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 (idle_policy(p->policy)) {
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 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2119 #ifdef CONFIG_NUMA_BALANCING
2121 void set_numabalancing_state(bool enabled)
2124 static_branch_enable(&sched_numa_balancing);
2126 static_branch_disable(&sched_numa_balancing);
2129 #ifdef CONFIG_PROC_SYSCTL
2130 int sysctl_numa_balancing(struct ctl_table *table, int write,
2131 void __user *buffer, size_t *lenp, loff_t *ppos)
2135 int state = static_branch_likely(&sched_numa_balancing);
2137 if (write && !capable(CAP_SYS_ADMIN))
2142 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2146 set_numabalancing_state(state);
2153 * fork()/clone()-time setup:
2155 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2157 unsigned long flags;
2158 int cpu = get_cpu();
2160 __sched_fork(clone_flags, p);
2162 * We mark the process as running here. This guarantees that
2163 * nobody will actually run it, and a signal or other external
2164 * event cannot wake it up and insert it on the runqueue either.
2166 p->state = TASK_RUNNING;
2169 * Make sure we do not leak PI boosting priority to the child.
2171 p->prio = current->normal_prio;
2174 * Revert to default priority/policy on fork if requested.
2176 if (unlikely(p->sched_reset_on_fork)) {
2177 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2178 p->policy = SCHED_NORMAL;
2179 p->static_prio = NICE_TO_PRIO(0);
2181 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2182 p->static_prio = NICE_TO_PRIO(0);
2184 p->prio = p->normal_prio = __normal_prio(p);
2188 * We don't need the reset flag anymore after the fork. It has
2189 * fulfilled its duty:
2191 p->sched_reset_on_fork = 0;
2194 if (dl_prio(p->prio)) {
2197 } else if (rt_prio(p->prio)) {
2198 p->sched_class = &rt_sched_class;
2200 p->sched_class = &fair_sched_class;
2203 if (p->sched_class->task_fork)
2204 p->sched_class->task_fork(p);
2207 * The child is not yet in the pid-hash so no cgroup attach races,
2208 * and the cgroup is pinned to this child due to cgroup_fork()
2209 * is ran before sched_fork().
2211 * Silence PROVE_RCU.
2213 raw_spin_lock_irqsave(&p->pi_lock, flags);
2214 set_task_cpu(p, cpu);
2215 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2217 #ifdef CONFIG_SCHED_INFO
2218 if (likely(sched_info_on()))
2219 memset(&p->sched_info, 0, sizeof(p->sched_info));
2221 #if defined(CONFIG_SMP)
2224 init_task_preempt_count(p);
2226 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2227 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2234 unsigned long to_ratio(u64 period, u64 runtime)
2236 if (runtime == RUNTIME_INF)
2240 * Doing this here saves a lot of checks in all
2241 * the calling paths, and returning zero seems
2242 * safe for them anyway.
2247 return div64_u64(runtime << 20, period);
2251 inline struct dl_bw *dl_bw_of(int i)
2253 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2254 "sched RCU must be held");
2255 return &cpu_rq(i)->rd->dl_bw;
2258 static inline int dl_bw_cpus(int i)
2260 struct root_domain *rd = cpu_rq(i)->rd;
2263 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2264 "sched RCU must be held");
2265 for_each_cpu_and(i, rd->span, cpu_active_mask)
2271 inline struct dl_bw *dl_bw_of(int i)
2273 return &cpu_rq(i)->dl.dl_bw;
2276 static inline int dl_bw_cpus(int i)
2283 * We must be sure that accepting a new task (or allowing changing the
2284 * parameters of an existing one) is consistent with the bandwidth
2285 * constraints. If yes, this function also accordingly updates the currently
2286 * allocated bandwidth to reflect the new situation.
2288 * This function is called while holding p's rq->lock.
2290 * XXX we should delay bw change until the task's 0-lag point, see
2293 static int dl_overflow(struct task_struct *p, int policy,
2294 const struct sched_attr *attr)
2297 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2298 u64 period = attr->sched_period ?: attr->sched_deadline;
2299 u64 runtime = attr->sched_runtime;
2300 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2303 if (new_bw == p->dl.dl_bw)
2307 * Either if a task, enters, leave, or stays -deadline but changes
2308 * its parameters, we may need to update accordingly the total
2309 * allocated bandwidth of the container.
2311 raw_spin_lock(&dl_b->lock);
2312 cpus = dl_bw_cpus(task_cpu(p));
2313 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2314 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2315 __dl_add(dl_b, new_bw);
2317 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2318 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2319 __dl_clear(dl_b, p->dl.dl_bw);
2320 __dl_add(dl_b, new_bw);
2322 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2323 __dl_clear(dl_b, p->dl.dl_bw);
2326 raw_spin_unlock(&dl_b->lock);
2331 extern void init_dl_bw(struct dl_bw *dl_b);
2334 * wake_up_new_task - wake up a newly created task for the first time.
2336 * This function will do some initial scheduler statistics housekeeping
2337 * that must be done for every newly created context, then puts the task
2338 * on the runqueue and wakes it.
2340 void wake_up_new_task(struct task_struct *p)
2342 unsigned long flags;
2345 raw_spin_lock_irqsave(&p->pi_lock, flags);
2346 /* Initialize new task's runnable average */
2347 init_entity_runnable_average(&p->se);
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 rq = __task_rq_lock(p);
2358 activate_task(rq, p, 0);
2359 p->on_rq = TASK_ON_RQ_QUEUED;
2360 trace_sched_wakeup_new(p);
2361 check_preempt_curr(rq, p, WF_FORK);
2363 if (p->sched_class->task_woken)
2364 p->sched_class->task_woken(rq, p);
2366 task_rq_unlock(rq, p, &flags);
2369 #ifdef CONFIG_PREEMPT_NOTIFIERS
2371 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2373 void preempt_notifier_inc(void)
2375 static_key_slow_inc(&preempt_notifier_key);
2377 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2379 void preempt_notifier_dec(void)
2381 static_key_slow_dec(&preempt_notifier_key);
2383 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2386 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2387 * @notifier: notifier struct to register
2389 void preempt_notifier_register(struct preempt_notifier *notifier)
2391 if (!static_key_false(&preempt_notifier_key))
2392 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2394 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2396 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2399 * preempt_notifier_unregister - no longer interested in preemption notifications
2400 * @notifier: notifier struct to unregister
2402 * This is *not* safe to call from within a preemption notifier.
2404 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2406 hlist_del(¬ifier->link);
2408 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2410 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2412 struct preempt_notifier *notifier;
2414 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2415 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2418 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2420 if (static_key_false(&preempt_notifier_key))
2421 __fire_sched_in_preempt_notifiers(curr);
2425 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2426 struct task_struct *next)
2428 struct preempt_notifier *notifier;
2430 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2431 notifier->ops->sched_out(notifier, next);
2434 static __always_inline void
2435 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2436 struct task_struct *next)
2438 if (static_key_false(&preempt_notifier_key))
2439 __fire_sched_out_preempt_notifiers(curr, next);
2442 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2444 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2449 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2450 struct task_struct *next)
2454 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2457 * prepare_task_switch - prepare to switch tasks
2458 * @rq: the runqueue preparing to switch
2459 * @prev: the current task that is being switched out
2460 * @next: the task we are going to switch to.
2462 * This is called with the rq lock held and interrupts off. It must
2463 * be paired with a subsequent finish_task_switch after the context
2466 * prepare_task_switch sets up locking and calls architecture specific
2470 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2471 struct task_struct *next)
2473 sched_info_switch(rq, prev, next);
2474 perf_event_task_sched_out(prev, next);
2475 fire_sched_out_preempt_notifiers(prev, next);
2476 prepare_lock_switch(rq, next);
2477 prepare_arch_switch(next);
2481 * finish_task_switch - clean up after a task-switch
2482 * @prev: the thread we just switched away from.
2484 * finish_task_switch must be called after the context switch, paired
2485 * with a prepare_task_switch call before the context switch.
2486 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2487 * and do any other architecture-specific cleanup actions.
2489 * Note that we may have delayed dropping an mm in context_switch(). If
2490 * so, we finish that here outside of the runqueue lock. (Doing it
2491 * with the lock held can cause deadlocks; see schedule() for
2494 * The context switch have flipped the stack from under us and restored the
2495 * local variables which were saved when this task called schedule() in the
2496 * past. prev == current is still correct but we need to recalculate this_rq
2497 * because prev may have moved to another CPU.
2499 static struct rq *finish_task_switch(struct task_struct *prev)
2500 __releases(rq->lock)
2502 struct rq *rq = this_rq();
2503 struct mm_struct *mm = rq->prev_mm;
2507 * The previous task will have left us with a preempt_count of 2
2508 * because it left us after:
2511 * preempt_disable(); // 1
2513 * raw_spin_lock_irq(&rq->lock) // 2
2515 * Also, see FORK_PREEMPT_COUNT.
2521 * A task struct has one reference for the use as "current".
2522 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2523 * schedule one last time. The schedule call will never return, and
2524 * the scheduled task must drop that reference.
2526 * We must observe prev->state before clearing prev->on_cpu (in
2527 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2528 * running on another CPU and we could rave with its RUNNING -> DEAD
2529 * transition, resulting in a double drop.
2531 prev_state = prev->state;
2532 vtime_task_switch(prev);
2533 perf_event_task_sched_in(prev, current);
2534 finish_lock_switch(rq, prev);
2535 finish_arch_post_lock_switch();
2537 fire_sched_in_preempt_notifiers(current);
2540 if (unlikely(prev_state == TASK_DEAD)) {
2541 if (prev->sched_class->task_dead)
2542 prev->sched_class->task_dead(prev);
2545 * Remove function-return probe instances associated with this
2546 * task and put them back on the free list.
2548 kprobe_flush_task(prev);
2549 put_task_struct(prev);
2552 tick_nohz_task_switch();
2558 /* rq->lock is NOT held, but preemption is disabled */
2559 static void __balance_callback(struct rq *rq)
2561 struct callback_head *head, *next;
2562 void (*func)(struct rq *rq);
2563 unsigned long flags;
2565 raw_spin_lock_irqsave(&rq->lock, flags);
2566 head = rq->balance_callback;
2567 rq->balance_callback = NULL;
2569 func = (void (*)(struct rq *))head->func;
2576 raw_spin_unlock_irqrestore(&rq->lock, flags);
2579 static inline void balance_callback(struct rq *rq)
2581 if (unlikely(rq->balance_callback))
2582 __balance_callback(rq);
2587 static inline void balance_callback(struct rq *rq)
2594 * schedule_tail - first thing a freshly forked thread must call.
2595 * @prev: the thread we just switched away from.
2597 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2598 __releases(rq->lock)
2603 * New tasks start with FORK_PREEMPT_COUNT, see there and
2604 * finish_task_switch() for details.
2606 * finish_task_switch() will drop rq->lock() and lower preempt_count
2607 * and the preempt_enable() will end up enabling preemption (on
2608 * PREEMPT_COUNT kernels).
2611 rq = finish_task_switch(prev);
2612 balance_callback(rq);
2615 if (current->set_child_tid)
2616 put_user(task_pid_vnr(current), current->set_child_tid);
2620 * context_switch - switch to the new MM and the new thread's register state.
2622 static inline struct rq *
2623 context_switch(struct rq *rq, struct task_struct *prev,
2624 struct task_struct *next)
2626 struct mm_struct *mm, *oldmm;
2628 prepare_task_switch(rq, prev, next);
2631 oldmm = prev->active_mm;
2633 * For paravirt, this is coupled with an exit in switch_to to
2634 * combine the page table reload and the switch backend into
2637 arch_start_context_switch(prev);
2640 next->active_mm = oldmm;
2641 atomic_inc(&oldmm->mm_count);
2642 enter_lazy_tlb(oldmm, next);
2644 switch_mm(oldmm, mm, next);
2647 prev->active_mm = NULL;
2648 rq->prev_mm = oldmm;
2651 * Since the runqueue lock will be released by the next
2652 * task (which is an invalid locking op but in the case
2653 * of the scheduler it's an obvious special-case), so we
2654 * do an early lockdep release here:
2656 lockdep_unpin_lock(&rq->lock);
2657 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2659 /* Here we just switch the register state and the stack. */
2660 switch_to(prev, next, prev);
2663 return finish_task_switch(prev);
2667 * nr_running and nr_context_switches:
2669 * externally visible scheduler statistics: current number of runnable
2670 * threads, total number of context switches performed since bootup.
2672 unsigned long nr_running(void)
2674 unsigned long i, sum = 0;
2676 for_each_online_cpu(i)
2677 sum += cpu_rq(i)->nr_running;
2683 * Check if only the current task is running on the cpu.
2685 * Caution: this function does not check that the caller has disabled
2686 * preemption, thus the result might have a time-of-check-to-time-of-use
2687 * race. The caller is responsible to use it correctly, for example:
2689 * - from a non-preemptable section (of course)
2691 * - from a thread that is bound to a single CPU
2693 * - in a loop with very short iterations (e.g. a polling loop)
2695 bool single_task_running(void)
2697 return raw_rq()->nr_running == 1;
2699 EXPORT_SYMBOL(single_task_running);
2701 unsigned long long nr_context_switches(void)
2704 unsigned long long sum = 0;
2706 for_each_possible_cpu(i)
2707 sum += cpu_rq(i)->nr_switches;
2712 unsigned long nr_iowait(void)
2714 unsigned long i, sum = 0;
2716 for_each_possible_cpu(i)
2717 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2722 unsigned long nr_iowait_cpu(int cpu)
2724 struct rq *this = cpu_rq(cpu);
2725 return atomic_read(&this->nr_iowait);
2728 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2730 struct rq *rq = this_rq();
2731 *nr_waiters = atomic_read(&rq->nr_iowait);
2732 *load = rq->load.weight;
2738 * sched_exec - execve() is a valuable balancing opportunity, because at
2739 * this point the task has the smallest effective memory and cache footprint.
2741 void sched_exec(void)
2743 struct task_struct *p = current;
2744 unsigned long flags;
2747 raw_spin_lock_irqsave(&p->pi_lock, flags);
2748 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2749 if (dest_cpu == smp_processor_id())
2752 if (likely(cpu_active(dest_cpu))) {
2753 struct migration_arg arg = { p, dest_cpu };
2755 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2756 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2760 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2765 DEFINE_PER_CPU(struct kernel_stat, kstat);
2766 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2768 EXPORT_PER_CPU_SYMBOL(kstat);
2769 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2772 * Return accounted runtime for the task.
2773 * In case the task is currently running, return the runtime plus current's
2774 * pending runtime that have not been accounted yet.
2776 unsigned long long task_sched_runtime(struct task_struct *p)
2778 unsigned long flags;
2782 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2784 * 64-bit doesn't need locks to atomically read a 64bit value.
2785 * So we have a optimization chance when the task's delta_exec is 0.
2786 * Reading ->on_cpu is racy, but this is ok.
2788 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2789 * If we race with it entering cpu, unaccounted time is 0. This is
2790 * indistinguishable from the read occurring a few cycles earlier.
2791 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2792 * been accounted, so we're correct here as well.
2794 if (!p->on_cpu || !task_on_rq_queued(p))
2795 return p->se.sum_exec_runtime;
2798 rq = task_rq_lock(p, &flags);
2800 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2801 * project cycles that may never be accounted to this
2802 * thread, breaking clock_gettime().
2804 if (task_current(rq, p) && task_on_rq_queued(p)) {
2805 update_rq_clock(rq);
2806 p->sched_class->update_curr(rq);
2808 ns = p->se.sum_exec_runtime;
2809 task_rq_unlock(rq, p, &flags);
2815 * This function gets called by the timer code, with HZ frequency.
2816 * We call it with interrupts disabled.
2818 void scheduler_tick(void)
2820 int cpu = smp_processor_id();
2821 struct rq *rq = cpu_rq(cpu);
2822 struct task_struct *curr = rq->curr;
2826 raw_spin_lock(&rq->lock);
2827 update_rq_clock(rq);
2828 curr->sched_class->task_tick(rq, curr, 0);
2829 update_cpu_load_active(rq);
2830 calc_global_load_tick(rq);
2831 raw_spin_unlock(&rq->lock);
2833 perf_event_task_tick();
2836 rq->idle_balance = idle_cpu(cpu);
2837 trigger_load_balance(rq);
2839 rq_last_tick_reset(rq);
2842 #ifdef CONFIG_NO_HZ_FULL
2844 * scheduler_tick_max_deferment
2846 * Keep at least one tick per second when a single
2847 * active task is running because the scheduler doesn't
2848 * yet completely support full dynticks environment.
2850 * This makes sure that uptime, CFS vruntime, load
2851 * balancing, etc... continue to move forward, even
2852 * with a very low granularity.
2854 * Return: Maximum deferment in nanoseconds.
2856 u64 scheduler_tick_max_deferment(void)
2858 struct rq *rq = this_rq();
2859 unsigned long next, now = READ_ONCE(jiffies);
2861 next = rq->last_sched_tick + HZ;
2863 if (time_before_eq(next, now))
2866 return jiffies_to_nsecs(next - now);
2870 notrace unsigned long get_parent_ip(unsigned long addr)
2872 if (in_lock_functions(addr)) {
2873 addr = CALLER_ADDR2;
2874 if (in_lock_functions(addr))
2875 addr = CALLER_ADDR3;
2880 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2881 defined(CONFIG_PREEMPT_TRACER))
2883 void preempt_count_add(int val)
2885 #ifdef CONFIG_DEBUG_PREEMPT
2889 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2892 __preempt_count_add(val);
2893 #ifdef CONFIG_DEBUG_PREEMPT
2895 * Spinlock count overflowing soon?
