_running(p_rq, p) || p->state)
4346 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4348 schedstat_inc(rq, yld_count);
4350 * Make p's CPU reschedule; pick_next_entity takes care of
4353 if (preempt && rq != p_rq)
4354 resched_task(p_rq->curr);
4358 double_rq_unlock(rq, p_rq);
4360 local_irq_restore(flags);
4367 EXPORT_SYMBOL_GPL(yield_to);
4370 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4371 * that process accounting knows that this is a task in IO wait state.
4373 void __sched io_schedule(void)
4375 struct rq *rq = raw_rq();
4377 delayacct_blkio_start();
4378 atomic_inc(&rq->nr_iowait);
4379 blk_flush_plug(current);
4380 current->in_iowait = 1;
4382 current->in_iowait = 0;
4383 atomic_dec(&rq->nr_iowait);
4384 delayacct_blkio_end();
4386 EXPORT_SYMBOL(io_schedule);
4388 long __sched io_schedule_timeout(long timeout)
4390 struct rq *rq = raw_rq();
4393 delayacct_blkio_start();
4394 atomic_inc(&rq->nr_iowait);
4395 blk_flush_plug(current);
4396 current->in_iowait = 1;
4397 ret = schedule_timeout(timeout);
4398 current->in_iowait = 0;
4399 atomic_dec(&rq->nr_iowait);
4400 delayacct_blkio_end();
4405 * sys_sched_get_priority_max - return maximum RT priority.
4406 * @policy: scheduling class.
4408 * Return: On success, this syscall returns the maximum
4409 * rt_priority that can be used by a given scheduling class.
4410 * On failure, a negative error code is returned.
4412 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4419 ret = MAX_USER_RT_PRIO-1;
4421 case SCHED_DEADLINE:
4432 * sys_sched_get_priority_min - return minimum RT priority.
4433 * @policy: scheduling class.
4435 * Return: On success, this syscall returns the minimum
4436 * rt_priority that can be used by a given scheduling class.
4437 * On failure, a negative error code is returned.
4439 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4448 case SCHED_DEADLINE:
4458 * sys_sched_rr_get_interval - return the default timeslice of a process.
4459 * @pid: pid of the process.
4460 * @interval: userspace pointer to the timeslice value.
4462 * this syscall writes the default timeslice value of a given process
4463 * into the user-space timespec buffer. A value of '0' means infinity.
4465 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4468 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4469 struct timespec __user *, interval)
4471 struct task_struct *p;
4472 unsigned int time_slice;
4473 unsigned long flags;
4483 p = find_process_by_pid(pid);
4487 retval = security_task_getscheduler(p);
4491 rq = task_rq_lock(p, &flags);
4493 if (p->sched_class->get_rr_interval)
4494 time_slice = p->sched_class->get_rr_interval(rq, p);
4495 task_rq_unlock(rq, p, &flags);
4498 jiffies_to_timespec(time_slice, &t);
4499 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4507 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4509 void sched_show_task(struct task_struct *p)
4511 unsigned long free = 0;
4515 state = p->state ? __ffs(p->state) + 1 : 0;
4516 printk(KERN_INFO "%-15.15s %c", p->comm,
4517 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4518 #if BITS_PER_LONG == 32
4519 if (state == TASK_RUNNING)
4520 printk(KERN_CONT " running ");
4522 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4524 if (state == TASK_RUNNING)
4525 printk(KERN_CONT " running task ");
4527 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4529 #ifdef CONFIG_DEBUG_STACK_USAGE
4530 free = stack_not_used(p);
4533 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4535 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4536 task_pid_nr(p), ppid,
4537 (unsigned long)task_thread_info(p)->flags);
4539 print_worker_info(KERN_INFO, p);
4540 show_stack(p, NULL);
4543 void show_state_filter(unsigned long state_filter)
4545 struct task_struct *g, *p;
4547 #if BITS_PER_LONG == 32
4549 " task PC stack pid father\n");
4552 " task PC stack pid father\n");
4555 do_each_thread(g, p) {
4557 * reset the NMI-timeout, listing all files on a slow
4558 * console might take a lot of time:
4559 * Also, reset softlockup watchdogs on all CPUs, because
4560 * another CPU might be blocked waiting for us to process
4563 touch_nmi_watchdog();
4564 touch_all_softlockup_watchdogs();
4565 if (!state_filter || (p->state & state_filter))
4567 } while_each_thread(g, p);
4569 #ifdef CONFIG_SCHED_DEBUG
4570 sysrq_sched_debug_show();
4574 * Only show locks if all tasks are dumped:
4577 debug_show_all_locks();
4580 void init_idle_bootup_task(struct task_struct *idle)
4582 idle->sched_class = &idle_sched_class;
4586 * init_idle - set up an idle thread for a given CPU
4587 * @idle: task in question
4588 * @cpu: cpu the idle task belongs to
4590 * NOTE: this function does not set the idle thread's NEED_RESCHED
4591 * flag, to make booting more robust.
4593 void init_idle(struct task_struct *idle, int cpu)
4595 struct rq *rq = cpu_rq(cpu);
4596 unsigned long flags;
4598 raw_spin_lock_irqsave(&rq->lock, flags);
4600 __sched_fork(0, idle);
4601 idle->state = TASK_RUNNING;
4602 idle->se.exec_start = sched_clock();
4604 do_set_cpus_allowed(idle, cpumask_of(cpu));
4606 * We're having a chicken and egg problem, even though we are
4607 * holding rq->lock, the cpu isn't yet set to this cpu so the
4608 * lockdep check in task_group() will fail.
4610 * Similar case to sched_fork(). / Alternatively we could
4611 * use task_rq_lock() here and obtain the other rq->lock.
4616 __set_task_cpu(idle, cpu);
4619 rq->curr = rq->idle = idle;
4621 #if defined(CONFIG_SMP)
4624 raw_spin_unlock_irqrestore(&rq->lock, flags);
4626 /* Set the preempt count _outside_ the spinlocks! */
4627 init_idle_preempt_count(idle, cpu);
4630 * The idle tasks have their own, simple scheduling class:
4632 idle->sched_class = &idle_sched_class;
4633 ftrace_graph_init_idle_task(idle, cpu);
4634 vtime_init_idle(idle, cpu);
4635 #if defined(CONFIG_SMP)
4636 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4641 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4643 if (p->sched_class && p->sched_class->set_cpus_allowed)
4644 p->sched_class->set_cpus_allowed(p, new_mask);
4646 cpumask_copy(&p->cpus_allowed, new_mask);
4647 p->nr_cpus_allowed = cpumask_weight(new_mask);
4651 * This is how migration works:
4653 * 1) we invoke migration_cpu_stop() on the target CPU using
4655 * 2) stopper starts to run (implicitly forcing the migrated thread
4657 * 3) it checks whether the migrated task is still in the wrong runqueue.
4658 * 4) if it's in the wrong runqueue then the migration thread removes
4659 * it and puts it into the right queue.
4660 * 5) stopper completes and stop_one_cpu() returns and the migration
4665 * Change a given task's CPU affinity. Migrate the thread to a
4666 * proper CPU and schedule it away if the CPU it's executing on
4667 * is removed from the allowed bitmask.
4669 * NOTE: the caller must have a valid reference to the task, the
4670 * task must not exit() & deallocate itself prematurely. The
4671 * call is not atomic; no spinlocks may be held.
4673 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4675 unsigned long flags;
4677 unsigned int dest_cpu;
4680 rq = task_rq_lock(p, &flags);
4682 if (cpumask_equal(&p->cpus_allowed, new_mask))
4685 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4690 do_set_cpus_allowed(p, new_mask);
4692 /* Can the task run on the task's current CPU? If so, we're done */
4693 if (cpumask_test_cpu(task_cpu(p), new_mask))
4696 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4698 struct migration_arg arg = { p, dest_cpu };
4699 /* Need help from migration thread: drop lock and wait. */
4700 task_rq_unlock(rq, p, &flags);
4701 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4702 tlb_migrate_finish(p->mm);
4706 task_rq_unlock(rq, p, &flags);
4710 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4713 * Move (not current) task off this cpu, onto dest cpu. We're doing
4714 * this because either it can't run here any more (set_cpus_allowed()
4715 * away from this CPU, or CPU going down), or because we're
4716 * attempting to rebalance this task on exec (sched_exec).
4718 * So we race with normal scheduler movements, but that's OK, as long
4719 * as the task is no longer on this CPU.
4721 * Returns non-zero if task was successfully migrated.
4723 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4725 struct rq *rq_dest, *rq_src;
4728 if (unlikely(!cpu_active(dest_cpu)))
4731 rq_src = cpu_rq(src_cpu);
4732 rq_dest = cpu_rq(dest_cpu);
4734 raw_spin_lock(&p->pi_lock);
4735 double_rq_lock(rq_src, rq_dest);
4736 /* Already moved. */
4737 if (task_cpu(p) != src_cpu)
4739 /* Affinity changed (again). */
4740 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4744 * If we're not on a rq, the next wake-up will ensure we're
4748 dequeue_task(rq_src, p, 0);
4749 set_task_cpu(p, dest_cpu);
4750 enqueue_task(rq_dest, p, 0);
4751 check_preempt_curr(rq_dest, p, 0);
4756 double_rq_unlock(rq_src, rq_dest);
4757 raw_spin_unlock(&p->pi_lock);
4761 #ifdef CONFIG_NUMA_BALANCING
4762 /* Migrate current task p to target_cpu */
4763 int migrate_task_to(struct task_struct *p, int target_cpu)
4765 struct migration_arg arg = { p, target_cpu };
4766 int curr_cpu = task_cpu(p);
4768 if (curr_cpu == target_cpu)
4771 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4774 /* TODO: This is not properly updating schedstats */
4776 trace_sched_move_numa(p, curr_cpu, target_cpu);
4777 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4781 * Requeue a task on a given node and accurately track the number of NUMA
4782 * tasks on the runqueues
4784 void sched_setnuma(struct task_struct *p, int nid)
4787 unsigned long flags;
4788 bool on_rq, running;
4790 rq = task_rq_lock(p, &flags);
4792 running = task_current(rq, p);
4795 dequeue_task(rq, p, 0);
4797 p->sched_class->put_prev_task(rq, p);
4799 p->numa_preferred_nid = nid;
4802 p->sched_class->set_curr_task(rq);
4804 enqueue_task(rq, p, 0);
4805 task_rq_unlock(rq, p, &flags);
4810 * migration_cpu_stop - this will be executed by a highprio stopper thread
4811 * and performs thread migration by bumping thread off CPU then
4812 * 'pushing' onto another runqueue.
