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)) {
633 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
640 if (!is_housekeeping_cpu(cpu))
641 cpu = housekeeping_any_cpu();
647 * When add_timer_on() enqueues a timer into the timer wheel of an
648 * idle CPU then this timer might expire before the next timer event
649 * which is scheduled to wake up that CPU. In case of a completely
650 * idle system the next event might even be infinite time into the
651 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
652 * leaves the inner idle loop so the newly added timer is taken into
653 * account when the CPU goes back to idle and evaluates the timer
654 * wheel for the next timer event.
656 static void wake_up_idle_cpu(int cpu)
658 struct rq *rq = cpu_rq(cpu);
660 if (cpu == smp_processor_id())
663 if (set_nr_and_not_polling(rq->idle))
664 smp_send_reschedule(cpu);
666 trace_sched_wake_idle_without_ipi(cpu);
669 static bool wake_up_full_nohz_cpu(int cpu)
672 * We just need the target to call irq_exit() and re-evaluate
673 * the next tick. The nohz full kick at least implies that.
674 * If needed we can still optimize that later with an
677 if (tick_nohz_full_cpu(cpu)) {
678 if (cpu != smp_processor_id() ||
679 tick_nohz_tick_stopped())
680 tick_nohz_full_kick_cpu(cpu);
687 void wake_up_nohz_cpu(int cpu)
689 if (!wake_up_full_nohz_cpu(cpu))
690 wake_up_idle_cpu(cpu);
693 static inline bool got_nohz_idle_kick(void)
695 int cpu = smp_processor_id();
697 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
700 if (idle_cpu(cpu) && !need_resched())
704 * We can't run Idle Load Balance on this CPU for this time so we
705 * cancel it and clear NOHZ_BALANCE_KICK
707 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
711 #else /* CONFIG_NO_HZ_COMMON */
713 static inline bool got_nohz_idle_kick(void)
718 #endif /* CONFIG_NO_HZ_COMMON */
720 #ifdef CONFIG_NO_HZ_FULL
721 bool sched_can_stop_tick(void)
724 * FIFO realtime policy runs the highest priority task. Other runnable
725 * tasks are of a lower priority. The scheduler tick does nothing.
727 if (current->policy == SCHED_FIFO)
731 * Round-robin realtime tasks time slice with other tasks at the same
732 * realtime priority. Is this task the only one at this priority?
734 if (current->policy == SCHED_RR) {
735 struct sched_rt_entity *rt_se = ¤t->rt;
737 return rt_se->run_list.prev == rt_se->run_list.next;
741 * More than one running task need preemption.
742 * nr_running update is assumed to be visible
743 * after IPI is sent from wakers.
745 if (this_rq()->nr_running > 1)
750 #endif /* CONFIG_NO_HZ_FULL */
752 void sched_avg_update(struct rq *rq)
754 s64 period = sched_avg_period();
756 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
758 * Inline assembly required to prevent the compiler
759 * optimising this loop into a divmod call.
760 * See __iter_div_u64_rem() for another example of this.
762 asm("" : "+rm" (rq->age_stamp));
763 rq->age_stamp += period;
768 #endif /* CONFIG_SMP */
770 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
771 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
773 * Iterate task_group tree rooted at *from, calling @down when first entering a
774 * node and @up when leaving it for the final time.
776 * Caller must hold rcu_lock or sufficient equivalent.
778 int walk_tg_tree_from(struct task_group *from,
779 tg_visitor down, tg_visitor up, void *data)
781 struct task_group *parent, *child;
787 ret = (*down)(parent, data);
790 list_for_each_entry_rcu(child, &parent->children, siblings) {
797 ret = (*up)(parent, data);
798 if (ret || parent == from)
802 parent = parent->parent;
809 int tg_nop(struct task_group *tg, void *data)
815 static void set_load_weight(struct task_struct *p)
817 int prio = p->static_prio - MAX_RT_PRIO;
818 struct load_weight *load = &p->se.load;
821 * SCHED_IDLE tasks get minimal weight:
823 if (idle_policy(p->policy)) {
824 load->weight = scale_load(WEIGHT_IDLEPRIO);
825 load->inv_weight = WMULT_IDLEPRIO;
829 load->weight = scale_load(prio_to_weight[prio]);
830 load->inv_weight = prio_to_wmult[prio];
833 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
836 if (!(flags & ENQUEUE_RESTORE))
837 sched_info_queued(rq, p);
838 p->sched_class->enqueue_task(rq, p, flags);
841 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
844 if (!(flags & DEQUEUE_SAVE))
845 sched_info_dequeued(rq, p);
846 p->sched_class->dequeue_task(rq, p, flags);
849 void activate_task(struct rq *rq, struct task_struct *p, int flags)
851 if (task_contributes_to_load(p))
852 rq->nr_uninterruptible--;
854 enqueue_task(rq, p, flags);
857 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
859 if (task_contributes_to_load(p))
860 rq->nr_uninterruptible++;
862 dequeue_task(rq, p, flags);
865 static void update_rq_clock_task(struct rq *rq, s64 delta)
868 * In theory, the compile should just see 0 here, and optimize out the call
869 * to sched_rt_avg_update. But I don't trust it...
871 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
872 s64 steal = 0, irq_delta = 0;
874 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
875 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
878 * Since irq_time is only updated on {soft,}irq_exit, we might run into
879 * this case when a previous update_rq_clock() happened inside a
882 * When this happens, we stop ->clock_task and only update the
883 * prev_irq_time stamp to account for the part that fit, so that a next
884 * update will consume the rest. This ensures ->clock_task is
887 * It does however cause some slight miss-attribution of {soft,}irq
888 * time, a more accurate solution would be to update the irq_time using
889 * the current rq->clock timestamp, except that would require using
892 if (irq_delta > delta)
895 rq->prev_irq_time += irq_delta;
898 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
899 if (static_key_false((¶virt_steal_rq_enabled))) {
900 steal = paravirt_steal_clock(cpu_of(rq));
901 steal -= rq->prev_steal_time_rq;
903 if (unlikely(steal > delta))
906 rq->prev_steal_time_rq += steal;
911 rq->clock_task += delta;
913 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
914 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
915 sched_rt_avg_update(rq, irq_delta + steal);
919 void sched_set_stop_task(int cpu, struct task_struct *stop)
921 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
922 struct task_struct *old_stop = cpu_rq(cpu)->stop;
926 * Make it appear like a SCHED_FIFO task, its something
927 * userspace knows about and won't get confused about.
929 * Also, it will make PI more or less work without too
930 * much confusion -- but then, stop work should not
931 * rely on PI working anyway.
933 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
935 stop->sched_class = &stop_sched_class;
938 cpu_rq(cpu)->stop = stop;
942 * Reset it back to a normal scheduling class so that
943 * it can die in pieces.
945 old_stop->sched_class = &rt_sched_class;
950 * __normal_prio - return the priority that is based on the static prio
952 static inline int __normal_prio(struct task_struct *p)
954 return p->static_prio;
958 * Calculate the expected normal priority: i.e. priority
959 * without taking RT-inheritance into account. Might be
960 * boosted by interactivity modifiers. Changes upon fork,
961 * setprio syscalls, and whenever the interactivity
962 * estimator recalculates.
964 static inline int normal_prio(struct task_struct *p)
968 if (task_has_dl_policy(p))
969 prio = MAX_DL_PRIO-1;
970 else if (task_has_rt_policy(p))
971 prio = MAX_RT_PRIO-1 - p->rt_priority;
973 prio = __normal_prio(p);
978 * Calculate the current priority, i.e. the priority
979 * taken into account by the scheduler. This value might
980 * be boosted by RT tasks, or might be boosted by
981 * interactivity modifiers. Will be RT if the task got
982 * RT-boosted. If not then it returns p->normal_prio.
984 static int effective_prio(struct task_struct *p)
986 p->normal_prio = normal_prio(p);
988 * If we are RT tasks or we were boosted to RT priority,
989 * keep the priority unchanged. Otherwise, update priority
990 * to the normal priority:
992 if (!rt_prio(p->prio))
993 return p->normal_prio;
998 * task_curr - is this task currently executing on a CPU?
999 * @p: the task in question.
1001 * Return: 1 if the task is currently executing. 0 otherwise.
1003 inline int task_curr(const struct task_struct *p)
1005 return cpu_curr(task_cpu(p)) == p;
1009 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1010 * use the balance_callback list if you want balancing.
1012 * this means any call to check_class_changed() must be followed by a call to
1013 * balance_callback().
1015 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1016 const struct sched_class *prev_class,
1019 if (prev_class != p->sched_class) {
1020 if (prev_class->switched_from)
1021 prev_class->switched_from(rq, p);
1023 p->sched_class->switched_to(rq, p);
1024 } else if (oldprio != p->prio || dl_task(p))
1025 p->sched_class->prio_changed(rq, p, oldprio);
1028 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1030 const struct sched_class *class;
1032 if (p->sched_class == rq->curr->sched_class) {
1033 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1035 for_each_class(class) {
1036 if (class == rq->curr->sched_class)
1038 if (class == p->sched_class) {
1046 * A queue event has occurred, and we're going to schedule. In
1047 * this case, we can save a useless back to back clock update.
1049 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1050 rq_clock_skip_update(rq, true);
1055 * This is how migration works:
1057 * 1) we invoke migration_cpu_stop() on the target CPU using
1059 * 2) stopper starts to run (implicitly forcing the migrated thread
1061 * 3) it checks whether the migrated task is still in the wrong runqueue.
1062 * 4) if it's in the wrong runqueue then the migration thread removes
1063 * it and puts it into the right queue.
1064 * 5) stopper completes and stop_one_cpu() returns and the migration
1069 * move_queued_task - move a queued task to new rq.
1071 * Returns (locked) new rq. Old rq's lock is released.
1073 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1075 lockdep_assert_held(&rq->lock);
1077 dequeue_task(rq, p, 0);
1078 p->on_rq = TASK_ON_RQ_MIGRATING;
1079 set_task_cpu(p, new_cpu);
1080 raw_spin_unlock(&rq->lock);
1082 rq = cpu_rq(new_cpu);
1084 raw_spin_lock(&rq->lock);
1085 BUG_ON(task_cpu(p) != new_cpu);
1086 p->on_rq = TASK_ON_RQ_QUEUED;
1087 enqueue_task(rq, p, 0);
1088 check_preempt_curr(rq, p, 0);
1093 struct migration_arg {
1094 struct task_struct *task;
1099 * Move (not current) task off this cpu, onto dest cpu. We're doing
1100 * this because either it can't run here any more (set_cpus_allowed()
1101 * away from this CPU, or CPU going down), or because we're
1102 * attempting to rebalance this task on exec (sched_exec).
1104 * So we race with normal scheduler movements, but that's OK, as long
1105 * as the task is no longer on this CPU.
1107 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1109 if (unlikely(!cpu_active(dest_cpu)))
1112 /* Affinity changed (again). */
1113 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1116 rq = move_queued_task(rq, p, dest_cpu);
1122 * migration_cpu_stop - this will be executed by a highprio stopper thread
1123 * and performs thread migration by bumping thread off CPU then
1124 * 'pushing' onto another runqueue.
1126 static int migration_cpu_stop(void *data)
1128 struct migration_arg *arg = data;
1129 struct task_struct *p = arg->task;
1130 struct rq *rq = this_rq();
1133 * The original target cpu might have gone down and we might
1134 * be on another cpu but it doesn't matter.
1136 local_irq_disable();
1138 * We need to explicitly wake pending tasks before running
1139 * __migrate_task() such that we will not miss enforcing cpus_allowed
1140 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1142 sched_ttwu_pending();
1144 raw_spin_lock(&p->pi_lock);
1145 raw_spin_lock(&rq->lock);
1147 * If task_rq(p) != rq, it cannot be migrated here, because we're
1148 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1149 * we're holding p->pi_lock.
1151 if (task_rq(p) == rq && task_on_rq_queued(p))
1152 rq = __migrate_task(rq, p, arg->dest_cpu);
1153 raw_spin_unlock(&rq->lock);
1154 raw_spin_unlock(&p->pi_lock);
1161 * sched_class::set_cpus_allowed must do the below, but is not required to
1162 * actually call this function.
1164 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1166 cpumask_copy(&p->cpus_allowed, new_mask);
1167 p->nr_cpus_allowed = cpumask_weight(new_mask);
1170 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1172 struct rq *rq = task_rq(p);
1173 bool queued, running;
1175 lockdep_assert_held(&p->pi_lock);
1177 queued = task_on_rq_queued(p);
1178 running = task_current(rq, p);
1182 * Because __kthread_bind() calls this on blocked tasks without
1185 lockdep_assert_held(&rq->lock);
1186 dequeue_task(rq, p, DEQUEUE_SAVE);
1189 put_prev_task(rq, p);
1191 p->sched_class->set_cpus_allowed(p, new_mask);
1194 p->sched_class->set_curr_task(rq);
1196 enqueue_task(rq, p, ENQUEUE_RESTORE);
1200 * Change a given task's CPU affinity. Migrate the thread to a
1201 * proper CPU and schedule it away if the CPU it's executing on
1202 * is removed from the allowed bitmask.
1204 * NOTE: the caller must have a valid reference to the task, the
1205 * task must not exit() & deallocate itself prematurely. The
1206 * call is not atomic; no spinlocks may be held.
1208 static int __set_cpus_allowed_ptr(struct task_struct *p,
1209 const struct cpumask *new_mask, bool check)
1211 unsigned long flags;
1213 unsigned int dest_cpu;
1216 rq = task_rq_lock(p, &flags);
1219 * Must re-check here, to close a race against __kthread_bind(),
1220 * sched_setaffinity() is not guaranteed to observe the flag.
1222 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1227 if (cpumask_equal(&p->cpus_allowed, new_mask))
1230 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1235 do_set_cpus_allowed(p, new_mask);
1237 /* Can the task run on the task's current CPU? If so, we're done */
1238 if (cpumask_test_cpu(task_cpu(p), new_mask))
1241 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1242 if (task_running(rq, p) || p->state == TASK_WAKING) {
1243 struct migration_arg arg = { p, dest_cpu };
1244 /* Need help from migration thread: drop lock and wait. */
1245 task_rq_unlock(rq, p, &flags);
1246 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1247 tlb_migrate_finish(p->mm);
1249 } else if (task_on_rq_queued(p)) {
1251 * OK, since we're going to drop the lock immediately
1252 * afterwards anyway.
1254 lockdep_unpin_lock(&rq->lock);
1255 rq = move_queued_task(rq, p, dest_cpu);
1256 lockdep_pin_lock(&rq->lock);
1259 task_rq_unlock(rq, p, &flags);
1264 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1266 return __set_cpus_allowed_ptr(p, new_mask, false);
1268 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1270 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1272 #ifdef CONFIG_SCHED_DEBUG
1274 * We should never call set_task_cpu() on a blocked task,
1275 * ttwu() will sort out the placement.
1277 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1280 #ifdef CONFIG_LOCKDEP
1282 * The caller should hold either p->pi_lock or rq->lock, when changing
1283 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1285 * sched_move_task() holds both and thus holding either pins the cgroup,
1288 * Furthermore, all task_rq users should acquire both locks, see
1291 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1292 lockdep_is_held(&task_rq(p)->lock)));
1296 trace_sched_migrate_task(p, new_cpu);
1298 if (task_cpu(p) != new_cpu) {
1299 if (p->sched_class->migrate_task_rq)
1300 p->sched_class->migrate_task_rq(p);
1301 p->se.nr_migrations++;
1302 perf_event_task_migrate(p);
1305 __set_task_cpu(p, new_cpu);
1308 static void __migrate_swap_task(struct task_struct *p, int cpu)
1310 if (task_on_rq_queued(p)) {
1311 struct rq *src_rq, *dst_rq;
1313 src_rq = task_rq(p);
1314 dst_rq = cpu_rq(cpu);
1316 deactivate_task(src_rq, p, 0);
1317 set_task_cpu(p, cpu);
1318 activate_task(dst_rq, p, 0);
1319 check_preempt_curr(dst_rq, p, 0);
1322 * Task isn't running anymore; make it appear like we migrated
1323 * it before it went to sleep. This means on wakeup we make the
1324 * previous cpu our targer instead of where it really is.
1330 struct migration_swap_arg {
1331 struct task_struct *src_task, *dst_task;
1332 int src_cpu, dst_cpu;
1335 static int migrate_swap_stop(void *data)
1337 struct migration_swap_arg *arg = data;
1338 struct rq *src_rq, *dst_rq;
1341 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1344 src_rq = cpu_rq(arg->src_cpu);
1345 dst_rq = cpu_rq(arg->dst_cpu);
1347 double_raw_lock(&arg->src_task->pi_lock,
1348 &arg->dst_task->pi_lock);
1349 double_rq_lock(src_rq, dst_rq);
1351 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1354 if (task_cpu(arg->src_task) != arg->src_cpu)
1357 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1360 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1363 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1364 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1369 double_rq_unlock(src_rq, dst_rq);
1370 raw_spin_unlock(&arg->dst_task->pi_lock);
1371 raw_spin_unlock(&arg->src_task->pi_lock);
1377 * Cross migrate two tasks
1379 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1381 struct migration_swap_arg arg;
1384 arg = (struct migration_swap_arg){
1386 .src_cpu = task_cpu(cur),
1388 .dst_cpu = task_cpu(p),
1391 if (arg.src_cpu == arg.dst_cpu)
1395 * These three tests are all lockless; this is OK since all of them
1396 * will be re-checked with proper locks held further down the line.
1398 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1401 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1404 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1407 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1408 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1415 * wait_task_inactive - wait for a thread to unschedule.
1417 * If @match_state is nonzero, it's the @p->state value just checked and
1418 * not expected to change. If it changes, i.e. @p might have woken up,
1419 * then return zero. When we succeed in waiting for @p to be off its CPU,
1420 * we return a positive number (its total switch count). If a second call
1421 * a short while later returns the same number, the caller can be sure that
1422 * @p has remained unscheduled the whole time.