2897 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2900 if (preempt_count() == val) {
2901 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2902 #ifdef CONFIG_DEBUG_PREEMPT
2903 current->preempt_disable_ip = ip;
2905 trace_preempt_off(CALLER_ADDR0, ip);
2908 EXPORT_SYMBOL(preempt_count_add);
2909 NOKPROBE_SYMBOL(preempt_count_add);
2911 void preempt_count_sub(int val)
2913 #ifdef CONFIG_DEBUG_PREEMPT
2917 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2920 * Is the spinlock portion underflowing?
2922 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2923 !(preempt_count() & PREEMPT_MASK)))
2927 if (preempt_count() == val)
2928 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2929 __preempt_count_sub(val);
2931 EXPORT_SYMBOL(preempt_count_sub);
2932 NOKPROBE_SYMBOL(preempt_count_sub);
2937 * Print scheduling while atomic bug:
2939 static noinline void __schedule_bug(struct task_struct *prev)
2941 if (oops_in_progress)
2944 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2945 prev->comm, prev->pid, preempt_count());
2947 debug_show_held_locks(prev);
2949 if (irqs_disabled())
2950 print_irqtrace_events(prev);
2951 #ifdef CONFIG_DEBUG_PREEMPT
2952 if (in_atomic_preempt_off()) {
2953 pr_err("Preemption disabled at:");
2954 print_ip_sym(current->preempt_disable_ip);
2959 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2963 * Various schedule()-time debugging checks and statistics:
2965 static inline void schedule_debug(struct task_struct *prev)
2967 #ifdef CONFIG_SCHED_STACK_END_CHECK
2968 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2971 if (unlikely(in_atomic_preempt_off())) {
2972 __schedule_bug(prev);
2973 preempt_count_set(PREEMPT_DISABLED);
2977 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2979 schedstat_inc(this_rq(), sched_count);
2983 * Pick up the highest-prio task:
2985 static inline struct task_struct *
2986 pick_next_task(struct rq *rq, struct task_struct *prev)
2988 const struct sched_class *class = &fair_sched_class;
2989 struct task_struct *p;
2992 * Optimization: we know that if all tasks are in
2993 * the fair class we can call that function directly:
2995 if (likely(prev->sched_class == class &&
2996 rq->nr_running == rq->cfs.h_nr_running)) {
2997 p = fair_sched_class.pick_next_task(rq, prev);
2998 if (unlikely(p == RETRY_TASK))
3001 /* assumes fair_sched_class->next == idle_sched_class */
3003 p = idle_sched_class.pick_next_task(rq, prev);
3009 for_each_class(class) {
3010 p = class->pick_next_task(rq, prev);
3012 if (unlikely(p == RETRY_TASK))
3018 BUG(); /* the idle class will always have a runnable task */
3022 * __schedule() is the main scheduler function.
3024 * The main means of driving the scheduler and thus entering this function are:
3026 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3028 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3029 * paths. For example, see arch/x86/entry_64.S.
3031 * To drive preemption between tasks, the scheduler sets the flag in timer
3032 * interrupt handler scheduler_tick().
3034 * 3. Wakeups don't really cause entry into schedule(). They add a
3035 * task to the run-queue and that's it.
3037 * Now, if the new task added to the run-queue preempts the current
3038 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3039 * called on the nearest possible occasion:
3041 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3043 * - in syscall or exception context, at the next outmost
3044 * preempt_enable(). (this might be as soon as the wake_up()'s
3047 * - in IRQ context, return from interrupt-handler to
3048 * preemptible context
3050 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3053 * - cond_resched() call
3054 * - explicit schedule() call
3055 * - return from syscall or exception to user-space
3056 * - return from interrupt-handler to user-space
3058 * WARNING: must be called with preemption disabled!
3060 static void __sched __schedule(bool preempt)
3062 struct task_struct *prev, *next;
3063 unsigned long *switch_count;
3067 cpu = smp_processor_id();
3069 rcu_note_context_switch();
3073 * do_exit() calls schedule() with preemption disabled as an exception;
3074 * however we must fix that up, otherwise the next task will see an
3075 * inconsistent (higher) preempt count.
3077 * It also avoids the below schedule_debug() test from complaining
3080 if (unlikely(prev->state == TASK_DEAD))
3081 preempt_enable_no_resched_notrace();
3083 schedule_debug(prev);
3085 if (sched_feat(HRTICK))
3089 * Make sure that signal_pending_state()->signal_pending() below
3090 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3091 * done by the caller to avoid the race with signal_wake_up().
3093 smp_mb__before_spinlock();
3094 raw_spin_lock_irq(&rq->lock);
3095 lockdep_pin_lock(&rq->lock);
3097 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3099 switch_count = &prev->nivcsw;
3100 if (!preempt && prev->state) {
3101 if (unlikely(signal_pending_state(prev->state, prev))) {
3102 prev->state = TASK_RUNNING;
3104 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3108 * If a worker went to sleep, notify and ask workqueue
3109 * whether it wants to wake up a task to maintain
3112 if (prev->flags & PF_WQ_WORKER) {
3113 struct task_struct *to_wakeup;
3115 to_wakeup = wq_worker_sleeping(prev, cpu);
3117 try_to_wake_up_local(to_wakeup);
3120 switch_count = &prev->nvcsw;
3123 if (task_on_rq_queued(prev))
3124 update_rq_clock(rq);
3126 next = pick_next_task(rq, prev);
3127 clear_tsk_need_resched(prev);
3128 clear_preempt_need_resched();
3129 rq->clock_skip_update = 0;
3131 if (likely(prev != next)) {
3136 trace_sched_switch(preempt, prev, next);
3137 rq = context_switch(rq, prev, next); /* unlocks the rq */
3140 lockdep_unpin_lock(&rq->lock);
3141 raw_spin_unlock_irq(&rq->lock);
3144 balance_callback(rq);
3147 static inline void sched_submit_work(struct task_struct *tsk)
3149 if (!tsk->state || tsk_is_pi_blocked(tsk))
3152 * If we are going to sleep and we have plugged IO queued,
3153 * make sure to submit it to avoid deadlocks.
3155 if (blk_needs_flush_plug(tsk))
3156 blk_schedule_flush_plug(tsk);
3159 asmlinkage __visible void __sched schedule(void)
3161 struct task_struct *tsk = current;
3163 sched_submit_work(tsk);
3167 sched_preempt_enable_no_resched();
3168 } while (need_resched());
3170 EXPORT_SYMBOL(schedule);
3172 #ifdef CONFIG_CONTEXT_TRACKING
3173 asmlinkage __visible void __sched schedule_user(void)
3176 * If we come here after a random call to set_need_resched(),
3177 * or we have been woken up remotely but the IPI has not yet arrived,
3178 * we haven't yet exited the RCU idle mode. Do it here manually until
3179 * we find a better solution.
3181 * NB: There are buggy callers of this function. Ideally we
3182 * should warn if prev_state != CONTEXT_USER, but that will trigger
3183 * too frequently to make sense yet.
3185 enum ctx_state prev_state = exception_enter();
3187 exception_exit(prev_state);
3192 * schedule_preempt_disabled - called with preemption disabled
3194 * Returns with preemption disabled. Note: preempt_count must be 1
3196 void __sched schedule_preempt_disabled(void)
3198 sched_preempt_enable_no_resched();
3203 static void __sched notrace preempt_schedule_common(void)
3208 sched_preempt_enable_no_resched();
3211 * Check again in case we missed a preemption opportunity
3212 * between schedule and now.
3214 } while (need_resched());
3217 #ifdef CONFIG_PREEMPT
3219 * this is the entry point to schedule() from in-kernel preemption
3220 * off of preempt_enable. Kernel preemptions off return from interrupt
3221 * occur there and call schedule directly.
3223 asmlinkage __visible void __sched notrace preempt_schedule(void)
3226 * If there is a non-zero preempt_count or interrupts are disabled,
3227 * we do not want to preempt the current task. Just return..
3229 if (likely(!preemptible()))
3232 preempt_schedule_common();
3234 NOKPROBE_SYMBOL(preempt_schedule);
3235 EXPORT_SYMBOL(preempt_schedule);
3238 * preempt_schedule_notrace - preempt_schedule called by tracing
3240 * The tracing infrastructure uses preempt_enable_notrace to prevent
3241 * recursion and tracing preempt enabling caused by the tracing
3242 * infrastructure itself. But as tracing can happen in areas coming
3243 * from userspace or just about to enter userspace, a preempt enable
3244 * can occur before user_exit() is called. This will cause the scheduler
3245 * to be called when the system is still in usermode.
3247 * To prevent this, the preempt_enable_notrace will use this function
3248 * instead of preempt_schedule() to exit user context if needed before
3249 * calling the scheduler.
3251 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3253 enum ctx_state prev_ctx;
3255 if (likely(!preemptible()))
3259 preempt_disable_notrace();
3261 * Needs preempt disabled in case user_exit() is traced
3262 * and the tracer calls preempt_enable_notrace() causing
3263 * an infinite recursion.
3265 prev_ctx = exception_enter();
3267 exception_exit(prev_ctx);
3269 preempt_enable_no_resched_notrace();
3270 } while (need_resched());
3272 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3274 #endif /* CONFIG_PREEMPT */
3277 * this is the entry point to schedule() from kernel preemption
3278 * off of irq context.
3279 * Note, that this is called and return with irqs disabled. This will
3280 * protect us against recursive calling from irq.
3282 asmlinkage __visible void __sched preempt_schedule_irq(void)
3284 enum ctx_state prev_state;
3286 /* Catch callers which need to be fixed */
3287 BUG_ON(preempt_count() || !irqs_disabled());
3289 prev_state = exception_enter();
3295 local_irq_disable();
3296 sched_preempt_enable_no_resched();
3297 } while (need_resched());
3299 exception_exit(prev_state);
3302 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3305 return try_to_wake_up(curr->private, mode, wake_flags);
3307 EXPORT_SYMBOL(default_wake_function);
3309 #ifdef CONFIG_RT_MUTEXES
3312 * rt_mutex_setprio - set the current priority of a task
3314 * @prio: prio value (kernel-internal form)
3316 * This function changes the 'effective' priority of a task. It does
3317 * not touch ->normal_prio like __setscheduler().
3319 * Used by the rt_mutex code to implement priority inheritance
3320 * logic. Call site only calls if the priority of the task changed.
3322 void rt_mutex_setprio(struct task_struct *p, int prio)
3324 int oldprio, queued, running, enqueue_flag = 0;
3326 const struct sched_class *prev_class;
3328 BUG_ON(prio > MAX_PRIO);
3330 rq = __task_rq_lock(p);
3333 * Idle task boosting is a nono in general. There is one
3334 * exception, when PREEMPT_RT and NOHZ is active:
3336 * The idle task calls get_next_timer_interrupt() and holds
3337 * the timer wheel base->lock on the CPU and another CPU wants
3338 * to access the timer (probably to cancel it). We can safely
3339 * ignore the boosting request, as the idle CPU runs this code
3340 * with interrupts disabled and will complete the lock
3341 * protected section without being interrupted. So there is no
3342 * real need to boost.
3344 if (unlikely(p == rq->idle)) {
3345 WARN_ON(p != rq->curr);
3346 WARN_ON(p->pi_blocked_on);
3350 trace_sched_pi_setprio(p, prio);
3352 prev_class = p->sched_class;
3353 queued = task_on_rq_queued(p);
3354 running = task_current(rq, p);
3356 dequeue_task(rq, p, 0);
3358 put_prev_task(rq, p);
3361 * Boosting condition are:
3362 * 1. -rt task is running and holds mutex A
3363 * --> -dl task blocks on mutex A
3365 * 2. -dl task is running and holds mutex A
3366 * --> -dl task blocks on mutex A and could preempt the
3369 if (dl_prio(prio)) {
3370 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3371 if (!dl_prio(p->normal_prio) ||
3372 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3373 p->dl.dl_boosted = 1;
3374 enqueue_flag = ENQUEUE_REPLENISH;
3376 p->dl.dl_boosted = 0;
3377 p->sched_class = &dl_sched_class;
3378 } else if (rt_prio(prio)) {
3379 if (dl_prio(oldprio))
3380 p->dl.dl_boosted = 0;
3382 enqueue_flag = ENQUEUE_HEAD;
3383 p->sched_class = &rt_sched_class;
3385 if (dl_prio(oldprio))
3386 p->dl.dl_boosted = 0;
3387 if (rt_prio(oldprio))
3389 p->sched_class = &fair_sched_class;
3395 p->sched_class->set_curr_task(rq);
3397 enqueue_task(rq, p, enqueue_flag);
3399 check_class_changed(rq, p, prev_class, oldprio);
3401 preempt_disable(); /* avoid rq from going away on us */
3402 __task_rq_unlock(rq);
3404 balance_callback(rq);
3409 void set_user_nice(struct task_struct *p, long nice)
3411 int old_prio, delta, queued;
3412 unsigned long flags;
3415 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3418 * We have to be careful, if called from sys_setpriority(),
3419 * the task might be in the middle of scheduling on another CPU.
3421 rq = task_rq_lock(p, &flags);
3423 * The RT priorities are set via sched_setscheduler(), but we still
3424 * allow the 'normal' nice value to be set - but as expected
3425 * it wont have any effect on scheduling until the task is
3426 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3428 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3429 p->static_prio = NICE_TO_PRIO(nice);
3432 queued = task_on_rq_queued(p);
3434 dequeue_task(rq, p, 0);
3436 p->static_prio = NICE_TO_PRIO(nice);
3439 p->prio = effective_prio(p);
3440 delta = p->prio - old_prio;
3443 enqueue_task(rq, p, 0);
3445 * If the task increased its priority or is running and
3446 * lowered its priority, then reschedule its CPU:
3448 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3452 task_rq_unlock(rq, p, &flags);
3454 EXPORT_SYMBOL(set_user_nice);
3457 * can_nice - check if a task can reduce its nice value
3461 int can_nice(const struct task_struct *p, const int nice)
3463 /* convert nice value [19,-20] to rlimit style value [1,40] */
3464 int nice_rlim = nice_to_rlimit(nice);
3466 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3467 capable(CAP_SYS_NICE));
3470 #ifdef __ARCH_WANT_SYS_NICE
3473 * sys_nice - change the priority of the current process.
3474 * @increment: priority increment
3476 * sys_setpriority is a more generic, but much slower function that
3477 * does similar things.
3479 SYSCALL_DEFINE1(nice, int, increment)
3484 * Setpriority might change our priority at the same moment.
3485 * We don't have to worry. Conceptually one call occurs first
3486 * and we have a single winner.
3488 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3489 nice = task_nice(current) + increment;
3491 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3492 if (increment < 0 && !can_nice(current, nice))
3495 retval = security_task_setnice(current, nice);
3499 set_user_nice(current, nice);
3506 * task_prio - return the priority value of a given task.
3507 * @p: the task in question.
3509 * Return: The priority value as seen by users in /proc.
3510 * RT tasks are offset by -200. Normal tasks are centered
3511 * around 0, value goes from -16 to +15.
3513 int task_prio(const struct task_struct *p)
3515 return p->prio - MAX_RT_PRIO;
3519 * idle_cpu - is a given cpu idle currently?
3520 * @cpu: the processor in question.
3522 * Return: 1 if the CPU is currently idle. 0 otherwise.
3524 int idle_cpu(int cpu)
3526 struct rq *rq = cpu_rq(cpu);
3528 if (rq->curr != rq->idle)
3535 if (!llist_empty(&rq->wake_list))
3543 * idle_task - return the idle task for a given cpu.
3544 * @cpu: the processor in question.
3546 * Return: The idle task for the cpu @cpu.
3548 struct task_struct *idle_task(int cpu)
3550 return cpu_rq(cpu)->idle;
3554 * find_process_by_pid - find a process with a matching PID value.
3555 * @pid: the pid in question.
3557 * The task of @pid, if found. %NULL otherwise.
3559 static struct task_struct *find_process_by_pid(pid_t pid)
3561 return pid ? find_task_by_vpid(pid) : current;
3565 * This function initializes the sched_dl_entity of a newly becoming
3566 * SCHED_DEADLINE task.
3568 * Only the static values are considered here, the actual runtime and the
3569 * absolute deadline will be properly calculated when the task is enqueued
3570 * for the first time with its new policy.
3573 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3575 struct sched_dl_entity *dl_se = &p->dl;
3577 dl_se->dl_runtime = attr->sched_runtime;
3578 dl_se->dl_deadline = attr->sched_deadline;
3579 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3580 dl_se->flags = attr->sched_flags;
3581 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3584 * Changing the parameters of a task is 'tricky' and we're not doing
3585 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3587 * What we SHOULD do is delay the bandwidth release until the 0-lag
3588 * point. This would include retaining the task_struct until that time
3589 * and change dl_overflow() to not immediately decrement the current
3592 * Instead we retain the current runtime/deadline and let the new
3593 * parameters take effect after the current reservation period lapses.
3594 * This is safe (albeit pessimistic) because the 0-lag point is always
3595 * before the current scheduling deadline.
3597 * We can still have temporary overloads because we do not delay the
3598 * change in bandwidth until that time; so admission control is
3599 * not on the safe side. It does however guarantee tasks will never
3600 * consume more than promised.
3605 * sched_setparam() passes in -1 for its policy, to let the functions
3606 * it calls know not to change it.