4814 static int migration_cpu_stop(void *data)
4816 struct migration_arg *arg = data;
4819 * The original target cpu might have gone down and we might
4820 * be on another cpu but it doesn't matter.
4822 local_irq_disable();
4823 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4828 #ifdef CONFIG_HOTPLUG_CPU
4831 * Ensures that the idle task is using init_mm right before its cpu goes
4834 void idle_task_exit(void)
4836 struct mm_struct *mm = current->active_mm;
4838 BUG_ON(cpu_online(smp_processor_id()));
4840 if (mm != &init_mm) {
4841 switch_mm(mm, &init_mm, current);
4842 finish_arch_post_lock_switch();
4848 * Since this CPU is going 'away' for a while, fold any nr_active delta
4849 * we might have. Assumes we're called after migrate_tasks() so that the
4850 * nr_active count is stable.
4852 * Also see the comment "Global load-average calculations".
4854 static void calc_load_migrate(struct rq *rq)
4856 long delta = calc_load_fold_active(rq);
4858 atomic_long_add(delta, &calc_load_tasks);
4861 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4865 static const struct sched_class fake_sched_class = {
4866 .put_prev_task = put_prev_task_fake,
4869 static struct task_struct fake_task = {
4871 * Avoid pull_{rt,dl}_task()
4873 .prio = MAX_PRIO + 1,
4874 .sched_class = &fake_sched_class,
4878 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4879 * try_to_wake_up()->select_task_rq().
4881 * Called with rq->lock held even though we'er in stop_machine() and
4882 * there's no concurrency possible, we hold the required locks anyway
4883 * because of lock validation efforts.
4885 static void migrate_tasks(unsigned int dead_cpu)
4887 struct rq *rq = cpu_rq(dead_cpu);
4888 struct task_struct *next, *stop = rq->stop;
4892 * Fudge the rq selection such that the below task selection loop
4893 * doesn't get stuck on the currently eligible stop task.
4895 * We're currently inside stop_machine() and the rq is either stuck
4896 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4897 * either way we should never end up calling schedule() until we're
4903 * put_prev_task() and pick_next_task() sched
4904 * class method both need to have an up-to-date
4905 * value of rq->clock[_task]
4907 update_rq_clock(rq);
4911 * There's this thread running, bail when that's the only
4914 if (rq->nr_running == 1)
4917 next = pick_next_task(rq, &fake_task);
4919 next->sched_class->put_prev_task(rq, next);
4921 /* Find suitable destination for @next, with force if needed. */
4922 dest_cpu = select_fallback_rq(dead_cpu, next);
4923 raw_spin_unlock(&rq->lock);
4925 __migrate_task(next, dead_cpu, dest_cpu);
4927 raw_spin_lock(&rq->lock);
4933 #endif /* CONFIG_HOTPLUG_CPU */
4935 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4937 static struct ctl_table sd_ctl_dir[] = {
4939 .procname = "sched_domain",
4945 static struct ctl_table sd_ctl_root[] = {
4947 .procname = "kernel",
4949 .child = sd_ctl_dir,
4954 static struct ctl_table *sd_alloc_ctl_entry(int n)
4956 struct ctl_table *entry =
4957 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4962 static void sd_free_ctl_entry(struct ctl_table **tablep)
4964 struct ctl_table *entry;
4967 * In the intermediate directories, both the child directory and
4968 * procname are dynamically allocated and could fail but the mode
4969 * will always be set. In the lowest directory the names are
4970 * static strings and all have proc handlers.
4972 for (entry = *tablep; entry->mode; entry++) {
4974 sd_free_ctl_entry(&entry->child);
4975 if (entry->proc_handler == NULL)
4976 kfree(entry->procname);
4983 static int min_load_idx = 0;
4984 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4987 set_table_entry(struct ctl_table *entry,
4988 const char *procname, void *data, int maxlen,
4989 umode_t mode, proc_handler *proc_handler,
4992 entry->procname = procname;
4994 entry->maxlen = maxlen;
4996 entry->proc_handler = proc_handler;
4999 entry->extra1 = &min_load_idx;
5000 entry->extra2 = &max_load_idx;
5004 static struct ctl_table *
5005 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5007 struct ctl_table *table = sd_alloc_ctl_entry(14);
5012 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5013 sizeof(long), 0644, proc_doulongvec_minmax, false);
5014 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5015 sizeof(long), 0644, proc_doulongvec_minmax, false);
5016 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5017 sizeof(int), 0644, proc_dointvec_minmax, true);
5018 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5019 sizeof(int), 0644, proc_dointvec_minmax, true);
5020 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5021 sizeof(int), 0644, proc_dointvec_minmax, true);
5022 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5023 sizeof(int), 0644, proc_dointvec_minmax, true);
5024 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5025 sizeof(int), 0644, proc_dointvec_minmax, true);
5026 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5027 sizeof(int), 0644, proc_dointvec_minmax, false);
5028 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5029 sizeof(int), 0644, proc_dointvec_minmax, false);
5030 set_table_entry(&table[9], "cache_nice_tries",
5031 &sd->cache_nice_tries,
5032 sizeof(int), 0644, proc_dointvec_minmax, false);
5033 set_table_entry(&table[10], "flags", &sd->flags,
5034 sizeof(int), 0644, proc_dointvec_minmax, false);
5035 set_table_entry(&table[11], "max_newidle_lb_cost",
5036 &sd->max_newidle_lb_cost,
5037 sizeof(long), 0644, proc_doulongvec_minmax, false);
5038 set_table_entry(&table[12], "name", sd->name,
5039 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5040 /* &table[13] is terminator */
5045 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5047 struct ctl_table *entry, *table;
5048 struct sched_domain *sd;
5049 int domain_num = 0, i;
5052 for_each_domain(cpu, sd)
5054 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5059 for_each_domain(cpu, sd) {
5060 snprintf(buf, 32, "domain%d", i);
5061 entry->procname = kstrdup(buf, GFP_KERNEL);
5063 entry->child = sd_alloc_ctl_domain_table(sd);
5070 static struct ctl_table_header *sd_sysctl_header;
5071 static void register_sched_domain_sysctl(void)
5073 int i, cpu_num = num_possible_cpus();
5074 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5077 WARN_ON(sd_ctl_dir[0].child);
5078 sd_ctl_dir[0].child = entry;
5083 for_each_possible_cpu(i) {
5084 snprintf(buf, 32, "cpu%d", i);
5085 entry->procname = kstrdup(buf, GFP_KERNEL);
5087 entry->child = sd_alloc_ctl_cpu_table(i);
5091 WARN_ON(sd_sysctl_header);
5092 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5095 /* may be called multiple times per register */
5096 static void unregister_sched_domain_sysctl(void)
5098 if (sd_sysctl_header)
5099 unregister_sysctl_table(sd_sysctl_header);
5100 sd_sysctl_header = NULL;
5101 if (sd_ctl_dir[0].child)
5102 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5105 static void register_sched_domain_sysctl(void)
5108 static void unregister_sched_domain_sysctl(void)
5113 static void set_rq_online(struct rq *rq)
5116 const struct sched_class *class;
5118 cpumask_set_cpu(rq->cpu, rq->rd->online);
5121 for_each_class(class) {
5122 if (class->rq_online)
5123 class->rq_online(rq);
5128 static void set_rq_offline(struct rq *rq)
5131 const struct sched_class *class;
5133 for_each_class(class) {
5134 if (class->rq_offline)
5135 class->rq_offline(rq);
5138 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5144 * migration_call - callback that gets triggered when a CPU is added.
5145 * Here we can start up the necessary migration thread for the new CPU.
5148 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5150 int cpu = (long)hcpu;
5151 unsigned long flags;
5152 struct rq *rq = cpu_rq(cpu);
5154 switch (action & ~CPU_TASKS_FROZEN) {
5156 case CPU_UP_PREPARE:
5157 rq->calc_load_update = calc_load_update;
5161 /* Update our root-domain */
5162 raw_spin_lock_irqsave(&rq->lock, flags);
5164 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5168 raw_spin_unlock_irqrestore(&rq->lock, flags);
5171 #ifdef CONFIG_HOTPLUG_CPU
5173 sched_ttwu_pending();
5174 /* Update our root-domain */
5175 raw_spin_lock_irqsave(&rq->lock, flags);
5177 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5181 BUG_ON(rq->nr_running != 1); /* the migration thread */
5182 raw_spin_unlock_irqrestore(&rq->lock, flags);
5186 calc_load_migrate(rq);
5191 update_max_interval();
5197 * Register at high priority so that task migration (migrate_all_tasks)
5198 * happens before everything else. This has to be lower priority than
5199 * the notifier in the perf_event subsystem, though.
5201 static struct notifier_block migration_notifier = {
5202 .notifier_call = migration_call,
5203 .priority = CPU_PRI_MIGRATION,
5206 static void __cpuinit set_cpu_rq_start_time(void)
5208 int cpu = smp_processor_id();
5209 struct rq *rq = cpu_rq(cpu);
5210 rq->age_stamp = sched_clock_cpu(cpu);
5213 #ifdef CONFIG_SCHED_SMT
5214 atomic_t sched_smt_present = ATOMIC_INIT(0);
5217 static int sched_cpu_active(struct notifier_block *nfb,
5218 unsigned long action, void *hcpu)
5220 switch (action & ~CPU_TASKS_FROZEN) {
5222 set_cpu_rq_start_time();
5226 * At this point a starting CPU has marked itself as online via
5227 * set_cpu_online(). But it might not yet have marked itself
5228 * as active, which is essential from here on.
5230 * Thus, fall-through and help the starting CPU along.
5232 case CPU_DOWN_FAILED:
5233 #ifdef CONFIG_SCHED_SMT
5235 * When going up, increment the number of cores with SMT present.
5237 if (cpumask_weight(cpu_smt_mask((long)hcpu)) == 2)
5238 atomic_inc(&sched_smt_present);
5240 set_cpu_active((long)hcpu, true);
5247 static int sched_cpu_inactive(struct notifier_block *nfb,
5248 unsigned long action, void *hcpu)
5250 unsigned long flags;
5251 long cpu = (long)hcpu;
5253 switch (action & ~CPU_TASKS_FROZEN) {
5254 case CPU_DOWN_PREPARE:
5255 set_cpu_active(cpu, false);
5257 #ifdef CONFIG_SCHED_SMT
5259 * When going down, decrement the number of cores with SMT present.
