4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
96 static void update_rq_clock_task(struct rq *rq, s64 delta);
98 void update_rq_clock(struct rq *rq)
102 lockdep_assert_held(&rq->lock);
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
111 update_rq_clock_task(rq, delta);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug unsigned int sysctl_sched_features =
122 #include "features.h"
127 #ifdef CONFIG_SCHED_DEBUG
128 #define SCHED_FEAT(name, enabled) \
131 static const char * const sched_feat_names[] = {
132 #include "features.h"
137 static int sched_feat_show(struct seq_file *m, void *v)
141 for (i = 0; i < __SCHED_FEAT_NR; i++) {
142 if (!(sysctl_sched_features & (1UL << i)))
144 seq_printf(m, "%s ", sched_feat_names[i]);
151 #ifdef HAVE_JUMP_LABEL
153 #define jump_label_key__true STATIC_KEY_INIT_TRUE
154 #define jump_label_key__false STATIC_KEY_INIT_FALSE
156 #define SCHED_FEAT(name, enabled) \
157 jump_label_key__##enabled ,
159 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
160 #include "features.h"
165 static void sched_feat_disable(int i)
167 static_key_disable(&sched_feat_keys[i]);
170 static void sched_feat_enable(int i)
172 static_key_enable(&sched_feat_keys[i]);
175 static void sched_feat_disable(int i) { };
176 static void sched_feat_enable(int i) { };
177 #endif /* HAVE_JUMP_LABEL */
179 static int sched_feat_set(char *cmp)
184 if (strncmp(cmp, "NO_", 3) == 0) {
189 for (i = 0; i < __SCHED_FEAT_NR; i++) {
190 if (strcmp(cmp, sched_feat_names[i]) == 0) {
192 sysctl_sched_features &= ~(1UL << i);
193 sched_feat_disable(i);
195 sysctl_sched_features |= (1UL << i);
196 sched_feat_enable(i);
206 sched_feat_write(struct file *filp, const char __user *ubuf,
207 size_t cnt, loff_t *ppos)
217 if (copy_from_user(&buf, ubuf, cnt))
223 /* Ensure the static_key remains in a consistent state */
224 inode = file_inode(filp);
225 mutex_lock(&inode->i_mutex);
226 i = sched_feat_set(cmp);
227 mutex_unlock(&inode->i_mutex);
228 if (i == __SCHED_FEAT_NR)
236 static int sched_feat_open(struct inode *inode, struct file *filp)
238 return single_open(filp, sched_feat_show, NULL);
241 static const struct file_operations sched_feat_fops = {
242 .open = sched_feat_open,
243 .write = sched_feat_write,
246 .release = single_release,
249 static __init int sched_init_debug(void)
251 debugfs_create_file("sched_features", 0644, NULL, NULL,
256 late_initcall(sched_init_debug);
257 #endif /* CONFIG_SCHED_DEBUG */
260 * Number of tasks to iterate in a single balance run.
261 * Limited because this is done with IRQs disabled.
263 const_debug unsigned int sysctl_sched_nr_migrate = 32;
266 * period over which we average the RT time consumption, measured
271 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
274 * period over which we measure -rt task cpu usage in us.
277 unsigned int sysctl_sched_rt_period = 1000000;
279 __read_mostly int scheduler_running;
282 * part of the period that we allow rt tasks to run in us.
285 int sysctl_sched_rt_runtime = 950000;
287 /* cpus with isolated domains */
288 cpumask_var_t cpu_isolated_map;
291 lock_rq_of(struct task_struct *p, unsigned long *flags)
293 return task_rq_lock(p, flags);
297 unlock_rq_of(struct rq *rq, struct task_struct *p, unsigned long *flags)
299 task_rq_unlock(rq, p, flags);
303 * this_rq_lock - lock this runqueue and disable interrupts.
305 static struct rq *this_rq_lock(void)
312 raw_spin_lock(&rq->lock);
317 #ifdef CONFIG_SCHED_HRTICK
319 * Use HR-timers to deliver accurate preemption points.
322 static void hrtick_clear(struct rq *rq)
324 if (hrtimer_active(&rq->hrtick_timer))
325 hrtimer_cancel(&rq->hrtick_timer);
329 * High-resolution timer tick.
330 * Runs from hardirq context with interrupts disabled.
332 static enum hrtimer_restart hrtick(struct hrtimer *timer)
334 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
336 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
338 raw_spin_lock(&rq->lock);
340 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
341 raw_spin_unlock(&rq->lock);
343 return HRTIMER_NORESTART;
348 static void __hrtick_restart(struct rq *rq)
350 struct hrtimer *timer = &rq->hrtick_timer;
352 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
356 * called from hardirq (IPI) context
358 static void __hrtick_start(void *arg)
362 raw_spin_lock(&rq->lock);
363 __hrtick_restart(rq);
364 rq->hrtick_csd_pending = 0;
365 raw_spin_unlock(&rq->lock);
369 * Called to set the hrtick timer state.
371 * called with rq->lock held and irqs disabled
373 void hrtick_start(struct rq *rq, u64 delay)
375 struct hrtimer *timer = &rq->hrtick_timer;
380 * Don't schedule slices shorter than 10000ns, that just
381 * doesn't make sense and can cause timer DoS.
383 delta = max_t(s64, delay, 10000LL);
384 time = ktime_add_ns(timer->base->get_time(), delta);
386 hrtimer_set_expires(timer, time);
388 if (rq == this_rq()) {
389 __hrtick_restart(rq);
390 } else if (!rq->hrtick_csd_pending) {
391 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
392 rq->hrtick_csd_pending = 1;
397 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
399 int cpu = (int)(long)hcpu;
402 case CPU_UP_CANCELED:
403 case CPU_UP_CANCELED_FROZEN:
404 case CPU_DOWN_PREPARE:
405 case CPU_DOWN_PREPARE_FROZEN:
407 case CPU_DEAD_FROZEN:
408 hrtick_clear(cpu_rq(cpu));
415 static __init void init_hrtick(void)
417 hotcpu_notifier(hotplug_hrtick, 0);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq *rq, u64 delay)
428 * Don't schedule slices shorter than 10000ns, that just
429 * doesn't make sense. Rely on vruntime for fairness.
431 delay = max_t(u64, delay, 10000LL);
432 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
433 HRTIMER_MODE_REL_PINNED);
436 static inline void init_hrtick(void)
439 #endif /* CONFIG_SMP */
441 static void init_rq_hrtick(struct rq *rq)
444 rq->hrtick_csd_pending = 0;
446 rq->hrtick_csd.flags = 0;
447 rq->hrtick_csd.func = __hrtick_start;
448 rq->hrtick_csd.info = rq;
451 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
452 rq->hrtick_timer.function = hrtick;
454 #else /* CONFIG_SCHED_HRTICK */
455 static inline void hrtick_clear(struct rq *rq)
459 static inline void init_rq_hrtick(struct rq *rq)
463 static inline void init_hrtick(void)
466 #endif /* CONFIG_SCHED_HRTICK */
469 * cmpxchg based fetch_or, macro so it works for different integer types
471 #define fetch_or(ptr, val) \
472 ({ typeof(*(ptr)) __old, __val = *(ptr); \
474 __old = cmpxchg((ptr), __val, __val | (val)); \
475 if (__old == __val) \
482 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
484 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
485 * this avoids any races wrt polling state changes and thereby avoids
488 static bool set_nr_and_not_polling(struct task_struct *p)
490 struct thread_info *ti = task_thread_info(p);
491 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
495 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
497 * If this returns true, then the idle task promises to call
498 * sched_ttwu_pending() and reschedule soon.
500 static bool set_nr_if_polling(struct task_struct *p)
502 struct thread_info *ti = task_thread_info(p);
503 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
506 if (!(val & _TIF_POLLING_NRFLAG))
508 if (val & _TIF_NEED_RESCHED)
510 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
519 static bool set_nr_and_not_polling(struct task_struct *p)
521 set_tsk_need_resched(p);
526 static bool set_nr_if_polling(struct task_struct *p)
533 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
535 struct wake_q_node *node = &task->wake_q;
538 * Atomically grab the task, if ->wake_q is !nil already it means
539 * its already queued (either by us or someone else) and will get the
540 * wakeup due to that.
542 * This cmpxchg() implies a full barrier, which pairs with the write
543 * barrier implied by the wakeup in wake_up_list().
545 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
548 get_task_struct(task);
551 * The head is context local, there can be no concurrency.
554 head->lastp = &node->next;
557 void wake_up_q(struct wake_q_head *head)
559 struct wake_q_node *node = head->first;
561 while (node != WAKE_Q_TAIL) {
562 struct task_struct *task;
564 task = container_of(node, struct task_struct, wake_q);
566 /* task can safely be re-inserted now */
568 task->wake_q.next = NULL;
571 * wake_up_process() implies a wmb() to pair with the queueing
572 * in wake_q_add() so as not to miss wakeups.
574 wake_up_process(task);
575 put_task_struct(task);
580 * resched_curr - mark rq's current task 'to be rescheduled now'.
582 * On UP this means the setting of the need_resched flag, on SMP it
583 * might also involve a cross-CPU call to trigger the scheduler on
586 void resched_curr(struct rq *rq)
588 struct task_struct *curr = rq->curr;
591 lockdep_assert_held(&rq->lock);
593 if (test_tsk_need_resched(curr))
598 if (cpu == smp_processor_id()) {
599 set_tsk_need_resched(curr);
600 set_preempt_need_resched();
604 if (set_nr_and_not_polling(curr))
605 smp_send_reschedule(cpu);
607 trace_sched_wake_idle_without_ipi(cpu);
610 void resched_cpu(int cpu)
612 struct rq *rq = cpu_rq(cpu);
615 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
618 raw_spin_unlock_irqrestore(&rq->lock, flags);
622 #ifdef CONFIG_NO_HZ_COMMON
624 * In the semi idle case, use the nearest busy cpu for migrating timers
625 * from an idle cpu. This is good for power-savings.
627 * We don't do similar optimization for completely idle system, as
628 * selecting an idle cpu will add more delays to the timers than intended
629 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
631 int get_nohz_timer_target(void)
633 int i, cpu = smp_processor_id();
634 struct sched_domain *sd;
636 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
640 for_each_domain(cpu, sd) {
641 for_each_cpu(i, sched_domain_span(sd)) {
642 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
649 if (!is_housekeeping_cpu(cpu))
650 cpu = housekeeping_any_cpu();
656 * When add_timer_on() enqueues a timer into the timer wheel of an
657 * idle CPU then this timer might expire before the next timer event
658 * which is scheduled to wake up that CPU. In case of a completely
659 * idle system the next event might even be infinite time into the
660 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
661 * leaves the inner idle loop so the newly added timer is taken into
662 * account when the CPU goes back to idle and evaluates the timer
663 * wheel for the next timer event.
665 static void wake_up_idle_cpu(int cpu)
667 struct rq *rq = cpu_rq(cpu);
669 if (cpu == smp_processor_id())
672 if (set_nr_and_not_polling(rq->idle))
673 smp_send_reschedule(cpu);
675 trace_sched_wake_idle_without_ipi(cpu);
678 static bool wake_up_full_nohz_cpu(int cpu)
681 * We just need the target to call irq_exit() and re-evaluate
682 * the next tick. The nohz full kick at least implies that.
683 * If needed we can still optimize that later with an
686 if (tick_nohz_full_cpu(cpu)) {
687 if (cpu != smp_processor_id() ||
688 tick_nohz_tick_stopped())
689 tick_nohz_full_kick_cpu(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)
733 * FIFO realtime policy runs the highest priority task. Other runnable
734 * tasks are of a lower priority. The scheduler tick does nothing.
736 if (current->policy == SCHED_FIFO)
740 * Round-robin realtime tasks time slice with other tasks at the same
741 * realtime priority. Is this task the only one at this priority?
743 if (current->policy == SCHED_RR) {
744 struct sched_rt_entity *rt_se = ¤t->rt;
746 return rt_se->run_list.prev == rt_se->run_list.next;
750 * More than one running task need preemption.
751 * nr_running update is assumed to be visible
752 * after IPI is sent from wakers.
754 if (this_rq()->nr_running > 1)
759 #endif /* CONFIG_NO_HZ_FULL */
761 void sched_avg_update(struct rq *rq)
763 s64 period = sched_avg_period();
765 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
767 * Inline assembly required to prevent the compiler
768 * optimising this loop into a divmod call.
769 * See __iter_div_u64_rem() for another example of this.
771 asm("" : "+rm" (rq->age_stamp));
772 rq->age_stamp += period;
777 #endif /* CONFIG_SMP */
779 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
780 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
782 * Iterate task_group tree rooted at *from, calling @down when first entering a
783 * node and @up when leaving it for the final time.
785 * Caller must hold rcu_lock or sufficient equivalent.
787 int walk_tg_tree_from(struct task_group *from,
788 tg_visitor down, tg_visitor up, void *data)
790 struct task_group *parent, *child;
796 ret = (*down)(parent, data);
799 list_for_each_entry_rcu(child, &parent->children, siblings) {
806 ret = (*up)(parent, data);
807 if (ret || parent == from)
811 parent = parent->parent;
818 int tg_nop(struct task_group *tg, void *data)
824 static void set_load_weight(struct task_struct *p)
826 int prio = p->static_prio - MAX_RT_PRIO;
827 struct load_weight *load = &p->se.load;
830 * SCHED_IDLE tasks get minimal weight:
832 if (idle_policy(p->policy)) {
833 load->weight = scale_load(WEIGHT_IDLEPRIO);
834 load->inv_weight = WMULT_IDLEPRIO;
838 load->weight = scale_load(prio_to_weight[prio]);
839 load->inv_weight = prio_to_wmult[prio];
842 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
845 if (!(flags & ENQUEUE_RESTORE))
846 sched_info_queued(rq, p);
847 p->sched_class->enqueue_task(rq, p, flags);
850 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
853 if (!(flags & DEQUEUE_SAVE))
854 sched_info_dequeued(rq, p);
855 p->sched_class->dequeue_task(rq, p, flags);
858 void activate_task(struct rq *rq, struct task_struct *p, int flags)
860 if (task_contributes_to_load(p))
861 rq->nr_uninterruptible--;
863 enqueue_task(rq, p, flags);
866 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
868 if (task_contributes_to_load(p))
869 rq->nr_uninterruptible++;
871 dequeue_task(rq, p, flags);
874 static void update_rq_clock_task(struct rq *rq, s64 delta)
877 * In theory, the compile should just see 0 here, and optimize out the call
878 * to sched_rt_avg_update. But I don't trust it...
880 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
881 s64 steal = 0, irq_delta = 0;
883 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
884 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
887 * Since irq_time is only updated on {soft,}irq_exit, we might run into
888 * this case when a previous update_rq_clock() happened inside a
891 * When this happens, we stop ->clock_task and only update the
892 * prev_irq_time stamp to account for the part that fit, so that a next
893 * update will consume the rest. This ensures ->clock_task is
896 * It does however cause some slight miss-attribution of {soft,}irq
897 * time, a more accurate solution would be to update the irq_time using
898 * the current rq->clock timestamp, except that would require using
901 if (irq_delta > delta)
904 rq->prev_irq_time += irq_delta;
907 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
908 if (static_key_false((¶virt_steal_rq_enabled))) {
909 steal = paravirt_steal_clock(cpu_of(rq));
910 steal -= rq->prev_steal_time_rq;
912 if (unlikely(steal > delta))
915 rq->prev_steal_time_rq += steal;
920 rq->clock_task += delta;
922 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
923 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
924 sched_rt_avg_update(rq, irq_delta + steal);
928 void sched_set_stop_task(int cpu, struct task_struct *stop)
930 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
931 struct task_struct *old_stop = cpu_rq(cpu)->stop;
935 * Make it appear like a SCHED_FIFO task, its something
936 * userspace knows about and won't get confused about.
938 * Also, it will make PI more or less work without too
939 * much confusion -- but then, stop work should not
940 * rely on PI working anyway.
942 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
944 stop->sched_class = &stop_sched_class;
947 cpu_rq(cpu)->stop = stop;
951 * Reset it back to a normal scheduling class so that
952 * it can die in pieces.
954 old_stop->sched_class = &rt_sched_class;
959 * __normal_prio - return the priority that is based on the static prio
961 static inline int __normal_prio(struct task_struct *p)
963 return p->static_prio;
967 * Calculate the expected normal priority: i.e. priority
968 * without taking RT-inheritance into account. Might be
969 * boosted by interactivity modifiers. Changes upon fork,
970 * setprio syscalls, and whenever the interactivity
971 * estimator recalculates.
973 static inline int normal_prio(struct task_struct *p)
977 if (task_has_dl_policy(p))
978 prio = MAX_DL_PRIO-1;
979 else if (task_has_rt_policy(p))
980 prio = MAX_RT_PRIO-1 - p->rt_priority;
982 prio = __normal_prio(p);
987 * Calculate the current priority, i.e. the priority
988 * taken into account by the scheduler. This value might
989 * be boosted by RT tasks, or might be boosted by
990 * interactivity modifiers. Will be RT if the task got
991 * RT-boosted. If not then it returns p->normal_prio.
993 static int effective_prio(struct task_struct *p)
995 p->normal_prio = normal_prio(p);
997 * If we are RT tasks or we were boosted to RT priority,
998 * keep the priority unchanged. Otherwise, update priority
999 * to the normal priority:
1001 if (!rt_prio(p->prio))
1002 return p->normal_prio;
1007 * task_curr - is this task currently executing on a CPU?
1008 * @p: the task in question.
1010 * Return: 1 if the task is currently executing. 0 otherwise.
1012 inline int task_curr(const struct task_struct *p)
1014 return cpu_curr(task_cpu(p)) == p;
1018 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1019 * use the balance_callback list if you want balancing.
1021 * this means any call to check_class_changed() must be followed by a call to
1022 * balance_callback().
1024 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1025 const struct sched_class *prev_class,
1028 if (prev_class != p->sched_class) {
1029 if (prev_class->switched_from)
1030 prev_class->switched_from(rq, p);
1032 p->sched_class->switched_to(rq, p);
1033 } else if (oldprio != p->prio || dl_task(p))
1034 p->sched_class->prio_changed(rq, p, oldprio);
1037 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1039 const struct sched_class *class;
1041 if (p->sched_class == rq->curr->sched_class) {
1042 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1044 for_each_class(class) {
1045 if (class == rq->curr->sched_class)
1047 if (class == p->sched_class) {
1055 * A queue event has occurred, and we're going to schedule. In
1056 * this case, we can save a useless back to back clock update.
1058 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1059 rq_clock_skip_update(rq, true);
1064 * This is how migration works:
1066 * 1) we invoke migration_cpu_stop() on the target CPU using
1068 * 2) stopper starts to run (implicitly forcing the migrated thread
1070 * 3) it checks whether the migrated task is still in the wrong runqueue.
1071 * 4) if it's in the wrong runqueue then the migration thread removes
1072 * it and puts it into the right queue.
1073 * 5) stopper completes and stop_one_cpu() returns and the migration
1078 * move_queued_task - move a queued task to new rq.
1080 * Returns (locked) new rq. Old rq's lock is released.
1082 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1084 lockdep_assert_held(&rq->lock);
1086 dequeue_task(rq, p, 0);
1087 p->on_rq = TASK_ON_RQ_MIGRATING;
1088 set_task_cpu(p, new_cpu);
1089 raw_spin_unlock(&rq->lock);
1091 rq = cpu_rq(new_cpu);
1093 raw_spin_lock(&rq->lock);
1094 BUG_ON(task_cpu(p) != new_cpu);
1095 p->on_rq = TASK_ON_RQ_QUEUED;
1096 enqueue_task(rq, p, 0);
1097 check_preempt_curr(rq, p, 0);
1102 struct migration_arg {
1103 struct task_struct *task;
1108 * Move (not current) task off this cpu, onto dest cpu. We're doing
1109 * this because either it can't run here any more (set_cpus_allowed()
1110 * away from this CPU, or CPU going down), or because we're
1111 * attempting to rebalance this task on exec (sched_exec).
1113 * So we race with normal scheduler movements, but that's OK, as long
1114 * as the task is no longer on this CPU.
1116 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1118 if (unlikely(!cpu_active(dest_cpu)))
1121 /* Affinity changed (again). */
1122 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1125 rq = move_queued_task(rq, p, dest_cpu);
1131 * migration_cpu_stop - this will be executed by a highprio stopper thread
1132 * and performs thread migration by bumping thread off CPU then
1133 * 'pushing' onto another runqueue.
1135 static int migration_cpu_stop(void *data)
1137 struct migration_arg *arg = data;
1138 struct task_struct *p = arg->task;
1139 struct rq *rq = this_rq();
1142 * The original target cpu might have gone down and we might
1143 * be on another cpu but it doesn't matter.
1145 local_irq_disable();
1147 * We need to explicitly wake pending tasks before running
1148 * __migrate_task() such that we will not miss enforcing cpus_allowed
1149 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1151 sched_ttwu_pending();
1153 raw_spin_lock(&p->pi_lock);
1154 raw_spin_lock(&rq->lock);
1156 * If task_rq(p) != rq, it cannot be migrated here, because we're
1157 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1158 * we're holding p->pi_lock.
1160 if (task_rq(p) == rq && task_on_rq_queued(p))
1161 rq = __migrate_task(rq, p, arg->dest_cpu);
1162 raw_spin_unlock(&rq->lock);
1163 raw_spin_unlock(&p->pi_lock);
1170 * sched_class::set_cpus_allowed must do the below, but is not required to
1171 * actually call this function.
1173 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1175 cpumask_copy(&p->cpus_allowed, new_mask);
1176 p->nr_cpus_allowed = cpumask_weight(new_mask);
1179 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1181 struct rq *rq = task_rq(p);
1182 bool queued, running;
1184 lockdep_assert_held(&p->pi_lock);
1186 queued = task_on_rq_queued(p);
1187 running = task_current(rq, p);
1191 * Because __kthread_bind() calls this on blocked tasks without
1194 lockdep_assert_held(&rq->lock);
1195 dequeue_task(rq, p, DEQUEUE_SAVE);
1198 put_prev_task(rq, p);
1200 p->sched_class->set_cpus_allowed(p, new_mask);
1203 p->sched_class->set_curr_task(rq);
1205 enqueue_task(rq, p, ENQUEUE_RESTORE);
1209 * Change a given task's CPU affinity. Migrate the thread to a
1210 * proper CPU and schedule it away if the CPU it's executing on
1211 * is removed from the allowed bitmask.
1213 * NOTE: the caller must have a valid reference to the task, the
1214 * task must not exit() & deallocate itself prematurely. The
1215 * call is not atomic; no spinlocks may be held.
1217 static int __set_cpus_allowed_ptr(struct task_struct *p,
1218 const struct cpumask *new_mask, bool check)
1220 unsigned long flags;
1222 unsigned int dest_cpu;
1225 rq = task_rq_lock(p, &flags);
1228 * Must re-check here, to close a race against __kthread_bind(),
1229 * sched_setaffinity() is not guaranteed to observe the flag.
1231 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1236 if (cpumask_equal(&p->cpus_allowed, new_mask))
1239 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1244 do_set_cpus_allowed(p, new_mask);
1246 /* Can the task run on the task's current CPU? If so, we're done */
1247 if (cpumask_test_cpu(task_cpu(p), new_mask))
1250 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1251 if (task_running(rq, p) || p->state == TASK_WAKING) {
1252 struct migration_arg arg = { p, dest_cpu };
1253 /* Need help from migration thread: drop lock and wait. */
1254 task_rq_unlock(rq, p, &flags);
1255 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1256 tlb_migrate_finish(p->mm);
1258 } else if (task_on_rq_queued(p)) {
1260 * OK, since we're going to drop the lock immediately
1261 * afterwards anyway.