3608 #define SETPARAM_POLICY -1
3610 static void __setscheduler_params(struct task_struct *p,
3611 const struct sched_attr *attr)
3613 int policy = attr->sched_policy;
3615 if (policy == SETPARAM_POLICY)
3620 if (dl_policy(policy))
3621 __setparam_dl(p, attr);
3622 else if (fair_policy(policy))
3623 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3626 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3627 * !rt_policy. Always setting this ensures that things like
3628 * getparam()/getattr() don't report silly values for !rt tasks.
3630 p->rt_priority = attr->sched_priority;
3631 p->normal_prio = normal_prio(p);
3635 /* Actually do priority change: must hold pi & rq lock. */
3636 static void __setscheduler(struct rq *rq, struct task_struct *p,
3637 const struct sched_attr *attr, bool keep_boost)
3639 __setscheduler_params(p, attr);
3642 * Keep a potential priority boosting if called from
3643 * sched_setscheduler().
3646 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3648 p->prio = normal_prio(p);
3650 if (dl_prio(p->prio))
3651 p->sched_class = &dl_sched_class;
3652 else if (rt_prio(p->prio))
3653 p->sched_class = &rt_sched_class;
3655 p->sched_class = &fair_sched_class;
3659 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3661 struct sched_dl_entity *dl_se = &p->dl;
3663 attr->sched_priority = p->rt_priority;
3664 attr->sched_runtime = dl_se->dl_runtime;
3665 attr->sched_deadline = dl_se->dl_deadline;
3666 attr->sched_period = dl_se->dl_period;
3667 attr->sched_flags = dl_se->flags;
3671 * This function validates the new parameters of a -deadline task.
3672 * We ask for the deadline not being zero, and greater or equal
3673 * than the runtime, as well as the period of being zero or
3674 * greater than deadline. Furthermore, we have to be sure that
3675 * user parameters are above the internal resolution of 1us (we
3676 * check sched_runtime only since it is always the smaller one) and
3677 * below 2^63 ns (we have to check both sched_deadline and
3678 * sched_period, as the latter can be zero).
3681 __checkparam_dl(const struct sched_attr *attr)
3684 if (attr->sched_deadline == 0)
3688 * Since we truncate DL_SCALE bits, make sure we're at least
3691 if (attr->sched_runtime < (1ULL << DL_SCALE))
3695 * Since we use the MSB for wrap-around and sign issues, make
3696 * sure it's not set (mind that period can be equal to zero).
3698 if (attr->sched_deadline & (1ULL << 63) ||
3699 attr->sched_period & (1ULL << 63))
3702 /* runtime <= deadline <= period (if period != 0) */
3703 if ((attr->sched_period != 0 &&
3704 attr->sched_period < attr->sched_deadline) ||
3705 attr->sched_deadline < attr->sched_runtime)
3712 * check the target process has a UID that matches the current process's
3714 static bool check_same_owner(struct task_struct *p)
3716 const struct cred *cred = current_cred(), *pcred;
3720 pcred = __task_cred(p);
3721 match = (uid_eq(cred->euid, pcred->euid) ||
3722 uid_eq(cred->euid, pcred->uid));
3727 static bool dl_param_changed(struct task_struct *p,
3728 const struct sched_attr *attr)
3730 struct sched_dl_entity *dl_se = &p->dl;
3732 if (dl_se->dl_runtime != attr->sched_runtime ||
3733 dl_se->dl_deadline != attr->sched_deadline ||
3734 dl_se->dl_period != attr->sched_period ||
3735 dl_se->flags != attr->sched_flags)
3741 static int __sched_setscheduler(struct task_struct *p,
3742 const struct sched_attr *attr,
3745 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3746 MAX_RT_PRIO - 1 - attr->sched_priority;
3747 int retval, oldprio, oldpolicy = -1, queued, running;
3748 int new_effective_prio, policy = attr->sched_policy;
3749 unsigned long flags;
3750 const struct sched_class *prev_class;
3754 /* may grab non-irq protected spin_locks */
3755 BUG_ON(in_interrupt());
3757 /* double check policy once rq lock held */
3759 reset_on_fork = p->sched_reset_on_fork;
3760 policy = oldpolicy = p->policy;
3762 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3764 if (!valid_policy(policy))
3768 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3772 * Valid priorities for SCHED_FIFO and SCHED_RR are
3773 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3774 * SCHED_BATCH and SCHED_IDLE is 0.
3776 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3777 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3779 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3780 (rt_policy(policy) != (attr->sched_priority != 0)))
3784 * Allow unprivileged RT tasks to decrease priority:
3786 if (user && !capable(CAP_SYS_NICE)) {
3787 if (fair_policy(policy)) {
3788 if (attr->sched_nice < task_nice(p) &&
3789 !can_nice(p, attr->sched_nice))
3793 if (rt_policy(policy)) {
3794 unsigned long rlim_rtprio =
3795 task_rlimit(p, RLIMIT_RTPRIO);
3797 /* can't set/change the rt policy */
3798 if (policy != p->policy && !rlim_rtprio)
3801 /* can't increase priority */
3802 if (attr->sched_priority > p->rt_priority &&
3803 attr->sched_priority > rlim_rtprio)
3808 * Can't set/change SCHED_DEADLINE policy at all for now
3809 * (safest behavior); in the future we would like to allow
3810 * unprivileged DL tasks to increase their relative deadline
3811 * or reduce their runtime (both ways reducing utilization)
3813 if (dl_policy(policy))
3817 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3818 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3820 if (idle_policy(p->policy) && !idle_policy(policy)) {
3821 if (!can_nice(p, task_nice(p)))
3825 /* can't change other user's priorities */
3826 if (!check_same_owner(p))
3829 /* Normal users shall not reset the sched_reset_on_fork flag */
3830 if (p->sched_reset_on_fork && !reset_on_fork)
3835 retval = security_task_setscheduler(p);
3841 * make sure no PI-waiters arrive (or leave) while we are
3842 * changing the priority of the task:
3844 * To be able to change p->policy safely, the appropriate
3845 * runqueue lock must be held.
3847 rq = task_rq_lock(p, &flags);
3850 * Changing the policy of the stop threads its a very bad idea
3852 if (p == rq->stop) {
3853 task_rq_unlock(rq, p, &flags);
3858 * If not changing anything there's no need to proceed further,
3859 * but store a possible modification of reset_on_fork.
3861 if (unlikely(policy == p->policy)) {
3862 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3864 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3866 if (dl_policy(policy) && dl_param_changed(p, attr))
3869 p->sched_reset_on_fork = reset_on_fork;
3870 task_rq_unlock(rq, p, &flags);
3876 #ifdef CONFIG_RT_GROUP_SCHED
3878 * Do not allow realtime tasks into groups that have no runtime
3881 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3882 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3883 !task_group_is_autogroup(task_group(p))) {
3884 task_rq_unlock(rq, p, &flags);
3889 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3890 cpumask_t *span = rq->rd->span;
3893 * Don't allow tasks with an affinity mask smaller than
3894 * the entire root_domain to become SCHED_DEADLINE. We
3895 * will also fail if there's no bandwidth available.
3897 if (!cpumask_subset(span, &p->cpus_allowed) ||
3898 rq->rd->dl_bw.bw == 0) {
3899 task_rq_unlock(rq, p, &flags);
3906 /* recheck policy now with rq lock held */
3907 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3908 policy = oldpolicy = -1;
3909 task_rq_unlock(rq, p, &flags);
3914 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3915 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3918 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3919 task_rq_unlock(rq, p, &flags);
3923 p->sched_reset_on_fork = reset_on_fork;
3928 * Take priority boosted tasks into account. If the new
3929 * effective priority is unchanged, we just store the new
3930 * normal parameters and do not touch the scheduler class and
3931 * the runqueue. This will be done when the task deboost
3934 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3935 if (new_effective_prio == oldprio) {
3936 __setscheduler_params(p, attr);
3937 task_rq_unlock(rq, p, &flags);
3942 queued = task_on_rq_queued(p);
3943 running = task_current(rq, p);
3945 dequeue_task(rq, p, 0);
3947 put_prev_task(rq, p);
3949 prev_class = p->sched_class;
3950 __setscheduler(rq, p, attr, pi);
3953 p->sched_class->set_curr_task(rq);
3956 * We enqueue to tail when the priority of a task is
3957 * increased (user space view).
3959 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3962 check_class_changed(rq, p, prev_class, oldprio);
3963 preempt_disable(); /* avoid rq from going away on us */
3964 task_rq_unlock(rq, p, &flags);
3967 rt_mutex_adjust_pi(p);
3970 * Run balance callbacks after we've adjusted the PI chain.
3972 balance_callback(rq);
3978 static int _sched_setscheduler(struct task_struct *p, int policy,
3979 const struct sched_param *param, bool check)
3981 struct sched_attr attr = {
3982 .sched_policy = policy,
3983 .sched_priority = param->sched_priority,
3984 .sched_nice = PRIO_TO_NICE(p->static_prio),
3987 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3988 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3989 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3990 policy &= ~SCHED_RESET_ON_FORK;
3991 attr.sched_policy = policy;
3994 return __sched_setscheduler(p, &attr, check, true);
3997 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3998 * @p: the task in question.
3999 * @policy: new policy.
4000 * @param: structure containing the new RT priority.
4002 * Return: 0 on success. An error code otherwise.
4004 * NOTE that the task may be already dead.
4006 int sched_setscheduler(struct task_struct *p, int policy,
4007 const struct sched_param *param)
4009 return _sched_setscheduler(p, policy, param, true);
4011 EXPORT_SYMBOL_GPL(sched_setscheduler);
4013 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4015 return __sched_setscheduler(p, attr, true, true);
4017 EXPORT_SYMBOL_GPL(sched_setattr);
4020 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4021 * @p: the task in question.
4022 * @policy: new policy.
4023 * @param: structure containing the new RT priority.
4025 * Just like sched_setscheduler, only don't bother checking if the
4026 * current context has permission. For example, this is needed in
4027 * stop_machine(): we create temporary high priority worker threads,
4028 * but our caller might not have that capability.
4030 * Return: 0 on success. An error code otherwise.
4032 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4033 const struct sched_param *param)
4035 return _sched_setscheduler(p, policy, param, false);
4039 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4041 struct sched_param lparam;
4042 struct task_struct *p;
4045 if (!param || pid < 0)
4047 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4052 p = find_process_by_pid(pid);
4054 retval = sched_setscheduler(p, policy, &lparam);
4061 * Mimics kernel/events/core.c perf_copy_attr().
4063 static int sched_copy_attr(struct sched_attr __user *uattr,
4064 struct sched_attr *attr)
4069 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4073 * zero the full structure, so that a short copy will be nice.
4075 memset(attr, 0, sizeof(*attr));
4077 ret = get_user(size, &uattr->size);
4081 if (size > PAGE_SIZE) /* silly large */
4084 if (!size) /* abi compat */
4085 size = SCHED_ATTR_SIZE_VER0;
4087 if (size < SCHED_ATTR_SIZE_VER0)
4091 * If we're handed a bigger struct than we know of,
4092 * ensure all the unknown bits are 0 - i.e. new
4093 * user-space does not rely on any kernel feature
4094 * extensions we dont know about yet.
4096 if (size > sizeof(*attr)) {
4097 unsigned char __user *addr;
4098 unsigned char __user *end;
4101 addr = (void __user *)uattr + sizeof(*attr);
4102 end = (void __user *)uattr + size;
4104 for (; addr < end; addr++) {
4105 ret = get_user(val, addr);
4111 size = sizeof(*attr);
4114 ret = copy_from_user(attr, uattr, size);
4119 * XXX: do we want to be lenient like existing syscalls; or do we want
4120 * to be strict and return an error on out-of-bounds values?
4122 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4127 put_user(sizeof(*attr), &uattr->size);
4132 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4133 * @pid: the pid in question.
4134 * @policy: new policy.
4135 * @param: structure containing the new RT priority.
4137 * Return: 0 on success. An error code otherwise.
4139 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4140 struct sched_param __user *, param)
4142 /* negative values for policy are not valid */
4146 return do_sched_setscheduler(pid, policy, param);
4150 * sys_sched_setparam - set/change the RT priority of a thread
4151 * @pid: the pid in question.
4152 * @param: structure containing the new RT priority.
4154 * Return: 0 on success. An error code otherwise.
4156 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4158 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4162 * sys_sched_setattr - same as above, but with extended sched_attr
4163 * @pid: the pid in question.
4164 * @uattr: structure containing the extended parameters.
4165 * @flags: for future extension.
4167 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4168 unsigned int, flags)
4170 struct sched_attr attr;
4171 struct task_struct *p;
4174 if (!uattr || pid < 0 || flags)
4177 retval = sched_copy_attr(uattr, &attr);
4181 if ((int)attr.sched_policy < 0)
4186 p = find_process_by_pid(pid);
4188 retval = sched_setattr(p, &attr);
4195 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4196 * @pid: the pid in question.
4198 * Return: On success, the policy of the thread. Otherwise, a negative error
4201 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4203 struct task_struct *p;
4211 p = find_process_by_pid(pid);
4213 retval = security_task_getscheduler(p);
4216 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4223 * sys_sched_getparam - get the RT priority of a thread
4224 * @pid: the pid in question.
4225 * @param: structure containing the RT priority.
4227 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4230 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4232 struct sched_param lp = { .sched_priority = 0 };
4233 struct task_struct *p;
4236 if (!param || pid < 0)
4240 p = find_process_by_pid(pid);
4245 retval = security_task_getscheduler(p);
4249 if (task_has_rt_policy(p))
4250 lp.sched_priority = p->rt_priority;
4254 * This one might sleep, we cannot do it with a spinlock held ...
4256 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4265 static int sched_read_attr(struct sched_attr __user *uattr,
4266 struct sched_attr *attr,
4271 if (!access_ok(VERIFY_WRITE, uattr, usize))
4275 * If we're handed a smaller struct than we know of,
4276 * ensure all the unknown bits are 0 - i.e. old
4277 * user-space does not get uncomplete information.
4279 if (usize < sizeof(*attr)) {
4280 unsigned char *addr;
4283 addr = (void *)attr + usize;
4284 end = (void *)attr + sizeof(*attr);
4286 for (; addr < end; addr++) {
4294 ret = copy_to_user(uattr, attr, attr->size);
4302 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4303 * @pid: the pid in question.
4304 * @uattr: structure containing the extended parameters.
4305 * @size: sizeof(attr) for fwd/bwd comp.
4306 * @flags: for future extension.
4308 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4309 unsigned int, size, unsigned int, flags)
4311 struct sched_attr attr = {
4312 .size = sizeof(struct sched_attr),
4314 struct task_struct *p;
4317 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4318 size < SCHED_ATTR_SIZE_VER0 || flags)
4322 p = find_process_by_pid(pid);
4327 retval = security_task_getscheduler(p);
4331 attr.sched_policy = p->policy;
4332 if (p->sched_reset_on_fork)
4333 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4334 if (task_has_dl_policy(p))
4335 __getparam_dl(p, &attr);
4336 else if (task_has_rt_policy(p))
4337 attr.sched_priority = p->rt_priority;
4339 attr.sched_nice = task_nice(p);
4343 retval = sched_read_attr(uattr, &attr, size);
4351 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4353 cpumask_var_t cpus_allowed, new_mask;
4354 struct task_struct *p;
4359 p = find_process_by_pid(pid);
4365 /* Prevent p going away */
4369 if (p->flags & PF_NO_SETAFFINITY) {
4373 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4377 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4379 goto out_free_cpus_allowed;
4382 if (!check_same_owner(p)) {
4384 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4386 goto out_free_new_mask;
4391 retval = security_task_setscheduler(p);
4393 goto out_free_new_mask;
4396 cpuset_cpus_allowed(p, cpus_allowed);
4397 cpumask_and(new_mask, in_mask, cpus_allowed);
4400 * Since bandwidth control happens on root_domain basis,
4401 * if admission test is enabled, we only admit -deadline
4402 * tasks allowed to run on all the CPUs in the task's
4406 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4408 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4411 goto out_free_new_mask;
4417 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4420 cpuset_cpus_allowed(p, cpus_allowed);
4421 if (!cpumask_subset(new_mask, cpus_allowed)) {
4423 * We must have raced with a concurrent cpuset
4424 * update. Just reset the cpus_allowed to the
4425 * cpuset's cpus_allowed
4427 cpumask_copy(new_mask, cpus_allowed);
4432 free_cpumask_var(new_mask);
4433 out_free_cpus_allowed:
4434 free_cpumask_var(cpus_allowed);
4440 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4441 struct cpumask *new_mask)
4443 if (len < cpumask_size())
4444 cpumask_clear(new_mask);
4445 else if (len > cpumask_size())
4446 len = cpumask_size();
4448 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4452 * sys_sched_setaffinity - set the cpu affinity of a process
4453 * @pid: pid of the process
4454 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4455 * @user_mask_ptr: user-space pointer to the new cpu mask
4457 * Return: 0 on success. An error code otherwise.
4459 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4460 unsigned long __user *, user_mask_ptr)
4462 cpumask_var_t new_mask;
4465 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4468 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4470 retval = sched_setaffinity(pid, new_mask);
4471 free_cpumask_var(new_mask);
4475 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4477 struct task_struct *p;
4478 unsigned long flags;
4484 p = find_process_by_pid(pid);
4488 retval = security_task_getscheduler(p);
4492 raw_spin_lock_irqsave(&p->pi_lock, flags);
4493 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4494 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4503 * sys_sched_getaffinity - get the cpu affinity of a process
4504 * @pid: pid of the process
4505 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4506 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4508 * Return: 0 on success. An error code otherwise.