5261 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
5262 atomic_dec(&sched_smt_present);
5265 /* explicitly allow suspend */
5266 if (!(action & CPU_TASKS_FROZEN)) {
5267 struct dl_bw *dl_b = dl_bw_of(cpu);
5271 raw_spin_lock_irqsave(&dl_b->lock, flags);
5272 cpus = dl_bw_cpus(cpu);
5273 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5274 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5277 return notifier_from_errno(-EBUSY);
5285 static int __init migration_init(void)
5287 void *cpu = (void *)(long)smp_processor_id();
5290 /* Initialize migration for the boot CPU */
5291 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5292 BUG_ON(err == NOTIFY_BAD);
5293 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5294 register_cpu_notifier(&migration_notifier);
5296 /* Register cpu active notifiers */
5297 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5298 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5302 early_initcall(migration_init);
5307 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5308 static cpumask_var_t sched_domains_tmpmask2;
5310 #ifdef CONFIG_SCHED_DEBUG
5312 static __read_mostly int sched_debug_enabled;
5314 static int __init sched_debug_setup(char *str)
5316 sched_debug_enabled = 1;
5320 early_param("sched_debug", sched_debug_setup);
5322 static inline bool sched_debug(void)
5324 return sched_debug_enabled;
5327 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5328 struct cpumask *groupmask)
5330 struct sched_group *group = sd->groups;
5333 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5334 cpumask_clear(groupmask);
5336 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5338 if (!(sd->flags & SD_LOAD_BALANCE)) {
5339 printk("does not load-balance\n");
5341 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5346 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5348 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5349 printk(KERN_ERR "ERROR: domain->span does not contain "
5352 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5353 printk(KERN_ERR "ERROR: domain->groups does not contain"
5357 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5361 printk(KERN_ERR "ERROR: group is NULL\n");
5366 * Even though we initialize ->capacity to something semi-sane,
5367 * we leave capacity_orig unset. This allows us to detect if
5368 * domain iteration is still funny without causing /0 traps.
5370 if (!group->sgc->capacity_orig) {
5371 printk(KERN_CONT "\n");
5372 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5376 if (!cpumask_weight(sched_group_cpus(group))) {
5377 printk(KERN_CONT "\n");
5378 printk(KERN_ERR "ERROR: empty group\n");
5382 if (!(sd->flags & SD_OVERLAP) &&
5383 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5384 printk(KERN_CONT "\n");
5385 printk(KERN_ERR "ERROR: repeated CPUs\n");
5389 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5391 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5393 printk(KERN_CONT " %s", str);
5394 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5395 printk(KERN_CONT " (cpu_capacity = %d)",
5396 group->sgc->capacity);
5399 group = group->next;
5400 } while (group != sd->groups);
5401 printk(KERN_CONT "\n");
5403 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5404 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5407 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5408 printk(KERN_ERR "ERROR: parent span is not a superset "
5409 "of domain->span\n");
5413 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5417 if (!sched_debug_enabled)
5421 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5425 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5428 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5436 #else /* !CONFIG_SCHED_DEBUG */
5437 # define sched_domain_debug(sd, cpu) do { } while (0)
5438 static inline bool sched_debug(void)
5442 #endif /* CONFIG_SCHED_DEBUG */
5444 static int sd_degenerate(struct sched_domain *sd)
5446 if (cpumask_weight(sched_domain_span(sd)) == 1)
5449 /* Following flags need at least 2 groups */
5450 if (sd->flags & (SD_LOAD_BALANCE |
5451 SD_BALANCE_NEWIDLE |
5454 SD_SHARE_CPUCAPACITY |
5455 SD_SHARE_PKG_RESOURCES |
5456 SD_SHARE_POWERDOMAIN)) {
5457 if (sd->groups != sd->groups->next)
5461 /* Following flags don't use groups */
5462 if (sd->flags & (SD_WAKE_AFFINE))
5469 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5471 unsigned long cflags = sd->flags, pflags = parent->flags;
5473 if (sd_degenerate(parent))
5476 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5479 /* Flags needing groups don't count if only 1 group in parent */
5480 if (parent->groups == parent->groups->next) {
5481 pflags &= ~(SD_LOAD_BALANCE |
5482 SD_BALANCE_NEWIDLE |
5485 SD_SHARE_CPUCAPACITY |
5486 SD_SHARE_PKG_RESOURCES |
5488 SD_SHARE_POWERDOMAIN);
5489 if (nr_node_ids == 1)
5490 pflags &= ~SD_SERIALIZE;
5492 if (~cflags & pflags)
5498 static void free_rootdomain(struct rcu_head *rcu)
5500 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5502 cpupri_cleanup(&rd->cpupri);
5503 cpudl_cleanup(&rd->cpudl);
5504 free_cpumask_var(rd->dlo_mask);
5505 free_cpumask_var(rd->rto_mask);
5506 free_cpumask_var(rd->online);
5507 free_cpumask_var(rd->span);
5511 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5513 struct root_domain *old_rd = NULL;
5514 unsigned long flags;
5516 raw_spin_lock_irqsave(&rq->lock, flags);
5521 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5524 cpumask_clear_cpu(rq->cpu, old_rd->span);
5527 * If we dont want to free the old_rd yet then
5528 * set old_rd to NULL to skip the freeing later
5531 if (!atomic_dec_and_test(&old_rd->refcount))
5535 atomic_inc(&rd->refcount);
5538 cpumask_set_cpu(rq->cpu, rd->span);
5539 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5542 raw_spin_unlock_irqrestore(&rq->lock, flags);
5545 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5548 static int init_rootdomain(struct root_domain *rd)
5550 memset(rd, 0, sizeof(*rd));
5552 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5554 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5556 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5558 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5561 init_dl_bw(&rd->dl_bw);
5562 if (cpudl_init(&rd->cpudl) != 0)
5565 if (cpupri_init(&rd->cpupri) != 0)
5570 free_cpumask_var(rd->rto_mask);
5572 free_cpumask_var(rd->dlo_mask);
5574 free_cpumask_var(rd->online);
5576 free_cpumask_var(rd->span);
5582 * By default the system creates a single root-domain with all cpus as
5583 * members (mimicking the global state we have today).
5585 struct root_domain def_root_domain;
5587 static void init_defrootdomain(void)
5589 init_rootdomain(&def_root_domain);
5591 atomic_set(&def_root_domain.refcount, 1);
5594 static struct root_domain *alloc_rootdomain(void)
5596 struct root_domain *rd;
5598 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5602 if (init_rootdomain(rd) != 0) {
5610 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5612 struct sched_group *tmp, *first;
5621 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5626 } while (sg != first);
5629 static void free_sched_domain(struct rcu_head *rcu)
5631 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5634 * If its an overlapping domain it has private groups, iterate and
5637 if (sd->flags & SD_OVERLAP) {
5638 free_sched_groups(sd->groups, 1);
5639 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5640 kfree(sd->groups->sgc);
5646 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5648 call_rcu(&sd->rcu, free_sched_domain);
5651 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5653 for (; sd; sd = sd->parent)
5654 destroy_sched_domain(sd, cpu);
5658 * Keep a special pointer to the highest sched_domain that has
5659 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5660 * allows us to avoid some pointer chasing select_idle_sibling().
5662 * Also keep a unique ID per domain (we use the first cpu number in
5663 * the cpumask of the domain), this allows us to quickly tell if
5664 * two cpus are in the same cache domain, see cpus_share_cache().
5666 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5667 DEFINE_PER_CPU(int, sd_llc_size);
5668 DEFINE_PER_CPU(int, sd_llc_id);
5669 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5670 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5671 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5673 static void update_top_cache_domain(int cpu)
5675 struct sched_domain *sd;
5676 struct sched_domain *busy_sd = NULL;
5680 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5682 id = cpumask_first(sched_domain_span(sd));
5683 size = cpumask_weight(sched_domain_span(sd));
5684 busy_sd = sd->parent; /* sd_busy */
5686 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5688 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5689 per_cpu(sd_llc_size, cpu) = size;
5690 per_cpu(sd_llc_id, cpu) = id;
5692 sd = lowest_flag_domain(cpu, SD_NUMA);
5693 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5695 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5696 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5700 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5701 * hold the hotplug lock.
5704 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5706 struct rq *rq = cpu_rq(cpu);
5707 struct sched_domain *tmp;
5709 /* Remove the sched domains which do not contribute to scheduling. */
5710 for (tmp = sd; tmp; ) {
5711 struct sched_domain *parent = tmp->parent;
5715 if (sd_parent_degenerate(tmp, parent)) {
5716 tmp->parent = parent->parent;
5718 parent->parent->child = tmp;
5720 * Transfer SD_PREFER_SIBLING down in case of a
5721 * degenerate parent; the spans match for this
5722 * so the property transfers.
5724 if (parent->flags & SD_PREFER_SIBLING)
5725 tmp->flags |= SD_PREFER_SIBLING;
5726 destroy_sched_domain(parent, cpu);
5731 if (sd && sd_degenerate(sd)) {
5734 destroy_sched_domain(tmp, cpu);
5739 sched_domain_debug(sd, cpu);
5741 rq_attach_root(rq, rd);
5743 rcu_assign_pointer(rq->sd, sd);
5744 destroy_sched_domains(tmp, cpu);
5746 update_top_cache_domain(cpu);
5749 /* cpus with isolated domains */
5750 static cpumask_var_t cpu_isolated_map;
5752 /* Setup the mask of cpus configured for isolated domains */
5753 static int __init isolated_cpu_setup(char *str)
5755 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5756 cpulist_parse(str, cpu_isolated_map);
5760 __setup("isolcpus=", isolated_cpu_setup);
5763 struct sched_domain ** __percpu sd;
5764 struct root_domain *rd;
5775 * Build an iteration mask that can exclude certain CPUs from the upwards
5778 * Asymmetric node setups can result in situations where the domain tree is of
5779 * unequal depth, make sure to skip domains that already cover the entire
5782 * In that case build_sched_domains() will have terminated the iteration early
5783 * and our sibling sd spans will be empty. Domains should always include the
5784 * cpu they're built on, so check that.
5786 * Only CPUs that can arrive at this group should be considered to continue
5790 build_group_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
5792 const struct cpumask *sg_span = sched_group_cpus(sg);
5793 struct sd_data *sdd = sd->private;
5794 struct sched_domain *sibling;
5797 cpumask_clear(mask);
5799 for_each_cpu(i, sg_span) {
5800 sibling = *per_cpu_ptr(sdd->sd, i);
5803 * Can happen in the asymmetric case, where these siblings are
5804 * unused. The mask will not be empty because those CPUs that
5805 * do have the top domain _should_ span the domain.
5807 if (!sibling->child)
5810 /* If we would not end up here, we can't continue from here */
5811 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
5814 cpumask_set_cpu(i, mask);
5817 /* We must not have empty masks here */
5818 WARN_ON_ONCE(cpumask_empty(mask));
5822 * Return the canonical balance cpu for this group, this is the first cpu
5823 * of this group that's also in the iteration mask.