1263 lockdep_unpin_lock(&rq->lock);
1264 rq = move_queued_task(rq, p, dest_cpu);
1265 lockdep_pin_lock(&rq->lock);
1268 task_rq_unlock(rq, p, &flags);
1273 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1275 return __set_cpus_allowed_ptr(p, new_mask, false);
1277 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1279 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1281 #ifdef CONFIG_SCHED_DEBUG
1283 * We should never call set_task_cpu() on a blocked task,
1284 * ttwu() will sort out the placement.
1286 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1289 #ifdef CONFIG_LOCKDEP
1291 * The caller should hold either p->pi_lock or rq->lock, when changing
1292 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1294 * sched_move_task() holds both and thus holding either pins the cgroup,
1297 * Furthermore, all task_rq users should acquire both locks, see
1300 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1301 lockdep_is_held(&task_rq(p)->lock)));
1305 trace_sched_migrate_task(p, new_cpu);
1307 if (task_cpu(p) != new_cpu) {
1308 if (p->sched_class->migrate_task_rq)
1309 p->sched_class->migrate_task_rq(p);
1310 p->se.nr_migrations++;
1311 perf_event_task_migrate(p);
1314 __set_task_cpu(p, new_cpu);
1317 static void __migrate_swap_task(struct task_struct *p, int cpu)
1319 if (task_on_rq_queued(p)) {
1320 struct rq *src_rq, *dst_rq;
1322 src_rq = task_rq(p);
1323 dst_rq = cpu_rq(cpu);
1325 deactivate_task(src_rq, p, 0);
1326 set_task_cpu(p, cpu);
1327 activate_task(dst_rq, p, 0);
1328 check_preempt_curr(dst_rq, p, 0);
1331 * Task isn't running anymore; make it appear like we migrated
1332 * it before it went to sleep. This means on wakeup we make the
1333 * previous cpu our targer instead of where it really is.
1339 struct migration_swap_arg {
1340 struct task_struct *src_task, *dst_task;
1341 int src_cpu, dst_cpu;
1344 static int migrate_swap_stop(void *data)
1346 struct migration_swap_arg *arg = data;
1347 struct rq *src_rq, *dst_rq;
1350 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1353 src_rq = cpu_rq(arg->src_cpu);
1354 dst_rq = cpu_rq(arg->dst_cpu);
1356 double_raw_lock(&arg->src_task->pi_lock,
1357 &arg->dst_task->pi_lock);
1358 double_rq_lock(src_rq, dst_rq);
1360 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1363 if (task_cpu(arg->src_task) != arg->src_cpu)
1366 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1369 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1372 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1373 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1378 double_rq_unlock(src_rq, dst_rq);
1379 raw_spin_unlock(&arg->dst_task->pi_lock);
1380 raw_spin_unlock(&arg->src_task->pi_lock);
1386 * Cross migrate two tasks
1388 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1390 struct migration_swap_arg arg;
1393 arg = (struct migration_swap_arg){
1395 .src_cpu = task_cpu(cur),
1397 .dst_cpu = task_cpu(p),
1400 if (arg.src_cpu == arg.dst_cpu)
1404 * These three tests are all lockless; this is OK since all of them
1405 * will be re-checked with proper locks held further down the line.
1407 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1410 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1413 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1416 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1417 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1424 * wait_task_inactive - wait for a thread to unschedule.
1426 * If @match_state is nonzero, it's the @p->state value just checked and
1427 * not expected to change. If it changes, i.e. @p might have woken up,
1428 * then return zero. When we succeed in waiting for @p to be off its CPU,
1429 * we return a positive number (its total switch count). If a second call
1430 * a short while later returns the same number, the caller can be sure that
1431 * @p has remained unscheduled the whole time.
1433 * The caller must ensure that the task *will* unschedule sometime soon,
1434 * else this function might spin for a *long* time. This function can't
1435 * be called with interrupts off, or it may introduce deadlock with
1436 * smp_call_function() if an IPI is sent by the same process we are
1437 * waiting to become inactive.
1439 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1441 unsigned long flags;
1442 int running, queued;
1448 * We do the initial early heuristics without holding
1449 * any task-queue locks at all. We'll only try to get
1450 * the runqueue lock when things look like they will
1456 * If the task is actively running on another CPU
1457 * still, just relax and busy-wait without holding
1460 * NOTE! Since we don't hold any locks, it's not
1461 * even sure that "rq" stays as the right runqueue!
1462 * But we don't care, since "task_running()" will
1463 * return false if the runqueue has changed and p
1464 * is actually now running somewhere else!
1466 while (task_running(rq, p)) {
1467 if (match_state && unlikely(p->state != match_state))
1473 * Ok, time to look more closely! We need the rq
1474 * lock now, to be *sure*. If we're wrong, we'll
1475 * just go back and repeat.
1477 rq = task_rq_lock(p, &flags);
1478 trace_sched_wait_task(p);
1479 running = task_running(rq, p);
1480 queued = task_on_rq_queued(p);
1482 if (!match_state || p->state == match_state)
1483 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1484 task_rq_unlock(rq, p, &flags);
1487 * If it changed from the expected state, bail out now.
1489 if (unlikely(!ncsw))
1493 * Was it really running after all now that we
1494 * checked with the proper locks actually held?
1496 * Oops. Go back and try again..
1498 if (unlikely(running)) {
1504 * It's not enough that it's not actively running,
1505 * it must be off the runqueue _entirely_, and not
1508 * So if it was still runnable (but just not actively
1509 * running right now), it's preempted, and we should
1510 * yield - it could be a while.
1512 if (unlikely(queued)) {
1513 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1515 set_current_state(TASK_UNINTERRUPTIBLE);
1516 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1521 * Ahh, all good. It wasn't running, and it wasn't
1522 * runnable, which means that it will never become
1523 * running in the future either. We're all done!
1532 * kick_process - kick a running thread to enter/exit the kernel
1533 * @p: the to-be-kicked thread
1535 * Cause a process which is running on another CPU to enter
1536 * kernel-mode, without any delay. (to get signals handled.)
1538 * NOTE: this function doesn't have to take the runqueue lock,
1539 * because all it wants to ensure is that the remote task enters
1540 * the kernel. If the IPI races and the task has been migrated
1541 * to another CPU then no harm is done and the purpose has been
1544 void kick_process(struct task_struct *p)
1550 if ((cpu != smp_processor_id()) && task_curr(p))
1551 smp_send_reschedule(cpu);
1554 EXPORT_SYMBOL_GPL(kick_process);
1557 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1559 static int select_fallback_rq(int cpu, struct task_struct *p)
1561 int nid = cpu_to_node(cpu);
1562 const struct cpumask *nodemask = NULL;
1563 enum { cpuset, possible, fail } state = cpuset;
1567 * If the node that the cpu is on has been offlined, cpu_to_node()
1568 * will return -1. There is no cpu on the node, and we should
1569 * select the cpu on the other node.
1572 nodemask = cpumask_of_node(nid);
1574 /* Look for allowed, online CPU in same node. */
1575 for_each_cpu(dest_cpu, nodemask) {
1576 if (!cpu_online(dest_cpu))
1578 if (!cpu_active(dest_cpu))
1580 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1586 /* Any allowed, online CPU? */
1587 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1588 if (!cpu_online(dest_cpu))
1590 if (!cpu_active(dest_cpu))
1595 /* No more Mr. Nice Guy. */
1598 if (IS_ENABLED(CONFIG_CPUSETS)) {
1599 cpuset_cpus_allowed_fallback(p);
1605 do_set_cpus_allowed(p, cpu_possible_mask);
1616 if (state != cpuset) {
1618 * Don't tell them about moving exiting tasks or
1619 * kernel threads (both mm NULL), since they never
1622 if (p->mm && printk_ratelimit()) {
1623 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1624 task_pid_nr(p), p->comm, cpu);
1632 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1635 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1637 lockdep_assert_held(&p->pi_lock);
1639 if (p->nr_cpus_allowed > 1)
1640 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1643 * In order not to call set_task_cpu() on a blocking task we need
1644 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1647 * Since this is common to all placement strategies, this lives here.
1649 * [ this allows ->select_task() to simply return task_cpu(p) and
1650 * not worry about this generic constraint ]
1652 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1654 cpu = select_fallback_rq(task_cpu(p), p);
1659 static void update_avg(u64 *avg, u64 sample)
1661 s64 diff = sample - *avg;
1667 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1668 const struct cpumask *new_mask, bool check)
1670 return set_cpus_allowed_ptr(p, new_mask);
1673 #endif /* CONFIG_SMP */
1676 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1678 #ifdef CONFIG_SCHEDSTATS
1679 struct rq *rq = this_rq();
1682 int this_cpu = smp_processor_id();
1684 if (cpu == this_cpu) {
1685 schedstat_inc(rq, ttwu_local);
1686 schedstat_inc(p, se.statistics.nr_wakeups_local);
1688 struct sched_domain *sd;
1690 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1692 for_each_domain(this_cpu, sd) {
1693 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1694 schedstat_inc(sd, ttwu_wake_remote);
1701 if (wake_flags & WF_MIGRATED)
1702 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1704 #endif /* CONFIG_SMP */
1706 schedstat_inc(rq, ttwu_count);
1707 schedstat_inc(p, se.statistics.nr_wakeups);
1709 if (wake_flags & WF_SYNC)
1710 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1712 #endif /* CONFIG_SCHEDSTATS */
1715 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1717 activate_task(rq, p, en_flags);
1718 p->on_rq = TASK_ON_RQ_QUEUED;
1720 /* if a worker is waking up, notify workqueue */
1721 if (p->flags & PF_WQ_WORKER)
1722 wq_worker_waking_up(p, cpu_of(rq));
1726 * Mark the task runnable and perform wakeup-preemption.
1729 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1731 check_preempt_curr(rq, p, wake_flags);
1732 p->state = TASK_RUNNING;
1733 trace_sched_wakeup(p);
1736 if (p->sched_class->task_woken) {
1738 * Our task @p is fully woken up and running; so its safe to
1739 * drop the rq->lock, hereafter rq is only used for statistics.
1741 lockdep_unpin_lock(&rq->lock);
1742 p->sched_class->task_woken(rq, p);
1743 lockdep_pin_lock(&rq->lock);
1746 if (rq->idle_stamp) {
1747 u64 delta = rq_clock(rq) - rq->idle_stamp;
1748 u64 max = 2*rq->max_idle_balance_cost;
1750 update_avg(&rq->avg_idle, delta);
1752 if (rq->avg_idle > max)
1761 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1763 lockdep_assert_held(&rq->lock);
1766 if (p->sched_contributes_to_load)
1767 rq->nr_uninterruptible--;
1770 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1771 ttwu_do_wakeup(rq, p, wake_flags);
1775 * Called in case the task @p isn't fully descheduled from its runqueue,
1776 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1777 * since all we need to do is flip p->state to TASK_RUNNING, since
1778 * the task is still ->on_rq.
1780 static int ttwu_remote(struct task_struct *p, int wake_flags)
1785 rq = __task_rq_lock(p);
1786 if (task_on_rq_queued(p)) {
1787 /* check_preempt_curr() may use rq clock */
1788 update_rq_clock(rq);
1789 ttwu_do_wakeup(rq, p, wake_flags);
1792 __task_rq_unlock(rq);
1798 void sched_ttwu_pending(void)
1800 struct rq *rq = this_rq();
1801 struct llist_node *llist = llist_del_all(&rq->wake_list);
1802 struct task_struct *p;
1803 unsigned long flags;
1808 raw_spin_lock_irqsave(&rq->lock, flags);
1809 lockdep_pin_lock(&rq->lock);
1812 p = llist_entry(llist, struct task_struct, wake_entry);
1813 llist = llist_next(llist);
1814 ttwu_do_activate(rq, p, 0);
1817 lockdep_unpin_lock(&rq->lock);
1818 raw_spin_unlock_irqrestore(&rq->lock, flags);
1821 void scheduler_ipi(void)
1824 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1825 * TIF_NEED_RESCHED remotely (for the first time) will also send
1828 preempt_fold_need_resched();
1830 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1834 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1835 * traditionally all their work was done from the interrupt return
1836 * path. Now that we actually do some work, we need to make sure
1839 * Some archs already do call them, luckily irq_enter/exit nest
1842 * Arguably we should visit all archs and update all handlers,
1843 * however a fair share of IPIs are still resched only so this would
1844 * somewhat pessimize the simple resched case.
1847 sched_ttwu_pending();
1850 * Check if someone kicked us for doing the nohz idle load balance.
1852 if (unlikely(got_nohz_idle_kick())) {
1853 this_rq()->idle_balance = 1;
1854 raise_softirq_irqoff(SCHED_SOFTIRQ);
1859 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1861 struct rq *rq = cpu_rq(cpu);
1863 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1864 if (!set_nr_if_polling(rq->idle))
1865 smp_send_reschedule(cpu);
1867 trace_sched_wake_idle_without_ipi(cpu);
1871 void wake_up_if_idle(int cpu)
1873 struct rq *rq = cpu_rq(cpu);
1874 unsigned long flags;
1878 if (!is_idle_task(rcu_dereference(rq->curr)))
1881 if (set_nr_if_polling(rq->idle)) {
1882 trace_sched_wake_idle_without_ipi(cpu);
1884 raw_spin_lock_irqsave(&rq->lock, flags);
1885 if (is_idle_task(rq->curr))
1886 smp_send_reschedule(cpu);
1887 /* Else cpu is not in idle, do nothing here */
1888 raw_spin_unlock_irqrestore(&rq->lock, flags);
1895 bool cpus_share_cache(int this_cpu, int that_cpu)
1897 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1899 #endif /* CONFIG_SMP */
1901 static void ttwu_queue(struct task_struct *p, int cpu)
1903 struct rq *rq = cpu_rq(cpu);
1905 #if defined(CONFIG_SMP)
1906 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1907 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1908 ttwu_queue_remote(p, cpu);
1913 raw_spin_lock(&rq->lock);
1914 lockdep_pin_lock(&rq->lock);
1915 ttwu_do_activate(rq, p, 0);
1916 lockdep_unpin_lock(&rq->lock);
1917 raw_spin_unlock(&rq->lock);
1921 * try_to_wake_up - wake up a thread
1922 * @p: the thread to be awakened
1923 * @state: the mask of task states that can be woken
1924 * @wake_flags: wake modifier flags (WF_*)
1926 * Put it on the run-queue if it's not already there. The "current"
1927 * thread is always on the run-queue (except when the actual
1928 * re-schedule is in progress), and as such you're allowed to do
1929 * the simpler "current->state = TASK_RUNNING" to mark yourself
1930 * runnable without the overhead of this.
1932 * Return: %true if @p was woken up, %false if it was already running.
1933 * or @state didn't match @p's state.
1936 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1938 unsigned long flags;
1939 int cpu, success = 0;
1942 * If we are going to wake up a thread waiting for CONDITION we
1943 * need to ensure that CONDITION=1 done by the caller can not be
1944 * reordered with p->state check below. This pairs with mb() in
1945 * set_current_state() the waiting thread does.
1947 smp_mb__before_spinlock();
1948 raw_spin_lock_irqsave(&p->pi_lock, flags);
1949 if (!(p->state & state))
1952 trace_sched_waking(p);
1954 success = 1; /* we're going to change ->state */
1957 if (p->on_rq && ttwu_remote(p, wake_flags))
1962 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1963 * possible to, falsely, observe p->on_cpu == 0.
1965 * One must be running (->on_cpu == 1) in order to remove oneself
1966 * from the runqueue.
1968 * [S] ->on_cpu = 1; [L] ->on_rq
1972 * [S] ->on_rq = 0; [L] ->on_cpu
1974 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1975 * from the consecutive calls to schedule(); the first switching to our
1976 * task, the second putting it to sleep.
1981 * If the owning (remote) cpu is still in the middle of schedule() with
1982 * this task as prev, wait until its done referencing the task.
1987 * Combined with the control dependency above, we have an effective
1988 * smp_load_acquire() without the need for full barriers.
1990 * Pairs with the smp_store_release() in finish_lock_switch().
1992 * This ensures that tasks getting woken will be fully ordered against
1993 * their previous state and preserve Program Order.
1997 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1998 p->state = TASK_WAKING;
2000 if (p->sched_class->task_waking)
2001 p->sched_class->task_waking(p);
2003 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2004 if (task_cpu(p) != cpu) {
2005 wake_flags |= WF_MIGRATED;
2006 set_task_cpu(p, cpu);
2008 #endif /* CONFIG_SMP */
2012 ttwu_stat(p, cpu, wake_flags);
2014 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2020 * try_to_wake_up_local - try to wake up a local task with rq lock held
2021 * @p: the thread to be awakened
2023 * Put @p on the run-queue if it's not already there. The caller must
2024 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2027 static void try_to_wake_up_local(struct task_struct *p)
2029 struct rq *rq = task_rq(p);
2031 if (WARN_ON_ONCE(rq != this_rq()) ||
2032 WARN_ON_ONCE(p == current))
2035 lockdep_assert_held(&rq->lock);
2037 if (!raw_spin_trylock(&p->pi_lock)) {
2039 * This is OK, because current is on_cpu, which avoids it being
2040 * picked for load-balance and preemption/IRQs are still
2041 * disabled avoiding further scheduler activity on it and we've
2042 * not yet picked a replacement task.
2044 lockdep_unpin_lock(&rq->lock);
2045 raw_spin_unlock(&rq->lock);
2046 raw_spin_lock(&p->pi_lock);
2047 raw_spin_lock(&rq->lock);
2048 lockdep_pin_lock(&rq->lock);
2051 if (!(p->state & TASK_NORMAL))
2054 trace_sched_waking(p);
2056 if (!task_on_rq_queued(p))
2057 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2059 ttwu_do_wakeup(rq, p, 0);
2060 ttwu_stat(p, smp_processor_id(), 0);
2062 raw_spin_unlock(&p->pi_lock);
2066 * wake_up_process - Wake up a specific process
2067 * @p: The process to be woken up.
2069 * Attempt to wake up the nominated process and move it to the set of runnable
2072 * Return: 1 if the process was woken up, 0 if it was already running.
2074 * It may be assumed that this function implies a write memory barrier before
2075 * changing the task state if and only if any tasks are woken up.
2077 int wake_up_process(struct task_struct *p)
2079 return try_to_wake_up(p, TASK_NORMAL, 0);
2081 EXPORT_SYMBOL(wake_up_process);
2083 int wake_up_state(struct task_struct *p, unsigned int state)
2085 return try_to_wake_up(p, state, 0);
2089 * This function clears the sched_dl_entity static params.
2091 void __dl_clear_params(struct task_struct *p)
2093 struct sched_dl_entity *dl_se = &p->dl;
2095 dl_se->dl_runtime = 0;
2096 dl_se->dl_deadline = 0;
2097 dl_se->dl_period = 0;
2101 dl_se->dl_throttled = 0;
2103 dl_se->dl_yielded = 0;
2107 * Perform scheduler related setup for a newly forked process p.
2108 * p is forked by current.
2110 * __sched_fork() is basic setup used by init_idle() too:
2112 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2117 p->se.exec_start = 0;
2118 p->se.sum_exec_runtime = 0;
2119 p->se.prev_sum_exec_runtime = 0;
2120 p->se.nr_migrations = 0;
2122 INIT_LIST_HEAD(&p->se.group_node);
2124 #ifdef CONFIG_SCHEDSTATS
2125 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2128 RB_CLEAR_NODE(&p->dl.rb_node);
2129 init_dl_task_timer(&p->dl);
2130 __dl_clear_params(p);
2132 INIT_LIST_HEAD(&p->rt.run_list);
2134 #ifdef CONFIG_PREEMPT_NOTIFIERS
2135 INIT_HLIST_HEAD(&p->preempt_notifiers);
2138 #ifdef CONFIG_NUMA_BALANCING
2139 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2140 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2141 p->mm->numa_scan_seq = 0;
2144 if (clone_flags & CLONE_VM)
2145 p->numa_preferred_nid = current->numa_preferred_nid;
2147 p->numa_preferred_nid = -1;
2149 p->node_stamp = 0ULL;
2150 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2151 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2152 p->numa_work.next = &p->numa_work;
2153 p->numa_faults = NULL;
2154 p->last_task_numa_placement = 0;
2155 p->last_sum_exec_runtime = 0;
2157 p->numa_group = NULL;
2158 #endif /* CONFIG_NUMA_BALANCING */
2161 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2163 #ifdef CONFIG_NUMA_BALANCING
2165 void set_numabalancing_state(bool enabled)
2168 static_branch_enable(&sched_numa_balancing);
2170 static_branch_disable(&sched_numa_balancing);
2173 #ifdef CONFIG_PROC_SYSCTL
2174 int sysctl_numa_balancing(struct ctl_table *table, int write,
2175 void __user *buffer, size_t *lenp, loff_t *ppos)
2179 int state = static_branch_likely(&sched_numa_balancing);
2181 if (write && !capable(CAP_SYS_ADMIN))
2186 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2190 set_numabalancing_state(state);
2197 * fork()/clone()-time setup:
2199 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2201 unsigned long flags;
2202 int cpu = get_cpu();
2204 __sched_fork(clone_flags, p);
2206 * We mark the process as running here. This guarantees that
2207 * nobody will actually run it, and a signal or other external
2208 * event cannot wake it up and insert it on the runqueue either.
2210 p->state = TASK_RUNNING;
2213 * Make sure we do not leak PI boosting priority to the child.
2215 p->prio = current->normal_prio;
2218 * Revert to default priority/policy on fork if requested.
2220 if (unlikely(p->sched_reset_on_fork)) {
2221 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2222 p->policy = SCHED_NORMAL;
2223 p->static_prio = NICE_TO_PRIO(0);
2225 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2226 p->static_prio = NICE_TO_PRIO(0);
2228 p->prio = p->normal_prio = __normal_prio(p);
2232 * We don't need the reset flag anymore after the fork. It has
2233 * fulfilled its duty:
2235 p->sched_reset_on_fork = 0;
2238 if (dl_prio(p->prio)) {
2241 } else if (rt_prio(p->prio)) {
2242 p->sched_class = &rt_sched_class;
2244 p->sched_class = &fair_sched_class;
2247 if (p->sched_class->task_fork)
2248 p->sched_class->task_fork(p);
2251 * The child is not yet in the pid-hash so no cgroup attach races,
2252 * and the cgroup is pinned to this child due to cgroup_fork()
2253 * is ran before sched_fork().
2255 * Silence PROVE_RCU.
2257 raw_spin_lock_irqsave(&p->pi_lock, flags);
2258 set_task_cpu(p, cpu);
2259 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2261 #ifdef CONFIG_SCHED_INFO
2262 if (likely(sched_info_on()))
2263 memset(&p->sched_info, 0, sizeof(p->sched_info));
2265 #if defined(CONFIG_SMP)
2268 init_task_preempt_count(p);
2270 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2271 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2278 unsigned long to_ratio(u64 period, u64 runtime)
2280 if (runtime == RUNTIME_INF)
2284 * Doing this here saves a lot of checks in all
2285 * the calling paths, and returning zero seems
2286 * safe for them anyway.