1424 * The caller must ensure that the task *will* unschedule sometime soon,
1425 * else this function might spin for a *long* time. This function can't
1426 * be called with interrupts off, or it may introduce deadlock with
1427 * smp_call_function() if an IPI is sent by the same process we are
1428 * waiting to become inactive.
1430 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1432 unsigned long flags;
1433 int running, queued;
1439 * We do the initial early heuristics without holding
1440 * any task-queue locks at all. We'll only try to get
1441 * the runqueue lock when things look like they will
1447 * If the task is actively running on another CPU
1448 * still, just relax and busy-wait without holding
1451 * NOTE! Since we don't hold any locks, it's not
1452 * even sure that "rq" stays as the right runqueue!
1453 * But we don't care, since "task_running()" will
1454 * return false if the runqueue has changed and p
1455 * is actually now running somewhere else!
1457 while (task_running(rq, p)) {
1458 if (match_state && unlikely(p->state != match_state))
1464 * Ok, time to look more closely! We need the rq
1465 * lock now, to be *sure*. If we're wrong, we'll
1466 * just go back and repeat.
1468 rq = task_rq_lock(p, &flags);
1469 trace_sched_wait_task(p);
1470 running = task_running(rq, p);
1471 queued = task_on_rq_queued(p);
1473 if (!match_state || p->state == match_state)
1474 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1475 task_rq_unlock(rq, p, &flags);
1478 * If it changed from the expected state, bail out now.
1480 if (unlikely(!ncsw))
1484 * Was it really running after all now that we
1485 * checked with the proper locks actually held?
1487 * Oops. Go back and try again..
1489 if (unlikely(running)) {
1495 * It's not enough that it's not actively running,
1496 * it must be off the runqueue _entirely_, and not
1499 * So if it was still runnable (but just not actively
1500 * running right now), it's preempted, and we should
1501 * yield - it could be a while.
1503 if (unlikely(queued)) {
1504 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1506 set_current_state(TASK_UNINTERRUPTIBLE);
1507 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1512 * Ahh, all good. It wasn't running, and it wasn't
1513 * runnable, which means that it will never become
1514 * running in the future either. We're all done!
1523 * kick_process - kick a running thread to enter/exit the kernel
1524 * @p: the to-be-kicked thread
1526 * Cause a process which is running on another CPU to enter
1527 * kernel-mode, without any delay. (to get signals handled.)
1529 * NOTE: this function doesn't have to take the runqueue lock,
1530 * because all it wants to ensure is that the remote task enters
1531 * the kernel. If the IPI races and the task has been migrated
1532 * to another CPU then no harm is done and the purpose has been
1535 void kick_process(struct task_struct *p)
1541 if ((cpu != smp_processor_id()) && task_curr(p))
1542 smp_send_reschedule(cpu);
1545 EXPORT_SYMBOL_GPL(kick_process);
1548 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1550 static int select_fallback_rq(int cpu, struct task_struct *p)
1552 int nid = cpu_to_node(cpu);
1553 const struct cpumask *nodemask = NULL;
1554 enum { cpuset, possible, fail } state = cpuset;
1558 * If the node that the cpu is on has been offlined, cpu_to_node()
1559 * will return -1. There is no cpu on the node, and we should
1560 * select the cpu on the other node.
1563 nodemask = cpumask_of_node(nid);
1565 /* Look for allowed, online CPU in same node. */
1566 for_each_cpu(dest_cpu, nodemask) {
1567 if (!cpu_online(dest_cpu))
1569 if (!cpu_active(dest_cpu))
1571 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1577 /* Any allowed, online CPU? */
1578 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1579 if (!cpu_online(dest_cpu))
1581 if (!cpu_active(dest_cpu))
1586 /* No more Mr. Nice Guy. */
1589 if (IS_ENABLED(CONFIG_CPUSETS)) {
1590 cpuset_cpus_allowed_fallback(p);
1596 do_set_cpus_allowed(p, cpu_possible_mask);
1607 if (state != cpuset) {
1609 * Don't tell them about moving exiting tasks or
1610 * kernel threads (both mm NULL), since they never
1613 if (p->mm && printk_ratelimit()) {
1614 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1615 task_pid_nr(p), p->comm, cpu);
1623 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1626 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1628 lockdep_assert_held(&p->pi_lock);
1630 if (p->nr_cpus_allowed > 1)
1631 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1634 * In order not to call set_task_cpu() on a blocking task we need
1635 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1638 * Since this is common to all placement strategies, this lives here.
1640 * [ this allows ->select_task() to simply return task_cpu(p) and
1641 * not worry about this generic constraint ]
1643 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1645 cpu = select_fallback_rq(task_cpu(p), p);
1650 static void update_avg(u64 *avg, u64 sample)
1652 s64 diff = sample - *avg;
1658 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1659 const struct cpumask *new_mask, bool check)
1661 return set_cpus_allowed_ptr(p, new_mask);
1664 #endif /* CONFIG_SMP */
1667 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1669 #ifdef CONFIG_SCHEDSTATS
1670 struct rq *rq = this_rq();
1673 int this_cpu = smp_processor_id();
1675 if (cpu == this_cpu) {
1676 schedstat_inc(rq, ttwu_local);
1677 schedstat_inc(p, se.statistics.nr_wakeups_local);
1679 struct sched_domain *sd;
1681 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1683 for_each_domain(this_cpu, sd) {
1684 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1685 schedstat_inc(sd, ttwu_wake_remote);
1692 if (wake_flags & WF_MIGRATED)
1693 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1695 #endif /* CONFIG_SMP */
1697 schedstat_inc(rq, ttwu_count);
1698 schedstat_inc(p, se.statistics.nr_wakeups);
1700 if (wake_flags & WF_SYNC)
1701 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1703 #endif /* CONFIG_SCHEDSTATS */
1706 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1708 activate_task(rq, p, en_flags);
1709 p->on_rq = TASK_ON_RQ_QUEUED;
1711 /* if a worker is waking up, notify workqueue */
1712 if (p->flags & PF_WQ_WORKER)
1713 wq_worker_waking_up(p, cpu_of(rq));
1717 * Mark the task runnable and perform wakeup-preemption.
1720 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1722 check_preempt_curr(rq, p, wake_flags);
1723 p->state = TASK_RUNNING;
1724 trace_sched_wakeup(p);
1727 if (p->sched_class->task_woken) {
1729 * Our task @p is fully woken up and running; so its safe to
1730 * drop the rq->lock, hereafter rq is only used for statistics.
1732 lockdep_unpin_lock(&rq->lock);
1733 p->sched_class->task_woken(rq, p);
1734 lockdep_pin_lock(&rq->lock);
1737 if (rq->idle_stamp) {
1738 u64 delta = rq_clock(rq) - rq->idle_stamp;
1739 u64 max = 2*rq->max_idle_balance_cost;
1741 update_avg(&rq->avg_idle, delta);
1743 if (rq->avg_idle > max)
1752 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1754 lockdep_assert_held(&rq->lock);
1757 if (p->sched_contributes_to_load)
1758 rq->nr_uninterruptible--;
1761 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1762 ttwu_do_wakeup(rq, p, wake_flags);
1766 * Called in case the task @p isn't fully descheduled from its runqueue,
1767 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1768 * since all we need to do is flip p->state to TASK_RUNNING, since
1769 * the task is still ->on_rq.
1771 static int ttwu_remote(struct task_struct *p, int wake_flags)
1776 rq = __task_rq_lock(p);
1777 if (task_on_rq_queued(p)) {
1778 /* check_preempt_curr() may use rq clock */
1779 update_rq_clock(rq);
1780 ttwu_do_wakeup(rq, p, wake_flags);
1783 __task_rq_unlock(rq);
1789 void sched_ttwu_pending(void)
1791 struct rq *rq = this_rq();
1792 struct llist_node *llist = llist_del_all(&rq->wake_list);
1793 struct task_struct *p;
1794 unsigned long flags;
1799 raw_spin_lock_irqsave(&rq->lock, flags);
1800 lockdep_pin_lock(&rq->lock);
1803 p = llist_entry(llist, struct task_struct, wake_entry);
1804 llist = llist_next(llist);
1805 ttwu_do_activate(rq, p, 0);
1808 lockdep_unpin_lock(&rq->lock);
1809 raw_spin_unlock_irqrestore(&rq->lock, flags);
1812 void scheduler_ipi(void)
1815 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1816 * TIF_NEED_RESCHED remotely (for the first time) will also send
1819 preempt_fold_need_resched();
1821 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1825 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1826 * traditionally all their work was done from the interrupt return
1827 * path. Now that we actually do some work, we need to make sure
1830 * Some archs already do call them, luckily irq_enter/exit nest
1833 * Arguably we should visit all archs and update all handlers,
1834 * however a fair share of IPIs are still resched only so this would
1835 * somewhat pessimize the simple resched case.
1838 sched_ttwu_pending();
1841 * Check if someone kicked us for doing the nohz idle load balance.
1843 if (unlikely(got_nohz_idle_kick())) {
1844 this_rq()->idle_balance = 1;
1845 raise_softirq_irqoff(SCHED_SOFTIRQ);
1850 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1852 struct rq *rq = cpu_rq(cpu);
1854 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1855 if (!set_nr_if_polling(rq->idle))
1856 smp_send_reschedule(cpu);
1858 trace_sched_wake_idle_without_ipi(cpu);
1862 void wake_up_if_idle(int cpu)
1864 struct rq *rq = cpu_rq(cpu);
1865 unsigned long flags;
1869 if (!is_idle_task(rcu_dereference(rq->curr)))
1872 if (set_nr_if_polling(rq->idle)) {
1873 trace_sched_wake_idle_without_ipi(cpu);
1875 raw_spin_lock_irqsave(&rq->lock, flags);
1876 if (is_idle_task(rq->curr))
1877 smp_send_reschedule(cpu);
1878 /* Else cpu is not in idle, do nothing here */
1879 raw_spin_unlock_irqrestore(&rq->lock, flags);
1886 bool cpus_share_cache(int this_cpu, int that_cpu)
1888 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1890 #endif /* CONFIG_SMP */
1892 static void ttwu_queue(struct task_struct *p, int cpu)
1894 struct rq *rq = cpu_rq(cpu);
1896 #if defined(CONFIG_SMP)
1897 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1898 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1899 ttwu_queue_remote(p, cpu);
1904 raw_spin_lock(&rq->lock);
1905 lockdep_pin_lock(&rq->lock);
1906 ttwu_do_activate(rq, p, 0);
1907 lockdep_unpin_lock(&rq->lock);
1908 raw_spin_unlock(&rq->lock);
1912 * try_to_wake_up - wake up a thread
1913 * @p: the thread to be awakened
1914 * @state: the mask of task states that can be woken
1915 * @wake_flags: wake modifier flags (WF_*)
1917 * Put it on the run-queue if it's not already there. The "current"
1918 * thread is always on the run-queue (except when the actual
1919 * re-schedule is in progress), and as such you're allowed to do
1920 * the simpler "current->state = TASK_RUNNING" to mark yourself
1921 * runnable without the overhead of this.
1923 * Return: %true if @p was woken up, %false if it was already running.
1924 * or @state didn't match @p's state.
1927 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1929 unsigned long flags;
1930 int cpu, success = 0;
1933 * If we are going to wake up a thread waiting for CONDITION we
1934 * need to ensure that CONDITION=1 done by the caller can not be
1935 * reordered with p->state check below. This pairs with mb() in
1936 * set_current_state() the waiting thread does.
1938 smp_mb__before_spinlock();
1939 raw_spin_lock_irqsave(&p->pi_lock, flags);
1940 if (!(p->state & state))
1943 trace_sched_waking(p);
1945 success = 1; /* we're going to change ->state */
1949 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1950 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1951 * in smp_cond_load_acquire() below.
1953 * sched_ttwu_pending() try_to_wake_up()
1954 * [S] p->on_rq = 1; [L] P->state
1955 * UNLOCK rq->lock -----.
1959 * LOCK rq->lock -----'
1963 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1965 * Pairs with the UNLOCK+LOCK on rq->lock from the
1966 * last wakeup of our task and the schedule that got our task
1970 if (p->on_rq && ttwu_remote(p, wake_flags))
1975 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1976 * possible to, falsely, observe p->on_cpu == 0.
1978 * One must be running (->on_cpu == 1) in order to remove oneself
1979 * from the runqueue.
1981 * [S] ->on_cpu = 1; [L] ->on_rq
1985 * [S] ->on_rq = 0; [L] ->on_cpu
1987 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1988 * from the consecutive calls to schedule(); the first switching to our
1989 * task, the second putting it to sleep.
1994 * If the owning (remote) cpu is still in the middle of schedule() with
1995 * this task as prev, wait until its done referencing the task.
2000 * Combined with the control dependency above, we have an effective
2001 * smp_load_acquire() without the need for full barriers.
2003 * Pairs with the smp_store_release() in finish_lock_switch().
2005 * This ensures that tasks getting woken will be fully ordered against
2006 * their previous state and preserve Program Order.
2010 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2011 p->state = TASK_WAKING;
2013 if (p->sched_class->task_waking)
2014 p->sched_class->task_waking(p);
2016 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2017 if (task_cpu(p) != cpu) {
2018 wake_flags |= WF_MIGRATED;
2019 set_task_cpu(p, cpu);
2021 #endif /* CONFIG_SMP */
2025 ttwu_stat(p, cpu, wake_flags);
2027 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2033 * try_to_wake_up_local - try to wake up a local task with rq lock held
2034 * @p: the thread to be awakened
2036 * Put @p on the run-queue if it's not already there. The caller must
2037 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2040 static void try_to_wake_up_local(struct task_struct *p)
2042 struct rq *rq = task_rq(p);
2044 if (WARN_ON_ONCE(rq != this_rq()) ||
2045 WARN_ON_ONCE(p == current))
2048 lockdep_assert_held(&rq->lock);
2050 if (!raw_spin_trylock(&p->pi_lock)) {
2052 * This is OK, because current is on_cpu, which avoids it being
2053 * picked for load-balance and preemption/IRQs are still
2054 * disabled avoiding further scheduler activity on it and we've
2055 * not yet picked a replacement task.
2057 lockdep_unpin_lock(&rq->lock);
2058 raw_spin_unlock(&rq->lock);
2059 raw_spin_lock(&p->pi_lock);
2060 raw_spin_lock(&rq->lock);
2061 lockdep_pin_lock(&rq->lock);
2064 if (!(p->state & TASK_NORMAL))
2067 trace_sched_waking(p);
2069 if (!task_on_rq_queued(p))
2070 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2072 ttwu_do_wakeup(rq, p, 0);
2073 ttwu_stat(p, smp_processor_id(), 0);
2075 raw_spin_unlock(&p->pi_lock);
2079 * wake_up_process - Wake up a specific process
2080 * @p: The process to be woken up.
2082 * Attempt to wake up the nominated process and move it to the set of runnable
2085 * Return: 1 if the process was woken up, 0 if it was already running.
2087 * It may be assumed that this function implies a write memory barrier before
2088 * changing the task state if and only if any tasks are woken up.
2090 int wake_up_process(struct task_struct *p)
2092 return try_to_wake_up(p, TASK_NORMAL, 0);
2094 EXPORT_SYMBOL(wake_up_process);
2096 int wake_up_state(struct task_struct *p, unsigned int state)
2098 return try_to_wake_up(p, state, 0);
2102 * This function clears the sched_dl_entity static params.
2104 void __dl_clear_params(struct task_struct *p)
2106 struct sched_dl_entity *dl_se = &p->dl;
2108 dl_se->dl_runtime = 0;
2109 dl_se->dl_deadline = 0;
2110 dl_se->dl_period = 0;
2114 dl_se->dl_throttled = 0;
2116 dl_se->dl_yielded = 0;
2120 * Perform scheduler related setup for a newly forked process p.
2121 * p is forked by current.
2123 * __sched_fork() is basic setup used by init_idle() too:
2125 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2130 p->se.exec_start = 0;
2131 p->se.sum_exec_runtime = 0;
2132 p->se.prev_sum_exec_runtime = 0;
2133 p->se.nr_migrations = 0;
2135 INIT_LIST_HEAD(&p->se.group_node);
2137 #ifdef CONFIG_SCHEDSTATS
2138 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2141 RB_CLEAR_NODE(&p->dl.rb_node);
2142 init_dl_task_timer(&p->dl);
2143 __dl_clear_params(p);
2145 INIT_LIST_HEAD(&p->rt.run_list);
2147 #ifdef CONFIG_PREEMPT_NOTIFIERS
2148 INIT_HLIST_HEAD(&p->preempt_notifiers);
2151 #ifdef CONFIG_NUMA_BALANCING
2152 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2153 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2154 p->mm->numa_scan_seq = 0;
2157 if (clone_flags & CLONE_VM)
2158 p->numa_preferred_nid = current->numa_preferred_nid;
2160 p->numa_preferred_nid = -1;
2162 p->node_stamp = 0ULL;
2163 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2164 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2165 p->numa_work.next = &p->numa_work;
2166 p->numa_faults = NULL;
2167 p->last_task_numa_placement = 0;
2168 p->last_sum_exec_runtime = 0;
2170 p->numa_group = NULL;
2171 #endif /* CONFIG_NUMA_BALANCING */
2174 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2176 #ifdef CONFIG_NUMA_BALANCING
2178 void set_numabalancing_state(bool enabled)
2181 static_branch_enable(&sched_numa_balancing);
2183 static_branch_disable(&sched_numa_balancing);
2186 #ifdef CONFIG_PROC_SYSCTL
2187 int sysctl_numa_balancing(struct ctl_table *table, int write,
2188 void __user *buffer, size_t *lenp, loff_t *ppos)
2192 int state = static_branch_likely(&sched_numa_balancing);
2194 if (write && !capable(CAP_SYS_ADMIN))
2199 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2203 set_numabalancing_state(state);
2210 * fork()/clone()-time setup:
2212 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2214 unsigned long flags;
2215 int cpu = get_cpu();
2217 __sched_fork(clone_flags, p);
2219 * We mark the process as running here. This guarantees that
2220 * nobody will actually run it, and a signal or other external
2221 * event cannot wake it up and insert it on the runqueue either.
2223 p->state = TASK_RUNNING;
2226 * Make sure we do not leak PI boosting priority to the child.
