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 * this_rq_lock - lock this runqueue and disable interrupts.
293 static struct rq *this_rq_lock(void)
300 raw_spin_lock(&rq->lock);
305 #ifdef CONFIG_SCHED_HRTICK
307 * Use HR-timers to deliver accurate preemption points.
310 static void hrtick_clear(struct rq *rq)
312 if (hrtimer_active(&rq->hrtick_timer))
313 hrtimer_cancel(&rq->hrtick_timer);
317 * High-resolution timer tick.
318 * Runs from hardirq context with interrupts disabled.
320 static enum hrtimer_restart hrtick(struct hrtimer *timer)
322 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
324 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
326 raw_spin_lock(&rq->lock);
328 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
329 raw_spin_unlock(&rq->lock);
331 return HRTIMER_NORESTART;
336 static void __hrtick_restart(struct rq *rq)
338 struct hrtimer *timer = &rq->hrtick_timer;
340 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
344 * called from hardirq (IPI) context
346 static void __hrtick_start(void *arg)
350 raw_spin_lock(&rq->lock);
351 __hrtick_restart(rq);
352 rq->hrtick_csd_pending = 0;
353 raw_spin_unlock(&rq->lock);
357 * Called to set the hrtick timer state.
359 * called with rq->lock held and irqs disabled
361 void hrtick_start(struct rq *rq, u64 delay)
363 struct hrtimer *timer = &rq->hrtick_timer;
368 * Don't schedule slices shorter than 10000ns, that just
369 * doesn't make sense and can cause timer DoS.
371 delta = max_t(s64, delay, 10000LL);
372 time = ktime_add_ns(timer->base->get_time(), delta);
374 hrtimer_set_expires(timer, time);
376 if (rq == this_rq()) {
377 __hrtick_restart(rq);
378 } else if (!rq->hrtick_csd_pending) {
379 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
380 rq->hrtick_csd_pending = 1;
385 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
387 int cpu = (int)(long)hcpu;
390 case CPU_UP_CANCELED:
391 case CPU_UP_CANCELED_FROZEN:
392 case CPU_DOWN_PREPARE:
393 case CPU_DOWN_PREPARE_FROZEN:
395 case CPU_DEAD_FROZEN:
396 hrtick_clear(cpu_rq(cpu));
403 static __init void init_hrtick(void)
405 hotcpu_notifier(hotplug_hrtick, 0);
409 * Called to set the hrtick timer state.
411 * called with rq->lock held and irqs disabled
413 void hrtick_start(struct rq *rq, u64 delay)
416 * Don't schedule slices shorter than 10000ns, that just
417 * doesn't make sense. Rely on vruntime for fairness.
419 delay = max_t(u64, delay, 10000LL);
420 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
421 HRTIMER_MODE_REL_PINNED);
424 static inline void init_hrtick(void)
427 #endif /* CONFIG_SMP */
429 static void init_rq_hrtick(struct rq *rq)
432 rq->hrtick_csd_pending = 0;
434 rq->hrtick_csd.flags = 0;
435 rq->hrtick_csd.func = __hrtick_start;
436 rq->hrtick_csd.info = rq;
439 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
440 rq->hrtick_timer.function = hrtick;
442 #else /* CONFIG_SCHED_HRTICK */
443 static inline void hrtick_clear(struct rq *rq)
447 static inline void init_rq_hrtick(struct rq *rq)
451 static inline void init_hrtick(void)
454 #endif /* CONFIG_SCHED_HRTICK */
457 * cmpxchg based fetch_or, macro so it works for different integer types
459 #define fetch_or(ptr, val) \
460 ({ typeof(*(ptr)) __old, __val = *(ptr); \
462 __old = cmpxchg((ptr), __val, __val | (val)); \
463 if (__old == __val) \
470 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
472 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
473 * this avoids any races wrt polling state changes and thereby avoids
476 static bool set_nr_and_not_polling(struct task_struct *p)
478 struct thread_info *ti = task_thread_info(p);
479 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
483 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
485 * If this returns true, then the idle task promises to call
486 * sched_ttwu_pending() and reschedule soon.
488 static bool set_nr_if_polling(struct task_struct *p)
490 struct thread_info *ti = task_thread_info(p);
491 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
494 if (!(val & _TIF_POLLING_NRFLAG))
496 if (val & _TIF_NEED_RESCHED)
498 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
507 static bool set_nr_and_not_polling(struct task_struct *p)
509 set_tsk_need_resched(p);
514 static bool set_nr_if_polling(struct task_struct *p)
521 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
523 struct wake_q_node *node = &task->wake_q;
526 * Atomically grab the task, if ->wake_q is !nil already it means
527 * its already queued (either by us or someone else) and will get the
528 * wakeup due to that.
530 * This cmpxchg() implies a full barrier, which pairs with the write
531 * barrier implied by the wakeup in wake_up_list().
533 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
536 get_task_struct(task);
539 * The head is context local, there can be no concurrency.
542 head->lastp = &node->next;
545 void wake_up_q(struct wake_q_head *head)
547 struct wake_q_node *node = head->first;
549 while (node != WAKE_Q_TAIL) {
550 struct task_struct *task;
552 task = container_of(node, struct task_struct, wake_q);
554 /* task can safely be re-inserted now */
556 task->wake_q.next = NULL;
559 * wake_up_process() implies a wmb() to pair with the queueing
560 * in wake_q_add() so as not to miss wakeups.
562 wake_up_process(task);
563 put_task_struct(task);
568 * resched_curr - mark rq's current task 'to be rescheduled now'.
570 * On UP this means the setting of the need_resched flag, on SMP it
571 * might also involve a cross-CPU call to trigger the scheduler on
574 void resched_curr(struct rq *rq)
576 struct task_struct *curr = rq->curr;
579 lockdep_assert_held(&rq->lock);
581 if (test_tsk_need_resched(curr))
586 if (cpu == smp_processor_id()) {
587 set_tsk_need_resched(curr);
588 set_preempt_need_resched();
592 if (set_nr_and_not_polling(curr))
593 smp_send_reschedule(cpu);
595 trace_sched_wake_idle_without_ipi(cpu);
598 void resched_cpu(int cpu)
600 struct rq *rq = cpu_rq(cpu);
603 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
606 raw_spin_unlock_irqrestore(&rq->lock, flags);
610 #ifdef CONFIG_NO_HZ_COMMON
612 * In the semi idle case, use the nearest busy cpu for migrating timers
613 * from an idle cpu. This is good for power-savings.
615 * We don't do similar optimization for completely idle system, as
616 * selecting an idle cpu will add more delays to the timers than intended
617 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
619 int get_nohz_timer_target(void)
621 int i, cpu = smp_processor_id();
622 struct sched_domain *sd;
624 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
628 for_each_domain(cpu, sd) {
629 for_each_cpu(i, sched_domain_span(sd)) {
630 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
637 if (!is_housekeeping_cpu(cpu))
638 cpu = housekeeping_any_cpu();
644 * When add_timer_on() enqueues a timer into the timer wheel of an
645 * idle CPU then this timer might expire before the next timer event
646 * which is scheduled to wake up that CPU. In case of a completely
647 * idle system the next event might even be infinite time into the
648 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
649 * leaves the inner idle loop so the newly added timer is taken into
650 * account when the CPU goes back to idle and evaluates the timer
651 * wheel for the next timer event.
653 static void wake_up_idle_cpu(int cpu)
655 struct rq *rq = cpu_rq(cpu);
657 if (cpu == smp_processor_id())
660 if (set_nr_and_not_polling(rq->idle))
661 smp_send_reschedule(cpu);
663 trace_sched_wake_idle_without_ipi(cpu);
666 static bool wake_up_full_nohz_cpu(int cpu)
669 * We just need the target to call irq_exit() and re-evaluate
670 * the next tick. The nohz full kick at least implies that.
671 * If needed we can still optimize that later with an
674 if (tick_nohz_full_cpu(cpu)) {
675 if (cpu != smp_processor_id() ||
676 tick_nohz_tick_stopped())
677 tick_nohz_full_kick_cpu(cpu);
684 void wake_up_nohz_cpu(int cpu)
686 if (!wake_up_full_nohz_cpu(cpu))
687 wake_up_idle_cpu(cpu);
690 static inline bool got_nohz_idle_kick(void)
692 int cpu = smp_processor_id();
694 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
697 if (idle_cpu(cpu) && !need_resched())
701 * We can't run Idle Load Balance on this CPU for this time so we
702 * cancel it and clear NOHZ_BALANCE_KICK
704 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
708 #else /* CONFIG_NO_HZ_COMMON */
710 static inline bool got_nohz_idle_kick(void)
715 #endif /* CONFIG_NO_HZ_COMMON */
717 #ifdef CONFIG_NO_HZ_FULL
718 bool sched_can_stop_tick(void)
721 * FIFO realtime policy runs the highest priority task. Other runnable
722 * tasks are of a lower priority. The scheduler tick does nothing.
724 if (current->policy == SCHED_FIFO)
728 * Round-robin realtime tasks time slice with other tasks at the same
729 * realtime priority. Is this task the only one at this priority?
731 if (current->policy == SCHED_RR) {
732 struct sched_rt_entity *rt_se = ¤t->rt;
734 return rt_se->run_list.prev == rt_se->run_list.next;
738 * More than one running task need preemption.
739 * nr_running update is assumed to be visible
740 * after IPI is sent from wakers.
742 if (this_rq()->nr_running > 1)
747 #endif /* CONFIG_NO_HZ_FULL */
749 void sched_avg_update(struct rq *rq)
751 s64 period = sched_avg_period();
753 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
755 * Inline assembly required to prevent the compiler
756 * optimising this loop into a divmod call.
757 * See __iter_div_u64_rem() for another example of this.
759 asm("" : "+rm" (rq->age_stamp));
760 rq->age_stamp += period;
765 #endif /* CONFIG_SMP */
767 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
768 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
770 * Iterate task_group tree rooted at *from, calling @down when first entering a
771 * node and @up when leaving it for the final time.
773 * Caller must hold rcu_lock or sufficient equivalent.
775 int walk_tg_tree_from(struct task_group *from,
776 tg_visitor down, tg_visitor up, void *data)
778 struct task_group *parent, *child;
784 ret = (*down)(parent, data);
787 list_for_each_entry_rcu(child, &parent->children, siblings) {
794 ret = (*up)(parent, data);
795 if (ret || parent == from)
799 parent = parent->parent;
806 int tg_nop(struct task_group *tg, void *data)
812 static void set_load_weight(struct task_struct *p)
814 int prio = p->static_prio - MAX_RT_PRIO;
815 struct load_weight *load = &p->se.load;
818 * SCHED_IDLE tasks get minimal weight:
820 if (p->policy == SCHED_IDLE) {
821 load->weight = scale_load(WEIGHT_IDLEPRIO);
822 load->inv_weight = WMULT_IDLEPRIO;
826 load->weight = scale_load(prio_to_weight[prio]);
827 load->inv_weight = prio_to_wmult[prio];
830 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
833 sched_info_queued(rq, p);
834 p->sched_class->enqueue_task(rq, p, flags);
837 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
840 sched_info_dequeued(rq, p);
841 p->sched_class->dequeue_task(rq, p, flags);
844 void activate_task(struct rq *rq, struct task_struct *p, int flags)
846 if (task_contributes_to_load(p))
847 rq->nr_uninterruptible--;
849 enqueue_task(rq, p, flags);
852 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
854 if (task_contributes_to_load(p))
855 rq->nr_uninterruptible++;
857 dequeue_task(rq, p, flags);
860 static void update_rq_clock_task(struct rq *rq, s64 delta)
863 * In theory, the compile should just see 0 here, and optimize out the call
864 * to sched_rt_avg_update. But I don't trust it...
866 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
867 s64 steal = 0, irq_delta = 0;
869 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
870 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
873 * Since irq_time is only updated on {soft,}irq_exit, we might run into
874 * this case when a previous update_rq_clock() happened inside a
877 * When this happens, we stop ->clock_task and only update the
878 * prev_irq_time stamp to account for the part that fit, so that a next
879 * update will consume the rest. This ensures ->clock_task is
882 * It does however cause some slight miss-attribution of {soft,}irq
883 * time, a more accurate solution would be to update the irq_time using
884 * the current rq->clock timestamp, except that would require using
887 if (irq_delta > delta)
890 rq->prev_irq_time += irq_delta;
893 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
894 if (static_key_false((¶virt_steal_rq_enabled))) {
895 steal = paravirt_steal_clock(cpu_of(rq));
896 steal -= rq->prev_steal_time_rq;
898 if (unlikely(steal > delta))
901 rq->prev_steal_time_rq += steal;
906 rq->clock_task += delta;
908 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
909 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
910 sched_rt_avg_update(rq, irq_delta + steal);
914 void sched_set_stop_task(int cpu, struct task_struct *stop)
916 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
917 struct task_struct *old_stop = cpu_rq(cpu)->stop;
921 * Make it appear like a SCHED_FIFO task, its something
922 * userspace knows about and won't get confused about.
924 * Also, it will make PI more or less work without too
925 * much confusion -- but then, stop work should not
926 * rely on PI working anyway.
928 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
930 stop->sched_class = &stop_sched_class;
933 cpu_rq(cpu)->stop = stop;
937 * Reset it back to a normal scheduling class so that
938 * it can die in pieces.
940 old_stop->sched_class = &rt_sched_class;
945 * __normal_prio - return the priority that is based on the static prio
947 static inline int __normal_prio(struct task_struct *p)
949 return p->static_prio;
953 * Calculate the expected normal priority: i.e. priority
954 * without taking RT-inheritance into account. Might be
955 * boosted by interactivity modifiers. Changes upon fork,
956 * setprio syscalls, and whenever the interactivity
957 * estimator recalculates.
959 static inline int normal_prio(struct task_struct *p)
963 if (task_has_dl_policy(p))
964 prio = MAX_DL_PRIO-1;
965 else if (task_has_rt_policy(p))
966 prio = MAX_RT_PRIO-1 - p->rt_priority;
968 prio = __normal_prio(p);
973 * Calculate the current priority, i.e. the priority
974 * taken into account by the scheduler. This value might
975 * be boosted by RT tasks, or might be boosted by
976 * interactivity modifiers. Will be RT if the task got
977 * RT-boosted. If not then it returns p->normal_prio.
979 static int effective_prio(struct task_struct *p)
981 p->normal_prio = normal_prio(p);
983 * If we are RT tasks or we were boosted to RT priority,
984 * keep the priority unchanged. Otherwise, update priority
985 * to the normal priority:
987 if (!rt_prio(p->prio))
988 return p->normal_prio;
993 * task_curr - is this task currently executing on a CPU?
994 * @p: the task in question.
996 * Return: 1 if the task is currently executing. 0 otherwise.
998 inline int task_curr(const struct task_struct *p)
1000 return cpu_curr(task_cpu(p)) == p;
1004 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1005 * use the balance_callback list if you want balancing.
1007 * this means any call to check_class_changed() must be followed by a call to
1008 * balance_callback().
1010 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1011 const struct sched_class *prev_class,
1014 if (prev_class != p->sched_class) {
1015 if (prev_class->switched_from)
1016 prev_class->switched_from(rq, p);
1018 p->sched_class->switched_to(rq, p);
1019 } else if (oldprio != p->prio || dl_task(p))
1020 p->sched_class->prio_changed(rq, p, oldprio);
1023 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1025 const struct sched_class *class;
1027 if (p->sched_class == rq->curr->sched_class) {
1028 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1030 for_each_class(class) {
1031 if (class == rq->curr->sched_class)
1033 if (class == p->sched_class) {
1041 * A queue event has occurred, and we're going to schedule. In
1042 * this case, we can save a useless back to back clock update.
1044 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1045 rq_clock_skip_update(rq, true);
1050 * This is how migration works:
1052 * 1) we invoke migration_cpu_stop() on the target CPU using
1054 * 2) stopper starts to run (implicitly forcing the migrated thread
1056 * 3) it checks whether the migrated task is still in the wrong runqueue.
1057 * 4) if it's in the wrong runqueue then the migration thread removes
1058 * it and puts it into the right queue.
1059 * 5) stopper completes and stop_one_cpu() returns and the migration
1064 * move_queued_task - move a queued task to new rq.
1066 * Returns (locked) new rq. Old rq's lock is released.
1068 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1070 lockdep_assert_held(&rq->lock);
1072 dequeue_task(rq, p, 0);
1073 p->on_rq = TASK_ON_RQ_MIGRATING;
1074 set_task_cpu(p, new_cpu);
1075 raw_spin_unlock(&rq->lock);
1077 rq = cpu_rq(new_cpu);
1079 raw_spin_lock(&rq->lock);
1080 BUG_ON(task_cpu(p) != new_cpu);
1081 p->on_rq = TASK_ON_RQ_QUEUED;
1082 enqueue_task(rq, p, 0);
1083 check_preempt_curr(rq, p, 0);
1088 struct migration_arg {
1089 struct task_struct *task;
1094 * Move (not current) task off this cpu, onto dest cpu. We're doing
1095 * this because either it can't run here any more (set_cpus_allowed()
1096 * away from this CPU, or CPU going down), or because we're
1097 * attempting to rebalance this task on exec (sched_exec).
1099 * So we race with normal scheduler movements, but that's OK, as long
1100 * as the task is no longer on this CPU.
1102 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1104 if (unlikely(!cpu_active(dest_cpu)))
1107 /* Affinity changed (again). */
1108 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1111 rq = move_queued_task(rq, p, dest_cpu);
1117 * migration_cpu_stop - this will be executed by a highprio stopper thread
1118 * and performs thread migration by bumping thread off CPU then
1119 * 'pushing' onto another runqueue.
1121 static int migration_cpu_stop(void *data)
1123 struct migration_arg *arg = data;
1124 struct task_struct *p = arg->task;
1125 struct rq *rq = this_rq();
1128 * The original target cpu might have gone down and we might
1129 * be on another cpu but it doesn't matter.
1131 local_irq_disable();
1133 * We need to explicitly wake pending tasks before running
1134 * __migrate_task() such that we will not miss enforcing cpus_allowed
1135 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1137 sched_ttwu_pending();
1139 raw_spin_lock(&p->pi_lock);
1140 raw_spin_lock(&rq->lock);
1142 * If task_rq(p) != rq, it cannot be migrated here, because we're
1143 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1144 * we're holding p->pi_lock.
1146 if (task_rq(p) == rq && task_on_rq_queued(p))
1147 rq = __migrate_task(rq, p, arg->dest_cpu);
1148 raw_spin_unlock(&rq->lock);
1149 raw_spin_unlock(&p->pi_lock);
1156 * sched_class::set_cpus_allowed must do the below, but is not required to
1157 * actually call this function.
1159 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1161 cpumask_copy(&p->cpus_allowed, new_mask);
1162 p->nr_cpus_allowed = cpumask_weight(new_mask);
1165 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1167 struct rq *rq = task_rq(p);
1168 bool queued, running;
1170 lockdep_assert_held(&p->pi_lock);
1172 queued = task_on_rq_queued(p);
1173 running = task_current(rq, p);
1177 * Because __kthread_bind() calls this on blocked tasks without
1180 lockdep_assert_held(&rq->lock);
1181 dequeue_task(rq, p, 0);
1184 put_prev_task(rq, p);
1186 p->sched_class->set_cpus_allowed(p, new_mask);
1189 p->sched_class->set_curr_task(rq);
1191 enqueue_task(rq, p, 0);
1195 * Change a given task's CPU affinity. Migrate the thread to a
1196 * proper CPU and schedule it away if the CPU it's executing on
1197 * is removed from the allowed bitmask.
1199 * NOTE: the caller must have a valid reference to the task, the
1200 * task must not exit() & deallocate itself prematurely. The
1201 * call is not atomic; no spinlocks may be held.
1203 static int __set_cpus_allowed_ptr(struct task_struct *p,
1204 const struct cpumask *new_mask, bool check)
1206 unsigned long flags;
1208 unsigned int dest_cpu;
1211 rq = task_rq_lock(p, &flags);
1214 * Must re-check here, to close a race against __kthread_bind(),
1215 * sched_setaffinity() is not guaranteed to observe the flag.
1217 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1222 if (cpumask_equal(&p->cpus_allowed, new_mask))
1225 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1230 do_set_cpus_allowed(p, new_mask);
1232 /* Can the task run on the task's current CPU? If so, we're done */
1233 if (cpumask_test_cpu(task_cpu(p), new_mask))
1236 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1237 if (task_running(rq, p) || p->state == TASK_WAKING) {
1238 struct migration_arg arg = { p, dest_cpu };
1239 /* Need help from migration thread: drop lock and wait. */
1240 task_rq_unlock(rq, p, &flags);
1241 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1242 tlb_migrate_finish(p->mm);
1244 } else if (task_on_rq_queued(p)) {
1246 * OK, since we're going to drop the lock immediately
1247 * afterwards anyway.
1249 lockdep_unpin_lock(&rq->lock);
1250 rq = move_queued_task(rq, p, dest_cpu);
1251 lockdep_pin_lock(&rq->lock);
1254 task_rq_unlock(rq, p, &flags);
1259 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1261 return __set_cpus_allowed_ptr(p, new_mask, false);
1263 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1265 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1267 #ifdef CONFIG_SCHED_DEBUG
1269 * We should never call set_task_cpu() on a blocked task,
1270 * ttwu() will sort out the placement.
1272 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1275 #ifdef CONFIG_LOCKDEP
1277 * The caller should hold either p->pi_lock or rq->lock, when changing
1278 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1280 * sched_move_task() holds both and thus holding either pins the cgroup,
1283 * Furthermore, all task_rq users should acquire both locks, see
1286 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1287 lockdep_is_held(&task_rq(p)->lock)));
1291 trace_sched_migrate_task(p, new_cpu);
1293 if (task_cpu(p) != new_cpu) {
1294 if (p->sched_class->migrate_task_rq)
1295 p->sched_class->migrate_task_rq(p, new_cpu);
1296 p->se.nr_migrations++;
1297 perf_event_task_migrate(p);
1300 __set_task_cpu(p, new_cpu);
1303 static void __migrate_swap_task(struct task_struct *p, int cpu)
1305 if (task_on_rq_queued(p)) {
1306 struct rq *src_rq, *dst_rq;
1308 src_rq = task_rq(p);
1309 dst_rq = cpu_rq(cpu);
1311 deactivate_task(src_rq, p, 0);
1312 set_task_cpu(p, cpu);
1313 activate_task(dst_rq, p, 0);
1314 check_preempt_curr(dst_rq, p, 0);
1317 * Task isn't running anymore; make it appear like we migrated
1318 * it before it went to sleep. This means on wakeup we make the
1319 * previous cpu our targer instead of where it really is.