2291 return div64_u64(runtime << 20, period);
2295 inline struct dl_bw *dl_bw_of(int i)
2297 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2298 "sched RCU must be held");
2299 return &cpu_rq(i)->rd->dl_bw;
2302 static inline int dl_bw_cpus(int i)
2304 struct root_domain *rd = cpu_rq(i)->rd;
2307 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2308 "sched RCU must be held");
2309 for_each_cpu_and(i, rd->span, cpu_active_mask)
2315 inline struct dl_bw *dl_bw_of(int i)
2317 return &cpu_rq(i)->dl.dl_bw;
2320 static inline int dl_bw_cpus(int i)
2327 * We must be sure that accepting a new task (or allowing changing the
2328 * parameters of an existing one) is consistent with the bandwidth
2329 * constraints. If yes, this function also accordingly updates the currently
2330 * allocated bandwidth to reflect the new situation.
2332 * This function is called while holding p's rq->lock.
2334 * XXX we should delay bw change until the task's 0-lag point, see
2337 static int dl_overflow(struct task_struct *p, int policy,
2338 const struct sched_attr *attr)
2341 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2342 u64 period = attr->sched_period ?: attr->sched_deadline;
2343 u64 runtime = attr->sched_runtime;
2344 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2347 if (new_bw == p->dl.dl_bw)
2351 * Either if a task, enters, leave, or stays -deadline but changes
2352 * its parameters, we may need to update accordingly the total
2353 * allocated bandwidth of the container.
2355 raw_spin_lock(&dl_b->lock);
2356 cpus = dl_bw_cpus(task_cpu(p));
2357 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2358 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2359 __dl_add(dl_b, new_bw);
2361 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2362 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2363 __dl_clear(dl_b, p->dl.dl_bw);
2364 __dl_add(dl_b, new_bw);
2366 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2367 __dl_clear(dl_b, p->dl.dl_bw);
2370 raw_spin_unlock(&dl_b->lock);
2375 extern void init_dl_bw(struct dl_bw *dl_b);
2378 * wake_up_new_task - wake up a newly created task for the first time.
2380 * This function will do some initial scheduler statistics housekeeping
2381 * that must be done for every newly created context, then puts the task
2382 * on the runqueue and wakes it.
2384 void wake_up_new_task(struct task_struct *p)
2386 unsigned long flags;
2389 raw_spin_lock_irqsave(&p->pi_lock, flags);
2390 /* Initialize new task's runnable average */
2391 init_entity_runnable_average(&p->se);
2394 * Fork balancing, do it here and not earlier because:
2395 * - cpus_allowed can change in the fork path
2396 * - any previously selected cpu might disappear through hotplug
2398 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2401 rq = __task_rq_lock(p);
2402 activate_task(rq, p, ENQUEUE_WAKEUP_NEW);
2403 p->on_rq = TASK_ON_RQ_QUEUED;
2404 trace_sched_wakeup_new(p);
2405 check_preempt_curr(rq, p, WF_FORK);
2407 if (p->sched_class->task_woken) {
2409 * Nothing relies on rq->lock after this, so its fine to
2412 lockdep_unpin_lock(&rq->lock);
2413 p->sched_class->task_woken(rq, p);
2414 lockdep_pin_lock(&rq->lock);
2417 task_rq_unlock(rq, p, &flags);
2420 #ifdef CONFIG_PREEMPT_NOTIFIERS
2422 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2424 void preempt_notifier_inc(void)
2426 static_key_slow_inc(&preempt_notifier_key);
2428 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2430 void preempt_notifier_dec(void)
2432 static_key_slow_dec(&preempt_notifier_key);
2434 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2437 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2438 * @notifier: notifier struct to register
2440 void preempt_notifier_register(struct preempt_notifier *notifier)
2442 if (!static_key_false(&preempt_notifier_key))
2443 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2445 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2447 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2450 * preempt_notifier_unregister - no longer interested in preemption notifications
2451 * @notifier: notifier struct to unregister
2453 * This is *not* safe to call from within a preemption notifier.
2455 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2457 hlist_del(¬ifier->link);
2459 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2461 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2463 struct preempt_notifier *notifier;
2465 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2466 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2469 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2471 if (static_key_false(&preempt_notifier_key))
2472 __fire_sched_in_preempt_notifiers(curr);
2476 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2477 struct task_struct *next)
2479 struct preempt_notifier *notifier;
2481 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2482 notifier->ops->sched_out(notifier, next);
2485 static __always_inline void
2486 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2487 struct task_struct *next)
2489 if (static_key_false(&preempt_notifier_key))
2490 __fire_sched_out_preempt_notifiers(curr, next);
2493 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2495 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2500 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2501 struct task_struct *next)
2505 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2508 * prepare_task_switch - prepare to switch tasks
2509 * @rq: the runqueue preparing to switch
2510 * @prev: the current task that is being switched out
2511 * @next: the task we are going to switch to.
2513 * This is called with the rq lock held and interrupts off. It must
2514 * be paired with a subsequent finish_task_switch after the context
2517 * prepare_task_switch sets up locking and calls architecture specific
2521 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2522 struct task_struct *next)
2524 sched_info_switch(rq, prev, next);
2525 perf_event_task_sched_out(prev, next);
2526 fire_sched_out_preempt_notifiers(prev, next);
2527 prepare_lock_switch(rq, next);
2528 prepare_arch_switch(next);
2532 * finish_task_switch - clean up after a task-switch
2533 * @prev: the thread we just switched away from.
2535 * finish_task_switch must be called after the context switch, paired
2536 * with a prepare_task_switch call before the context switch.
2537 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2538 * and do any other architecture-specific cleanup actions.
2540 * Note that we may have delayed dropping an mm in context_switch(). If
2541 * so, we finish that here outside of the runqueue lock. (Doing it
2542 * with the lock held can cause deadlocks; see schedule() for
2545 * The context switch have flipped the stack from under us and restored the
2546 * local variables which were saved when this task called schedule() in the
2547 * past. prev == current is still correct but we need to recalculate this_rq
2548 * because prev may have moved to another CPU.
2550 static struct rq *finish_task_switch(struct task_struct *prev)
2551 __releases(rq->lock)
2553 struct rq *rq = this_rq();
2554 struct mm_struct *mm = rq->prev_mm;
2558 * The previous task will have left us with a preempt_count of 2
2559 * because it left us after:
2562 * preempt_disable(); // 1
2564 * raw_spin_lock_irq(&rq->lock) // 2
2566 * Also, see FORK_PREEMPT_COUNT.
2568 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2569 "corrupted preempt_count: %s/%d/0x%x\n",
2570 current->comm, current->pid, preempt_count()))
2571 preempt_count_set(FORK_PREEMPT_COUNT);
2576 * A task struct has one reference for the use as "current".
2577 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2578 * schedule one last time. The schedule call will never return, and
2579 * the scheduled task must drop that reference.
2581 * We must observe prev->state before clearing prev->on_cpu (in
2582 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2583 * running on another CPU and we could rave with its RUNNING -> DEAD
2584 * transition, resulting in a double drop.
2586 prev_state = prev->state;
2587 vtime_task_switch(prev);
2588 perf_event_task_sched_in(prev, current);
2589 finish_lock_switch(rq, prev);
2590 finish_arch_post_lock_switch();
2592 fire_sched_in_preempt_notifiers(current);
2595 if (unlikely(prev_state == TASK_DEAD)) {
2596 if (prev->sched_class->task_dead)
2597 prev->sched_class->task_dead(prev);
2600 * Remove function-return probe instances associated with this
2601 * task and put them back on the free list.
2603 kprobe_flush_task(prev);
2604 put_task_struct(prev);
2607 tick_nohz_task_switch();
2613 /* rq->lock is NOT held, but preemption is disabled */
2614 static void __balance_callback(struct rq *rq)
2616 struct callback_head *head, *next;
2617 void (*func)(struct rq *rq);
2618 unsigned long flags;
2620 raw_spin_lock_irqsave(&rq->lock, flags);
2621 head = rq->balance_callback;
2622 rq->balance_callback = NULL;
2624 func = (void (*)(struct rq *))head->func;
2631 raw_spin_unlock_irqrestore(&rq->lock, flags);
2634 static inline void balance_callback(struct rq *rq)
2636 if (unlikely(rq->balance_callback))
2637 __balance_callback(rq);
2642 static inline void balance_callback(struct rq *rq)
2649 * schedule_tail - first thing a freshly forked thread must call.
2650 * @prev: the thread we just switched away from.
2652 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2653 __releases(rq->lock)
2658 * New tasks start with FORK_PREEMPT_COUNT, see there and
2659 * finish_task_switch() for details.
2661 * finish_task_switch() will drop rq->lock() and lower preempt_count
2662 * and the preempt_enable() will end up enabling preemption (on
2663 * PREEMPT_COUNT kernels).
2666 rq = finish_task_switch(prev);
2667 balance_callback(rq);
2670 if (current->set_child_tid)
2671 put_user(task_pid_vnr(current), current->set_child_tid);
2675 * context_switch - switch to the new MM and the new thread's register state.
2677 static inline struct rq *
2678 context_switch(struct rq *rq, struct task_struct *prev,
2679 struct task_struct *next)
2681 struct mm_struct *mm, *oldmm;
2683 prepare_task_switch(rq, prev, next);
2686 oldmm = prev->active_mm;
2688 * For paravirt, this is coupled with an exit in switch_to to
2689 * combine the page table reload and the switch backend into
2692 arch_start_context_switch(prev);
2695 next->active_mm = oldmm;
2696 atomic_inc(&oldmm->mm_count);
2697 enter_lazy_tlb(oldmm, next);
2699 switch_mm(oldmm, mm, next);
2702 prev->active_mm = NULL;
2703 rq->prev_mm = oldmm;
2706 * Since the runqueue lock will be released by the next
2707 * task (which is an invalid locking op but in the case
2708 * of the scheduler it's an obvious special-case), so we
2709 * do an early lockdep release here:
2711 lockdep_unpin_lock(&rq->lock);
2712 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2714 /* Here we just switch the register state and the stack. */
2715 switch_to(prev, next, prev);
2718 return finish_task_switch(prev);
2722 * nr_running and nr_context_switches:
2724 * externally visible scheduler statistics: current number of runnable
2725 * threads, total number of context switches performed since bootup.
2727 unsigned long nr_running(void)
2729 unsigned long i, sum = 0;
2731 for_each_online_cpu(i)
2732 sum += cpu_rq(i)->nr_running;
2738 * Check if only the current task is running on the cpu.
2740 * Caution: this function does not check that the caller has disabled
2741 * preemption, thus the result might have a time-of-check-to-time-of-use
2742 * race. The caller is responsible to use it correctly, for example:
2744 * - from a non-preemptable section (of course)
2746 * - from a thread that is bound to a single CPU
2748 * - in a loop with very short iterations (e.g. a polling loop)
2750 bool single_task_running(void)
2752 return raw_rq()->nr_running == 1;
2754 EXPORT_SYMBOL(single_task_running);
2756 unsigned long long nr_context_switches(void)
2759 unsigned long long sum = 0;
2761 for_each_possible_cpu(i)
2762 sum += cpu_rq(i)->nr_switches;
2767 unsigned long nr_iowait(void)
2769 unsigned long i, sum = 0;
2771 for_each_possible_cpu(i)
2772 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2777 unsigned long nr_iowait_cpu(int cpu)
2779 struct rq *this = cpu_rq(cpu);
2780 return atomic_read(&this->nr_iowait);
2783 #ifdef CONFIG_CPU_QUIET
2784 u64 nr_running_integral(unsigned int cpu)
2786 unsigned int seqcnt;
2790 if (cpu >= nr_cpu_ids)
2796 * Update average to avoid reading stalled value if there were
2797 * no run-queue changes for a long time. On the other hand if
2798 * the changes are happening right now, just read current value
2802 seqcnt = read_seqcount_begin(&q->ave_seqcnt);
2803 integral = do_nr_running_integral(q);
2804 if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
2805 read_seqcount_begin(&q->ave_seqcnt);
2806 integral = q->nr_running_integral;
2813 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2815 struct rq *rq = this_rq();
2816 *nr_waiters = atomic_read(&rq->nr_iowait);
2817 *load = rq->load.weight;
2823 * sched_exec - execve() is a valuable balancing opportunity, because at
2824 * this point the task has the smallest effective memory and cache footprint.
2826 void sched_exec(void)
2828 struct task_struct *p = current;
2829 unsigned long flags;
2832 raw_spin_lock_irqsave(&p->pi_lock, flags);
2833 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2834 if (dest_cpu == smp_processor_id())
2837 if (likely(cpu_active(dest_cpu))) {
2838 struct migration_arg arg = { p, dest_cpu };
2840 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2841 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2845 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2850 DEFINE_PER_CPU(struct kernel_stat, kstat);
2851 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2853 EXPORT_PER_CPU_SYMBOL(kstat);
2854 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2857 * Return accounted runtime for the task.
2858 * In case the task is currently running, return the runtime plus current's
2859 * pending runtime that have not been accounted yet.
2861 unsigned long long task_sched_runtime(struct task_struct *p)
2863 unsigned long flags;
2867 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2869 * 64-bit doesn't need locks to atomically read a 64bit value.
2870 * So we have a optimization chance when the task's delta_exec is 0.
2871 * Reading ->on_cpu is racy, but this is ok.
2873 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2874 * If we race with it entering cpu, unaccounted time is 0. This is
2875 * indistinguishable from the read occurring a few cycles earlier.
2876 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2877 * been accounted, so we're correct here as well.
2879 if (!p->on_cpu || !task_on_rq_queued(p))
2880 return p->se.sum_exec_runtime;
2883 rq = task_rq_lock(p, &flags);
2885 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2886 * project cycles that may never be accounted to this
2887 * thread, breaking clock_gettime().
2889 if (task_current(rq, p) && task_on_rq_queued(p)) {
2890 update_rq_clock(rq);
2891 p->sched_class->update_curr(rq);
2893 ns = p->se.sum_exec_runtime;
2894 task_rq_unlock(rq, p, &flags);
2899 #ifdef CONFIG_CPU_FREQ_GOV_SCHED
2900 static unsigned long sum_capacity_reqs(unsigned long cfs_cap,
2901 struct sched_capacity_reqs *scr)
2903 unsigned long total = cfs_cap + scr->rt;
2905 total = total * capacity_margin;
2906 total /= SCHED_CAPACITY_SCALE;
2911 static void sched_freq_tick(int cpu)
2913 struct sched_capacity_reqs *scr;
2914 unsigned long capacity_orig, capacity_curr;
2919 capacity_orig = capacity_orig_of(cpu);
2920 capacity_curr = capacity_curr_of(cpu);
2921 if (capacity_curr == capacity_orig)
2925 * To make free room for a task that is building up its "real"
2926 * utilization and to harm its performance the least, request
2927 * a jump to max OPP as soon as the margin of free capacity is
2928 * impacted (specified by capacity_margin).
2930 scr = &per_cpu(cpu_sched_capacity_reqs, cpu);
2931 if (capacity_curr < sum_capacity_reqs(cpu_util(cpu), scr))
2932 set_cfs_cpu_capacity(cpu, true, capacity_max);
2935 static inline void sched_freq_tick(int cpu) { }
2939 * This function gets called by the timer code, with HZ frequency.
2940 * We call it with interrupts disabled.
2942 void scheduler_tick(void)
2944 int cpu = smp_processor_id();
2945 struct rq *rq = cpu_rq(cpu);
2946 struct task_struct *curr = rq->curr;
2950 raw_spin_lock(&rq->lock);
2951 update_rq_clock(rq);
2952 curr->sched_class->task_tick(rq, curr, 0);
2953 update_cpu_load_active(rq);
2954 calc_global_load_tick(rq);
2955 sched_freq_tick(cpu);
2956 raw_spin_unlock(&rq->lock);
2958 perf_event_task_tick();
2961 rq->idle_balance = idle_cpu(cpu);
2962 trigger_load_balance(rq);
2964 rq_last_tick_reset(rq);
2967 #ifdef CONFIG_NO_HZ_FULL
2969 * scheduler_tick_max_deferment
2971 * Keep at least one tick per second when a single
2972 * active task is running because the scheduler doesn't
2973 * yet completely support full dynticks environment.
2975 * This makes sure that uptime, CFS vruntime, load
2976 * balancing, etc... continue to move forward, even
2977 * with a very low granularity.
2979 * Return: Maximum deferment in nanoseconds.
2981 u64 scheduler_tick_max_deferment(void)
2983 struct rq *rq = this_rq();
2984 unsigned long next, now = READ_ONCE(jiffies);
2986 next = rq->last_sched_tick + HZ;
2988 if (time_before_eq(next, now))
2991 return jiffies_to_nsecs(next - now);
2995 notrace unsigned long get_parent_ip(unsigned long addr)
2997 if (in_lock_functions(addr)) {
2998 addr = CALLER_ADDR2;
2999 if (in_lock_functions(addr))
3000 addr = CALLER_ADDR3;
3005 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3006 defined(CONFIG_PREEMPT_TRACER))
3008 void preempt_count_add(int val)
3010 #ifdef CONFIG_DEBUG_PREEMPT
3014 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3017 __preempt_count_add(val);
3018 #ifdef CONFIG_DEBUG_PREEMPT
3020 * Spinlock count overflowing soon?
3022 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3025 if (preempt_count() == val) {
3026 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3027 #ifdef CONFIG_DEBUG_PREEMPT
3028 current->preempt_disable_ip = ip;
3030 trace_preempt_off(CALLER_ADDR0, ip);
3033 EXPORT_SYMBOL(preempt_count_add);
3034 NOKPROBE_SYMBOL(preempt_count_add);
3036 void preempt_count_sub(int val)
3038 #ifdef CONFIG_DEBUG_PREEMPT
3042 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3045 * Is the spinlock portion underflowing?
3047 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3048 !(preempt_count() & PREEMPT_MASK)))
3052 if (preempt_count() == val)
3053 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3054 __preempt_count_sub(val);
3056 EXPORT_SYMBOL(preempt_count_sub);
3057 NOKPROBE_SYMBOL(preempt_count_sub);
3062 * Print scheduling while atomic bug:
3064 static noinline void __schedule_bug(struct task_struct *prev)
3066 if (oops_in_progress)
3069 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3070 prev->comm, prev->pid, preempt_count());
3072 debug_show_held_locks(prev);
3074 if (irqs_disabled())
3075 print_irqtrace_events(prev);
3076 #ifdef CONFIG_DEBUG_PREEMPT
3077 if (in_atomic_preempt_off()) {
3078 pr_err("Preemption disabled at:");
3079 print_ip_sym(current->preempt_disable_ip);
3084 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3088 * Various schedule()-time debugging checks and statistics:
3090 static inline void schedule_debug(struct task_struct *prev)
3092 #ifdef CONFIG_SCHED_STACK_END_CHECK
3093 if (task_stack_end_corrupted(prev))
3094 panic("corrupted stack end detected inside scheduler\n");
3097 if (unlikely(in_atomic_preempt_off())) {
3098 __schedule_bug(prev);
3099 preempt_count_set(PREEMPT_DISABLED);
3103 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3105 schedstat_inc(this_rq(), sched_count);
3109 * Pick up the highest-prio task:
3111 static inline struct task_struct *
3112 pick_next_task(struct rq *rq, struct task_struct *prev)
3114 const struct sched_class *class = &fair_sched_class;
3115 struct task_struct *p;
3118 * Optimization: we know that if all tasks are in
3119 * the fair class we can call that function directly:
3121 if (likely(prev->sched_class == class &&
3122 rq->nr_running == rq->cfs.h_nr_running)) {
3123 p = fair_sched_class.pick_next_task(rq, prev);
3124 if (unlikely(p == RETRY_TASK))
3127 /* assumes fair_sched_class->next == idle_sched_class */
3129 p = idle_sched_class.pick_next_task(rq, prev);
3135 for_each_class(class) {
3136 p = class->pick_next_task(rq, prev);
3138 if (unlikely(p == RETRY_TASK))
3144 BUG(); /* the idle class will always have a runnable task */
3148 * __schedule() is the main scheduler function.
3150 * The main means of driving the scheduler and thus entering this function are:
3152 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3154 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3155 * paths. For example, see arch/x86/entry_64.S.
3157 * To drive preemption between tasks, the scheduler sets the flag in timer
3158 * interrupt handler scheduler_tick().
3160 * 3. Wakeups don't really cause entry into schedule(). They add a
3161 * task to the run-queue and that's it.
3163 * Now, if the new task added to the run-queue preempts the current
3164 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3165 * called on the nearest possible occasion:
3167 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3169 * - in syscall or exception context, at the next outmost
3170 * preempt_enable(). (this might be as soon as the wake_up()'s
3173 * - in IRQ context, return from interrupt-handler to
3174 * preemptible context
3176 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3179 * - cond_resched() call
3180 * - explicit schedule() call
3181 * - return from syscall or exception to user-space
3182 * - return from interrupt-handler to user-space
3184 * WARNING: must be called with preemption disabled!
3186 static void __sched notrace __schedule(bool preempt)
3188 struct task_struct *prev, *next;
3189 unsigned long *switch_count;
3193 cpu = smp_processor_id();
3195 rcu_note_context_switch();
3199 * do_exit() calls schedule() with preemption disabled as an exception;
3200 * however we must fix that up, otherwise the next task will see an
3201 * inconsistent (higher) preempt count.
3203 * It also avoids the below schedule_debug() test from complaining
3206 if (unlikely(prev->state == TASK_DEAD))
3207 preempt_enable_no_resched_notrace();
3209 schedule_debug(prev);
3211 if (sched_feat(HRTICK))
3215 * Make sure that signal_pending_state()->signal_pending() below
3216 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3217 * done by the caller to avoid the race with signal_wake_up().
3219 smp_mb__before_spinlock();
3220 raw_spin_lock_irq(&rq->lock);
3221 lockdep_pin_lock(&rq->lock);
3223 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3225 switch_count = &prev->nivcsw;
3226 if (!preempt && prev->state) {
3227 if (unlikely(signal_pending_state(prev->state, prev))) {
3228 prev->state = TASK_RUNNING;
3230 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3234 * If a worker went to sleep, notify and ask workqueue
3235 * whether it wants to wake up a task to maintain
3238 if (prev->flags & PF_WQ_WORKER) {
3239 struct task_struct *to_wakeup;
3241 to_wakeup = wq_worker_sleeping(prev, cpu);
3243 try_to_wake_up_local(to_wakeup);
3246 switch_count = &prev->nvcsw;
3249 if (task_on_rq_queued(prev))
3250 update_rq_clock(rq);
3252 next = pick_next_task(rq, prev);
3253 clear_tsk_need_resched(prev);
3254 clear_preempt_need_resched();
3255 rq->clock_skip_update = 0;
3257 if (likely(prev != next)) {
3262 trace_sched_switch(preempt, prev, next);
3263 rq = context_switch(rq, prev, next); /* unlocks the rq */
3266 lockdep_unpin_lock(&rq->lock);
3267 raw_spin_unlock_irq(&rq->lock);
3270 balance_callback(rq);
3273 static inline void sched_submit_work(struct task_struct *tsk)
3275 if (!tsk->state || tsk_is_pi_blocked(tsk))
3278 * If we are going to sleep and we have plugged IO queued,
3279 * make sure to submit it to avoid deadlocks.
3281 if (blk_needs_flush_plug(tsk))
3282 blk_schedule_flush_plug(tsk);
3285 asmlinkage __visible void __sched schedule(void)
3287 struct task_struct *tsk = current;
3289 sched_submit_work(tsk);
3293 sched_preempt_enable_no_resched();
3294 } while (need_resched());
3296 EXPORT_SYMBOL(schedule);
3298 #ifdef CONFIG_CONTEXT_TRACKING
3299 asmlinkage __visible void __sched schedule_user(void)
3302 * If we come here after a random call to set_need_resched(),
3303 * or we have been woken up remotely but the IPI has not yet arrived,
3304 * we haven't yet exited the RCU idle mode. Do it here manually until
3305 * we find a better solution.