2228 p->prio = current->normal_prio;
2231 * Revert to default priority/policy on fork if requested.
2233 if (unlikely(p->sched_reset_on_fork)) {
2234 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2235 p->policy = SCHED_NORMAL;
2236 p->static_prio = NICE_TO_PRIO(0);
2238 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2239 p->static_prio = NICE_TO_PRIO(0);
2241 p->prio = p->normal_prio = __normal_prio(p);
2245 * We don't need the reset flag anymore after the fork. It has
2246 * fulfilled its duty:
2248 p->sched_reset_on_fork = 0;
2251 if (dl_prio(p->prio)) {
2254 } else if (rt_prio(p->prio)) {
2255 p->sched_class = &rt_sched_class;
2257 p->sched_class = &fair_sched_class;
2260 if (p->sched_class->task_fork)
2261 p->sched_class->task_fork(p);
2264 * The child is not yet in the pid-hash so no cgroup attach races,
2265 * and the cgroup is pinned to this child due to cgroup_fork()
2266 * is ran before sched_fork().
2268 * Silence PROVE_RCU.
2270 raw_spin_lock_irqsave(&p->pi_lock, flags);
2271 set_task_cpu(p, cpu);
2272 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2274 #ifdef CONFIG_SCHED_INFO
2275 if (likely(sched_info_on()))
2276 memset(&p->sched_info, 0, sizeof(p->sched_info));
2278 #if defined(CONFIG_SMP)
2281 init_task_preempt_count(p);
2283 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2284 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2291 unsigned long to_ratio(u64 period, u64 runtime)
2293 if (runtime == RUNTIME_INF)
2297 * Doing this here saves a lot of checks in all
2298 * the calling paths, and returning zero seems
2299 * safe for them anyway.
2304 return div64_u64(runtime << 20, period);
2308 inline struct dl_bw *dl_bw_of(int i)
2310 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2311 "sched RCU must be held");
2312 return &cpu_rq(i)->rd->dl_bw;
2315 static inline int dl_bw_cpus(int i)
2317 struct root_domain *rd = cpu_rq(i)->rd;
2320 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2321 "sched RCU must be held");
2322 for_each_cpu_and(i, rd->span, cpu_active_mask)
2328 inline struct dl_bw *dl_bw_of(int i)
2330 return &cpu_rq(i)->dl.dl_bw;
2333 static inline int dl_bw_cpus(int i)
2340 * We must be sure that accepting a new task (or allowing changing the
2341 * parameters of an existing one) is consistent with the bandwidth
2342 * constraints. If yes, this function also accordingly updates the currently
2343 * allocated bandwidth to reflect the new situation.
2345 * This function is called while holding p's rq->lock.
2347 * XXX we should delay bw change until the task's 0-lag point, see
2350 static int dl_overflow(struct task_struct *p, int policy,
2351 const struct sched_attr *attr)
2354 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2355 u64 period = attr->sched_period ?: attr->sched_deadline;
2356 u64 runtime = attr->sched_runtime;
2357 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2360 if (new_bw == p->dl.dl_bw)
2364 * Either if a task, enters, leave, or stays -deadline but changes
2365 * its parameters, we may need to update accordingly the total
2366 * allocated bandwidth of the container.
2368 raw_spin_lock(&dl_b->lock);
2369 cpus = dl_bw_cpus(task_cpu(p));
2370 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2371 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2372 __dl_add(dl_b, new_bw);
2374 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2375 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2376 __dl_clear(dl_b, p->dl.dl_bw);
2377 __dl_add(dl_b, new_bw);
2379 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2380 __dl_clear(dl_b, p->dl.dl_bw);
2383 raw_spin_unlock(&dl_b->lock);
2388 extern void init_dl_bw(struct dl_bw *dl_b);
2391 * wake_up_new_task - wake up a newly created task for the first time.
2393 * This function will do some initial scheduler statistics housekeeping
2394 * that must be done for every newly created context, then puts the task
2395 * on the runqueue and wakes it.
2397 void wake_up_new_task(struct task_struct *p)
2399 unsigned long flags;
2402 raw_spin_lock_irqsave(&p->pi_lock, flags);
2403 /* Initialize new task's runnable average */
2404 init_entity_runnable_average(&p->se);
2407 * Fork balancing, do it here and not earlier because:
2408 * - cpus_allowed can change in the fork path
2409 * - any previously selected cpu might disappear through hotplug
2411 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2414 rq = __task_rq_lock(p);
2415 activate_task(rq, p, 0);
2416 p->on_rq = TASK_ON_RQ_QUEUED;
2417 trace_sched_wakeup_new(p);
2418 check_preempt_curr(rq, p, WF_FORK);
2420 if (p->sched_class->task_woken) {
2422 * Nothing relies on rq->lock after this, so its fine to
2425 lockdep_unpin_lock(&rq->lock);
2426 p->sched_class->task_woken(rq, p);
2427 lockdep_pin_lock(&rq->lock);
2430 task_rq_unlock(rq, p, &flags);
2433 #ifdef CONFIG_PREEMPT_NOTIFIERS
2435 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2437 void preempt_notifier_inc(void)
2439 static_key_slow_inc(&preempt_notifier_key);
2441 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2443 void preempt_notifier_dec(void)
2445 static_key_slow_dec(&preempt_notifier_key);
2447 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2450 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2451 * @notifier: notifier struct to register
2453 void preempt_notifier_register(struct preempt_notifier *notifier)
2455 if (!static_key_false(&preempt_notifier_key))
2456 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2458 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2460 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2463 * preempt_notifier_unregister - no longer interested in preemption notifications
2464 * @notifier: notifier struct to unregister
2466 * This is *not* safe to call from within a preemption notifier.
2468 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2470 hlist_del(¬ifier->link);
2472 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2474 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2476 struct preempt_notifier *notifier;
2478 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2479 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2482 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2484 if (static_key_false(&preempt_notifier_key))
2485 __fire_sched_in_preempt_notifiers(curr);
2489 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2490 struct task_struct *next)
2492 struct preempt_notifier *notifier;
2494 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2495 notifier->ops->sched_out(notifier, next);
2498 static __always_inline void
2499 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2500 struct task_struct *next)
2502 if (static_key_false(&preempt_notifier_key))
2503 __fire_sched_out_preempt_notifiers(curr, next);
2506 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2508 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2513 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2514 struct task_struct *next)
2518 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2521 * prepare_task_switch - prepare to switch tasks
2522 * @rq: the runqueue preparing to switch
2523 * @prev: the current task that is being switched out
2524 * @next: the task we are going to switch to.
2526 * This is called with the rq lock held and interrupts off. It must
2527 * be paired with a subsequent finish_task_switch after the context
2530 * prepare_task_switch sets up locking and calls architecture specific
2534 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2535 struct task_struct *next)
2537 sched_info_switch(rq, prev, next);
2538 perf_event_task_sched_out(prev, next);
2539 fire_sched_out_preempt_notifiers(prev, next);
2540 prepare_lock_switch(rq, next);
2541 prepare_arch_switch(next);
2545 * finish_task_switch - clean up after a task-switch
2546 * @prev: the thread we just switched away from.
2548 * finish_task_switch must be called after the context switch, paired
2549 * with a prepare_task_switch call before the context switch.
2550 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2551 * and do any other architecture-specific cleanup actions.
2553 * Note that we may have delayed dropping an mm in context_switch(). If
2554 * so, we finish that here outside of the runqueue lock. (Doing it
2555 * with the lock held can cause deadlocks; see schedule() for
2558 * The context switch have flipped the stack from under us and restored the
2559 * local variables which were saved when this task called schedule() in the
2560 * past. prev == current is still correct but we need to recalculate this_rq
2561 * because prev may have moved to another CPU.
2563 static struct rq *finish_task_switch(struct task_struct *prev)
2564 __releases(rq->lock)
2566 struct rq *rq = this_rq();
2567 struct mm_struct *mm = rq->prev_mm;
2571 * The previous task will have left us with a preempt_count of 2
2572 * because it left us after:
2575 * preempt_disable(); // 1
2577 * raw_spin_lock_irq(&rq->lock) // 2
2579 * Also, see FORK_PREEMPT_COUNT.
2581 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2582 "corrupted preempt_count: %s/%d/0x%x\n",
2583 current->comm, current->pid, preempt_count()))
2584 preempt_count_set(FORK_PREEMPT_COUNT);
2589 * A task struct has one reference for the use as "current".
2590 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2591 * schedule one last time. The schedule call will never return, and
2592 * the scheduled task must drop that reference.
2594 * We must observe prev->state before clearing prev->on_cpu (in
2595 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2596 * running on another CPU and we could rave with its RUNNING -> DEAD
2597 * transition, resulting in a double drop.
2599 prev_state = prev->state;
2600 vtime_task_switch(prev);
2601 perf_event_task_sched_in(prev, current);
2602 finish_lock_switch(rq, prev);
2603 finish_arch_post_lock_switch();
2605 fire_sched_in_preempt_notifiers(current);
2608 if (unlikely(prev_state == TASK_DEAD)) {
2609 if (prev->sched_class->task_dead)
2610 prev->sched_class->task_dead(prev);
2613 * Remove function-return probe instances associated with this
2614 * task and put them back on the free list.
2616 kprobe_flush_task(prev);
2617 put_task_struct(prev);
2620 tick_nohz_task_switch();
2626 /* rq->lock is NOT held, but preemption is disabled */
2627 static void __balance_callback(struct rq *rq)
2629 struct callback_head *head, *next;
2630 void (*func)(struct rq *rq);
2631 unsigned long flags;
2633 raw_spin_lock_irqsave(&rq->lock, flags);
2634 head = rq->balance_callback;
2635 rq->balance_callback = NULL;
2637 func = (void (*)(struct rq *))head->func;
2644 raw_spin_unlock_irqrestore(&rq->lock, flags);
2647 static inline void balance_callback(struct rq *rq)
2649 if (unlikely(rq->balance_callback))
2650 __balance_callback(rq);
2655 static inline void balance_callback(struct rq *rq)
2662 * schedule_tail - first thing a freshly forked thread must call.
2663 * @prev: the thread we just switched away from.
2665 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2666 __releases(rq->lock)
2671 * New tasks start with FORK_PREEMPT_COUNT, see there and
2672 * finish_task_switch() for details.
2674 * finish_task_switch() will drop rq->lock() and lower preempt_count
2675 * and the preempt_enable() will end up enabling preemption (on
2676 * PREEMPT_COUNT kernels).
2679 rq = finish_task_switch(prev);
2680 balance_callback(rq);
2683 if (current->set_child_tid)
2684 put_user(task_pid_vnr(current), current->set_child_tid);
2688 * context_switch - switch to the new MM and the new thread's register state.
2690 static inline struct rq *
2691 context_switch(struct rq *rq, struct task_struct *prev,
2692 struct task_struct *next)
2694 struct mm_struct *mm, *oldmm;
2696 prepare_task_switch(rq, prev, next);
2699 oldmm = prev->active_mm;
2701 * For paravirt, this is coupled with an exit in switch_to to
2702 * combine the page table reload and the switch backend into
2705 arch_start_context_switch(prev);
2708 next->active_mm = oldmm;
2709 atomic_inc(&oldmm->mm_count);
2710 enter_lazy_tlb(oldmm, next);
2712 switch_mm(oldmm, mm, next);
2715 prev->active_mm = NULL;
2716 rq->prev_mm = oldmm;
2719 * Since the runqueue lock will be released by the next
2720 * task (which is an invalid locking op but in the case
2721 * of the scheduler it's an obvious special-case), so we
2722 * do an early lockdep release here:
2724 lockdep_unpin_lock(&rq->lock);
2725 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2727 /* Here we just switch the register state and the stack. */
2728 switch_to(prev, next, prev);
2731 return finish_task_switch(prev);
2735 * nr_running and nr_context_switches:
2737 * externally visible scheduler statistics: current number of runnable
2738 * threads, total number of context switches performed since bootup.
2740 unsigned long nr_running(void)
2742 unsigned long i, sum = 0;
2744 for_each_online_cpu(i)
2745 sum += cpu_rq(i)->nr_running;
2751 * Check if only the current task is running on the cpu.
2753 * Caution: this function does not check that the caller has disabled
2754 * preemption, thus the result might have a time-of-check-to-time-of-use
2755 * race. The caller is responsible to use it correctly, for example:
2757 * - from a non-preemptable section (of course)
2759 * - from a thread that is bound to a single CPU
2761 * - in a loop with very short iterations (e.g. a polling loop)
2763 bool single_task_running(void)
2765 return raw_rq()->nr_running == 1;
2767 EXPORT_SYMBOL(single_task_running);
2769 unsigned long long nr_context_switches(void)
2772 unsigned long long sum = 0;
2774 for_each_possible_cpu(i)
2775 sum += cpu_rq(i)->nr_switches;
2780 unsigned long nr_iowait(void)
2782 unsigned long i, sum = 0;
2784 for_each_possible_cpu(i)
2785 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2790 unsigned long nr_iowait_cpu(int cpu)
2792 struct rq *this = cpu_rq(cpu);
2793 return atomic_read(&this->nr_iowait);
2796 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2798 struct rq *rq = this_rq();
2799 *nr_waiters = atomic_read(&rq->nr_iowait);
2800 *load = rq->load.weight;
2806 * sched_exec - execve() is a valuable balancing opportunity, because at
2807 * this point the task has the smallest effective memory and cache footprint.
2809 void sched_exec(void)
2811 struct task_struct *p = current;
2812 unsigned long flags;
2815 raw_spin_lock_irqsave(&p->pi_lock, flags);
2816 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2817 if (dest_cpu == smp_processor_id())
2820 if (likely(cpu_active(dest_cpu))) {
2821 struct migration_arg arg = { p, dest_cpu };
2823 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2824 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2828 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2833 DEFINE_PER_CPU(struct kernel_stat, kstat);
2834 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2836 EXPORT_PER_CPU_SYMBOL(kstat);
2837 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2840 * Return accounted runtime for the task.
2841 * In case the task is currently running, return the runtime plus current's
2842 * pending runtime that have not been accounted yet.
2844 unsigned long long task_sched_runtime(struct task_struct *p)
2846 unsigned long flags;
2850 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2852 * 64-bit doesn't need locks to atomically read a 64bit value.
2853 * So we have a optimization chance when the task's delta_exec is 0.
2854 * Reading ->on_cpu is racy, but this is ok.
2856 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2857 * If we race with it entering cpu, unaccounted time is 0. This is
2858 * indistinguishable from the read occurring a few cycles earlier.
2859 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2860 * been accounted, so we're correct here as well.
2862 if (!p->on_cpu || !task_on_rq_queued(p))
2863 return p->se.sum_exec_runtime;
2866 rq = task_rq_lock(p, &flags);
2868 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2869 * project cycles that may never be accounted to this
2870 * thread, breaking clock_gettime().
2872 if (task_current(rq, p) && task_on_rq_queued(p)) {
2873 update_rq_clock(rq);
2874 p->sched_class->update_curr(rq);
2876 ns = p->se.sum_exec_runtime;
2877 task_rq_unlock(rq, p, &flags);
2883 * This function gets called by the timer code, with HZ frequency.
2884 * We call it with interrupts disabled.
2886 void scheduler_tick(void)
2888 int cpu = smp_processor_id();
2889 struct rq *rq = cpu_rq(cpu);
2890 struct task_struct *curr = rq->curr;
2894 raw_spin_lock(&rq->lock);
2895 update_rq_clock(rq);
2896 curr->sched_class->task_tick(rq, curr, 0);
2897 update_cpu_load_active(rq);
2898 calc_global_load_tick(rq);
2899 raw_spin_unlock(&rq->lock);
2901 perf_event_task_tick();
2904 rq->idle_balance = idle_cpu(cpu);
2905 trigger_load_balance(rq);
2907 rq_last_tick_reset(rq);
2910 #ifdef CONFIG_NO_HZ_FULL
2912 * scheduler_tick_max_deferment
2914 * Keep at least one tick per second when a single
2915 * active task is running because the scheduler doesn't
2916 * yet completely support full dynticks environment.
2918 * This makes sure that uptime, CFS vruntime, load
2919 * balancing, etc... continue to move forward, even
2920 * with a very low granularity.
2922 * Return: Maximum deferment in nanoseconds.
2924 u64 scheduler_tick_max_deferment(void)
2926 struct rq *rq = this_rq();
2927 unsigned long next, now = READ_ONCE(jiffies);
2929 next = rq->last_sched_tick + HZ;
2931 if (time_before_eq(next, now))
2934 return jiffies_to_nsecs(next - now);
2938 notrace unsigned long get_parent_ip(unsigned long addr)
2940 if (in_lock_functions(addr)) {
2941 addr = CALLER_ADDR2;
2942 if (in_lock_functions(addr))
2943 addr = CALLER_ADDR3;
2948 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2949 defined(CONFIG_PREEMPT_TRACER))
2951 void preempt_count_add(int val)
2953 #ifdef CONFIG_DEBUG_PREEMPT
2957 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2960 __preempt_count_add(val);
2961 #ifdef CONFIG_DEBUG_PREEMPT
2963 * Spinlock count overflowing soon?
2965 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2968 if (preempt_count() == val) {
2969 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2970 #ifdef CONFIG_DEBUG_PREEMPT
2971 current->preempt_disable_ip = ip;
2973 trace_preempt_off(CALLER_ADDR0, ip);
2976 EXPORT_SYMBOL(preempt_count_add);
2977 NOKPROBE_SYMBOL(preempt_count_add);
2979 void preempt_count_sub(int val)
2981 #ifdef CONFIG_DEBUG_PREEMPT
2985 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2988 * Is the spinlock portion underflowing?