1325 struct migration_swap_arg {
1326 struct task_struct *src_task, *dst_task;
1327 int src_cpu, dst_cpu;
1330 static int migrate_swap_stop(void *data)
1332 struct migration_swap_arg *arg = data;
1333 struct rq *src_rq, *dst_rq;
1336 src_rq = cpu_rq(arg->src_cpu);
1337 dst_rq = cpu_rq(arg->dst_cpu);
1339 double_raw_lock(&arg->src_task->pi_lock,
1340 &arg->dst_task->pi_lock);
1341 double_rq_lock(src_rq, dst_rq);
1342 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1345 if (task_cpu(arg->src_task) != arg->src_cpu)
1348 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1351 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1354 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1355 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1360 double_rq_unlock(src_rq, dst_rq);
1361 raw_spin_unlock(&arg->dst_task->pi_lock);
1362 raw_spin_unlock(&arg->src_task->pi_lock);
1368 * Cross migrate two tasks
1370 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1372 struct migration_swap_arg arg;
1375 arg = (struct migration_swap_arg){
1377 .src_cpu = task_cpu(cur),
1379 .dst_cpu = task_cpu(p),
1382 if (arg.src_cpu == arg.dst_cpu)
1386 * These three tests are all lockless; this is OK since all of them
1387 * will be re-checked with proper locks held further down the line.
1389 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1392 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1395 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1398 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1399 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1406 * wait_task_inactive - wait for a thread to unschedule.
1408 * If @match_state is nonzero, it's the @p->state value just checked and
1409 * not expected to change. If it changes, i.e. @p might have woken up,
1410 * then return zero. When we succeed in waiting for @p to be off its CPU,
1411 * we return a positive number (its total switch count). If a second call
1412 * a short while later returns the same number, the caller can be sure that
1413 * @p has remained unscheduled the whole time.
1415 * The caller must ensure that the task *will* unschedule sometime soon,
1416 * else this function might spin for a *long* time. This function can't
1417 * be called with interrupts off, or it may introduce deadlock with
1418 * smp_call_function() if an IPI is sent by the same process we are
1419 * waiting to become inactive.
1421 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1423 unsigned long flags;
1424 int running, queued;
1430 * We do the initial early heuristics without holding
1431 * any task-queue locks at all. We'll only try to get
1432 * the runqueue lock when things look like they will
1438 * If the task is actively running on another CPU
1439 * still, just relax and busy-wait without holding
1442 * NOTE! Since we don't hold any locks, it's not
1443 * even sure that "rq" stays as the right runqueue!
1444 * But we don't care, since "task_running()" will
1445 * return false if the runqueue has changed and p
1446 * is actually now running somewhere else!
1448 while (task_running(rq, p)) {
1449 if (match_state && unlikely(p->state != match_state))
1455 * Ok, time to look more closely! We need the rq
1456 * lock now, to be *sure*. If we're wrong, we'll
1457 * just go back and repeat.
1459 rq = task_rq_lock(p, &flags);
1460 trace_sched_wait_task(p);
1461 running = task_running(rq, p);
1462 queued = task_on_rq_queued(p);
1464 if (!match_state || p->state == match_state)
1465 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1466 task_rq_unlock(rq, p, &flags);
1469 * If it changed from the expected state, bail out now.
1471 if (unlikely(!ncsw))
1475 * Was it really running after all now that we
1476 * checked with the proper locks actually held?
1478 * Oops. Go back and try again..
1480 if (unlikely(running)) {
1486 * It's not enough that it's not actively running,
1487 * it must be off the runqueue _entirely_, and not
1490 * So if it was still runnable (but just not actively
1491 * running right now), it's preempted, and we should
1492 * yield - it could be a while.
1494 if (unlikely(queued)) {
1495 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1497 set_current_state(TASK_UNINTERRUPTIBLE);
1498 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1503 * Ahh, all good. It wasn't running, and it wasn't
1504 * runnable, which means that it will never become
1505 * running in the future either. We're all done!
1514 * kick_process - kick a running thread to enter/exit the kernel
1515 * @p: the to-be-kicked thread
1517 * Cause a process which is running on another CPU to enter
1518 * kernel-mode, without any delay. (to get signals handled.)
1520 * NOTE: this function doesn't have to take the runqueue lock,
1521 * because all it wants to ensure is that the remote task enters
1522 * the kernel. If the IPI races and the task has been migrated
1523 * to another CPU then no harm is done and the purpose has been
1526 void kick_process(struct task_struct *p)
1532 if ((cpu != smp_processor_id()) && task_curr(p))
1533 smp_send_reschedule(cpu);
1536 EXPORT_SYMBOL_GPL(kick_process);
1539 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1541 static int select_fallback_rq(int cpu, struct task_struct *p)
1543 int nid = cpu_to_node(cpu);
1544 const struct cpumask *nodemask = NULL;
1545 enum { cpuset, possible, fail } state = cpuset;
1549 * If the node that the cpu is on has been offlined, cpu_to_node()
1550 * will return -1. There is no cpu on the node, and we should
1551 * select the cpu on the other node.
1554 nodemask = cpumask_of_node(nid);
1556 /* Look for allowed, online CPU in same node. */
1557 for_each_cpu(dest_cpu, nodemask) {
1558 if (!cpu_online(dest_cpu))
1560 if (!cpu_active(dest_cpu))
1562 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1568 /* Any allowed, online CPU? */
1569 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1570 if (!cpu_online(dest_cpu))
1572 if (!cpu_active(dest_cpu))
1579 /* No more Mr. Nice Guy. */
1580 cpuset_cpus_allowed_fallback(p);
1585 do_set_cpus_allowed(p, cpu_possible_mask);
1596 if (state != cpuset) {
1598 * Don't tell them about moving exiting tasks or
1599 * kernel threads (both mm NULL), since they never
1602 if (p->mm && printk_ratelimit()) {
1603 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1604 task_pid_nr(p), p->comm, cpu);
1612 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1615 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1617 lockdep_assert_held(&p->pi_lock);
1619 if (p->nr_cpus_allowed > 1)
1620 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1623 * In order not to call set_task_cpu() on a blocking task we need
1624 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1627 * Since this is common to all placement strategies, this lives here.
1629 * [ this allows ->select_task() to simply return task_cpu(p) and
1630 * not worry about this generic constraint ]
1632 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1634 cpu = select_fallback_rq(task_cpu(p), p);
1639 static void update_avg(u64 *avg, u64 sample)
1641 s64 diff = sample - *avg;
1647 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1648 const struct cpumask *new_mask, bool check)
1650 return set_cpus_allowed_ptr(p, new_mask);
1653 #endif /* CONFIG_SMP */
1656 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1658 #ifdef CONFIG_SCHEDSTATS
1659 struct rq *rq = this_rq();
1662 int this_cpu = smp_processor_id();
1664 if (cpu == this_cpu) {
1665 schedstat_inc(rq, ttwu_local);
1666 schedstat_inc(p, se.statistics.nr_wakeups_local);
1668 struct sched_domain *sd;
1670 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1672 for_each_domain(this_cpu, sd) {
1673 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1674 schedstat_inc(sd, ttwu_wake_remote);
1681 if (wake_flags & WF_MIGRATED)
1682 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1684 #endif /* CONFIG_SMP */
1686 schedstat_inc(rq, ttwu_count);
1687 schedstat_inc(p, se.statistics.nr_wakeups);
1689 if (wake_flags & WF_SYNC)
1690 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1692 #endif /* CONFIG_SCHEDSTATS */
1695 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1697 activate_task(rq, p, en_flags);
1698 p->on_rq = TASK_ON_RQ_QUEUED;
1700 /* if a worker is waking up, notify workqueue */
1701 if (p->flags & PF_WQ_WORKER)
1702 wq_worker_waking_up(p, cpu_of(rq));
1706 * Mark the task runnable and perform wakeup-preemption.
1709 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1711 check_preempt_curr(rq, p, wake_flags);
1712 p->state = TASK_RUNNING;
1713 trace_sched_wakeup(p);
1716 if (p->sched_class->task_woken) {
1718 * Our task @p is fully woken up and running; so its safe to
1719 * drop the rq->lock, hereafter rq is only used for statistics.
1721 lockdep_unpin_lock(&rq->lock);
1722 p->sched_class->task_woken(rq, p);
1723 lockdep_pin_lock(&rq->lock);
1726 if (rq->idle_stamp) {
1727 u64 delta = rq_clock(rq) - rq->idle_stamp;
1728 u64 max = 2*rq->max_idle_balance_cost;
1730 update_avg(&rq->avg_idle, delta);
1732 if (rq->avg_idle > max)
1741 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1743 lockdep_assert_held(&rq->lock);
1746 if (p->sched_contributes_to_load)
1747 rq->nr_uninterruptible--;
1750 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1751 ttwu_do_wakeup(rq, p, wake_flags);
1755 * Called in case the task @p isn't fully descheduled from its runqueue,
1756 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1757 * since all we need to do is flip p->state to TASK_RUNNING, since
1758 * the task is still ->on_rq.
1760 static int ttwu_remote(struct task_struct *p, int wake_flags)
1765 rq = __task_rq_lock(p);
1766 if (task_on_rq_queued(p)) {
1767 /* check_preempt_curr() may use rq clock */
1768 update_rq_clock(rq);
1769 ttwu_do_wakeup(rq, p, wake_flags);
1772 __task_rq_unlock(rq);
1778 void sched_ttwu_pending(void)
1780 struct rq *rq = this_rq();
1781 struct llist_node *llist = llist_del_all(&rq->wake_list);
1782 struct task_struct *p;
1783 unsigned long flags;
1788 raw_spin_lock_irqsave(&rq->lock, flags);
1789 lockdep_pin_lock(&rq->lock);
1792 p = llist_entry(llist, struct task_struct, wake_entry);
1793 llist = llist_next(llist);
1794 ttwu_do_activate(rq, p, 0);
1797 lockdep_unpin_lock(&rq->lock);
1798 raw_spin_unlock_irqrestore(&rq->lock, flags);
1801 void scheduler_ipi(void)
1804 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1805 * TIF_NEED_RESCHED remotely (for the first time) will also send
1808 preempt_fold_need_resched();
1810 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1814 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1815 * traditionally all their work was done from the interrupt return
1816 * path. Now that we actually do some work, we need to make sure
1819 * Some archs already do call them, luckily irq_enter/exit nest
1822 * Arguably we should visit all archs and update all handlers,
1823 * however a fair share of IPIs are still resched only so this would
1824 * somewhat pessimize the simple resched case.
1827 sched_ttwu_pending();
1830 * Check if someone kicked us for doing the nohz idle load balance.
1832 if (unlikely(got_nohz_idle_kick())) {
1833 this_rq()->idle_balance = 1;
1834 raise_softirq_irqoff(SCHED_SOFTIRQ);
1839 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1841 struct rq *rq = cpu_rq(cpu);
1843 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1844 if (!set_nr_if_polling(rq->idle))
1845 smp_send_reschedule(cpu);
1847 trace_sched_wake_idle_without_ipi(cpu);
1851 void wake_up_if_idle(int cpu)
1853 struct rq *rq = cpu_rq(cpu);
1854 unsigned long flags;
1858 if (!is_idle_task(rcu_dereference(rq->curr)))
1861 if (set_nr_if_polling(rq->idle)) {
1862 trace_sched_wake_idle_without_ipi(cpu);
1864 raw_spin_lock_irqsave(&rq->lock, flags);
1865 if (is_idle_task(rq->curr))
1866 smp_send_reschedule(cpu);
1867 /* Else cpu is not in idle, do nothing here */
1868 raw_spin_unlock_irqrestore(&rq->lock, flags);
1875 bool cpus_share_cache(int this_cpu, int that_cpu)
1877 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1879 #endif /* CONFIG_SMP */
1881 static void ttwu_queue(struct task_struct *p, int cpu)
1883 struct rq *rq = cpu_rq(cpu);
1885 #if defined(CONFIG_SMP)
1886 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1887 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1888 ttwu_queue_remote(p, cpu);
1893 raw_spin_lock(&rq->lock);
1894 lockdep_pin_lock(&rq->lock);
1895 ttwu_do_activate(rq, p, 0);
1896 lockdep_unpin_lock(&rq->lock);
1897 raw_spin_unlock(&rq->lock);
1901 * try_to_wake_up - wake up a thread
1902 * @p: the thread to be awakened
1903 * @state: the mask of task states that can be woken
1904 * @wake_flags: wake modifier flags (WF_*)
1906 * Put it on the run-queue if it's not already there. The "current"
1907 * thread is always on the run-queue (except when the actual
1908 * re-schedule is in progress), and as such you're allowed to do
1909 * the simpler "current->state = TASK_RUNNING" to mark yourself
1910 * runnable without the overhead of this.
1912 * Return: %true if @p was woken up, %false if it was already running.
1913 * or @state didn't match @p's state.
1916 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1918 unsigned long flags;
1919 int cpu, success = 0;
1922 * If we are going to wake up a thread waiting for CONDITION we
1923 * need to ensure that CONDITION=1 done by the caller can not be
1924 * reordered with p->state check below. This pairs with mb() in
1925 * set_current_state() the waiting thread does.
1927 smp_mb__before_spinlock();
1928 raw_spin_lock_irqsave(&p->pi_lock, flags);
1929 if (!(p->state & state))
1932 trace_sched_waking(p);
1934 success = 1; /* we're going to change ->state */
1937 if (p->on_rq && ttwu_remote(p, wake_flags))
1942 * If the owning (remote) cpu is still in the middle of schedule() with
1943 * this task as prev, wait until its done referencing the task.
1948 * Pairs with the smp_wmb() in finish_lock_switch().
1952 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1953 p->state = TASK_WAKING;
1955 if (p->sched_class->task_waking)
1956 p->sched_class->task_waking(p);
1958 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1959 if (task_cpu(p) != cpu) {
1960 wake_flags |= WF_MIGRATED;
1961 set_task_cpu(p, cpu);
1963 #endif /* CONFIG_SMP */
1967 ttwu_stat(p, cpu, wake_flags);
1969 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1975 * try_to_wake_up_local - try to wake up a local task with rq lock held
1976 * @p: the thread to be awakened
1978 * Put @p on the run-queue if it's not already there. The caller must
1979 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1982 static void try_to_wake_up_local(struct task_struct *p)
1984 struct rq *rq = task_rq(p);
1986 if (WARN_ON_ONCE(rq != this_rq()) ||
1987 WARN_ON_ONCE(p == current))
1990 lockdep_assert_held(&rq->lock);
1992 if (!raw_spin_trylock(&p->pi_lock)) {
1994 * This is OK, because current is on_cpu, which avoids it being
1995 * picked for load-balance and preemption/IRQs are still
1996 * disabled avoiding further scheduler activity on it and we've
1997 * not yet picked a replacement task.
1999 lockdep_unpin_lock(&rq->lock);
2000 raw_spin_unlock(&rq->lock);
2001 raw_spin_lock(&p->pi_lock);
2002 raw_spin_lock(&rq->lock);
2003 lockdep_pin_lock(&rq->lock);
2006 if (!(p->state & TASK_NORMAL))
2009 trace_sched_waking(p);
2011 if (!task_on_rq_queued(p))
2012 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2014 ttwu_do_wakeup(rq, p, 0);
2015 ttwu_stat(p, smp_processor_id(), 0);
2017 raw_spin_unlock(&p->pi_lock);
2021 * wake_up_process - Wake up a specific process
2022 * @p: The process to be woken up.
2024 * Attempt to wake up the nominated process and move it to the set of runnable
2027 * Return: 1 if the process was woken up, 0 if it was already running.
2029 * It may be assumed that this function implies a write memory barrier before
2030 * changing the task state if and only if any tasks are woken up.
2032 int wake_up_process(struct task_struct *p)
2034 WARN_ON(task_is_stopped_or_traced(p));
2035 return try_to_wake_up(p, TASK_NORMAL, 0);
2037 EXPORT_SYMBOL(wake_up_process);
2039 int wake_up_state(struct task_struct *p, unsigned int state)
2041 return try_to_wake_up(p, state, 0);
2045 * This function clears the sched_dl_entity static params.
2047 void __dl_clear_params(struct task_struct *p)
2049 struct sched_dl_entity *dl_se = &p->dl;
2051 dl_se->dl_runtime = 0;
2052 dl_se->dl_deadline = 0;
2053 dl_se->dl_period = 0;
2057 dl_se->dl_throttled = 0;
2059 dl_se->dl_yielded = 0;
2063 * Perform scheduler related setup for a newly forked process p.
2064 * p is forked by current.
2066 * __sched_fork() is basic setup used by init_idle() too:
2068 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2073 p->se.exec_start = 0;
2074 p->se.sum_exec_runtime = 0;
2075 p->se.prev_sum_exec_runtime = 0;
2076 p->se.nr_migrations = 0;
2078 INIT_LIST_HEAD(&p->se.group_node);
2080 #ifdef CONFIG_SCHEDSTATS
2081 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2084 RB_CLEAR_NODE(&p->dl.rb_node);
2085 init_dl_task_timer(&p->dl);
2086 __dl_clear_params(p);
2088 INIT_LIST_HEAD(&p->rt.run_list);
2090 #ifdef CONFIG_PREEMPT_NOTIFIERS
2091 INIT_HLIST_HEAD(&p->preempt_notifiers);
2094 #ifdef CONFIG_NUMA_BALANCING
2095 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2096 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2097 p->mm->numa_scan_seq = 0;
2100 if (clone_flags & CLONE_VM)
2101 p->numa_preferred_nid = current->numa_preferred_nid;
2103 p->numa_preferred_nid = -1;
2105 p->node_stamp = 0ULL;
2106 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2107 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2108 p->numa_work.next = &p->numa_work;
2109 p->numa_faults = NULL;
2110 p->last_task_numa_placement = 0;
2111 p->last_sum_exec_runtime = 0;
2113 p->numa_group = NULL;
2114 #endif /* CONFIG_NUMA_BALANCING */
2117 #ifdef CONFIG_NUMA_BALANCING
2118 __read_mostly bool sched_numa_balancing;
2120 void set_numabalancing_state(bool enabled)
2122 sched_numa_balancing = enabled;
2123 #ifdef CONFIG_SCHED_DEBUG
2125 sched_feat_set("NUMA");
2127 sched_feat_set("NO_NUMA");
2128 #endif /* CONFIG_SCHED_DEBUG */
2131 #ifdef CONFIG_PROC_SYSCTL
2132 int sysctl_numa_balancing(struct ctl_table *table, int write,
2133 void __user *buffer, size_t *lenp, loff_t *ppos)
2137 int state = sched_numa_balancing;
2139 if (write && !capable(CAP_SYS_ADMIN))
2144 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2148 set_numabalancing_state(state);
2155 * fork()/clone()-time setup:
2157 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2159 unsigned long flags;
2160 int cpu = get_cpu();
2162 __sched_fork(clone_flags, p);
2164 * We mark the process as running here. This guarantees that
2165 * nobody will actually run it, and a signal or other external
2166 * event cannot wake it up and insert it on the runqueue either.
2168 p->state = TASK_RUNNING;
2171 * Make sure we do not leak PI boosting priority to the child.
2173 p->prio = current->normal_prio;
2176 * Revert to default priority/policy on fork if requested.
2178 if (unlikely(p->sched_reset_on_fork)) {
2179 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2180 p->policy = SCHED_NORMAL;
2181 p->static_prio = NICE_TO_PRIO(0);
2183 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2184 p->static_prio = NICE_TO_PRIO(0);
2186 p->prio = p->normal_prio = __normal_prio(p);
2190 * We don't need the reset flag anymore after the fork. It has
2191 * fulfilled its duty:
2193 p->sched_reset_on_fork = 0;
2196 if (dl_prio(p->prio)) {
2199 } else if (rt_prio(p->prio)) {
2200 p->sched_class = &rt_sched_class;
2202 p->sched_class = &fair_sched_class;
2205 if (p->sched_class->task_fork)
2206 p->sched_class->task_fork(p);
2209 * The child is not yet in the pid-hash so no cgroup attach races,
2210 * and the cgroup is pinned to this child due to cgroup_fork()
2211 * is ran before sched_fork().
2213 * Silence PROVE_RCU.
2215 raw_spin_lock_irqsave(&p->pi_lock, flags);
2216 set_task_cpu(p, cpu);
2217 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2219 #ifdef CONFIG_SCHED_INFO
2220 if (likely(sched_info_on()))
2221 memset(&p->sched_info, 0, sizeof(p->sched_info));
2223 #if defined(CONFIG_SMP)
2226 init_task_preempt_count(p);
2228 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2229 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2236 unsigned long to_ratio(u64 period, u64 runtime)
2238 if (runtime == RUNTIME_INF)
2242 * Doing this here saves a lot of checks in all
2243 * the calling paths, and returning zero seems
2244 * safe for them anyway.
2249 return div64_u64(runtime << 20, period);
2253 inline struct dl_bw *dl_bw_of(int i)
2255 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2256 "sched RCU must be held");
2257 return &cpu_rq(i)->rd->dl_bw;
2260 static inline int dl_bw_cpus(int i)
2262 struct root_domain *rd = cpu_rq(i)->rd;
2265 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2266 "sched RCU must be held");
2267 for_each_cpu_and(i, rd->span, cpu_active_mask)
2273 inline struct dl_bw *dl_bw_of(int i)
2275 return &cpu_rq(i)->dl.dl_bw;
2278 static inline int dl_bw_cpus(int i)
2285 * We must be sure that accepting a new task (or allowing changing the
2286 * parameters of an existing one) is consistent with the bandwidth
2287 * constraints. If yes, this function also accordingly updates the currently
2288 * allocated bandwidth to reflect the new situation.
2290 * This function is called while holding p's rq->lock.
2292 * XXX we should delay bw change until the task's 0-lag point, see
2295 static int dl_overflow(struct task_struct *p, int policy,
2296 const struct sched_attr *attr)
2299 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2300 u64 period = attr->sched_period ?: attr->sched_deadline;
2301 u64 runtime = attr->sched_runtime;
2302 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2305 if (new_bw == p->dl.dl_bw)
2309 * Either if a task, enters, leave, or stays -deadline but changes
2310 * its parameters, we may need to update accordingly the total
2311 * allocated bandwidth of the container.