4510 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4511 unsigned long __user *, user_mask_ptr)
4516 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4518 if (len & (sizeof(unsigned long)-1))
4521 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4524 ret = sched_getaffinity(pid, mask);
4526 size_t retlen = min_t(size_t, len, cpumask_size());
4528 if (copy_to_user(user_mask_ptr, mask, retlen))
4533 free_cpumask_var(mask);
4539 * sys_sched_yield - yield the current processor to other threads.
4541 * This function yields the current CPU to other tasks. If there are no
4542 * other threads running on this CPU then this function will return.
4546 SYSCALL_DEFINE0(sched_yield)
4548 struct rq *rq = this_rq_lock();
4550 schedstat_inc(rq, yld_count);
4551 current->sched_class->yield_task(rq);
4554 * Since we are going to call schedule() anyway, there's
4555 * no need to preempt or enable interrupts:
4557 __release(rq->lock);
4558 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4559 do_raw_spin_unlock(&rq->lock);
4560 sched_preempt_enable_no_resched();
4567 int __sched _cond_resched(void)
4569 if (should_resched(0)) {
4570 preempt_schedule_common();
4575 EXPORT_SYMBOL(_cond_resched);
4578 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4579 * call schedule, and on return reacquire the lock.
4581 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4582 * operations here to prevent schedule() from being called twice (once via
4583 * spin_unlock(), once by hand).
4585 int __cond_resched_lock(spinlock_t *lock)
4587 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4590 lockdep_assert_held(lock);
4592 if (spin_needbreak(lock) || resched) {
4595 preempt_schedule_common();
4603 EXPORT_SYMBOL(__cond_resched_lock);
4605 int __sched __cond_resched_softirq(void)
4607 BUG_ON(!in_softirq());
4609 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4611 preempt_schedule_common();
4617 EXPORT_SYMBOL(__cond_resched_softirq);
4620 * yield - yield the current processor to other threads.
4622 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4624 * The scheduler is at all times free to pick the calling task as the most
4625 * eligible task to run, if removing the yield() call from your code breaks
4626 * it, its already broken.
4628 * Typical broken usage is:
4633 * where one assumes that yield() will let 'the other' process run that will
4634 * make event true. If the current task is a SCHED_FIFO task that will never
4635 * happen. Never use yield() as a progress guarantee!!
4637 * If you want to use yield() to wait for something, use wait_event().
4638 * If you want to use yield() to be 'nice' for others, use cond_resched().
4639 * If you still want to use yield(), do not!
4641 void __sched yield(void)
4643 set_current_state(TASK_RUNNING);
4646 EXPORT_SYMBOL(yield);
4649 * yield_to - yield the current processor to another thread in
4650 * your thread group, or accelerate that thread toward the
4651 * processor it's on.
4653 * @preempt: whether task preemption is allowed or not
4655 * It's the caller's job to ensure that the target task struct
4656 * can't go away on us before we can do any checks.
4659 * true (>0) if we indeed boosted the target task.
4660 * false (0) if we failed to boost the target.
4661 * -ESRCH if there's no task to yield to.
4663 int __sched yield_to(struct task_struct *p, bool preempt)
4665 struct task_struct *curr = current;
4666 struct rq *rq, *p_rq;
4667 unsigned long flags;
4670 local_irq_save(flags);
4676 * If we're the only runnable task on the rq and target rq also
4677 * has only one task, there's absolutely no point in yielding.
4679 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4684 double_rq_lock(rq, p_rq);
4685 if (task_rq(p) != p_rq) {
4686 double_rq_unlock(rq, p_rq);
4690 if (!curr->sched_class->yield_to_task)
4693 if (curr->sched_class != p->sched_class)
4696 if (task_running(p_rq, p) || p->state)
4699 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4701 schedstat_inc(rq, yld_count);
4703 * Make p's CPU reschedule; pick_next_entity takes care of
4706 if (preempt && rq != p_rq)
4711 double_rq_unlock(rq, p_rq);
4713 local_irq_restore(flags);
4720 EXPORT_SYMBOL_GPL(yield_to);
4723 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4724 * that process accounting knows that this is a task in IO wait state.
4726 long __sched io_schedule_timeout(long timeout)
4728 int old_iowait = current->in_iowait;
4732 current->in_iowait = 1;
4733 blk_schedule_flush_plug(current);
4735 delayacct_blkio_start();
4737 atomic_inc(&rq->nr_iowait);
4738 ret = schedule_timeout(timeout);
4739 current->in_iowait = old_iowait;
4740 atomic_dec(&rq->nr_iowait);
4741 delayacct_blkio_end();
4745 EXPORT_SYMBOL(io_schedule_timeout);
4748 * sys_sched_get_priority_max - return maximum RT priority.
4749 * @policy: scheduling class.
4751 * Return: On success, this syscall returns the maximum
4752 * rt_priority that can be used by a given scheduling class.
4753 * On failure, a negative error code is returned.
4755 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4762 ret = MAX_USER_RT_PRIO-1;
4764 case SCHED_DEADLINE:
4775 * sys_sched_get_priority_min - return minimum RT priority.
4776 * @policy: scheduling class.
4778 * Return: On success, this syscall returns the minimum
4779 * rt_priority that can be used by a given scheduling class.
4780 * On failure, a negative error code is returned.
4782 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4791 case SCHED_DEADLINE:
4801 * sys_sched_rr_get_interval - return the default timeslice of a process.
4802 * @pid: pid of the process.
4803 * @interval: userspace pointer to the timeslice value.
4805 * this syscall writes the default timeslice value of a given process
4806 * into the user-space timespec buffer. A value of '0' means infinity.
4808 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4811 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4812 struct timespec __user *, interval)
4814 struct task_struct *p;
4815 unsigned int time_slice;
4816 unsigned long flags;
4826 p = find_process_by_pid(pid);
4830 retval = security_task_getscheduler(p);
4834 rq = task_rq_lock(p, &flags);
4836 if (p->sched_class->get_rr_interval)
4837 time_slice = p->sched_class->get_rr_interval(rq, p);
4838 task_rq_unlock(rq, p, &flags);
4841 jiffies_to_timespec(time_slice, &t);
4842 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4850 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4852 void sched_show_task(struct task_struct *p)
4854 unsigned long free = 0;
4856 unsigned long state = p->state;
4859 state = __ffs(state) + 1;
4860 printk(KERN_INFO "%-15.15s %c", p->comm,
4861 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4862 #if BITS_PER_LONG == 32
4863 if (state == TASK_RUNNING)
4864 printk(KERN_CONT " running ");
4866 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4868 if (state == TASK_RUNNING)
4869 printk(KERN_CONT " running task ");
4871 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4873 #ifdef CONFIG_DEBUG_STACK_USAGE
4874 free = stack_not_used(p);
4879 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4881 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4882 task_pid_nr(p), ppid,
4883 (unsigned long)task_thread_info(p)->flags);
4885 print_worker_info(KERN_INFO, p);
4886 show_stack(p, NULL);
4889 void show_state_filter(unsigned long state_filter)
4891 struct task_struct *g, *p;
4893 #if BITS_PER_LONG == 32
4895 " task PC stack pid father\n");
4898 " task PC stack pid father\n");
4901 for_each_process_thread(g, p) {
4903 * reset the NMI-timeout, listing all files on a slow
4904 * console might take a lot of time:
4906 touch_nmi_watchdog();
4907 if (!state_filter || (p->state & state_filter))
4911 touch_all_softlockup_watchdogs();
4913 #ifdef CONFIG_SCHED_DEBUG
4914 sysrq_sched_debug_show();
4918 * Only show locks if all tasks are dumped:
4921 debug_show_all_locks();
4924 void init_idle_bootup_task(struct task_struct *idle)
4926 idle->sched_class = &idle_sched_class;
4930 * init_idle - set up an idle thread for a given CPU
4931 * @idle: task in question
4932 * @cpu: cpu the idle task belongs to
4934 * NOTE: this function does not set the idle thread's NEED_RESCHED
4935 * flag, to make booting more robust.
4937 void init_idle(struct task_struct *idle, int cpu)
4939 struct rq *rq = cpu_rq(cpu);
4940 unsigned long flags;
4942 raw_spin_lock_irqsave(&idle->pi_lock, flags);
4943 raw_spin_lock(&rq->lock);
4945 __sched_fork(0, idle);
4946 idle->state = TASK_RUNNING;
4947 idle->se.exec_start = sched_clock();
4951 * Its possible that init_idle() gets called multiple times on a task,
4952 * in that case do_set_cpus_allowed() will not do the right thing.
4954 * And since this is boot we can forgo the serialization.
4956 set_cpus_allowed_common(idle, cpumask_of(cpu));
4959 * We're having a chicken and egg problem, even though we are
4960 * holding rq->lock, the cpu isn't yet set to this cpu so the
4961 * lockdep check in task_group() will fail.
4963 * Similar case to sched_fork(). / Alternatively we could
4964 * use task_rq_lock() here and obtain the other rq->lock.
4969 __set_task_cpu(idle, cpu);
4972 rq->curr = rq->idle = idle;
4973 idle->on_rq = TASK_ON_RQ_QUEUED;
4977 raw_spin_unlock(&rq->lock);
4978 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
4980 /* Set the preempt count _outside_ the spinlocks! */
4981 init_idle_preempt_count(idle, cpu);
4984 * The idle tasks have their own, simple scheduling class:
4986 idle->sched_class = &idle_sched_class;
4987 ftrace_graph_init_idle_task(idle, cpu);
4988 vtime_init_idle(idle, cpu);
4990 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4994 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4995 const struct cpumask *trial)
4997 int ret = 1, trial_cpus;
4998 struct dl_bw *cur_dl_b;
4999 unsigned long flags;
5001 if (!cpumask_weight(cur))
5004 rcu_read_lock_sched();
5005 cur_dl_b = dl_bw_of(cpumask_any(cur));
5006 trial_cpus = cpumask_weight(trial);
5008 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5009 if (cur_dl_b->bw != -1 &&
5010 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5012 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5013 rcu_read_unlock_sched();
5018 int task_can_attach(struct task_struct *p,
5019 const struct cpumask *cs_cpus_allowed)
5024 * Kthreads which disallow setaffinity shouldn't be moved
5025 * to a new cpuset; we don't want to change their cpu
5026 * affinity and isolating such threads by their set of
5027 * allowed nodes is unnecessary. Thus, cpusets are not
5028 * applicable for such threads. This prevents checking for
5029 * success of set_cpus_allowed_ptr() on all attached tasks
5030 * before cpus_allowed may be changed.
5032 if (p->flags & PF_NO_SETAFFINITY) {
5038 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5040 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5045 unsigned long flags;
5047 rcu_read_lock_sched();
5048 dl_b = dl_bw_of(dest_cpu);
5049 raw_spin_lock_irqsave(&dl_b->lock, flags);
5050 cpus = dl_bw_cpus(dest_cpu);
5051 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5056 * We reserve space for this task in the destination
5057 * root_domain, as we can't fail after this point.
5058 * We will free resources in the source root_domain
5059 * later on (see set_cpus_allowed_dl()).
5061 __dl_add(dl_b, p->dl.dl_bw);
5063 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5064 rcu_read_unlock_sched();
5074 #ifdef CONFIG_NUMA_BALANCING
5075 /* Migrate current task p to target_cpu */
5076 int migrate_task_to(struct task_struct *p, int target_cpu)
5078 struct migration_arg arg = { p, target_cpu };
5079 int curr_cpu = task_cpu(p);
5081 if (curr_cpu == target_cpu)
5084 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5087 /* TODO: This is not properly updating schedstats */
5089 trace_sched_move_numa(p, curr_cpu, target_cpu);
5090 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5094 * Requeue a task on a given node and accurately track the number of NUMA
5095 * tasks on the runqueues
5097 void sched_setnuma(struct task_struct *p, int nid)
5100 unsigned long flags;
5101 bool queued, running;
5103 rq = task_rq_lock(p, &flags);
5104 queued = task_on_rq_queued(p);
5105 running = task_current(rq, p);
5108 dequeue_task(rq, p, 0);
5110 put_prev_task(rq, p);
5112 p->numa_preferred_nid = nid;
5115 p->sched_class->set_curr_task(rq);
5117 enqueue_task(rq, p, 0);
5118 task_rq_unlock(rq, p, &flags);
5120 #endif /* CONFIG_NUMA_BALANCING */
5122 #ifdef CONFIG_HOTPLUG_CPU
5124 * Ensures that the idle task is using init_mm right before its cpu goes
5127 void idle_task_exit(void)
5129 struct mm_struct *mm = current->active_mm;
5131 BUG_ON(cpu_online(smp_processor_id()));
5133 if (mm != &init_mm) {
5134 switch_mm(mm, &init_mm, current);
5135 finish_arch_post_lock_switch();
5141 * Since this CPU is going 'away' for a while, fold any nr_active delta
5142 * we might have. Assumes we're called after migrate_tasks() so that the
5143 * nr_active count is stable.
5145 * Also see the comment "Global load-average calculations".
5147 static void calc_load_migrate(struct rq *rq)
5149 long delta = calc_load_fold_active(rq);
5151 atomic_long_add(delta, &calc_load_tasks);
5154 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5158 static const struct sched_class fake_sched_class = {
5159 .put_prev_task = put_prev_task_fake,
5162 static struct task_struct fake_task = {
5164 * Avoid pull_{rt,dl}_task()
5166 .prio = MAX_PRIO + 1,
5167 .sched_class = &fake_sched_class,
5171 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5172 * try_to_wake_up()->select_task_rq().
5174 * Called with rq->lock held even though we'er in stop_machine() and
5175 * there's no concurrency possible, we hold the required locks anyway
5176 * because of lock validation efforts.
5178 static void migrate_tasks(struct rq *dead_rq)
5180 struct rq *rq = dead_rq;
5181 struct task_struct *next, *stop = rq->stop;
5185 * Fudge the rq selection such that the below task selection loop
5186 * doesn't get stuck on the currently eligible stop task.
5188 * We're currently inside stop_machine() and the rq is either stuck
5189 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5190 * either way we should never end up calling schedule() until we're
5196 * put_prev_task() and pick_next_task() sched
5197 * class method both need to have an up-to-date
5198 * value of rq->clock[_task]
5200 update_rq_clock(rq);
5204 * There's this thread running, bail when that's the only
5207 if (rq->nr_running == 1)
5211 * pick_next_task assumes pinned rq->lock.
5213 lockdep_pin_lock(&rq->lock);
5214 next = pick_next_task(rq, &fake_task);
5216 next->sched_class->put_prev_task(rq, next);
5219 * Rules for changing task_struct::cpus_allowed are holding
5220 * both pi_lock and rq->lock, such that holding either
5221 * stabilizes the mask.
5223 * Drop rq->lock is not quite as disastrous as it usually is
5224 * because !cpu_active at this point, which means load-balance
5225 * will not interfere. Also, stop-machine.
5227 lockdep_unpin_lock(&rq->lock);
5228 raw_spin_unlock(&rq->lock);
5229 raw_spin_lock(&next->pi_lock);
5230 raw_spin_lock(&rq->lock);
5233 * Since we're inside stop-machine, _nothing_ should have
5234 * changed the task, WARN if weird stuff happened, because in
5235 * that case the above rq->lock drop is a fail too.
5237 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5238 raw_spin_unlock(&next->pi_lock);
5242 /* Find suitable destination for @next, with force if needed. */
5243 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5245 rq = __migrate_task(rq, next, dest_cpu);
5246 if (rq != dead_rq) {
5247 raw_spin_unlock(&rq->lock);
5249 raw_spin_lock(&rq->lock);
5251 raw_spin_unlock(&next->pi_lock);
5256 #endif /* CONFIG_HOTPLUG_CPU */
5258 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5260 static struct ctl_table sd_ctl_dir[] = {
5262 .procname = "sched_domain",
5268 static struct ctl_table sd_ctl_root[] = {
5270 .procname = "kernel",
5272 .child = sd_ctl_dir,
5277 static struct ctl_table *sd_alloc_ctl_entry(int n)
5279 struct ctl_table *entry =
5280 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5285 static void sd_free_ctl_entry(struct ctl_table **tablep)
5287 struct ctl_table *entry;
5290 * In the intermediate directories, both the child directory and
5291 * procname are dynamically allocated and could fail but the mode
5292 * will always be set. In the lowest directory the names are
5293 * static strings and all have proc handlers.