2990 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2991 !(preempt_count() & PREEMPT_MASK)))
2995 if (preempt_count() == val)
2996 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2997 __preempt_count_sub(val);
2999 EXPORT_SYMBOL(preempt_count_sub);
3000 NOKPROBE_SYMBOL(preempt_count_sub);
3005 * Print scheduling while atomic bug:
3007 static noinline void __schedule_bug(struct task_struct *prev)
3009 if (oops_in_progress)
3012 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3013 prev->comm, prev->pid, preempt_count());
3015 debug_show_held_locks(prev);
3017 if (irqs_disabled())
3018 print_irqtrace_events(prev);
3019 #ifdef CONFIG_DEBUG_PREEMPT
3020 if (in_atomic_preempt_off()) {
3021 pr_err("Preemption disabled at:");
3022 print_ip_sym(current->preempt_disable_ip);
3027 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3031 * Various schedule()-time debugging checks and statistics:
3033 static inline void schedule_debug(struct task_struct *prev)
3035 #ifdef CONFIG_SCHED_STACK_END_CHECK
3036 if (task_stack_end_corrupted(prev))
3037 panic("corrupted stack end detected inside scheduler\n");
3040 if (unlikely(in_atomic_preempt_off())) {
3041 __schedule_bug(prev);
3042 preempt_count_set(PREEMPT_DISABLED);
3046 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3048 schedstat_inc(this_rq(), sched_count);
3052 * Pick up the highest-prio task:
3054 static inline struct task_struct *
3055 pick_next_task(struct rq *rq, struct task_struct *prev)
3057 const struct sched_class *class = &fair_sched_class;
3058 struct task_struct *p;
3061 * Optimization: we know that if all tasks are in
3062 * the fair class we can call that function directly:
3064 if (likely(prev->sched_class == class &&
3065 rq->nr_running == rq->cfs.h_nr_running)) {
3066 p = fair_sched_class.pick_next_task(rq, prev);
3067 if (unlikely(p == RETRY_TASK))
3070 /* assumes fair_sched_class->next == idle_sched_class */
3072 p = idle_sched_class.pick_next_task(rq, prev);
3078 for_each_class(class) {
3079 p = class->pick_next_task(rq, prev);
3081 if (unlikely(p == RETRY_TASK))
3087 BUG(); /* the idle class will always have a runnable task */
3091 * __schedule() is the main scheduler function.
3093 * The main means of driving the scheduler and thus entering this function are:
3095 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3097 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3098 * paths. For example, see arch/x86/entry_64.S.
3100 * To drive preemption between tasks, the scheduler sets the flag in timer
3101 * interrupt handler scheduler_tick().
3103 * 3. Wakeups don't really cause entry into schedule(). They add a
3104 * task to the run-queue and that's it.
3106 * Now, if the new task added to the run-queue preempts the current
3107 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3108 * called on the nearest possible occasion:
3110 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3112 * - in syscall or exception context, at the next outmost
3113 * preempt_enable(). (this might be as soon as the wake_up()'s
3116 * - in IRQ context, return from interrupt-handler to
3117 * preemptible context
3119 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3122 * - cond_resched() call
3123 * - explicit schedule() call
3124 * - return from syscall or exception to user-space
3125 * - return from interrupt-handler to user-space
3127 * WARNING: must be called with preemption disabled!
3129 static void __sched notrace __schedule(bool preempt)
3131 struct task_struct *prev, *next;
3132 unsigned long *switch_count;
3136 cpu = smp_processor_id();
3138 rcu_note_context_switch();
3142 * do_exit() calls schedule() with preemption disabled as an exception;
3143 * however we must fix that up, otherwise the next task will see an
3144 * inconsistent (higher) preempt count.
3146 * It also avoids the below schedule_debug() test from complaining
3149 if (unlikely(prev->state == TASK_DEAD))
3150 preempt_enable_no_resched_notrace();
3152 schedule_debug(prev);
3154 if (sched_feat(HRTICK))
3158 * Make sure that signal_pending_state()->signal_pending() below
3159 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3160 * done by the caller to avoid the race with signal_wake_up().
3162 smp_mb__before_spinlock();
3163 raw_spin_lock_irq(&rq->lock);
3164 lockdep_pin_lock(&rq->lock);
3166 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3168 switch_count = &prev->nivcsw;
3169 if (!preempt && prev->state) {
3170 if (unlikely(signal_pending_state(prev->state, prev))) {
3171 prev->state = TASK_RUNNING;
3173 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3177 * If a worker went to sleep, notify and ask workqueue
3178 * whether it wants to wake up a task to maintain
3181 if (prev->flags & PF_WQ_WORKER) {
3182 struct task_struct *to_wakeup;
3184 to_wakeup = wq_worker_sleeping(prev, cpu);
3186 try_to_wake_up_local(to_wakeup);
3189 switch_count = &prev->nvcsw;
3192 if (task_on_rq_queued(prev))
3193 update_rq_clock(rq);
3195 next = pick_next_task(rq, prev);
3196 clear_tsk_need_resched(prev);
3197 clear_preempt_need_resched();
3198 rq->clock_skip_update = 0;
3200 if (likely(prev != next)) {
3205 trace_sched_switch(preempt, prev, next);
3206 rq = context_switch(rq, prev, next); /* unlocks the rq */
3209 lockdep_unpin_lock(&rq->lock);
3210 raw_spin_unlock_irq(&rq->lock);
3213 balance_callback(rq);
3216 static inline void sched_submit_work(struct task_struct *tsk)
3218 if (!tsk->state || tsk_is_pi_blocked(tsk))
3221 * If we are going to sleep and we have plugged IO queued,
3222 * make sure to submit it to avoid deadlocks.
3224 if (blk_needs_flush_plug(tsk))
3225 blk_schedule_flush_plug(tsk);
3228 asmlinkage __visible void __sched schedule(void)
3230 struct task_struct *tsk = current;
3232 sched_submit_work(tsk);
3236 sched_preempt_enable_no_resched();
3237 } while (need_resched());
3239 EXPORT_SYMBOL(schedule);
3241 #ifdef CONFIG_CONTEXT_TRACKING
3242 asmlinkage __visible void __sched schedule_user(void)
3245 * If we come here after a random call to set_need_resched(),
3246 * or we have been woken up remotely but the IPI has not yet arrived,
3247 * we haven't yet exited the RCU idle mode. Do it here manually until
3248 * we find a better solution.
3250 * NB: There are buggy callers of this function. Ideally we
3251 * should warn if prev_state != CONTEXT_USER, but that will trigger
3252 * too frequently to make sense yet.
3254 enum ctx_state prev_state = exception_enter();
3256 exception_exit(prev_state);
3261 * schedule_preempt_disabled - called with preemption disabled
3263 * Returns with preemption disabled. Note: preempt_count must be 1
3265 void __sched schedule_preempt_disabled(void)
3267 sched_preempt_enable_no_resched();
3272 static void __sched notrace preempt_schedule_common(void)
3275 preempt_disable_notrace();
3277 preempt_enable_no_resched_notrace();
3280 * Check again in case we missed a preemption opportunity
3281 * between schedule and now.
3283 } while (need_resched());
3286 #ifdef CONFIG_PREEMPT
3288 * this is the entry point to schedule() from in-kernel preemption
3289 * off of preempt_enable. Kernel preemptions off return from interrupt
3290 * occur there and call schedule directly.
3292 asmlinkage __visible void __sched notrace preempt_schedule(void)
3295 * If there is a non-zero preempt_count or interrupts are disabled,
3296 * we do not want to preempt the current task. Just return..
3298 if (likely(!preemptible()))
3301 preempt_schedule_common();
3303 NOKPROBE_SYMBOL(preempt_schedule);
3304 EXPORT_SYMBOL(preempt_schedule);
3307 * preempt_schedule_notrace - preempt_schedule called by tracing
3309 * The tracing infrastructure uses preempt_enable_notrace to prevent
3310 * recursion and tracing preempt enabling caused by the tracing
3311 * infrastructure itself. But as tracing can happen in areas coming
3312 * from userspace or just about to enter userspace, a preempt enable
3313 * can occur before user_exit() is called. This will cause the scheduler
3314 * to be called when the system is still in usermode.
3316 * To prevent this, the preempt_enable_notrace will use this function
3317 * instead of preempt_schedule() to exit user context if needed before
3318 * calling the scheduler.
3320 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3322 enum ctx_state prev_ctx;
3324 if (likely(!preemptible()))
3328 preempt_disable_notrace();
3330 * Needs preempt disabled in case user_exit() is traced
3331 * and the tracer calls preempt_enable_notrace() causing
3332 * an infinite recursion.
3334 prev_ctx = exception_enter();
3336 exception_exit(prev_ctx);
3338 preempt_enable_no_resched_notrace();
3339 } while (need_resched());
3341 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3343 #endif /* CONFIG_PREEMPT */
3346 * this is the entry point to schedule() from kernel preemption
3347 * off of irq context.
3348 * Note, that this is called and return with irqs disabled. This will
3349 * protect us against recursive calling from irq.
3351 asmlinkage __visible void __sched preempt_schedule_irq(void)
3353 enum ctx_state prev_state;
3355 /* Catch callers which need to be fixed */
3356 BUG_ON(preempt_count() || !irqs_disabled());
3358 prev_state = exception_enter();
3364 local_irq_disable();
3365 sched_preempt_enable_no_resched();
3366 } while (need_resched());
3368 exception_exit(prev_state);
3371 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3374 return try_to_wake_up(curr->private, mode, wake_flags);
3376 EXPORT_SYMBOL(default_wake_function);
3378 #ifdef CONFIG_RT_MUTEXES
3381 * rt_mutex_setprio - set the current priority of a task
3383 * @prio: prio value (kernel-internal form)
3385 * This function changes the 'effective' priority of a task. It does
3386 * not touch ->normal_prio like __setscheduler().
3388 * Used by the rt_mutex code to implement priority inheritance
3389 * logic. Call site only calls if the priority of the task changed.
3391 void rt_mutex_setprio(struct task_struct *p, int prio)
3393 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3395 const struct sched_class *prev_class;
3397 BUG_ON(prio > MAX_PRIO);
3399 rq = __task_rq_lock(p);
3402 * Idle task boosting is a nono in general. There is one
3403 * exception, when PREEMPT_RT and NOHZ is active:
3405 * The idle task calls get_next_timer_interrupt() and holds
3406 * the timer wheel base->lock on the CPU and another CPU wants
3407 * to access the timer (probably to cancel it). We can safely
3408 * ignore the boosting request, as the idle CPU runs this code
3409 * with interrupts disabled and will complete the lock
3410 * protected section without being interrupted. So there is no
3411 * real need to boost.
3413 if (unlikely(p == rq->idle)) {
3414 WARN_ON(p != rq->curr);
3415 WARN_ON(p->pi_blocked_on);
3419 trace_sched_pi_setprio(p, prio);
3421 prev_class = p->sched_class;
3422 queued = task_on_rq_queued(p);
3423 running = task_current(rq, p);
3425 dequeue_task(rq, p, DEQUEUE_SAVE);
3427 put_prev_task(rq, p);
3430 * Boosting condition are:
3431 * 1. -rt task is running and holds mutex A
3432 * --> -dl task blocks on mutex A
3434 * 2. -dl task is running and holds mutex A
3435 * --> -dl task blocks on mutex A and could preempt the
3438 if (dl_prio(prio)) {
3439 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3440 if (!dl_prio(p->normal_prio) ||
3441 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3442 p->dl.dl_boosted = 1;
3443 enqueue_flag |= ENQUEUE_REPLENISH;
3445 p->dl.dl_boosted = 0;
3446 p->sched_class = &dl_sched_class;
3447 } else if (rt_prio(prio)) {
3448 if (dl_prio(oldprio))
3449 p->dl.dl_boosted = 0;
3451 enqueue_flag |= ENQUEUE_HEAD;
3452 p->sched_class = &rt_sched_class;
3454 if (dl_prio(oldprio))
3455 p->dl.dl_boosted = 0;
3456 if (rt_prio(oldprio))
3458 p->sched_class = &fair_sched_class;
3464 p->sched_class->set_curr_task(rq);
3466 enqueue_task(rq, p, enqueue_flag);
3468 check_class_changed(rq, p, prev_class, oldprio);
3470 preempt_disable(); /* avoid rq from going away on us */
3471 __task_rq_unlock(rq);
3473 balance_callback(rq);
3478 void set_user_nice(struct task_struct *p, long nice)
3480 int old_prio, delta, queued;
3481 unsigned long flags;
3484 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3487 * We have to be careful, if called from sys_setpriority(),
3488 * the task might be in the middle of scheduling on another CPU.
3490 rq = task_rq_lock(p, &flags);
3492 * The RT priorities are set via sched_setscheduler(), but we still
3493 * allow the 'normal' nice value to be set - but as expected
3494 * it wont have any effect on scheduling until the task is
3495 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3497 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3498 p->static_prio = NICE_TO_PRIO(nice);
3501 queued = task_on_rq_queued(p);
3503 dequeue_task(rq, p, DEQUEUE_SAVE);
3505 p->static_prio = NICE_TO_PRIO(nice);
3508 p->prio = effective_prio(p);
3509 delta = p->prio - old_prio;
3512 enqueue_task(rq, p, ENQUEUE_RESTORE);
3514 * If the task increased its priority or is running and
3515 * lowered its priority, then reschedule its CPU:
3517 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3521 task_rq_unlock(rq, p, &flags);
3523 EXPORT_SYMBOL(set_user_nice);
3526 * can_nice - check if a task can reduce its nice value
3530 int can_nice(const struct task_struct *p, const int nice)
3532 /* convert nice value [19,-20] to rlimit style value [1,40] */
3533 int nice_rlim = nice_to_rlimit(nice);
3535 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3536 capable(CAP_SYS_NICE));
3539 #ifdef __ARCH_WANT_SYS_NICE
3542 * sys_nice - change the priority of the current process.
3543 * @increment: priority increment
3545 * sys_setpriority is a more generic, but much slower function that
3546 * does similar things.
3548 SYSCALL_DEFINE1(nice, int, increment)
3553 * Setpriority might change our priority at the same moment.
3554 * We don't have to worry. Conceptually one call occurs first
3555 * and we have a single winner.
3557 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3558 nice = task_nice(current) + increment;
3560 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3561 if (increment < 0 && !can_nice(current, nice))
3564 retval = security_task_setnice(current, nice);
3568 set_user_nice(current, nice);
3575 * task_prio - return the priority value of a given task.
3576 * @p: the task in question.
3578 * Return: The priority value as seen by users in /proc.
3579 * RT tasks are offset by -200. Normal tasks are centered
3580 * around 0, value goes from -16 to +15.
3582 int task_prio(const struct task_struct *p)
3584 return p->prio - MAX_RT_PRIO;
3588 * idle_cpu - is a given cpu idle currently?
3589 * @cpu: the processor in question.
3591 * Return: 1 if the CPU is currently idle. 0 otherwise.
3593 int idle_cpu(int cpu)
3595 struct rq *rq = cpu_rq(cpu);
3597 if (rq->curr != rq->idle)
3604 if (!llist_empty(&rq->wake_list))
3612 * idle_task - return the idle task for a given cpu.
3613 * @cpu: the processor in question.
3615 * Return: The idle task for the cpu @cpu.
3617 struct task_struct *idle_task(int cpu)
3619 return cpu_rq(cpu)->idle;
3623 * find_process_by_pid - find a process with a matching PID value.
3624 * @pid: the pid in question.
3626 * The task of @pid, if found. %NULL otherwise.
3628 static struct task_struct *find_process_by_pid(pid_t pid)
3630 return pid ? find_task_by_vpid(pid) : current;
3634 * This function initializes the sched_dl_entity of a newly becoming
3635 * SCHED_DEADLINE task.
3637 * Only the static values are considered here, the actual runtime and the
3638 * absolute deadline will be properly calculated when the task is enqueued
3639 * for the first time with its new policy.
3642 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3644 struct sched_dl_entity *dl_se = &p->dl;
3646 dl_se->dl_runtime = attr->sched_runtime;
3647 dl_se->dl_deadline = attr->sched_deadline;
3648 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3649 dl_se->flags = attr->sched_flags;
3650 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3653 * Changing the parameters of a task is 'tricky' and we're not doing
3654 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3656 * What we SHOULD do is delay the bandwidth release until the 0-lag
3657 * point. This would include retaining the task_struct until that time
3658 * and change dl_overflow() to not immediately decrement the current
3661 * Instead we retain the current runtime/deadline and let the new
3662 * parameters take effect after the current reservation period lapses.
3663 * This is safe (albeit pessimistic) because the 0-lag point is always
3664 * before the current scheduling deadline.
3666 * We can still have temporary overloads because we do not delay the
3667 * change in bandwidth until that time; so admission control is
3668 * not on the safe side. It does however guarantee tasks will never
3669 * consume more than promised.
3674 * sched_setparam() passes in -1 for its policy, to let the functions
3675 * it calls know not to change it.
3677 #define SETPARAM_POLICY -1
3679 static void __setscheduler_params(struct task_struct *p,
3680 const struct sched_attr *attr)
3682 int policy = attr->sched_policy;
3684 if (policy == SETPARAM_POLICY)
3689 if (dl_policy(policy))
3690 __setparam_dl(p, attr);
3691 else if (fair_policy(policy))
3692 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3695 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3696 * !rt_policy. Always setting this ensures that things like
3697 * getparam()/getattr() don't report silly values for !rt tasks.
3699 p->rt_priority = attr->sched_priority;
3700 p->normal_prio = normal_prio(p);
3704 /* Actually do priority change: must hold pi & rq lock. */
3705 static void __setscheduler(struct rq *rq, struct task_struct *p,
3706 const struct sched_attr *attr, bool keep_boost)
3708 __setscheduler_params(p, attr);
3711 * Keep a potential priority boosting if called from
3712 * sched_setscheduler().
3715 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3717 p->prio = normal_prio(p);
3719 if (dl_prio(p->prio))
3720 p->sched_class = &dl_sched_class;
3721 else if (rt_prio(p->prio))
3722 p->sched_class = &rt_sched_class;
3724 p->sched_class = &fair_sched_class;
3728 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3730 struct sched_dl_entity *dl_se = &p->dl;
3732 attr->sched_priority = p->rt_priority;
3733 attr->sched_runtime = dl_se->dl_runtime;
3734 attr->sched_deadline = dl_se->dl_deadline;
3735 attr->sched_period = dl_se->dl_period;
3736 attr->sched_flags = dl_se->flags;
3740 * This function validates the new parameters of a -deadline task.