2313 raw_spin_lock(&dl_b->lock);
2314 cpus = dl_bw_cpus(task_cpu(p));
2315 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2316 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2317 __dl_add(dl_b, new_bw);
2319 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2320 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2321 __dl_clear(dl_b, p->dl.dl_bw);
2322 __dl_add(dl_b, new_bw);
2324 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2325 __dl_clear(dl_b, p->dl.dl_bw);
2328 raw_spin_unlock(&dl_b->lock);
2333 extern void init_dl_bw(struct dl_bw *dl_b);
2336 * wake_up_new_task - wake up a newly created task for the first time.
2338 * This function will do some initial scheduler statistics housekeeping
2339 * that must be done for every newly created context, then puts the task
2340 * on the runqueue and wakes it.
2342 void wake_up_new_task(struct task_struct *p)
2344 unsigned long flags;
2347 raw_spin_lock_irqsave(&p->pi_lock, flags);
2350 * Fork balancing, do it here and not earlier because:
2351 * - cpus_allowed can change in the fork path
2352 * - any previously selected cpu might disappear through hotplug
2354 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2357 /* Initialize new task's runnable average */
2358 init_entity_runnable_average(&p->se);
2359 rq = __task_rq_lock(p);
2360 activate_task(rq, p, 0);
2361 p->on_rq = TASK_ON_RQ_QUEUED;
2362 trace_sched_wakeup_new(p);
2363 check_preempt_curr(rq, p, WF_FORK);
2365 if (p->sched_class->task_woken)
2366 p->sched_class->task_woken(rq, p);
2368 task_rq_unlock(rq, p, &flags);
2371 #ifdef CONFIG_PREEMPT_NOTIFIERS
2373 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2375 void preempt_notifier_inc(void)
2377 static_key_slow_inc(&preempt_notifier_key);
2379 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2381 void preempt_notifier_dec(void)
2383 static_key_slow_dec(&preempt_notifier_key);
2385 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2388 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2389 * @notifier: notifier struct to register
2391 void preempt_notifier_register(struct preempt_notifier *notifier)
2393 if (!static_key_false(&preempt_notifier_key))
2394 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2396 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2398 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2401 * preempt_notifier_unregister - no longer interested in preemption notifications
2402 * @notifier: notifier struct to unregister
2404 * This is *not* safe to call from within a preemption notifier.
2406 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2408 hlist_del(¬ifier->link);
2410 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2412 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2414 struct preempt_notifier *notifier;
2416 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2417 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2420 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2422 if (static_key_false(&preempt_notifier_key))
2423 __fire_sched_in_preempt_notifiers(curr);
2427 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2428 struct task_struct *next)
2430 struct preempt_notifier *notifier;
2432 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2433 notifier->ops->sched_out(notifier, next);
2436 static __always_inline void
2437 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2438 struct task_struct *next)
2440 if (static_key_false(&preempt_notifier_key))
2441 __fire_sched_out_preempt_notifiers(curr, next);
2444 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2446 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2451 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2452 struct task_struct *next)
2456 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2459 * prepare_task_switch - prepare to switch tasks
2460 * @rq: the runqueue preparing to switch
2461 * @prev: the current task that is being switched out
2462 * @next: the task we are going to switch to.
2464 * This is called with the rq lock held and interrupts off. It must
2465 * be paired with a subsequent finish_task_switch after the context
2468 * prepare_task_switch sets up locking and calls architecture specific
2472 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2473 struct task_struct *next)
2475 trace_sched_switch(prev, next);
2476 sched_info_switch(rq, prev, next);
2477 perf_event_task_sched_out(prev, next);
2478 fire_sched_out_preempt_notifiers(prev, next);
2479 prepare_lock_switch(rq, next);
2480 prepare_arch_switch(next);
2484 * finish_task_switch - clean up after a task-switch
2485 * @prev: the thread we just switched away from.
2487 * finish_task_switch must be called after the context switch, paired
2488 * with a prepare_task_switch call before the context switch.
2489 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2490 * and do any other architecture-specific cleanup actions.
2492 * Note that we may have delayed dropping an mm in context_switch(). If
2493 * so, we finish that here outside of the runqueue lock. (Doing it
2494 * with the lock held can cause deadlocks; see schedule() for
2497 * The context switch have flipped the stack from under us and restored the
2498 * local variables which were saved when this task called schedule() in the
2499 * past. prev == current is still correct but we need to recalculate this_rq
2500 * because prev may have moved to another CPU.
2502 static struct rq *finish_task_switch(struct task_struct *prev)
2503 __releases(rq->lock)
2505 struct rq *rq = this_rq();
2506 struct mm_struct *mm = rq->prev_mm;
2512 * A task struct has one reference for the use as "current".
2513 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2514 * schedule one last time. The schedule call will never return, and
2515 * the scheduled task must drop that reference.
2516 * The test for TASK_DEAD must occur while the runqueue locks are
2517 * still held, otherwise prev could be scheduled on another cpu, die
2518 * there before we look at prev->state, and then the reference would
2520 * Manfred Spraul <manfred@colorfullife.com>
2522 prev_state = prev->state;
2523 vtime_task_switch(prev);
2524 perf_event_task_sched_in(prev, current);
2525 finish_lock_switch(rq, prev);
2526 finish_arch_post_lock_switch();
2528 fire_sched_in_preempt_notifiers(current);
2531 if (unlikely(prev_state == TASK_DEAD)) {
2532 if (prev->sched_class->task_dead)
2533 prev->sched_class->task_dead(prev);
2536 * Remove function-return probe instances associated with this
2537 * task and put them back on the free list.
2539 kprobe_flush_task(prev);
2540 put_task_struct(prev);
2543 tick_nohz_task_switch();
2549 /* rq->lock is NOT held, but preemption is disabled */
2550 static void __balance_callback(struct rq *rq)
2552 struct callback_head *head, *next;
2553 void (*func)(struct rq *rq);
2554 unsigned long flags;
2556 raw_spin_lock_irqsave(&rq->lock, flags);
2557 head = rq->balance_callback;
2558 rq->balance_callback = NULL;
2560 func = (void (*)(struct rq *))head->func;
2567 raw_spin_unlock_irqrestore(&rq->lock, flags);
2570 static inline void balance_callback(struct rq *rq)
2572 if (unlikely(rq->balance_callback))
2573 __balance_callback(rq);
2578 static inline void balance_callback(struct rq *rq)
2585 * schedule_tail - first thing a freshly forked thread must call.
2586 * @prev: the thread we just switched away from.
2588 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2589 __releases(rq->lock)
2593 /* finish_task_switch() drops rq->lock and enables preemtion */
2595 rq = finish_task_switch(prev);
2596 balance_callback(rq);
2599 if (current->set_child_tid)
2600 put_user(task_pid_vnr(current), current->set_child_tid);
2604 * context_switch - switch to the new MM and the new thread's register state.
2606 static inline struct rq *
2607 context_switch(struct rq *rq, struct task_struct *prev,
2608 struct task_struct *next)
2610 struct mm_struct *mm, *oldmm;
2612 prepare_task_switch(rq, prev, next);
2615 oldmm = prev->active_mm;
2617 * For paravirt, this is coupled with an exit in switch_to to
2618 * combine the page table reload and the switch backend into
2621 arch_start_context_switch(prev);
2624 next->active_mm = oldmm;
2625 atomic_inc(&oldmm->mm_count);
2626 enter_lazy_tlb(oldmm, next);
2628 switch_mm(oldmm, mm, next);
2631 prev->active_mm = NULL;
2632 rq->prev_mm = oldmm;
2635 * Since the runqueue lock will be released by the next
2636 * task (which is an invalid locking op but in the case
2637 * of the scheduler it's an obvious special-case), so we
2638 * do an early lockdep release here:
2640 lockdep_unpin_lock(&rq->lock);
2641 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2643 /* Here we just switch the register state and the stack. */
2644 switch_to(prev, next, prev);
2647 return finish_task_switch(prev);
2651 * nr_running and nr_context_switches:
2653 * externally visible scheduler statistics: current number of runnable
2654 * threads, total number of context switches performed since bootup.
2656 unsigned long nr_running(void)
2658 unsigned long i, sum = 0;
2660 for_each_online_cpu(i)
2661 sum += cpu_rq(i)->nr_running;
2667 * Check if only the current task is running on the cpu.
2669 bool single_task_running(void)
2671 if (cpu_rq(smp_processor_id())->nr_running == 1)
2676 EXPORT_SYMBOL(single_task_running);
2678 unsigned long long nr_context_switches(void)
2681 unsigned long long sum = 0;
2683 for_each_possible_cpu(i)
2684 sum += cpu_rq(i)->nr_switches;
2689 unsigned long nr_iowait(void)
2691 unsigned long i, sum = 0;
2693 for_each_possible_cpu(i)
2694 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2699 unsigned long nr_iowait_cpu(int cpu)
2701 struct rq *this = cpu_rq(cpu);
2702 return atomic_read(&this->nr_iowait);
2705 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2707 struct rq *rq = this_rq();
2708 *nr_waiters = atomic_read(&rq->nr_iowait);
2709 *load = rq->load.weight;
2715 * sched_exec - execve() is a valuable balancing opportunity, because at
2716 * this point the task has the smallest effective memory and cache footprint.
2718 void sched_exec(void)
2720 struct task_struct *p = current;
2721 unsigned long flags;
2724 raw_spin_lock_irqsave(&p->pi_lock, flags);
2725 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2726 if (dest_cpu == smp_processor_id())
2729 if (likely(cpu_active(dest_cpu))) {
2730 struct migration_arg arg = { p, dest_cpu };
2732 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2733 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2737 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2742 DEFINE_PER_CPU(struct kernel_stat, kstat);
2743 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2745 EXPORT_PER_CPU_SYMBOL(kstat);
2746 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2749 * Return accounted runtime for the task.
2750 * In case the task is currently running, return the runtime plus current's
2751 * pending runtime that have not been accounted yet.
2753 unsigned long long task_sched_runtime(struct task_struct *p)
2755 unsigned long flags;
2759 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2761 * 64-bit doesn't need locks to atomically read a 64bit value.
2762 * So we have a optimization chance when the task's delta_exec is 0.
2763 * Reading ->on_cpu is racy, but this is ok.
2765 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2766 * If we race with it entering cpu, unaccounted time is 0. This is
2767 * indistinguishable from the read occurring a few cycles earlier.
2768 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2769 * been accounted, so we're correct here as well.
2771 if (!p->on_cpu || !task_on_rq_queued(p))
2772 return p->se.sum_exec_runtime;
2775 rq = task_rq_lock(p, &flags);
2777 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2778 * project cycles that may never be accounted to this
2779 * thread, breaking clock_gettime().
2781 if (task_current(rq, p) && task_on_rq_queued(p)) {
2782 update_rq_clock(rq);
2783 p->sched_class->update_curr(rq);
2785 ns = p->se.sum_exec_runtime;
2786 task_rq_unlock(rq, p, &flags);
2792 * This function gets called by the timer code, with HZ frequency.
2793 * We call it with interrupts disabled.
2795 void scheduler_tick(void)
2797 int cpu = smp_processor_id();
2798 struct rq *rq = cpu_rq(cpu);
2799 struct task_struct *curr = rq->curr;
2803 raw_spin_lock(&rq->lock);
2804 update_rq_clock(rq);
2805 curr->sched_class->task_tick(rq, curr, 0);
2806 update_cpu_load_active(rq);
2807 calc_global_load_tick(rq);
2808 raw_spin_unlock(&rq->lock);
2810 perf_event_task_tick();
2813 rq->idle_balance = idle_cpu(cpu);
2814 trigger_load_balance(rq);
2816 rq_last_tick_reset(rq);
2819 #ifdef CONFIG_NO_HZ_FULL
2821 * scheduler_tick_max_deferment
2823 * Keep at least one tick per second when a single
2824 * active task is running because the scheduler doesn't
2825 * yet completely support full dynticks environment.
2827 * This makes sure that uptime, CFS vruntime, load
2828 * balancing, etc... continue to move forward, even
2829 * with a very low granularity.
2831 * Return: Maximum deferment in nanoseconds.
2833 u64 scheduler_tick_max_deferment(void)
2835 struct rq *rq = this_rq();
2836 unsigned long next, now = READ_ONCE(jiffies);
2838 next = rq->last_sched_tick + HZ;
2840 if (time_before_eq(next, now))
2843 return jiffies_to_nsecs(next - now);
2847 notrace unsigned long get_parent_ip(unsigned long addr)
2849 if (in_lock_functions(addr)) {
2850 addr = CALLER_ADDR2;
2851 if (in_lock_functions(addr))
2852 addr = CALLER_ADDR3;
2857 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2858 defined(CONFIG_PREEMPT_TRACER))
2860 void preempt_count_add(int val)
2862 #ifdef CONFIG_DEBUG_PREEMPT
2866 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2869 __preempt_count_add(val);
2870 #ifdef CONFIG_DEBUG_PREEMPT
2872 * Spinlock count overflowing soon?
2874 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2877 if (preempt_count() == val) {
2878 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2879 #ifdef CONFIG_DEBUG_PREEMPT
2880 current->preempt_disable_ip = ip;
2882 trace_preempt_off(CALLER_ADDR0, ip);
2885 EXPORT_SYMBOL(preempt_count_add);
2886 NOKPROBE_SYMBOL(preempt_count_add);
2888 void preempt_count_sub(int val)
2890 #ifdef CONFIG_DEBUG_PREEMPT
2894 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2897 * Is the spinlock portion underflowing?
2899 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2900 !(preempt_count() & PREEMPT_MASK)))
2904 if (preempt_count() == val)
2905 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2906 __preempt_count_sub(val);
2908 EXPORT_SYMBOL(preempt_count_sub);
2909 NOKPROBE_SYMBOL(preempt_count_sub);
2914 * Print scheduling while atomic bug:
2916 static noinline void __schedule_bug(struct task_struct *prev)
2918 if (oops_in_progress)
2921 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2922 prev->comm, prev->pid, preempt_count());
2924 debug_show_held_locks(prev);
2926 if (irqs_disabled())
2927 print_irqtrace_events(prev);
2928 #ifdef CONFIG_DEBUG_PREEMPT
2929 if (in_atomic_preempt_off()) {
2930 pr_err("Preemption disabled at:");
2931 print_ip_sym(current->preempt_disable_ip);
2936 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2940 * Various schedule()-time debugging checks and statistics:
2942 static inline void schedule_debug(struct task_struct *prev)
2944 #ifdef CONFIG_SCHED_STACK_END_CHECK
2945 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2948 * Test if we are atomic. Since do_exit() needs to call into
2949 * schedule() atomically, we ignore that path. Otherwise whine
2950 * if we are scheduling when we should not.
2952 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2953 __schedule_bug(prev);
2956 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2958 schedstat_inc(this_rq(), sched_count);
2962 * Pick up the highest-prio task:
2964 static inline struct task_struct *
2965 pick_next_task(struct rq *rq, struct task_struct *prev)
2967 const struct sched_class *class = &fair_sched_class;
2968 struct task_struct *p;
2971 * Optimization: we know that if all tasks are in
2972 * the fair class we can call that function directly:
2974 if (likely(prev->sched_class == class &&
2975 rq->nr_running == rq->cfs.h_nr_running)) {
2976 p = fair_sched_class.pick_next_task(rq, prev);
2977 if (unlikely(p == RETRY_TASK))
2980 /* assumes fair_sched_class->next == idle_sched_class */
2982 p = idle_sched_class.pick_next_task(rq, prev);
2988 for_each_class(class) {
2989 p = class->pick_next_task(rq, prev);
2991 if (unlikely(p == RETRY_TASK))
2997 BUG(); /* the idle class will always have a runnable task */
3001 * __schedule() is the main scheduler function.
3003 * The main means of driving the scheduler and thus entering this function are:
3005 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3007 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3008 * paths. For example, see arch/x86/entry_64.S.
3010 * To drive preemption between tasks, the scheduler sets the flag in timer
3011 * interrupt handler scheduler_tick().
3013 * 3. Wakeups don't really cause entry into schedule(). They add a
3014 * task to the run-queue and that's it.
3016 * Now, if the new task added to the run-queue preempts the current
3017 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3018 * called on the nearest possible occasion:
3020 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3022 * - in syscall or exception context, at the next outmost
3023 * preempt_enable(). (this might be as soon as the wake_up()'s
3026 * - in IRQ context, return from interrupt-handler to
3027 * preemptible context
3029 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3032 * - cond_resched() call
3033 * - explicit schedule() call
3034 * - return from syscall or exception to user-space
3035 * - return from interrupt-handler to user-space
3037 * WARNING: must be called with preemption disabled!
3039 static void __sched __schedule(void)
3041 struct task_struct *prev, *next;
3042 unsigned long *switch_count;
3046 cpu = smp_processor_id();
3048 rcu_note_context_switch();
3051 schedule_debug(prev);
3053 if (sched_feat(HRTICK))
3057 * Make sure that signal_pending_state()->signal_pending() below
3058 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3059 * done by the caller to avoid the race with signal_wake_up().
3061 smp_mb__before_spinlock();
3062 raw_spin_lock_irq(&rq->lock);
3063 lockdep_pin_lock(&rq->lock);
3065 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3067 switch_count = &prev->nivcsw;
3068 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3069 if (unlikely(signal_pending_state(prev->state, prev))) {
3070 prev->state = TASK_RUNNING;
3072 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3076 * If a worker went to sleep, notify and ask workqueue
3077 * whether it wants to wake up a task to maintain
3080 if (prev->flags & PF_WQ_WORKER) {
3081 struct task_struct *to_wakeup;
3083 to_wakeup = wq_worker_sleeping(prev, cpu);
3085 try_to_wake_up_local(to_wakeup);
3088 switch_count = &prev->nvcsw;
3091 if (task_on_rq_queued(prev))
3092 update_rq_clock(rq);
3094 next = pick_next_task(rq, prev);
3095 clear_tsk_need_resched(prev);
3096 clear_preempt_need_resched();
3097 rq->clock_skip_update = 0;
3099 if (likely(prev != next)) {
3104 rq = context_switch(rq, prev, next); /* unlocks the rq */
3107 lockdep_unpin_lock(&rq->lock);
3108 raw_spin_unlock_irq(&rq->lock);
3111 balance_callback(rq);
3114 static inline void sched_submit_work(struct task_struct *tsk)
3116 if (!tsk->state || tsk_is_pi_blocked(tsk))
3119 * If we are going to sleep and we have plugged IO queued,
3120 * make sure to submit it to avoid deadlocks.
3122 if (blk_needs_flush_plug(tsk))
3123 blk_schedule_flush_plug(tsk);
3126 asmlinkage __visible void __sched schedule(void)
3128 struct task_struct *tsk = current;
3130 sched_submit_work(tsk);
3134 sched_preempt_enable_no_resched();
3135 } while (need_resched());
3137 EXPORT_SYMBOL(schedule);
3139 #ifdef CONFIG_CONTEXT_TRACKING
3140 asmlinkage __visible void __sched schedule_user(void)
3143 * If we come here after a random call to set_need_resched(),
3144 * or we have been woken up remotely but the IPI has not yet arrived,
3145 * we haven't yet exited the RCU idle mode. Do it here manually until
3146 * we find a better solution.
3148 * NB: There are buggy callers of this function. Ideally we
3149 * should warn if prev_state != CONTEXT_USER, but that will trigger
3150 * too frequently to make sense yet.
3152 enum ctx_state prev_state = exception_enter();
3154 exception_exit(prev_state);
3159 * schedule_preempt_disabled - called with preemption disabled
3161 * Returns with preemption disabled. Note: preempt_count must be 1
3163 void __sched schedule_preempt_disabled(void)
3165 sched_preempt_enable_no_resched();
3170 static void __sched notrace preempt_schedule_common(void)
3173 preempt_active_enter();
3175 preempt_active_exit();
3178 * Check again in case we missed a preemption opportunity
3179 * between schedule and now.
3181 } while (need_resched());
3184 #ifdef CONFIG_PREEMPT
3186 * this is the entry point to schedule() from in-kernel preemption
3187 * off of preempt_enable. Kernel preemptions off return from interrupt
3188 * occur there and call schedule directly.
3190 asmlinkage __visible void __sched notrace preempt_schedule(void)
3193 * If there is a non-zero preempt_count or interrupts are disabled,
3194 * we do not want to preempt the current task. Just return..
3196 if (likely(!preemptible()))
3199 preempt_schedule_common();
3201 NOKPROBE_SYMBOL(preempt_schedule);
3202 EXPORT_SYMBOL(preempt_schedule);
3205 * preempt_schedule_notrace - preempt_schedule called by tracing
3207 * The tracing infrastructure uses preempt_enable_notrace to prevent
3208 * recursion and tracing preempt enabling caused by the tracing
3209 * infrastructure itself. But as tracing can happen in areas coming
3210 * from userspace or just about to enter userspace, a preempt enable
3211 * can occur before user_exit() is called. This will cause the scheduler
3212 * to be called when the system is still in usermode.
3214 * To prevent this, the preempt_enable_notrace will use this function
3215 * instead of preempt_schedule() to exit user context if needed before
3216 * calling the scheduler.
3218 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3220 enum ctx_state prev_ctx;
3222 if (likely(!preemptible()))
3227 * Use raw __prempt_count() ops that don't call function.
3228 * We can't call functions before disabling preemption which
3229 * disarm preemption tracing recursions.
3231 __preempt_count_add(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
3234 * Needs preempt disabled in case user_exit() is traced
3235 * and the tracer calls preempt_enable_notrace() causing
3236 * an infinite recursion.
3238 prev_ctx = exception_enter();
3240 exception_exit(prev_ctx);
3243 __preempt_count_sub(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
3244 } while (need_resched());
3246 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3248 #endif /* CONFIG_PREEMPT */
3251 * this is the entry point to schedule() from kernel preemption
3252 * off of irq context.
3253 * Note, that this is called and return with irqs disabled. This will
3254 * protect us against recursive calling from irq.