3307 * NB: There are buggy callers of this function. Ideally we
3308 * should warn if prev_state != CONTEXT_USER, but that will trigger
3309 * too frequently to make sense yet.
3311 enum ctx_state prev_state = exception_enter();
3313 exception_exit(prev_state);
3318 * schedule_preempt_disabled - called with preemption disabled
3320 * Returns with preemption disabled. Note: preempt_count must be 1
3322 void __sched schedule_preempt_disabled(void)
3324 sched_preempt_enable_no_resched();
3329 static void __sched notrace preempt_schedule_common(void)
3332 preempt_disable_notrace();
3334 preempt_enable_no_resched_notrace();
3337 * Check again in case we missed a preemption opportunity
3338 * between schedule and now.
3340 } while (need_resched());
3343 #ifdef CONFIG_PREEMPT
3345 * this is the entry point to schedule() from in-kernel preemption
3346 * off of preempt_enable. Kernel preemptions off return from interrupt
3347 * occur there and call schedule directly.
3349 asmlinkage __visible void __sched notrace preempt_schedule(void)
3352 * If there is a non-zero preempt_count or interrupts are disabled,
3353 * we do not want to preempt the current task. Just return..
3355 if (likely(!preemptible()))
3358 preempt_schedule_common();
3360 NOKPROBE_SYMBOL(preempt_schedule);
3361 EXPORT_SYMBOL(preempt_schedule);
3364 * preempt_schedule_notrace - preempt_schedule called by tracing
3366 * The tracing infrastructure uses preempt_enable_notrace to prevent
3367 * recursion and tracing preempt enabling caused by the tracing
3368 * infrastructure itself. But as tracing can happen in areas coming
3369 * from userspace or just about to enter userspace, a preempt enable
3370 * can occur before user_exit() is called. This will cause the scheduler
3371 * to be called when the system is still in usermode.
3373 * To prevent this, the preempt_enable_notrace will use this function
3374 * instead of preempt_schedule() to exit user context if needed before
3375 * calling the scheduler.
3377 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3379 enum ctx_state prev_ctx;
3381 if (likely(!preemptible()))
3385 preempt_disable_notrace();
3387 * Needs preempt disabled in case user_exit() is traced
3388 * and the tracer calls preempt_enable_notrace() causing
3389 * an infinite recursion.
3391 prev_ctx = exception_enter();
3393 exception_exit(prev_ctx);
3395 preempt_enable_no_resched_notrace();
3396 } while (need_resched());
3398 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3400 #endif /* CONFIG_PREEMPT */
3403 * this is the entry point to schedule() from kernel preemption
3404 * off of irq context.
3405 * Note, that this is called and return with irqs disabled. This will
3406 * protect us against recursive calling from irq.
3408 asmlinkage __visible void __sched preempt_schedule_irq(void)
3410 enum ctx_state prev_state;
3412 /* Catch callers which need to be fixed */
3413 BUG_ON(preempt_count() || !irqs_disabled());
3415 prev_state = exception_enter();
3421 local_irq_disable();
3422 sched_preempt_enable_no_resched();
3423 } while (need_resched());
3425 exception_exit(prev_state);
3428 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3431 return try_to_wake_up(curr->private, mode, wake_flags);
3433 EXPORT_SYMBOL(default_wake_function);
3435 #ifdef CONFIG_RT_MUTEXES
3438 * rt_mutex_setprio - set the current priority of a task
3440 * @prio: prio value (kernel-internal form)
3442 * This function changes the 'effective' priority of a task. It does
3443 * not touch ->normal_prio like __setscheduler().
3445 * Used by the rt_mutex code to implement priority inheritance
3446 * logic. Call site only calls if the priority of the task changed.
3448 void rt_mutex_setprio(struct task_struct *p, int prio)
3450 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3452 const struct sched_class *prev_class;
3454 BUG_ON(prio > MAX_PRIO);
3456 rq = __task_rq_lock(p);
3459 * Idle task boosting is a nono in general. There is one
3460 * exception, when PREEMPT_RT and NOHZ is active:
3462 * The idle task calls get_next_timer_interrupt() and holds
3463 * the timer wheel base->lock on the CPU and another CPU wants
3464 * to access the timer (probably to cancel it). We can safely
3465 * ignore the boosting request, as the idle CPU runs this code
3466 * with interrupts disabled and will complete the lock
3467 * protected section without being interrupted. So there is no
3468 * real need to boost.
3470 if (unlikely(p == rq->idle)) {
3471 WARN_ON(p != rq->curr);
3472 WARN_ON(p->pi_blocked_on);
3476 trace_sched_pi_setprio(p, prio);
3478 prev_class = p->sched_class;
3479 queued = task_on_rq_queued(p);
3480 running = task_current(rq, p);
3482 dequeue_task(rq, p, DEQUEUE_SAVE);
3484 put_prev_task(rq, p);
3487 * Boosting condition are:
3488 * 1. -rt task is running and holds mutex A
3489 * --> -dl task blocks on mutex A
3491 * 2. -dl task is running and holds mutex A
3492 * --> -dl task blocks on mutex A and could preempt the
3495 if (dl_prio(prio)) {
3496 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3497 if (!dl_prio(p->normal_prio) ||
3498 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3499 p->dl.dl_boosted = 1;
3500 enqueue_flag |= ENQUEUE_REPLENISH;
3502 p->dl.dl_boosted = 0;
3503 p->sched_class = &dl_sched_class;
3504 } else if (rt_prio(prio)) {
3505 if (dl_prio(oldprio))
3506 p->dl.dl_boosted = 0;
3508 enqueue_flag |= ENQUEUE_HEAD;
3509 p->sched_class = &rt_sched_class;
3511 if (dl_prio(oldprio))
3512 p->dl.dl_boosted = 0;
3513 if (rt_prio(oldprio))
3515 p->sched_class = &fair_sched_class;
3521 p->sched_class->set_curr_task(rq);
3523 enqueue_task(rq, p, enqueue_flag);
3525 check_class_changed(rq, p, prev_class, oldprio);
3527 preempt_disable(); /* avoid rq from going away on us */
3528 __task_rq_unlock(rq);
3530 balance_callback(rq);
3535 void set_user_nice(struct task_struct *p, long nice)
3537 int old_prio, delta, queued;
3538 unsigned long flags;
3541 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3544 * We have to be careful, if called from sys_setpriority(),
3545 * the task might be in the middle of scheduling on another CPU.
3547 rq = task_rq_lock(p, &flags);
3549 * The RT priorities are set via sched_setscheduler(), but we still
3550 * allow the 'normal' nice value to be set - but as expected
3551 * it wont have any effect on scheduling until the task is
3552 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3554 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3555 p->static_prio = NICE_TO_PRIO(nice);
3558 queued = task_on_rq_queued(p);
3560 dequeue_task(rq, p, DEQUEUE_SAVE);
3562 p->static_prio = NICE_TO_PRIO(nice);
3565 p->prio = effective_prio(p);
3566 delta = p->prio - old_prio;
3569 enqueue_task(rq, p, ENQUEUE_RESTORE);
3571 * If the task increased its priority or is running and
3572 * lowered its priority, then reschedule its CPU:
3574 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3578 task_rq_unlock(rq, p, &flags);
3580 EXPORT_SYMBOL(set_user_nice);
3583 * can_nice - check if a task can reduce its nice value
3587 int can_nice(const struct task_struct *p, const int nice)
3589 /* convert nice value [19,-20] to rlimit style value [1,40] */
3590 int nice_rlim = nice_to_rlimit(nice);
3592 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3593 capable(CAP_SYS_NICE));
3596 #ifdef __ARCH_WANT_SYS_NICE
3599 * sys_nice - change the priority of the current process.
3600 * @increment: priority increment
3602 * sys_setpriority is a more generic, but much slower function that
3603 * does similar things.
3605 SYSCALL_DEFINE1(nice, int, increment)
3610 * Setpriority might change our priority at the same moment.
3611 * We don't have to worry. Conceptually one call occurs first
3612 * and we have a single winner.
3614 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3615 nice = task_nice(current) + increment;
3617 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3618 if (increment < 0 && !can_nice(current, nice))
3621 retval = security_task_setnice(current, nice);
3625 set_user_nice(current, nice);
3632 * task_prio - return the priority value of a given task.
3633 * @p: the task in question.
3635 * Return: The priority value as seen by users in /proc.
3636 * RT tasks are offset by -200. Normal tasks are centered
3637 * around 0, value goes from -16 to +15.
3639 int task_prio(const struct task_struct *p)
3641 return p->prio - MAX_RT_PRIO;
3645 * idle_cpu - is a given cpu idle currently?
3646 * @cpu: the processor in question.
3648 * Return: 1 if the CPU is currently idle. 0 otherwise.
3650 int idle_cpu(int cpu)
3652 struct rq *rq = cpu_rq(cpu);
3654 if (rq->curr != rq->idle)
3661 if (!llist_empty(&rq->wake_list))
3669 * idle_task - return the idle task for a given cpu.
3670 * @cpu: the processor in question.
3672 * Return: The idle task for the cpu @cpu.
3674 struct task_struct *idle_task(int cpu)
3676 return cpu_rq(cpu)->idle;
3680 * find_process_by_pid - find a process with a matching PID value.
3681 * @pid: the pid in question.
3683 * The task of @pid, if found. %NULL otherwise.
3685 static struct task_struct *find_process_by_pid(pid_t pid)
3687 return pid ? find_task_by_vpid(pid) : current;
3691 * This function initializes the sched_dl_entity of a newly becoming
3692 * SCHED_DEADLINE task.
3694 * Only the static values are considered here, the actual runtime and the
3695 * absolute deadline will be properly calculated when the task is enqueued
3696 * for the first time with its new policy.
3699 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3701 struct sched_dl_entity *dl_se = &p->dl;
3703 dl_se->dl_runtime = attr->sched_runtime;
3704 dl_se->dl_deadline = attr->sched_deadline;
3705 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3706 dl_se->flags = attr->sched_flags;
3707 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3710 * Changing the parameters of a task is 'tricky' and we're not doing
3711 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3713 * What we SHOULD do is delay the bandwidth release until the 0-lag
3714 * point. This would include retaining the task_struct until that time
3715 * and change dl_overflow() to not immediately decrement the current
3718 * Instead we retain the current runtime/deadline and let the new
3719 * parameters take effect after the current reservation period lapses.
3720 * This is safe (albeit pessimistic) because the 0-lag point is always
3721 * before the current scheduling deadline.
3723 * We can still have temporary overloads because we do not delay the
3724 * change in bandwidth until that time; so admission control is
3725 * not on the safe side. It does however guarantee tasks will never
3726 * consume more than promised.
3731 * sched_setparam() passes in -1 for its policy, to let the functions
3732 * it calls know not to change it.
3734 #define SETPARAM_POLICY -1
3736 static void __setscheduler_params(struct task_struct *p,
3737 const struct sched_attr *attr)
3739 int policy = attr->sched_policy;
3741 if (policy == SETPARAM_POLICY)
3746 if (dl_policy(policy))
3747 __setparam_dl(p, attr);
3748 else if (fair_policy(policy))
3749 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3752 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3753 * !rt_policy. Always setting this ensures that things like
3754 * getparam()/getattr() don't report silly values for !rt tasks.
3756 p->rt_priority = attr->sched_priority;
3757 p->normal_prio = normal_prio(p);
3761 /* Actually do priority change: must hold pi & rq lock. */
3762 static void __setscheduler(struct rq *rq, struct task_struct *p,
3763 const struct sched_attr *attr, bool keep_boost)
3765 __setscheduler_params(p, attr);
3768 * Keep a potential priority boosting if called from
3769 * sched_setscheduler().
3772 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3774 p->prio = normal_prio(p);
3776 if (dl_prio(p->prio))
3777 p->sched_class = &dl_sched_class;
3778 else if (rt_prio(p->prio))
3779 p->sched_class = &rt_sched_class;
3781 p->sched_class = &fair_sched_class;
3785 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3787 struct sched_dl_entity *dl_se = &p->dl;
3789 attr->sched_priority = p->rt_priority;
3790 attr->sched_runtime = dl_se->dl_runtime;
3791 attr->sched_deadline = dl_se->dl_deadline;
3792 attr->sched_period = dl_se->dl_period;
3793 attr->sched_flags = dl_se->flags;
3797 * This function validates the new parameters of a -deadline task.
3798 * We ask for the deadline not being zero, and greater or equal
3799 * than the runtime, as well as the period of being zero or
3800 * greater than deadline. Furthermore, we have to be sure that
3801 * user parameters are above the internal resolution of 1us (we
3802 * check sched_runtime only since it is always the smaller one) and
3803 * below 2^63 ns (we have to check both sched_deadline and
3804 * sched_period, as the latter can be zero).
3807 __checkparam_dl(const struct sched_attr *attr)
3810 if (attr->sched_deadline == 0)
3814 * Since we truncate DL_SCALE bits, make sure we're at least
3817 if (attr->sched_runtime < (1ULL << DL_SCALE))
3821 * Since we use the MSB for wrap-around and sign issues, make
3822 * sure it's not set (mind that period can be equal to zero).
3824 if (attr->sched_deadline & (1ULL << 63) ||
3825 attr->sched_period & (1ULL << 63))
3828 /* runtime <= deadline <= period (if period != 0) */
3829 if ((attr->sched_period != 0 &&
3830 attr->sched_period < attr->sched_deadline) ||
3831 attr->sched_deadline < attr->sched_runtime)
3838 * check the target process has a UID that matches the current process's
3840 static bool check_same_owner(struct task_struct *p)
3842 const struct cred *cred = current_cred(), *pcred;
3846 pcred = __task_cred(p);
3847 match = (uid_eq(cred->euid, pcred->euid) ||
3848 uid_eq(cred->euid, pcred->uid));
3853 static bool dl_param_changed(struct task_struct *p,
3854 const struct sched_attr *attr)
3856 struct sched_dl_entity *dl_se = &p->dl;
3858 if (dl_se->dl_runtime != attr->sched_runtime ||
3859 dl_se->dl_deadline != attr->sched_deadline ||
3860 dl_se->dl_period != attr->sched_period ||
3861 dl_se->flags != attr->sched_flags)
3867 static int __sched_setscheduler(struct task_struct *p,
3868 const struct sched_attr *attr,
3871 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3872 MAX_RT_PRIO - 1 - attr->sched_priority;
3873 int retval, oldprio, oldpolicy = -1, queued, running;
3874 int new_effective_prio, policy = attr->sched_policy;
3875 unsigned long flags;
3876 const struct sched_class *prev_class;
3880 /* may grab non-irq protected spin_locks */
3881 BUG_ON(in_interrupt());
3883 /* double check policy once rq lock held */
3885 reset_on_fork = p->sched_reset_on_fork;
3886 policy = oldpolicy = p->policy;
3888 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3890 if (!valid_policy(policy))
3894 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3898 * Valid priorities for SCHED_FIFO and SCHED_RR are
3899 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3900 * SCHED_BATCH and SCHED_IDLE is 0.
3902 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3903 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3905 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3906 (rt_policy(policy) != (attr->sched_priority != 0)))
3910 * Allow unprivileged RT tasks to decrease priority:
3912 if (user && !capable(CAP_SYS_NICE)) {
3913 if (fair_policy(policy)) {
3914 if (attr->sched_nice < task_nice(p) &&
3915 !can_nice(p, attr->sched_nice))
3919 if (rt_policy(policy)) {
3920 unsigned long rlim_rtprio =
3921 task_rlimit(p, RLIMIT_RTPRIO);
3923 /* can't set/change the rt policy */
3924 if (policy != p->policy && !rlim_rtprio)
3927 /* can't increase priority */
3928 if (attr->sched_priority > p->rt_priority &&
3929 attr->sched_priority > rlim_rtprio)
3934 * Can't set/change SCHED_DEADLINE policy at all for now
3935 * (safest behavior); in the future we would like to allow
3936 * unprivileged DL tasks to increase their relative deadline
3937 * or reduce their runtime (both ways reducing utilization)
3939 if (dl_policy(policy))
3943 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3944 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3946 if (idle_policy(p->policy) && !idle_policy(policy)) {
3947 if (!can_nice(p, task_nice(p)))
3951 /* can't change other user's priorities */
3952 if (!check_same_owner(p))
3955 /* Normal users shall not reset the sched_reset_on_fork flag */
3956 if (p->sched_reset_on_fork && !reset_on_fork)
3961 retval = security_task_setscheduler(p);
3967 * make sure no PI-waiters arrive (or leave) while we are
3968 * changing the priority of the task:
3970 * To be able to change p->policy safely, the appropriate
3971 * runqueue lock must be held.
3973 rq = task_rq_lock(p, &flags);
3976 * Changing the policy of the stop threads its a very bad idea
3978 if (p == rq->stop) {
3979 task_rq_unlock(rq, p, &flags);
3984 * If not changing anything there's no need to proceed further,
3985 * but store a possible modification of reset_on_fork.
3987 if (unlikely(policy == p->policy)) {
3988 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3990 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3992 if (dl_policy(policy) && dl_param_changed(p, attr))
3995 p->sched_reset_on_fork = reset_on_fork;
3996 task_rq_unlock(rq, p, &flags);
4002 #ifdef CONFIG_RT_GROUP_SCHED
4004 * Do not allow realtime tasks into groups that have no runtime
4007 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4008 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4009 !task_group_is_autogroup(task_group(p))) {
4010 task_rq_unlock(rq, p, &flags);
4015 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4016 cpumask_t *span = rq->rd->span;
4019 * Don't allow tasks with an affinity mask smaller than
4020 * the entire root_domain to become SCHED_DEADLINE. We
4021 * will also fail if there's no bandwidth available.
4023 if (!cpumask_subset(span, &p->cpus_allowed) ||
4024 rq->rd->dl_bw.bw == 0) {
4025 task_rq_unlock(rq, p, &flags);
4032 /* recheck policy now with rq lock held */
4033 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4034 policy = oldpolicy = -1;
4035 task_rq_unlock(rq, p, &flags);
4040 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4041 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4044 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4045 task_rq_unlock(rq, p, &flags);
4049 p->sched_reset_on_fork = reset_on_fork;
4054 * Take priority boosted tasks into account. If the new
4055 * effective priority is unchanged, we just store the new
4056 * normal parameters and do not touch the scheduler class and
4057 * the runqueue. This will be done when the task deboost
4060 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4061 if (new_effective_prio == oldprio) {
4062 __setscheduler_params(p, attr);
4063 task_rq_unlock(rq, p, &flags);
4068 queued = task_on_rq_queued(p);
4069 running = task_current(rq, p);
4071 dequeue_task(rq, p, DEQUEUE_SAVE);
4073 put_prev_task(rq, p);
4075 prev_class = p->sched_class;
4076 __setscheduler(rq, p, attr, pi);
4079 p->sched_class->set_curr_task(rq);
4081 int enqueue_flags = ENQUEUE_RESTORE;
4083 * We enqueue to tail when the priority of a task is
4084 * increased (user space view).
4086 if (oldprio <= p->prio)
4087 enqueue_flags |= ENQUEUE_HEAD;
4089 enqueue_task(rq, p, enqueue_flags);
4092 check_class_changed(rq, p, prev_class, oldprio);
4093 preempt_disable(); /* avoid rq from going away on us */
4094 task_rq_unlock(rq, p, &flags);
4097 rt_mutex_adjust_pi(p);
4100 * Run balance callbacks after we've adjusted the PI chain.
4102 balance_callback(rq);
4108 static int _sched_setscheduler(struct task_struct *p, int policy,
4109 const struct sched_param *param, bool check)
4111 struct sched_attr attr = {
4112 .sched_policy = policy,
4113 .sched_priority = param->sched_priority,
4114 .sched_nice = PRIO_TO_NICE(p->static_prio),
4117 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4118 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4119 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4120 policy &= ~SCHED_RESET_ON_FORK;
4121 attr.sched_policy = policy;
4124 return __sched_setscheduler(p, &attr, check, true);
4127 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4128 * @p: the task in question.
4129 * @policy: new policy.
4130 * @param: structure containing the new RT priority.
4132 * Return: 0 on success. An error code otherwise.
4134 * NOTE that the task may be already dead.
4136 int sched_setscheduler(struct task_struct *p, int policy,
4137 const struct sched_param *param)
4139 return _sched_setscheduler(p, policy, param, true);
4141 EXPORT_SYMBOL_GPL(sched_setscheduler);
4143 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4145 return __sched_setscheduler(p, attr, true, true);
4147 EXPORT_SYMBOL_GPL(sched_setattr);
4150 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4151 * @p: the task in question.
4152 * @policy: new policy.
4153 * @param: structure containing the new RT priority.
4155 * Just like sched_setscheduler, only don't bother checking if the
4156 * current context has permission. For example, this is needed in
4157 * stop_machine(): we create temporary high priority worker threads,
4158 * but our caller might not have that capability.
4160 * Return: 0 on success. An error code otherwise.
4162 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4163 const struct sched_param *param)
4165 return _sched_setscheduler(p, policy, param, false);
4167 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4170 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4172 struct sched_param lparam;
4173 struct task_struct *p;
4176 if (!param || pid < 0)
4178 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4183 p = find_process_by_pid(pid);
4185 retval = sched_setscheduler(p, policy, &lparam);
4192 * Mimics kernel/events/core.c perf_copy_attr().
4194 static int sched_copy_attr(struct sched_attr __user *uattr,
4195 struct sched_attr *attr)
4200 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4204 * zero the full structure, so that a short copy will be nice.
4206 memset(attr, 0, sizeof(*attr));
4208 ret = get_user(size, &uattr->size);
4212 if (size > PAGE_SIZE) /* silly large */
4215 if (!size) /* abi compat */
4216 size = SCHED_ATTR_SIZE_VER0;
4218 if (size < SCHED_ATTR_SIZE_VER0)
4222 * If we're handed a bigger struct than we know of,
4223 * ensure all the unknown bits are 0 - i.e. new
4224 * user-space does not rely on any kernel feature
4225 * extensions we dont know about yet.
4227 if (size > sizeof(*attr)) {
4228 unsigned char __user *addr;
4229 unsigned char __user *end;
4232 addr = (void __user *)uattr + sizeof(*attr);
4233 end = (void __user *)uattr + size;
4235 for (; addr < end; addr++) {
4236 ret = get_user(val, addr);
4242 size = sizeof(*attr);
4245 ret = copy_from_user(attr, uattr, size);
4250 * XXX: do we want to be lenient like existing syscalls; or do we want
4251 * to be strict and return an error on out-of-bounds values?
4253 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4258 put_user(sizeof(*attr), &uattr->size);
4263 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4264 * @pid: the pid in question.
4265 * @policy: new policy.
4266 * @param: structure containing the new RT priority.
4268 * Return: 0 on success. An error code otherwise.
4270 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4271 struct sched_param __user *, param)
4273 /* negative values for policy are not valid */
4277 return do_sched_setscheduler(pid, policy, param);
4281 * sys_sched_setparam - set/change the RT priority of a thread
4282 * @pid: the pid in question.
4283 * @param: structure containing the new RT priority.
4285 * Return: 0 on success. An error code otherwise.
4287 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4289 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4293 * sys_sched_setattr - same as above, but with extended sched_attr
4294 * @pid: the pid in question.
4295 * @uattr: structure containing the extended parameters.
4296 * @flags: for future extension.
4298 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4299 unsigned int, flags)
4301 struct sched_attr attr;
4302 struct task_struct *p;
4305 if (!uattr || pid < 0 || flags)
4308 retval = sched_copy_attr(uattr, &attr);
4312 if ((int)attr.sched_policy < 0)
4317 p = find_process_by_pid(pid);
4319 retval = sched_setattr(p, &attr);
4326 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4327 * @pid: the pid in question.