3741 * We ask for the deadline not being zero, and greater or equal
3742 * than the runtime, as well as the period of being zero or
3743 * greater than deadline. Furthermore, we have to be sure that
3744 * user parameters are above the internal resolution of 1us (we
3745 * check sched_runtime only since it is always the smaller one) and
3746 * below 2^63 ns (we have to check both sched_deadline and
3747 * sched_period, as the latter can be zero).
3750 __checkparam_dl(const struct sched_attr *attr)
3753 if (attr->sched_deadline == 0)
3757 * Since we truncate DL_SCALE bits, make sure we're at least
3760 if (attr->sched_runtime < (1ULL << DL_SCALE))
3764 * Since we use the MSB for wrap-around and sign issues, make
3765 * sure it's not set (mind that period can be equal to zero).
3767 if (attr->sched_deadline & (1ULL << 63) ||
3768 attr->sched_period & (1ULL << 63))
3771 /* runtime <= deadline <= period (if period != 0) */
3772 if ((attr->sched_period != 0 &&
3773 attr->sched_period < attr->sched_deadline) ||
3774 attr->sched_deadline < attr->sched_runtime)
3781 * check the target process has a UID that matches the current process's
3783 static bool check_same_owner(struct task_struct *p)
3785 const struct cred *cred = current_cred(), *pcred;
3789 pcred = __task_cred(p);
3790 match = (uid_eq(cred->euid, pcred->euid) ||
3791 uid_eq(cred->euid, pcred->uid));
3796 static bool dl_param_changed(struct task_struct *p,
3797 const struct sched_attr *attr)
3799 struct sched_dl_entity *dl_se = &p->dl;
3801 if (dl_se->dl_runtime != attr->sched_runtime ||
3802 dl_se->dl_deadline != attr->sched_deadline ||
3803 dl_se->dl_period != attr->sched_period ||
3804 dl_se->flags != attr->sched_flags)
3810 static int __sched_setscheduler(struct task_struct *p,
3811 const struct sched_attr *attr,
3814 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3815 MAX_RT_PRIO - 1 - attr->sched_priority;
3816 int retval, oldprio, oldpolicy = -1, queued, running;
3817 int new_effective_prio, policy = attr->sched_policy;
3818 unsigned long flags;
3819 const struct sched_class *prev_class;
3823 /* may grab non-irq protected spin_locks */
3824 BUG_ON(in_interrupt());
3826 /* double check policy once rq lock held */
3828 reset_on_fork = p->sched_reset_on_fork;
3829 policy = oldpolicy = p->policy;
3831 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3833 if (!valid_policy(policy))
3837 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3841 * Valid priorities for SCHED_FIFO and SCHED_RR are
3842 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3843 * SCHED_BATCH and SCHED_IDLE is 0.
3845 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3846 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3848 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3849 (rt_policy(policy) != (attr->sched_priority != 0)))
3853 * Allow unprivileged RT tasks to decrease priority:
3855 if (user && !capable(CAP_SYS_NICE)) {
3856 if (fair_policy(policy)) {
3857 if (attr->sched_nice < task_nice(p) &&
3858 !can_nice(p, attr->sched_nice))
3862 if (rt_policy(policy)) {
3863 unsigned long rlim_rtprio =
3864 task_rlimit(p, RLIMIT_RTPRIO);
3866 /* can't set/change the rt policy */
3867 if (policy != p->policy && !rlim_rtprio)
3870 /* can't increase priority */
3871 if (attr->sched_priority > p->rt_priority &&
3872 attr->sched_priority > rlim_rtprio)
3877 * Can't set/change SCHED_DEADLINE policy at all for now
3878 * (safest behavior); in the future we would like to allow
3879 * unprivileged DL tasks to increase their relative deadline
3880 * or reduce their runtime (both ways reducing utilization)
3882 if (dl_policy(policy))
3886 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3887 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3889 if (idle_policy(p->policy) && !idle_policy(policy)) {
3890 if (!can_nice(p, task_nice(p)))
3894 /* can't change other user's priorities */
3895 if (!check_same_owner(p))
3898 /* Normal users shall not reset the sched_reset_on_fork flag */
3899 if (p->sched_reset_on_fork && !reset_on_fork)
3904 retval = security_task_setscheduler(p);
3910 * make sure no PI-waiters arrive (or leave) while we are
3911 * changing the priority of the task:
3913 * To be able to change p->policy safely, the appropriate
3914 * runqueue lock must be held.
3916 rq = task_rq_lock(p, &flags);
3919 * Changing the policy of the stop threads its a very bad idea
3921 if (p == rq->stop) {
3922 task_rq_unlock(rq, p, &flags);
3927 * If not changing anything there's no need to proceed further,
3928 * but store a possible modification of reset_on_fork.
3930 if (unlikely(policy == p->policy)) {
3931 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3933 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3935 if (dl_policy(policy) && dl_param_changed(p, attr))
3938 p->sched_reset_on_fork = reset_on_fork;
3939 task_rq_unlock(rq, p, &flags);
3945 #ifdef CONFIG_RT_GROUP_SCHED
3947 * Do not allow realtime tasks into groups that have no runtime
3950 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3951 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3952 !task_group_is_autogroup(task_group(p))) {
3953 task_rq_unlock(rq, p, &flags);
3958 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3959 cpumask_t *span = rq->rd->span;
3962 * Don't allow tasks with an affinity mask smaller than
3963 * the entire root_domain to become SCHED_DEADLINE. We
3964 * will also fail if there's no bandwidth available.
3966 if (!cpumask_subset(span, &p->cpus_allowed) ||
3967 rq->rd->dl_bw.bw == 0) {
3968 task_rq_unlock(rq, p, &flags);
3975 /* recheck policy now with rq lock held */
3976 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3977 policy = oldpolicy = -1;
3978 task_rq_unlock(rq, p, &flags);
3983 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3984 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3987 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3988 task_rq_unlock(rq, p, &flags);
3992 p->sched_reset_on_fork = reset_on_fork;
3997 * Take priority boosted tasks into account. If the new
3998 * effective priority is unchanged, we just store the new
3999 * normal parameters and do not touch the scheduler class and
4000 * the runqueue. This will be done when the task deboost
4003 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4004 if (new_effective_prio == oldprio) {
4005 __setscheduler_params(p, attr);
4006 task_rq_unlock(rq, p, &flags);
4011 queued = task_on_rq_queued(p);
4012 running = task_current(rq, p);
4014 dequeue_task(rq, p, DEQUEUE_SAVE);
4016 put_prev_task(rq, p);
4018 prev_class = p->sched_class;
4019 __setscheduler(rq, p, attr, pi);
4022 p->sched_class->set_curr_task(rq);
4024 int enqueue_flags = ENQUEUE_RESTORE;
4026 * We enqueue to tail when the priority of a task is
4027 * increased (user space view).
4029 if (oldprio <= p->prio)
4030 enqueue_flags |= ENQUEUE_HEAD;
4032 enqueue_task(rq, p, enqueue_flags);
4035 check_class_changed(rq, p, prev_class, oldprio);
4036 preempt_disable(); /* avoid rq from going away on us */
4037 task_rq_unlock(rq, p, &flags);
4040 rt_mutex_adjust_pi(p);
4043 * Run balance callbacks after we've adjusted the PI chain.
4045 balance_callback(rq);
4051 static int _sched_setscheduler(struct task_struct *p, int policy,
4052 const struct sched_param *param, bool check)
4054 struct sched_attr attr = {
4055 .sched_policy = policy,
4056 .sched_priority = param->sched_priority,
4057 .sched_nice = PRIO_TO_NICE(p->static_prio),
4060 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4061 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4062 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4063 policy &= ~SCHED_RESET_ON_FORK;
4064 attr.sched_policy = policy;
4067 return __sched_setscheduler(p, &attr, check, true);
4070 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4071 * @p: the task in question.
4072 * @policy: new policy.
4073 * @param: structure containing the new RT priority.
4075 * Return: 0 on success. An error code otherwise.
4077 * NOTE that the task may be already dead.
4079 int sched_setscheduler(struct task_struct *p, int policy,
4080 const struct sched_param *param)
4082 return _sched_setscheduler(p, policy, param, true);
4084 EXPORT_SYMBOL_GPL(sched_setscheduler);
4086 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4088 return __sched_setscheduler(p, attr, true, true);
4090 EXPORT_SYMBOL_GPL(sched_setattr);
4093 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4094 * @p: the task in question.
4095 * @policy: new policy.
4096 * @param: structure containing the new RT priority.
4098 * Just like sched_setscheduler, only don't bother checking if the
4099 * current context has permission. For example, this is needed in
4100 * stop_machine(): we create temporary high priority worker threads,
4101 * but our caller might not have that capability.
4103 * Return: 0 on success. An error code otherwise.
4105 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4106 const struct sched_param *param)
4108 return _sched_setscheduler(p, policy, param, false);
4110 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4113 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4115 struct sched_param lparam;
4116 struct task_struct *p;
4119 if (!param || pid < 0)
4121 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4126 p = find_process_by_pid(pid);
4128 retval = sched_setscheduler(p, policy, &lparam);
4135 * Mimics kernel/events/core.c perf_copy_attr().
4137 static int sched_copy_attr(struct sched_attr __user *uattr,
4138 struct sched_attr *attr)
4143 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4147 * zero the full structure, so that a short copy will be nice.
4149 memset(attr, 0, sizeof(*attr));
4151 ret = get_user(size, &uattr->size);
4155 if (size > PAGE_SIZE) /* silly large */
4158 if (!size) /* abi compat */
4159 size = SCHED_ATTR_SIZE_VER0;
4161 if (size < SCHED_ATTR_SIZE_VER0)
4165 * If we're handed a bigger struct than we know of,
4166 * ensure all the unknown bits are 0 - i.e. new
4167 * user-space does not rely on any kernel feature
4168 * extensions we dont know about yet.
4170 if (size > sizeof(*attr)) {
4171 unsigned char __user *addr;
4172 unsigned char __user *end;
4175 addr = (void __user *)uattr + sizeof(*attr);
4176 end = (void __user *)uattr + size;
4178 for (; addr < end; addr++) {
4179 ret = get_user(val, addr);
4185 size = sizeof(*attr);
4188 ret = copy_from_user(attr, uattr, size);
4193 * XXX: do we want to be lenient like existing syscalls; or do we want
4194 * to be strict and return an error on out-of-bounds values?
4196 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4201 put_user(sizeof(*attr), &uattr->size);
4206 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4207 * @pid: the pid in question.
4208 * @policy: new policy.
4209 * @param: structure containing the new RT priority.
4211 * Return: 0 on success. An error code otherwise.
4213 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4214 struct sched_param __user *, param)
4216 /* negative values for policy are not valid */
4220 return do_sched_setscheduler(pid, policy, param);
4224 * sys_sched_setparam - set/change the RT priority of a thread
4225 * @pid: the pid in question.
4226 * @param: structure containing the new RT priority.
4228 * Return: 0 on success. An error code otherwise.
4230 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4232 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4236 * sys_sched_setattr - same as above, but with extended sched_attr
4237 * @pid: the pid in question.
4238 * @uattr: structure containing the extended parameters.
4239 * @flags: for future extension.
4241 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4242 unsigned int, flags)
4244 struct sched_attr attr;
4245 struct task_struct *p;
4248 if (!uattr || pid < 0 || flags)
4251 retval = sched_copy_attr(uattr, &attr);
4255 if ((int)attr.sched_policy < 0)
4260 p = find_process_by_pid(pid);
4262 retval = sched_setattr(p, &attr);
4269 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4270 * @pid: the pid in question.
4272 * Return: On success, the policy of the thread. Otherwise, a negative error
4275 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4277 struct task_struct *p;
4285 p = find_process_by_pid(pid);
4287 retval = security_task_getscheduler(p);
4290 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4297 * sys_sched_getparam - get the RT priority of a thread
4298 * @pid: the pid in question.
4299 * @param: structure containing the RT priority.
4301 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4304 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4306 struct sched_param lp = { .sched_priority = 0 };
4307 struct task_struct *p;
4310 if (!param || pid < 0)
4314 p = find_process_by_pid(pid);
4319 retval = security_task_getscheduler(p);
4323 if (task_has_rt_policy(p))
4324 lp.sched_priority = p->rt_priority;
4328 * This one might sleep, we cannot do it with a spinlock held ...
4330 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4339 static int sched_read_attr(struct sched_attr __user *uattr,
4340 struct sched_attr *attr,
4345 if (!access_ok(VERIFY_WRITE, uattr, usize))
4349 * If we're handed a smaller struct than we know of,
4350 * ensure all the unknown bits are 0 - i.e. old
4351 * user-space does not get uncomplete information.
4353 if (usize < sizeof(*attr)) {
4354 unsigned char *addr;
4357 addr = (void *)attr + usize;
4358 end = (void *)attr + sizeof(*attr);
4360 for (; addr < end; addr++) {
4368 ret = copy_to_user(uattr, attr, attr->size);
4376 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4377 * @pid: the pid in question.
4378 * @uattr: structure containing the extended parameters.
4379 * @size: sizeof(attr) for fwd/bwd comp.
4380 * @flags: for future extension.
4382 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4383 unsigned int, size, unsigned int, flags)
4385 struct sched_attr attr = {
4386 .size = sizeof(struct sched_attr),
4388 struct task_struct *p;
4391 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4392 size < SCHED_ATTR_SIZE_VER0 || flags)
4396 p = find_process_by_pid(pid);
4401 retval = security_task_getscheduler(p);
4405 attr.sched_policy = p->policy;
4406 if (p->sched_reset_on_fork)
4407 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4408 if (task_has_dl_policy(p))
4409 __getparam_dl(p, &attr);
4410 else if (task_has_rt_policy(p))
4411 attr.sched_priority = p->rt_priority;
4413 attr.sched_nice = task_nice(p);
4417 retval = sched_read_attr(uattr, &attr, size);
4425 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4427 cpumask_var_t cpus_allowed, new_mask;
4428 struct task_struct *p;
4433 p = find_process_by_pid(pid);
4439 /* Prevent p going away */
4443 if (p->flags & PF_NO_SETAFFINITY) {
4447 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4451 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4453 goto out_free_cpus_allowed;
4456 if (!check_same_owner(p)) {
4458 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4460 goto out_free_new_mask;
4465 retval = security_task_setscheduler(p);
4467 goto out_free_new_mask;
4470 cpuset_cpus_allowed(p, cpus_allowed);
4471 cpumask_and(new_mask, in_mask, cpus_allowed);
4474 * Since bandwidth control happens on root_domain basis,
4475 * if admission test is enabled, we only admit -deadline
4476 * tasks allowed to run on all the CPUs in the task's
4480 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4482 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4485 goto out_free_new_mask;
4491 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4494 cpuset_cpus_allowed(p, cpus_allowed);
4495 if (!cpumask_subset(new_mask, cpus_allowed)) {
4497 * We must have raced with a concurrent cpuset
4498 * update. Just reset the cpus_allowed to the
4499 * cpuset's cpus_allowed
4501 cpumask_copy(new_mask, cpus_allowed);
4506 free_cpumask_var(new_mask);
4507 out_free_cpus_allowed:
4508 free_cpumask_var(cpus_allowed);
4514 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4515 struct cpumask *new_mask)
4517 if (len < cpumask_size())
4518 cpumask_clear(new_mask);
4519 else if (len > cpumask_size())
4520 len = cpumask_size();
4522 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4526 * sys_sched_setaffinity - set the cpu affinity of a process
4527 * @pid: pid of the process
4528 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4529 * @user_mask_ptr: user-space pointer to the new cpu mask
4531 * Return: 0 on success. An error code otherwise.
4533 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4534 unsigned long __user *, user_mask_ptr)
4536 cpumask_var_t new_mask;
4539 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4542 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4544 retval = sched_setaffinity(pid, new_mask);
4545 free_cpumask_var(new_mask);
4549 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4551 struct task_struct *p;
4552 unsigned long flags;
4558 p = find_process_by_pid(pid);
4562 retval = security_task_getscheduler(p);
4566 raw_spin_lock_irqsave(&p->pi_lock, flags);
4567 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4568 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4577 * sys_sched_getaffinity - get the cpu affinity of a process
4578 * @pid: pid of the process
4579 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4580 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4582 * Return: 0 on success. An error code otherwise.
4584 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4585 unsigned long __user *, user_mask_ptr)
4590 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4592 if (len & (sizeof(unsigned long)-1))
4595 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4598 ret = sched_getaffinity(pid, mask);
4600 size_t retlen = min_t(size_t, len, cpumask_size());
4602 if (copy_to_user(user_mask_ptr, mask, retlen))
4607 free_cpumask_var(mask);
4613 * sys_sched_yield - yield the current processor to other threads.
4615 * This function yields the current CPU to other tasks. If there are no
4616 * other threads running on this CPU then this function will return.
4620 SYSCALL_DEFINE0(sched_yield)
4622 struct rq *rq = this_rq_lock();
4624 schedstat_inc(rq, yld_count);
4625 current->sched_class->yield_task(rq);
4628 * Since we are going to call schedule() anyway, there's
4629 * no need to preempt or enable interrupts:
4631 __release(rq->lock);
4632 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4633 do_raw_spin_unlock(&rq->lock);
4634 sched_preempt_enable_no_resched();
4641 int __sched _cond_resched(void)
4643 if (should_resched(0)) {
4644 preempt_schedule_common();
4649 EXPORT_SYMBOL(_cond_resched);
4652 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4653 * call schedule, and on return reacquire the lock.
4655 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4656 * operations here to prevent schedule() from being called twice (once via
4657 * spin_unlock(), once by hand).
4659 int __cond_resched_lock(spinlock_t *lock)
4661 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4664 lockdep_assert_held(lock);
4666 if (spin_needbreak(lock) || resched) {
4669 preempt_schedule_common();
4677 EXPORT_SYMBOL(__cond_resched_lock);
4679 int __sched __cond_resched_softirq(void)
4681 BUG_ON(!in_softirq());
4683 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4685 preempt_schedule_common();
4691 EXPORT_SYMBOL(__cond_resched_softirq);
4694 * yield - yield the current processor to other threads.