3256 asmlinkage __visible void __sched preempt_schedule_irq(void)
3258 enum ctx_state prev_state;
3260 /* Catch callers which need to be fixed */
3261 BUG_ON(preempt_count() || !irqs_disabled());
3263 prev_state = exception_enter();
3266 preempt_active_enter();
3269 local_irq_disable();
3270 preempt_active_exit();
3271 } while (need_resched());
3273 exception_exit(prev_state);
3276 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3279 return try_to_wake_up(curr->private, mode, wake_flags);
3281 EXPORT_SYMBOL(default_wake_function);
3283 #ifdef CONFIG_RT_MUTEXES
3286 * rt_mutex_setprio - set the current priority of a task
3288 * @prio: prio value (kernel-internal form)
3290 * This function changes the 'effective' priority of a task. It does
3291 * not touch ->normal_prio like __setscheduler().
3293 * Used by the rt_mutex code to implement priority inheritance
3294 * logic. Call site only calls if the priority of the task changed.
3296 void rt_mutex_setprio(struct task_struct *p, int prio)
3298 int oldprio, queued, running, enqueue_flag = 0;
3300 const struct sched_class *prev_class;
3302 BUG_ON(prio > MAX_PRIO);
3304 rq = __task_rq_lock(p);
3307 * Idle task boosting is a nono in general. There is one
3308 * exception, when PREEMPT_RT and NOHZ is active:
3310 * The idle task calls get_next_timer_interrupt() and holds
3311 * the timer wheel base->lock on the CPU and another CPU wants
3312 * to access the timer (probably to cancel it). We can safely
3313 * ignore the boosting request, as the idle CPU runs this code
3314 * with interrupts disabled and will complete the lock
3315 * protected section without being interrupted. So there is no
3316 * real need to boost.
3318 if (unlikely(p == rq->idle)) {
3319 WARN_ON(p != rq->curr);
3320 WARN_ON(p->pi_blocked_on);
3324 trace_sched_pi_setprio(p, prio);
3326 prev_class = p->sched_class;
3327 queued = task_on_rq_queued(p);
3328 running = task_current(rq, p);
3330 dequeue_task(rq, p, 0);
3332 put_prev_task(rq, p);
3335 * Boosting condition are:
3336 * 1. -rt task is running and holds mutex A
3337 * --> -dl task blocks on mutex A
3339 * 2. -dl task is running and holds mutex A
3340 * --> -dl task blocks on mutex A and could preempt the
3343 if (dl_prio(prio)) {
3344 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3345 if (!dl_prio(p->normal_prio) ||
3346 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3347 p->dl.dl_boosted = 1;
3348 enqueue_flag = ENQUEUE_REPLENISH;
3350 p->dl.dl_boosted = 0;
3351 p->sched_class = &dl_sched_class;
3352 } else if (rt_prio(prio)) {
3353 if (dl_prio(oldprio))
3354 p->dl.dl_boosted = 0;
3356 enqueue_flag = ENQUEUE_HEAD;
3357 p->sched_class = &rt_sched_class;
3359 if (dl_prio(oldprio))
3360 p->dl.dl_boosted = 0;
3361 if (rt_prio(oldprio))
3363 p->sched_class = &fair_sched_class;
3369 p->sched_class->set_curr_task(rq);
3371 enqueue_task(rq, p, enqueue_flag);
3373 check_class_changed(rq, p, prev_class, oldprio);
3375 preempt_disable(); /* avoid rq from going away on us */
3376 __task_rq_unlock(rq);
3378 balance_callback(rq);
3383 void set_user_nice(struct task_struct *p, long nice)
3385 int old_prio, delta, queued;
3386 unsigned long flags;
3389 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3392 * We have to be careful, if called from sys_setpriority(),
3393 * the task might be in the middle of scheduling on another CPU.
3395 rq = task_rq_lock(p, &flags);
3397 * The RT priorities are set via sched_setscheduler(), but we still
3398 * allow the 'normal' nice value to be set - but as expected
3399 * it wont have any effect on scheduling until the task is
3400 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3402 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3403 p->static_prio = NICE_TO_PRIO(nice);
3406 queued = task_on_rq_queued(p);
3408 dequeue_task(rq, p, 0);
3410 p->static_prio = NICE_TO_PRIO(nice);
3413 p->prio = effective_prio(p);
3414 delta = p->prio - old_prio;
3417 enqueue_task(rq, p, 0);
3419 * If the task increased its priority or is running and
3420 * lowered its priority, then reschedule its CPU:
3422 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3426 task_rq_unlock(rq, p, &flags);
3428 EXPORT_SYMBOL(set_user_nice);
3431 * can_nice - check if a task can reduce its nice value
3435 int can_nice(const struct task_struct *p, const int nice)
3437 /* convert nice value [19,-20] to rlimit style value [1,40] */
3438 int nice_rlim = nice_to_rlimit(nice);
3440 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3441 capable(CAP_SYS_NICE));
3444 #ifdef __ARCH_WANT_SYS_NICE
3447 * sys_nice - change the priority of the current process.
3448 * @increment: priority increment
3450 * sys_setpriority is a more generic, but much slower function that
3451 * does similar things.
3453 SYSCALL_DEFINE1(nice, int, increment)
3458 * Setpriority might change our priority at the same moment.
3459 * We don't have to worry. Conceptually one call occurs first
3460 * and we have a single winner.
3462 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3463 nice = task_nice(current) + increment;
3465 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3466 if (increment < 0 && !can_nice(current, nice))
3469 retval = security_task_setnice(current, nice);
3473 set_user_nice(current, nice);
3480 * task_prio - return the priority value of a given task.
3481 * @p: the task in question.
3483 * Return: The priority value as seen by users in /proc.
3484 * RT tasks are offset by -200. Normal tasks are centered
3485 * around 0, value goes from -16 to +15.
3487 int task_prio(const struct task_struct *p)
3489 return p->prio - MAX_RT_PRIO;
3493 * idle_cpu - is a given cpu idle currently?
3494 * @cpu: the processor in question.
3496 * Return: 1 if the CPU is currently idle. 0 otherwise.
3498 int idle_cpu(int cpu)
3500 struct rq *rq = cpu_rq(cpu);
3502 if (rq->curr != rq->idle)
3509 if (!llist_empty(&rq->wake_list))
3517 * idle_task - return the idle task for a given cpu.
3518 * @cpu: the processor in question.
3520 * Return: The idle task for the cpu @cpu.
3522 struct task_struct *idle_task(int cpu)
3524 return cpu_rq(cpu)->idle;
3528 * find_process_by_pid - find a process with a matching PID value.
3529 * @pid: the pid in question.
3531 * The task of @pid, if found. %NULL otherwise.
3533 static struct task_struct *find_process_by_pid(pid_t pid)
3535 return pid ? find_task_by_vpid(pid) : current;
3539 * This function initializes the sched_dl_entity of a newly becoming
3540 * SCHED_DEADLINE task.
3542 * Only the static values are considered here, the actual runtime and the
3543 * absolute deadline will be properly calculated when the task is enqueued
3544 * for the first time with its new policy.
3547 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3549 struct sched_dl_entity *dl_se = &p->dl;
3551 dl_se->dl_runtime = attr->sched_runtime;
3552 dl_se->dl_deadline = attr->sched_deadline;
3553 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3554 dl_se->flags = attr->sched_flags;
3555 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3558 * Changing the parameters of a task is 'tricky' and we're not doing
3559 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3561 * What we SHOULD do is delay the bandwidth release until the 0-lag
3562 * point. This would include retaining the task_struct until that time
3563 * and change dl_overflow() to not immediately decrement the current
3566 * Instead we retain the current runtime/deadline and let the new
3567 * parameters take effect after the current reservation period lapses.
3568 * This is safe (albeit pessimistic) because the 0-lag point is always
3569 * before the current scheduling deadline.
3571 * We can still have temporary overloads because we do not delay the
3572 * change in bandwidth until that time; so admission control is
3573 * not on the safe side. It does however guarantee tasks will never
3574 * consume more than promised.
3579 * sched_setparam() passes in -1 for its policy, to let the functions
3580 * it calls know not to change it.
3582 #define SETPARAM_POLICY -1
3584 static void __setscheduler_params(struct task_struct *p,
3585 const struct sched_attr *attr)
3587 int policy = attr->sched_policy;
3589 if (policy == SETPARAM_POLICY)
3594 if (dl_policy(policy))
3595 __setparam_dl(p, attr);
3596 else if (fair_policy(policy))
3597 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3600 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3601 * !rt_policy. Always setting this ensures that things like
3602 * getparam()/getattr() don't report silly values for !rt tasks.
3604 p->rt_priority = attr->sched_priority;
3605 p->normal_prio = normal_prio(p);
3609 /* Actually do priority change: must hold pi & rq lock. */
3610 static void __setscheduler(struct rq *rq, struct task_struct *p,
3611 const struct sched_attr *attr, bool keep_boost)
3613 __setscheduler_params(p, attr);
3616 * Keep a potential priority boosting if called from
3617 * sched_setscheduler().
3620 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3622 p->prio = normal_prio(p);
3624 if (dl_prio(p->prio))
3625 p->sched_class = &dl_sched_class;
3626 else if (rt_prio(p->prio))
3627 p->sched_class = &rt_sched_class;
3629 p->sched_class = &fair_sched_class;
3633 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3635 struct sched_dl_entity *dl_se = &p->dl;
3637 attr->sched_priority = p->rt_priority;
3638 attr->sched_runtime = dl_se->dl_runtime;
3639 attr->sched_deadline = dl_se->dl_deadline;
3640 attr->sched_period = dl_se->dl_period;
3641 attr->sched_flags = dl_se->flags;
3645 * This function validates the new parameters of a -deadline task.
3646 * We ask for the deadline not being zero, and greater or equal
3647 * than the runtime, as well as the period of being zero or
3648 * greater than deadline. Furthermore, we have to be sure that
3649 * user parameters are above the internal resolution of 1us (we
3650 * check sched_runtime only since it is always the smaller one) and
3651 * below 2^63 ns (we have to check both sched_deadline and
3652 * sched_period, as the latter can be zero).
3655 __checkparam_dl(const struct sched_attr *attr)
3658 if (attr->sched_deadline == 0)
3662 * Since we truncate DL_SCALE bits, make sure we're at least
3665 if (attr->sched_runtime < (1ULL << DL_SCALE))
3669 * Since we use the MSB for wrap-around and sign issues, make
3670 * sure it's not set (mind that period can be equal to zero).
3672 if (attr->sched_deadline & (1ULL << 63) ||
3673 attr->sched_period & (1ULL << 63))
3676 /* runtime <= deadline <= period (if period != 0) */
3677 if ((attr->sched_period != 0 &&
3678 attr->sched_period < attr->sched_deadline) ||
3679 attr->sched_deadline < attr->sched_runtime)
3686 * check the target process has a UID that matches the current process's
3688 static bool check_same_owner(struct task_struct *p)
3690 const struct cred *cred = current_cred(), *pcred;
3694 pcred = __task_cred(p);
3695 match = (uid_eq(cred->euid, pcred->euid) ||
3696 uid_eq(cred->euid, pcred->uid));
3701 static bool dl_param_changed(struct task_struct *p,
3702 const struct sched_attr *attr)
3704 struct sched_dl_entity *dl_se = &p->dl;
3706 if (dl_se->dl_runtime != attr->sched_runtime ||
3707 dl_se->dl_deadline != attr->sched_deadline ||
3708 dl_se->dl_period != attr->sched_period ||
3709 dl_se->flags != attr->sched_flags)
3715 static int __sched_setscheduler(struct task_struct *p,
3716 const struct sched_attr *attr,
3719 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3720 MAX_RT_PRIO - 1 - attr->sched_priority;
3721 int retval, oldprio, oldpolicy = -1, queued, running;
3722 int new_effective_prio, policy = attr->sched_policy;
3723 unsigned long flags;
3724 const struct sched_class *prev_class;
3728 /* may grab non-irq protected spin_locks */
3729 BUG_ON(in_interrupt());
3731 /* double check policy once rq lock held */
3733 reset_on_fork = p->sched_reset_on_fork;
3734 policy = oldpolicy = p->policy;
3736 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3738 if (policy != SCHED_DEADLINE &&
3739 policy != SCHED_FIFO && policy != SCHED_RR &&
3740 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3741 policy != SCHED_IDLE)
3745 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3749 * Valid priorities for SCHED_FIFO and SCHED_RR are
3750 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3751 * SCHED_BATCH and SCHED_IDLE is 0.
3753 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3754 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3756 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3757 (rt_policy(policy) != (attr->sched_priority != 0)))
3761 * Allow unprivileged RT tasks to decrease priority:
3763 if (user && !capable(CAP_SYS_NICE)) {
3764 if (fair_policy(policy)) {
3765 if (attr->sched_nice < task_nice(p) &&
3766 !can_nice(p, attr->sched_nice))
3770 if (rt_policy(policy)) {
3771 unsigned long rlim_rtprio =
3772 task_rlimit(p, RLIMIT_RTPRIO);
3774 /* can't set/change the rt policy */
3775 if (policy != p->policy && !rlim_rtprio)
3778 /* can't increase priority */
3779 if (attr->sched_priority > p->rt_priority &&
3780 attr->sched_priority > rlim_rtprio)
3785 * Can't set/change SCHED_DEADLINE policy at all for now
3786 * (safest behavior); in the future we would like to allow
3787 * unprivileged DL tasks to increase their relative deadline
3788 * or reduce their runtime (both ways reducing utilization)
3790 if (dl_policy(policy))
3794 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3795 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3797 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3798 if (!can_nice(p, task_nice(p)))
3802 /* can't change other user's priorities */
3803 if (!check_same_owner(p))
3806 /* Normal users shall not reset the sched_reset_on_fork flag */
3807 if (p->sched_reset_on_fork && !reset_on_fork)
3812 retval = security_task_setscheduler(p);
3818 * make sure no PI-waiters arrive (or leave) while we are
3819 * changing the priority of the task:
3821 * To be able to change p->policy safely, the appropriate
3822 * runqueue lock must be held.
3824 rq = task_rq_lock(p, &flags);
3827 * Changing the policy of the stop threads its a very bad idea
3829 if (p == rq->stop) {
3830 task_rq_unlock(rq, p, &flags);
3835 * If not changing anything there's no need to proceed further,
3836 * but store a possible modification of reset_on_fork.
3838 if (unlikely(policy == p->policy)) {
3839 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3841 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3843 if (dl_policy(policy) && dl_param_changed(p, attr))
3846 p->sched_reset_on_fork = reset_on_fork;
3847 task_rq_unlock(rq, p, &flags);
3853 #ifdef CONFIG_RT_GROUP_SCHED
3855 * Do not allow realtime tasks into groups that have no runtime
3858 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3859 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3860 !task_group_is_autogroup(task_group(p))) {
3861 task_rq_unlock(rq, p, &flags);
3866 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3867 cpumask_t *span = rq->rd->span;
3870 * Don't allow tasks with an affinity mask smaller than
3871 * the entire root_domain to become SCHED_DEADLINE. We
3872 * will also fail if there's no bandwidth available.
3874 if (!cpumask_subset(span, &p->cpus_allowed) ||
3875 rq->rd->dl_bw.bw == 0) {
3876 task_rq_unlock(rq, p, &flags);
3883 /* recheck policy now with rq lock held */
3884 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3885 policy = oldpolicy = -1;
3886 task_rq_unlock(rq, p, &flags);
3891 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3892 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3895 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3896 task_rq_unlock(rq, p, &flags);
3900 p->sched_reset_on_fork = reset_on_fork;
3905 * Take priority boosted tasks into account. If the new
3906 * effective priority is unchanged, we just store the new
3907 * normal parameters and do not touch the scheduler class and
3908 * the runqueue. This will be done when the task deboost
3911 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3912 if (new_effective_prio == oldprio) {
3913 __setscheduler_params(p, attr);
3914 task_rq_unlock(rq, p, &flags);
3919 queued = task_on_rq_queued(p);
3920 running = task_current(rq, p);
3922 dequeue_task(rq, p, 0);
3924 put_prev_task(rq, p);
3926 prev_class = p->sched_class;
3927 __setscheduler(rq, p, attr, pi);
3930 p->sched_class->set_curr_task(rq);
3933 * We enqueue to tail when the priority of a task is
3934 * increased (user space view).
3936 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3939 check_class_changed(rq, p, prev_class, oldprio);
3940 preempt_disable(); /* avoid rq from going away on us */
3941 task_rq_unlock(rq, p, &flags);
3944 rt_mutex_adjust_pi(p);
3947 * Run balance callbacks after we've adjusted the PI chain.
3949 balance_callback(rq);
3955 static int _sched_setscheduler(struct task_struct *p, int policy,
3956 const struct sched_param *param, bool check)
3958 struct sched_attr attr = {
3959 .sched_policy = policy,
3960 .sched_priority = param->sched_priority,
3961 .sched_nice = PRIO_TO_NICE(p->static_prio),
3964 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3965 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3966 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3967 policy &= ~SCHED_RESET_ON_FORK;
3968 attr.sched_policy = policy;
3971 return __sched_setscheduler(p, &attr, check, true);
3974 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3975 * @p: the task in question.
3976 * @policy: new policy.
3977 * @param: structure containing the new RT priority.
3979 * Return: 0 on success. An error code otherwise.
3981 * NOTE that the task may be already dead.
3983 int sched_setscheduler(struct task_struct *p, int policy,
3984 const struct sched_param *param)
3986 return _sched_setscheduler(p, policy, param, true);
3988 EXPORT_SYMBOL_GPL(sched_setscheduler);
3990 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3992 return __sched_setscheduler(p, attr, true, true);
3994 EXPORT_SYMBOL_GPL(sched_setattr);
3997 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3998 * @p: the task in question.
3999 * @policy: new policy.
4000 * @param: structure containing the new RT priority.
4002 * Just like sched_setscheduler, only don't bother checking if the
4003 * current context has permission. For example, this is needed in
4004 * stop_machine(): we create temporary high priority worker threads,
4005 * but our caller might not have that capability.
4007 * Return: 0 on success. An error code otherwise.
4009 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4010 const struct sched_param *param)
4012 return _sched_setscheduler(p, policy, param, false);
4016 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4018 struct sched_param lparam;
4019 struct task_struct *p;
4022 if (!param || pid < 0)
4024 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4029 p = find_process_by_pid(pid);
4031 retval = sched_setscheduler(p, policy, &lparam);
4038 * Mimics kernel/events/core.c perf_copy_attr().
4040 static int sched_copy_attr(struct sched_attr __user *uattr,
4041 struct sched_attr *attr)
4046 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4050 * zero the full structure, so that a short copy will be nice.
4052 memset(attr, 0, sizeof(*attr));
4054 ret = get_user(size, &uattr->size);
4058 if (size > PAGE_SIZE) /* silly large */
4061 if (!size) /* abi compat */
4062 size = SCHED_ATTR_SIZE_VER0;
4064 if (size < SCHED_ATTR_SIZE_VER0)
4068 * If we're handed a bigger struct than we know of,
4069 * ensure all the unknown bits are 0 - i.e. new
4070 * user-space does not rely on any kernel feature
4071 * extensions we dont know about yet.
4073 if (size > sizeof(*attr)) {
4074 unsigned char __user *addr;
4075 unsigned char __user *end;
4078 addr = (void __user *)uattr + sizeof(*attr);
4079 end = (void __user *)uattr + size;
4081 for (; addr < end; addr++) {
4082 ret = get_user(val, addr);
4088 size = sizeof(*attr);
4091 ret = copy_from_user(attr, uattr, size);
4096 * XXX: do we want to be lenient like existing syscalls; or do we want
4097 * to be strict and return an error on out-of-bounds values?
4099 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4104 put_user(sizeof(*attr), &uattr->size);
4109 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4110 * @pid: the pid in question.
4111 * @policy: new policy.
4112 * @param: structure containing the new RT priority.
4114 * Return: 0 on success. An error code otherwise.
4116 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4117 struct sched_param __user *, param)
4119 /* negative values for policy are not valid */
4123 return do_sched_setscheduler(pid, policy, param);
4127 * sys_sched_setparam - set/change the RT priority of a thread
4128 * @pid: the pid in question.
4129 * @param: structure containing the new RT priority.
4131 * Return: 0 on success. An error code otherwise.
4133 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4135 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4139 * sys_sched_setattr - same as above, but with extended sched_attr
4140 * @pid: the pid in question.
4141 * @uattr: structure containing the extended parameters.
4142 * @flags: for future extension.
4144 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4145 unsigned int, flags)
4147 struct sched_attr attr;
4148 struct task_struct *p;
4151 if (!uattr || pid < 0 || flags)
4154 retval = sched_copy_attr(uattr, &attr);
4158 if ((int)attr.sched_policy < 0)
4163 p = find_process_by_pid(pid);
4165 retval = sched_setattr(p, &attr);
4172 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4173 * @pid: the pid in question.
4175 * Return: On success, the policy of the thread. Otherwise, a negative error
4178 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4180 struct task_struct *p;
4188 p = find_process_by_pid(pid);
4190 retval = security_task_getscheduler(p);
4193 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4200 * sys_sched_getparam - get the RT priority of a thread
4201 * @pid: the pid in question.
4202 * @param: structure containing the RT priority.
4204 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4207 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4209 struct sched_param lp = { .sched_priority = 0 };
4210 struct task_struct *p;
4213 if (!param || pid < 0)
4217 p = find_process_by_pid(pid);
4222 retval = security_task_getscheduler(p);
4226 if (task_has_rt_policy(p))
4227 lp.sched_priority = p->rt_priority;
4231 * This one might sleep, we cannot do it with a spinlock held ...
4233 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4242 static int sched_read_attr(struct sched_attr __user *uattr,
4243 struct sched_attr *attr,
4248 if (!access_ok(VERIFY_WRITE, uattr, usize))
4252 * If we're handed a smaller struct than we know of,
4253 * ensure all the unknown bits are 0 - i.e. old
4254 * user-space does not get uncomplete information.
4256 if (usize < sizeof(*attr)) {
4257 unsigned char *addr;
4260 addr = (void *)attr + usize;
4261 end = (void *)attr + sizeof(*attr);
4263 for (; addr < end; addr++) {
4271 ret = copy_to_user(uattr, attr, attr->size);
4279 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4280 * @pid: the pid in question.
4281 * @uattr: structure containing the extended parameters.
4282 * @size: sizeof(attr) for fwd/bwd comp.
4283 * @flags: for future extension.