4329 * Return: On success, the policy of the thread. Otherwise, a negative error
4332 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4334 struct task_struct *p;
4342 p = find_process_by_pid(pid);
4344 retval = security_task_getscheduler(p);
4347 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4354 * sys_sched_getparam - get the RT priority of a thread
4355 * @pid: the pid in question.
4356 * @param: structure containing the RT priority.
4358 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4361 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4363 struct sched_param lp = { .sched_priority = 0 };
4364 struct task_struct *p;
4367 if (!param || pid < 0)
4371 p = find_process_by_pid(pid);
4376 retval = security_task_getscheduler(p);
4380 if (task_has_rt_policy(p))
4381 lp.sched_priority = p->rt_priority;
4385 * This one might sleep, we cannot do it with a spinlock held ...
4387 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4396 static int sched_read_attr(struct sched_attr __user *uattr,
4397 struct sched_attr *attr,
4402 if (!access_ok(VERIFY_WRITE, uattr, usize))
4406 * If we're handed a smaller struct than we know of,
4407 * ensure all the unknown bits are 0 - i.e. old
4408 * user-space does not get uncomplete information.
4410 if (usize < sizeof(*attr)) {
4411 unsigned char *addr;
4414 addr = (void *)attr + usize;
4415 end = (void *)attr + sizeof(*attr);
4417 for (; addr < end; addr++) {
4425 ret = copy_to_user(uattr, attr, attr->size);
4433 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4434 * @pid: the pid in question.
4435 * @uattr: structure containing the extended parameters.
4436 * @size: sizeof(attr) for fwd/bwd comp.
4437 * @flags: for future extension.
4439 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4440 unsigned int, size, unsigned int, flags)
4442 struct sched_attr attr = {
4443 .size = sizeof(struct sched_attr),
4445 struct task_struct *p;
4448 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4449 size < SCHED_ATTR_SIZE_VER0 || flags)
4453 p = find_process_by_pid(pid);
4458 retval = security_task_getscheduler(p);
4462 attr.sched_policy = p->policy;
4463 if (p->sched_reset_on_fork)
4464 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4465 if (task_has_dl_policy(p))
4466 __getparam_dl(p, &attr);
4467 else if (task_has_rt_policy(p))
4468 attr.sched_priority = p->rt_priority;
4470 attr.sched_nice = task_nice(p);
4474 retval = sched_read_attr(uattr, &attr, size);
4482 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4484 cpumask_var_t cpus_allowed, new_mask;
4485 struct task_struct *p;
4490 p = find_process_by_pid(pid);
4496 /* Prevent p going away */
4500 if (p->flags & PF_NO_SETAFFINITY) {
4504 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4508 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4510 goto out_free_cpus_allowed;
4513 if (!check_same_owner(p)) {
4515 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4517 goto out_free_new_mask;
4522 retval = security_task_setscheduler(p);
4524 goto out_free_new_mask;
4527 cpuset_cpus_allowed(p, cpus_allowed);
4528 cpumask_and(new_mask, in_mask, cpus_allowed);
4531 * Since bandwidth control happens on root_domain basis,
4532 * if admission test is enabled, we only admit -deadline
4533 * tasks allowed to run on all the CPUs in the task's
4537 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4539 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4542 goto out_free_new_mask;
4548 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4551 cpuset_cpus_allowed(p, cpus_allowed);
4552 if (!cpumask_subset(new_mask, cpus_allowed)) {
4554 * We must have raced with a concurrent cpuset
4555 * update. Just reset the cpus_allowed to the
4556 * cpuset's cpus_allowed
4558 cpumask_copy(new_mask, cpus_allowed);
4563 free_cpumask_var(new_mask);
4564 out_free_cpus_allowed:
4565 free_cpumask_var(cpus_allowed);
4571 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4572 struct cpumask *new_mask)
4574 if (len < cpumask_size())
4575 cpumask_clear(new_mask);
4576 else if (len > cpumask_size())
4577 len = cpumask_size();
4579 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4583 * sys_sched_setaffinity - set the cpu affinity of a process
4584 * @pid: pid of the process
4585 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4586 * @user_mask_ptr: user-space pointer to the new cpu mask
4588 * Return: 0 on success. An error code otherwise.
4590 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4591 unsigned long __user *, user_mask_ptr)
4593 cpumask_var_t new_mask;
4596 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4599 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4601 retval = sched_setaffinity(pid, new_mask);
4602 free_cpumask_var(new_mask);
4606 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4608 struct task_struct *p;
4609 unsigned long flags;
4615 p = find_process_by_pid(pid);
4619 retval = security_task_getscheduler(p);
4623 raw_spin_lock_irqsave(&p->pi_lock, flags);
4624 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4625 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4634 * sys_sched_getaffinity - get the cpu affinity of a process
4635 * @pid: pid of the process
4636 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4637 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4639 * Return: 0 on success. An error code otherwise.
4641 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4642 unsigned long __user *, user_mask_ptr)
4647 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4649 if (len & (sizeof(unsigned long)-1))
4652 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4655 ret = sched_getaffinity(pid, mask);
4657 size_t retlen = min_t(size_t, len, cpumask_size());
4659 if (copy_to_user(user_mask_ptr, mask, retlen))
4664 free_cpumask_var(mask);
4670 * sys_sched_yield - yield the current processor to other threads.
4672 * This function yields the current CPU to other tasks. If there are no
4673 * other threads running on this CPU then this function will return.
4677 SYSCALL_DEFINE0(sched_yield)
4679 struct rq *rq = this_rq_lock();
4681 schedstat_inc(rq, yld_count);
4682 current->sched_class->yield_task(rq);
4685 * Since we are going to call schedule() anyway, there's
4686 * no need to preempt or enable interrupts:
4688 __release(rq->lock);
4689 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4690 do_raw_spin_unlock(&rq->lock);
4691 sched_preempt_enable_no_resched();
4698 int __sched _cond_resched(void)
4700 if (should_resched(0)) {
4701 preempt_schedule_common();
4706 EXPORT_SYMBOL(_cond_resched);
4709 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4710 * call schedule, and on return reacquire the lock.
4712 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4713 * operations here to prevent schedule() from being called twice (once via
4714 * spin_unlock(), once by hand).
4716 int __cond_resched_lock(spinlock_t *lock)
4718 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4721 lockdep_assert_held(lock);
4723 if (spin_needbreak(lock) || resched) {
4726 preempt_schedule_common();
4734 EXPORT_SYMBOL(__cond_resched_lock);
4736 int __sched __cond_resched_softirq(void)
4738 BUG_ON(!in_softirq());
4740 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4742 preempt_schedule_common();
4748 EXPORT_SYMBOL(__cond_resched_softirq);
4751 * yield - yield the current processor to other threads.
4753 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4755 * The scheduler is at all times free to pick the calling task as the most
4756 * eligible task to run, if removing the yield() call from your code breaks
4757 * it, its already broken.
4759 * Typical broken usage is:
4764 * where one assumes that yield() will let 'the other' process run that will
4765 * make event true. If the current task is a SCHED_FIFO task that will never
4766 * happen. Never use yield() as a progress guarantee!!
4768 * If you want to use yield() to wait for something, use wait_event().
4769 * If you want to use yield() to be 'nice' for others, use cond_resched().
4770 * If you still want to use yield(), do not!
4772 void __sched yield(void)
4774 set_current_state(TASK_RUNNING);
4777 EXPORT_SYMBOL(yield);
4780 * yield_to - yield the current processor to another thread in
4781 * your thread group, or accelerate that thread toward the
4782 * processor it's on.
4784 * @preempt: whether task preemption is allowed or not
4786 * It's the caller's job to ensure that the target task struct
4787 * can't go away on us before we can do any checks.
4790 * true (>0) if we indeed boosted the target task.
4791 * false (0) if we failed to boost the target.
4792 * -ESRCH if there's no task to yield to.
4794 int __sched yield_to(struct task_struct *p, bool preempt)
4796 struct task_struct *curr = current;
4797 struct rq *rq, *p_rq;
4798 unsigned long flags;
4801 local_irq_save(flags);
4807 * If we're the only runnable task on the rq and target rq also
4808 * has only one task, there's absolutely no point in yielding.
4810 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4815 double_rq_lock(rq, p_rq);
4816 if (task_rq(p) != p_rq) {
4817 double_rq_unlock(rq, p_rq);
4821 if (!curr->sched_class->yield_to_task)
4824 if (curr->sched_class != p->sched_class)
4827 if (task_running(p_rq, p) || p->state)
4830 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4832 schedstat_inc(rq, yld_count);
4834 * Make p's CPU reschedule; pick_next_entity takes care of
4837 if (preempt && rq != p_rq)
4842 double_rq_unlock(rq, p_rq);
4844 local_irq_restore(flags);
4851 EXPORT_SYMBOL_GPL(yield_to);
4854 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4855 * that process accounting knows that this is a task in IO wait state.
4857 long __sched io_schedule_timeout(long timeout)
4859 int old_iowait = current->in_iowait;
4863 current->in_iowait = 1;
4864 blk_schedule_flush_plug(current);
4866 delayacct_blkio_start();
4868 atomic_inc(&rq->nr_iowait);
4869 ret = schedule_timeout(timeout);
4870 current->in_iowait = old_iowait;
4871 atomic_dec(&rq->nr_iowait);
4872 delayacct_blkio_end();
4876 EXPORT_SYMBOL(io_schedule_timeout);
4879 * sys_sched_get_priority_max - return maximum RT priority.
4880 * @policy: scheduling class.
4882 * Return: On success, this syscall returns the maximum
4883 * rt_priority that can be used by a given scheduling class.
4884 * On failure, a negative error code is returned.
4886 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4893 ret = MAX_USER_RT_PRIO-1;
4895 case SCHED_DEADLINE:
4906 * sys_sched_get_priority_min - return minimum RT priority.
4907 * @policy: scheduling class.
4909 * Return: On success, this syscall returns the minimum
4910 * rt_priority that can be used by a given scheduling class.
4911 * On failure, a negative error code is returned.
4913 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4922 case SCHED_DEADLINE:
4932 * sys_sched_rr_get_interval - return the default timeslice of a process.
4933 * @pid: pid of the process.
4934 * @interval: userspace pointer to the timeslice value.
4936 * this syscall writes the default timeslice value of a given process
4937 * into the user-space timespec buffer. A value of '0' means infinity.
4939 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4942 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4943 struct timespec __user *, interval)
4945 struct task_struct *p;
4946 unsigned int time_slice;
4947 unsigned long flags;
4957 p = find_process_by_pid(pid);
4961 retval = security_task_getscheduler(p);
4965 rq = task_rq_lock(p, &flags);
4967 if (p->sched_class->get_rr_interval)
4968 time_slice = p->sched_class->get_rr_interval(rq, p);
4969 task_rq_unlock(rq, p, &flags);
4972 jiffies_to_timespec(time_slice, &t);
4973 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4981 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4983 void sched_show_task(struct task_struct *p)
4985 unsigned long free = 0;
4987 unsigned long state = p->state;
4990 state = __ffs(state) + 1;
4991 printk(KERN_INFO "%-15.15s %c", p->comm,
4992 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4993 #if BITS_PER_LONG == 32
4994 if (state == TASK_RUNNING)
4995 printk(KERN_CONT " running ");
4997 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4999 if (state == TASK_RUNNING)
5000 printk(KERN_CONT " running task ");
5002 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5004 #ifdef CONFIG_DEBUG_STACK_USAGE
5005 free = stack_not_used(p);
5010 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5012 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5013 task_pid_nr(p), ppid,
5014 (unsigned long)task_thread_info(p)->flags);
5016 print_worker_info(KERN_INFO, p);
5017 show_stack(p, NULL);
5020 void show_state_filter(unsigned long state_filter)
5022 struct task_struct *g, *p;
5024 #if BITS_PER_LONG == 32
5026 " task PC stack pid father\n");
5029 " task PC stack pid father\n");
5032 for_each_process_thread(g, p) {
5034 * reset the NMI-timeout, listing all files on a slow
5035 * console might take a lot of time:
5036 * Also, reset softlockup watchdogs on all CPUs, because
5037 * another CPU might be blocked waiting for us to process
5040 touch_nmi_watchdog();
5041 touch_all_softlockup_watchdogs();
5042 if (!state_filter || (p->state & state_filter))
5046 #ifdef CONFIG_SCHED_DEBUG
5047 sysrq_sched_debug_show();
5051 * Only show locks if all tasks are dumped:
5054 debug_show_all_locks();
5057 void init_idle_bootup_task(struct task_struct *idle)
5059 idle->sched_class = &idle_sched_class;
5063 * init_idle - set up an idle thread for a given CPU
5064 * @idle: task in question
5065 * @cpu: cpu the idle task belongs to
5067 * NOTE: this function does not set the idle thread's NEED_RESCHED
5068 * flag, to make booting more robust.
5070 void init_idle(struct task_struct *idle, int cpu)
5072 struct rq *rq = cpu_rq(cpu);
5073 unsigned long flags;
5075 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5076 raw_spin_lock(&rq->lock);
5078 __sched_fork(0, idle);
5079 idle->state = TASK_RUNNING;
5080 idle->se.exec_start = sched_clock();
5084 * Its possible that init_idle() gets called multiple times on a task,
5085 * in that case do_set_cpus_allowed() will not do the right thing.
5087 * And since this is boot we can forgo the serialization.
5089 set_cpus_allowed_common(idle, cpumask_of(cpu));
5092 * We're having a chicken and egg problem, even though we are
5093 * holding rq->lock, the cpu isn't yet set to this cpu so the
5094 * lockdep check in task_group() will fail.
5096 * Similar case to sched_fork(). / Alternatively we could
5097 * use task_rq_lock() here and obtain the other rq->lock.
5102 __set_task_cpu(idle, cpu);
5105 rq->curr = rq->idle = idle;
5106 idle->on_rq = TASK_ON_RQ_QUEUED;
5110 raw_spin_unlock(&rq->lock);
5111 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5113 /* Set the preempt count _outside_ the spinlocks! */
5114 init_idle_preempt_count(idle, cpu);
5117 * The idle tasks have their own, simple scheduling class:
5119 idle->sched_class = &idle_sched_class;
5120 ftrace_graph_init_idle_task(idle, cpu);
5121 vtime_init_idle(idle, cpu);
5123 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5127 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5128 const struct cpumask *trial)
5130 int ret = 1, trial_cpus;
5131 struct dl_bw *cur_dl_b;
5132 unsigned long flags;
5134 if (!cpumask_weight(cur))
5137 rcu_read_lock_sched();
5138 cur_dl_b = dl_bw_of(cpumask_any(cur));
5139 trial_cpus = cpumask_weight(trial);
5141 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5142 if (cur_dl_b->bw != -1 &&
5143 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5145 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5146 rcu_read_unlock_sched();
5151 int task_can_attach(struct task_struct *p,
5152 const struct cpumask *cs_cpus_allowed)
5157 * Kthreads which disallow setaffinity shouldn't be moved
5158 * to a new cpuset; we don't want to change their cpu
5159 * affinity and isolating such threads by their set of
5160 * allowed nodes is unnecessary. Thus, cpusets are not
5161 * applicable for such threads. This prevents checking for
5162 * success of set_cpus_allowed_ptr() on all attached tasks
5163 * before cpus_allowed may be changed.
5165 if (p->flags & PF_NO_SETAFFINITY) {
5171 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5173 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5178 unsigned long flags;
5180 rcu_read_lock_sched();
5181 dl_b = dl_bw_of(dest_cpu);
5182 raw_spin_lock_irqsave(&dl_b->lock, flags);
5183 cpus = dl_bw_cpus(dest_cpu);
5184 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5189 * We reserve space for this task in the destination
5190 * root_domain, as we can't fail after this point.
5191 * We will free resources in the source root_domain
5192 * later on (see set_cpus_allowed_dl()).
5194 __dl_add(dl_b, p->dl.dl_bw);
5196 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5197 rcu_read_unlock_sched();
5207 #ifdef CONFIG_NUMA_BALANCING
5208 /* Migrate current task p to target_cpu */
5209 int migrate_task_to(struct task_struct *p, int target_cpu)
5211 struct migration_arg arg = { p, target_cpu };
5212 int curr_cpu = task_cpu(p);
5214 if (curr_cpu == target_cpu)
5217 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5220 /* TODO: This is not properly updating schedstats */
5222 trace_sched_move_numa(p, curr_cpu, target_cpu);
5223 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5227 * Requeue a task on a given node and accurately track the number of NUMA
5228 * tasks on the runqueues
5230 void sched_setnuma(struct task_struct *p, int nid)
5233 unsigned long flags;
5234 bool queued, running;
5236 rq = task_rq_lock(p, &flags);
5237 queued = task_on_rq_queued(p);
5238 running = task_current(rq, p);
5241 dequeue_task(rq, p, DEQUEUE_SAVE);
5243 put_prev_task(rq, p);
5245 p->numa_preferred_nid = nid;
5248 p->sched_class->set_curr_task(rq);
5250 enqueue_task(rq, p, ENQUEUE_RESTORE);
5251 task_rq_unlock(rq, p, &flags);
5253 #endif /* CONFIG_NUMA_BALANCING */
5255 #ifdef CONFIG_HOTPLUG_CPU
5257 * Ensures that the idle task is using init_mm right before its cpu goes
5260 void idle_task_exit(void)
5262 struct mm_struct *mm = current->active_mm;
5264 BUG_ON(cpu_online(smp_processor_id()));
5266 if (mm != &init_mm) {
5267 switch_mm(mm, &init_mm, current);
5268 finish_arch_post_lock_switch();
5274 * Since this CPU is going 'away' for a while, fold any nr_active delta
5275 * we might have. Assumes we're called after migrate_tasks() so that the
5276 * nr_active count is stable.
5278 * Also see the comment "Global load-average calculations".
5280 static void calc_load_migrate(struct rq *rq)
5282 long delta = calc_load_fold_active(rq);
5284 atomic_long_add(delta, &calc_load_tasks);
5287 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5291 static const struct sched_class fake_sched_class = {
5292 .put_prev_task = put_prev_task_fake,
5295 static struct task_struct fake_task = {
5297 * Avoid pull_{rt,dl}_task()
5299 .prio = MAX_PRIO + 1,
5300 .sched_class = &fake_sched_class,
5304 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5305 * try_to_wake_up()->select_task_rq().
5307 * Called with rq->lock held even though we'er in stop_machine() and
5308 * there's no concurrency possible, we hold the required locks anyway
5309 * because of lock validation efforts.
5311 static void migrate_tasks(struct rq *dead_rq)
5313 struct rq *rq = dead_rq;
5314 struct task_struct *next, *stop = rq->stop;
5318 * Fudge the rq selection such that the below task selection loop
5319 * doesn't get stuck on the currently eligible stop task.
5321 * We're currently inside stop_machine() and the rq is either stuck
5322 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5323 * either way we should never end up calling schedule() until we're
5329 * put_prev_task() and pick_next_task() sched
5330 * class method both need to have an up-to-date
5331 * value of rq->clock[_task]
5333 update_rq_clock(rq);
5337 * There's this thread running, bail when that's the only
5340 if (rq->nr_running == 1)
5344 * pick_next_task assumes pinned rq->lock.
5346 lockdep_pin_lock(&rq->lock);
5347 next = pick_next_task(rq, &fake_task);
5349 next->sched_class->put_prev_task(rq, next);
5352 * Rules for changing task_struct::cpus_allowed are holding
5353 * both pi_lock and rq->lock, such that holding either
5354 * stabilizes the mask.
5356 * Drop rq->lock is not quite as disastrous as it usually is
5357 * because !cpu_active at this point, which means load-balance
5358 * will not interfere. Also, stop-machine.
5360 lockdep_unpin_lock(&rq->lock);
5361 raw_spin_unlock(&rq->lock);
5362 raw_spin_lock(&next->pi_lock);
5363 raw_spin_lock(&rq->lock);
5366 * Since we're inside stop-machine, _nothing_ should have
5367 * changed the task, WARN if weird stuff happened, because in
5368 * that case the above rq->lock drop is a fail too.
5370 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5371 raw_spin_unlock(&next->pi_lock);
5375 /* Find suitable destination for @next, with force if needed. */
5376 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5378 rq = __migrate_task(rq, next, dest_cpu);
5379 if (rq != dead_rq) {
5380 raw_spin_unlock(&rq->lock);
5382 raw_spin_lock(&rq->lock);
5384 raw_spin_unlock(&next->pi_lock);
5389 #endif /* CONFIG_HOTPLUG_CPU */
5391 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5393 static struct ctl_table sd_ctl_dir[] = {
5395 .procname = "sched_domain",
5401 static struct ctl_table sd_ctl_root[] = {
5403 .procname = "kernel",
5405 .child = sd_ctl_dir,
5410 static struct ctl_table *sd_alloc_ctl_entry(int n)
5412 struct ctl_table *entry =
5413 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5418 static void sd_free_ctl_entry(struct ctl_table **tablep)
5420 struct ctl_table *entry;
5423 * In the intermediate directories, both the child directory and
5424 * procname are dynamically allocated and could fail but the mode
5425 * will always be set. In the lowest directory the names are
5426 * static strings and all have proc handlers.