5295 for (entry = *tablep; entry->mode; entry++) {
5297 sd_free_ctl_entry(&entry->child);
5298 if (entry->proc_handler == NULL)
5299 kfree(entry->procname);
5306 static int min_load_idx = 0;
5307 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5310 set_table_entry(struct ctl_table *entry,
5311 const char *procname, void *data, int maxlen,
5312 umode_t mode, proc_handler *proc_handler,
5315 entry->procname = procname;
5317 entry->maxlen = maxlen;
5319 entry->proc_handler = proc_handler;
5322 entry->extra1 = &min_load_idx;
5323 entry->extra2 = &max_load_idx;
5327 static struct ctl_table *
5328 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5330 struct ctl_table *table = sd_alloc_ctl_entry(14);
5335 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5336 sizeof(long), 0644, proc_doulongvec_minmax, false);
5337 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5338 sizeof(long), 0644, proc_doulongvec_minmax, false);
5339 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5340 sizeof(int), 0644, proc_dointvec_minmax, true);
5341 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5342 sizeof(int), 0644, proc_dointvec_minmax, true);
5343 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5344 sizeof(int), 0644, proc_dointvec_minmax, true);
5345 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5346 sizeof(int), 0644, proc_dointvec_minmax, true);
5347 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5348 sizeof(int), 0644, proc_dointvec_minmax, true);
5349 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5350 sizeof(int), 0644, proc_dointvec_minmax, false);
5351 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5352 sizeof(int), 0644, proc_dointvec_minmax, false);
5353 set_table_entry(&table[9], "cache_nice_tries",
5354 &sd->cache_nice_tries,
5355 sizeof(int), 0644, proc_dointvec_minmax, false);
5356 set_table_entry(&table[10], "flags", &sd->flags,
5357 sizeof(int), 0644, proc_dointvec_minmax, false);
5358 set_table_entry(&table[11], "max_newidle_lb_cost",
5359 &sd->max_newidle_lb_cost,
5360 sizeof(long), 0644, proc_doulongvec_minmax, false);
5361 set_table_entry(&table[12], "name", sd->name,
5362 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5363 /* &table[13] is terminator */
5368 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5370 struct ctl_table *entry, *table;
5371 struct sched_domain *sd;
5372 int domain_num = 0, i;
5375 for_each_domain(cpu, sd)
5377 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5382 for_each_domain(cpu, sd) {
5383 snprintf(buf, 32, "domain%d", i);
5384 entry->procname = kstrdup(buf, GFP_KERNEL);
5386 entry->child = sd_alloc_ctl_domain_table(sd);
5393 static struct ctl_table_header *sd_sysctl_header;
5394 static void register_sched_domain_sysctl(void)
5396 int i, cpu_num = num_possible_cpus();
5397 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5400 WARN_ON(sd_ctl_dir[0].child);
5401 sd_ctl_dir[0].child = entry;
5406 for_each_possible_cpu(i) {
5407 snprintf(buf, 32, "cpu%d", i);
5408 entry->procname = kstrdup(buf, GFP_KERNEL);
5410 entry->child = sd_alloc_ctl_cpu_table(i);
5414 WARN_ON(sd_sysctl_header);
5415 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5418 /* may be called multiple times per register */
5419 static void unregister_sched_domain_sysctl(void)
5421 unregister_sysctl_table(sd_sysctl_header);
5422 sd_sysctl_header = NULL;
5423 if (sd_ctl_dir[0].child)
5424 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5427 static void register_sched_domain_sysctl(void)
5430 static void unregister_sched_domain_sysctl(void)
5433 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5435 static void set_rq_online(struct rq *rq)
5438 const struct sched_class *class;
5440 cpumask_set_cpu(rq->cpu, rq->rd->online);
5443 for_each_class(class) {
5444 if (class->rq_online)
5445 class->rq_online(rq);
5450 static void set_rq_offline(struct rq *rq)
5453 const struct sched_class *class;
5455 for_each_class(class) {
5456 if (class->rq_offline)
5457 class->rq_offline(rq);
5460 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5466 * migration_call - callback that gets triggered when a CPU is added.
5467 * Here we can start up the necessary migration thread for the new CPU.
5470 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5472 int cpu = (long)hcpu;
5473 unsigned long flags;
5474 struct rq *rq = cpu_rq(cpu);
5476 switch (action & ~CPU_TASKS_FROZEN) {
5478 case CPU_UP_PREPARE:
5479 rq->calc_load_update = calc_load_update;
5483 /* Update our root-domain */
5484 raw_spin_lock_irqsave(&rq->lock, flags);
5486 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5490 raw_spin_unlock_irqrestore(&rq->lock, flags);
5493 #ifdef CONFIG_HOTPLUG_CPU
5495 sched_ttwu_pending();
5496 /* Update our root-domain */
5497 raw_spin_lock_irqsave(&rq->lock, flags);
5499 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5503 BUG_ON(rq->nr_running != 1); /* the migration thread */
5504 raw_spin_unlock_irqrestore(&rq->lock, flags);
5508 calc_load_migrate(rq);
5513 update_max_interval();
5519 * Register at high priority so that task migration (migrate_all_tasks)
5520 * happens before everything else. This has to be lower priority than
5521 * the notifier in the perf_event subsystem, though.
5523 static struct notifier_block migration_notifier = {
5524 .notifier_call = migration_call,
5525 .priority = CPU_PRI_MIGRATION,
5528 static void set_cpu_rq_start_time(void)
5530 int cpu = smp_processor_id();
5531 struct rq *rq = cpu_rq(cpu);
5532 rq->age_stamp = sched_clock_cpu(cpu);
5535 static int sched_cpu_active(struct notifier_block *nfb,
5536 unsigned long action, void *hcpu)
5538 switch (action & ~CPU_TASKS_FROZEN) {
5540 set_cpu_rq_start_time();
5544 * At this point a starting CPU has marked itself as online via
5545 * set_cpu_online(). But it might not yet have marked itself
5546 * as active, which is essential from here on.
5548 * Thus, fall-through and help the starting CPU along.
5550 case CPU_DOWN_FAILED:
5551 set_cpu_active((long)hcpu, true);
5558 static int sched_cpu_inactive(struct notifier_block *nfb,
5559 unsigned long action, void *hcpu)
5561 switch (action & ~CPU_TASKS_FROZEN) {
5562 case CPU_DOWN_PREPARE:
5563 set_cpu_active((long)hcpu, false);
5570 static int __init migration_init(void)
5572 void *cpu = (void *)(long)smp_processor_id();
5575 /* Initialize migration for the boot CPU */
5576 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5577 BUG_ON(err == NOTIFY_BAD);
5578 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5579 register_cpu_notifier(&migration_notifier);
5581 /* Register cpu active notifiers */
5582 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5583 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5587 early_initcall(migration_init);
5589 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5591 #ifdef CONFIG_SCHED_DEBUG
5593 static __read_mostly int sched_debug_enabled;
5595 static int __init sched_debug_setup(char *str)
5597 sched_debug_enabled = 1;
5601 early_param("sched_debug", sched_debug_setup);
5603 static inline bool sched_debug(void)
5605 return sched_debug_enabled;
5608 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5609 struct cpumask *groupmask)
5611 struct sched_group *group = sd->groups;
5613 cpumask_clear(groupmask);
5615 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5617 if (!(sd->flags & SD_LOAD_BALANCE)) {
5618 printk("does not load-balance\n");
5620 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5625 printk(KERN_CONT "span %*pbl level %s\n",
5626 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5628 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5629 printk(KERN_ERR "ERROR: domain->span does not contain "
5632 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5633 printk(KERN_ERR "ERROR: domain->groups does not contain"
5637 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5641 printk(KERN_ERR "ERROR: group is NULL\n");
5645 if (!cpumask_weight(sched_group_cpus(group))) {
5646 printk(KERN_CONT "\n");
5647 printk(KERN_ERR "ERROR: empty group\n");
5651 if (!(sd->flags & SD_OVERLAP) &&
5652 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5653 printk(KERN_CONT "\n");
5654 printk(KERN_ERR "ERROR: repeated CPUs\n");
5658 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5660 printk(KERN_CONT " %*pbl",
5661 cpumask_pr_args(sched_group_cpus(group)));
5662 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5663 printk(KERN_CONT " (cpu_capacity = %d)",
5664 group->sgc->capacity);
5667 group = group->next;
5668 } while (group != sd->groups);
5669 printk(KERN_CONT "\n");
5671 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5672 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5675 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5676 printk(KERN_ERR "ERROR: parent span is not a superset "
5677 "of domain->span\n");
5681 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5685 if (!sched_debug_enabled)
5689 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5693 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5696 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5704 #else /* !CONFIG_SCHED_DEBUG */
5705 # define sched_domain_debug(sd, cpu) do { } while (0)
5706 static inline bool sched_debug(void)
5710 #endif /* CONFIG_SCHED_DEBUG */
5712 static int sd_degenerate(struct sched_domain *sd)
5714 if (cpumask_weight(sched_domain_span(sd)) == 1)
5717 /* Following flags need at least 2 groups */
5718 if (sd->flags & (SD_LOAD_BALANCE |
5719 SD_BALANCE_NEWIDLE |
5722 SD_SHARE_CPUCAPACITY |
5723 SD_SHARE_PKG_RESOURCES |
5724 SD_SHARE_POWERDOMAIN)) {
5725 if (sd->groups != sd->groups->next)
5729 /* Following flags don't use groups */
5730 if (sd->flags & (SD_WAKE_AFFINE))
5737 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5739 unsigned long cflags = sd->flags, pflags = parent->flags;
5741 if (sd_degenerate(parent))
5744 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5747 /* Flags needing groups don't count if only 1 group in parent */
5748 if (parent->groups == parent->groups->next) {
5749 pflags &= ~(SD_LOAD_BALANCE |
5750 SD_BALANCE_NEWIDLE |
5753 SD_SHARE_CPUCAPACITY |
5754 SD_SHARE_PKG_RESOURCES |
5756 SD_SHARE_POWERDOMAIN);
5757 if (nr_node_ids == 1)
5758 pflags &= ~SD_SERIALIZE;
5760 if (~cflags & pflags)
5766 static void free_rootdomain(struct rcu_head *rcu)
5768 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5770 cpupri_cleanup(&rd->cpupri);
5771 cpudl_cleanup(&rd->cpudl);
5772 free_cpumask_var(rd->dlo_mask);
5773 free_cpumask_var(rd->rto_mask);
5774 free_cpumask_var(rd->online);
5775 free_cpumask_var(rd->span);
5779 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5781 struct root_domain *old_rd = NULL;
5782 unsigned long flags;
5784 raw_spin_lock_irqsave(&rq->lock, flags);
5789 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5792 cpumask_clear_cpu(rq->cpu, old_rd->span);
5795 * If we dont want to free the old_rd yet then
5796 * set old_rd to NULL to skip the freeing later
5799 if (!atomic_dec_and_test(&old_rd->refcount))
5803 atomic_inc(&rd->refcount);
5806 cpumask_set_cpu(rq->cpu, rd->span);
5807 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5810 raw_spin_unlock_irqrestore(&rq->lock, flags);
5813 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5816 static int init_rootdomain(struct root_domain *rd)
5818 memset(rd, 0, sizeof(*rd));
5820 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5822 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5824 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5826 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5829 init_dl_bw(&rd->dl_bw);
5830 if (cpudl_init(&rd->cpudl) != 0)
5833 if (cpupri_init(&rd->cpupri) != 0)
5838 free_cpumask_var(rd->rto_mask);
5840 free_cpumask_var(rd->dlo_mask);
5842 free_cpumask_var(rd->online);
5844 free_cpumask_var(rd->span);
5850 * By default the system creates a single root-domain with all cpus as
5851 * members (mimicking the global state we have today).
5853 struct root_domain def_root_domain;
5855 static void init_defrootdomain(void)
5857 init_rootdomain(&def_root_domain);
5859 atomic_set(&def_root_domain.refcount, 1);
5862 static struct root_domain *alloc_rootdomain(void)
5864 struct root_domain *rd;
5866 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5870 if (init_rootdomain(rd) != 0) {
5878 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5880 struct sched_group *tmp, *first;
5889 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5894 } while (sg != first);
5897 static void free_sched_domain(struct rcu_head *rcu)
5899 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5902 * If its an overlapping domain it has private groups, iterate and
5905 if (sd->flags & SD_OVERLAP) {
5906 free_sched_groups(sd->groups, 1);
5907 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5908 kfree(sd->groups->sgc);
5914 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5916 call_rcu(&sd->rcu, free_sched_domain);
5919 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5921 for (; sd; sd = sd->parent)
5922 destroy_sched_domain(sd, cpu);
5926 * Keep a special pointer to the highest sched_domain that has
5927 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5928 * allows us to avoid some pointer chasing select_idle_sibling().
5930 * Also keep a unique ID per domain (we use the first cpu number in
5931 * the cpumask of the domain), this allows us to quickly tell if
5932 * two cpus are in the same cache domain, see cpus_share_cache().
5934 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5935 DEFINE_PER_CPU(int, sd_llc_size);
5936 DEFINE_PER_CPU(int, sd_llc_id);
5937 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5938 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5939 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5941 static void update_top_cache_domain(int cpu)
5943 struct sched_domain *sd;
5944 struct sched_domain *busy_sd = NULL;
5948 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5950 id = cpumask_first(sched_domain_span(sd));
5951 size = cpumask_weight(sched_domain_span(sd));
5952 busy_sd = sd->parent; /* sd_busy */
5954 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5956 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5957 per_cpu(sd_llc_size, cpu) = size;
5958 per_cpu(sd_llc_id, cpu) = id;
5960 sd = lowest_flag_domain(cpu, SD_NUMA);
5961 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5963 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5964 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5968 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5969 * hold the hotplug lock.
5972 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5974 struct rq *rq = cpu_rq(cpu);
5975 struct sched_domain *tmp;
5977 /* Remove the sched domains which do not contribute to scheduling. */
5978 for (tmp = sd; tmp; ) {
5979 struct sched_domain *parent = tmp->parent;
5983 if (sd_parent_degenerate(tmp, parent)) {
5984 tmp->parent = parent->parent;
5986 parent->parent->child = tmp;
5988 * Transfer SD_PREFER_SIBLING down in case of a
5989 * degenerate parent; the spans match for this
5990 * so the property transfers.
5992 if (parent->flags & SD_PREFER_SIBLING)
5993 tmp->flags |= SD_PREFER_SIBLING;
5994 destroy_sched_domain(parent, cpu);
5999 if (sd && sd_degenerate(sd)) {
6002 destroy_sched_domain(tmp, cpu);
6007 sched_domain_debug(sd, cpu);
6009 rq_attach_root(rq, rd);
6011 rcu_assign_pointer(rq->sd, sd);
6012 destroy_sched_domains(tmp, cpu);
6014 update_top_cache_domain(cpu);
6017 /* Setup the mask of cpus configured for isolated domains */
6018 static int __init isolated_cpu_setup(char *str)
6020 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6021 cpulist_parse(str, cpu_isolated_map);
6025 __setup("isolcpus=", isolated_cpu_setup);
6028 struct sched_domain ** __percpu sd;
6029 struct root_domain *rd;
6040 * Build an iteration mask that can exclude certain CPUs from the upwards
6043 * Asymmetric node setups can result in situations where the domain tree is of
6044 * unequal depth, make sure to skip domains that already cover the entire
6047 * In that case build_sched_domains() will have terminated the iteration early
6048 * and our sibling sd spans will be empty. Domains should always include the
6049 * cpu they're built on, so check that.
6052 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6054 const struct cpumask *span = sched_domain_span(sd);
6055 struct sd_data *sdd = sd->private;
6056 struct sched_domain *sibling;
6059 for_each_cpu(i, span) {
6060 sibling = *per_cpu_ptr(sdd->sd, i);
6061 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6064 cpumask_set_cpu(i, sched_group_mask(sg));
6069 * Return the canonical balance cpu for this group, this is the first cpu
6070 * of this group that's also in the iteration mask.
6072 int group_balance_cpu(struct sched_group *sg)
6074 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6078 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6080 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6081 const struct cpumask *span = sched_domain_span(sd);
6082 struct cpumask *covered = sched_domains_tmpmask;
6083 struct sd_data *sdd = sd->private;
6084 struct sched_domain *sibling;
6087 cpumask_clear(covered);
6089 for_each_cpu(i, span) {
6090 struct cpumask *sg_span;
6092 if (cpumask_test_cpu(i, covered))
6095 sibling = *per_cpu_ptr(sdd->sd, i);
6097 /* See the comment near build_group_mask(). */
6098 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6101 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6102 GFP_KERNEL, cpu_to_node(cpu));
6107 sg_span = sched_group_cpus(sg);
6109 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6111 cpumask_set_cpu(i, sg_span);
6113 cpumask_or(covered, covered, sg_span);
6115 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6116 if (atomic_inc_return(&sg->sgc->ref) == 1)
6117 build_group_mask(sd, sg);
6120 * Initialize sgc->capacity such that even if we mess up the
6121 * domains and no possible iteration will get us here, we won't
6124 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6127 * Make sure the first group of this domain contains the
6128 * canonical balance cpu. Otherwise the sched_domain iteration
6129 * breaks. See update_sg_lb_stats().
6131 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6132 group_balance_cpu(sg) == cpu)
6142 sd->groups = groups;
6147 free_sched_groups(first, 0);
6152 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6154 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6155 struct sched_domain *child = sd->child;
6158 cpu = cpumask_first(sched_domain_span(child));
6161 *sg = *per_cpu_ptr(sdd->sg, cpu);
6162 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6163 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6170 * build_sched_groups will build a circular linked list of the groups
6171 * covered by the given span, and will set each group's ->cpumask correctly,
6172 * and ->cpu_capacity to 0.
6174 * Assumes the sched_domain tree is fully constructed
6177 build_sched_groups(struct sched_domain *sd, int cpu)
6179 struct sched_group *first = NULL, *last = NULL;
6180 struct sd_data *sdd = sd->private;
6181 const struct cpumask *span = sched_domain_span(sd);
6182 struct cpumask *covered;
6185 get_group(cpu, sdd, &sd->groups);
6186 atomic_inc(&sd->groups->ref);
6188 if (cpu != cpumask_first(span))
6191 lockdep_assert_held(&sched_domains_mutex);
6192 covered = sched_domains_tmpmask;
6194 cpumask_clear(covered);
6196 for_each_cpu(i, span) {
6197 struct sched_group *sg;
6200 if (cpumask_test_cpu(i, covered))
6203 group = get_group(i, sdd, &sg);
6204 cpumask_setall(sched_group_mask(sg));
6206 for_each_cpu(j, span) {
6207 if (get_group(j, sdd, NULL) != group)
6210 cpumask_set_cpu(j, covered);
6211 cpumask_set_cpu(j, sched_group_cpus(sg));
6226 * Initialize sched groups cpu_capacity.