4696 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4698 * The scheduler is at all times free to pick the calling task as the most
4699 * eligible task to run, if removing the yield() call from your code breaks
4700 * it, its already broken.
4702 * Typical broken usage is:
4707 * where one assumes that yield() will let 'the other' process run that will
4708 * make event true. If the current task is a SCHED_FIFO task that will never
4709 * happen. Never use yield() as a progress guarantee!!
4711 * If you want to use yield() to wait for something, use wait_event().
4712 * If you want to use yield() to be 'nice' for others, use cond_resched().
4713 * If you still want to use yield(), do not!
4715 void __sched yield(void)
4717 set_current_state(TASK_RUNNING);
4720 EXPORT_SYMBOL(yield);
4723 * yield_to - yield the current processor to another thread in
4724 * your thread group, or accelerate that thread toward the
4725 * processor it's on.
4727 * @preempt: whether task preemption is allowed or not
4729 * It's the caller's job to ensure that the target task struct
4730 * can't go away on us before we can do any checks.
4733 * true (>0) if we indeed boosted the target task.
4734 * false (0) if we failed to boost the target.
4735 * -ESRCH if there's no task to yield to.
4737 int __sched yield_to(struct task_struct *p, bool preempt)
4739 struct task_struct *curr = current;
4740 struct rq *rq, *p_rq;
4741 unsigned long flags;
4744 local_irq_save(flags);
4750 * If we're the only runnable task on the rq and target rq also
4751 * has only one task, there's absolutely no point in yielding.
4753 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4758 double_rq_lock(rq, p_rq);
4759 if (task_rq(p) != p_rq) {
4760 double_rq_unlock(rq, p_rq);
4764 if (!curr->sched_class->yield_to_task)
4767 if (curr->sched_class != p->sched_class)
4770 if (task_running(p_rq, p) || p->state)
4773 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4775 schedstat_inc(rq, yld_count);
4777 * Make p's CPU reschedule; pick_next_entity takes care of
4780 if (preempt && rq != p_rq)
4785 double_rq_unlock(rq, p_rq);
4787 local_irq_restore(flags);
4794 EXPORT_SYMBOL_GPL(yield_to);
4797 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4798 * that process accounting knows that this is a task in IO wait state.
4800 long __sched io_schedule_timeout(long timeout)
4802 int old_iowait = current->in_iowait;
4806 current->in_iowait = 1;
4807 blk_schedule_flush_plug(current);
4809 delayacct_blkio_start();
4811 atomic_inc(&rq->nr_iowait);
4812 ret = schedule_timeout(timeout);
4813 current->in_iowait = old_iowait;
4814 atomic_dec(&rq->nr_iowait);
4815 delayacct_blkio_end();
4819 EXPORT_SYMBOL(io_schedule_timeout);
4822 * sys_sched_get_priority_max - return maximum RT priority.
4823 * @policy: scheduling class.
4825 * Return: On success, this syscall returns the maximum
4826 * rt_priority that can be used by a given scheduling class.
4827 * On failure, a negative error code is returned.
4829 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4836 ret = MAX_USER_RT_PRIO-1;
4838 case SCHED_DEADLINE:
4849 * sys_sched_get_priority_min - return minimum RT priority.
4850 * @policy: scheduling class.
4852 * Return: On success, this syscall returns the minimum
4853 * rt_priority that can be used by a given scheduling class.
4854 * On failure, a negative error code is returned.
4856 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4865 case SCHED_DEADLINE:
4875 * sys_sched_rr_get_interval - return the default timeslice of a process.
4876 * @pid: pid of the process.
4877 * @interval: userspace pointer to the timeslice value.
4879 * this syscall writes the default timeslice value of a given process
4880 * into the user-space timespec buffer. A value of '0' means infinity.
4882 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4885 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4886 struct timespec __user *, interval)
4888 struct task_struct *p;
4889 unsigned int time_slice;
4890 unsigned long flags;
4900 p = find_process_by_pid(pid);
4904 retval = security_task_getscheduler(p);
4908 rq = task_rq_lock(p, &flags);
4910 if (p->sched_class->get_rr_interval)
4911 time_slice = p->sched_class->get_rr_interval(rq, p);
4912 task_rq_unlock(rq, p, &flags);
4915 jiffies_to_timespec(time_slice, &t);
4916 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4924 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4926 void sched_show_task(struct task_struct *p)
4928 unsigned long free = 0;
4930 unsigned long state = p->state;
4933 state = __ffs(state) + 1;
4934 printk(KERN_INFO "%-15.15s %c", p->comm,
4935 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4936 #if BITS_PER_LONG == 32
4937 if (state == TASK_RUNNING)
4938 printk(KERN_CONT " running ");
4940 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4942 if (state == TASK_RUNNING)
4943 printk(KERN_CONT " running task ");
4945 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4947 #ifdef CONFIG_DEBUG_STACK_USAGE
4948 free = stack_not_used(p);
4953 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4955 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4956 task_pid_nr(p), ppid,
4957 (unsigned long)task_thread_info(p)->flags);
4959 print_worker_info(KERN_INFO, p);
4960 show_stack(p, NULL);
4963 void show_state_filter(unsigned long state_filter)
4965 struct task_struct *g, *p;
4967 #if BITS_PER_LONG == 32
4969 " task PC stack pid father\n");
4972 " task PC stack pid father\n");
4975 for_each_process_thread(g, p) {
4977 * reset the NMI-timeout, listing all files on a slow
4978 * console might take a lot of time:
4979 * Also, reset softlockup watchdogs on all CPUs, because
4980 * another CPU might be blocked waiting for us to process
4983 touch_nmi_watchdog();
4984 touch_all_softlockup_watchdogs();
4985 if (!state_filter || (p->state & state_filter))
4989 #ifdef CONFIG_SCHED_DEBUG
4990 sysrq_sched_debug_show();
4994 * Only show locks if all tasks are dumped:
4997 debug_show_all_locks();
5000 void init_idle_bootup_task(struct task_struct *idle)
5002 idle->sched_class = &idle_sched_class;
5006 * init_idle - set up an idle thread for a given CPU
5007 * @idle: task in question
5008 * @cpu: cpu the idle task belongs to
5010 * NOTE: this function does not set the idle thread's NEED_RESCHED
5011 * flag, to make booting more robust.
5013 void init_idle(struct task_struct *idle, int cpu)
5015 struct rq *rq = cpu_rq(cpu);
5016 unsigned long flags;
5018 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5019 raw_spin_lock(&rq->lock);
5021 __sched_fork(0, idle);
5022 idle->state = TASK_RUNNING;
5023 idle->se.exec_start = sched_clock();
5027 * Its possible that init_idle() gets called multiple times on a task,
5028 * in that case do_set_cpus_allowed() will not do the right thing.
5030 * And since this is boot we can forgo the serialization.
5032 set_cpus_allowed_common(idle, cpumask_of(cpu));
5035 * We're having a chicken and egg problem, even though we are
5036 * holding rq->lock, the cpu isn't yet set to this cpu so the
5037 * lockdep check in task_group() will fail.
5039 * Similar case to sched_fork(). / Alternatively we could
5040 * use task_rq_lock() here and obtain the other rq->lock.
5045 __set_task_cpu(idle, cpu);
5048 rq->curr = rq->idle = idle;
5049 idle->on_rq = TASK_ON_RQ_QUEUED;
5053 raw_spin_unlock(&rq->lock);
5054 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5056 /* Set the preempt count _outside_ the spinlocks! */
5057 init_idle_preempt_count(idle, cpu);
5060 * The idle tasks have their own, simple scheduling class:
5062 idle->sched_class = &idle_sched_class;
5063 ftrace_graph_init_idle_task(idle, cpu);
5064 vtime_init_idle(idle, cpu);
5066 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5070 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5071 const struct cpumask *trial)
5073 int ret = 1, trial_cpus;
5074 struct dl_bw *cur_dl_b;
5075 unsigned long flags;
5077 if (!cpumask_weight(cur))
5080 rcu_read_lock_sched();
5081 cur_dl_b = dl_bw_of(cpumask_any(cur));
5082 trial_cpus = cpumask_weight(trial);
5084 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5085 if (cur_dl_b->bw != -1 &&
5086 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5088 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5089 rcu_read_unlock_sched();
5094 int task_can_attach(struct task_struct *p,
5095 const struct cpumask *cs_cpus_allowed)
5100 * Kthreads which disallow setaffinity shouldn't be moved
5101 * to a new cpuset; we don't want to change their cpu
5102 * affinity and isolating such threads by their set of
5103 * allowed nodes is unnecessary. Thus, cpusets are not
5104 * applicable for such threads. This prevents checking for
5105 * success of set_cpus_allowed_ptr() on all attached tasks
5106 * before cpus_allowed may be changed.
5108 if (p->flags & PF_NO_SETAFFINITY) {
5114 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5116 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5121 unsigned long flags;
5123 rcu_read_lock_sched();
5124 dl_b = dl_bw_of(dest_cpu);
5125 raw_spin_lock_irqsave(&dl_b->lock, flags);
5126 cpus = dl_bw_cpus(dest_cpu);
5127 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5132 * We reserve space for this task in the destination
5133 * root_domain, as we can't fail after this point.
5134 * We will free resources in the source root_domain
5135 * later on (see set_cpus_allowed_dl()).
5137 __dl_add(dl_b, p->dl.dl_bw);
5139 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5140 rcu_read_unlock_sched();
5150 #ifdef CONFIG_NUMA_BALANCING
5151 /* Migrate current task p to target_cpu */
5152 int migrate_task_to(struct task_struct *p, int target_cpu)
5154 struct migration_arg arg = { p, target_cpu };
5155 int curr_cpu = task_cpu(p);
5157 if (curr_cpu == target_cpu)
5160 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5163 /* TODO: This is not properly updating schedstats */
5165 trace_sched_move_numa(p, curr_cpu, target_cpu);
5166 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5170 * Requeue a task on a given node and accurately track the number of NUMA
5171 * tasks on the runqueues
5173 void sched_setnuma(struct task_struct *p, int nid)
5176 unsigned long flags;
5177 bool queued, running;
5179 rq = task_rq_lock(p, &flags);
5180 queued = task_on_rq_queued(p);
5181 running = task_current(rq, p);
5184 dequeue_task(rq, p, DEQUEUE_SAVE);
5186 put_prev_task(rq, p);
5188 p->numa_preferred_nid = nid;
5191 p->sched_class->set_curr_task(rq);
5193 enqueue_task(rq, p, ENQUEUE_RESTORE);
5194 task_rq_unlock(rq, p, &flags);
5196 #endif /* CONFIG_NUMA_BALANCING */
5198 #ifdef CONFIG_HOTPLUG_CPU
5200 * Ensures that the idle task is using init_mm right before its cpu goes
5203 void idle_task_exit(void)
5205 struct mm_struct *mm = current->active_mm;
5207 BUG_ON(cpu_online(smp_processor_id()));
5209 if (mm != &init_mm) {
5210 switch_mm(mm, &init_mm, current);
5211 finish_arch_post_lock_switch();
5217 * Since this CPU is going 'away' for a while, fold any nr_active delta
5218 * we might have. Assumes we're called after migrate_tasks() so that the
5219 * nr_active count is stable.
5221 * Also see the comment "Global load-average calculations".
5223 static void calc_load_migrate(struct rq *rq)
5225 long delta = calc_load_fold_active(rq);
5227 atomic_long_add(delta, &calc_load_tasks);
5230 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5234 static const struct sched_class fake_sched_class = {
5235 .put_prev_task = put_prev_task_fake,
5238 static struct task_struct fake_task = {
5240 * Avoid pull_{rt,dl}_task()
5242 .prio = MAX_PRIO + 1,
5243 .sched_class = &fake_sched_class,
5247 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5248 * try_to_wake_up()->select_task_rq().
5250 * Called with rq->lock held even though we'er in stop_machine() and
5251 * there's no concurrency possible, we hold the required locks anyway
5252 * because of lock validation efforts.
5254 static void migrate_tasks(struct rq *dead_rq)
5256 struct rq *rq = dead_rq;
5257 struct task_struct *next, *stop = rq->stop;
5261 * Fudge the rq selection such that the below task selection loop
5262 * doesn't get stuck on the currently eligible stop task.
5264 * We're currently inside stop_machine() and the rq is either stuck
5265 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5266 * either way we should never end up calling schedule() until we're
5272 * put_prev_task() and pick_next_task() sched
5273 * class method both need to have an up-to-date
5274 * value of rq->clock[_task]
5276 update_rq_clock(rq);
5280 * There's this thread running, bail when that's the only
5283 if (rq->nr_running == 1)
5287 * pick_next_task assumes pinned rq->lock.
5289 lockdep_pin_lock(&rq->lock);
5290 next = pick_next_task(rq, &fake_task);
5292 next->sched_class->put_prev_task(rq, next);
5295 * Rules for changing task_struct::cpus_allowed are holding
5296 * both pi_lock and rq->lock, such that holding either
5297 * stabilizes the mask.
5299 * Drop rq->lock is not quite as disastrous as it usually is
5300 * because !cpu_active at this point, which means load-balance
5301 * will not interfere. Also, stop-machine.
5303 lockdep_unpin_lock(&rq->lock);
5304 raw_spin_unlock(&rq->lock);
5305 raw_spin_lock(&next->pi_lock);
5306 raw_spin_lock(&rq->lock);
5309 * Since we're inside stop-machine, _nothing_ should have
5310 * changed the task, WARN if weird stuff happened, because in
5311 * that case the above rq->lock drop is a fail too.
5313 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5314 raw_spin_unlock(&next->pi_lock);
5318 /* Find suitable destination for @next, with force if needed. */
5319 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5321 rq = __migrate_task(rq, next, dest_cpu);
5322 if (rq != dead_rq) {
5323 raw_spin_unlock(&rq->lock);
5325 raw_spin_lock(&rq->lock);
5327 raw_spin_unlock(&next->pi_lock);
5332 #endif /* CONFIG_HOTPLUG_CPU */
5334 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5336 static struct ctl_table sd_ctl_dir[] = {
5338 .procname = "sched_domain",
5344 static struct ctl_table sd_ctl_root[] = {
5346 .procname = "kernel",
5348 .child = sd_ctl_dir,
5353 static struct ctl_table *sd_alloc_ctl_entry(int n)
5355 struct ctl_table *entry =
5356 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5361 static void sd_free_ctl_entry(struct ctl_table **tablep)
5363 struct ctl_table *entry;
5366 * In the intermediate directories, both the child directory and
5367 * procname are dynamically allocated and could fail but the mode
5368 * will always be set. In the lowest directory the names are
5369 * static strings and all have proc handlers.
5371 for (entry = *tablep; entry->mode; entry++) {
5373 sd_free_ctl_entry(&entry->child);
5374 if (entry->proc_handler == NULL)
5375 kfree(entry->procname);
5382 static int min_load_idx = 0;
5383 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5386 set_table_entry(struct ctl_table *entry,
5387 const char *procname, void *data, int maxlen,
5388 umode_t mode, proc_handler *proc_handler,
5391 entry->procname = procname;
5393 entry->maxlen = maxlen;
5395 entry->proc_handler = proc_handler;
5398 entry->extra1 = &min_load_idx;
5399 entry->extra2 = &max_load_idx;
5403 static struct ctl_table *
5404 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5406 struct ctl_table *table = sd_alloc_ctl_entry(14);
5411 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5412 sizeof(long), 0644, proc_doulongvec_minmax, false);
5413 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5414 sizeof(long), 0644, proc_doulongvec_minmax, false);
5415 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5416 sizeof(int), 0644, proc_dointvec_minmax, true);
5417 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5418 sizeof(int), 0644, proc_dointvec_minmax, true);
5419 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5420 sizeof(int), 0644, proc_dointvec_minmax, true);
5421 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5422 sizeof(int), 0644, proc_dointvec_minmax, true);
5423 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5424 sizeof(int), 0644, proc_dointvec_minmax, true);
5425 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5426 sizeof(int), 0644, proc_dointvec_minmax, false);
5427 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5428 sizeof(int), 0644, proc_dointvec_minmax, false);
5429 set_table_entry(&table[9], "cache_nice_tries",
5430 &sd->cache_nice_tries,
5431 sizeof(int), 0644, proc_dointvec_minmax, false);
5432 set_table_entry(&table[10], "flags", &sd->flags,
5433 sizeof(int), 0644, proc_dointvec_minmax, false);
5434 set_table_entry(&table[11], "max_newidle_lb_cost",
5435 &sd->max_newidle_lb_cost,
5436 sizeof(long), 0644, proc_doulongvec_minmax, false);
5437 set_table_entry(&table[12], "name", sd->name,
5438 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5439 /* &table[13] is terminator */
5444 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5446 struct ctl_table *entry, *table;
5447 struct sched_domain *sd;
5448 int domain_num = 0, i;
5451 for_each_domain(cpu, sd)
5453 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5458 for_each_domain(cpu, sd) {
5459 snprintf(buf, 32, "domain%d", i);
5460 entry->procname = kstrdup(buf, GFP_KERNEL);
5462 entry->child = sd_alloc_ctl_domain_table(sd);
5469 static struct ctl_table_header *sd_sysctl_header;
5470 static void register_sched_domain_sysctl(void)
5472 int i, cpu_num = num_possible_cpus();
5473 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5476 WARN_ON(sd_ctl_dir[0].child);
5477 sd_ctl_dir[0].child = entry;
5482 for_each_possible_cpu(i) {
5483 snprintf(buf, 32, "cpu%d", i);
5484 entry->procname = kstrdup(buf, GFP_KERNEL);
5486 entry->child = sd_alloc_ctl_cpu_table(i);
5490 WARN_ON(sd_sysctl_header);
5491 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5494 /* may be called multiple times per register */
5495 static void unregister_sched_domain_sysctl(void)
5497 unregister_sysctl_table(sd_sysctl_header);
5498 sd_sysctl_header = NULL;
5499 if (sd_ctl_dir[0].child)
5500 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5503 static void register_sched_domain_sysctl(void)
5506 static void unregister_sched_domain_sysctl(void)
5509 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5511 static void set_rq_online(struct rq *rq)
5514 const struct sched_class *class;
5516 cpumask_set_cpu(rq->cpu, rq->rd->online);
5519 for_each_class(class) {
5520 if (class->rq_online)
5521 class->rq_online(rq);
5526 static void set_rq_offline(struct rq *rq)
5529 const struct sched_class *class;
5531 for_each_class(class) {
5532 if (class->rq_offline)
5533 class->rq_offline(rq);
5536 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5542 * migration_call - callback that gets triggered when a CPU is added.