5825 int group_balance_cpu(struct sched_group *sg)
5827 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5830 static struct sched_group *
5831 build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
5833 struct sched_group *sg;
5834 struct cpumask *sg_span;
5836 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5837 GFP_KERNEL, cpu_to_node(cpu));
5842 sg_span = sched_group_cpus(sg);
5844 cpumask_copy(sg_span, sched_domain_span(sd->child));
5846 cpumask_copy(sg_span, sched_domain_span(sd));
5851 static void init_overlap_sched_group(struct sched_domain *sd,
5852 struct sched_group *sg)
5854 struct cpumask *mask = sched_domains_tmpmask2;
5855 struct sd_data *sdd = sd->private;
5856 struct cpumask *sg_span;
5859 build_group_mask(sd, sg, mask);
5860 cpu = cpumask_first_and(sched_group_cpus(sg), mask);
5862 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5863 if (atomic_inc_return(&sg->sgc->ref) == 1)
5864 cpumask_copy(sched_group_mask(sg), mask);
5867 * Initialize sgc->capacity such that even if we mess up the
5868 * domains and no possible iteration will get us here, we won't
5871 sg_span = sched_group_cpus(sg);
5872 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5873 sg->sgc->capacity_orig = sg->sgc->capacity;
5877 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5879 struct sched_group *first = NULL, *last = NULL, *sg;
5880 const struct cpumask *span = sched_domain_span(sd);
5881 struct cpumask *covered = sched_domains_tmpmask;
5882 struct sd_data *sdd = sd->private;
5883 struct sched_domain *sibling;
5886 cpumask_clear(covered);
5888 for_each_cpu_wrap(i, span, cpu) {
5889 struct cpumask *sg_span;
5891 if (cpumask_test_cpu(i, covered))
5894 sibling = *per_cpu_ptr(sdd->sd, i);
5896 /* See the comment near build_group_mask(). */
5897 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5900 sg = build_group_from_child_sched_domain(sibling, cpu);
5904 sg_span = sched_group_cpus(sg);
5905 cpumask_or(covered, covered, sg_span);
5907 init_overlap_sched_group(sd, sg);
5921 free_sched_groups(first, 0);
5926 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5928 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5929 struct sched_domain *child = sd->child;
5932 cpu = cpumask_first(sched_domain_span(child));
5935 *sg = *per_cpu_ptr(sdd->sg, cpu);
5936 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5937 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5944 * build_sched_groups will build a circular linked list of the groups
5945 * covered by the given span, and will set each group's ->cpumask correctly,
5946 * and ->cpu_capacity to 0.
5948 * Assumes the sched_domain tree is fully constructed
5951 build_sched_groups(struct sched_domain *sd, int cpu)
5953 struct sched_group *first = NULL, *last = NULL;
5954 struct sd_data *sdd = sd->private;
5955 const struct cpumask *span = sched_domain_span(sd);
5956 struct cpumask *covered;
5959 get_group(cpu, sdd, &sd->groups);
5960 atomic_inc(&sd->groups->ref);
5962 if (cpu != cpumask_first(span))
5965 lockdep_assert_held(&sched_domains_mutex);
5966 covered = sched_domains_tmpmask;
5968 cpumask_clear(covered);
5970 for_each_cpu(i, span) {
5971 struct sched_group *sg;
5974 if (cpumask_test_cpu(i, covered))
5977 group = get_group(i, sdd, &sg);
5978 cpumask_setall(sched_group_mask(sg));
5980 for_each_cpu(j, span) {
5981 if (get_group(j, sdd, NULL) != group)
5984 cpumask_set_cpu(j, covered);
5985 cpumask_set_cpu(j, sched_group_cpus(sg));
6000 * Initialize sched groups cpu_capacity.
6002 * cpu_capacity indicates the capacity of sched group, which is used while
6003 * distributing the load between different sched groups in a sched domain.
6004 * Typically cpu_capacity for all the groups in a sched domain will be same
6005 * unless there are asymmetries in the topology. If there are asymmetries,
6006 * group having more cpu_capacity will pickup more load compared to the
6007 * group having less cpu_capacity.
6009 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6011 struct sched_group *sg = sd->groups;
6016 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6018 } while (sg != sd->groups);
6020 if (cpu != group_balance_cpu(sg))
6023 update_group_capacity(sd, cpu);
6024 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6028 * Initializers for schedule domains
6029 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6032 static int default_relax_domain_level = -1;
6033 int sched_domain_level_max;
6035 static int __init setup_relax_domain_level(char *str)
6037 if (kstrtoint(str, 0, &default_relax_domain_level))
6038 pr_warn("Unable to set relax_domain_level\n");
6042 __setup("relax_domain_level=", setup_relax_domain_level);
6044 static void set_domain_attribute(struct sched_domain *sd,
6045 struct sched_domain_attr *attr)
6049 if (!attr || attr->relax_domain_level < 0) {
6050 if (default_relax_domain_level < 0)
6053 request = default_relax_domain_level;
6055 request = attr->relax_domain_level;
6056 if (request < sd->level) {
6057 /* turn off idle balance on this domain */
6058 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6060 /* turn on idle balance on this domain */
6061 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6065 static void __sdt_free(const struct cpumask *cpu_map);
6066 static int __sdt_alloc(const struct cpumask *cpu_map);
6068 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6069 const struct cpumask *cpu_map)
6073 if (!atomic_read(&d->rd->refcount))
6074 free_rootdomain(&d->rd->rcu); /* fall through */
6076 free_percpu(d->sd); /* fall through */
6078 __sdt_free(cpu_map); /* fall through */
6084 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6085 const struct cpumask *cpu_map)
6087 memset(d, 0, sizeof(*d));
6089 if (__sdt_alloc(cpu_map))
6090 return sa_sd_storage;
6091 d->sd = alloc_percpu(struct sched_domain *);
6093 return sa_sd_storage;
6094 d->rd = alloc_rootdomain();
6097 return sa_rootdomain;
6101 * NULL the sd_data elements we've used to build the sched_domain and
6102 * sched_group structure so that the subsequent __free_domain_allocs()
6103 * will not free the data we're using.
6105 static void claim_allocations(int cpu, struct sched_domain *sd)
6107 struct sd_data *sdd = sd->private;
6109 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6110 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6112 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6113 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6115 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6116 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6120 static int sched_domains_numa_levels;
6121 static int *sched_domains_numa_distance;
6122 static struct cpumask ***sched_domains_numa_masks;
6123 static int sched_domains_curr_level;
6127 * SD_flags allowed in topology descriptions.
6129 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6130 * SD_SHARE_PKG_RESOURCES - describes shared caches
6131 * SD_NUMA - describes NUMA topologies
6132 * SD_SHARE_POWERDOMAIN - describes shared power domain
6135 * SD_ASYM_PACKING - describes SMT quirks
6137 #define TOPOLOGY_SD_FLAGS \
6138 (SD_SHARE_CPUCAPACITY | \
6139 SD_SHARE_PKG_RESOURCES | \
6142 SD_SHARE_POWERDOMAIN)
6144 static struct sched_domain *
6145 sd_init(struct sched_domain_topology_level *tl, int cpu)
6147 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6148 int sd_weight, sd_flags = 0;
6152 * Ugly hack to pass state to sd_numa_mask()...
6154 sched_domains_curr_level = tl->numa_level;
6157 sd_weight = cpumask_weight(tl->mask(cpu));
6160 sd_flags = (*tl->sd_flags)();
6161 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6162 "wrong sd_flags in topology description\n"))
6163 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6165 *sd = (struct sched_domain){
6166 .min_interval = sd_weight,
6167 .max_interval = 2*sd_weight,
6169 .imbalance_pct = 125,
6171 .cache_nice_tries = 0,
6178 .flags = 1*SD_LOAD_BALANCE
6179 | 1*SD_BALANCE_NEWIDLE
6184 | 0*SD_SHARE_CPUCAPACITY
6185 | 0*SD_SHARE_PKG_RESOURCES
6187 | 0*SD_PREFER_SIBLING
6192 .last_balance = jiffies,
6193 .balance_interval = sd_weight,
6195 .max_newidle_lb_cost = 0,
6196 .next_decay_max_lb_cost = jiffies,
6197 #ifdef CONFIG_SCHED_DEBUG
6203 * Convert topological properties into behaviour.
6206 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6207 sd->imbalance_pct = 110;
6208 sd->smt_gain = 1178; /* ~15% */
6210 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6211 sd->imbalance_pct = 117;
6212 sd->cache_nice_tries = 1;
6216 } else if (sd->flags & SD_NUMA) {
6217 sd->cache_nice_tries = 2;
6221 sd->flags |= SD_SERIALIZE;
6222 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6223 sd->flags &= ~(SD_BALANCE_EXEC |
6230 sd->flags |= SD_PREFER_SIBLING;
6231 sd->cache_nice_tries = 1;
6236 sd->private = &tl->data;
6242 * Topology list, bottom-up.
6244 static struct sched_domain_topology_level default_topology[] = {
6245 #ifdef CONFIG_SCHED_SMT
6246 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6248 #ifdef CONFIG_SCHED_MC
6249 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6251 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6255 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6257 #define for_each_sd_topology(tl) \
6258 for (tl = sched_domain_topology; tl->mask; tl++)
6260 void set_sched_topology(struct sched_domain_topology_level *tl)
6262 sched_domain_topology = tl;
6267 static const struct cpumask *sd_numa_mask(int cpu)
6269 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6272 static void sched_numa_warn(const char *str)
6274 static int done = false;
6282 printk(KERN_WARNING "ERROR: %s\n\n", str);
6284 for (i = 0; i < nr_node_ids; i++) {
6285 printk(KERN_WARNING " ");
6286 for (j = 0; j < nr_node_ids; j++)
6287 printk(KERN_CONT "%02d ", node_distance(i,j));
6288 printk(KERN_CONT "\n");
6290 printk(KERN_WARNING "\n");
6293 static bool find_numa_distance(int distance)
6297 if (distance == node_distance(0, 0))
6300 for (i = 0; i < sched_domains_numa_levels; i++) {
6301 if (sched_domains_numa_distance[i] == distance)
6308 static void sched_init_numa(void)
6310 int next_distance, curr_distance = node_distance(0, 0);
6311 struct sched_domain_topology_level *tl;
6315 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6316 if (!sched_domains_numa_distance)
6320 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6321 * unique distances in the node_distance() table.
6323 * Assumes node_distance(0,j) includes all distances in
6324 * node_distance(i,j) in order to avoid cubic time.