6228 * cpu_capacity indicates the capacity of sched group, which is used while
6229 * distributing the load between different sched groups in a sched domain.
6230 * Typically cpu_capacity for all the groups in a sched domain will be same
6231 * unless there are asymmetries in the topology. If there are asymmetries,
6232 * group having more cpu_capacity will pickup more load compared to the
6233 * group having less cpu_capacity.
6235 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6237 struct sched_group *sg = sd->groups;
6242 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6244 } while (sg != sd->groups);
6246 if (cpu != group_balance_cpu(sg))
6249 update_group_capacity(sd, cpu);
6250 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6254 * Initializers for schedule domains
6255 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6258 static int default_relax_domain_level = -1;
6259 int sched_domain_level_max;
6261 static int __init setup_relax_domain_level(char *str)
6263 if (kstrtoint(str, 0, &default_relax_domain_level))
6264 pr_warn("Unable to set relax_domain_level\n");
6268 __setup("relax_domain_level=", setup_relax_domain_level);
6270 static void set_domain_attribute(struct sched_domain *sd,
6271 struct sched_domain_attr *attr)
6275 if (!attr || attr->relax_domain_level < 0) {
6276 if (default_relax_domain_level < 0)
6279 request = default_relax_domain_level;
6281 request = attr->relax_domain_level;
6282 if (request < sd->level) {
6283 /* turn off idle balance on this domain */
6284 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6286 /* turn on idle balance on this domain */
6287 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6291 static void __sdt_free(const struct cpumask *cpu_map);
6292 static int __sdt_alloc(const struct cpumask *cpu_map);
6294 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6295 const struct cpumask *cpu_map)
6299 if (!atomic_read(&d->rd->refcount))
6300 free_rootdomain(&d->rd->rcu); /* fall through */
6302 free_percpu(d->sd); /* fall through */
6304 __sdt_free(cpu_map); /* fall through */
6310 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6311 const struct cpumask *cpu_map)
6313 memset(d, 0, sizeof(*d));
6315 if (__sdt_alloc(cpu_map))
6316 return sa_sd_storage;
6317 d->sd = alloc_percpu(struct sched_domain *);
6319 return sa_sd_storage;
6320 d->rd = alloc_rootdomain();
6323 return sa_rootdomain;
6327 * NULL the sd_data elements we've used to build the sched_domain and
6328 * sched_group structure so that the subsequent __free_domain_allocs()
6329 * will not free the data we're using.
6331 static void claim_allocations(int cpu, struct sched_domain *sd)
6333 struct sd_data *sdd = sd->private;
6335 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6336 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6338 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6339 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6341 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6342 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6346 static int sched_domains_numa_levels;
6347 enum numa_topology_type sched_numa_topology_type;
6348 static int *sched_domains_numa_distance;
6349 int sched_max_numa_distance;
6350 static struct cpumask ***sched_domains_numa_masks;
6351 static int sched_domains_curr_level;
6355 * SD_flags allowed in topology descriptions.
6357 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6358 * SD_SHARE_PKG_RESOURCES - describes shared caches
6359 * SD_NUMA - describes NUMA topologies
6360 * SD_SHARE_POWERDOMAIN - describes shared power domain
6363 * SD_ASYM_PACKING - describes SMT quirks
6365 #define TOPOLOGY_SD_FLAGS \
6366 (SD_SHARE_CPUCAPACITY | \
6367 SD_SHARE_PKG_RESOURCES | \
6370 SD_SHARE_POWERDOMAIN)
6372 static struct sched_domain *
6373 sd_init(struct sched_domain_topology_level *tl, int cpu)
6375 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6376 int sd_weight, sd_flags = 0;
6380 * Ugly hack to pass state to sd_numa_mask()...
6382 sched_domains_curr_level = tl->numa_level;
6385 sd_weight = cpumask_weight(tl->mask(cpu));
6388 sd_flags = (*tl->sd_flags)();
6389 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6390 "wrong sd_flags in topology description\n"))
6391 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6393 *sd = (struct sched_domain){
6394 .min_interval = sd_weight,
6395 .max_interval = 2*sd_weight,
6397 .imbalance_pct = 125,
6399 .cache_nice_tries = 0,
6406 .flags = 1*SD_LOAD_BALANCE
6407 | 1*SD_BALANCE_NEWIDLE
6412 | 0*SD_SHARE_CPUCAPACITY
6413 | 0*SD_SHARE_PKG_RESOURCES
6415 | 0*SD_PREFER_SIBLING
6420 .last_balance = jiffies,
6421 .balance_interval = sd_weight,
6423 .max_newidle_lb_cost = 0,
6424 .next_decay_max_lb_cost = jiffies,
6425 #ifdef CONFIG_SCHED_DEBUG
6431 * Convert topological properties into behaviour.
6434 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6435 sd->flags |= SD_PREFER_SIBLING;
6436 sd->imbalance_pct = 110;
6437 sd->smt_gain = 1178; /* ~15% */
6439 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6440 sd->imbalance_pct = 117;
6441 sd->cache_nice_tries = 1;
6445 } else if (sd->flags & SD_NUMA) {
6446 sd->cache_nice_tries = 2;
6450 sd->flags |= SD_SERIALIZE;
6451 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6452 sd->flags &= ~(SD_BALANCE_EXEC |
6459 sd->flags |= SD_PREFER_SIBLING;
6460 sd->cache_nice_tries = 1;
6465 sd->private = &tl->data;
6471 * Topology list, bottom-up.
6473 static struct sched_domain_topology_level default_topology[] = {
6474 #ifdef CONFIG_SCHED_SMT
6475 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6477 #ifdef CONFIG_SCHED_MC
6478 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6480 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6484 static struct sched_domain_topology_level *sched_domain_topology =
6487 #define for_each_sd_topology(tl) \
6488 for (tl = sched_domain_topology; tl->mask; tl++)
6490 void set_sched_topology(struct sched_domain_topology_level *tl)
6492 sched_domain_topology = tl;
6497 static const struct cpumask *sd_numa_mask(int cpu)
6499 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6502 static void sched_numa_warn(const char *str)
6504 static int done = false;
6512 printk(KERN_WARNING "ERROR: %s\n\n", str);
6514 for (i = 0; i < nr_node_ids; i++) {
6515 printk(KERN_WARNING " ");
6516 for (j = 0; j < nr_node_ids; j++)
6517 printk(KERN_CONT "%02d ", node_distance(i,j));
6518 printk(KERN_CONT "\n");
6520 printk(KERN_WARNING "\n");
6523 bool find_numa_distance(int distance)
6527 if (distance == node_distance(0, 0))
6530 for (i = 0; i < sched_domains_numa_levels; i++) {
6531 if (sched_domains_numa_distance[i] == distance)
6539 * A system can have three types of NUMA topology:
6540 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6541 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6542 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6544 * The difference between a glueless mesh topology and a backplane
6545 * topology lies in whether communication between not directly
6546 * connected nodes goes through intermediary nodes (where programs
6547 * could run), or through backplane controllers. This affects
6548 * placement of programs.
6550 * The type of topology can be discerned with the following tests:
6551 * - If the maximum distance between any nodes is 1 hop, the system
6552 * is directly connected.
6553 * - If for two nodes A and B, located N > 1 hops away from each other,
6554 * there is an intermediary node C, which is < N hops away from both
6555 * nodes A and B, the system is a glueless mesh.
6557 static void init_numa_topology_type(void)
6561 n = sched_max_numa_distance;
6563 if (sched_domains_numa_levels <= 1) {
6564 sched_numa_topology_type = NUMA_DIRECT;
6568 for_each_online_node(a) {
6569 for_each_online_node(b) {
6570 /* Find two nodes furthest removed from each other. */
6571 if (node_distance(a, b) < n)
6574 /* Is there an intermediary node between a and b? */
6575 for_each_online_node(c) {
6576 if (node_distance(a, c) < n &&
6577 node_distance(b, c) < n) {
6578 sched_numa_topology_type =
6584 sched_numa_topology_type = NUMA_BACKPLANE;
6590 static void sched_init_numa(void)
6592 int next_distance, curr_distance = node_distance(0, 0);
6593 struct sched_domain_topology_level *tl;
6597 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6598 if (!sched_domains_numa_distance)
6602 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6603 * unique distances in the node_distance() table.
6605 * Assumes node_distance(0,j) includes all distances in
6606 * node_distance(i,j) in order to avoid cubic time.
6608 next_distance = curr_distance;
6609 for (i = 0; i < nr_node_ids; i++) {
6610 for (j = 0; j < nr_node_ids; j++) {
6611 for (k = 0; k < nr_node_ids; k++) {
6612 int distance = node_distance(i, k);
6614 if (distance > curr_distance &&
6615 (distance < next_distance ||
6616 next_distance == curr_distance))
6617 next_distance = distance;
6620 * While not a strong assumption it would be nice to know
6621 * about cases where if node A is connected to B, B is not
6622 * equally connected to A.
6624 if (sched_debug() && node_distance(k, i) != distance)
6625 sched_numa_warn("Node-distance not symmetric");
6627 if (sched_debug() && i && !find_numa_distance(distance))
6628 sched_numa_warn("Node-0 not representative");
6630 if (next_distance != curr_distance) {
6631 sched_domains_numa_distance[level++] = next_distance;
6632 sched_domains_numa_levels = level;
6633 curr_distance = next_distance;
6638 * In case of sched_debug() we verify the above assumption.
6648 * 'level' contains the number of unique distances, excluding the
6649 * identity distance node_distance(i,i).
6651 * The sched_domains_numa_distance[] array includes the actual distance
6656 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6657 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6658 * the array will contain less then 'level' members. This could be
6659 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6660 * in other functions.
6662 * We reset it to 'level' at the end of this function.
6664 sched_domains_numa_levels = 0;
6666 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6667 if (!sched_domains_numa_masks)
6671 * Now for each level, construct a mask per node which contains all
6672 * cpus of nodes that are that many hops away from us.
6674 for (i = 0; i < level; i++) {
6675 sched_domains_numa_masks[i] =
6676 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6677 if (!sched_domains_numa_masks[i])
6680 for (j = 0; j < nr_node_ids; j++) {
6681 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6685 sched_domains_numa_masks[i][j] = mask;
6687 for (k = 0; k < nr_node_ids; k++) {
6688 if (node_distance(j, k) > sched_domains_numa_distance[i])
6691 cpumask_or(mask, mask, cpumask_of_node(k));
6696 /* Compute default topology size */
6697 for (i = 0; sched_domain_topology[i].mask; i++);
6699 tl = kzalloc((i + level + 1) *
6700 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6705 * Copy the default topology bits..
6707 for (i = 0; sched_domain_topology[i].mask; i++)
6708 tl[i] = sched_domain_topology[i];
6711 * .. and append 'j' levels of NUMA goodness.
6713 for (j = 0; j < level; i++, j++) {
6714 tl[i] = (struct sched_domain_topology_level){
6715 .mask = sd_numa_mask,
6716 .sd_flags = cpu_numa_flags,
6717 .flags = SDTL_OVERLAP,
6723 sched_domain_topology = tl;
6725 sched_domains_numa_levels = level;
6726 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6728 init_numa_topology_type();
6731 static void sched_domains_numa_masks_set(int cpu)
6734 int node = cpu_to_node(cpu);
6736 for (i = 0; i < sched_domains_numa_levels; i++) {
6737 for (j = 0; j < nr_node_ids; j++) {
6738 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6739 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6744 static void sched_domains_numa_masks_clear(int cpu)
6747 for (i = 0; i < sched_domains_numa_levels; i++) {
6748 for (j = 0; j < nr_node_ids; j++)
6749 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6754 * Update sched_domains_numa_masks[level][node] array when new cpus
6757 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6758 unsigned long action,
6761 int cpu = (long)hcpu;
6763 switch (action & ~CPU_TASKS_FROZEN) {
6765 sched_domains_numa_masks_set(cpu);
6769 sched_domains_numa_masks_clear(cpu);
6779 static inline void sched_init_numa(void)
6783 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6784 unsigned long action,
6789 #endif /* CONFIG_NUMA */
6791 static int __sdt_alloc(const struct cpumask *cpu_map)
6793 struct sched_domain_topology_level *tl;
6796 for_each_sd_topology(tl) {
6797 struct sd_data *sdd = &tl->data;
6799 sdd->sd = alloc_percpu(struct sched_domain *);
6803 sdd->sg = alloc_percpu(struct sched_group *);
6807 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6811 for_each_cpu(j, cpu_map) {
6812 struct sched_domain *sd;
6813 struct sched_group *sg;
6814 struct sched_group_capacity *sgc;
6816 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6817 GFP_KERNEL, cpu_to_node(j));
6821 *per_cpu_ptr(sdd->sd, j) = sd;
6823 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6824 GFP_KERNEL, cpu_to_node(j));
6830 *per_cpu_ptr(sdd->sg, j) = sg;
6832 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6833 GFP_KERNEL, cpu_to_node(j));
6837 *per_cpu_ptr(sdd->sgc, j) = sgc;
6844 static void __sdt_free(const struct cpumask *cpu_map)
6846 struct sched_domain_topology_level *tl;
6849 for_each_sd_topology(tl) {
6850 struct sd_data *sdd = &tl->data;
6852 for_each_cpu(j, cpu_map) {
6853 struct sched_domain *sd;
6856 sd = *per_cpu_ptr(sdd->sd, j);
6857 if (sd && (sd->flags & SD_OVERLAP))
6858 free_sched_groups(sd->groups, 0);
6859 kfree(*per_cpu_ptr(sdd->sd, j));
6863 kfree(*per_cpu_ptr(sdd->sg, j));
6865 kfree(*per_cpu_ptr(sdd->sgc, j));
6867 free_percpu(sdd->sd);
6869 free_percpu(sdd->sg);
6871 free_percpu(sdd->sgc);
6876 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6877 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6878 struct sched_domain *child, int cpu)
6880 struct sched_domain *sd = sd_init(tl, cpu);
6884 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6886 sd->level = child->level + 1;
6887 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6891 if (!cpumask_subset(sched_domain_span(child),
6892 sched_domain_span(sd))) {
6893 pr_err("BUG: arch topology borken\n");
6894 #ifdef CONFIG_SCHED_DEBUG
6895 pr_err(" the %s domain not a subset of the %s domain\n",
6896 child->name, sd->name);
6898 /* Fixup, ensure @sd has at least @child cpus. */
6899 cpumask_or(sched_domain_span(sd),
6900 sched_domain_span(sd),
6901 sched_domain_span(child));
6905 set_domain_attribute(sd, attr);
6911 * Build sched domains for a given set of cpus and attach the sched domains
6912 * to the individual cpus
6914 static int build_sched_domains(const struct cpumask *cpu_map,
6915 struct sched_domain_attr *attr)
6917 enum s_alloc alloc_state;
6918 struct sched_domain *sd;
6920 int i, ret = -ENOMEM;
6922 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6923 if (alloc_state != sa_rootdomain)
6926 /* Set up domains for cpus specified by the cpu_map. */
6927 for_each_cpu(i, cpu_map) {
6928 struct sched_domain_topology_level *tl;
6931 for_each_sd_topology(tl) {
6932 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6933 if (tl == sched_domain_topology)
6934 *per_cpu_ptr(d.sd, i) = sd;
6935 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6936 sd->flags |= SD_OVERLAP;
6937 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6942 /* Build the groups for the domains */
6943 for_each_cpu(i, cpu_map) {
6944 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6945 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6946 if (sd->flags & SD_OVERLAP) {
6947 if (build_overlap_sched_groups(sd, i))
6950 if (build_sched_groups(sd, i))
6956 /* Calculate CPU capacity for physical packages and nodes */
6957 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6958 if (!cpumask_test_cpu(i, cpu_map))
6961 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6962 claim_allocations(i, sd);
6963 init_sched_groups_capacity(i, sd);
6967 /* Attach the domains */
6969 for_each_cpu(i, cpu_map) {
6970 sd = *per_cpu_ptr(d.sd, i);
6971 cpu_attach_domain(sd, d.rd, i);
6977 __free_domain_allocs(&d, alloc_state, cpu_map);
6981 static cpumask_var_t *doms_cur; /* current sched domains */
6982 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6983 static struct sched_domain_attr *dattr_cur;
6984 /* attribues of custom domains in 'doms_cur' */
6987 * Special case: If a kmalloc of a doms_cur partition (array of
6988 * cpumask) fails, then fallback to a single sched domain,
6989 * as determined by the single cpumask fallback_doms.
6991 static cpumask_var_t fallback_doms;
6994 * arch_update_cpu_topology lets virtualized architectures update the
6995 * cpu core maps. It is supposed to return 1 if the topology changed
6996 * or 0 if it stayed the same.
6998 int __weak arch_update_cpu_topology(void)
7003 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7006 cpumask_var_t *doms;
7008 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7011 for (i = 0; i < ndoms; i++) {
7012 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7013 free_sched_domains(doms, i);
7020 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7023 for (i = 0; i < ndoms; i++)
7024 free_cpumask_var(doms[i]);
7029 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7030 * For now this just excludes isolated cpus, but could be used to
7031 * exclude other special cases in the future.