5428 for (entry = *tablep; entry->mode; entry++) {
5430 sd_free_ctl_entry(&entry->child);
5431 if (entry->proc_handler == NULL)
5432 kfree(entry->procname);
5439 static int min_load_idx = 0;
5440 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5443 set_table_entry(struct ctl_table *entry,
5444 const char *procname, void *data, int maxlen,
5445 umode_t mode, proc_handler *proc_handler,
5448 entry->procname = procname;
5450 entry->maxlen = maxlen;
5452 entry->proc_handler = proc_handler;
5455 entry->extra1 = &min_load_idx;
5456 entry->extra2 = &max_load_idx;
5460 static struct ctl_table *
5461 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5463 struct ctl_table *table = sd_alloc_ctl_entry(5);
5468 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5469 sizeof(int), 0644, proc_dointvec_minmax, false);
5470 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5471 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5472 proc_doulongvec_minmax, false);
5473 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5474 sizeof(int), 0644, proc_dointvec_minmax, false);
5475 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5476 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5477 proc_doulongvec_minmax, false);
5482 static struct ctl_table *
5483 sd_alloc_ctl_group_table(struct sched_group *sg)
5485 struct ctl_table *table = sd_alloc_ctl_entry(2);
5490 table->procname = kstrdup("energy", GFP_KERNEL);
5492 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5497 static struct ctl_table *
5498 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5500 struct ctl_table *table;
5501 unsigned int nr_entries = 14;
5504 struct sched_group *sg = sd->groups;
5509 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5511 nr_entries += nr_sgs;
5514 table = sd_alloc_ctl_entry(nr_entries);
5519 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5520 sizeof(long), 0644, proc_doulongvec_minmax, false);
5521 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5522 sizeof(long), 0644, proc_doulongvec_minmax, false);
5523 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5524 sizeof(int), 0644, proc_dointvec_minmax, true);
5525 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5526 sizeof(int), 0644, proc_dointvec_minmax, true);
5527 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5528 sizeof(int), 0644, proc_dointvec_minmax, true);
5529 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5530 sizeof(int), 0644, proc_dointvec_minmax, true);
5531 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5532 sizeof(int), 0644, proc_dointvec_minmax, true);
5533 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5534 sizeof(int), 0644, proc_dointvec_minmax, false);
5535 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5536 sizeof(int), 0644, proc_dointvec_minmax, false);
5537 set_table_entry(&table[9], "cache_nice_tries",
5538 &sd->cache_nice_tries,
5539 sizeof(int), 0644, proc_dointvec_minmax, false);
5540 set_table_entry(&table[10], "flags", &sd->flags,
5541 sizeof(int), 0644, proc_dointvec_minmax, false);
5542 set_table_entry(&table[11], "max_newidle_lb_cost",
5543 &sd->max_newidle_lb_cost,
5544 sizeof(long), 0644, proc_doulongvec_minmax, false);
5545 set_table_entry(&table[12], "name", sd->name,
5546 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5550 struct ctl_table *entry = &table[13];
5553 snprintf(buf, 32, "group%d", i);
5554 entry->procname = kstrdup(buf, GFP_KERNEL);
5556 entry->child = sd_alloc_ctl_group_table(sg);
5557 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5559 /* &table[nr_entries-1] is terminator */
5564 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5566 struct ctl_table *entry, *table;
5567 struct sched_domain *sd;
5568 int domain_num = 0, i;
5571 for_each_domain(cpu, sd)
5573 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5578 for_each_domain(cpu, sd) {
5579 snprintf(buf, 32, "domain%d", i);
5580 entry->procname = kstrdup(buf, GFP_KERNEL);
5582 entry->child = sd_alloc_ctl_domain_table(sd);
5589 static struct ctl_table_header *sd_sysctl_header;
5590 static void register_sched_domain_sysctl(void)
5592 int i, cpu_num = num_possible_cpus();
5593 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5596 WARN_ON(sd_ctl_dir[0].child);
5597 sd_ctl_dir[0].child = entry;
5602 for_each_possible_cpu(i) {
5603 snprintf(buf, 32, "cpu%d", i);
5604 entry->procname = kstrdup(buf, GFP_KERNEL);
5606 entry->child = sd_alloc_ctl_cpu_table(i);
5610 WARN_ON(sd_sysctl_header);
5611 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5614 /* may be called multiple times per register */
5615 static void unregister_sched_domain_sysctl(void)
5617 unregister_sysctl_table(sd_sysctl_header);
5618 sd_sysctl_header = NULL;
5619 if (sd_ctl_dir[0].child)
5620 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5623 static void register_sched_domain_sysctl(void)
5626 static void unregister_sched_domain_sysctl(void)
5629 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5631 static void set_rq_online(struct rq *rq)
5634 const struct sched_class *class;
5636 cpumask_set_cpu(rq->cpu, rq->rd->online);
5639 for_each_class(class) {
5640 if (class->rq_online)
5641 class->rq_online(rq);
5646 static void set_rq_offline(struct rq *rq)
5649 const struct sched_class *class;
5651 for_each_class(class) {
5652 if (class->rq_offline)
5653 class->rq_offline(rq);
5656 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5662 * migration_call - callback that gets triggered when a CPU is added.
5663 * Here we can start up the necessary migration thread for the new CPU.
5666 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5668 int cpu = (long)hcpu;
5669 unsigned long flags;
5670 struct rq *rq = cpu_rq(cpu);
5672 switch (action & ~CPU_TASKS_FROZEN) {
5674 case CPU_UP_PREPARE:
5675 rq->calc_load_update = calc_load_update;
5676 account_reset_rq(rq);
5680 /* Update our root-domain */
5681 raw_spin_lock_irqsave(&rq->lock, flags);
5683 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5687 raw_spin_unlock_irqrestore(&rq->lock, flags);
5690 #ifdef CONFIG_HOTPLUG_CPU
5692 sched_ttwu_pending();
5693 /* Update our root-domain */
5694 raw_spin_lock_irqsave(&rq->lock, flags);
5696 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5700 BUG_ON(rq->nr_running != 1); /* the migration thread */
5701 raw_spin_unlock_irqrestore(&rq->lock, flags);
5705 calc_load_migrate(rq);
5710 update_max_interval();
5716 * Register at high priority so that task migration (migrate_all_tasks)
5717 * happens before everything else. This has to be lower priority than
5718 * the notifier in the perf_event subsystem, though.
5720 static struct notifier_block migration_notifier = {
5721 .notifier_call = migration_call,
5722 .priority = CPU_PRI_MIGRATION,
5725 static void set_cpu_rq_start_time(void)
5727 int cpu = smp_processor_id();
5728 struct rq *rq = cpu_rq(cpu);
5729 rq->age_stamp = sched_clock_cpu(cpu);
5732 static int sched_cpu_active(struct notifier_block *nfb,
5733 unsigned long action, void *hcpu)
5735 int cpu = (long)hcpu;
5737 switch (action & ~CPU_TASKS_FROZEN) {
5739 set_cpu_rq_start_time();
5744 * At this point a starting CPU has marked itself as online via
5745 * set_cpu_online(). But it might not yet have marked itself
5746 * as active, which is essential from here on.
5748 set_cpu_active(cpu, true);
5749 stop_machine_unpark(cpu);
5752 case CPU_DOWN_FAILED:
5753 set_cpu_active(cpu, true);
5761 static int sched_cpu_inactive(struct notifier_block *nfb,
5762 unsigned long action, void *hcpu)
5764 switch (action & ~CPU_TASKS_FROZEN) {
5765 case CPU_DOWN_PREPARE:
5766 set_cpu_active((long)hcpu, false);
5773 static int __init migration_init(void)
5775 void *cpu = (void *)(long)smp_processor_id();
5778 /* Initialize migration for the boot CPU */
5779 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5780 BUG_ON(err == NOTIFY_BAD);
5781 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5782 register_cpu_notifier(&migration_notifier);
5784 /* Register cpu active notifiers */
5785 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5786 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5790 early_initcall(migration_init);
5792 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5794 #ifdef CONFIG_SCHED_DEBUG
5796 static __read_mostly int sched_debug_enabled;
5798 static int __init sched_debug_setup(char *str)
5800 sched_debug_enabled = 1;
5804 early_param("sched_debug", sched_debug_setup);
5806 static inline bool sched_debug(void)
5808 return sched_debug_enabled;
5811 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5812 struct cpumask *groupmask)
5814 struct sched_group *group = sd->groups;
5816 cpumask_clear(groupmask);
5818 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5820 if (!(sd->flags & SD_LOAD_BALANCE)) {
5821 printk("does not load-balance\n");
5823 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5828 printk(KERN_CONT "span %*pbl level %s\n",
5829 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5831 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5832 printk(KERN_ERR "ERROR: domain->span does not contain "
5835 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5836 printk(KERN_ERR "ERROR: domain->groups does not contain"
5840 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5844 printk(KERN_ERR "ERROR: group is NULL\n");
5848 if (!cpumask_weight(sched_group_cpus(group))) {
5849 printk(KERN_CONT "\n");
5850 printk(KERN_ERR "ERROR: empty group\n");
5854 if (!(sd->flags & SD_OVERLAP) &&
5855 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5856 printk(KERN_CONT "\n");
5857 printk(KERN_ERR "ERROR: repeated CPUs\n");
5861 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5863 printk(KERN_CONT " %*pbl",
5864 cpumask_pr_args(sched_group_cpus(group)));
5865 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5866 printk(KERN_CONT " (cpu_capacity = %lu)",
5867 group->sgc->capacity);
5870 group = group->next;
5871 } while (group != sd->groups);
5872 printk(KERN_CONT "\n");
5874 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5875 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5878 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5879 printk(KERN_ERR "ERROR: parent span is not a superset "
5880 "of domain->span\n");
5884 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5888 if (!sched_debug_enabled)
5892 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5896 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5899 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5907 #else /* !CONFIG_SCHED_DEBUG */
5908 # define sched_domain_debug(sd, cpu) do { } while (0)
5909 static inline bool sched_debug(void)
5913 #endif /* CONFIG_SCHED_DEBUG */
5915 static int sd_degenerate(struct sched_domain *sd)
5917 if (cpumask_weight(sched_domain_span(sd)) == 1)
5920 /* Following flags need at least 2 groups */
5921 if (sd->flags & (SD_LOAD_BALANCE |
5922 SD_BALANCE_NEWIDLE |
5925 SD_SHARE_CPUCAPACITY |
5926 SD_SHARE_PKG_RESOURCES |
5927 SD_SHARE_POWERDOMAIN |
5928 SD_SHARE_CAP_STATES)) {
5929 if (sd->groups != sd->groups->next)
5933 /* Following flags don't use groups */
5934 if (sd->flags & (SD_WAKE_AFFINE))
5941 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5943 unsigned long cflags = sd->flags, pflags = parent->flags;
5945 if (sd_degenerate(parent))
5948 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5951 /* Flags needing groups don't count if only 1 group in parent */
5952 if (parent->groups == parent->groups->next) {
5953 pflags &= ~(SD_LOAD_BALANCE |
5954 SD_BALANCE_NEWIDLE |
5957 SD_SHARE_CPUCAPACITY |
5958 SD_SHARE_PKG_RESOURCES |
5960 SD_SHARE_POWERDOMAIN |
5961 SD_SHARE_CAP_STATES);
5962 if (nr_node_ids == 1)
5963 pflags &= ~SD_SERIALIZE;
5965 if (~cflags & pflags)
5971 static void free_rootdomain(struct rcu_head *rcu)
5973 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5975 cpupri_cleanup(&rd->cpupri);
5976 cpudl_cleanup(&rd->cpudl);
5977 free_cpumask_var(rd->dlo_mask);
5978 free_cpumask_var(rd->rto_mask);
5979 free_cpumask_var(rd->online);
5980 free_cpumask_var(rd->span);
5984 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5986 struct root_domain *old_rd = NULL;
5987 unsigned long flags;
5989 raw_spin_lock_irqsave(&rq->lock, flags);
5994 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5997 cpumask_clear_cpu(rq->cpu, old_rd->span);
6000 * If we dont want to free the old_rd yet then
6001 * set old_rd to NULL to skip the freeing later
6004 if (!atomic_dec_and_test(&old_rd->refcount))
6008 atomic_inc(&rd->refcount);
6011 cpumask_set_cpu(rq->cpu, rd->span);
6012 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6015 raw_spin_unlock_irqrestore(&rq->lock, flags);
6018 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6021 static int init_rootdomain(struct root_domain *rd)
6023 memset(rd, 0, sizeof(*rd));
6025 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6027 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6029 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6031 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6034 init_dl_bw(&rd->dl_bw);
6035 if (cpudl_init(&rd->cpudl) != 0)
6038 if (cpupri_init(&rd->cpupri) != 0)
6041 init_max_cpu_capacity(&rd->max_cpu_capacity);
6045 free_cpumask_var(rd->rto_mask);
6047 free_cpumask_var(rd->dlo_mask);
6049 free_cpumask_var(rd->online);
6051 free_cpumask_var(rd->span);
6057 * By default the system creates a single root-domain with all cpus as
6058 * members (mimicking the global state we have today).
6060 struct root_domain def_root_domain;
6062 static void init_defrootdomain(void)
6064 init_rootdomain(&def_root_domain);
6066 atomic_set(&def_root_domain.refcount, 1);
6069 static struct root_domain *alloc_rootdomain(void)
6071 struct root_domain *rd;
6073 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6077 if (init_rootdomain(rd) != 0) {
6085 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6087 struct sched_group *tmp, *first;
6096 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6101 } while (sg != first);
6104 static void free_sched_domain(struct rcu_head *rcu)
6106 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6109 * If its an overlapping domain it has private groups, iterate and
6112 if (sd->flags & SD_OVERLAP) {
6113 free_sched_groups(sd->groups, 1);
6114 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6115 kfree(sd->groups->sgc);
6121 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6123 call_rcu(&sd->rcu, free_sched_domain);
6126 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6128 for (; sd; sd = sd->parent)
6129 destroy_sched_domain(sd, cpu);
6133 * Keep a special pointer to the highest sched_domain that has
6134 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6135 * allows us to avoid some pointer chasing select_idle_sibling().
6137 * Also keep a unique ID per domain (we use the first cpu number in
6138 * the cpumask of the domain), this allows us to quickly tell if
6139 * two cpus are in the same cache domain, see cpus_share_cache().
6141 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6142 DEFINE_PER_CPU(int, sd_llc_size);
6143 DEFINE_PER_CPU(int, sd_llc_id);
6144 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6145 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6146 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6147 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6148 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6150 static void update_top_cache_domain(int cpu)
6152 struct sched_domain *sd;
6153 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6157 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6159 id = cpumask_first(sched_domain_span(sd));
6160 size = cpumask_weight(sched_domain_span(sd));
6161 busy_sd = sd->parent; /* sd_busy */
6163 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6165 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6166 per_cpu(sd_llc_size, cpu) = size;
6167 per_cpu(sd_llc_id, cpu) = id;
6169 sd = lowest_flag_domain(cpu, SD_NUMA);
6170 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6172 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6173 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6175 for_each_domain(cpu, sd) {
6176 if (sd->groups->sge)
6181 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6183 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6184 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6188 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6189 * hold the hotplug lock.
6192 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6194 struct rq *rq = cpu_rq(cpu);
6195 struct sched_domain *tmp;
6197 /* Remove the sched domains which do not contribute to scheduling. */
6198 for (tmp = sd; tmp; ) {
6199 struct sched_domain *parent = tmp->parent;
6203 if (sd_parent_degenerate(tmp, parent)) {
6204 tmp->parent = parent->parent;
6206 parent->parent->child = tmp;
6208 * Transfer SD_PREFER_SIBLING down in case of a
6209 * degenerate parent; the spans match for this
6210 * so the property transfers.
6212 if (parent->flags & SD_PREFER_SIBLING)
6213 tmp->flags |= SD_PREFER_SIBLING;
6214 destroy_sched_domain(parent, cpu);
6219 if (sd && sd_degenerate(sd)) {
6222 destroy_sched_domain(tmp, cpu);
6227 sched_domain_debug(sd, cpu);
6229 rq_attach_root(rq, rd);
6231 rcu_assign_pointer(rq->sd, sd);
6232 destroy_sched_domains(tmp, cpu);
6234 update_top_cache_domain(cpu);
6237 /* Setup the mask of cpus configured for isolated domains */
6238 static int __init isolated_cpu_setup(char *str)
6240 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6241 cpulist_parse(str, cpu_isolated_map);
6245 __setup("isolcpus=", isolated_cpu_setup);
6248 struct sched_domain ** __percpu sd;
6249 struct root_domain *rd;
6260 * Build an iteration mask that can exclude certain CPUs from the upwards
6263 * Asymmetric node setups can result in situations where the domain tree is of
6264 * unequal depth, make sure to skip domains that already cover the entire
6267 * In that case build_sched_domains() will have terminated the iteration early
6268 * and our sibling sd spans will be empty. Domains should always include the
6269 * cpu they're built on, so check that.
6272 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6274 const struct cpumask *span = sched_domain_span(sd);
6275 struct sd_data *sdd = sd->private;
6276 struct sched_domain *sibling;
6279 for_each_cpu(i, span) {
6280 sibling = *per_cpu_ptr(sdd->sd, i);
6281 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6284 cpumask_set_cpu(i, sched_group_mask(sg));
6289 * Return the canonical balance cpu for this group, this is the first cpu
6290 * of this group that's also in the iteration mask.
6292 int group_balance_cpu(struct sched_group *sg)
6294 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6298 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6300 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6301 const struct cpumask *span = sched_domain_span(sd);
6302 struct cpumask *covered = sched_domains_tmpmask;
6303 struct sd_data *sdd = sd->private;
6304 struct sched_domain *sibling;
6307 cpumask_clear(covered);
6309 for_each_cpu(i, span) {
6310 struct cpumask *sg_span;
6312 if (cpumask_test_cpu(i, covered))
6315 sibling = *per_cpu_ptr(sdd->sd, i);
6317 /* See the comment near build_group_mask(). */
6318 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6321 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6322 GFP_KERNEL, cpu_to_node(cpu));
6327 sg_span = sched_group_cpus(sg);
6329 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6331 cpumask_set_cpu(i, sg_span);
6333 cpumask_or(covered, covered, sg_span);
6335 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6336 if (atomic_inc_return(&sg->sgc->ref) == 1)
6337 build_group_mask(sd, sg);
6340 * Initialize sgc->capacity such that even if we mess up the
6341 * domains and no possible iteration will get us here, we won't
6344 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6345 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6348 * Make sure the first group of this domain contains the
6349 * canonical balance cpu. Otherwise the sched_domain iteration
6350 * breaks. See update_sg_lb_stats().
6352 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6353 group_balance_cpu(sg) == cpu)
6363 sd->groups = groups;
6368 free_sched_groups(first, 0);
6373 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6375 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6376 struct sched_domain *child = sd->child;
6379 cpu = cpumask_first(sched_domain_span(child));
6382 *sg = *per_cpu_ptr(sdd->sg, cpu);
6383 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6384 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6391 * build_sched_groups will build a circular linked list of the groups
6392 * covered by the given span, and will set each group's ->cpumask correctly,
6393 * and ->cpu_capacity to 0.
6395 * Assumes the sched_domain tree is fully constructed
6398 build_sched_groups(struct sched_domain *sd, int cpu)
6400 struct sched_group *first = NULL, *last = NULL;
6401 struct sd_data *sdd = sd->private;
6402 const struct cpumask *span = sched_domain_span(sd);
6403 struct cpumask *covered;
6406 get_group(cpu, sdd, &sd->groups);
6407 atomic_inc(&sd->groups->ref);
6409 if (cpu != cpumask_first(span))
6412 lockdep_assert_held(&sched_domains_mutex);
6413 covered = sched_domains_tmpmask;
6415 cpumask_clear(covered);
6417 for_each_cpu(i, span) {
6418 struct sched_group *sg;
6421 if (cpumask_test_cpu(i, covered))
6424 group = get_group(i, sdd, &sg);
6425 cpumask_setall(sched_group_mask(sg));
6427 for_each_cpu(j, span) {
6428 if (get_group(j, sdd, NULL) != group)
6431 cpumask_set_cpu(j, covered);
6432 cpumask_set_cpu(j, sched_group_cpus(sg));
6447 * Initialize sched groups cpu_capacity.
6449 * cpu_capacity indicates the capacity of sched group, which is used while
6450 * distributing the load between different sched groups in a sched domain.
6451 * Typically cpu_capacity for all the groups in a sched domain will be same
6452 * unless there are asymmetries in the topology. If there are asymmetries,
6453 * group having more cpu_capacity will pickup more load compared to the
6454 * group having less cpu_capacity.
6456 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6458 struct sched_group *sg = sd->groups;
6463 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6465 } while (sg != sd->groups);
6467 if (cpu != group_balance_cpu(sg))
6470 update_group_capacity(sd, cpu);
6471 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6475 * Check that the per-cpu provided sd energy data is consistent for all cpus
6478 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6479 const struct cpumask *cpumask)
6481 const struct sched_group_energy * const sge = fn(cpu);
6482 struct cpumask mask;
6485 if (cpumask_weight(cpumask) <= 1)
6488 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6490 for_each_cpu(i, &mask) {
6491 const struct sched_group_energy * const e = fn(i);
6494 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6496 for (y = 0; y < (e->nr_idle_states); y++) {
6497 BUG_ON(e->idle_states[y].power !=
6498 sge->idle_states[y].power);
6501 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6503 for (y = 0; y < (e->nr_cap_states); y++) {
6504 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6505 BUG_ON(e->cap_states[y].power !=
6506 sge->cap_states[y].power);
6511 static void init_sched_energy(int cpu, struct sched_domain *sd,
6512 sched_domain_energy_f fn)
6514 if (!(fn && fn(cpu)))
6517 if (cpu != group_balance_cpu(sd->groups))
6520 if (sd->child && !sd->child->groups->sge) {
6521 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6522 #ifdef CONFIG_SCHED_DEBUG
6523 pr_err(" energy data on %s but not on %s domain\n",
6524 sd->name, sd->child->name);
6529 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6531 sd->groups->sge = fn(cpu);
6535 * Initializers for schedule domains
6536 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6539 static int default_relax_domain_level = -1;
6540 int sched_domain_level_max;
6542 static int __init setup_relax_domain_level(char *str)
6544 if (kstrtoint(str, 0, &default_relax_domain_level))
6545 pr_warn("Unable to set relax_domain_level\n");
6549 __setup("relax_domain_level=", setup_relax_domain_level);
6551 static void set_domain_attribute(struct sched_domain *sd,
6552 struct sched_domain_attr *attr)
6556 if (!attr || attr->relax_domain_level < 0) {
6557 if (default_relax_domain_level < 0)
6560 request = default_relax_domain_level;
6562 request = attr->relax_domain_level;
6563 if (request < sd->level) {
6564 /* turn off idle balance on this domain */
6565 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6567 /* turn on idle balance on this domain */
6568 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6572 static void __sdt_free(const struct cpumask *cpu_map);
6573 static int __sdt_alloc(const struct cpumask *cpu_map);
6575 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6576 const struct cpumask *cpu_map)
6580 if (!atomic_read(&d->rd->refcount))
6581 free_rootdomain(&d->rd->rcu); /* fall through */
6583 free_percpu(d->sd); /* fall through */
6585 __sdt_free(cpu_map); /* fall through */
6591 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6592 const struct cpumask *cpu_map)
6594 memset(d, 0, sizeof(*d));
6596 if (__sdt_alloc(cpu_map))
6597 return sa_sd_storage;
6598 d->sd = alloc_percpu(struct sched_domain *);
6600 return sa_sd_storage;
6601 d->rd = alloc_rootdomain();
6604 return sa_rootdomain;
6608 * NULL the sd_data elements we've used to build the sched_domain and
6609 * sched_group structure so that the subsequent __free_domain_allocs()
6610 * will not free the data we're using.
6612 static void claim_allocations(int cpu, struct sched_domain *sd)
6614 struct sd_data *sdd = sd->private;
6616 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6617 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6619 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6620 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6622 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6623 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6627 static int sched_domains_numa_levels;
6628 enum numa_topology_type sched_numa_topology_type;
6629 static int *sched_domains_numa_distance;
6630 int sched_max_numa_distance;
6631 static struct cpumask ***sched_domains_numa_masks;
6632 static int sched_domains_curr_level;
6636 * SD_flags allowed in topology descriptions.
6638 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6639 * SD_SHARE_PKG_RESOURCES - describes shared caches
6640 * SD_NUMA - describes NUMA topologies
6641 * SD_SHARE_POWERDOMAIN - describes shared power domain
6642 * SD_SHARE_CAP_STATES - describes shared capacity states
6645 * SD_ASYM_PACKING - describes SMT quirks
6647 #define TOPOLOGY_SD_FLAGS \
6648 (SD_SHARE_CPUCAPACITY | \
6649 SD_SHARE_PKG_RESOURCES | \
6652 SD_SHARE_POWERDOMAIN | \
6653 SD_SHARE_CAP_STATES)
6655 static struct sched_domain *
6656 sd_init(struct sched_domain_topology_level *tl, int cpu)
6658 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6659 int sd_weight, sd_flags = 0;
6663 * Ugly hack to pass state to sd_numa_mask()...
6665 sched_domains_curr_level = tl->numa_level;
6668 sd_weight = cpumask_weight(tl->mask(cpu));
6671 sd_flags = (*tl->sd_flags)();
6672 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6673 "wrong sd_flags in topology description\n"))
6674 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6676 *sd = (struct sched_domain){
6677 .min_interval = sd_weight,
6678 .max_interval = 2*sd_weight,
6680 .imbalance_pct = 125,
6682 .cache_nice_tries = 0,
6689 .flags = 1*SD_LOAD_BALANCE
6690 | 1*SD_BALANCE_NEWIDLE
6695 | 0*SD_SHARE_CPUCAPACITY
6696 | 0*SD_SHARE_PKG_RESOURCES
6698 | 0*SD_PREFER_SIBLING
6703 .last_balance = jiffies,
6704 .balance_interval = sd_weight,
6706 .max_newidle_lb_cost = 0,
6707 .next_decay_max_lb_cost = jiffies,
6708 #ifdef CONFIG_SCHED_DEBUG
6714 * Convert topological properties into behaviour.