6326 next_distance = curr_distance;
6327 for (i = 0; i < nr_node_ids; i++) {
6328 for (j = 0; j < nr_node_ids; j++) {
6329 for (k = 0; k < nr_node_ids; k++) {
6330 int distance = node_distance(i, k);
6332 if (distance > curr_distance &&
6333 (distance < next_distance ||
6334 next_distance == curr_distance))
6335 next_distance = distance;
6338 * While not a strong assumption it would be nice to know
6339 * about cases where if node A is connected to B, B is not
6340 * equally connected to A.
6342 if (sched_debug() && node_distance(k, i) != distance)
6343 sched_numa_warn("Node-distance not symmetric");
6345 if (sched_debug() && i && !find_numa_distance(distance))
6346 sched_numa_warn("Node-0 not representative");
6348 if (next_distance != curr_distance) {
6349 sched_domains_numa_distance[level++] = next_distance;
6350 sched_domains_numa_levels = level;
6351 curr_distance = next_distance;
6356 * In case of sched_debug() we verify the above assumption.
6362 * 'level' contains the number of unique distances, excluding the
6363 * identity distance node_distance(i,i).
6365 * The sched_domains_numa_distance[] array includes the actual distance
6370 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6371 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6372 * the array will contain less then 'level' members. This could be
6373 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6374 * in other functions.
6376 * We reset it to 'level' at the end of this function.
6378 sched_domains_numa_levels = 0;
6380 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6381 if (!sched_domains_numa_masks)
6385 * Now for each level, construct a mask per node which contains all
6386 * cpus of nodes that are that many hops away from us.
6388 for (i = 0; i < level; i++) {
6389 sched_domains_numa_masks[i] =
6390 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6391 if (!sched_domains_numa_masks[i])
6394 for (j = 0; j < nr_node_ids; j++) {
6395 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6399 sched_domains_numa_masks[i][j] = mask;
6401 for (k = 0; k < nr_node_ids; k++) {
6402 if (node_distance(j, k) > sched_domains_numa_distance[i])
6405 cpumask_or(mask, mask, cpumask_of_node(k));
6410 /* Compute default topology size */
6411 for (i = 0; sched_domain_topology[i].mask; i++);
6413 tl = kzalloc((i + level + 1) *
6414 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6419 * Copy the default topology bits..
6421 for (i = 0; sched_domain_topology[i].mask; i++)
6422 tl[i] = sched_domain_topology[i];
6425 * .. and append 'j' levels of NUMA goodness.
6427 for (j = 0; j < level; i++, j++) {
6428 tl[i] = (struct sched_domain_topology_level){
6429 .mask = sd_numa_mask,
6430 .sd_flags = cpu_numa_flags,
6431 .flags = SDTL_OVERLAP,
6437 sched_domain_topology = tl;
6439 sched_domains_numa_levels = level;
6442 static void sched_domains_numa_masks_set(int cpu)
6445 int node = cpu_to_node(cpu);
6447 for (i = 0; i < sched_domains_numa_levels; i++) {
6448 for (j = 0; j < nr_node_ids; j++) {
6449 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6450 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6455 static void sched_domains_numa_masks_clear(int cpu)
6458 for (i = 0; i < sched_domains_numa_levels; i++) {
6459 for (j = 0; j < nr_node_ids; j++)
6460 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6465 * Update sched_domains_numa_masks[level][node] array when new cpus
6468 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6469 unsigned long action,
6472 int cpu = (long)hcpu;
6474 switch (action & ~CPU_TASKS_FROZEN) {
6476 sched_domains_numa_masks_set(cpu);
6480 sched_domains_numa_masks_clear(cpu);
6490 static inline void sched_init_numa(void)
6494 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6495 unsigned long action,
6500 #endif /* CONFIG_NUMA */
6502 static int __sdt_alloc(const struct cpumask *cpu_map)
6504 struct sched_domain_topology_level *tl;
6507 for_each_sd_topology(tl) {
6508 struct sd_data *sdd = &tl->data;
6510 sdd->sd = alloc_percpu(struct sched_domain *);
6514 sdd->sg = alloc_percpu(struct sched_group *);
6518 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6522 for_each_cpu(j, cpu_map) {
6523 struct sched_domain *sd;
6524 struct sched_group *sg;
6525 struct sched_group_capacity *sgc;
6527 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6528 GFP_KERNEL, cpu_to_node(j));
6532 *per_cpu_ptr(sdd->sd, j) = sd;
6534 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6535 GFP_KERNEL, cpu_to_node(j));
6541 *per_cpu_ptr(sdd->sg, j) = sg;
6543 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6544 GFP_KERNEL, cpu_to_node(j));
6548 *per_cpu_ptr(sdd->sgc, j) = sgc;
6555 static void __sdt_free(const struct cpumask *cpu_map)
6557 struct sched_domain_topology_level *tl;
6560 for_each_sd_topology(tl) {
6561 struct sd_data *sdd = &tl->data;
6563 for_each_cpu(j, cpu_map) {
6564 struct sched_domain *sd;
6567 sd = *per_cpu_ptr(sdd->sd, j);
6568 if (sd && (sd->flags & SD_OVERLAP))
6569 free_sched_groups(sd->groups, 0);
6570 kfree(*per_cpu_ptr(sdd->sd, j));
6574 kfree(*per_cpu_ptr(sdd->sg, j));
6576 kfree(*per_cpu_ptr(sdd->sgc, j));
6578 free_percpu(sdd->sd);
6580 free_percpu(sdd->sg);
6582 free_percpu(sdd->sgc);
6587 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6588 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6589 struct sched_domain *child, int cpu)
6591 struct sched_domain *sd = sd_init(tl, cpu);
6595 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6597 sd->level = child->level + 1;
6598 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6602 set_domain_attribute(sd, attr);
6608 * Build sched domains for a given set of cpus and attach the sched domains
6609 * to the individual cpus
6611 static int build_sched_domains(const struct cpumask *cpu_map,
6612 struct sched_domain_attr *attr)
6614 enum s_alloc alloc_state;
6615 struct sched_domain *sd;
6617 int i, ret = -ENOMEM;
6619 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6620 if (alloc_state != sa_rootdomain)
6623 /* Set up domains for cpus specified by the cpu_map. */
6624 for_each_cpu(i, cpu_map) {
6625 struct sched_domain_topology_level *tl;
6628 for_each_sd_topology(tl) {
6629 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6630 if (tl == sched_domain_topology)
6631 *per_cpu_ptr(d.sd, i) = sd;
6632 if (tl->flags & SDTL_OVERLAP)
6633 sd->flags |= SD_OVERLAP;
6634 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6639 /* Build the groups for the domains */
6640 for_each_cpu(i, cpu_map) {
6641 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6642 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6643 if (sd->flags & SD_OVERLAP) {
6644 if (build_overlap_sched_groups(sd, i))
6647 if (build_sched_groups(sd, i))
6653 /* Calculate CPU capacity for physical packages and nodes */
6654 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6655 if (!cpumask_test_cpu(i, cpu_map))
6658 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6659 claim_allocations(i, sd);
6660 init_sched_groups_capacity(i, sd);
6664 /* Attach the domains */
6666 for_each_cpu(i, cpu_map) {
6667 sd = *per_cpu_ptr(d.sd, i);
6668 cpu_attach_domain(sd, d.rd, i);
6674 __free_domain_allocs(&d, alloc_state, cpu_map);
6678 static cpumask_var_t *doms_cur; /* current sched domains */
6679 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6680 static struct sched_domain_attr *dattr_cur;
6681 /* attribues of custom domains in 'doms_cur' */
6684 * Special case: If a kmalloc of a doms_cur partition (array of
6685 * cpumask) fails, then fallback to a single sched domain,
6686 * as determined by the single cpumask fallback_doms.
6688 static cpumask_var_t fallback_doms;
6691 * arch_update_cpu_topology lets virtualized architectures update the
6692 * cpu core maps. It is supposed to return 1 if the topology changed
6693 * or 0 if it stayed the same.
6695 int __weak arch_update_cpu_topology(void)
6700 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6703 cpumask_var_t *doms;
6705 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6708 for (i = 0; i < ndoms; i++) {
6709 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6710 free_sched_domains(doms, i);
6717 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6720 for (i = 0; i < ndoms; i++)
6721 free_cpumask_var(doms[i]);
6726 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6727 * For now this just excludes isolated cpus, but could be used to
6728 * exclude other special cases in the future.
6730 static int init_sched_domains(const struct cpumask *cpu_map)
6734 arch_update_cpu_topology();
6736 doms_cur = alloc_sched_domains(ndoms_cur);
6738 doms_cur = &fallback_doms;
6739 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6740 err = build_sched_domains(doms_cur[0], NULL);
6741 register_sched_domain_sysctl();
6747 * Detach sched domains from a group of cpus specified in cpu_map
6748 * These cpus will now be attached to the NULL domain
6750 static void detach_destroy_domains(const struct cpumask *cpu_map)
6755 for_each_cpu(i, cpu_map)
6756 cpu_attach_domain(NULL, &def_root_domain, i);
6760 /* handle null as "default" */
6761 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6762 struct sched_domain_attr *new, int idx_new)
6764 struct sched_domain_attr tmp;
6771 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6772 new ? (new + idx_new) : &tmp,
6773 sizeof(struct sched_domain_attr));
6777 * Partition sched domains as specified by the 'ndoms_new'
6778 * cpumasks in the array doms_new[] of cpumasks. This compares
6779 * doms_new[] to the current sched domain partitioning, doms_cur[].
6780 * It destroys each deleted domain and builds each new domain.
6782 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6783 * The masks don't intersect (don't overlap.) We should setup one
6784 * sched domain for each mask. CPUs not in any of the cpumasks will
6785 * not be load balanced. If the same cpumask appears both in the
6786 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6789 * The passed in 'doms_new' should be allocated using
6790 * alloc_sched_domains. This routine takes ownership of it and will
6791 * free_sched_domains it when done with it. If the caller failed the
6792 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6793 * and partition_sched_domains() will fallback to the single partition
6794 * 'fallback_doms', it also forces the domains to be rebuilt.
6796 * If doms_new == NULL it will be replaced with cpu_online_mask.
6797 * ndoms_new == 0 is a special case for destroying existing domains,
6798 * and it will not create the default domain.