7033 static int init_sched_domains(const struct cpumask *cpu_map)
7037 arch_update_cpu_topology();
7039 doms_cur = alloc_sched_domains(ndoms_cur);
7041 doms_cur = &fallback_doms;
7042 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7043 err = build_sched_domains(doms_cur[0], NULL);
7044 register_sched_domain_sysctl();
7050 * Detach sched domains from a group of cpus specified in cpu_map
7051 * These cpus will now be attached to the NULL domain
7053 static void detach_destroy_domains(const struct cpumask *cpu_map)
7058 for_each_cpu(i, cpu_map)
7059 cpu_attach_domain(NULL, &def_root_domain, i);
7063 /* handle null as "default" */
7064 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7065 struct sched_domain_attr *new, int idx_new)
7067 struct sched_domain_attr tmp;
7074 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7075 new ? (new + idx_new) : &tmp,
7076 sizeof(struct sched_domain_attr));
7080 * Partition sched domains as specified by the 'ndoms_new'
7081 * cpumasks in the array doms_new[] of cpumasks. This compares
7082 * doms_new[] to the current sched domain partitioning, doms_cur[].
7083 * It destroys each deleted domain and builds each new domain.
7085 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7086 * The masks don't intersect (don't overlap.) We should setup one
7087 * sched domain for each mask. CPUs not in any of the cpumasks will
7088 * not be load balanced. If the same cpumask appears both in the
7089 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7092 * The passed in 'doms_new' should be allocated using
7093 * alloc_sched_domains. This routine takes ownership of it and will
7094 * free_sched_domains it when done with it. If the caller failed the
7095 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7096 * and partition_sched_domains() will fallback to the single partition
7097 * 'fallback_doms', it also forces the domains to be rebuilt.
7099 * If doms_new == NULL it will be replaced with cpu_online_mask.
7100 * ndoms_new == 0 is a special case for destroying existing domains,
7101 * and it will not create the default domain.
7103 * Call with hotplug lock held
7105 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7106 struct sched_domain_attr *dattr_new)
7111 mutex_lock(&sched_domains_mutex);
7113 /* always unregister in case we don't destroy any domains */
7114 unregister_sched_domain_sysctl();
7116 /* Let architecture update cpu core mappings. */
7117 new_topology = arch_update_cpu_topology();
7119 n = doms_new ? ndoms_new : 0;
7121 /* Destroy deleted domains */
7122 for (i = 0; i < ndoms_cur; i++) {
7123 for (j = 0; j < n && !new_topology; j++) {
7124 if (cpumask_equal(doms_cur[i], doms_new[j])
7125 && dattrs_equal(dattr_cur, i, dattr_new, j))
7128 /* no match - a current sched domain not in new doms_new[] */
7129 detach_destroy_domains(doms_cur[i]);
7135 if (doms_new == NULL) {
7137 doms_new = &fallback_doms;
7138 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7139 WARN_ON_ONCE(dattr_new);
7142 /* Build new domains */
7143 for (i = 0; i < ndoms_new; i++) {
7144 for (j = 0; j < n && !new_topology; j++) {
7145 if (cpumask_equal(doms_new[i], doms_cur[j])
7146 && dattrs_equal(dattr_new, i, dattr_cur, j))
7149 /* no match - add a new doms_new */
7150 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7155 /* Remember the new sched domains */
7156 if (doms_cur != &fallback_doms)
7157 free_sched_domains(doms_cur, ndoms_cur);
7158 kfree(dattr_cur); /* kfree(NULL) is safe */
7159 doms_cur = doms_new;
7160 dattr_cur = dattr_new;
7161 ndoms_cur = ndoms_new;
7163 register_sched_domain_sysctl();
7165 mutex_unlock(&sched_domains_mutex);
7168 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7171 * Update cpusets according to cpu_active mask. If cpusets are
7172 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7173 * around partition_sched_domains().
7175 * If we come here as part of a suspend/resume, don't touch cpusets because we
7176 * want to restore it back to its original state upon resume anyway.
7178 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7182 case CPU_ONLINE_FROZEN:
7183 case CPU_DOWN_FAILED_FROZEN:
7186 * num_cpus_frozen tracks how many CPUs are involved in suspend
7187 * resume sequence. As long as this is not the last online
7188 * operation in the resume sequence, just build a single sched
7189 * domain, ignoring cpusets.
7192 if (likely(num_cpus_frozen)) {
7193 partition_sched_domains(1, NULL, NULL);
7198 * This is the last CPU online operation. So fall through and
7199 * restore the original sched domains by considering the
7200 * cpuset configurations.
7204 cpuset_update_active_cpus(true);
7212 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7215 unsigned long flags;
7216 long cpu = (long)hcpu;
7222 case CPU_DOWN_PREPARE:
7223 rcu_read_lock_sched();
7224 dl_b = dl_bw_of(cpu);
7226 raw_spin_lock_irqsave(&dl_b->lock, flags);
7227 cpus = dl_bw_cpus(cpu);
7228 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7229 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7231 rcu_read_unlock_sched();
7234 return notifier_from_errno(-EBUSY);
7235 cpuset_update_active_cpus(false);
7237 case CPU_DOWN_PREPARE_FROZEN:
7239 partition_sched_domains(1, NULL, NULL);
7247 void __init sched_init_smp(void)
7249 cpumask_var_t non_isolated_cpus;
7251 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7252 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7254 /* nohz_full won't take effect without isolating the cpus. */
7255 tick_nohz_full_add_cpus_to(cpu_isolated_map);
7260 * There's no userspace yet to cause hotplug operations; hence all the
7261 * cpu masks are stable and all blatant races in the below code cannot
7264 mutex_lock(&sched_domains_mutex);
7265 init_sched_domains(cpu_active_mask);
7266 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7267 if (cpumask_empty(non_isolated_cpus))
7268 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7269 mutex_unlock(&sched_domains_mutex);
7271 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7272 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7273 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7277 /* Move init over to a non-isolated CPU */
7278 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7280 sched_init_granularity();
7281 free_cpumask_var(non_isolated_cpus);
7283 init_sched_rt_class();
7284 init_sched_dl_class();
7287 void __init sched_init_smp(void)
7289 sched_init_granularity();
7291 #endif /* CONFIG_SMP */
7293 int in_sched_functions(unsigned long addr)
7295 return in_lock_functions(addr) ||
7296 (addr >= (unsigned long)__sched_text_start
7297 && addr < (unsigned long)__sched_text_end);
7300 #ifdef CONFIG_CGROUP_SCHED
7302 * Default task group.
7303 * Every task in system belongs to this group at bootup.
7305 struct task_group root_task_group;
7306 LIST_HEAD(task_groups);
7309 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7311 void __init sched_init(void)
7314 unsigned long alloc_size = 0, ptr;
7316 #ifdef CONFIG_FAIR_GROUP_SCHED
7317 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7319 #ifdef CONFIG_RT_GROUP_SCHED
7320 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7323 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7325 #ifdef CONFIG_FAIR_GROUP_SCHED
7326 root_task_group.se = (struct sched_entity **)ptr;
7327 ptr += nr_cpu_ids * sizeof(void **);
7329 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7330 ptr += nr_cpu_ids * sizeof(void **);
7332 #endif /* CONFIG_FAIR_GROUP_SCHED */
7333 #ifdef CONFIG_RT_GROUP_SCHED
7334 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7335 ptr += nr_cpu_ids * sizeof(void **);
7337 root_task_group.rt_rq = (struct rt_rq **)ptr;
7338 ptr += nr_cpu_ids * sizeof(void **);
7340 #endif /* CONFIG_RT_GROUP_SCHED */
7342 #ifdef CONFIG_CPUMASK_OFFSTACK
7343 for_each_possible_cpu(i) {
7344 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7345 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7347 #endif /* CONFIG_CPUMASK_OFFSTACK */
7349 init_rt_bandwidth(&def_rt_bandwidth,
7350 global_rt_period(), global_rt_runtime());
7351 init_dl_bandwidth(&def_dl_bandwidth,
7352 global_rt_period(), global_rt_runtime());
7355 init_defrootdomain();
7358 #ifdef CONFIG_RT_GROUP_SCHED
7359 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7360 global_rt_period(), global_rt_runtime());
7361 #endif /* CONFIG_RT_GROUP_SCHED */
7363 #ifdef CONFIG_CGROUP_SCHED
7364 list_add(&root_task_group.list, &task_groups);
7365 INIT_LIST_HEAD(&root_task_group.children);
7366 INIT_LIST_HEAD(&root_task_group.siblings);
7367 autogroup_init(&init_task);
7369 #endif /* CONFIG_CGROUP_SCHED */
7371 for_each_possible_cpu(i) {
7375 raw_spin_lock_init(&rq->lock);
7377 rq->calc_load_active = 0;
7378 rq->calc_load_update = jiffies + LOAD_FREQ;
7379 init_cfs_rq(&rq->cfs);
7380 init_rt_rq(&rq->rt);
7381 init_dl_rq(&rq->dl);
7382 #ifdef CONFIG_FAIR_GROUP_SCHED
7383 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7384 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7386 * How much cpu bandwidth does root_task_group get?
7388 * In case of task-groups formed thr' the cgroup filesystem, it
7389 * gets 100% of the cpu resources in the system. This overall
7390 * system cpu resource is divided among the tasks of
7391 * root_task_group and its child task-groups in a fair manner,
7392 * based on each entity's (task or task-group's) weight
7393 * (se->load.weight).
7395 * In other words, if root_task_group has 10 tasks of weight
7396 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7397 * then A0's share of the cpu resource is:
7399 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7401 * We achieve this by letting root_task_group's tasks sit
7402 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7404 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7405 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7406 #endif /* CONFIG_FAIR_GROUP_SCHED */
7408 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7409 #ifdef CONFIG_RT_GROUP_SCHED
7410 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7413 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7414 rq->cpu_load[j] = 0;
7416 rq->last_load_update_tick = jiffies;
7421 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7422 rq->balance_callback = NULL;
7423 rq->active_balance = 0;
7424 rq->next_balance = jiffies;
7429 rq->avg_idle = 2*sysctl_sched_migration_cost;
7430 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7432 INIT_LIST_HEAD(&rq->cfs_tasks);
7434 rq_attach_root(rq, &def_root_domain);
7435 #ifdef CONFIG_NO_HZ_COMMON
7438 #ifdef CONFIG_NO_HZ_FULL
7439 rq->last_sched_tick = 0;
7443 atomic_set(&rq->nr_iowait, 0);
7446 set_load_weight(&init_task);
7448 #ifdef CONFIG_PREEMPT_NOTIFIERS
7449 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7453 * The boot idle thread does lazy MMU switching as well:
7455 atomic_inc(&init_mm.mm_count);
7456 enter_lazy_tlb(&init_mm, current);
7459 * During early bootup we pretend to be a normal task:
7461 current->sched_class = &fair_sched_class;
7464 * Make us the idle thread. Technically, schedule() should not be
7465 * called from this thread, however somewhere below it might be,
7466 * but because we are the idle thread, we just pick up running again
7467 * when this runqueue becomes "idle".
7469 init_idle(current, smp_processor_id());
7471 calc_load_update = jiffies + LOAD_FREQ;
7474 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7475 /* May be allocated at isolcpus cmdline parse time */
7476 if (cpu_isolated_map == NULL)
7477 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7478 idle_thread_set_boot_cpu();
7479 set_cpu_rq_start_time();
7481 init_sched_fair_class();
7483 scheduler_running = 1;
7486 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7487 static inline int preempt_count_equals(int preempt_offset)
7489 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7491 return (nested == preempt_offset);
7494 void __might_sleep(const char *file, int line, int preempt_offset)
7497 * Blocking primitives will set (and therefore destroy) current->state,
7498 * since we will exit with TASK_RUNNING make sure we enter with it,
7499 * otherwise we will destroy state.
7501 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7502 "do not call blocking ops when !TASK_RUNNING; "
7503 "state=%lx set at [<%p>] %pS\n",
7505 (void *)current->task_state_change,
7506 (void *)current->task_state_change);
7508 ___might_sleep(file, line, preempt_offset);
7510 EXPORT_SYMBOL(__might_sleep);
7512 void ___might_sleep(const char *file, int line, int preempt_offset)
7514 static unsigned long prev_jiffy; /* ratelimiting */
7516 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7517 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7518 !is_idle_task(current)) ||
7519 system_state != SYSTEM_RUNNING || oops_in_progress)
7521 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7523 prev_jiffy = jiffies;
7526 "BUG: sleeping function called from invalid context at %s:%d\n",
7529 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7530 in_atomic(), irqs_disabled(),
7531 current->pid, current->comm);
7533 if (task_stack_end_corrupted(current))
7534 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7536 debug_show_held_locks(current);
7537 if (irqs_disabled())
7538 print_irqtrace_events(current);
7539 #ifdef CONFIG_DEBUG_PREEMPT
7540 if (!preempt_count_equals(preempt_offset)) {
7541 pr_err("Preemption disabled at:");
7542 print_ip_sym(current->preempt_disable_ip);
7548 EXPORT_SYMBOL(___might_sleep);
7551 #ifdef CONFIG_MAGIC_SYSRQ
7552 void normalize_rt_tasks(void)
7554 struct task_struct *g, *p;
7555 struct sched_attr attr = {
7556 .sched_policy = SCHED_NORMAL,
7559 read_lock(&tasklist_lock);
7560 for_each_process_thread(g, p) {
7562 * Only normalize user tasks:
7564 if (p->flags & PF_KTHREAD)
7567 p->se.exec_start = 0;
7568 #ifdef CONFIG_SCHEDSTATS
7569 p->se.statistics.wait_start = 0;
7570 p->se.statistics.sleep_start = 0;
7571 p->se.statistics.block_start = 0;
7574 if (!dl_task(p) && !rt_task(p)) {
7576 * Renice negative nice level userspace
7579 if (task_nice(p) < 0)
7580 set_user_nice(p, 0);
7584 __sched_setscheduler(p, &attr, false, false);
7586 read_unlock(&tasklist_lock);
7589 #endif /* CONFIG_MAGIC_SYSRQ */
7591 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7593 * These functions are only useful for the IA64 MCA handling, or kdb.
7595 * They can only be called when the whole system has been
7596 * stopped - every CPU needs to be quiescent, and no scheduling
7597 * activity can take place. Using them for anything else would
7598 * be a serious bug, and as a result, they aren't even visible
7599 * under any other configuration.
7603 * curr_task - return the current task for a given cpu.
7604 * @cpu: the processor in question.
7606 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7608 * Return: The current task for @cpu.
7610 struct task_struct *curr_task(int cpu)
7612 return cpu_curr(cpu);
7615 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7619 * set_curr_task - set the current task for a given cpu.
7620 * @cpu: the processor in question.
7621 * @p: the task pointer to set.
7623 * Description: This function must only be used when non-maskable interrupts
7624 * are serviced on a separate stack. It allows the architecture to switch the
7625 * notion of the current task on a cpu in a non-blocking manner. This function
7626 * must be called with all CPU's synchronized, and interrupts disabled, the
7627 * and caller must save the original value of the current task (see
7628 * curr_task() above) and restore that value before reenabling interrupts and
7629 * re-starting the system.
7631 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7633 void set_curr_task(int cpu, struct task_struct *p)
7640 #ifdef CONFIG_CGROUP_SCHED
7641 /* task_group_lock serializes the addition/removal of task groups */
7642 static DEFINE_SPINLOCK(task_group_lock);
7644 static void free_sched_group(struct task_group *tg)
7646 free_fair_sched_group(tg);
7647 free_rt_sched_group(tg);
7652 /* allocate runqueue etc for a new task group */
7653 struct task_group *sched_create_group(struct task_group *parent)
7655 struct task_group *tg;
7657 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7659 return ERR_PTR(-ENOMEM);
7661 if (!alloc_fair_sched_group(tg, parent))
7664 if (!alloc_rt_sched_group(tg, parent))
7670 free_sched_group(tg);
7671 return ERR_PTR(-ENOMEM);
7674 void sched_online_group(struct task_group *tg, struct task_group *parent)
7676 unsigned long flags;
7678 spin_lock_irqsave(&task_group_lock, flags);
7679 list_add_rcu(&tg->list, &task_groups);
7681 WARN_ON(!parent); /* root should already exist */
7683 tg->parent = parent;
7684 INIT_LIST_HEAD(&tg->children);
7685 list_add_rcu(&tg->siblings, &parent->children);
7686 spin_unlock_irqrestore(&task_group_lock, flags);
7689 /* rcu callback to free various structures associated with a task group */
7690 static void free_sched_group_rcu(struct rcu_head *rhp)
7692 /* now it should be safe to free those cfs_rqs */
7693 free_sched_group(container_of(rhp, struct task_group, rcu));
7696 /* Destroy runqueue etc associated with a task group */
7697 void sched_destroy_group(struct task_group *tg)
7699 /* wait for possible concurrent references to cfs_rqs complete */
7700 call_rcu(&tg->rcu, free_sched_group_rcu);
7703 void sched_offline_group(struct task_group *tg)
7705 unsigned long flags;
7708 /* end participation in shares distribution */
7709 for_each_possible_cpu(i)
7710 unregister_fair_sched_group(tg, i);
7712 spin_lock_irqsave(&task_group_lock, flags);
7713 list_del_rcu(&tg->list);
7714 list_del_rcu(&tg->siblings);
7715 spin_unlock_irqrestore(&task_group_lock, flags);
7718 /* change task's runqueue when it moves between groups.