5543 * Here we can start up the necessary migration thread for the new CPU.
5546 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5548 int cpu = (long)hcpu;
5549 unsigned long flags;
5550 struct rq *rq = cpu_rq(cpu);
5552 switch (action & ~CPU_TASKS_FROZEN) {
5554 case CPU_UP_PREPARE:
5555 rq->calc_load_update = calc_load_update;
5556 account_reset_rq(rq);
5560 /* Update our root-domain */
5561 raw_spin_lock_irqsave(&rq->lock, flags);
5563 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5567 raw_spin_unlock_irqrestore(&rq->lock, flags);
5570 #ifdef CONFIG_HOTPLUG_CPU
5572 sched_ttwu_pending();
5573 /* Update our root-domain */
5574 raw_spin_lock_irqsave(&rq->lock, flags);
5576 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5580 BUG_ON(rq->nr_running != 1); /* the migration thread */
5581 raw_spin_unlock_irqrestore(&rq->lock, flags);
5585 calc_load_migrate(rq);
5590 update_max_interval();
5596 * Register at high priority so that task migration (migrate_all_tasks)
5597 * happens before everything else. This has to be lower priority than
5598 * the notifier in the perf_event subsystem, though.
5600 static struct notifier_block migration_notifier = {
5601 .notifier_call = migration_call,
5602 .priority = CPU_PRI_MIGRATION,
5605 static void set_cpu_rq_start_time(void)
5607 int cpu = smp_processor_id();
5608 struct rq *rq = cpu_rq(cpu);
5609 rq->age_stamp = sched_clock_cpu(cpu);
5612 static int sched_cpu_active(struct notifier_block *nfb,
5613 unsigned long action, void *hcpu)
5615 int cpu = (long)hcpu;
5617 switch (action & ~CPU_TASKS_FROZEN) {
5619 set_cpu_rq_start_time();
5624 * At this point a starting CPU has marked itself as online via
5625 * set_cpu_online(). But it might not yet have marked itself
5626 * as active, which is essential from here on.
5628 set_cpu_active(cpu, true);
5629 stop_machine_unpark(cpu);
5632 case CPU_DOWN_FAILED:
5633 set_cpu_active(cpu, true);
5641 static int sched_cpu_inactive(struct notifier_block *nfb,
5642 unsigned long action, void *hcpu)
5644 switch (action & ~CPU_TASKS_FROZEN) {
5645 case CPU_DOWN_PREPARE:
5646 set_cpu_active((long)hcpu, false);
5653 static int __init migration_init(void)
5655 void *cpu = (void *)(long)smp_processor_id();
5658 /* Initialize migration for the boot CPU */
5659 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5660 BUG_ON(err == NOTIFY_BAD);
5661 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5662 register_cpu_notifier(&migration_notifier);
5664 /* Register cpu active notifiers */
5665 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5666 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5670 early_initcall(migration_init);
5672 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5674 #ifdef CONFIG_SCHED_DEBUG
5676 static __read_mostly int sched_debug_enabled;
5678 static int __init sched_debug_setup(char *str)
5680 sched_debug_enabled = 1;
5684 early_param("sched_debug", sched_debug_setup);
5686 static inline bool sched_debug(void)
5688 return sched_debug_enabled;
5691 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5692 struct cpumask *groupmask)
5694 struct sched_group *group = sd->groups;
5696 cpumask_clear(groupmask);
5698 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5700 if (!(sd->flags & SD_LOAD_BALANCE)) {
5701 printk("does not load-balance\n");
5703 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5708 printk(KERN_CONT "span %*pbl level %s\n",
5709 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5711 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5712 printk(KERN_ERR "ERROR: domain->span does not contain "
5715 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5716 printk(KERN_ERR "ERROR: domain->groups does not contain"
5720 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5724 printk(KERN_ERR "ERROR: group is NULL\n");
5728 if (!cpumask_weight(sched_group_cpus(group))) {
5729 printk(KERN_CONT "\n");
5730 printk(KERN_ERR "ERROR: empty group\n");
5734 if (!(sd->flags & SD_OVERLAP) &&
5735 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5736 printk(KERN_CONT "\n");
5737 printk(KERN_ERR "ERROR: repeated CPUs\n");
5741 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5743 printk(KERN_CONT " %*pbl",
5744 cpumask_pr_args(sched_group_cpus(group)));
5745 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5746 printk(KERN_CONT " (cpu_capacity = %d)",
5747 group->sgc->capacity);
5750 group = group->next;
5751 } while (group != sd->groups);
5752 printk(KERN_CONT "\n");
5754 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5755 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5758 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5759 printk(KERN_ERR "ERROR: parent span is not a superset "
5760 "of domain->span\n");
5764 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5768 if (!sched_debug_enabled)
5772 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5776 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5779 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5787 #else /* !CONFIG_SCHED_DEBUG */
5788 # define sched_domain_debug(sd, cpu) do { } while (0)
5789 static inline bool sched_debug(void)
5793 #endif /* CONFIG_SCHED_DEBUG */
5795 static int sd_degenerate(struct sched_domain *sd)
5797 if (cpumask_weight(sched_domain_span(sd)) == 1)
5800 /* Following flags need at least 2 groups */
5801 if (sd->flags & (SD_LOAD_BALANCE |
5802 SD_BALANCE_NEWIDLE |
5805 SD_SHARE_CPUCAPACITY |
5806 SD_SHARE_PKG_RESOURCES |
5807 SD_SHARE_POWERDOMAIN)) {
5808 if (sd->groups != sd->groups->next)
5812 /* Following flags don't use groups */
5813 if (sd->flags & (SD_WAKE_AFFINE))
5820 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5822 unsigned long cflags = sd->flags, pflags = parent->flags;
5824 if (sd_degenerate(parent))
5827 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5830 /* Flags needing groups don't count if only 1 group in parent */
5831 if (parent->groups == parent->groups->next) {
5832 pflags &= ~(SD_LOAD_BALANCE |
5833 SD_BALANCE_NEWIDLE |
5836 SD_SHARE_CPUCAPACITY |
5837 SD_SHARE_PKG_RESOURCES |
5839 SD_SHARE_POWERDOMAIN);
5840 if (nr_node_ids == 1)
5841 pflags &= ~SD_SERIALIZE;
5843 if (~cflags & pflags)
5849 static void free_rootdomain(struct rcu_head *rcu)
5851 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5853 cpupri_cleanup(&rd->cpupri);
5854 cpudl_cleanup(&rd->cpudl);
5855 free_cpumask_var(rd->dlo_mask);
5856 free_cpumask_var(rd->rto_mask);
5857 free_cpumask_var(rd->online);
5858 free_cpumask_var(rd->span);
5862 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5864 struct root_domain *old_rd = NULL;
5865 unsigned long flags;
5867 raw_spin_lock_irqsave(&rq->lock, flags);
5872 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5875 cpumask_clear_cpu(rq->cpu, old_rd->span);
5878 * If we dont want to free the old_rd yet then
5879 * set old_rd to NULL to skip the freeing later
5882 if (!atomic_dec_and_test(&old_rd->refcount))
5886 atomic_inc(&rd->refcount);
5889 cpumask_set_cpu(rq->cpu, rd->span);
5890 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5893 raw_spin_unlock_irqrestore(&rq->lock, flags);
5896 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5899 static int init_rootdomain(struct root_domain *rd)
5901 memset(rd, 0, sizeof(*rd));
5903 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5905 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5907 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5909 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5912 init_dl_bw(&rd->dl_bw);
5913 if (cpudl_init(&rd->cpudl) != 0)
5916 if (cpupri_init(&rd->cpupri) != 0)
5921 free_cpumask_var(rd->rto_mask);
5923 free_cpumask_var(rd->dlo_mask);
5925 free_cpumask_var(rd->online);
5927 free_cpumask_var(rd->span);
5933 * By default the system creates a single root-domain with all cpus as
5934 * members (mimicking the global state we have today).
5936 struct root_domain def_root_domain;
5938 static void init_defrootdomain(void)
5940 init_rootdomain(&def_root_domain);
5942 atomic_set(&def_root_domain.refcount, 1);
5945 static struct root_domain *alloc_rootdomain(void)
5947 struct root_domain *rd;
5949 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5953 if (init_rootdomain(rd) != 0) {
5961 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5963 struct sched_group *tmp, *first;
5972 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5977 } while (sg != first);
5980 static void free_sched_domain(struct rcu_head *rcu)
5982 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5985 * If its an overlapping domain it has private groups, iterate and
5988 if (sd->flags & SD_OVERLAP) {
5989 free_sched_groups(sd->groups, 1);
5990 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5991 kfree(sd->groups->sgc);
5997 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5999 call_rcu(&sd->rcu, free_sched_domain);
6002 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6004 for (; sd; sd = sd->parent)
6005 destroy_sched_domain(sd, cpu);
6009 * Keep a special pointer to the highest sched_domain that has
6010 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6011 * allows us to avoid some pointer chasing select_idle_sibling().
6013 * Also keep a unique ID per domain (we use the first cpu number in
6014 * the cpumask of the domain), this allows us to quickly tell if
6015 * two cpus are in the same cache domain, see cpus_share_cache().
6017 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6018 DEFINE_PER_CPU(int, sd_llc_size);
6019 DEFINE_PER_CPU(int, sd_llc_id);
6020 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6021 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6022 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6024 static void update_top_cache_domain(int cpu)
6026 struct sched_domain *sd;
6027 struct sched_domain *busy_sd = NULL;
6031 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6033 id = cpumask_first(sched_domain_span(sd));
6034 size = cpumask_weight(sched_domain_span(sd));
6035 busy_sd = sd->parent; /* sd_busy */
6037 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6039 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6040 per_cpu(sd_llc_size, cpu) = size;
6041 per_cpu(sd_llc_id, cpu) = id;
6043 sd = lowest_flag_domain(cpu, SD_NUMA);
6044 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6046 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6047 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6051 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6052 * hold the hotplug lock.
6055 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6057 struct rq *rq = cpu_rq(cpu);
6058 struct sched_domain *tmp;
6060 /* Remove the sched domains which do not contribute to scheduling. */
6061 for (tmp = sd; tmp; ) {
6062 struct sched_domain *parent = tmp->parent;
6066 if (sd_parent_degenerate(tmp, parent)) {
6067 tmp->parent = parent->parent;
6069 parent->parent->child = tmp;
6071 * Transfer SD_PREFER_SIBLING down in case of a
6072 * degenerate parent; the spans match for this
6073 * so the property transfers.
6075 if (parent->flags & SD_PREFER_SIBLING)
6076 tmp->flags |= SD_PREFER_SIBLING;
6077 destroy_sched_domain(parent, cpu);
6082 if (sd && sd_degenerate(sd)) {
6085 destroy_sched_domain(tmp, cpu);
6090 sched_domain_debug(sd, cpu);
6092 rq_attach_root(rq, rd);
6094 rcu_assign_pointer(rq->sd, sd);
6095 destroy_sched_domains(tmp, cpu);
6097 update_top_cache_domain(cpu);
6100 /* Setup the mask of cpus configured for isolated domains */
6101 static int __init isolated_cpu_setup(char *str)
6103 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6104 cpulist_parse(str, cpu_isolated_map);
6108 __setup("isolcpus=", isolated_cpu_setup);
6111 struct sched_domain ** __percpu sd;
6112 struct root_domain *rd;
6123 * Build an iteration mask that can exclude certain CPUs from the upwards
6126 * Asymmetric node setups can result in situations where the domain tree is of
6127 * unequal depth, make sure to skip domains that already cover the entire
6130 * In that case build_sched_domains() will have terminated the iteration early
6131 * and our sibling sd spans will be empty. Domains should always include the
6132 * cpu they're built on, so check that.
6135 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6137 const struct cpumask *span = sched_domain_span(sd);
6138 struct sd_data *sdd = sd->private;
6139 struct sched_domain *sibling;
6142 for_each_cpu(i, span) {
6143 sibling = *per_cpu_ptr(sdd->sd, i);
6144 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6147 cpumask_set_cpu(i, sched_group_mask(sg));
6152 * Return the canonical balance cpu for this group, this is the first cpu
6153 * of this group that's also in the iteration mask.
6155 int group_balance_cpu(struct sched_group *sg)
6157 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6161 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6163 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6164 const struct cpumask *span = sched_domain_span(sd);
6165 struct cpumask *covered = sched_domains_tmpmask;
6166 struct sd_data *sdd = sd->private;
6167 struct sched_domain *sibling;
6170 cpumask_clear(covered);
6172 for_each_cpu(i, span) {
6173 struct cpumask *sg_span;
6175 if (cpumask_test_cpu(i, covered))
6178 sibling = *per_cpu_ptr(sdd->sd, i);
6180 /* See the comment near build_group_mask(). */
6181 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6184 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6185 GFP_KERNEL, cpu_to_node(cpu));
6190 sg_span = sched_group_cpus(sg);
6192 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6194 cpumask_set_cpu(i, sg_span);
6196 cpumask_or(covered, covered, sg_span);
6198 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6199 if (atomic_inc_return(&sg->sgc->ref) == 1)
6200 build_group_mask(sd, sg);
6203 * Initialize sgc->capacity such that even if we mess up the
6204 * domains and no possible iteration will get us here, we won't
6207 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6210 * Make sure the first group of this domain contains the
6211 * canonical balance cpu. Otherwise the sched_domain iteration
6212 * breaks. See update_sg_lb_stats().
6214 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6215 group_balance_cpu(sg) == cpu)
6225 sd->groups = groups;
6230 free_sched_groups(first, 0);
6235 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6237 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6238 struct sched_domain *child = sd->child;
6241 cpu = cpumask_first(sched_domain_span(child));
6244 *sg = *per_cpu_ptr(sdd->sg, cpu);
6245 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6246 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6253 * build_sched_groups will build a circular linked list of the groups
6254 * covered by the given span, and will set each group's ->cpumask correctly,
6255 * and ->cpu_capacity to 0.
6257 * Assumes the sched_domain tree is fully constructed
6260 build_sched_groups(struct sched_domain *sd, int cpu)
6262 struct sched_group *first = NULL, *last = NULL;
6263 struct sd_data *sdd = sd->private;
6264 const struct cpumask *span = sched_domain_span(sd);
6265 struct cpumask *covered;
6268 get_group(cpu, sdd, &sd->groups);
6269 atomic_inc(&sd->groups->ref);
6271 if (cpu != cpumask_first(span))
6274 lockdep_assert_held(&sched_domains_mutex);
6275 covered = sched_domains_tmpmask;
6277 cpumask_clear(covered);
6279 for_each_cpu(i, span) {
6280 struct sched_group *sg;
6283 if (cpumask_test_cpu(i, covered))
6286 group = get_group(i, sdd, &sg);
6287 cpumask_setall(sched_group_mask(sg));
6289 for_each_cpu(j, span) {
6290 if (get_group(j, sdd, NULL) != group)
6293 cpumask_set_cpu(j, covered);
6294 cpumask_set_cpu(j, sched_group_cpus(sg));
6309 * Initialize sched groups cpu_capacity.
6311 * cpu_capacity indicates the capacity of sched group, which is used while
6312 * distributing the load between different sched groups in a sched domain.
6313 * Typically cpu_capacity for all the groups in a sched domain will be same
6314 * unless there are asymmetries in the topology. If there are asymmetries,
6315 * group having more cpu_capacity will pickup more load compared to the
6316 * group having less cpu_capacity.
6318 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6320 struct sched_group *sg = sd->groups;
6325 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6327 } while (sg != sd->groups);
6329 if (cpu != group_balance_cpu(sg))
6332 update_group_capacity(sd, cpu);
6333 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6337 * Initializers for schedule domains
6338 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6341 static int default_relax_domain_level = -1;
6342 int sched_domain_level_max;
6344 static int __init setup_relax_domain_level(char *str)
6346 if (kstrtoint(str, 0, &default_relax_domain_level))
6347 pr_warn("Unable to set relax_domain_level\n");
6351 __setup("relax_domain_level=", setup_relax_domain_level);
6353 static void set_domain_attribute(struct sched_domain *sd,
6354 struct sched_domain_attr *attr)
6358 if (!attr || attr->relax_domain_level < 0) {
6359 if (default_relax_domain_level < 0)
6362 request = default_relax_domain_level;
6364 request = attr->relax_domain_level;
6365 if (request < sd->level) {
6366 /* turn off idle balance on this domain */
6367 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6369 /* turn on idle balance on this domain */
6370 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6374 static void __sdt_free(const struct cpumask *cpu_map);
6375 static int __sdt_alloc(const struct cpumask *cpu_map);
6377 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6378 const struct cpumask *cpu_map)
6382 if (!atomic_read(&d->rd->refcount))
6383 free_rootdomain(&d->rd->rcu); /* fall through */
6385 free_percpu(d->sd); /* fall through */
6387 __sdt_free(cpu_map); /* fall through */
6393 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6394 const struct cpumask *cpu_map)
6396 memset(d, 0, sizeof(*d));
6398 if (__sdt_alloc(cpu_map))
6399 return sa_sd_storage;
6400 d->sd = alloc_percpu(struct sched_domain *);
6402 return sa_sd_storage;
6403 d->rd = alloc_rootdomain();
6406 return sa_rootdomain;
6410 * NULL the sd_data elements we've used to build the sched_domain and
6411 * sched_group structure so that the subsequent __free_domain_allocs()
6412 * will not free the data we're using.