6800 * Call with hotplug lock held
6802 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6803 struct sched_domain_attr *dattr_new)
6808 mutex_lock(&sched_domains_mutex);
6810 /* always unregister in case we don't destroy any domains */
6811 unregister_sched_domain_sysctl();
6813 /* Let architecture update cpu core mappings. */
6814 new_topology = arch_update_cpu_topology();
6816 n = doms_new ? ndoms_new : 0;
6818 /* Destroy deleted domains */
6819 for (i = 0; i < ndoms_cur; i++) {
6820 for (j = 0; j < n && !new_topology; j++) {
6821 if (cpumask_equal(doms_cur[i], doms_new[j])
6822 && dattrs_equal(dattr_cur, i, dattr_new, j))
6825 /* no match - a current sched domain not in new doms_new[] */
6826 detach_destroy_domains(doms_cur[i]);
6832 if (doms_new == NULL) {
6834 doms_new = &fallback_doms;
6835 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6836 WARN_ON_ONCE(dattr_new);
6839 /* Build new domains */
6840 for (i = 0; i < ndoms_new; i++) {
6841 for (j = 0; j < n && !new_topology; j++) {
6842 if (cpumask_equal(doms_new[i], doms_cur[j])
6843 && dattrs_equal(dattr_new, i, dattr_cur, j))
6846 /* no match - add a new doms_new */
6847 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6852 /* Remember the new sched domains */
6853 if (doms_cur != &fallback_doms)
6854 free_sched_domains(doms_cur, ndoms_cur);
6855 kfree(dattr_cur); /* kfree(NULL) is safe */
6856 doms_cur = doms_new;
6857 dattr_cur = dattr_new;
6858 ndoms_cur = ndoms_new;
6860 register_sched_domain_sysctl();
6862 mutex_unlock(&sched_domains_mutex);
6865 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6868 * Update cpusets according to cpu_active mask. If cpusets are
6869 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6870 * around partition_sched_domains().
6872 * If we come here as part of a suspend/resume, don't touch cpusets because we
6873 * want to restore it back to its original state upon resume anyway.
6875 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6879 case CPU_ONLINE_FROZEN:
6880 case CPU_DOWN_FAILED_FROZEN:
6883 * num_cpus_frozen tracks how many CPUs are involved in suspend
6884 * resume sequence. As long as this is not the last online
6885 * operation in the resume sequence, just build a single sched
6886 * domain, ignoring cpusets.
6889 if (likely(num_cpus_frozen)) {
6890 partition_sched_domains(1, NULL, NULL);
6895 * This is the last CPU online operation. So fall through and
6896 * restore the original sched domains by considering the
6897 * cpuset configurations.
6901 case CPU_DOWN_FAILED:
6902 cpuset_update_active_cpus(true);
6910 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6914 case CPU_DOWN_PREPARE:
6915 cpuset_update_active_cpus(false);
6917 case CPU_DOWN_PREPARE_FROZEN:
6919 partition_sched_domains(1, NULL, NULL);
6927 void __init sched_init_smp(void)
6929 cpumask_var_t non_isolated_cpus;
6931 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6932 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6937 * There's no userspace yet to cause hotplug operations; hence all the
6938 * cpu masks are stable and all blatant races in the below code cannot
6941 mutex_lock(&sched_domains_mutex);
6942 init_sched_domains(cpu_active_mask);
6943 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6944 if (cpumask_empty(non_isolated_cpus))
6945 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6946 mutex_unlock(&sched_domains_mutex);
6948 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6949 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6950 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6954 /* Move init over to a non-isolated CPU */
6955 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6957 sched_init_granularity();
6958 free_cpumask_var(non_isolated_cpus);
6960 init_sched_rt_class();
6961 init_sched_dl_class();
6964 void __init sched_init_smp(void)
6966 sched_init_granularity();
6968 #endif /* CONFIG_SMP */
6970 const_debug unsigned int sysctl_timer_migration = 1;
6972 int in_sched_functions(unsigned long addr)
6974 return in_lock_functions(addr) ||
6975 (addr >= (unsigned long)__sched_text_start
6976 && addr < (unsigned long)__sched_text_end);
6979 #ifdef CONFIG_CGROUP_SCHED
6981 * Default task group.
6982 * Every task in system belongs to this group at bootup.
6984 struct task_group root_task_group;
6985 LIST_HEAD(task_groups);
6988 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6990 void __init sched_init(void)
6993 unsigned long alloc_size = 0, ptr;
6995 #ifdef CONFIG_FAIR_GROUP_SCHED
6996 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6998 #ifdef CONFIG_RT_GROUP_SCHED
6999 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7001 #ifdef CONFIG_CPUMASK_OFFSTACK
7002 alloc_size += num_possible_cpus() * cpumask_size();
7005 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7007 #ifdef CONFIG_FAIR_GROUP_SCHED
7008 root_task_group.se = (struct sched_entity **)ptr;
7009 ptr += nr_cpu_ids * sizeof(void **);
7011 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7012 ptr += nr_cpu_ids * sizeof(void **);
7014 #endif /* CONFIG_FAIR_GROUP_SCHED */
7015 #ifdef CONFIG_RT_GROUP_SCHED
7016 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7017 ptr += nr_cpu_ids * sizeof(void **);
7019 root_task_group.rt_rq = (struct rt_rq **)ptr;
7020 ptr += nr_cpu_ids * sizeof(void **);
7022 #endif /* CONFIG_RT_GROUP_SCHED */
7023 #ifdef CONFIG_CPUMASK_OFFSTACK
7024 for_each_possible_cpu(i) {
7025 per_cpu(load_balance_mask, i) = (void *)ptr;
7026 ptr += cpumask_size();
7028 #endif /* CONFIG_CPUMASK_OFFSTACK */
7031 init_rt_bandwidth(&def_rt_bandwidth,
7032 global_rt_period(), global_rt_runtime());
7033 init_dl_bandwidth(&def_dl_bandwidth,
7034 global_rt_period(), global_rt_runtime());
7037 init_defrootdomain();
7040 #ifdef CONFIG_RT_GROUP_SCHED
7041 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7042 global_rt_period(), global_rt_runtime());
7043 #endif /* CONFIG_RT_GROUP_SCHED */
7045 #ifdef CONFIG_CGROUP_SCHED
7046 list_add(&root_task_group.list, &task_groups);
7047 INIT_LIST_HEAD(&root_task_group.children);
7048 INIT_LIST_HEAD(&root_task_group.siblings);
7049 autogroup_init(&init_task);
7051 #endif /* CONFIG_CGROUP_SCHED */
7053 for_each_possible_cpu(i) {
7057 raw_spin_lock_init(&rq->lock);
7059 rq->calc_load_active = 0;
7060 rq->calc_load_update = jiffies + LOAD_FREQ;
7061 init_cfs_rq(&rq->cfs);
7062 init_rt_rq(&rq->rt, rq);
7063 init_dl_rq(&rq->dl, rq);
7064 #ifdef CONFIG_FAIR_GROUP_SCHED
7065 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7066 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7068 * How much cpu bandwidth does root_task_group get?
7070 * In case of task-groups formed thr' the cgroup filesystem, it
7071 * gets 100% of the cpu resources in the system. This overall
7072 * system cpu resource is divided among the tasks of
7073 * root_task_group and its child task-groups in a fair manner,
7074 * based on each entity's (task or task-group's) weight
7075 * (se->load.weight).
7077 * In other words, if root_task_group has 10 tasks of weight
7078 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7079 * then A0's share of the cpu resource is:
7081 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7083 * We achieve this by letting root_task_group's tasks sit
7084 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7086 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7087 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7088 #endif /* CONFIG_FAIR_GROUP_SCHED */
7090 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7091 #ifdef CONFIG_RT_GROUP_SCHED
7092 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7095 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7096 rq->cpu_load[j] = 0;
7098 rq->last_load_update_tick = jiffies;
7103 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7104 rq->balance_callback = NULL;
7105 rq->active_balance = 0;
7106 rq->next_balance = jiffies;
7111 rq->avg_idle = 2*sysctl_sched_migration_cost;
7112 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7114 INIT_LIST_HEAD(&rq->cfs_tasks);
7116 rq_attach_root(rq, &def_root_domain);
7117 #ifdef CONFIG_NO_HZ_COMMON
7120 #ifdef CONFIG_NO_HZ_FULL
7121 rq->last_sched_tick = 0;
7125 atomic_set(&rq->nr_iowait, 0);
7128 set_load_weight(&init_task);
7130 #ifdef CONFIG_PREEMPT_NOTIFIERS
7131 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7135 * The boot idle thread does lazy MMU switching as well:
7137 atomic_inc(&init_mm.mm_count);
7138 enter_lazy_tlb(&init_mm, current);
7141 * Make us the idle thread. Technically, schedule() should not be
7142 * called from this thread, however somewhere below it might be,
7143 * but because we are the idle thread, we just pick up running again
7144 * when this runqueue becomes "idle".
7146 init_idle(current, smp_processor_id());
7148 calc_load_update = jiffies + LOAD_FREQ;
7151 * During early bootup we pretend to be a normal task:
7153 current->sched_class = &fair_sched_class;
7156 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7157 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_NOWAIT);
7158 /* May be allocated at isolcpus cmdline parse time */
7159 if (cpu_isolated_map == NULL)
7160 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7161 idle_thread_set_boot_cpu();
7162 set_cpu_rq_start_time();
7164 init_sched_fair_class();
7166 scheduler_running = 1;
7169 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7170 static inline int preempt_count_equals(int preempt_offset)
7172 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7174 return (nested == preempt_offset);
7177 void __might_sleep(const char *file, int line, int preempt_offset)
7179 static unsigned long prev_jiffy; /* ratelimiting */
7181 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7182 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7183 !is_idle_task(current)) ||
7184 system_state != SYSTEM_RUNNING || oops_in_progress)
7186 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7188 prev_jiffy = jiffies;
7191 "BUG: sleeping function called from invalid context at %s:%d\n",
7194 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7195 in_atomic(), irqs_disabled(),
7196 current->pid, current->comm);
7198 debug_show_held_locks(current);
7199 if (irqs_disabled())
7200 print_irqtrace_events(current);
7201 #ifdef CONFIG_DEBUG_PREEMPT
7202 if (!preempt_count_equals(preempt_offset)) {
7203 pr_err("Preemption disabled at:");
7204 print_ip_sym(current->preempt_disable_ip);
7210 EXPORT_SYMBOL(__might_sleep);
7213 #ifdef CONFIG_MAGIC_SYSRQ
7214 static void normalize_task(struct rq *rq, struct task_struct *p)
7216 const struct sched_class *prev_class = p->sched_class;
7217 struct sched_attr attr = {
7218 .sched_policy = SCHED_NORMAL,
7220 int old_prio = p->prio;
7225 dequeue_task(rq, p, 0);
7226 __setscheduler(rq, p, &attr, false);
7228 enqueue_task(rq, p, 0);
7229 resched_task(rq->curr);
7232 check_class_changed(rq, p, prev_class, old_prio);
7235 void normalize_rt_tasks(void)
7237 struct task_struct *g, *p;
7238 unsigned long flags;
7241 read_lock_irqsave(&tasklist_lock, flags);
7242 do_each_thread(g, p) {
7244 * Only normalize user tasks:
7249 p->se.exec_start = 0;
7250 #ifdef CONFIG_SCHEDSTATS
7251 p->se.statistics.wait_start = 0;
7252 p->se.statistics.sleep_start = 0;
7253 p->se.statistics.block_start = 0;
7256 if (!dl_task(p) && !rt_task(p)) {
7258 * Renice negative nice level userspace
7261 if (task_nice(p) < 0 && p->mm)
7262 set_user_nice(p, 0);
7266 raw_spin_lock(&p->pi_lock);
7267 rq = __task_rq_lock(p);
7269 normalize_task(rq, p);
7271 __task_rq_unlock(rq);
7272 raw_spin_unlock(&p->pi_lock);
7273 } while_each_thread(g, p);
7275 read_unlock_irqrestore(&tasklist_lock, flags);
7278 #endif /* CONFIG_MAGIC_SYSRQ */
7280 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7282 * These functions are only useful for the IA64 MCA handling, or kdb.