7719 * The caller of this function should have put the task in its new group
7720 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7721 * reflect its new group.
7723 void sched_move_task(struct task_struct *tsk)
7725 struct task_group *tg;
7726 int queued, running;
7727 unsigned long flags;
7730 rq = task_rq_lock(tsk, &flags);
7732 running = task_current(rq, tsk);
7733 queued = task_on_rq_queued(tsk);
7736 dequeue_task(rq, tsk, 0);
7737 if (unlikely(running))
7738 put_prev_task(rq, tsk);
7741 * All callers are synchronized by task_rq_lock(); we do not use RCU
7742 * which is pointless here. Thus, we pass "true" to task_css_check()
7743 * to prevent lockdep warnings.
7745 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7746 struct task_group, css);
7747 tg = autogroup_task_group(tsk, tg);
7748 tsk->sched_task_group = tg;
7750 #ifdef CONFIG_FAIR_GROUP_SCHED
7751 if (tsk->sched_class->task_move_group)
7752 tsk->sched_class->task_move_group(tsk);
7755 set_task_rq(tsk, task_cpu(tsk));
7757 if (unlikely(running))
7758 tsk->sched_class->set_curr_task(rq);
7760 enqueue_task(rq, tsk, 0);
7762 task_rq_unlock(rq, tsk, &flags);
7764 #endif /* CONFIG_CGROUP_SCHED */
7766 #ifdef CONFIG_RT_GROUP_SCHED
7768 * Ensure that the real time constraints are schedulable.
7770 static DEFINE_MUTEX(rt_constraints_mutex);
7772 /* Must be called with tasklist_lock held */
7773 static inline int tg_has_rt_tasks(struct task_group *tg)
7775 struct task_struct *g, *p;
7778 * Autogroups do not have RT tasks; see autogroup_create().
7780 if (task_group_is_autogroup(tg))
7783 for_each_process_thread(g, p) {
7784 if (rt_task(p) && task_group(p) == tg)
7791 struct rt_schedulable_data {
7792 struct task_group *tg;
7797 static int tg_rt_schedulable(struct task_group *tg, void *data)
7799 struct rt_schedulable_data *d = data;
7800 struct task_group *child;
7801 unsigned long total, sum = 0;
7802 u64 period, runtime;
7804 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7805 runtime = tg->rt_bandwidth.rt_runtime;
7808 period = d->rt_period;
7809 runtime = d->rt_runtime;
7813 * Cannot have more runtime than the period.
7815 if (runtime > period && runtime != RUNTIME_INF)
7819 * Ensure we don't starve existing RT tasks.
7821 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7824 total = to_ratio(period, runtime);
7827 * Nobody can have more than the global setting allows.
7829 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7833 * The sum of our children's runtime should not exceed our own.
7835 list_for_each_entry_rcu(child, &tg->children, siblings) {
7836 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7837 runtime = child->rt_bandwidth.rt_runtime;
7839 if (child == d->tg) {
7840 period = d->rt_period;
7841 runtime = d->rt_runtime;
7844 sum += to_ratio(period, runtime);
7853 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7857 struct rt_schedulable_data data = {
7859 .rt_period = period,
7860 .rt_runtime = runtime,
7864 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7870 static int tg_set_rt_bandwidth(struct task_group *tg,
7871 u64 rt_period, u64 rt_runtime)
7876 * Disallowing the root group RT runtime is BAD, it would disallow the
7877 * kernel creating (and or operating) RT threads.
7879 if (tg == &root_task_group && rt_runtime == 0)
7882 /* No period doesn't make any sense. */
7886 mutex_lock(&rt_constraints_mutex);
7887 read_lock(&tasklist_lock);
7888 err = __rt_schedulable(tg, rt_period, rt_runtime);
7892 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7893 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7894 tg->rt_bandwidth.rt_runtime = rt_runtime;
7896 for_each_possible_cpu(i) {
7897 struct rt_rq *rt_rq = tg->rt_rq[i];
7899 raw_spin_lock(&rt_rq->rt_runtime_lock);
7900 rt_rq->rt_runtime = rt_runtime;
7901 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7903 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7905 read_unlock(&tasklist_lock);
7906 mutex_unlock(&rt_constraints_mutex);
7911 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7913 u64 rt_runtime, rt_period;
7915 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7916 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7917 if (rt_runtime_us < 0)
7918 rt_runtime = RUNTIME_INF;
7920 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7923 static long sched_group_rt_runtime(struct task_group *tg)
7927 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7930 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7931 do_div(rt_runtime_us, NSEC_PER_USEC);
7932 return rt_runtime_us;
7935 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7937 u64 rt_runtime, rt_period;
7939 rt_period = rt_period_us * NSEC_PER_USEC;
7940 rt_runtime = tg->rt_bandwidth.rt_runtime;
7942 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7945 static long sched_group_rt_period(struct task_group *tg)
7949 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7950 do_div(rt_period_us, NSEC_PER_USEC);
7951 return rt_period_us;
7953 #endif /* CONFIG_RT_GROUP_SCHED */
7955 #ifdef CONFIG_RT_GROUP_SCHED
7956 static int sched_rt_global_constraints(void)
7960 mutex_lock(&rt_constraints_mutex);
7961 read_lock(&tasklist_lock);
7962 ret = __rt_schedulable(NULL, 0, 0);
7963 read_unlock(&tasklist_lock);
7964 mutex_unlock(&rt_constraints_mutex);
7969 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7971 /* Don't accept realtime tasks when there is no way for them to run */
7972 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7978 #else /* !CONFIG_RT_GROUP_SCHED */
7979 static int sched_rt_global_constraints(void)
7981 unsigned long flags;
7984 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7985 for_each_possible_cpu(i) {
7986 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7988 raw_spin_lock(&rt_rq->rt_runtime_lock);
7989 rt_rq->rt_runtime = global_rt_runtime();
7990 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7992 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7996 #endif /* CONFIG_RT_GROUP_SCHED */
7998 static int sched_dl_global_validate(void)
8000 u64 runtime = global_rt_runtime();
8001 u64 period = global_rt_period();
8002 u64 new_bw = to_ratio(period, runtime);
8005 unsigned long flags;
8008 * Here we want to check the bandwidth not being set to some
8009 * value smaller than the currently allocated bandwidth in
8010 * any of the root_domains.
8012 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8013 * cycling on root_domains... Discussion on different/better
8014 * solutions is welcome!
8016 for_each_possible_cpu(cpu) {
8017 rcu_read_lock_sched();
8018 dl_b = dl_bw_of(cpu);
8020 raw_spin_lock_irqsave(&dl_b->lock, flags);
8021 if (new_bw < dl_b->total_bw)
8023 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8025 rcu_read_unlock_sched();
8034 static void sched_dl_do_global(void)
8039 unsigned long flags;
8041 def_dl_bandwidth.dl_period = global_rt_period();
8042 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8044 if (global_rt_runtime() != RUNTIME_INF)
8045 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8048 * FIXME: As above...
8050 for_each_possible_cpu(cpu) {
8051 rcu_read_lock_sched();
8052 dl_b = dl_bw_of(cpu);
8054 raw_spin_lock_irqsave(&dl_b->lock, flags);
8056 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8058 rcu_read_unlock_sched();
8062 static int sched_rt_global_validate(void)
8064 if (sysctl_sched_rt_period <= 0)
8067 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8068 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8074 static void sched_rt_do_global(void)
8076 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8077 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8080 int sched_rt_handler(struct ctl_table *table, int write,
8081 void __user *buffer, size_t *lenp,
8084 int old_period, old_runtime;
8085 static DEFINE_MUTEX(mutex);
8089 old_period = sysctl_sched_rt_period;
8090 old_runtime = sysctl_sched_rt_runtime;
8092 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8094 if (!ret && write) {
8095 ret = sched_rt_global_validate();
8099 ret = sched_dl_global_validate();
8103 ret = sched_rt_global_constraints();
8107 sched_rt_do_global();
8108 sched_dl_do_global();
8112 sysctl_sched_rt_period = old_period;
8113 sysctl_sched_rt_runtime = old_runtime;
8115 mutex_unlock(&mutex);
8120 int sched_rr_handler(struct ctl_table *table, int write,
8121 void __user *buffer, size_t *lenp,
8125 static DEFINE_MUTEX(mutex);
8128 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8129 /* make sure that internally we keep jiffies */
8130 /* also, writing zero resets timeslice to default */
8131 if (!ret && write) {
8132 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8133 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8135 mutex_unlock(&mutex);
8139 #ifdef CONFIG_CGROUP_SCHED
8141 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8143 return css ? container_of(css, struct task_group, css) : NULL;
8146 static struct cgroup_subsys_state *
8147 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8149 struct task_group *parent = css_tg(parent_css);
8150 struct task_group *tg;
8153 /* This is early initialization for the top cgroup */
8154 return &root_task_group.css;
8157 tg = sched_create_group(parent);
8159 return ERR_PTR(-ENOMEM);
8164 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8166 struct task_group *tg = css_tg(css);
8167 struct task_group *parent = css_tg(css->parent);
8170 sched_online_group(tg, parent);
8174 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8176 struct task_group *tg = css_tg(css);
8178 sched_destroy_group(tg);
8181 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8183 struct task_group *tg = css_tg(css);
8185 sched_offline_group(tg);
8188 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8190 sched_move_task(task);
8193 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8194 struct cgroup_taskset *tset)
8196 struct task_struct *task;
8198 cgroup_taskset_for_each(task, tset) {
8199 #ifdef CONFIG_RT_GROUP_SCHED
8200 if (!sched_rt_can_attach(css_tg(css), task))
8203 /* We don't support RT-tasks being in separate groups */
8204 if (task->sched_class != &fair_sched_class)
8211 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8212 struct cgroup_taskset *tset)
8214 struct task_struct *task;
8216 cgroup_taskset_for_each(task, tset)
8217 sched_move_task(task);
8220 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8221 struct cgroup_subsys_state *old_css,
8222 struct task_struct *task)
8224 sched_move_task(task);
8227 #ifdef CONFIG_FAIR_GROUP_SCHED
8228 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8229 struct cftype *cftype, u64 shareval)
8231 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8234 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8237 struct task_group *tg = css_tg(css);
8239 return (u64) scale_load_down(tg->shares);
8242 #ifdef CONFIG_CFS_BANDWIDTH
8243 static DEFINE_MUTEX(cfs_constraints_mutex);
8245 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8246 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8248 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8250 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8252 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8253 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8255 if (tg == &root_task_group)
8259 * Ensure we have at some amount of bandwidth every period. This is
8260 * to prevent reaching a state of large arrears when throttled via
8261 * entity_tick() resulting in prolonged exit starvation.
8263 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8267 * Likewise, bound things on the otherside by preventing insane quota
8268 * periods. This also allows us to normalize in computing quota
8271 if (period > max_cfs_quota_period)
8275 * Prevent race between setting of cfs_rq->runtime_enabled and
8276 * unthrottle_offline_cfs_rqs().
8279 mutex_lock(&cfs_constraints_mutex);
8280 ret = __cfs_schedulable(tg, period, quota);
8284 runtime_enabled = quota != RUNTIME_INF;
8285 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8287 * If we need to toggle cfs_bandwidth_used, off->on must occur
8288 * before making related changes, and on->off must occur afterwards
8290 if (runtime_enabled && !runtime_was_enabled)
8291 cfs_bandwidth_usage_inc();
8292 raw_spin_lock_irq(&cfs_b->lock);
8293 cfs_b->period = ns_to_ktime(period);
8294 cfs_b->quota = quota;
8296 __refill_cfs_bandwidth_runtime(cfs_b);
8297 /* restart the period timer (if active) to handle new period expiry */
8298 if (runtime_enabled)
8299 start_cfs_bandwidth(cfs_b);
8300 raw_spin_unlock_irq(&cfs_b->lock);
8302 for_each_online_cpu(i) {
8303 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8304 struct rq *rq = cfs_rq->rq;
8306 raw_spin_lock_irq(&rq->lock);
8307 cfs_rq->runtime_enabled = runtime_enabled;
8308 cfs_rq->runtime_remaining = 0;
8310 if (cfs_rq->throttled)
8311 unthrottle_cfs_rq(cfs_rq);
8312 raw_spin_unlock_irq(&rq->lock);
8314 if (runtime_was_enabled && !runtime_enabled)
8315 cfs_bandwidth_usage_dec();
8317 mutex_unlock(&cfs_constraints_mutex);
8323 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8327 period = ktime_to_ns(tg->cfs_bandwidth.period);
8328 if (cfs_quota_us < 0)
8329 quota = RUNTIME_INF;
8331 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8333 return tg_set_cfs_bandwidth(tg, period, quota);
8336 long tg_get_cfs_quota(struct task_group *tg)
8340 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8343 quota_us = tg->cfs_bandwidth.quota;
8344 do_div(quota_us, NSEC_PER_USEC);
8349 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8353 period = (u64)cfs_period_us * NSEC_PER_USEC;
8354 quota = tg->cfs_bandwidth.quota;
8356 return tg_set_cfs_bandwidth(tg, period, quota);
8359 long tg_get_cfs_period(struct task_group *tg)
8363 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8364 do_div(cfs_period_us, NSEC_PER_USEC);
8366 return cfs_period_us;
8369 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8372 return tg_get_cfs_quota(css_tg(css));
8375 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8376 struct cftype *cftype, s64 cfs_quota_us)
8378 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8381 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8384 return tg_get_cfs_period(css_tg(css));
8387 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8388 struct cftype *cftype, u64 cfs_period_us)
8390 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8393 struct cfs_schedulable_data {
8394 struct task_group *tg;
8399 * normalize group quota/period to be quota/max_period
8400 * note: units are usecs
8402 static u64 normalize_cfs_quota(struct task_group *tg,
8403 struct cfs_schedulable_data *d)
8411 period = tg_get_cfs_period(tg);
8412 quota = tg_get_cfs_quota(tg);
8415 /* note: these should typically be equivalent */
8416 if (quota == RUNTIME_INF || quota == -1)
8419 return to_ratio(period, quota);
8422 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8424 struct cfs_schedulable_data *d = data;
8425 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8426 s64 quota = 0, parent_quota = -1;
8429 quota = RUNTIME_INF;
8431 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8433 quota = normalize_cfs_quota(tg, d);
8434 parent_quota = parent_b->hierarchical_quota;
8437 * ensure max(child_quota) <= parent_quota, inherit when no
8440 if (quota == RUNTIME_INF)
8441 quota = parent_quota;
8442 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8445 cfs_b->hierarchical_quota = quota;
8450 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8453 struct cfs_schedulable_data data = {
8459 if (quota != RUNTIME_INF) {
8460 do_div(data.period, NSEC_PER_USEC);
8461 do_div(data.quota, NSEC_PER_USEC);
8465 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8471 static int cpu_stats_show(struct seq_file *sf, void *v)
8473 struct task_group *tg = css_tg(seq_css(sf));
8474 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8476 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8477 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8478 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8482 #endif /* CONFIG_CFS_BANDWIDTH */
8483 #endif /* CONFIG_FAIR_GROUP_SCHED */
8485 #ifdef CONFIG_RT_GROUP_SCHED
8486 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8487 struct cftype *cft, s64 val)
8489 return sched_group_set_rt_runtime(css_tg(css), val);
8492 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8495 return sched_group_rt_runtime(css_tg(css));
8498 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8499 struct cftype *cftype, u64 rt_period_us)
8501 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8504 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8507 return sched_group_rt_period(css_tg(css));
8509 #endif /* CONFIG_RT_GROUP_SCHED */
8511 static struct cftype cpu_files[] = {
8512 #ifdef CONFIG_FAIR_GROUP_SCHED
8515 .read_u64 = cpu_shares_read_u64,
8516 .write_u64 = cpu_shares_write_u64,
8519 #ifdef CONFIG_CFS_BANDWIDTH
8521 .name = "cfs_quota_us",
8522 .read_s64 = cpu_cfs_quota_read_s64,
8523 .write_s64 = cpu_cfs_quota_write_s64,
8526 .name = "cfs_period_us",
8527 .read_u64 = cpu_cfs_period_read_u64,
8528 .write_u64 = cpu_cfs_period_write_u64,
8532 .seq_show = cpu_stats_show,
8535 #ifdef CONFIG_RT_GROUP_SCHED
8537 .name = "rt_runtime_us",
8538 .read_s64 = cpu_rt_runtime_read,
8539 .write_s64 = cpu_rt_runtime_write,
8542 .name = "rt_period_us",
8543 .read_u64 = cpu_rt_period_read_uint,
8544 .write_u64 = cpu_rt_period_write_uint,
8550 struct cgroup_subsys cpu_cgrp_subsys = {
8551 .css_alloc = cpu_cgroup_css_alloc,
8552 .css_free = cpu_cgroup_css_free,
8553 .css_online = cpu_cgroup_css_online,
8554 .css_offline = cpu_cgroup_css_offline,
8555 .fork = cpu_cgroup_fork,
8556 .can_attach = cpu_cgroup_can_attach,
8557 .attach = cpu_cgroup_attach,
8558 .exit = cpu_cgroup_exit,
8559 .legacy_cftypes = cpu_files,
8563 #endif /* CONFIG_CGROUP_SCHED */
8565 void dump_cpu_task(int cpu)
8567 pr_info("Task dump for CPU %d:\n", cpu);
8568 sched_show_task(cpu_curr(cpu));