4285 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4286 unsigned int, size, unsigned int, flags)
4288 struct sched_attr attr = {
4289 .size = sizeof(struct sched_attr),
4291 struct task_struct *p;
4294 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4295 size < SCHED_ATTR_SIZE_VER0 || flags)
4299 p = find_process_by_pid(pid);
4304 retval = security_task_getscheduler(p);
4308 attr.sched_policy = p->policy;
4309 if (p->sched_reset_on_fork)
4310 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4311 if (task_has_dl_policy(p))
4312 __getparam_dl(p, &attr);
4313 else if (task_has_rt_policy(p))
4314 attr.sched_priority = p->rt_priority;
4316 attr.sched_nice = task_nice(p);
4320 retval = sched_read_attr(uattr, &attr, size);
4328 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4330 cpumask_var_t cpus_allowed, new_mask;
4331 struct task_struct *p;
4336 p = find_process_by_pid(pid);
4342 /* Prevent p going away */
4346 if (p->flags & PF_NO_SETAFFINITY) {
4350 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4354 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4356 goto out_free_cpus_allowed;
4359 if (!check_same_owner(p)) {
4361 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4363 goto out_free_new_mask;
4368 retval = security_task_setscheduler(p);
4370 goto out_free_new_mask;
4373 cpuset_cpus_allowed(p, cpus_allowed);
4374 cpumask_and(new_mask, in_mask, cpus_allowed);
4377 * Since bandwidth control happens on root_domain basis,
4378 * if admission test is enabled, we only admit -deadline
4379 * tasks allowed to run on all the CPUs in the task's
4383 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4385 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4388 goto out_free_new_mask;
4394 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4397 cpuset_cpus_allowed(p, cpus_allowed);
4398 if (!cpumask_subset(new_mask, cpus_allowed)) {
4400 * We must have raced with a concurrent cpuset
4401 * update. Just reset the cpus_allowed to the
4402 * cpuset's cpus_allowed
4404 cpumask_copy(new_mask, cpus_allowed);
4409 free_cpumask_var(new_mask);
4410 out_free_cpus_allowed:
4411 free_cpumask_var(cpus_allowed);
4417 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4418 struct cpumask *new_mask)
4420 if (len < cpumask_size())
4421 cpumask_clear(new_mask);
4422 else if (len > cpumask_size())
4423 len = cpumask_size();
4425 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4429 * sys_sched_setaffinity - set the cpu affinity of a process
4430 * @pid: pid of the process
4431 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4432 * @user_mask_ptr: user-space pointer to the new cpu mask
4434 * Return: 0 on success. An error code otherwise.
4436 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4437 unsigned long __user *, user_mask_ptr)
4439 cpumask_var_t new_mask;
4442 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4445 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4447 retval = sched_setaffinity(pid, new_mask);
4448 free_cpumask_var(new_mask);
4452 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4454 struct task_struct *p;
4455 unsigned long flags;
4461 p = find_process_by_pid(pid);
4465 retval = security_task_getscheduler(p);
4469 raw_spin_lock_irqsave(&p->pi_lock, flags);
4470 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4471 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4480 * sys_sched_getaffinity - get the cpu affinity of a process
4481 * @pid: pid of the process
4482 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4483 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4485 * Return: 0 on success. An error code otherwise.
4487 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4488 unsigned long __user *, user_mask_ptr)
4493 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4495 if (len & (sizeof(unsigned long)-1))
4498 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4501 ret = sched_getaffinity(pid, mask);
4503 size_t retlen = min_t(size_t, len, cpumask_size());
4505 if (copy_to_user(user_mask_ptr, mask, retlen))
4510 free_cpumask_var(mask);
4516 * sys_sched_yield - yield the current processor to other threads.
4518 * This function yields the current CPU to other tasks. If there are no
4519 * other threads running on this CPU then this function will return.
4523 SYSCALL_DEFINE0(sched_yield)
4525 struct rq *rq = this_rq_lock();
4527 schedstat_inc(rq, yld_count);
4528 current->sched_class->yield_task(rq);
4531 * Since we are going to call schedule() anyway, there's
4532 * no need to preempt or enable interrupts:
4534 __release(rq->lock);
4535 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4536 do_raw_spin_unlock(&rq->lock);
4537 sched_preempt_enable_no_resched();
4544 int __sched _cond_resched(void)
4546 if (should_resched(0)) {
4547 preempt_schedule_common();
4552 EXPORT_SYMBOL(_cond_resched);
4555 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4556 * call schedule, and on return reacquire the lock.
4558 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4559 * operations here to prevent schedule() from being called twice (once via
4560 * spin_unlock(), once by hand).
4562 int __cond_resched_lock(spinlock_t *lock)
4564 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4567 lockdep_assert_held(lock);
4569 if (spin_needbreak(lock) || resched) {
4572 preempt_schedule_common();
4580 EXPORT_SYMBOL(__cond_resched_lock);
4582 int __sched __cond_resched_softirq(void)
4584 BUG_ON(!in_softirq());
4586 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4588 preempt_schedule_common();
4594 EXPORT_SYMBOL(__cond_resched_softirq);
4597 * yield - yield the current processor to other threads.
4599 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4601 * The scheduler is at all times free to pick the calling task as the most
4602 * eligible task to run, if removing the yield() call from your code breaks
4603 * it, its already broken.
4605 * Typical broken usage is:
4610 * where one assumes that yield() will let 'the other' process run that will
4611 * make event true. If the current task is a SCHED_FIFO task that will never
4612 * happen. Never use yield() as a progress guarantee!!
4614 * If you want to use yield() to wait for something, use wait_event().
4615 * If you want to use yield() to be 'nice' for others, use cond_resched().
4616 * If you still want to use yield(), do not!
4618 void __sched yield(void)
4620 set_current_state(TASK_RUNNING);
4623 EXPORT_SYMBOL(yield);
4626 * yield_to - yield the current processor to another thread in
4627 * your thread group, or accelerate that thread toward the
4628 * processor it's on.
4630 * @preempt: whether task preemption is allowed or not
4632 * It's the caller's job to ensure that the target task struct
4633 * can't go away on us before we can do any checks.
4636 * true (>0) if we indeed boosted the target task.
4637 * false (0) if we failed to boost the target.
4638 * -ESRCH if there's no task to yield to.
4640 int __sched yield_to(struct task_struct *p, bool preempt)
4642 struct task_struct *curr = current;
4643 struct rq *rq, *p_rq;
4644 unsigned long flags;
4647 local_irq_save(flags);
4653 * If we're the only runnable task on the rq and target rq also
4654 * has only one task, there's absolutely no point in yielding.
4656 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4661 double_rq_lock(rq, p_rq);
4662 if (task_rq(p) != p_rq) {
4663 double_rq_unlock(rq, p_rq);
4667 if (!curr->sched_class->yield_to_task)
4670 if (curr->sched_class != p->sched_class)
4673 if (task_running(p_rq, p) || p->state)
4676 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4678 schedstat_inc(rq, yld_count);
4680 * Make p's CPU reschedule; pick_next_entity takes care of
4683 if (preempt && rq != p_rq)
4688 double_rq_unlock(rq, p_rq);
4690 local_irq_restore(flags);
4697 EXPORT_SYMBOL_GPL(yield_to);
4700 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4701 * that process accounting knows that this is a task in IO wait state.
4703 long __sched io_schedule_timeout(long timeout)
4705 int old_iowait = current->in_iowait;
4709 current->in_iowait = 1;
4710 blk_schedule_flush_plug(current);
4712 delayacct_blkio_start();
4714 atomic_inc(&rq->nr_iowait);
4715 ret = schedule_timeout(timeout);
4716 current->in_iowait = old_iowait;
4717 atomic_dec(&rq->nr_iowait);
4718 delayacct_blkio_end();
4722 EXPORT_SYMBOL(io_schedule_timeout);
4725 * sys_sched_get_priority_max - return maximum RT priority.
4726 * @policy: scheduling class.
4728 * Return: On success, this syscall returns the maximum
4729 * rt_priority that can be used by a given scheduling class.
4730 * On failure, a negative error code is returned.
4732 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4739 ret = MAX_USER_RT_PRIO-1;
4741 case SCHED_DEADLINE:
4752 * sys_sched_get_priority_min - return minimum RT priority.
4753 * @policy: scheduling class.
4755 * Return: On success, this syscall returns the minimum
4756 * rt_priority that can be used by a given scheduling class.
4757 * On failure, a negative error code is returned.
4759 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4768 case SCHED_DEADLINE:
4778 * sys_sched_rr_get_interval - return the default timeslice of a process.
4779 * @pid: pid of the process.
4780 * @interval: userspace pointer to the timeslice value.
4782 * this syscall writes the default timeslice value of a given process
4783 * into the user-space timespec buffer. A value of '0' means infinity.
4785 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4788 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4789 struct timespec __user *, interval)
4791 struct task_struct *p;
4792 unsigned int time_slice;
4793 unsigned long flags;
4803 p = find_process_by_pid(pid);
4807 retval = security_task_getscheduler(p);
4811 rq = task_rq_lock(p, &flags);
4813 if (p->sched_class->get_rr_interval)
4814 time_slice = p->sched_class->get_rr_interval(rq, p);
4815 task_rq_unlock(rq, p, &flags);
4818 jiffies_to_timespec(time_slice, &t);
4819 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4827 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4829 void sched_show_task(struct task_struct *p)
4831 unsigned long free = 0;
4833 unsigned long state = p->state;
4836 state = __ffs(state) + 1;
4837 printk(KERN_INFO "%-15.15s %c", p->comm,
4838 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4839 #if BITS_PER_LONG == 32
4840 if (state == TASK_RUNNING)
4841 printk(KERN_CONT " running ");
4843 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4845 if (state == TASK_RUNNING)
4846 printk(KERN_CONT " running task ");
4848 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4850 #ifdef CONFIG_DEBUG_STACK_USAGE
4851 free = stack_not_used(p);
4856 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4858 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4859 task_pid_nr(p), ppid,
4860 (unsigned long)task_thread_info(p)->flags);
4862 print_worker_info(KERN_INFO, p);
4863 show_stack(p, NULL);
4866 void show_state_filter(unsigned long state_filter)
4868 struct task_struct *g, *p;
4870 #if BITS_PER_LONG == 32
4872 " task PC stack pid father\n");
4875 " task PC stack pid father\n");
4878 for_each_process_thread(g, p) {
4880 * reset the NMI-timeout, listing all files on a slow
4881 * console might take a lot of time:
4883 touch_nmi_watchdog();
4884 if (!state_filter || (p->state & state_filter))
4888 touch_all_softlockup_watchdogs();
4890 #ifdef CONFIG_SCHED_DEBUG
4891 sysrq_sched_debug_show();
4895 * Only show locks if all tasks are dumped:
4898 debug_show_all_locks();
4901 void init_idle_bootup_task(struct task_struct *idle)
4903 idle->sched_class = &idle_sched_class;
4907 * init_idle - set up an idle thread for a given CPU
4908 * @idle: task in question
4909 * @cpu: cpu the idle task belongs to
4911 * NOTE: this function does not set the idle thread's NEED_RESCHED
4912 * flag, to make booting more robust.
4914 void init_idle(struct task_struct *idle, int cpu)
4916 struct rq *rq = cpu_rq(cpu);
4917 unsigned long flags;
4919 raw_spin_lock_irqsave(&idle->pi_lock, flags);
4920 raw_spin_lock(&rq->lock);
4922 __sched_fork(0, idle);
4923 idle->state = TASK_RUNNING;
4924 idle->se.exec_start = sched_clock();
4926 do_set_cpus_allowed(idle, cpumask_of(cpu));
4928 * We're having a chicken and egg problem, even though we are
4929 * holding rq->lock, the cpu isn't yet set to this cpu so the
4930 * lockdep check in task_group() will fail.
4932 * Similar case to sched_fork(). / Alternatively we could
4933 * use task_rq_lock() here and obtain the other rq->lock.
4938 __set_task_cpu(idle, cpu);
4941 rq->curr = rq->idle = idle;
4942 idle->on_rq = TASK_ON_RQ_QUEUED;
4943 #if defined(CONFIG_SMP)
4946 raw_spin_unlock(&rq->lock);
4947 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
4949 /* Set the preempt count _outside_ the spinlocks! */
4950 init_idle_preempt_count(idle, cpu);
4953 * The idle tasks have their own, simple scheduling class:
4955 idle->sched_class = &idle_sched_class;
4956 ftrace_graph_init_idle_task(idle, cpu);
4957 vtime_init_idle(idle, cpu);
4958 #if defined(CONFIG_SMP)
4959 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4963 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4964 const struct cpumask *trial)
4966 int ret = 1, trial_cpus;
4967 struct dl_bw *cur_dl_b;
4968 unsigned long flags;
4970 if (!cpumask_weight(cur))
4973 rcu_read_lock_sched();
4974 cur_dl_b = dl_bw_of(cpumask_any(cur));
4975 trial_cpus = cpumask_weight(trial);
4977 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4978 if (cur_dl_b->bw != -1 &&
4979 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4981 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4982 rcu_read_unlock_sched();
4987 int task_can_attach(struct task_struct *p,
4988 const struct cpumask *cs_cpus_allowed)
4993 * Kthreads which disallow setaffinity shouldn't be moved
4994 * to a new cpuset; we don't want to change their cpu
4995 * affinity and isolating such threads by their set of
4996 * allowed nodes is unnecessary. Thus, cpusets are not
4997 * applicable for such threads. This prevents checking for
4998 * success of set_cpus_allowed_ptr() on all attached tasks
4999 * before cpus_allowed may be changed.
5001 if (p->flags & PF_NO_SETAFFINITY) {
5007 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5009 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5014 unsigned long flags;
5016 rcu_read_lock_sched();
5017 dl_b = dl_bw_of(dest_cpu);
5018 raw_spin_lock_irqsave(&dl_b->lock, flags);
5019 cpus = dl_bw_cpus(dest_cpu);
5020 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5025 * We reserve space for this task in the destination
5026 * root_domain, as we can't fail after this point.
5027 * We will free resources in the source root_domain
5028 * later on (see set_cpus_allowed_dl()).
5030 __dl_add(dl_b, p->dl.dl_bw);
5032 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5033 rcu_read_unlock_sched();
5043 #ifdef CONFIG_NUMA_BALANCING
5044 /* Migrate current task p to target_cpu */
5045 int migrate_task_to(struct task_struct *p, int target_cpu)
5047 struct migration_arg arg = { p, target_cpu };
5048 int curr_cpu = task_cpu(p);
5050 if (curr_cpu == target_cpu)
5053 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5056 /* TODO: This is not properly updating schedstats */
5058 trace_sched_move_numa(p, curr_cpu, target_cpu);
5059 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5063 * Requeue a task on a given node and accurately track the number of NUMA
5064 * tasks on the runqueues
5066 void sched_setnuma(struct task_struct *p, int nid)
5069 unsigned long flags;
5070 bool queued, running;
5072 rq = task_rq_lock(p, &flags);
5073 queued = task_on_rq_queued(p);
5074 running = task_current(rq, p);
5077 dequeue_task(rq, p, 0);
5079 put_prev_task(rq, p);
5081 p->numa_preferred_nid = nid;
5084 p->sched_class->set_curr_task(rq);
5086 enqueue_task(rq, p, 0);
5087 task_rq_unlock(rq, p, &flags);
5089 #endif /* CONFIG_NUMA_BALANCING */
5091 #ifdef CONFIG_HOTPLUG_CPU
5093 * Ensures that the idle task is using init_mm right before its cpu goes
5096 void idle_task_exit(void)
5098 struct mm_struct *mm = current->active_mm;
5100 BUG_ON(cpu_online(smp_processor_id()));
5102 if (mm != &init_mm) {
5103 switch_mm(mm, &init_mm, current);
5104 finish_arch_post_lock_switch();
5110 * Since this CPU is going 'away' for a while, fold any nr_active delta
5111 * we might have. Assumes we're called after migrate_tasks() so that the
5112 * nr_active count is stable.
5114 * Also see the comment "Global load-average calculations".
5116 static void calc_load_migrate(struct rq *rq)
5118 long delta = calc_load_fold_active(rq);
5120 atomic_long_add(delta, &calc_load_tasks);
5123 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5127 static const struct sched_class fake_sched_class = {
5128 .put_prev_task = put_prev_task_fake,
5131 static struct task_struct fake_task = {
5133 * Avoid pull_{rt,dl}_task()
5135 .prio = MAX_PRIO + 1,
5136 .sched_class = &fake_sched_class,
5140 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5141 * try_to_wake_up()->select_task_rq().
5143 * Called with rq->lock held even though we'er in stop_machine() and
5144 * there's no concurrency possible, we hold the required locks anyway
5145 * because of lock validation efforts.
5147 static void migrate_tasks(struct rq *dead_rq)
5149 struct rq *rq = dead_rq;
5150 struct task_struct *next, *stop = rq->stop;
5154 * Fudge the rq selection such that the below task selection loop
5155 * doesn't get stuck on the currently eligible stop task.
5157 * We're currently inside stop_machine() and the rq is either stuck
5158 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5159 * either way we should never end up calling schedule() until we're
5165 * put_prev_task() and pick_next_task() sched
5166 * class method both need to have an up-to-date
5167 * value of rq->clock[_task]
5169 update_rq_clock(rq);
5173 * There's this thread running, bail when that's the only
5176 if (rq->nr_running == 1)
5180 * pick_next_task assumes pinned rq->lock.
5182 lockdep_pin_lock(&rq->lock);
5183 next = pick_next_task(rq, &fake_task);
5185 next->sched_class->put_prev_task(rq, next);
5188 * Rules for changing task_struct::cpus_allowed are holding
5189 * both pi_lock and rq->lock, such that holding either
5190 * stabilizes the mask.
5192 * Drop rq->lock is not quite as disastrous as it usually is
5193 * because !cpu_active at this point, which means load-balance
5194 * will not interfere. Also, stop-machine.
5196 lockdep_unpin_lock(&rq->lock);
5197 raw_spin_unlock(&rq->lock);
5198 raw_spin_lock(&next->pi_lock);
5199 raw_spin_lock(&rq->lock);
5202 * Since we're inside stop-machine, _nothing_ should have
5203 * changed the task, WARN if weird stuff happened, because in
5204 * that case the above rq->lock drop is a fail too.
5206 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5207 raw_spin_unlock(&next->pi_lock);
5211 /* Find suitable destination for @next, with force if needed. */
5212 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5214 rq = __migrate_task(rq, next, dest_cpu);
5215 if (rq != dead_rq) {
5216 raw_spin_unlock(&rq->lock);
5218 raw_spin_lock(&rq->lock);
5220 raw_spin_unlock(&next->pi_lock);
5225 #endif /* CONFIG_HOTPLUG_CPU */
5227 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5229 static struct ctl_table sd_ctl_dir[] = {
5231 .procname = "sched_domain",
5237 static struct ctl_table sd_ctl_root[] = {
5239 .procname = "kernel",
5241 .child = sd_ctl_dir,
5246 static struct ctl_table *sd_alloc_ctl_entry(int n)
5248 struct ctl_table *entry =
5249 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5254 static void sd_free_ctl_entry(struct ctl_table **tablep)
5256 struct ctl_table *entry;
5259 * In the intermediate directories, both the child directory and
5260 * procname are dynamically allocated and could fail but the mode
5261 * will always be set. In the lowest directory the names are
5262 * static strings and all have proc handlers.
5264 for (entry = *tablep; entry->mode; entry++) {
5266 sd_free_ctl_entry(&entry->child);
5267 if (entry->proc_handler == NULL)
5268 kfree(entry->procname);
5275 static int min_load_idx = 0;
5276 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5279 set_table_entry(struct ctl_table *entry,
5280 const char *procname, void *data, int maxlen,
5281 umode_t mode, proc_handler *proc_handler,
5284 entry->procname = procname;
5286 entry->maxlen = maxlen;
5288 entry->proc_handler = proc_handler;
5291 entry->extra1 = &min_load_idx;
5292 entry->extra2 = &max_load_idx;
5296 static struct ctl_table *
5297 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5299 struct ctl_table *table = sd_alloc_ctl_entry(14);
5304 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5305 sizeof(long), 0644, proc_doulongvec_minmax, false);
5306 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5307 sizeof(long), 0644, proc_doulongvec_minmax, false);
5308 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5309 sizeof(int), 0644, proc_dointvec_minmax, true);
5310 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5311 sizeof(int), 0644, proc_dointvec_minmax, true);
5312 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5313 sizeof(int), 0644, proc_dointvec_minmax, true);
5314 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5315 sizeof(int), 0644, proc_dointvec_minmax, true);
5316 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5317 sizeof(int), 0644, proc_dointvec_minmax, true);
5318 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5319 sizeof(int), 0644, proc_dointvec_minmax, false);
5320 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5321 sizeof(int), 0644, proc_dointvec_minmax, false);
5322 set_table_entry(&table[9], "cache_nice_tries",
5323 &sd->cache_nice_tries,
5324 sizeof(int), 0644, proc_dointvec_minmax, false);
5325 set_table_entry(&table[10], "flags", &sd->flags,
5326 sizeof(int), 0644, proc_dointvec_minmax, false);
5327 set_table_entry(&table[11], "max_newidle_lb_cost",
5328 &sd->max_newidle_lb_cost,
5329 sizeof(long), 0644, proc_doulongvec_minmax, false);
5330 set_table_entry(&table[12], "name", sd->name,
5331 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5332 /* &table[13] is terminator */
5337 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5339 struct ctl_table *entry, *table;
5340 struct sched_domain *sd;
5341 int domain_num = 0, i;
5344 for_each_domain(cpu, sd)
5346 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5351 for_each_domain(cpu, sd) {
5352 snprintf(buf, 32, "domain%d", i);
5353 entry->procname = kstrdup(buf, GFP_KERNEL);
5355 entry->child = sd_alloc_ctl_domain_table(sd);
5362 static struct ctl_table_header *sd_sysctl_header;
5363 static void register_sched_domain_sysctl(void)
5365 int i, cpu_num = num_possible_cpus();
5366 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5369 WARN_ON(sd_ctl_dir[0].child);
5370 sd_ctl_dir[0].child = entry;
5375 for_each_possible_cpu(i) {
5376 snprintf(buf, 32, "cpu%d", i);
5377 entry->procname = kstrdup(buf, GFP_KERNEL);
5379 entry->child = sd_alloc_ctl_cpu_table(i);
5383 WARN_ON(sd_sysctl_header);
5384 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5387 /* may be called multiple times per register */
5388 static void unregister_sched_domain_sysctl(void)
5390 unregister_sysctl_table(sd_sysctl_header);
5391 sd_sysctl_header = NULL;
5392 if (sd_ctl_dir[0].child)
5393 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5396 static void register_sched_domain_sysctl(void)
5399 static void unregister_sched_domain_sysctl(void)
5402 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5404 static void set_rq_online(struct rq *rq)
5407 const struct sched_class *class;
5409 cpumask_set_cpu(rq->cpu, rq->rd->online);
5412 for_each_class(class) {
5413 if (class->rq_online)
5414 class->rq_online(rq);
5419 static void set_rq_offline(struct rq *rq)
5422 const struct sched_class *class;
5424 for_each_class(class) {
5425 if (class->rq_offline)
5426 class->rq_offline(rq);
5429 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5435 * migration_call - callback that gets triggered when a CPU is added.