6717 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6718 sd->flags |= SD_PREFER_SIBLING;
6719 sd->imbalance_pct = 110;
6720 sd->smt_gain = 1178; /* ~15% */
6722 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6723 sd->imbalance_pct = 117;
6724 sd->cache_nice_tries = 1;
6728 } else if (sd->flags & SD_NUMA) {
6729 sd->cache_nice_tries = 2;
6733 sd->flags |= SD_SERIALIZE;
6734 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6735 sd->flags &= ~(SD_BALANCE_EXEC |
6742 sd->flags |= SD_PREFER_SIBLING;
6743 sd->cache_nice_tries = 1;
6748 sd->private = &tl->data;
6754 * Topology list, bottom-up.
6756 static struct sched_domain_topology_level default_topology[] = {
6757 #ifdef CONFIG_SCHED_SMT
6758 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6760 #ifdef CONFIG_SCHED_MC
6761 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6763 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6767 static struct sched_domain_topology_level *sched_domain_topology =
6770 #define for_each_sd_topology(tl) \
6771 for (tl = sched_domain_topology; tl->mask; tl++)
6773 void set_sched_topology(struct sched_domain_topology_level *tl)
6775 sched_domain_topology = tl;
6780 static const struct cpumask *sd_numa_mask(int cpu)
6782 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6785 static void sched_numa_warn(const char *str)
6787 static int done = false;
6795 printk(KERN_WARNING "ERROR: %s\n\n", str);
6797 for (i = 0; i < nr_node_ids; i++) {
6798 printk(KERN_WARNING " ");
6799 for (j = 0; j < nr_node_ids; j++)
6800 printk(KERN_CONT "%02d ", node_distance(i,j));
6801 printk(KERN_CONT "\n");
6803 printk(KERN_WARNING "\n");
6806 bool find_numa_distance(int distance)
6810 if (distance == node_distance(0, 0))
6813 for (i = 0; i < sched_domains_numa_levels; i++) {
6814 if (sched_domains_numa_distance[i] == distance)
6822 * A system can have three types of NUMA topology:
6823 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6824 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6825 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6827 * The difference between a glueless mesh topology and a backplane
6828 * topology lies in whether communication between not directly
6829 * connected nodes goes through intermediary nodes (where programs
6830 * could run), or through backplane controllers. This affects
6831 * placement of programs.
6833 * The type of topology can be discerned with the following tests:
6834 * - If the maximum distance between any nodes is 1 hop, the system
6835 * is directly connected.
6836 * - If for two nodes A and B, located N > 1 hops away from each other,
6837 * there is an intermediary node C, which is < N hops away from both
6838 * nodes A and B, the system is a glueless mesh.
6840 static void init_numa_topology_type(void)
6844 n = sched_max_numa_distance;
6846 if (sched_domains_numa_levels <= 1) {
6847 sched_numa_topology_type = NUMA_DIRECT;
6851 for_each_online_node(a) {
6852 for_each_online_node(b) {
6853 /* Find two nodes furthest removed from each other. */
6854 if (node_distance(a, b) < n)
6857 /* Is there an intermediary node between a and b? */
6858 for_each_online_node(c) {
6859 if (node_distance(a, c) < n &&
6860 node_distance(b, c) < n) {
6861 sched_numa_topology_type =
6867 sched_numa_topology_type = NUMA_BACKPLANE;
6873 static void sched_init_numa(void)
6875 int next_distance, curr_distance = node_distance(0, 0);
6876 struct sched_domain_topology_level *tl;
6880 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6881 if (!sched_domains_numa_distance)
6885 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6886 * unique distances in the node_distance() table.
6888 * Assumes node_distance(0,j) includes all distances in
6889 * node_distance(i,j) in order to avoid cubic time.
6891 next_distance = curr_distance;
6892 for (i = 0; i < nr_node_ids; i++) {
6893 for (j = 0; j < nr_node_ids; j++) {
6894 for (k = 0; k < nr_node_ids; k++) {
6895 int distance = node_distance(i, k);
6897 if (distance > curr_distance &&
6898 (distance < next_distance ||
6899 next_distance == curr_distance))
6900 next_distance = distance;
6903 * While not a strong assumption it would be nice to know
6904 * about cases where if node A is connected to B, B is not
6905 * equally connected to A.
6907 if (sched_debug() && node_distance(k, i) != distance)
6908 sched_numa_warn("Node-distance not symmetric");
6910 if (sched_debug() && i && !find_numa_distance(distance))
6911 sched_numa_warn("Node-0 not representative");
6913 if (next_distance != curr_distance) {
6914 sched_domains_numa_distance[level++] = next_distance;
6915 sched_domains_numa_levels = level;
6916 curr_distance = next_distance;
6921 * In case of sched_debug() we verify the above assumption.
6931 * 'level' contains the number of unique distances, excluding the
6932 * identity distance node_distance(i,i).
6934 * The sched_domains_numa_distance[] array includes the actual distance
6939 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6940 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6941 * the array will contain less then 'level' members. This could be
6942 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6943 * in other functions.
6945 * We reset it to 'level' at the end of this function.
6947 sched_domains_numa_levels = 0;
6949 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6950 if (!sched_domains_numa_masks)
6954 * Now for each level, construct a mask per node which contains all
6955 * cpus of nodes that are that many hops away from us.
6957 for (i = 0; i < level; i++) {
6958 sched_domains_numa_masks[i] =
6959 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6960 if (!sched_domains_numa_masks[i])
6963 for (j = 0; j < nr_node_ids; j++) {
6964 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6968 sched_domains_numa_masks[i][j] = mask;
6971 if (node_distance(j, k) > sched_domains_numa_distance[i])
6974 cpumask_or(mask, mask, cpumask_of_node(k));
6979 /* Compute default topology size */
6980 for (i = 0; sched_domain_topology[i].mask; i++);
6982 tl = kzalloc((i + level + 1) *
6983 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6988 * Copy the default topology bits..
6990 for (i = 0; sched_domain_topology[i].mask; i++)
6991 tl[i] = sched_domain_topology[i];
6994 * .. and append 'j' levels of NUMA goodness.
6996 for (j = 0; j < level; i++, j++) {
6997 tl[i] = (struct sched_domain_topology_level){
6998 .mask = sd_numa_mask,
6999 .sd_flags = cpu_numa_flags,
7000 .flags = SDTL_OVERLAP,
7006 sched_domain_topology = tl;
7008 sched_domains_numa_levels = level;
7009 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7011 init_numa_topology_type();
7014 static void sched_domains_numa_masks_set(int cpu)
7017 int node = cpu_to_node(cpu);
7019 for (i = 0; i < sched_domains_numa_levels; i++) {
7020 for (j = 0; j < nr_node_ids; j++) {
7021 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7022 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7027 static void sched_domains_numa_masks_clear(int cpu)
7030 for (i = 0; i < sched_domains_numa_levels; i++) {
7031 for (j = 0; j < nr_node_ids; j++)
7032 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7037 * Update sched_domains_numa_masks[level][node] array when new cpus
7040 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7041 unsigned long action,
7044 int cpu = (long)hcpu;
7046 switch (action & ~CPU_TASKS_FROZEN) {
7048 sched_domains_numa_masks_set(cpu);
7052 sched_domains_numa_masks_clear(cpu);
7062 static inline void sched_init_numa(void)
7066 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7067 unsigned long action,
7072 #endif /* CONFIG_NUMA */
7074 static int __sdt_alloc(const struct cpumask *cpu_map)
7076 struct sched_domain_topology_level *tl;
7079 for_each_sd_topology(tl) {
7080 struct sd_data *sdd = &tl->data;
7082 sdd->sd = alloc_percpu(struct sched_domain *);
7086 sdd->sg = alloc_percpu(struct sched_group *);
7090 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7094 for_each_cpu(j, cpu_map) {
7095 struct sched_domain *sd;
7096 struct sched_group *sg;
7097 struct sched_group_capacity *sgc;
7099 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7100 GFP_KERNEL, cpu_to_node(j));
7104 *per_cpu_ptr(sdd->sd, j) = sd;
7106 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7107 GFP_KERNEL, cpu_to_node(j));
7113 *per_cpu_ptr(sdd->sg, j) = sg;
7115 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7116 GFP_KERNEL, cpu_to_node(j));
7120 *per_cpu_ptr(sdd->sgc, j) = sgc;
7127 static void __sdt_free(const struct cpumask *cpu_map)
7129 struct sched_domain_topology_level *tl;
7132 for_each_sd_topology(tl) {
7133 struct sd_data *sdd = &tl->data;
7135 for_each_cpu(j, cpu_map) {
7136 struct sched_domain *sd;
7139 sd = *per_cpu_ptr(sdd->sd, j);
7140 if (sd && (sd->flags & SD_OVERLAP))
7141 free_sched_groups(sd->groups, 0);
7142 kfree(*per_cpu_ptr(sdd->sd, j));
7146 kfree(*per_cpu_ptr(sdd->sg, j));
7148 kfree(*per_cpu_ptr(sdd->sgc, j));
7150 free_percpu(sdd->sd);
7152 free_percpu(sdd->sg);
7154 free_percpu(sdd->sgc);
7159 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7160 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7161 struct sched_domain *child, int cpu)
7163 struct sched_domain *sd = sd_init(tl, cpu);
7167 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7169 sd->level = child->level + 1;
7170 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7174 if (!cpumask_subset(sched_domain_span(child),
7175 sched_domain_span(sd))) {
7176 pr_err("BUG: arch topology borken\n");
7177 #ifdef CONFIG_SCHED_DEBUG
7178 pr_err(" the %s domain not a subset of the %s domain\n",
7179 child->name, sd->name);
7181 /* Fixup, ensure @sd has at least @child cpus. */
7182 cpumask_or(sched_domain_span(sd),
7183 sched_domain_span(sd),
7184 sched_domain_span(child));
7188 set_domain_attribute(sd, attr);
7194 * Build sched domains for a given set of cpus and attach the sched domains
7195 * to the individual cpus
7197 static int build_sched_domains(const struct cpumask *cpu_map,
7198 struct sched_domain_attr *attr)
7200 enum s_alloc alloc_state;
7201 struct sched_domain *sd;
7203 struct rq *rq = NULL;
7204 int i, ret = -ENOMEM;
7206 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7207 if (alloc_state != sa_rootdomain)
7210 /* Set up domains for cpus specified by the cpu_map. */
7211 for_each_cpu(i, cpu_map) {
7212 struct sched_domain_topology_level *tl;
7215 for_each_sd_topology(tl) {
7216 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7217 if (tl == sched_domain_topology)
7218 *per_cpu_ptr(d.sd, i) = sd;
7219 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7220 sd->flags |= SD_OVERLAP;
7221 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7226 /* Build the groups for the domains */
7227 for_each_cpu(i, cpu_map) {
7228 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7229 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7230 if (sd->flags & SD_OVERLAP) {
7231 if (build_overlap_sched_groups(sd, i))
7234 if (build_sched_groups(sd, i))
7240 /* Calculate CPU capacity for physical packages and nodes */
7241 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7242 struct sched_domain_topology_level *tl = sched_domain_topology;
7244 if (!cpumask_test_cpu(i, cpu_map))
7247 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7248 init_sched_energy(i, sd, tl->energy);
7249 claim_allocations(i, sd);
7250 init_sched_groups_capacity(i, sd);
7254 /* Attach the domains */
7256 for_each_cpu(i, cpu_map) {
7258 sd = *per_cpu_ptr(d.sd, i);
7259 cpu_attach_domain(sd, d.rd, i);
7265 __free_domain_allocs(&d, alloc_state, cpu_map);
7269 static cpumask_var_t *doms_cur; /* current sched domains */
7270 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7271 static struct sched_domain_attr *dattr_cur;
7272 /* attribues of custom domains in 'doms_cur' */
7275 * Special case: If a kmalloc of a doms_cur partition (array of
7276 * cpumask) fails, then fallback to a single sched domain,
7277 * as determined by the single cpumask fallback_doms.
7279 static cpumask_var_t fallback_doms;
7282 * arch_update_cpu_topology lets virtualized architectures update the
7283 * cpu core maps. It is supposed to return 1 if the topology changed
7284 * or 0 if it stayed the same.
7286 int __weak arch_update_cpu_topology(void)
7291 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7294 cpumask_var_t *doms;
7296 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7299 for (i = 0; i < ndoms; i++) {
7300 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7301 free_sched_domains(doms, i);
7308 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7311 for (i = 0; i < ndoms; i++)
7312 free_cpumask_var(doms[i]);
7317 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7318 * For now this just excludes isolated cpus, but could be used to
7319 * exclude other special cases in the future.
7321 static int init_sched_domains(const struct cpumask *cpu_map)
7325 arch_update_cpu_topology();
7327 doms_cur = alloc_sched_domains(ndoms_cur);
7329 doms_cur = &fallback_doms;
7330 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7331 err = build_sched_domains(doms_cur[0], NULL);
7332 register_sched_domain_sysctl();
7338 * Detach sched domains from a group of cpus specified in cpu_map
7339 * These cpus will now be attached to the NULL domain
7341 static void detach_destroy_domains(const struct cpumask *cpu_map)
7346 for_each_cpu(i, cpu_map)
7347 cpu_attach_domain(NULL, &def_root_domain, i);
7351 /* handle null as "default" */
7352 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7353 struct sched_domain_attr *new, int idx_new)
7355 struct sched_domain_attr tmp;
7362 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7363 new ? (new + idx_new) : &tmp,
7364 sizeof(struct sched_domain_attr));
7368 * Partition sched domains as specified by the 'ndoms_new'
7369 * cpumasks in the array doms_new[] of cpumasks. This compares
7370 * doms_new[] to the current sched domain partitioning, doms_cur[].
7371 * It destroys each deleted domain and builds each new domain.
7373 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7374 * The masks don't intersect (don't overlap.) We should setup one
7375 * sched domain for each mask. CPUs not in any of the cpumasks will
7376 * not be load balanced. If the same cpumask appears both in the
7377 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7380 * The passed in 'doms_new' should be allocated using
7381 * alloc_sched_domains. This routine takes ownership of it and will
7382 * free_sched_domains it when done with it. If the caller failed the
7383 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7384 * and partition_sched_domains() will fallback to the single partition
7385 * 'fallback_doms', it also forces the domains to be rebuilt.
7387 * If doms_new == NULL it will be replaced with cpu_online_mask.
7388 * ndoms_new == 0 is a special case for destroying existing domains,
7389 * and it will not create the default domain.
7391 * Call with hotplug lock held
7393 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7394 struct sched_domain_attr *dattr_new)
7399 mutex_lock(&sched_domains_mutex);
7401 /* always unregister in case we don't destroy any domains */
7402 unregister_sched_domain_sysctl();
7404 /* Let architecture update cpu core mappings. */
7405 new_topology = arch_update_cpu_topology();
7407 n = doms_new ? ndoms_new : 0;
7409 /* Destroy deleted domains */
7410 for (i = 0; i < ndoms_cur; i++) {
7411 for (j = 0; j < n && !new_topology; j++) {
7412 if (cpumask_equal(doms_cur[i], doms_new[j])
7413 && dattrs_equal(dattr_cur, i, dattr_new, j))
7416 /* no match - a current sched domain not in new doms_new[] */
7417 detach_destroy_domains(doms_cur[i]);
7423 if (doms_new == NULL) {
7425 doms_new = &fallback_doms;
7426 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7427 WARN_ON_ONCE(dattr_new);
7430 /* Build new domains */
7431 for (i = 0; i < ndoms_new; i++) {
7432 for (j = 0; j < n && !new_topology; j++) {
7433 if (cpumask_equal(doms_new[i], doms_cur[j])
7434 && dattrs_equal(dattr_new, i, dattr_cur, j))
7437 /* no match - add a new doms_new */
7438 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7443 /* Remember the new sched domains */
7444 if (doms_cur != &fallback_doms)
7445 free_sched_domains(doms_cur, ndoms_cur);
7446 kfree(dattr_cur); /* kfree(NULL) is safe */
7447 doms_cur = doms_new;
7448 dattr_cur = dattr_new;
7449 ndoms_cur = ndoms_new;
7451 register_sched_domain_sysctl();
7453 mutex_unlock(&sched_domains_mutex);
7456 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7459 * Update cpusets according to cpu_active mask. If cpusets are
7460 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7461 * around partition_sched_domains().
7463 * If we come here as part of a suspend/resume, don't touch cpusets because we
7464 * want to restore it back to its original state upon resume anyway.
7466 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7470 case CPU_ONLINE_FROZEN:
7471 case CPU_DOWN_FAILED_FROZEN:
7474 * num_cpus_frozen tracks how many CPUs are involved in suspend
7475 * resume sequence. As long as this is not the last online
7476 * operation in the resume sequence, just build a single sched
7477 * domain, ignoring cpusets.
7480 if (likely(num_cpus_frozen)) {
7481 partition_sched_domains(1, NULL, NULL);
7486 * This is the last CPU online operation. So fall through and
7487 * restore the original sched domains by considering the
7488 * cpuset configurations.
7492 cpuset_update_active_cpus(true);
7500 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7503 unsigned long flags;
7504 long cpu = (long)hcpu;
7510 case CPU_DOWN_PREPARE:
7511 rcu_read_lock_sched();
7512 dl_b = dl_bw_of(cpu);
7514 raw_spin_lock_irqsave(&dl_b->lock, flags);
7515 cpus = dl_bw_cpus(cpu);
7516 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7517 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7519 rcu_read_unlock_sched();
7522 return notifier_from_errno(-EBUSY);
7523 cpuset_update_active_cpus(false);
7525 case CPU_DOWN_PREPARE_FROZEN:
7527 partition_sched_domains(1, NULL, NULL);
7535 void __init sched_init_smp(void)
7537 cpumask_var_t non_isolated_cpus;
7539 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7540 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7545 * There's no userspace yet to cause hotplug operations; hence all the
7546 * cpu masks are stable and all blatant races in the below code cannot
7549 mutex_lock(&sched_domains_mutex);
7550 init_sched_domains(cpu_active_mask);
7551 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7552 if (cpumask_empty(non_isolated_cpus))
7553 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7554 mutex_unlock(&sched_domains_mutex);
7556 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7557 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7558 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7562 /* Move init over to a non-isolated CPU */
7563 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7565 sched_init_granularity();
7566 free_cpumask_var(non_isolated_cpus);
7568 init_sched_rt_class();
7569 init_sched_dl_class();
7572 void __init sched_init_smp(void)
7574 sched_init_granularity();
7576 #endif /* CONFIG_SMP */
7578 int in_sched_functions(unsigned long addr)
7580 return in_lock_functions(addr) ||
7581 (addr >= (unsigned long)__sched_text_start
7582 && addr < (unsigned long)__sched_text_end);
7585 #ifdef CONFIG_CGROUP_SCHED
7587 * Default task group.
7588 * Every task in system belongs to this group at bootup.
7590 struct task_group root_task_group;
7591 LIST_HEAD(task_groups);
7594 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7596 void __init sched_init(void)
7599 unsigned long alloc_size = 0, ptr;
7601 #ifdef CONFIG_FAIR_GROUP_SCHED
7602 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7604 #ifdef CONFIG_RT_GROUP_SCHED
7605 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7608 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7610 #ifdef CONFIG_FAIR_GROUP_SCHED
7611 root_task_group.se = (struct sched_entity **)ptr;
7612 ptr += nr_cpu_ids * sizeof(void **);
7614 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7615 ptr += nr_cpu_ids * sizeof(void **);
7617 #endif /* CONFIG_FAIR_GROUP_SCHED */
7618 #ifdef CONFIG_RT_GROUP_SCHED
7619 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7620 ptr += nr_cpu_ids * sizeof(void **);
7622 root_task_group.rt_rq = (struct rt_rq **)ptr;
7623 ptr += nr_cpu_ids * sizeof(void **);
7625 #endif /* CONFIG_RT_GROUP_SCHED */
7627 #ifdef CONFIG_CPUMASK_OFFSTACK
7628 for_each_possible_cpu(i) {
7629 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7630 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7632 #endif /* CONFIG_CPUMASK_OFFSTACK */
7634 init_rt_bandwidth(&def_rt_bandwidth,
7635 global_rt_period(), global_rt_runtime());
7636 init_dl_bandwidth(&def_dl_bandwidth,
7637 global_rt_period(), global_rt_runtime());
7640 init_defrootdomain();
7643 #ifdef CONFIG_RT_GROUP_SCHED
7644 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7645 global_rt_period(), global_rt_runtime());
7646 #endif /* CONFIG_RT_GROUP_SCHED */
7648 #ifdef CONFIG_CGROUP_SCHED
7649 list_add(&root_task_group.list, &task_groups);
7650 INIT_LIST_HEAD(&root_task_group.children);
7651 INIT_LIST_HEAD(&root_task_group.siblings);
7652 autogroup_init(&init_task);
7654 #endif /* CONFIG_CGROUP_SCHED */
7656 for_each_possible_cpu(i) {
7660 raw_spin_lock_init(&rq->lock);
7662 rq->calc_load_active = 0;
7663 rq->calc_load_update = jiffies + LOAD_FREQ;
7664 init_cfs_rq(&rq->cfs);
7665 init_rt_rq(&rq->rt);
7666 init_dl_rq(&rq->dl);
7667 #ifdef CONFIG_FAIR_GROUP_SCHED
7668 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7669 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7671 * How much cpu bandwidth does root_task_group get?
7673 * In case of task-groups formed thr' the cgroup filesystem, it
7674 * gets 100% of the cpu resources in the system. This overall
7675 * system cpu resource is divided among the tasks of
7676 * root_task_group and its child task-groups in a fair manner,
7677 * based on each entity's (task or task-group's) weight
7678 * (se->load.weight).
7680 * In other words, if root_task_group has 10 tasks of weight
7681 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7682 * then A0's share of the cpu resource is:
7684 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7686 * We achieve this by letting root_task_group's tasks sit
7687 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7689 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7690 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7691 #endif /* CONFIG_FAIR_GROUP_SCHED */
7693 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7694 #ifdef CONFIG_RT_GROUP_SCHED
7695 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7698 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7699 rq->cpu_load[j] = 0;
7701 rq->last_load_update_tick = jiffies;
7706 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7707 rq->balance_callback = NULL;
7708 rq->active_balance = 0;
7709 rq->next_balance = jiffies;
7714 rq->avg_idle = 2*sysctl_sched_migration_cost;
7715 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7717 INIT_LIST_HEAD(&rq->cfs_tasks);
7719 rq_attach_root(rq, &def_root_domain);
7720 #ifdef CONFIG_NO_HZ_COMMON
7723 #ifdef CONFIG_NO_HZ_FULL
7724 rq->last_sched_tick = 0;
7728 atomic_set(&rq->nr_iowait, 0);
7731 set_load_weight(&init_task);
7733 #ifdef CONFIG_PREEMPT_NOTIFIERS
7734 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7738 * The boot idle thread does lazy MMU switching as well:
7740 atomic_inc(&init_mm.mm_count);
7741 enter_lazy_tlb(&init_mm, current);
7744 * During early bootup we pretend to be a normal task:
7746 current->sched_class = &fair_sched_class;
7749 * Make us the idle thread. Technically, schedule() should not be
7750 * called from this thread, however somewhere below it might be,
7751 * but because we are the idle thread, we just pick up running again
7752 * when this runqueue becomes "idle".