7284 * They can only be called when the whole system has been
7285 * stopped - every CPU needs to be quiescent, and no scheduling
7286 * activity can take place. Using them for anything else would
7287 * be a serious bug, and as a result, they aren't even visible
7288 * under any other configuration.
7292 * curr_task - return the current task for a given cpu.
7293 * @cpu: the processor in question.
7295 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7297 * Return: The current task for @cpu.
7299 struct task_struct *curr_task(int cpu)
7301 return cpu_curr(cpu);
7304 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7308 * set_curr_task - set the current task for a given cpu.
7309 * @cpu: the processor in question.
7310 * @p: the task pointer to set.
7312 * Description: This function must only be used when non-maskable interrupts
7313 * are serviced on a separate stack. It allows the architecture to switch the
7314 * notion of the current task on a cpu in a non-blocking manner. This function
7315 * must be called with all CPU's synchronized, and interrupts disabled, the
7316 * and caller must save the original value of the current task (see
7317 * curr_task() above) and restore that value before reenabling interrupts and
7318 * re-starting the system.
7320 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7322 void set_curr_task(int cpu, struct task_struct *p)
7329 #ifdef CONFIG_CGROUP_SCHED
7330 /* task_group_lock serializes the addition/removal of task groups */
7331 static DEFINE_SPINLOCK(task_group_lock);
7333 static void free_sched_group(struct task_group *tg)
7335 free_fair_sched_group(tg);
7336 free_rt_sched_group(tg);
7341 /* allocate runqueue etc for a new task group */
7342 struct task_group *sched_create_group(struct task_group *parent)
7344 struct task_group *tg;
7346 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7348 return ERR_PTR(-ENOMEM);
7350 if (!alloc_fair_sched_group(tg, parent))
7353 if (!alloc_rt_sched_group(tg, parent))
7359 free_sched_group(tg);
7360 return ERR_PTR(-ENOMEM);
7363 void sched_online_group(struct task_group *tg, struct task_group *parent)
7365 unsigned long flags;
7367 spin_lock_irqsave(&task_group_lock, flags);
7368 list_add_rcu(&tg->list, &task_groups);
7370 WARN_ON(!parent); /* root should already exist */
7372 tg->parent = parent;
7373 INIT_LIST_HEAD(&tg->children);
7374 list_add_rcu(&tg->siblings, &parent->children);
7375 spin_unlock_irqrestore(&task_group_lock, flags);
7378 /* rcu callback to free various structures associated with a task group */
7379 static void free_sched_group_rcu(struct rcu_head *rhp)
7381 /* now it should be safe to free those cfs_rqs */
7382 free_sched_group(container_of(rhp, struct task_group, rcu));
7385 /* Destroy runqueue etc associated with a task group */
7386 void sched_destroy_group(struct task_group *tg)
7388 /* wait for possible concurrent references to cfs_rqs complete */
7389 call_rcu(&tg->rcu, free_sched_group_rcu);
7392 void sched_offline_group(struct task_group *tg)
7394 unsigned long flags;
7397 /* end participation in shares distribution */
7398 for_each_possible_cpu(i)
7399 unregister_fair_sched_group(tg, i);
7401 spin_lock_irqsave(&task_group_lock, flags);
7402 list_del_rcu(&tg->list);
7403 list_del_rcu(&tg->siblings);
7404 spin_unlock_irqrestore(&task_group_lock, flags);
7407 /* change task's runqueue when it moves between groups.
7408 * The caller of this function should have put the task in its new group
7409 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7410 * reflect its new group.
7412 void sched_move_task(struct task_struct *tsk)
7414 struct task_group *tg;
7416 unsigned long flags;
7419 rq = task_rq_lock(tsk, &flags);
7421 running = task_current(rq, tsk);
7425 dequeue_task(rq, tsk, 0);
7426 if (unlikely(running))
7427 tsk->sched_class->put_prev_task(rq, tsk);
7429 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7430 lockdep_is_held(&tsk->sighand->siglock)),
7431 struct task_group, css);
7432 tg = autogroup_task_group(tsk, tg);
7433 tsk->sched_task_group = tg;
7435 #ifdef CONFIG_FAIR_GROUP_SCHED
7436 if (tsk->sched_class->task_move_group)
7437 tsk->sched_class->task_move_group(tsk, on_rq);
7440 set_task_rq(tsk, task_cpu(tsk));
7442 if (unlikely(running))
7443 tsk->sched_class->set_curr_task(rq);
7445 enqueue_task(rq, tsk, 0);
7447 task_rq_unlock(rq, tsk, &flags);
7449 #endif /* CONFIG_CGROUP_SCHED */
7451 #ifdef CONFIG_RT_GROUP_SCHED
7453 * Ensure that the real time constraints are schedulable.
7455 static DEFINE_MUTEX(rt_constraints_mutex);
7457 /* Must be called with tasklist_lock held */
7458 static inline int tg_has_rt_tasks(struct task_group *tg)
7460 struct task_struct *g, *p;
7463 * Autogroups do not have RT tasks; see autogroup_create().
7465 if (task_group_is_autogroup(tg))
7468 do_each_thread(g, p) {
7469 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7471 } while_each_thread(g, p);
7476 struct rt_schedulable_data {
7477 struct task_group *tg;
7482 static int tg_rt_schedulable(struct task_group *tg, void *data)
7484 struct rt_schedulable_data *d = data;
7485 struct task_group *child;
7486 unsigned long total, sum = 0;
7487 u64 period, runtime;
7489 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7490 runtime = tg->rt_bandwidth.rt_runtime;
7493 period = d->rt_period;
7494 runtime = d->rt_runtime;
7498 * Cannot have more runtime than the period.
7500 if (runtime > period && runtime != RUNTIME_INF)
7504 * Ensure we don't starve existing RT tasks.
7506 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7509 total = to_ratio(period, runtime);
7512 * Nobody can have more than the global setting allows.
7514 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7518 * The sum of our children's runtime should not exceed our own.
7520 list_for_each_entry_rcu(child, &tg->children, siblings) {
7521 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7522 runtime = child->rt_bandwidth.rt_runtime;
7524 if (child == d->tg) {
7525 period = d->rt_period;
7526 runtime = d->rt_runtime;
7529 sum += to_ratio(period, runtime);
7538 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7542 struct rt_schedulable_data data = {
7544 .rt_period = period,
7545 .rt_runtime = runtime,
7549 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7555 static int tg_set_rt_bandwidth(struct task_group *tg,
7556 u64 rt_period, u64 rt_runtime)
7560 mutex_lock(&rt_constraints_mutex);
7561 read_lock(&tasklist_lock);
7562 err = __rt_schedulable(tg, rt_period, rt_runtime);
7566 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7567 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7568 tg->rt_bandwidth.rt_runtime = rt_runtime;
7570 for_each_possible_cpu(i) {
7571 struct rt_rq *rt_rq = tg->rt_rq[i];
7573 raw_spin_lock(&rt_rq->rt_runtime_lock);
7574 rt_rq->rt_runtime = rt_runtime;
7575 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7577 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7579 read_unlock(&tasklist_lock);
7580 mutex_unlock(&rt_constraints_mutex);
7585 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7587 u64 rt_runtime, rt_period;
7589 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7590 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7591 if (rt_runtime_us < 0)
7592 rt_runtime = RUNTIME_INF;
7594 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7597 static long sched_group_rt_runtime(struct task_group *tg)
7601 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7604 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7605 do_div(rt_runtime_us, NSEC_PER_USEC);
7606 return rt_runtime_us;
7609 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7611 u64 rt_runtime, rt_period;
7613 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7614 rt_runtime = tg->rt_bandwidth.rt_runtime;
7619 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7622 static long sched_group_rt_period(struct task_group *tg)
7626 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7627 do_div(rt_period_us, NSEC_PER_USEC);
7628 return rt_period_us;
7630 #endif /* CONFIG_RT_GROUP_SCHED */
7632 #ifdef CONFIG_RT_GROUP_SCHED
7633 static int sched_rt_global_constraints(void)
7637 mutex_lock(&rt_constraints_mutex);
7638 read_lock(&tasklist_lock);
7639 ret = __rt_schedulable(NULL, 0, 0);
7640 read_unlock(&tasklist_lock);
7641 mutex_unlock(&rt_constraints_mutex);
7646 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7648 /* Don't accept realtime tasks when there is no way for them to run */
7649 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7655 #else /* !CONFIG_RT_GROUP_SCHED */
7656 static int sched_rt_global_constraints(void)
7658 unsigned long flags;
7661 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7662 for_each_possible_cpu(i) {
7663 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7665 raw_spin_lock(&rt_rq->rt_runtime_lock);
7666 rt_rq->rt_runtime = global_rt_runtime();
7667 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7669 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7673 #endif /* CONFIG_RT_GROUP_SCHED */
7675 static int sched_dl_global_constraints(void)
7677 u64 runtime = global_rt_runtime();
7678 u64 period = global_rt_period();
7679 u64 new_bw = to_ratio(period, runtime);
7681 unsigned long flags;
7686 * Here we want to check the bandwidth not being set to some
7687 * value smaller than the currently allocated bandwidth in
7688 * any of the root_domains.
7690 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7691 * cycling on root_domains... Discussion on different/better
7692 * solutions is welcome!
7694 for_each_possible_cpu(cpu) {
7695 struct dl_bw *dl_b = dl_bw_of(cpu);
7697 raw_spin_lock_irqsave(&dl_b->lock, flags);
7698 if (new_bw < dl_b->total_bw)
7700 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7711 static void sched_dl_do_global(void)
7715 unsigned long flags;
7717 def_dl_bandwidth.dl_period = global_rt_period();
7718 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7720 if (global_rt_runtime() != RUNTIME_INF)
7721 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7725 * FIXME: As above...