5436 * Here we can start up the necessary migration thread for the new CPU.
5439 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5441 int cpu = (long)hcpu;
5442 unsigned long flags;
5443 struct rq *rq = cpu_rq(cpu);
5445 switch (action & ~CPU_TASKS_FROZEN) {
5447 case CPU_UP_PREPARE:
5448 rq->calc_load_update = calc_load_update;
5452 /* Update our root-domain */
5453 raw_spin_lock_irqsave(&rq->lock, flags);
5455 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5459 raw_spin_unlock_irqrestore(&rq->lock, flags);
5462 #ifdef CONFIG_HOTPLUG_CPU
5464 sched_ttwu_pending();
5465 /* Update our root-domain */
5466 raw_spin_lock_irqsave(&rq->lock, flags);
5468 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5472 BUG_ON(rq->nr_running != 1); /* the migration thread */
5473 raw_spin_unlock_irqrestore(&rq->lock, flags);
5477 calc_load_migrate(rq);
5482 update_max_interval();
5488 * Register at high priority so that task migration (migrate_all_tasks)
5489 * happens before everything else. This has to be lower priority than
5490 * the notifier in the perf_event subsystem, though.
5492 static struct notifier_block migration_notifier = {
5493 .notifier_call = migration_call,
5494 .priority = CPU_PRI_MIGRATION,
5497 static void set_cpu_rq_start_time(void)
5499 int cpu = smp_processor_id();
5500 struct rq *rq = cpu_rq(cpu);
5501 rq->age_stamp = sched_clock_cpu(cpu);
5504 static int sched_cpu_active(struct notifier_block *nfb,
5505 unsigned long action, void *hcpu)
5507 switch (action & ~CPU_TASKS_FROZEN) {
5509 set_cpu_rq_start_time();
5513 * At this point a starting CPU has marked itself as online via
5514 * set_cpu_online(). But it might not yet have marked itself
5515 * as active, which is essential from here on.
5517 * Thus, fall-through and help the starting CPU along.
5519 case CPU_DOWN_FAILED:
5520 set_cpu_active((long)hcpu, true);
5527 static int sched_cpu_inactive(struct notifier_block *nfb,
5528 unsigned long action, void *hcpu)
5530 switch (action & ~CPU_TASKS_FROZEN) {
5531 case CPU_DOWN_PREPARE:
5532 set_cpu_active((long)hcpu, false);
5539 static int __init migration_init(void)
5541 void *cpu = (void *)(long)smp_processor_id();
5544 /* Initialize migration for the boot CPU */
5545 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5546 BUG_ON(err == NOTIFY_BAD);
5547 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5548 register_cpu_notifier(&migration_notifier);
5550 /* Register cpu active notifiers */
5551 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5552 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5556 early_initcall(migration_init);
5558 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5560 #ifdef CONFIG_SCHED_DEBUG
5562 static __read_mostly int sched_debug_enabled;
5564 static int __init sched_debug_setup(char *str)
5566 sched_debug_enabled = 1;
5570 early_param("sched_debug", sched_debug_setup);
5572 static inline bool sched_debug(void)
5574 return sched_debug_enabled;
5577 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5578 struct cpumask *groupmask)
5580 struct sched_group *group = sd->groups;
5582 cpumask_clear(groupmask);
5584 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5586 if (!(sd->flags & SD_LOAD_BALANCE)) {
5587 printk("does not load-balance\n");
5589 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5594 printk(KERN_CONT "span %*pbl level %s\n",
5595 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5597 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5598 printk(KERN_ERR "ERROR: domain->span does not contain "
5601 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5602 printk(KERN_ERR "ERROR: domain->groups does not contain"
5606 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5610 printk(KERN_ERR "ERROR: group is NULL\n");
5614 if (!cpumask_weight(sched_group_cpus(group))) {
5615 printk(KERN_CONT "\n");
5616 printk(KERN_ERR "ERROR: empty group\n");
5620 if (!(sd->flags & SD_OVERLAP) &&
5621 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5622 printk(KERN_CONT "\n");
5623 printk(KERN_ERR "ERROR: repeated CPUs\n");
5627 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5629 printk(KERN_CONT " %*pbl",
5630 cpumask_pr_args(sched_group_cpus(group)));
5631 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5632 printk(KERN_CONT " (cpu_capacity = %d)",
5633 group->sgc->capacity);
5636 group = group->next;
5637 } while (group != sd->groups);
5638 printk(KERN_CONT "\n");
5640 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5641 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5644 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5645 printk(KERN_ERR "ERROR: parent span is not a superset "
5646 "of domain->span\n");
5650 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5654 if (!sched_debug_enabled)
5658 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5662 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5665 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5673 #else /* !CONFIG_SCHED_DEBUG */
5674 # define sched_domain_debug(sd, cpu) do { } while (0)
5675 static inline bool sched_debug(void)
5679 #endif /* CONFIG_SCHED_DEBUG */
5681 static int sd_degenerate(struct sched_domain *sd)
5683 if (cpumask_weight(sched_domain_span(sd)) == 1)
5686 /* Following flags need at least 2 groups */
5687 if (sd->flags & (SD_LOAD_BALANCE |
5688 SD_BALANCE_NEWIDLE |
5691 SD_SHARE_CPUCAPACITY |
5692 SD_SHARE_PKG_RESOURCES |
5693 SD_SHARE_POWERDOMAIN)) {
5694 if (sd->groups != sd->groups->next)
5698 /* Following flags don't use groups */
5699 if (sd->flags & (SD_WAKE_AFFINE))
5706 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5708 unsigned long cflags = sd->flags, pflags = parent->flags;
5710 if (sd_degenerate(parent))
5713 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5716 /* Flags needing groups don't count if only 1 group in parent */
5717 if (parent->groups == parent->groups->next) {
5718 pflags &= ~(SD_LOAD_BALANCE |
5719 SD_BALANCE_NEWIDLE |
5722 SD_SHARE_CPUCAPACITY |
5723 SD_SHARE_PKG_RESOURCES |
5725 SD_SHARE_POWERDOMAIN);
5726 if (nr_node_ids == 1)
5727 pflags &= ~SD_SERIALIZE;
5729 if (~cflags & pflags)
5735 static void free_rootdomain(struct rcu_head *rcu)
5737 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5739 cpupri_cleanup(&rd->cpupri);
5740 cpudl_cleanup(&rd->cpudl);
5741 free_cpumask_var(rd->dlo_mask);
5742 free_cpumask_var(rd->rto_mask);
5743 free_cpumask_var(rd->online);
5744 free_cpumask_var(rd->span);
5748 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5750 struct root_domain *old_rd = NULL;
5751 unsigned long flags;
5753 raw_spin_lock_irqsave(&rq->lock, flags);
5758 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5761 cpumask_clear_cpu(rq->cpu, old_rd->span);
5764 * If we dont want to free the old_rd yet then
5765 * set old_rd to NULL to skip the freeing later
5768 if (!atomic_dec_and_test(&old_rd->refcount))
5772 atomic_inc(&rd->refcount);
5775 cpumask_set_cpu(rq->cpu, rd->span);
5776 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5779 raw_spin_unlock_irqrestore(&rq->lock, flags);
5782 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5785 static int init_rootdomain(struct root_domain *rd)
5787 memset(rd, 0, sizeof(*rd));
5789 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5791 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5793 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5795 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5798 init_dl_bw(&rd->dl_bw);
5799 if (cpudl_init(&rd->cpudl) != 0)
5802 if (cpupri_init(&rd->cpupri) != 0)
5807 free_cpumask_var(rd->rto_mask);
5809 free_cpumask_var(rd->dlo_mask);
5811 free_cpumask_var(rd->online);
5813 free_cpumask_var(rd->span);
5819 * By default the system creates a single root-domain with all cpus as
5820 * members (mimicking the global state we have today).
5822 struct root_domain def_root_domain;
5824 static void init_defrootdomain(void)
5826 init_rootdomain(&def_root_domain);
5828 atomic_set(&def_root_domain.refcount, 1);
5831 static struct root_domain *alloc_rootdomain(void)
5833 struct root_domain *rd;
5835 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5839 if (init_rootdomain(rd) != 0) {
5847 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5849 struct sched_group *tmp, *first;
5858 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5863 } while (sg != first);
5866 static void free_sched_domain(struct rcu_head *rcu)
5868 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5871 * If its an overlapping domain it has private groups, iterate and
5874 if (sd->flags & SD_OVERLAP) {
5875 free_sched_groups(sd->groups, 1);
5876 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5877 kfree(sd->groups->sgc);
5883 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5885 call_rcu(&sd->rcu, free_sched_domain);
5888 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5890 for (; sd; sd = sd->parent)
5891 destroy_sched_domain(sd, cpu);
5895 * Keep a special pointer to the highest sched_domain that has
5896 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5897 * allows us to avoid some pointer chasing select_idle_sibling().
5899 * Also keep a unique ID per domain (we use the first cpu number in
5900 * the cpumask of the domain), this allows us to quickly tell if
5901 * two cpus are in the same cache domain, see cpus_share_cache().
5903 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5904 DEFINE_PER_CPU(int, sd_llc_size);
5905 DEFINE_PER_CPU(int, sd_llc_id);
5906 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5907 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5908 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5910 static void update_top_cache_domain(int cpu)
5912 struct sched_domain *sd;
5913 struct sched_domain *busy_sd = NULL;
5917 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5919 id = cpumask_first(sched_domain_span(sd));
5920 size = cpumask_weight(sched_domain_span(sd));
5921 busy_sd = sd->parent; /* sd_busy */
5923 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5925 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5926 per_cpu(sd_llc_size, cpu) = size;
5927 per_cpu(sd_llc_id, cpu) = id;
5929 sd = lowest_flag_domain(cpu, SD_NUMA);
5930 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5932 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5933 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5937 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5938 * hold the hotplug lock.
5941 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5943 struct rq *rq = cpu_rq(cpu);
5944 struct sched_domain *tmp;
5946 /* Remove the sched domains which do not contribute to scheduling. */
5947 for (tmp = sd; tmp; ) {
5948 struct sched_domain *parent = tmp->parent;
5952 if (sd_parent_degenerate(tmp, parent)) {
5953 tmp->parent = parent->parent;
5955 parent->parent->child = tmp;
5957 * Transfer SD_PREFER_SIBLING down in case of a
5958 * degenerate parent; the spans match for this
5959 * so the property transfers.
5961 if (parent->flags & SD_PREFER_SIBLING)
5962 tmp->flags |= SD_PREFER_SIBLING;
5963 destroy_sched_domain(parent, cpu);
5968 if (sd && sd_degenerate(sd)) {
5971 destroy_sched_domain(tmp, cpu);
5976 sched_domain_debug(sd, cpu);
5978 rq_attach_root(rq, rd);
5980 rcu_assign_pointer(rq->sd, sd);
5981 destroy_sched_domains(tmp, cpu);
5983 update_top_cache_domain(cpu);
5986 /* Setup the mask of cpus configured for isolated domains */
5987 static int __init isolated_cpu_setup(char *str)
5989 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5990 cpulist_parse(str, cpu_isolated_map);
5994 __setup("isolcpus=", isolated_cpu_setup);
5997 struct sched_domain ** __percpu sd;
5998 struct root_domain *rd;
6009 * Build an iteration mask that can exclude certain CPUs from the upwards
6012 * Asymmetric node setups can result in situations where the domain tree is of
6013 * unequal depth, make sure to skip domains that already cover the entire
6016 * In that case build_sched_domains() will have terminated the iteration early
6017 * and our sibling sd spans will be empty. Domains should always include the
6018 * cpu they're built on, so check that.
6021 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6023 const struct cpumask *span = sched_domain_span(sd);
6024 struct sd_data *sdd = sd->private;
6025 struct sched_domain *sibling;
6028 for_each_cpu(i, span) {
6029 sibling = *per_cpu_ptr(sdd->sd, i);
6030 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6033 cpumask_set_cpu(i, sched_group_mask(sg));
6038 * Return the canonical balance cpu for this group, this is the first cpu
6039 * of this group that's also in the iteration mask.
6041 int group_balance_cpu(struct sched_group *sg)
6043 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6047 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6049 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6050 const struct cpumask *span = sched_domain_span(sd);
6051 struct cpumask *covered = sched_domains_tmpmask;
6052 struct sd_data *sdd = sd->private;
6053 struct sched_domain *sibling;
6056 cpumask_clear(covered);
6058 for_each_cpu(i, span) {
6059 struct cpumask *sg_span;
6061 if (cpumask_test_cpu(i, covered))
6064 sibling = *per_cpu_ptr(sdd->sd, i);
6066 /* See the comment near build_group_mask(). */
6067 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6070 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6071 GFP_KERNEL, cpu_to_node(cpu));
6076 sg_span = sched_group_cpus(sg);
6078 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6080 cpumask_set_cpu(i, sg_span);
6082 cpumask_or(covered, covered, sg_span);
6084 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6085 if (atomic_inc_return(&sg->sgc->ref) == 1)
6086 build_group_mask(sd, sg);
6089 * Initialize sgc->capacity such that even if we mess up the
6090 * domains and no possible iteration will get us here, we won't
6093 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6096 * Make sure the first group of this domain contains the
6097 * canonical balance cpu. Otherwise the sched_domain iteration
6098 * breaks. See update_sg_lb_stats().
6100 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6101 group_balance_cpu(sg) == cpu)
6111 sd->groups = groups;
6116 free_sched_groups(first, 0);
6121 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6123 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6124 struct sched_domain *child = sd->child;
6127 cpu = cpumask_first(sched_domain_span(child));
6130 *sg = *per_cpu_ptr(sdd->sg, cpu);
6131 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6132 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6139 * build_sched_groups will build a circular linked list of the groups
6140 * covered by the given span, and will set each group's ->cpumask correctly,
6141 * and ->cpu_capacity to 0.
6143 * Assumes the sched_domain tree is fully constructed
6146 build_sched_groups(struct sched_domain *sd, int cpu)
6148 struct sched_group *first = NULL, *last = NULL;
6149 struct sd_data *sdd = sd->private;
6150 const struct cpumask *span = sched_domain_span(sd);
6151 struct cpumask *covered;
6154 get_group(cpu, sdd, &sd->groups);
6155 atomic_inc(&sd->groups->ref);
6157 if (cpu != cpumask_first(span))
6160 lockdep_assert_held(&sched_domains_mutex);
6161 covered = sched_domains_tmpmask;
6163 cpumask_clear(covered);
6165 for_each_cpu(i, span) {
6166 struct sched_group *sg;
6169 if (cpumask_test_cpu(i, covered))
6172 group = get_group(i, sdd, &sg);
6173 cpumask_setall(sched_group_mask(sg));
6175 for_each_cpu(j, span) {
6176 if (get_group(j, sdd, NULL) != group)
6179 cpumask_set_cpu(j, covered);
6180 cpumask_set_cpu(j, sched_group_cpus(sg));
6195 * Initialize sched groups cpu_capacity.
6197 * cpu_capacity indicates the capacity of sched group, which is used while
6198 * distributing the load between different sched groups in a sched domain.
6199 * Typically cpu_capacity for all the groups in a sched domain will be same
6200 * unless there are asymmetries in the topology. If there are asymmetries,
6201 * group having more cpu_capacity will pickup more load compared to the
6202 * group having less cpu_capacity.
6204 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6206 struct sched_group *sg = sd->groups;
6211 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6213 } while (sg != sd->groups);
6215 if (cpu != group_balance_cpu(sg))
6218 update_group_capacity(sd, cpu);
6219 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6223 * Initializers for schedule domains
6224 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6227 static int default_relax_domain_level = -1;
6228 int sched_domain_level_max;
6230 static int __init setup_relax_domain_level(char *str)
6232 if (kstrtoint(str, 0, &default_relax_domain_level))
6233 pr_warn("Unable to set relax_domain_level\n");
6237 __setup("relax_domain_level=", setup_relax_domain_level);
6239 static void set_domain_attribute(struct sched_domain *sd,
6240 struct sched_domain_attr *attr)
6244 if (!attr || attr->relax_domain_level < 0) {
6245 if (default_relax_domain_level < 0)
6248 request = default_relax_domain_level;
6250 request = attr->relax_domain_level;
6251 if (request < sd->level) {
6252 /* turn off idle balance on this domain */
6253 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6255 /* turn on idle balance on this domain */
6256 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6260 static void __sdt_free(const struct cpumask *cpu_map);
6261 static int __sdt_alloc(const struct cpumask *cpu_map);
6263 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6264 const struct cpumask *cpu_map)
6268 if (!atomic_read(&d->rd->refcount))
6269 free_rootdomain(&d->rd->rcu); /* fall through */
6271 free_percpu(d->sd); /* fall through */
6273 __sdt_free(cpu_map); /* fall through */
6279 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6280 const struct cpumask *cpu_map)
6282 memset(d, 0, sizeof(*d));
6284 if (__sdt_alloc(cpu_map))
6285 return sa_sd_storage;
6286 d->sd = alloc_percpu(struct sched_domain *);
6288 return sa_sd_storage;
6289 d->rd = alloc_rootdomain();
6292 return sa_rootdomain;
6296 * NULL the sd_data elements we've used to build the sched_domain and
6297 * sched_group structure so that the subsequent __free_domain_allocs()
6298 * will not free the data we're using.
6300 static void claim_allocations(int cpu, struct sched_domain *sd)
6302 struct sd_data *sdd = sd->private;
6304 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6305 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6307 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6308 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6310 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6311 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6315 static int sched_domains_numa_levels;
6316 enum numa_topology_type sched_numa_topology_type;
6317 static int *sched_domains_numa_distance;
6318 int sched_max_numa_distance;
6319 static struct cpumask ***sched_domains_numa_masks;
6320 static int sched_domains_curr_level;
6324 * SD_flags allowed in topology descriptions.
6326 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6327 * SD_SHARE_PKG_RESOURCES - describes shared caches
6328 * SD_NUMA - describes NUMA topologies
6329 * SD_SHARE_POWERDOMAIN - describes shared power domain
6332 * SD_ASYM_PACKING - describes SMT quirks
6334 #define TOPOLOGY_SD_FLAGS \
6335 (SD_SHARE_CPUCAPACITY | \
6336 SD_SHARE_PKG_RESOURCES | \
6339 SD_SHARE_POWERDOMAIN)
6341 static struct sched_domain *
6342 sd_init(struct sched_domain_topology_level *tl, int cpu)
6344 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6345 int sd_weight, sd_flags = 0;
6349 * Ugly hack to pass state to sd_numa_mask()...
6351 sched_domains_curr_level = tl->numa_level;
6354 sd_weight = cpumask_weight(tl->mask(cpu));
6357 sd_flags = (*tl->sd_flags)();
6358 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6359 "wrong sd_flags in topology description\n"))
6360 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6362 *sd = (struct sched_domain){
6363 .min_interval = sd_weight,
6364 .max_interval = 2*sd_weight,
6366 .imbalance_pct = 125,
6368 .cache_nice_tries = 0,
6375 .flags = 1*SD_LOAD_BALANCE
6376 | 1*SD_BALANCE_NEWIDLE
6381 | 0*SD_SHARE_CPUCAPACITY
6382 | 0*SD_SHARE_PKG_RESOURCES
6384 | 0*SD_PREFER_SIBLING
6389 .last_balance = jiffies,
6390 .balance_interval = sd_weight,
6392 .max_newidle_lb_cost = 0,
6393 .next_decay_max_lb_cost = jiffies,
6394 #ifdef CONFIG_SCHED_DEBUG
6400 * Convert topological properties into behaviour.
6403 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6404 sd->flags |= SD_PREFER_SIBLING;
6405 sd->imbalance_pct = 110;
6406 sd->smt_gain = 1178; /* ~15% */
6408 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6409 sd->imbalance_pct = 117;
6410 sd->cache_nice_tries = 1;
6414 } else if (sd->flags & SD_NUMA) {
6415 sd->cache_nice_tries = 2;
6419 sd->flags |= SD_SERIALIZE;
6420 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6421 sd->flags &= ~(SD_BALANCE_EXEC |
6428 sd->flags |= SD_PREFER_SIBLING;
6429 sd->cache_nice_tries = 1;
6434 sd->private = &tl->data;
6440 * Topology list, bottom-up.
6442 static struct sched_domain_topology_level default_topology[] = {
6443 #ifdef CONFIG_SCHED_SMT
6444 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6446 #ifdef CONFIG_SCHED_MC
6447 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6449 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6453 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6455 #define for_each_sd_topology(tl) \
6456 for (tl = sched_domain_topology; tl->mask; tl++)
6458 void set_sched_topology(struct sched_domain_topology_level *tl)
6460 sched_domain_topology = tl;
6465 static const struct cpumask *sd_numa_mask(int cpu)
6467 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6470 static void sched_numa_warn(const char *str)
6472 static int done = false;
6480 printk(KERN_WARNING "ERROR: %s\n\n", str);
6482 for (i = 0; i < nr_node_ids; i++) {
6483 printk(KERN_WARNING " ");
6484 for (j = 0; j < nr_node_ids; j++)
6485 printk(KERN_CONT "%02d ", node_distance(i,j));
6486 printk(KERN_CONT "\n");
6488 printk(KERN_WARNING "\n");
6491 bool find_numa_distance(int distance)
6495 if (distance == node_distance(0, 0))
6498 for (i = 0; i < sched_domains_numa_levels; i++) {
6499 if (sched_domains_numa_distance[i] == distance)
6507 * A system can have three types of NUMA topology:
6508 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6509 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6510 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6512 * The difference between a glueless mesh topology and a backplane
6513 * topology lies in whether communication between not directly
6514 * connected nodes goes through intermediary nodes (where programs
6515 * could run), or through backplane controllers. This affects
6516 * placement of programs.