7754 init_idle(current, smp_processor_id());
7756 calc_load_update = jiffies + LOAD_FREQ;
7759 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7760 /* May be allocated at isolcpus cmdline parse time */
7761 if (cpu_isolated_map == NULL)
7762 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7763 idle_thread_set_boot_cpu();
7764 set_cpu_rq_start_time();
7766 init_sched_fair_class();
7768 scheduler_running = 1;
7771 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7772 static inline int preempt_count_equals(int preempt_offset)
7774 int nested = preempt_count() + rcu_preempt_depth();
7776 return (nested == preempt_offset);
7779 static int __might_sleep_init_called;
7780 int __init __might_sleep_init(void)
7782 __might_sleep_init_called = 1;
7785 early_initcall(__might_sleep_init);
7787 void __might_sleep(const char *file, int line, int preempt_offset)
7790 * Blocking primitives will set (and therefore destroy) current->state,
7791 * since we will exit with TASK_RUNNING make sure we enter with it,
7792 * otherwise we will destroy state.
7794 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7795 "do not call blocking ops when !TASK_RUNNING; "
7796 "state=%lx set at [<%p>] %pS\n",
7798 (void *)current->task_state_change,
7799 (void *)current->task_state_change);
7801 ___might_sleep(file, line, preempt_offset);
7803 EXPORT_SYMBOL(__might_sleep);
7805 void ___might_sleep(const char *file, int line, int preempt_offset)
7807 static unsigned long prev_jiffy; /* ratelimiting */
7809 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7810 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7811 !is_idle_task(current)) || oops_in_progress)
7813 if (system_state != SYSTEM_RUNNING &&
7814 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
7816 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7818 prev_jiffy = jiffies;
7821 "BUG: sleeping function called from invalid context at %s:%d\n",
7824 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7825 in_atomic(), irqs_disabled(),
7826 current->pid, current->comm);
7828 if (task_stack_end_corrupted(current))
7829 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7831 debug_show_held_locks(current);
7832 if (irqs_disabled())
7833 print_irqtrace_events(current);
7834 #ifdef CONFIG_DEBUG_PREEMPT
7835 if (!preempt_count_equals(preempt_offset)) {
7836 pr_err("Preemption disabled at:");
7837 print_ip_sym(current->preempt_disable_ip);
7843 EXPORT_SYMBOL(___might_sleep);
7846 #ifdef CONFIG_MAGIC_SYSRQ
7847 void normalize_rt_tasks(void)
7849 struct task_struct *g, *p;
7850 struct sched_attr attr = {
7851 .sched_policy = SCHED_NORMAL,
7854 read_lock(&tasklist_lock);
7855 for_each_process_thread(g, p) {
7857 * Only normalize user tasks:
7859 if (p->flags & PF_KTHREAD)
7862 p->se.exec_start = 0;
7863 #ifdef CONFIG_SCHEDSTATS
7864 p->se.statistics.wait_start = 0;
7865 p->se.statistics.sleep_start = 0;
7866 p->se.statistics.block_start = 0;
7869 if (!dl_task(p) && !rt_task(p)) {
7871 * Renice negative nice level userspace
7874 if (task_nice(p) < 0)
7875 set_user_nice(p, 0);
7879 __sched_setscheduler(p, &attr, false, false);
7881 read_unlock(&tasklist_lock);
7884 #endif /* CONFIG_MAGIC_SYSRQ */
7886 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7888 * These functions are only useful for the IA64 MCA handling, or kdb.
7890 * They can only be called when the whole system has been
7891 * stopped - every CPU needs to be quiescent, and no scheduling
7892 * activity can take place. Using them for anything else would
7893 * be a serious bug, and as a result, they aren't even visible
7894 * under any other configuration.
7898 * curr_task - return the current task for a given cpu.
7899 * @cpu: the processor in question.
7901 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7903 * Return: The current task for @cpu.
7905 struct task_struct *curr_task(int cpu)
7907 return cpu_curr(cpu);
7910 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7914 * set_curr_task - set the current task for a given cpu.
7915 * @cpu: the processor in question.
7916 * @p: the task pointer to set.
7918 * Description: This function must only be used when non-maskable interrupts
7919 * are serviced on a separate stack. It allows the architecture to switch the
7920 * notion of the current task on a cpu in a non-blocking manner. This function
7921 * must be called with all CPU's synchronized, and interrupts disabled, the
7922 * and caller must save the original value of the current task (see
7923 * curr_task() above) and restore that value before reenabling interrupts and
7924 * re-starting the system.
7926 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7928 void set_curr_task(int cpu, struct task_struct *p)
7935 #ifdef CONFIG_CGROUP_SCHED
7936 /* task_group_lock serializes the addition/removal of task groups */
7937 static DEFINE_SPINLOCK(task_group_lock);
7939 static void sched_free_group(struct task_group *tg)
7941 free_fair_sched_group(tg);
7942 free_rt_sched_group(tg);
7947 /* allocate runqueue etc for a new task group */
7948 struct task_group *sched_create_group(struct task_group *parent)
7950 struct task_group *tg;
7952 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7954 return ERR_PTR(-ENOMEM);
7956 if (!alloc_fair_sched_group(tg, parent))
7959 if (!alloc_rt_sched_group(tg, parent))
7965 sched_free_group(tg);
7966 return ERR_PTR(-ENOMEM);
7969 void sched_online_group(struct task_group *tg, struct task_group *parent)
7971 unsigned long flags;
7973 spin_lock_irqsave(&task_group_lock, flags);
7974 list_add_rcu(&tg->list, &task_groups);
7976 WARN_ON(!parent); /* root should already exist */
7978 tg->parent = parent;
7979 INIT_LIST_HEAD(&tg->children);
7980 list_add_rcu(&tg->siblings, &parent->children);
7981 spin_unlock_irqrestore(&task_group_lock, flags);
7984 /* rcu callback to free various structures associated with a task group */
7985 static void sched_free_group_rcu(struct rcu_head *rhp)
7987 /* now it should be safe to free those cfs_rqs */
7988 sched_free_group(container_of(rhp, struct task_group, rcu));
7991 void sched_destroy_group(struct task_group *tg)
7993 /* wait for possible concurrent references to cfs_rqs complete */
7994 call_rcu(&tg->rcu, sched_free_group_rcu);
7997 void sched_offline_group(struct task_group *tg)
7999 unsigned long flags;
8002 /* end participation in shares distribution */
8003 for_each_possible_cpu(i)
8004 unregister_fair_sched_group(tg, i);
8006 spin_lock_irqsave(&task_group_lock, flags);
8007 list_del_rcu(&tg->list);
8008 list_del_rcu(&tg->siblings);
8009 spin_unlock_irqrestore(&task_group_lock, flags);
8012 /* change task's runqueue when it moves between groups.
8013 * The caller of this function should have put the task in its new group
8014 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8015 * reflect its new group.
8017 void sched_move_task(struct task_struct *tsk)
8019 struct task_group *tg;
8020 int queued, running;
8021 unsigned long flags;
8024 rq = task_rq_lock(tsk, &flags);
8026 running = task_current(rq, tsk);
8027 queued = task_on_rq_queued(tsk);
8030 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8031 if (unlikely(running))
8032 put_prev_task(rq, tsk);
8035 * All callers are synchronized by task_rq_lock(); we do not use RCU
8036 * which is pointless here. Thus, we pass "true" to task_css_check()
8037 * to prevent lockdep warnings.
8039 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8040 struct task_group, css);
8041 tg = autogroup_task_group(tsk, tg);
8042 tsk->sched_task_group = tg;
8044 #ifdef CONFIG_FAIR_GROUP_SCHED
8045 if (tsk->sched_class->task_move_group)
8046 tsk->sched_class->task_move_group(tsk);
8049 set_task_rq(tsk, task_cpu(tsk));
8051 if (unlikely(running))
8052 tsk->sched_class->set_curr_task(rq);
8054 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8056 task_rq_unlock(rq, tsk, &flags);
8058 #endif /* CONFIG_CGROUP_SCHED */
8060 #ifdef CONFIG_RT_GROUP_SCHED
8062 * Ensure that the real time constraints are schedulable.
8064 static DEFINE_MUTEX(rt_constraints_mutex);
8066 /* Must be called with tasklist_lock held */
8067 static inline int tg_has_rt_tasks(struct task_group *tg)
8069 struct task_struct *g, *p;
8072 * Autogroups do not have RT tasks; see autogroup_create().
8074 if (task_group_is_autogroup(tg))
8077 for_each_process_thread(g, p) {
8078 if (rt_task(p) && task_group(p) == tg)
8085 struct rt_schedulable_data {
8086 struct task_group *tg;
8091 static int tg_rt_schedulable(struct task_group *tg, void *data)
8093 struct rt_schedulable_data *d = data;
8094 struct task_group *child;
8095 unsigned long total, sum = 0;
8096 u64 period, runtime;
8098 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8099 runtime = tg->rt_bandwidth.rt_runtime;
8102 period = d->rt_period;
8103 runtime = d->rt_runtime;
8107 * Cannot have more runtime than the period.
8109 if (runtime > period && runtime != RUNTIME_INF)
8113 * Ensure we don't starve existing RT tasks.
8115 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8118 total = to_ratio(period, runtime);
8121 * Nobody can have more than the global setting allows.
8123 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8127 * The sum of our children's runtime should not exceed our own.
8129 list_for_each_entry_rcu(child, &tg->children, siblings) {
8130 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8131 runtime = child->rt_bandwidth.rt_runtime;
8133 if (child == d->tg) {
8134 period = d->rt_period;
8135 runtime = d->rt_runtime;
8138 sum += to_ratio(period, runtime);
8147 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8151 struct rt_schedulable_data data = {
8153 .rt_period = period,
8154 .rt_runtime = runtime,
8158 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8164 static int tg_set_rt_bandwidth(struct task_group *tg,
8165 u64 rt_period, u64 rt_runtime)
8170 * Disallowing the root group RT runtime is BAD, it would disallow the
8171 * kernel creating (and or operating) RT threads.
8173 if (tg == &root_task_group && rt_runtime == 0)
8176 /* No period doesn't make any sense. */
8180 mutex_lock(&rt_constraints_mutex);
8181 read_lock(&tasklist_lock);
8182 err = __rt_schedulable(tg, rt_period, rt_runtime);
8186 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8187 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8188 tg->rt_bandwidth.rt_runtime = rt_runtime;
8190 for_each_possible_cpu(i) {
8191 struct rt_rq *rt_rq = tg->rt_rq[i];
8193 raw_spin_lock(&rt_rq->rt_runtime_lock);
8194 rt_rq->rt_runtime = rt_runtime;
8195 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8197 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8199 read_unlock(&tasklist_lock);
8200 mutex_unlock(&rt_constraints_mutex);
8205 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8207 u64 rt_runtime, rt_period;
8209 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8210 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8211 if (rt_runtime_us < 0)
8212 rt_runtime = RUNTIME_INF;
8214 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8217 static long sched_group_rt_runtime(struct task_group *tg)
8221 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8224 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8225 do_div(rt_runtime_us, NSEC_PER_USEC);
8226 return rt_runtime_us;
8229 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8231 u64 rt_runtime, rt_period;
8233 rt_period = rt_period_us * NSEC_PER_USEC;
8234 rt_runtime = tg->rt_bandwidth.rt_runtime;
8236 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8239 static long sched_group_rt_period(struct task_group *tg)
8243 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8244 do_div(rt_period_us, NSEC_PER_USEC);
8245 return rt_period_us;
8247 #endif /* CONFIG_RT_GROUP_SCHED */
8249 #ifdef CONFIG_RT_GROUP_SCHED
8250 static int sched_rt_global_constraints(void)
8254 mutex_lock(&rt_constraints_mutex);
8255 read_lock(&tasklist_lock);
8256 ret = __rt_schedulable(NULL, 0, 0);
8257 read_unlock(&tasklist_lock);
8258 mutex_unlock(&rt_constraints_mutex);
8263 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8265 /* Don't accept realtime tasks when there is no way for them to run */
8266 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8272 #else /* !CONFIG_RT_GROUP_SCHED */
8273 static int sched_rt_global_constraints(void)
8275 unsigned long flags;
8278 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8279 for_each_possible_cpu(i) {
8280 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8282 raw_spin_lock(&rt_rq->rt_runtime_lock);
8283 rt_rq->rt_runtime = global_rt_runtime();
8284 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8286 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8290 #endif /* CONFIG_RT_GROUP_SCHED */
8292 static int sched_dl_global_validate(void)
8294 u64 runtime = global_rt_runtime();
8295 u64 period = global_rt_period();
8296 u64 new_bw = to_ratio(period, runtime);
8299 unsigned long flags;
8302 * Here we want to check the bandwidth not being set to some
8303 * value smaller than the currently allocated bandwidth in
8304 * any of the root_domains.
8306 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8307 * cycling on root_domains... Discussion on different/better
8308 * solutions is welcome!
8310 for_each_possible_cpu(cpu) {
8311 rcu_read_lock_sched();
8312 dl_b = dl_bw_of(cpu);
8314 raw_spin_lock_irqsave(&dl_b->lock, flags);
8315 if (new_bw < dl_b->total_bw)
8317 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8319 rcu_read_unlock_sched();
8328 static void sched_dl_do_global(void)
8333 unsigned long flags;
8335 def_dl_bandwidth.dl_period = global_rt_period();
8336 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8338 if (global_rt_runtime() != RUNTIME_INF)
8339 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8342 * FIXME: As above...
8344 for_each_possible_cpu(cpu) {
8345 rcu_read_lock_sched();
8346 dl_b = dl_bw_of(cpu);
8348 raw_spin_lock_irqsave(&dl_b->lock, flags);
8350 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8352 rcu_read_unlock_sched();
8356 static int sched_rt_global_validate(void)
8358 if (sysctl_sched_rt_period <= 0)
8361 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8362 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8368 static void sched_rt_do_global(void)
8370 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8371 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8374 int sched_rt_handler(struct ctl_table *table, int write,
8375 void __user *buffer, size_t *lenp,
8378 int old_period, old_runtime;
8379 static DEFINE_MUTEX(mutex);
8383 old_period = sysctl_sched_rt_period;
8384 old_runtime = sysctl_sched_rt_runtime;
8386 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8388 if (!ret && write) {
8389 ret = sched_rt_global_validate();
8393 ret = sched_dl_global_validate();
8397 ret = sched_rt_global_constraints();
8401 sched_rt_do_global();
8402 sched_dl_do_global();
8406 sysctl_sched_rt_period = old_period;
8407 sysctl_sched_rt_runtime = old_runtime;
8409 mutex_unlock(&mutex);
8414 int sched_rr_handler(struct ctl_table *table, int write,
8415 void __user *buffer, size_t *lenp,
8419 static DEFINE_MUTEX(mutex);
8422 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8423 /* make sure that internally we keep jiffies */
8424 /* also, writing zero resets timeslice to default */
8425 if (!ret && write) {
8426 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8427 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8429 mutex_unlock(&mutex);
8433 #ifdef CONFIG_CGROUP_SCHED
8435 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8437 return css ? container_of(css, struct task_group, css) : NULL;
8440 static struct cgroup_subsys_state *
8441 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8443 struct task_group *parent = css_tg(parent_css);
8444 struct task_group *tg;
8447 /* This is early initialization for the top cgroup */
8448 return &root_task_group.css;
8451 tg = sched_create_group(parent);
8453 return ERR_PTR(-ENOMEM);
8455 sched_online_group(tg, parent);
8460 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8462 struct task_group *tg = css_tg(css);
8464 sched_offline_group(tg);
8467 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8469 struct task_group *tg = css_tg(css);
8472 * Relies on the RCU grace period between css_released() and this.
8474 sched_free_group(tg);
8477 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8479 sched_move_task(task);
8482 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8484 struct task_struct *task;
8485 struct cgroup_subsys_state *css;
8487 cgroup_taskset_for_each(task, css, tset) {
8488 #ifdef CONFIG_RT_GROUP_SCHED
8489 if (!sched_rt_can_attach(css_tg(css), task))
8492 /* We don't support RT-tasks being in separate groups */
8493 if (task->sched_class != &fair_sched_class)
8500 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8502 struct task_struct *task;
8503 struct cgroup_subsys_state *css;
8505 cgroup_taskset_for_each(task, css, tset)
8506 sched_move_task(task);
8509 #ifdef CONFIG_FAIR_GROUP_SCHED
8510 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8511 struct cftype *cftype, u64 shareval)
8513 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8516 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8519 struct task_group *tg = css_tg(css);
8521 return (u64) scale_load_down(tg->shares);
8524 #ifdef CONFIG_CFS_BANDWIDTH
8525 static DEFINE_MUTEX(cfs_constraints_mutex);
8527 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8528 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8530 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8532 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8534 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8535 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8537 if (tg == &root_task_group)
8541 * Ensure we have at some amount of bandwidth every period. This is
8542 * to prevent reaching a state of large arrears when throttled via
8543 * entity_tick() resulting in prolonged exit starvation.
8545 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8549 * Likewise, bound things on the otherside by preventing insane quota
8550 * periods. This also allows us to normalize in computing quota
8553 if (period > max_cfs_quota_period)
8557 * Prevent race between setting of cfs_rq->runtime_enabled and
8558 * unthrottle_offline_cfs_rqs().
8561 mutex_lock(&cfs_constraints_mutex);
8562 ret = __cfs_schedulable(tg, period, quota);
8566 runtime_enabled = quota != RUNTIME_INF;
8567 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8569 * If we need to toggle cfs_bandwidth_used, off->on must occur
8570 * before making related changes, and on->off must occur afterwards
8572 if (runtime_enabled && !runtime_was_enabled)
8573 cfs_bandwidth_usage_inc();
8574 raw_spin_lock_irq(&cfs_b->lock);
8575 cfs_b->period = ns_to_ktime(period);
8576 cfs_b->quota = quota;
8578 __refill_cfs_bandwidth_runtime(cfs_b);
8579 /* restart the period timer (if active) to handle new period expiry */
8580 if (runtime_enabled)
8581 start_cfs_bandwidth(cfs_b);
8582 raw_spin_unlock_irq(&cfs_b->lock);
8584 for_each_online_cpu(i) {
8585 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8586 struct rq *rq = cfs_rq->rq;
8588 raw_spin_lock_irq(&rq->lock);
8589 cfs_rq->runtime_enabled = runtime_enabled;
8590 cfs_rq->runtime_remaining = 0;
8592 if (cfs_rq->throttled)
8593 unthrottle_cfs_rq(cfs_rq);
8594 raw_spin_unlock_irq(&rq->lock);
8596 if (runtime_was_enabled && !runtime_enabled)
8597 cfs_bandwidth_usage_dec();
8599 mutex_unlock(&cfs_constraints_mutex);
8605 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8609 period = ktime_to_ns(tg->cfs_bandwidth.period);
8610 if (cfs_quota_us < 0)
8611 quota = RUNTIME_INF;
8613 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8615 return tg_set_cfs_bandwidth(tg, period, quota);
8618 long tg_get_cfs_quota(struct task_group *tg)
8622 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8625 quota_us = tg->cfs_bandwidth.quota;
8626 do_div(quota_us, NSEC_PER_USEC);
8631 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8635 period = (u64)cfs_period_us * NSEC_PER_USEC;
8636 quota = tg->cfs_bandwidth.quota;
8638 return tg_set_cfs_bandwidth(tg, period, quota);
8641 long tg_get_cfs_period(struct task_group *tg)
8645 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8646 do_div(cfs_period_us, NSEC_PER_USEC);
8648 return cfs_period_us;
8651 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8654 return tg_get_cfs_quota(css_tg(css));
8657 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8658 struct cftype *cftype, s64 cfs_quota_us)
8660 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8663 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8666 return tg_get_cfs_period(css_tg(css));
8669 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8670 struct cftype *cftype, u64 cfs_period_us)
8672 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8675 struct cfs_schedulable_data {
8676 struct task_group *tg;
8681 * normalize group quota/period to be quota/max_period
8682 * note: units are usecs
8684 static u64 normalize_cfs_quota(struct task_group *tg,
8685 struct cfs_schedulable_data *d)
8693 period = tg_get_cfs_period(tg);
8694 quota = tg_get_cfs_quota(tg);
8697 /* note: these should typically be equivalent */
8698 if (quota == RUNTIME_INF || quota == -1)
8701 return to_ratio(period, quota);
8704 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8706 struct cfs_schedulable_data *d = data;
8707 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8708 s64 quota = 0, parent_quota = -1;
8711 quota = RUNTIME_INF;
8713 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8715 quota = normalize_cfs_quota(tg, d);
8716 parent_quota = parent_b->hierarchical_quota;
8719 * ensure max(child_quota) <= parent_quota, inherit when no
8722 if (quota == RUNTIME_INF)
8723 quota = parent_quota;
8724 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8727 cfs_b->hierarchical_quota = quota;
8732 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8735 struct cfs_schedulable_data data = {
8741 if (quota != RUNTIME_INF) {
8742 do_div(data.period, NSEC_PER_USEC);
8743 do_div(data.quota, NSEC_PER_USEC);
8747 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8753 static int cpu_stats_show(struct seq_file *sf, void *v)
8755 struct task_group *tg = css_tg(seq_css(sf));
8756 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8758 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8759 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8760 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8764 #endif /* CONFIG_CFS_BANDWIDTH */
8765 #endif /* CONFIG_FAIR_GROUP_SCHED */
8767 #ifdef CONFIG_RT_GROUP_SCHED
8768 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8769 struct cftype *cft, s64 val)
8771 return sched_group_set_rt_runtime(css_tg(css), val);
8774 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8777 return sched_group_rt_runtime(css_tg(css));
8780 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8781 struct cftype *cftype, u64 rt_period_us)
8783 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8786 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8789 return sched_group_rt_period(css_tg(css));
8791 #endif /* CONFIG_RT_GROUP_SCHED */
8793 static struct cftype cpu_files[] = {
8794 #ifdef CONFIG_FAIR_GROUP_SCHED
8797 .read_u64 = cpu_shares_read_u64,
8798 .write_u64 = cpu_shares_write_u64,
8801 #ifdef CONFIG_CFS_BANDWIDTH
8803 .name = "cfs_quota_us",
8804 .read_s64 = cpu_cfs_quota_read_s64,
8805 .write_s64 = cpu_cfs_quota_write_s64,
8808 .name = "cfs_period_us",
8809 .read_u64 = cpu_cfs_period_read_u64,
8810 .write_u64 = cpu_cfs_period_write_u64,
8814 .seq_show = cpu_stats_show,
8817 #ifdef CONFIG_RT_GROUP_SCHED
8819 .name = "rt_runtime_us",
8820 .read_s64 = cpu_rt_runtime_read,
8821 .write_s64 = cpu_rt_runtime_write,
8824 .name = "rt_period_us",
8825 .read_u64 = cpu_rt_period_read_uint,
8826 .write_u64 = cpu_rt_period_write_uint,
8832 struct cgroup_subsys cpu_cgrp_subsys = {
8833 .css_alloc = cpu_cgroup_css_alloc,
8834 .css_released = cpu_cgroup_css_released,
8835 .css_free = cpu_cgroup_css_free,
8836 .fork = cpu_cgroup_fork,
8837 .can_attach = cpu_cgroup_can_attach,
8838 .attach = cpu_cgroup_attach,
8839 .allow_attach = subsys_cgroup_allow_attach,
8840 .legacy_cftypes = cpu_files,
8844 #endif /* CONFIG_CGROUP_SCHED */
8846 void dump_cpu_task(int cpu)
8848 pr_info("Task dump for CPU %d:\n", cpu);
8849 sched_show_task(cpu_curr(cpu));