7727 for_each_possible_cpu(cpu) {
7728 struct dl_bw *dl_b = dl_bw_of(cpu);
7730 raw_spin_lock_irqsave(&dl_b->lock, flags);
7732 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7737 static int sched_rt_global_validate(void)
7739 if (sysctl_sched_rt_period <= 0)
7742 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7743 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7749 static void sched_rt_do_global(void)
7751 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7752 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7755 int sched_rt_handler(struct ctl_table *table, int write,
7756 void __user *buffer, size_t *lenp,
7759 int old_period, old_runtime;
7760 static DEFINE_MUTEX(mutex);
7764 old_period = sysctl_sched_rt_period;
7765 old_runtime = sysctl_sched_rt_runtime;
7767 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7769 if (!ret && write) {
7770 ret = sched_rt_global_validate();
7774 ret = sched_rt_global_constraints();
7778 ret = sched_dl_global_constraints();
7782 sched_rt_do_global();
7783 sched_dl_do_global();
7787 sysctl_sched_rt_period = old_period;
7788 sysctl_sched_rt_runtime = old_runtime;
7790 mutex_unlock(&mutex);
7795 int sched_rr_handler(struct ctl_table *table, int write,
7796 void __user *buffer, size_t *lenp,
7800 static DEFINE_MUTEX(mutex);
7803 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7804 /* make sure that internally we keep jiffies */
7805 /* also, writing zero resets timeslice to default */
7806 if (!ret && write) {
7807 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7808 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7810 mutex_unlock(&mutex);
7814 #ifdef CONFIG_CGROUP_SCHED
7816 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7818 return css ? container_of(css, struct task_group, css) : NULL;
7821 static struct cgroup_subsys_state *
7822 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7824 struct task_group *parent = css_tg(parent_css);
7825 struct task_group *tg;
7828 /* This is early initialization for the top cgroup */
7829 return &root_task_group.css;
7832 tg = sched_create_group(parent);
7834 return ERR_PTR(-ENOMEM);
7839 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7841 struct task_group *tg = css_tg(css);
7842 struct task_group *parent = css_tg(css->parent);
7845 sched_online_group(tg, parent);
7849 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7851 struct task_group *tg = css_tg(css);
7853 sched_destroy_group(tg);
7856 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7858 struct task_group *tg = css_tg(css);
7860 sched_offline_group(tg);
7863 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7864 struct cgroup_taskset *tset)
7866 struct task_struct *task;
7868 cgroup_taskset_for_each(task, tset) {
7869 #ifdef CONFIG_RT_GROUP_SCHED
7870 if (!sched_rt_can_attach(css_tg(css), task))
7873 /* We don't support RT-tasks being in separate groups */
7874 if (task->sched_class != &fair_sched_class)
7881 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7882 struct cgroup_taskset *tset)
7884 struct task_struct *task;
7886 cgroup_taskset_for_each(task, tset)
7887 sched_move_task(task);
7890 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7891 struct cgroup_subsys_state *old_css,
7892 struct task_struct *task)
7895 * cgroup_exit() is called in the copy_process() failure path.
7896 * Ignore this case since the task hasn't ran yet, this avoids
7897 * trying to poke a half freed task state from generic code.
7899 if (!(task->flags & PF_EXITING))
7902 sched_move_task(task);
7905 #ifdef CONFIG_FAIR_GROUP_SCHED
7906 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7907 struct cftype *cftype, u64 shareval)
7909 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7912 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7915 struct task_group *tg = css_tg(css);
7917 return (u64) scale_load_down(tg->shares);
7920 #ifdef CONFIG_CFS_BANDWIDTH
7921 static DEFINE_MUTEX(cfs_constraints_mutex);
7923 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7924 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7926 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7928 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7930 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7931 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7933 if (tg == &root_task_group)
7937 * Ensure we have at some amount of bandwidth every period. This is
7938 * to prevent reaching a state of large arrears when throttled via
7939 * entity_tick() resulting in prolonged exit starvation.
7941 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7945 * Likewise, bound things on the otherside by preventing insane quota
7946 * periods. This also allows us to normalize in computing quota
7949 if (period > max_cfs_quota_period)
7952 mutex_lock(&cfs_constraints_mutex);
7953 ret = __cfs_schedulable(tg, period, quota);
7957 runtime_enabled = quota != RUNTIME_INF;
7958 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7960 * If we need to toggle cfs_bandwidth_used, off->on must occur
7961 * before making related changes, and on->off must occur afterwards
7963 if (runtime_enabled && !runtime_was_enabled)
7964 cfs_bandwidth_usage_inc();
7965 raw_spin_lock_irq(&cfs_b->lock);
7966 cfs_b->period = ns_to_ktime(period);
7967 cfs_b->quota = quota;
7969 __refill_cfs_bandwidth_runtime(cfs_b);
7970 /* restart the period timer (if active) to handle new period expiry */
7971 if (runtime_enabled && cfs_b->timer_active) {
7972 /* force a reprogram */
7973 __start_cfs_bandwidth(cfs_b, true);
7975 raw_spin_unlock_irq(&cfs_b->lock);
7977 for_each_possible_cpu(i) {
7978 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7979 struct rq *rq = cfs_rq->rq;
7981 raw_spin_lock_irq(&rq->lock);
7982 cfs_rq->runtime_enabled = runtime_enabled;
7983 cfs_rq->runtime_remaining = 0;
7985 if (cfs_rq->throttled)
7986 unthrottle_cfs_rq(cfs_rq);
7987 raw_spin_unlock_irq(&rq->lock);
7989 if (runtime_was_enabled && !runtime_enabled)
7990 cfs_bandwidth_usage_dec();
7992 mutex_unlock(&cfs_constraints_mutex);
7997 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8001 period = ktime_to_ns(tg->cfs_bandwidth.period);
8002 if (cfs_quota_us < 0)
8003 quota = RUNTIME_INF;
8005 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8007 return tg_set_cfs_bandwidth(tg, period, quota);
8010 long tg_get_cfs_quota(struct task_group *tg)
8014 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8017 quota_us = tg->cfs_bandwidth.quota;
8018 do_div(quota_us, NSEC_PER_USEC);
8023 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8027 period = (u64)cfs_period_us * NSEC_PER_USEC;
8028 quota = tg->cfs_bandwidth.quota;
8030 return tg_set_cfs_bandwidth(tg, period, quota);
8033 long tg_get_cfs_period(struct task_group *tg)
8037 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8038 do_div(cfs_period_us, NSEC_PER_USEC);
8040 return cfs_period_us;
8043 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8046 return tg_get_cfs_quota(css_tg(css));
8049 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8050 struct cftype *cftype, s64 cfs_quota_us)
8052 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8055 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8058 return tg_get_cfs_period(css_tg(css));
8061 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8062 struct cftype *cftype, u64 cfs_period_us)
8064 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8067 struct cfs_schedulable_data {
8068 struct task_group *tg;
8073 * normalize group quota/period to be quota/max_period
8074 * note: units are usecs
8076 static u64 normalize_cfs_quota(struct task_group *tg,
8077 struct cfs_schedulable_data *d)
8085 period = tg_get_cfs_period(tg);
8086 quota = tg_get_cfs_quota(tg);
8089 /* note: these should typically be equivalent */
8090 if (quota == RUNTIME_INF || quota == -1)
8093 return to_ratio(period, quota);
8096 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8098 struct cfs_schedulable_data *d = data;
8099 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8100 s64 quota = 0, parent_quota = -1;
8103 quota = RUNTIME_INF;
8105 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8107 quota = normalize_cfs_quota(tg, d);
8108 parent_quota = parent_b->hierarchal_quota;
8111 * ensure max(child_quota) <= parent_quota, inherit when no
8114 if (quota == RUNTIME_INF)
8115 quota = parent_quota;
8116 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8119 cfs_b->hierarchal_quota = quota;
8124 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8127 struct cfs_schedulable_data data = {
8133 if (quota != RUNTIME_INF) {
8134 do_div(data.period, NSEC_PER_USEC);
8135 do_div(data.quota, NSEC_PER_USEC);
8139 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8145 static int cpu_stats_show(struct seq_file *sf, void *v)
8147 struct task_group *tg = css_tg(seq_css(sf));
8148 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8150 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8151 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8152 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8156 #endif /* CONFIG_CFS_BANDWIDTH */
8157 #endif /* CONFIG_FAIR_GROUP_SCHED */
8159 #ifdef CONFIG_RT_GROUP_SCHED
8160 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8161 struct cftype *cft, s64 val)
8163 return sched_group_set_rt_runtime(css_tg(css), val);
8166 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8169 return sched_group_rt_runtime(css_tg(css));
8172 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8173 struct cftype *cftype, u64 rt_period_us)
8175 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8178 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8181 return sched_group_rt_period(css_tg(css));
8183 #endif /* CONFIG_RT_GROUP_SCHED */
8185 static struct cftype cpu_files[] = {
8186 #ifdef CONFIG_FAIR_GROUP_SCHED
8189 .read_u64 = cpu_shares_read_u64,
8190 .write_u64 = cpu_shares_write_u64,
8193 #ifdef CONFIG_CFS_BANDWIDTH
8195 .name = "cfs_quota_us",
8196 .read_s64 = cpu_cfs_quota_read_s64,
8197 .write_s64 = cpu_cfs_quota_write_s64,
8200 .name = "cfs_period_us",
8201 .read_u64 = cpu_cfs_period_read_u64,
8202 .write_u64 = cpu_cfs_period_write_u64,
8206 .seq_show = cpu_stats_show,
8209 #ifdef CONFIG_RT_GROUP_SCHED
8211 .name = "rt_runtime_us",
8212 .read_s64 = cpu_rt_runtime_read,
8213 .write_s64 = cpu_rt_runtime_write,
8216 .name = "rt_period_us",
8217 .read_u64 = cpu_rt_period_read_uint,
8218 .write_u64 = cpu_rt_period_write_uint,
8224 struct cgroup_subsys cpu_cgrp_subsys = {
8225 .css_alloc = cpu_cgroup_css_alloc,
8226 .css_free = cpu_cgroup_css_free,
8227 .css_online = cpu_cgroup_css_online,
8228 .css_offline = cpu_cgroup_css_offline,
8229 .can_attach = cpu_cgroup_can_attach,
8230 .attach = cpu_cgroup_attach,
8231 .exit = cpu_cgroup_exit,
8232 .base_cftypes = cpu_files,
8236 #endif /* CONFIG_CGROUP_SCHED */
8238 void dump_cpu_task(int cpu)
8240 pr_info("Task dump for CPU %d:\n", cpu);
8241 sched_show_task(cpu_curr(cpu));