6518 * The type of topology can be discerned with the following tests:
6519 * - If the maximum distance between any nodes is 1 hop, the system
6520 * is directly connected.
6521 * - If for two nodes A and B, located N > 1 hops away from each other,
6522 * there is an intermediary node C, which is < N hops away from both
6523 * nodes A and B, the system is a glueless mesh.
6525 static void init_numa_topology_type(void)
6529 n = sched_max_numa_distance;
6531 if (sched_domains_numa_levels <= 1) {
6532 sched_numa_topology_type = NUMA_DIRECT;
6536 for_each_online_node(a) {
6537 for_each_online_node(b) {
6538 /* Find two nodes furthest removed from each other. */
6539 if (node_distance(a, b) < n)
6542 /* Is there an intermediary node between a and b? */
6543 for_each_online_node(c) {
6544 if (node_distance(a, c) < n &&
6545 node_distance(b, c) < n) {
6546 sched_numa_topology_type =
6552 sched_numa_topology_type = NUMA_BACKPLANE;
6558 static void sched_init_numa(void)
6560 int next_distance, curr_distance = node_distance(0, 0);
6561 struct sched_domain_topology_level *tl;
6565 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6566 if (!sched_domains_numa_distance)
6570 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6571 * unique distances in the node_distance() table.
6573 * Assumes node_distance(0,j) includes all distances in
6574 * node_distance(i,j) in order to avoid cubic time.
6576 next_distance = curr_distance;
6577 for (i = 0; i < nr_node_ids; i++) {
6578 for (j = 0; j < nr_node_ids; j++) {
6579 for (k = 0; k < nr_node_ids; k++) {
6580 int distance = node_distance(i, k);
6582 if (distance > curr_distance &&
6583 (distance < next_distance ||
6584 next_distance == curr_distance))
6585 next_distance = distance;
6588 * While not a strong assumption it would be nice to know
6589 * about cases where if node A is connected to B, B is not
6590 * equally connected to A.
6592 if (sched_debug() && node_distance(k, i) != distance)
6593 sched_numa_warn("Node-distance not symmetric");
6595 if (sched_debug() && i && !find_numa_distance(distance))
6596 sched_numa_warn("Node-0 not representative");
6598 if (next_distance != curr_distance) {
6599 sched_domains_numa_distance[level++] = next_distance;
6600 sched_domains_numa_levels = level;
6601 curr_distance = next_distance;
6606 * In case of sched_debug() we verify the above assumption.
6616 * 'level' contains the number of unique distances, excluding the
6617 * identity distance node_distance(i,i).
6619 * The sched_domains_numa_distance[] array includes the actual distance
6624 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6625 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6626 * the array will contain less then 'level' members. This could be
6627 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6628 * in other functions.
6630 * We reset it to 'level' at the end of this function.
6632 sched_domains_numa_levels = 0;
6634 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6635 if (!sched_domains_numa_masks)
6639 * Now for each level, construct a mask per node which contains all
6640 * cpus of nodes that are that many hops away from us.
6642 for (i = 0; i < level; i++) {
6643 sched_domains_numa_masks[i] =
6644 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6645 if (!sched_domains_numa_masks[i])
6648 for (j = 0; j < nr_node_ids; j++) {
6649 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6653 sched_domains_numa_masks[i][j] = mask;
6655 for (k = 0; k < nr_node_ids; k++) {
6656 if (node_distance(j, k) > sched_domains_numa_distance[i])
6659 cpumask_or(mask, mask, cpumask_of_node(k));
6664 /* Compute default topology size */
6665 for (i = 0; sched_domain_topology[i].mask; i++);
6667 tl = kzalloc((i + level + 1) *
6668 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6673 * Copy the default topology bits..
6675 for (i = 0; sched_domain_topology[i].mask; i++)
6676 tl[i] = sched_domain_topology[i];
6679 * .. and append 'j' levels of NUMA goodness.
6681 for (j = 0; j < level; i++, j++) {
6682 tl[i] = (struct sched_domain_topology_level){
6683 .mask = sd_numa_mask,
6684 .sd_flags = cpu_numa_flags,
6685 .flags = SDTL_OVERLAP,
6691 sched_domain_topology = tl;
6693 sched_domains_numa_levels = level;
6694 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6696 init_numa_topology_type();
6699 static void sched_domains_numa_masks_set(int cpu)
6702 int node = cpu_to_node(cpu);
6704 for (i = 0; i < sched_domains_numa_levels; i++) {
6705 for (j = 0; j < nr_node_ids; j++) {
6706 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6707 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6712 static void sched_domains_numa_masks_clear(int cpu)
6715 for (i = 0; i < sched_domains_numa_levels; i++) {
6716 for (j = 0; j < nr_node_ids; j++)
6717 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6722 * Update sched_domains_numa_masks[level][node] array when new cpus
6725 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6726 unsigned long action,
6729 int cpu = (long)hcpu;
6731 switch (action & ~CPU_TASKS_FROZEN) {
6733 sched_domains_numa_masks_set(cpu);
6737 sched_domains_numa_masks_clear(cpu);
6747 static inline void sched_init_numa(void)
6751 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6752 unsigned long action,
6757 #endif /* CONFIG_NUMA */
6759 static int __sdt_alloc(const struct cpumask *cpu_map)
6761 struct sched_domain_topology_level *tl;
6764 for_each_sd_topology(tl) {
6765 struct sd_data *sdd = &tl->data;
6767 sdd->sd = alloc_percpu(struct sched_domain *);
6771 sdd->sg = alloc_percpu(struct sched_group *);
6775 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6779 for_each_cpu(j, cpu_map) {
6780 struct sched_domain *sd;
6781 struct sched_group *sg;
6782 struct sched_group_capacity *sgc;
6784 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6785 GFP_KERNEL, cpu_to_node(j));
6789 *per_cpu_ptr(sdd->sd, j) = sd;
6791 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6792 GFP_KERNEL, cpu_to_node(j));
6798 *per_cpu_ptr(sdd->sg, j) = sg;
6800 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6801 GFP_KERNEL, cpu_to_node(j));
6805 *per_cpu_ptr(sdd->sgc, j) = sgc;
6812 static void __sdt_free(const struct cpumask *cpu_map)
6814 struct sched_domain_topology_level *tl;
6817 for_each_sd_topology(tl) {
6818 struct sd_data *sdd = &tl->data;
6820 for_each_cpu(j, cpu_map) {
6821 struct sched_domain *sd;
6824 sd = *per_cpu_ptr(sdd->sd, j);
6825 if (sd && (sd->flags & SD_OVERLAP))
6826 free_sched_groups(sd->groups, 0);
6827 kfree(*per_cpu_ptr(sdd->sd, j));
6831 kfree(*per_cpu_ptr(sdd->sg, j));
6833 kfree(*per_cpu_ptr(sdd->sgc, j));
6835 free_percpu(sdd->sd);
6837 free_percpu(sdd->sg);
6839 free_percpu(sdd->sgc);
6844 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6845 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6846 struct sched_domain *child, int cpu)
6848 struct sched_domain *sd = sd_init(tl, cpu);
6852 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6854 sd->level = child->level + 1;
6855 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6859 if (!cpumask_subset(sched_domain_span(child),
6860 sched_domain_span(sd))) {
6861 pr_err("BUG: arch topology borken\n");
6862 #ifdef CONFIG_SCHED_DEBUG
6863 pr_err(" the %s domain not a subset of the %s domain\n",
6864 child->name, sd->name);
6866 /* Fixup, ensure @sd has at least @child cpus. */
6867 cpumask_or(sched_domain_span(sd),
6868 sched_domain_span(sd),
6869 sched_domain_span(child));
6873 set_domain_attribute(sd, attr);
6879 * Build sched domains for a given set of cpus and attach the sched domains
6880 * to the individual cpus
6882 static int build_sched_domains(const struct cpumask *cpu_map,
6883 struct sched_domain_attr *attr)
6885 enum s_alloc alloc_state;
6886 struct sched_domain *sd;
6888 int i, ret = -ENOMEM;
6890 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6891 if (alloc_state != sa_rootdomain)
6894 /* Set up domains for cpus specified by the cpu_map. */
6895 for_each_cpu(i, cpu_map) {
6896 struct sched_domain_topology_level *tl;
6899 for_each_sd_topology(tl) {
6900 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6901 if (tl == sched_domain_topology)
6902 *per_cpu_ptr(d.sd, i) = sd;
6903 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6904 sd->flags |= SD_OVERLAP;
6905 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6910 /* Build the groups for the domains */
6911 for_each_cpu(i, cpu_map) {
6912 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6913 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6914 if (sd->flags & SD_OVERLAP) {
6915 if (build_overlap_sched_groups(sd, i))
6918 if (build_sched_groups(sd, i))
6924 /* Calculate CPU capacity for physical packages and nodes */
6925 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6926 if (!cpumask_test_cpu(i, cpu_map))
6929 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6930 claim_allocations(i, sd);
6931 init_sched_groups_capacity(i, sd);
6935 /* Attach the domains */
6937 for_each_cpu(i, cpu_map) {
6938 sd = *per_cpu_ptr(d.sd, i);
6939 cpu_attach_domain(sd, d.rd, i);
6945 __free_domain_allocs(&d, alloc_state, cpu_map);
6949 static cpumask_var_t *doms_cur; /* current sched domains */
6950 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6951 static struct sched_domain_attr *dattr_cur;
6952 /* attribues of custom domains in 'doms_cur' */
6955 * Special case: If a kmalloc of a doms_cur partition (array of
6956 * cpumask) fails, then fallback to a single sched domain,
6957 * as determined by the single cpumask fallback_doms.
6959 static cpumask_var_t fallback_doms;
6962 * arch_update_cpu_topology lets virtualized architectures update the
6963 * cpu core maps. It is supposed to return 1 if the topology changed
6964 * or 0 if it stayed the same.
6966 int __weak arch_update_cpu_topology(void)
6971 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6974 cpumask_var_t *doms;
6976 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6979 for (i = 0; i < ndoms; i++) {
6980 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6981 free_sched_domains(doms, i);
6988 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6991 for (i = 0; i < ndoms; i++)
6992 free_cpumask_var(doms[i]);
6997 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6998 * For now this just excludes isolated cpus, but could be used to
6999 * exclude other special cases in the future.
7001 static int init_sched_domains(const struct cpumask *cpu_map)
7005 arch_update_cpu_topology();
7007 doms_cur = alloc_sched_domains(ndoms_cur);
7009 doms_cur = &fallback_doms;
7010 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7011 err = build_sched_domains(doms_cur[0], NULL);
7012 register_sched_domain_sysctl();
7018 * Detach sched domains from a group of cpus specified in cpu_map
7019 * These cpus will now be attached to the NULL domain
7021 static void detach_destroy_domains(const struct cpumask *cpu_map)
7026 for_each_cpu(i, cpu_map)
7027 cpu_attach_domain(NULL, &def_root_domain, i);
7031 /* handle null as "default" */
7032 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7033 struct sched_domain_attr *new, int idx_new)
7035 struct sched_domain_attr tmp;
7042 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7043 new ? (new + idx_new) : &tmp,
7044 sizeof(struct sched_domain_attr));
7048 * Partition sched domains as specified by the 'ndoms_new'
7049 * cpumasks in the array doms_new[] of cpumasks. This compares
7050 * doms_new[] to the current sched domain partitioning, doms_cur[].
7051 * It destroys each deleted domain and builds each new domain.
7053 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7054 * The masks don't intersect (don't overlap.) We should setup one
7055 * sched domain for each mask. CPUs not in any of the cpumasks will
7056 * not be load balanced. If the same cpumask appears both in the
7057 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7060 * The passed in 'doms_new' should be allocated using
7061 * alloc_sched_domains. This routine takes ownership of it and will
7062 * free_sched_domains it when done with it. If the caller failed the
7063 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7064 * and partition_sched_domains() will fallback to the single partition
7065 * 'fallback_doms', it also forces the domains to be rebuilt.
7067 * If doms_new == NULL it will be replaced with cpu_online_mask.
7068 * ndoms_new == 0 is a special case for destroying existing domains,
7069 * and it will not create the default domain.
7071 * Call with hotplug lock held
7073 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7074 struct sched_domain_attr *dattr_new)
7079 mutex_lock(&sched_domains_mutex);
7081 /* always unregister in case we don't destroy any domains */
7082 unregister_sched_domain_sysctl();
7084 /* Let architecture update cpu core mappings. */
7085 new_topology = arch_update_cpu_topology();
7087 n = doms_new ? ndoms_new : 0;
7089 /* Destroy deleted domains */
7090 for (i = 0; i < ndoms_cur; i++) {
7091 for (j = 0; j < n && !new_topology; j++) {
7092 if (cpumask_equal(doms_cur[i], doms_new[j])
7093 && dattrs_equal(dattr_cur, i, dattr_new, j))
7096 /* no match - a current sched domain not in new doms_new[] */
7097 detach_destroy_domains(doms_cur[i]);
7103 if (doms_new == NULL) {
7105 doms_new = &fallback_doms;
7106 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7107 WARN_ON_ONCE(dattr_new);
7110 /* Build new domains */
7111 for (i = 0; i < ndoms_new; i++) {
7112 for (j = 0; j < n && !new_topology; j++) {
7113 if (cpumask_equal(doms_new[i], doms_cur[j])
7114 && dattrs_equal(dattr_new, i, dattr_cur, j))
7117 /* no match - add a new doms_new */
7118 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7123 /* Remember the new sched domains */
7124 if (doms_cur != &fallback_doms)
7125 free_sched_domains(doms_cur, ndoms_cur);
7126 kfree(dattr_cur); /* kfree(NULL) is safe */
7127 doms_cur = doms_new;
7128 dattr_cur = dattr_new;
7129 ndoms_cur = ndoms_new;
7131 register_sched_domain_sysctl();
7133 mutex_unlock(&sched_domains_mutex);
7136 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7139 * Update cpusets according to cpu_active mask. If cpusets are
7140 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7141 * around partition_sched_domains().
7143 * If we come here as part of a suspend/resume, don't touch cpusets because we
7144 * want to restore it back to its original state upon resume anyway.
7146 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7150 case CPU_ONLINE_FROZEN:
7151 case CPU_DOWN_FAILED_FROZEN:
7154 * num_cpus_frozen tracks how many CPUs are involved in suspend
7155 * resume sequence. As long as this is not the last online
7156 * operation in the resume sequence, just build a single sched
7157 * domain, ignoring cpusets.
7160 if (likely(num_cpus_frozen)) {
7161 partition_sched_domains(1, NULL, NULL);
7166 * This is the last CPU online operation. So fall through and
7167 * restore the original sched domains by considering the
7168 * cpuset configurations.
7172 cpuset_update_active_cpus(true);
7180 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7183 unsigned long flags;
7184 long cpu = (long)hcpu;
7190 case CPU_DOWN_PREPARE:
7191 rcu_read_lock_sched();
7192 dl_b = dl_bw_of(cpu);
7194 raw_spin_lock_irqsave(&dl_b->lock, flags);
7195 cpus = dl_bw_cpus(cpu);
7196 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7197 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7199 rcu_read_unlock_sched();
7202 return notifier_from_errno(-EBUSY);
7203 cpuset_update_active_cpus(false);
7205 case CPU_DOWN_PREPARE_FROZEN:
7207 partition_sched_domains(1, NULL, NULL);
7215 void __init sched_init_smp(void)
7217 cpumask_var_t non_isolated_cpus;
7219 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7220 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7222 /* nohz_full won't take effect without isolating the cpus. */
7223 tick_nohz_full_add_cpus_to(cpu_isolated_map);
7228 * There's no userspace yet to cause hotplug operations; hence all the
7229 * cpu masks are stable and all blatant races in the below code cannot
7232 mutex_lock(&sched_domains_mutex);
7233 init_sched_domains(cpu_active_mask);
7234 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7235 if (cpumask_empty(non_isolated_cpus))
7236 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7237 mutex_unlock(&sched_domains_mutex);
7239 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7240 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7241 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7245 /* Move init over to a non-isolated CPU */
7246 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7248 sched_init_granularity();
7249 free_cpumask_var(non_isolated_cpus);
7251 init_sched_rt_class();
7252 init_sched_dl_class();
7255 void __init sched_init_smp(void)
7257 sched_init_granularity();
7259 #endif /* CONFIG_SMP */
7261 int in_sched_functions(unsigned long addr)
7263 return in_lock_functions(addr) ||
7264 (addr >= (unsigned long)__sched_text_start
7265 && addr < (unsigned long)__sched_text_end);
7268 #ifdef CONFIG_CGROUP_SCHED
7270 * Default task group.
7271 * Every task in system belongs to this group at bootup.
7273 struct task_group root_task_group;
7274 LIST_HEAD(task_groups);
7277 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7279 void __init sched_init(void)
7282 unsigned long alloc_size = 0, ptr;
7284 #ifdef CONFIG_FAIR_GROUP_SCHED
7285 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7287 #ifdef CONFIG_RT_GROUP_SCHED
7288 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7291 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7293 #ifdef CONFIG_FAIR_GROUP_SCHED
7294 root_task_group.se = (struct sched_entity **)ptr;
7295 ptr += nr_cpu_ids * sizeof(void **);
7297 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7298 ptr += nr_cpu_ids * sizeof(void **);
7300 #endif /* CONFIG_FAIR_GROUP_SCHED */
7301 #ifdef CONFIG_RT_GROUP_SCHED
7302 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7303 ptr += nr_cpu_ids * sizeof(void **);
7305 root_task_group.rt_rq = (struct rt_rq **)ptr;
7306 ptr += nr_cpu_ids * sizeof(void **);
7308 #endif /* CONFIG_RT_GROUP_SCHED */
7310 #ifdef CONFIG_CPUMASK_OFFSTACK
7311 for_each_possible_cpu(i) {
7312 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7313 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7315 #endif /* CONFIG_CPUMASK_OFFSTACK */
7317 init_rt_bandwidth(&def_rt_bandwidth,
7318 global_rt_period(), global_rt_runtime());
7319 init_dl_bandwidth(&def_dl_bandwidth,
7320 global_rt_period(), global_rt_runtime());
7323 init_defrootdomain();
7326 #ifdef CONFIG_RT_GROUP_SCHED
7327 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7328 global_rt_period(), global_rt_runtime());
7329 #endif /* CONFIG_RT_GROUP_SCHED */
7331 #ifdef CONFIG_CGROUP_SCHED
7332 list_add(&root_task_group.list, &task_groups);
7333 INIT_LIST_HEAD(&root_task_group.children);
7334 INIT_LIST_HEAD(&root_task_group.siblings);
7335 autogroup_init(&init_task);
7337 #endif /* CONFIG_CGROUP_SCHED */
7339 for_each_possible_cpu(i) {
7343 raw_spin_lock_init(&rq->lock);
7345 rq->calc_load_active = 0;
7346 rq->calc_load_update = jiffies + LOAD_FREQ;
7347 init_cfs_rq(&rq->cfs);
7348 init_rt_rq(&rq->rt);
7349 init_dl_rq(&rq->dl);
7350 #ifdef CONFIG_FAIR_GROUP_SCHED
7351 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7352 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7354 * How much cpu bandwidth does root_task_group get?
7356 * In case of task-groups formed thr' the cgroup filesystem, it
7357 * gets 100% of the cpu resources in the system. This overall
7358 * system cpu resource is divided among the tasks of
7359 * root_task_group and its child task-groups in a fair manner,
7360 * based on each entity's (task or task-group's) weight
7361 * (se->load.weight).
7363 * In other words, if root_task_group has 10 tasks of weight
7364 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7365 * then A0's share of the cpu resource is:
7367 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7369 * We achieve this by letting root_task_group's tasks sit
7370 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7372 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7373 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7374 #endif /* CONFIG_FAIR_GROUP_SCHED */
7376 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7377 #ifdef CONFIG_RT_GROUP_SCHED
7378 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7381 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7382 rq->cpu_load[j] = 0;
7384 rq->last_load_update_tick = jiffies;
7389 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7390 rq->balance_callback = NULL;
7391 rq->active_balance = 0;
7392 rq->next_balance = jiffies;
7397 rq->avg_idle = 2*sysctl_sched_migration_cost;
7398 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7400 INIT_LIST_HEAD(&rq->cfs_tasks);
7402 rq_attach_root(rq, &def_root_domain);
7403 #ifdef CONFIG_NO_HZ_COMMON
7406 #ifdef CONFIG_NO_HZ_FULL
7407 rq->last_sched_tick = 0;
7411 atomic_set(&rq->nr_iowait, 0);
7414 set_load_weight(&init_task);
7416 #ifdef CONFIG_PREEMPT_NOTIFIERS
7417 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7421 * The boot idle thread does lazy MMU switching as well:
7423 atomic_inc(&init_mm.mm_count);
7424 enter_lazy_tlb(&init_mm, current);
7427 * During early bootup we pretend to be a normal task:
7429 current->sched_class = &fair_sched_class;
7432 * Make us the idle thread. Technically, schedule() should not be
7433 * called from this thread, however somewhere below it might be,
7434 * but because we are the idle thread, we just pick up running again
7435 * when this runqueue becomes "idle".
7437 init_idle(current, smp_processor_id());
7439 calc_load_update = jiffies + LOAD_FREQ;
7442 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7443 /* May be allocated at isolcpus cmdline parse time */
7444 if (cpu_isolated_map == NULL)
7445 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7446 idle_thread_set_boot_cpu();
7447 set_cpu_rq_start_time();
7449 init_sched_fair_class();
7451 scheduler_running = 1;
7454 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7455 static inline int preempt_count_equals(int preempt_offset)
7457 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7459 return (nested == preempt_offset);
7462 void __might_sleep(const char *file, int line, int preempt_offset)
7465 * Blocking primitives will set (and therefore destroy) current->state,
7466 * since we will exit with TASK_RUNNING make sure we enter with it,
7467 * otherwise we will destroy state.
7469 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7470 "do not call blocking ops when !TASK_RUNNING; "
7471 "state=%lx set at [<%p>] %pS\n",
7473 (void *)current->task_state_change,
7474 (void *)current->task_state_change);
7476 ___might_sleep(file, line, preempt_offset);
7478 EXPORT_SYMBOL(__might_sleep);
7480 void ___might_sleep(const char *file, int line, int preempt_offset)
7482 static unsigned long prev_jiffy; /* ratelimiting */
7484 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7485 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7486 !is_idle_task(current)) ||
7487 system_state != SYSTEM_RUNNING || oops_in_progress)
7489 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7491 prev_jiffy = jiffies;
7494 "BUG: sleeping function called from invalid context at %s:%d\n",
7497 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7498 in_atomic(), irqs_disabled(),
7499 current->pid, current->comm);
7501 if (task_stack_end_corrupted(current))
7502 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7504 debug_show_held_locks(current);
7505 if (irqs_disabled())
7506 print_irqtrace_events(current);
7507 #ifdef CONFIG_DEBUG_PREEMPT
7508 if (!preempt_count_equals(preempt_offset)) {
7509 pr_err("Preemption disabled at:");
7510 print_ip_sym(current->preempt_disable_ip);
7516 EXPORT_SYMBOL(___might_sleep);
7519 #ifdef CONFIG_MAGIC_SYSRQ
7520 void normalize_rt_tasks(void)
7522 struct task_struct *g, *p;
7523 struct sched_attr attr = {
7524 .sched_policy = SCHED_NORMAL,
7527 read_lock(&tasklist_lock);
7528 for_each_process_thread(g, p) {
7530 * Only normalize user tasks:
7532 if (p->flags & PF_KTHREAD)
7535 p->se.exec_start = 0;
7536 #ifdef CONFIG_SCHEDSTATS
7537 p->se.statistics.wait_start = 0;
7538 p->se.statistics.sleep_start = 0;
7539 p->se.statistics.block_start = 0;
7542 if (!dl_task(p) && !rt_task(p)) {
7544 * Renice negative nice level userspace
7547 if (task_nice(p) < 0)
7548 set_user_nice(p, 0);
7552 __sched_setscheduler(p, &attr, false, false);
7554 read_unlock(&tasklist_lock);
7557 #endif /* CONFIG_MAGIC_SYSRQ */
7559 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7561 * These functions are only useful for the IA64 MCA handling, or kdb.
7563 * They can only be called when the whole system has been
7564 * stopped - every CPU needs to be quiescent, and no scheduling
7565 * activity can take place. Using them for anything else would
7566 * be a serious bug, and as a result, they aren't even visible
7567 * under any other configuration.
7571 * curr_task - return the current task for a given cpu.
7572 * @cpu: the processor in question.
7574 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7576 * Return: The current task for @cpu.
7578 struct task_struct *curr_task(int cpu)
7580 return cpu_curr(cpu);
7583 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7587 * set_curr_task - set the current task for a given cpu.
7588 * @cpu: the processor in question.
7589 * @p: the task pointer to set.
7591 * Description: This function must only be used when non-maskable interrupts
7592 * are serviced on a separate stack. It allows the architecture to switch the
7593 * notion of the current task on a cpu in a non-blocking manner. This function
7594 * must be called with all CPU's synchronized, and interrupts disabled, the
7595 * and caller must save the original value of the current task (see
7596 * curr_task() above) and restore that value before reenabling interrupts and
7597 * re-starting the system.
7599 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7601 void set_curr_task(int cpu, struct task_struct *p)
7608 #ifdef CONFIG_CGROUP_SCHED
7609 /* task_group_lock serializes the addition/removal of task groups */
7610 static DEFINE_SPINLOCK(task_group_lock);
7612 static void free_sched_group(struct task_group *tg)
7614 free_fair_sched_group(tg);
7615 free_rt_sched_group(tg);
7620 /* allocate runqueue etc for a new task group */
7621 struct task_group *sched_create_group(struct task_group *parent)
7623 struct task_group *tg;
7625 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7627 return ERR_PTR(-ENOMEM);
7629 if (!alloc_fair_sched_group(tg, parent))
7632 if (!alloc_rt_sched_group(tg, parent))
7638 free_sched_group(tg);
7639 return ERR_PTR(-ENOMEM);
7642 void sched_online_group(struct task_group *tg, struct task_group *parent)
7644 unsigned long flags;
7646 spin_lock_irqsave(&task_group_lock, flags);
7647 list_add_rcu(&tg->list, &task_groups);
7649 WARN_ON(!parent); /* root should already exist */
7651 tg->parent = parent;
7652 INIT_LIST_HEAD(&tg->children);
7653 list_add_rcu(&tg->siblings, &parent->children);
7654 spin_unlock_irqrestore(&task_group_lock, flags);
7657 /* rcu callback to free various structures associated with a task group */
7658 static void free_sched_group_rcu(struct rcu_head *rhp)
7660 /* now it should be safe to free those cfs_rqs */
7661 free_sched_group(container_of(rhp, struct task_group, rcu));
7664 /* Destroy runqueue etc associated with a task group */
7665 void sched_destroy_group(struct task_group *tg)
7667 /* wait for possible concurrent references to cfs_rqs complete */
7668 call_rcu(&tg->rcu, free_sched_group_rcu);
7671 void sched_offline_group(struct task_group *tg)
7673 unsigned long flags;
7676 /* end participation in shares distribution */
7677 for_each_possible_cpu(i)
7678 unregister_fair_sched_group(tg, i);
7680 spin_lock_irqsave(&task_group_lock, flags);
7681 list_del_rcu(&tg->list);
7682 list_del_rcu(&tg->siblings);
7683 spin_unlock_irqrestore(&task_group_lock, flags);
7686 /* change task's runqueue when it moves between groups.
7687 * The caller of this function should have put the task in its new group
7688 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7689 * reflect its new group.
7691 void sched_move_task(struct task_struct *tsk)
7693 struct task_group *tg;
7694 int queued, running;
7695 unsigned long flags;
7698 rq = task_rq_lock(tsk, &flags);
7700 running = task_current(rq, tsk);
7701 queued = task_on_rq_queued(tsk);
7704 dequeue_task(rq, tsk, 0);
7705 if (unlikely(running))
7706 put_prev_task(rq, tsk);
7709 * All callers are synchronized by task_rq_lock(); we do not use RCU
7710 * which is pointless here. Thus, we pass "true" to task_css_check()
7711 * to prevent lockdep warnings.
7713 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7714 struct task_group, css);
7715 tg = autogroup_task_group(tsk, tg);
7716 tsk->sched_task_group = tg;
7718 #ifdef CONFIG_FAIR_GROUP_SCHED
7719 if (tsk->sched_class->task_move_group)
7720 tsk->sched_class->task_move_group(tsk);
7723 set_task_rq(tsk, task_cpu(tsk));
7725 if (unlikely(running))
7726 tsk->sched_class->set_curr_task(rq);
7728 enqueue_task(rq, tsk, 0);
7730 task_rq_unlock(rq, tsk, &flags);
7732 #endif /* CONFIG_CGROUP_SCHED */
7734 #ifdef CONFIG_RT_GROUP_SCHED
7736 * Ensure that the real time constraints are schedulable.
7738 static DEFINE_MUTEX(rt_constraints_mutex);
7740 /* Must be called with tasklist_lock held */
7741 static inline int tg_has_rt_tasks(struct task_group *tg)
7743 struct task_struct *g, *p;
7746 * Autogroups do not have RT tasks; see autogroup_create().
7748 if (task_group_is_autogroup(tg))
7751 for_each_process_thread(g, p) {
7752 if (rt_task(p) && task_group(p) == tg)
7759 struct rt_schedulable_data {
7760 struct task_group *tg;
7765 static int tg_rt_schedulable(struct task_group *tg, void *data)
7767 struct rt_schedulable_data *d = data;
7768 struct task_group *child;
7769 unsigned long total, sum = 0;
7770 u64 period, runtime;
7772 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7773 runtime = tg->rt_bandwidth.rt_runtime;
7776 period = d->rt_period;
7777 runtime = d->rt_runtime;
7781 * Cannot have more runtime than the period.
7783 if (runtime > period && runtime != RUNTIME_INF)
7787 * Ensure we don't starve existing RT tasks.
7789 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7792 total = to_ratio(period, runtime);
7795 * Nobody can have more than the global setting allows.
7797 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7801 * The sum of our children's runtime should not exceed our own.
7803 list_for_each_entry_rcu(child, &tg->children, siblings) {
7804 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7805 runtime = child->rt_bandwidth.rt_runtime;
7807 if (child == d->tg) {
7808 period = d->rt_period;
7809 runtime = d->rt_runtime;
7812 sum += to_ratio(period, runtime);
7821 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7825 struct rt_schedulable_data data = {
7827 .rt_period = period,
7828 .rt_runtime = runtime,
7832 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7838 static int tg_set_rt_bandwidth(struct task_group *tg,
7839 u64 rt_period, u64 rt_runtime)
7844 * Disallowing the root group RT runtime is BAD, it would disallow the
7845 * kernel creating (and or operating) RT threads.
7847 if (tg == &root_task_group && rt_runtime == 0)
7850 /* No period doesn't make any sense. */
7854 mutex_lock(&rt_constraints_mutex);
7855 read_lock(&tasklist_lock);
7856 err = __rt_schedulable(tg, rt_period, rt_runtime);
7860 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7861 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7862 tg->rt_bandwidth.rt_runtime = rt_runtime;
7864 for_each_possible_cpu(i) {
7865 struct rt_rq *rt_rq = tg->rt_rq[i];
7867 raw_spin_lock(&rt_rq->rt_runtime_lock);
7868 rt_rq->rt_runtime = rt_runtime;
7869 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7871 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7873 read_unlock(&tasklist_lock);
7874 mutex_unlock(&rt_constraints_mutex);
7879 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7881 u64 rt_runtime, rt_period;
7883 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7884 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7885 if (rt_runtime_us < 0)
7886 rt_runtime = RUNTIME_INF;
7888 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7891 static long sched_group_rt_runtime(struct task_group *tg)
7895 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7898 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7899 do_div(rt_runtime_us, NSEC_PER_USEC);
7900 return rt_runtime_us;
7903 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7905 u64 rt_runtime, rt_period;
7907 rt_period = rt_period_us * NSEC_PER_USEC;
7908 rt_runtime = tg->rt_bandwidth.rt_runtime;
7910 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7913 static long sched_group_rt_period(struct task_group *tg)
7917 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7918 do_div(rt_period_us, NSEC_PER_USEC);
7919 return rt_period_us;
7921 #endif /* CONFIG_RT_GROUP_SCHED */
7923 #ifdef CONFIG_RT_GROUP_SCHED
7924 static int sched_rt_global_constraints(void)
7928 mutex_lock(&rt_constraints_mutex);
7929 read_lock(&tasklist_lock);
7930 ret = __rt_schedulable(NULL, 0, 0);
7931 read_unlock(&tasklist_lock);
7932 mutex_unlock(&rt_constraints_mutex);
7937 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7939 /* Don't accept realtime tasks when there is no way for them to run */
7940 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7946 #else /* !CONFIG_RT_GROUP_SCHED */
7947 static int sched_rt_global_constraints(void)
7949 unsigned long flags;
7952 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7953 for_each_possible_cpu(i) {
7954 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7956 raw_spin_lock(&rt_rq->rt_runtime_lock);
7957 rt_rq->rt_runtime = global_rt_runtime();
7958 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7960 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7964 #endif /* CONFIG_RT_GROUP_SCHED */
7966 static int sched_dl_global_validate(void)
7968 u64 runtime = global_rt_runtime();
7969 u64 period = global_rt_period();
7970 u64 new_bw = to_ratio(period, runtime);
7973 unsigned long flags;
7976 * Here we want to check the bandwidth not being set to some
7977 * value smaller than the currently allocated bandwidth in
7978 * any of the root_domains.
7980 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7981 * cycling on root_domains... Discussion on different/better
7982 * solutions is welcome!
7984 for_each_possible_cpu(cpu) {
7985 rcu_read_lock_sched();
7986 dl_b = dl_bw_of(cpu);
7988 raw_spin_lock_irqsave(&dl_b->lock, flags);
7989 if (new_bw < dl_b->total_bw)
7991 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7993 rcu_read_unlock_sched();
8002 static void sched_dl_do_global(void)
8007 unsigned long flags;
8009 def_dl_bandwidth.dl_period = global_rt_period();
8010 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8012 if (global_rt_runtime() != RUNTIME_INF)
8013 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8016 * FIXME: As above...
8018 for_each_possible_cpu(cpu) {
8019 rcu_read_lock_sched();
8020 dl_b = dl_bw_of(cpu);
8022 raw_spin_lock_irqsave(&dl_b->lock, flags);
8024 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8026 rcu_read_unlock_sched();
8030 static int sched_rt_global_validate(void)
8032 if (sysctl_sched_rt_period <= 0)
8035 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8036 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8042 static void sched_rt_do_global(void)
8044 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8045 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8048 int sched_rt_handler(struct ctl_table *table, int write,
8049 void __user *buffer, size_t *lenp,
8052 int old_period, old_runtime;
8053 static DEFINE_MUTEX(mutex);
8057 old_period = sysctl_sched_rt_period;
8058 old_runtime = sysctl_sched_rt_runtime;
8060 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8062 if (!ret && write) {
8063 ret = sched_rt_global_validate();
8067 ret = sched_dl_global_validate();
8071 ret = sched_rt_global_constraints();
8075 sched_rt_do_global();
8076 sched_dl_do_global();
8080 sysctl_sched_rt_period = old_period;
8081 sysctl_sched_rt_runtime = old_runtime;
8083 mutex_unlock(&mutex);
8088 int sched_rr_handler(struct ctl_table *table, int write,
8089 void __user *buffer, size_t *lenp,
8093 static DEFINE_MUTEX(mutex);
8096 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8097 /* make sure that internally we keep jiffies */
8098 /* also, writing zero resets timeslice to default */
8099 if (!ret && write) {
8100 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8101 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8103 mutex_unlock(&mutex);
8107 #ifdef CONFIG_CGROUP_SCHED
8109 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8111 return css ? container_of(css, struct task_group, css) : NULL;
8114 static struct cgroup_subsys_state *
8115 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8117 struct task_group *parent = css_tg(parent_css);
8118 struct task_group *tg;
8121 /* This is early initialization for the top cgroup */
8122 return &root_task_group.css;
8125 tg = sched_create_group(parent);
8127 return ERR_PTR(-ENOMEM);
8132 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8134 struct task_group *tg = css_tg(css);
8135 struct task_group *parent = css_tg(css->parent);
8138 sched_online_group(tg, parent);
8142 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8144 struct task_group *tg = css_tg(css);
8146 sched_destroy_group(tg);
8149 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8151 struct task_group *tg = css_tg(css);
8153 sched_offline_group(tg);
8156 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8158 sched_move_task(task);
8161 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8162 struct cgroup_taskset *tset)
8164 struct task_struct *task;
8166 cgroup_taskset_for_each(task, tset) {
8167 #ifdef CONFIG_RT_GROUP_SCHED
8168 if (!sched_rt_can_attach(css_tg(css), task))
8171 /* We don't support RT-tasks being in separate groups */
8172 if (task->sched_class != &fair_sched_class)
8179 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8180 struct cgroup_taskset *tset)
8182 struct task_struct *task;
8184 cgroup_taskset_for_each(task, tset)
8185 sched_move_task(task);
8188 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8189 struct cgroup_subsys_state *old_css,
8190 struct task_struct *task)
8192 sched_move_task(task);
8195 #ifdef CONFIG_FAIR_GROUP_SCHED
8196 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8197 struct cftype *cftype, u64 shareval)
8199 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8202 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8205 struct task_group *tg = css_tg(css);
8207 return (u64) scale_load_down(tg->shares);
8210 #ifdef CONFIG_CFS_BANDWIDTH
8211 static DEFINE_MUTEX(cfs_constraints_mutex);
8213 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8214 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8216 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8218 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8220 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8221 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8223 if (tg == &root_task_group)
8227 * Ensure we have at some amount of bandwidth every period. This is
8228 * to prevent reaching a state of large arrears when throttled via
8229 * entity_tick() resulting in prolonged exit starvation.
8231 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8235 * Likewise, bound things on the otherside by preventing insane quota
8236 * periods. This also allows us to normalize in computing quota
8239 if (period > max_cfs_quota_period)
8243 * Prevent race between setting of cfs_rq->runtime_enabled and
8244 * unthrottle_offline_cfs_rqs().
8247 mutex_lock(&cfs_constraints_mutex);
8248 ret = __cfs_schedulable(tg, period, quota);
8252 runtime_enabled = quota != RUNTIME_INF;
8253 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8255 * If we need to toggle cfs_bandwidth_used, off->on must occur
8256 * before making related changes, and on->off must occur afterwards
8258 if (runtime_enabled && !runtime_was_enabled)
8259 cfs_bandwidth_usage_inc();
8260 raw_spin_lock_irq(&cfs_b->lock);
8261 cfs_b->period = ns_to_ktime(period);
8262 cfs_b->quota = quota;
8264 __refill_cfs_bandwidth_runtime(cfs_b);
8265 /* restart the period timer (if active) to handle new period expiry */
8266 if (runtime_enabled)
8267 start_cfs_bandwidth(cfs_b);
8268 raw_spin_unlock_irq(&cfs_b->lock);
8270 for_each_online_cpu(i) {
8271 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8272 struct rq *rq = cfs_rq->rq;
8274 raw_spin_lock_irq(&rq->lock);
8275 cfs_rq->runtime_enabled = runtime_enabled;
8276 cfs_rq->runtime_remaining = 0;
8278 if (cfs_rq->throttled)
8279 unthrottle_cfs_rq(cfs_rq);
8280 raw_spin_unlock_irq(&rq->lock);
8282 if (runtime_was_enabled && !runtime_enabled)
8283 cfs_bandwidth_usage_dec();
8285 mutex_unlock(&cfs_constraints_mutex);
8291 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8295 period = ktime_to_ns(tg->cfs_bandwidth.period);
8296 if (cfs_quota_us < 0)
8297 quota = RUNTIME_INF;
8299 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8301 return tg_set_cfs_bandwidth(tg, period, quota);
8304 long tg_get_cfs_quota(struct task_group *tg)
8308 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8311 quota_us = tg->cfs_bandwidth.quota;
8312 do_div(quota_us, NSEC_PER_USEC);
8317 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8321 period = (u64)cfs_period_us * NSEC_PER_USEC;
8322 quota = tg->cfs_bandwidth.quota;
8324 return tg_set_cfs_bandwidth(tg, period, quota);
8327 long tg_get_cfs_period(struct task_group *tg)
8331 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8332 do_div(cfs_period_us, NSEC_PER_USEC);
8334 return cfs_period_us;
8337 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8340 return tg_get_cfs_quota(css_tg(css));
8343 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8344 struct cftype *cftype, s64 cfs_quota_us)
8346 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8349 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8352 return tg_get_cfs_period(css_tg(css));
8355 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8356 struct cftype *cftype, u64 cfs_period_us)
8358 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8361 struct cfs_schedulable_data {
8362 struct task_group *tg;
8367 * normalize group quota/period to be quota/max_period
8368 * note: units are usecs
8370 static u64 normalize_cfs_quota(struct task_group *tg,
8371 struct cfs_schedulable_data *d)
8379 period = tg_get_cfs_period(tg);
8380 quota = tg_get_cfs_quota(tg);
8383 /* note: these should typically be equivalent */
8384 if (quota == RUNTIME_INF || quota == -1)
8387 return to_ratio(period, quota);
8390 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8392 struct cfs_schedulable_data *d = data;
8393 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8394 s64 quota = 0, parent_quota = -1;
8397 quota = RUNTIME_INF;
8399 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8401 quota = normalize_cfs_quota(tg, d);
8402 parent_quota = parent_b->hierarchical_quota;
8405 * ensure max(child_quota) <= parent_quota, inherit when no
8408 if (quota == RUNTIME_INF)
8409 quota = parent_quota;
8410 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8413 cfs_b->hierarchical_quota = quota;
8418 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8421 struct cfs_schedulable_data data = {
8427 if (quota != RUNTIME_INF) {
8428 do_div(data.period, NSEC_PER_USEC);
8429 do_div(data.quota, NSEC_PER_USEC);
8433 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8439 static int cpu_stats_show(struct seq_file *sf, void *v)
8441 struct task_group *tg = css_tg(seq_css(sf));
8442 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8444 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8445 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8446 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8450 #endif /* CONFIG_CFS_BANDWIDTH */
8451 #endif /* CONFIG_FAIR_GROUP_SCHED */
8453 #ifdef CONFIG_RT_GROUP_SCHED
8454 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8455 struct cftype *cft, s64 val)
8457 return sched_group_set_rt_runtime(css_tg(css), val);
8460 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8463 return sched_group_rt_runtime(css_tg(css));
8466 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8467 struct cftype *cftype, u64 rt_period_us)
8469 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8472 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8475 return sched_group_rt_period(css_tg(css));
8477 #endif /* CONFIG_RT_GROUP_SCHED */
8479 static struct cftype cpu_files[] = {
8480 #ifdef CONFIG_FAIR_GROUP_SCHED
8483 .read_u64 = cpu_shares_read_u64,
8484 .write_u64 = cpu_shares_write_u64,
8487 #ifdef CONFIG_CFS_BANDWIDTH
8489 .name = "cfs_quota_us",
8490 .read_s64 = cpu_cfs_quota_read_s64,
8491 .write_s64 = cpu_cfs_quota_write_s64,
8494 .name = "cfs_period_us",
8495 .read_u64 = cpu_cfs_period_read_u64,
8496 .write_u64 = cpu_cfs_period_write_u64,
8500 .seq_show = cpu_stats_show,
8503 #ifdef CONFIG_RT_GROUP_SCHED
8505 .name = "rt_runtime_us",
8506 .read_s64 = cpu_rt_runtime_read,
8507 .write_s64 = cpu_rt_runtime_write,
8510 .name = "rt_period_us",
8511 .read_u64 = cpu_rt_period_read_uint,
8512 .write_u64 = cpu_rt_period_write_uint,
8518 struct cgroup_subsys cpu_cgrp_subsys = {
8519 .css_alloc = cpu_cgroup_css_alloc,
8520 .css_free = cpu_cgroup_css_free,
8521 .css_online = cpu_cgroup_css_online,
8522 .css_offline = cpu_cgroup_css_offline,
8523 .fork = cpu_cgroup_fork,
8524 .can_attach = cpu_cgroup_can_attach,
8525 .attach = cpu_cgroup_attach,
8526 .exit = cpu_cgroup_exit,
8527 .legacy_cftypes = cpu_files,
8531 #endif /* CONFIG_CGROUP_SCHED */
8533 void dump_cpu_task(int cpu)
8535 pr_info("Task dump for CPU %d:\n", cpu);
8536 sched_show_task(cpu_curr(cpu));