6414 static void claim_allocations(int cpu, struct sched_domain *sd)
6416 struct sd_data *sdd = sd->private;
6418 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6419 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6421 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6422 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6424 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6425 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6429 static int sched_domains_numa_levels;
6430 enum numa_topology_type sched_numa_topology_type;
6431 static int *sched_domains_numa_distance;
6432 int sched_max_numa_distance;
6433 static struct cpumask ***sched_domains_numa_masks;
6434 static int sched_domains_curr_level;
6438 * SD_flags allowed in topology descriptions.
6440 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6441 * SD_SHARE_PKG_RESOURCES - describes shared caches
6442 * SD_NUMA - describes NUMA topologies
6443 * SD_SHARE_POWERDOMAIN - describes shared power domain
6446 * SD_ASYM_PACKING - describes SMT quirks
6448 #define TOPOLOGY_SD_FLAGS \
6449 (SD_SHARE_CPUCAPACITY | \
6450 SD_SHARE_PKG_RESOURCES | \
6453 SD_SHARE_POWERDOMAIN)
6455 static struct sched_domain *
6456 sd_init(struct sched_domain_topology_level *tl, int cpu)
6458 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6459 int sd_weight, sd_flags = 0;
6463 * Ugly hack to pass state to sd_numa_mask()...
6465 sched_domains_curr_level = tl->numa_level;
6468 sd_weight = cpumask_weight(tl->mask(cpu));
6471 sd_flags = (*tl->sd_flags)();
6472 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6473 "wrong sd_flags in topology description\n"))
6474 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6476 *sd = (struct sched_domain){
6477 .min_interval = sd_weight,
6478 .max_interval = 2*sd_weight,
6480 .imbalance_pct = 125,
6482 .cache_nice_tries = 0,
6489 .flags = 1*SD_LOAD_BALANCE
6490 | 1*SD_BALANCE_NEWIDLE
6495 | 0*SD_SHARE_CPUCAPACITY
6496 | 0*SD_SHARE_PKG_RESOURCES
6498 | 0*SD_PREFER_SIBLING
6503 .last_balance = jiffies,
6504 .balance_interval = sd_weight,
6506 .max_newidle_lb_cost = 0,
6507 .next_decay_max_lb_cost = jiffies,
6508 #ifdef CONFIG_SCHED_DEBUG
6514 * Convert topological properties into behaviour.
6517 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6518 sd->flags |= SD_PREFER_SIBLING;
6519 sd->imbalance_pct = 110;
6520 sd->smt_gain = 1178; /* ~15% */
6522 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6523 sd->imbalance_pct = 117;
6524 sd->cache_nice_tries = 1;
6528 } else if (sd->flags & SD_NUMA) {
6529 sd->cache_nice_tries = 2;
6533 sd->flags |= SD_SERIALIZE;
6534 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6535 sd->flags &= ~(SD_BALANCE_EXEC |
6542 sd->flags |= SD_PREFER_SIBLING;
6543 sd->cache_nice_tries = 1;
6548 sd->private = &tl->data;
6554 * Topology list, bottom-up.
6556 static struct sched_domain_topology_level default_topology[] = {
6557 #ifdef CONFIG_SCHED_SMT
6558 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6560 #ifdef CONFIG_SCHED_MC
6561 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6563 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6567 static struct sched_domain_topology_level *sched_domain_topology =
6570 #define for_each_sd_topology(tl) \
6571 for (tl = sched_domain_topology; tl->mask; tl++)
6573 void set_sched_topology(struct sched_domain_topology_level *tl)
6575 sched_domain_topology = tl;
6580 static const struct cpumask *sd_numa_mask(int cpu)
6582 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6585 static void sched_numa_warn(const char *str)
6587 static int done = false;
6595 printk(KERN_WARNING "ERROR: %s\n\n", str);
6597 for (i = 0; i < nr_node_ids; i++) {
6598 printk(KERN_WARNING " ");
6599 for (j = 0; j < nr_node_ids; j++)
6600 printk(KERN_CONT "%02d ", node_distance(i,j));
6601 printk(KERN_CONT "\n");
6603 printk(KERN_WARNING "\n");
6606 bool find_numa_distance(int distance)
6610 if (distance == node_distance(0, 0))
6613 for (i = 0; i < sched_domains_numa_levels; i++) {
6614 if (sched_domains_numa_distance[i] == distance)
6622 * A system can have three types of NUMA topology:
6623 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6624 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6625 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6627 * The difference between a glueless mesh topology and a backplane
6628 * topology lies in whether communication between not directly
6629 * connected nodes goes through intermediary nodes (where programs
6630 * could run), or through backplane controllers. This affects
6631 * placement of programs.
6633 * The type of topology can be discerned with the following tests:
6634 * - If the maximum distance between any nodes is 1 hop, the system
6635 * is directly connected.
6636 * - If for two nodes A and B, located N > 1 hops away from each other,
6637 * there is an intermediary node C, which is < N hops away from both
6638 * nodes A and B, the system is a glueless mesh.
6640 static void init_numa_topology_type(void)
6644 n = sched_max_numa_distance;
6646 if (sched_domains_numa_levels <= 1) {
6647 sched_numa_topology_type = NUMA_DIRECT;
6651 for_each_online_node(a) {
6652 for_each_online_node(b) {
6653 /* Find two nodes furthest removed from each other. */
6654 if (node_distance(a, b) < n)
6657 /* Is there an intermediary node between a and b? */
6658 for_each_online_node(c) {
6659 if (node_distance(a, c) < n &&
6660 node_distance(b, c) < n) {
6661 sched_numa_topology_type =
6667 sched_numa_topology_type = NUMA_BACKPLANE;
6673 static void sched_init_numa(void)
6675 int next_distance, curr_distance = node_distance(0, 0);
6676 struct sched_domain_topology_level *tl;
6680 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6681 if (!sched_domains_numa_distance)
6685 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6686 * unique distances in the node_distance() table.
6688 * Assumes node_distance(0,j) includes all distances in
6689 * node_distance(i,j) in order to avoid cubic time.
6691 next_distance = curr_distance;
6692 for (i = 0; i < nr_node_ids; i++) {
6693 for (j = 0; j < nr_node_ids; j++) {
6694 for (k = 0; k < nr_node_ids; k++) {
6695 int distance = node_distance(i, k);
6697 if (distance > curr_distance &&
6698 (distance < next_distance ||
6699 next_distance == curr_distance))
6700 next_distance = distance;
6703 * While not a strong assumption it would be nice to know
6704 * about cases where if node A is connected to B, B is not
6705 * equally connected to A.
6707 if (sched_debug() && node_distance(k, i) != distance)
6708 sched_numa_warn("Node-distance not symmetric");
6710 if (sched_debug() && i && !find_numa_distance(distance))
6711 sched_numa_warn("Node-0 not representative");
6713 if (next_distance != curr_distance) {
6714 sched_domains_numa_distance[level++] = next_distance;
6715 sched_domains_numa_levels = level;
6716 curr_distance = next_distance;
6721 * In case of sched_debug() we verify the above assumption.
6731 * 'level' contains the number of unique distances, excluding the
6732 * identity distance node_distance(i,i).
6734 * The sched_domains_numa_distance[] array includes the actual distance
6739 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6740 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6741 * the array will contain less then 'level' members. This could be
6742 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6743 * in other functions.
6745 * We reset it to 'level' at the end of this function.
6747 sched_domains_numa_levels = 0;
6749 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6750 if (!sched_domains_numa_masks)
6754 * Now for each level, construct a mask per node which contains all
6755 * cpus of nodes that are that many hops away from us.
6757 for (i = 0; i < level; i++) {
6758 sched_domains_numa_masks[i] =
6759 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6760 if (!sched_domains_numa_masks[i])
6763 for (j = 0; j < nr_node_ids; j++) {
6764 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6768 sched_domains_numa_masks[i][j] = mask;
6771 if (node_distance(j, k) > sched_domains_numa_distance[i])
6774 cpumask_or(mask, mask, cpumask_of_node(k));
6779 /* Compute default topology size */
6780 for (i = 0; sched_domain_topology[i].mask; i++);
6782 tl = kzalloc((i + level + 1) *
6783 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6788 * Copy the default topology bits..
6790 for (i = 0; sched_domain_topology[i].mask; i++)
6791 tl[i] = sched_domain_topology[i];
6794 * .. and append 'j' levels of NUMA goodness.
6796 for (j = 0; j < level; i++, j++) {
6797 tl[i] = (struct sched_domain_topology_level){
6798 .mask = sd_numa_mask,
6799 .sd_flags = cpu_numa_flags,
6800 .flags = SDTL_OVERLAP,
6806 sched_domain_topology = tl;
6808 sched_domains_numa_levels = level;
6809 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6811 init_numa_topology_type();
6814 static void sched_domains_numa_masks_set(int cpu)
6817 int node = cpu_to_node(cpu);
6819 for (i = 0; i < sched_domains_numa_levels; i++) {
6820 for (j = 0; j < nr_node_ids; j++) {
6821 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6822 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6827 static void sched_domains_numa_masks_clear(int cpu)
6830 for (i = 0; i < sched_domains_numa_levels; i++) {
6831 for (j = 0; j < nr_node_ids; j++)
6832 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6837 * Update sched_domains_numa_masks[level][node] array when new cpus
6840 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6841 unsigned long action,
6844 int cpu = (long)hcpu;
6846 switch (action & ~CPU_TASKS_FROZEN) {
6848 sched_domains_numa_masks_set(cpu);
6852 sched_domains_numa_masks_clear(cpu);
6862 static inline void sched_init_numa(void)
6866 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6867 unsigned long action,
6872 #endif /* CONFIG_NUMA */
6874 static int __sdt_alloc(const struct cpumask *cpu_map)
6876 struct sched_domain_topology_level *tl;
6879 for_each_sd_topology(tl) {
6880 struct sd_data *sdd = &tl->data;
6882 sdd->sd = alloc_percpu(struct sched_domain *);
6886 sdd->sg = alloc_percpu(struct sched_group *);
6890 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6894 for_each_cpu(j, cpu_map) {
6895 struct sched_domain *sd;
6896 struct sched_group *sg;
6897 struct sched_group_capacity *sgc;
6899 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6900 GFP_KERNEL, cpu_to_node(j));
6904 *per_cpu_ptr(sdd->sd, j) = sd;
6906 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6907 GFP_KERNEL, cpu_to_node(j));
6913 *per_cpu_ptr(sdd->sg, j) = sg;
6915 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6916 GFP_KERNEL, cpu_to_node(j));
6920 *per_cpu_ptr(sdd->sgc, j) = sgc;
6927 static void __sdt_free(const struct cpumask *cpu_map)
6929 struct sched_domain_topology_level *tl;
6932 for_each_sd_topology(tl) {
6933 struct sd_data *sdd = &tl->data;
6935 for_each_cpu(j, cpu_map) {
6936 struct sched_domain *sd;
6939 sd = *per_cpu_ptr(sdd->sd, j);
6940 if (sd && (sd->flags & SD_OVERLAP))
6941 free_sched_groups(sd->groups, 0);
6942 kfree(*per_cpu_ptr(sdd->sd, j));
6946 kfree(*per_cpu_ptr(sdd->sg, j));
6948 kfree(*per_cpu_ptr(sdd->sgc, j));
6950 free_percpu(sdd->sd);
6952 free_percpu(sdd->sg);
6954 free_percpu(sdd->sgc);
6959 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6960 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6961 struct sched_domain *child, int cpu)
6963 struct sched_domain *sd = sd_init(tl, cpu);
6967 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6969 sd->level = child->level + 1;
6970 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6974 if (!cpumask_subset(sched_domain_span(child),
6975 sched_domain_span(sd))) {
6976 pr_err("BUG: arch topology borken\n");
6977 #ifdef CONFIG_SCHED_DEBUG
6978 pr_err(" the %s domain not a subset of the %s domain\n",
6979 child->name, sd->name);
6981 /* Fixup, ensure @sd has at least @child cpus. */
6982 cpumask_or(sched_domain_span(sd),
6983 sched_domain_span(sd),
6984 sched_domain_span(child));
6988 set_domain_attribute(sd, attr);
6994 * Build sched domains for a given set of cpus and attach the sched domains
6995 * to the individual cpus
6997 static int build_sched_domains(const struct cpumask *cpu_map,
6998 struct sched_domain_attr *attr)
7000 enum s_alloc alloc_state;
7001 struct sched_domain *sd;
7003 int i, ret = -ENOMEM;
7005 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7006 if (alloc_state != sa_rootdomain)
7009 /* Set up domains for cpus specified by the cpu_map. */
7010 for_each_cpu(i, cpu_map) {
7011 struct sched_domain_topology_level *tl;
7014 for_each_sd_topology(tl) {
7015 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7016 if (tl == sched_domain_topology)
7017 *per_cpu_ptr(d.sd, i) = sd;
7018 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7019 sd->flags |= SD_OVERLAP;
7020 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7025 /* Build the groups for the domains */
7026 for_each_cpu(i, cpu_map) {
7027 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7028 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7029 if (sd->flags & SD_OVERLAP) {
7030 if (build_overlap_sched_groups(sd, i))
7033 if (build_sched_groups(sd, i))
7039 /* Calculate CPU capacity for physical packages and nodes */
7040 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7041 if (!cpumask_test_cpu(i, cpu_map))
7044 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7045 claim_allocations(i, sd);
7046 init_sched_groups_capacity(i, sd);
7050 /* Attach the domains */
7052 for_each_cpu(i, cpu_map) {
7053 sd = *per_cpu_ptr(d.sd, i);
7054 cpu_attach_domain(sd, d.rd, i);
7060 __free_domain_allocs(&d, alloc_state, cpu_map);
7064 static cpumask_var_t *doms_cur; /* current sched domains */
7065 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7066 static struct sched_domain_attr *dattr_cur;
7067 /* attribues of custom domains in 'doms_cur' */
7070 * Special case: If a kmalloc of a doms_cur partition (array of
7071 * cpumask) fails, then fallback to a single sched domain,
7072 * as determined by the single cpumask fallback_doms.
7074 static cpumask_var_t fallback_doms;
7077 * arch_update_cpu_topology lets virtualized architectures update the
7078 * cpu core maps. It is supposed to return 1 if the topology changed
7079 * or 0 if it stayed the same.
7081 int __weak arch_update_cpu_topology(void)
7086 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7089 cpumask_var_t *doms;
7091 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7094 for (i = 0; i < ndoms; i++) {
7095 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7096 free_sched_domains(doms, i);
7103 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7106 for (i = 0; i < ndoms; i++)
7107 free_cpumask_var(doms[i]);
7112 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7113 * For now this just excludes isolated cpus, but could be used to
7114 * exclude other special cases in the future.
7116 static int init_sched_domains(const struct cpumask *cpu_map)
7120 arch_update_cpu_topology();
7122 doms_cur = alloc_sched_domains(ndoms_cur);
7124 doms_cur = &fallback_doms;
7125 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7126 err = build_sched_domains(doms_cur[0], NULL);
7127 register_sched_domain_sysctl();
7133 * Detach sched domains from a group of cpus specified in cpu_map
7134 * These cpus will now be attached to the NULL domain
7136 static void detach_destroy_domains(const struct cpumask *cpu_map)
7141 for_each_cpu(i, cpu_map)
7142 cpu_attach_domain(NULL, &def_root_domain, i);
7146 /* handle null as "default" */
7147 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7148 struct sched_domain_attr *new, int idx_new)
7150 struct sched_domain_attr tmp;
7157 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7158 new ? (new + idx_new) : &tmp,
7159 sizeof(struct sched_domain_attr));
7163 * Partition sched domains as specified by the 'ndoms_new'
7164 * cpumasks in the array doms_new[] of cpumasks. This compares
7165 * doms_new[] to the current sched domain partitioning, doms_cur[].
7166 * It destroys each deleted domain and builds each new domain.
7168 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7169 * The masks don't intersect (don't overlap.) We should setup one
7170 * sched domain for each mask. CPUs not in any of the cpumasks will
7171 * not be load balanced. If the same cpumask appears both in the
7172 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7175 * The passed in 'doms_new' should be allocated using
7176 * alloc_sched_domains. This routine takes ownership of it and will
7177 * free_sched_domains it when done with it. If the caller failed the
7178 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7179 * and partition_sched_domains() will fallback to the single partition
7180 * 'fallback_doms', it also forces the domains to be rebuilt.
7182 * If doms_new == NULL it will be replaced with cpu_online_mask.
7183 * ndoms_new == 0 is a special case for destroying existing domains,
7184 * and it will not create the default domain.
7186 * Call with hotplug lock held
7188 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7189 struct sched_domain_attr *dattr_new)
7194 mutex_lock(&sched_domains_mutex);
7196 /* always unregister in case we don't destroy any domains */
7197 unregister_sched_domain_sysctl();
7199 /* Let architecture update cpu core mappings. */
7200 new_topology = arch_update_cpu_topology();
7202 n = doms_new ? ndoms_new : 0;
7204 /* Destroy deleted domains */
7205 for (i = 0; i < ndoms_cur; i++) {
7206 for (j = 0; j < n && !new_topology; j++) {
7207 if (cpumask_equal(doms_cur[i], doms_new[j])
7208 && dattrs_equal(dattr_cur, i, dattr_new, j))
7211 /* no match - a current sched domain not in new doms_new[] */
7212 detach_destroy_domains(doms_cur[i]);
7218 if (doms_new == NULL) {
7220 doms_new = &fallback_doms;
7221 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7222 WARN_ON_ONCE(dattr_new);
7225 /* Build new domains */
7226 for (i = 0; i < ndoms_new; i++) {
7227 for (j = 0; j < n && !new_topology; j++) {
7228 if (cpumask_equal(doms_new[i], doms_cur[j])
7229 && dattrs_equal(dattr_new, i, dattr_cur, j))
7232 /* no match - add a new doms_new */
7233 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7238 /* Remember the new sched domains */
7239 if (doms_cur != &fallback_doms)
7240 free_sched_domains(doms_cur, ndoms_cur);
7241 kfree(dattr_cur); /* kfree(NULL) is safe */
7242 doms_cur = doms_new;
7243 dattr_cur = dattr_new;
7244 ndoms_cur = ndoms_new;
7246 register_sched_domain_sysctl();
7248 mutex_unlock(&sched_domains_mutex);
7251 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7254 * Update cpusets according to cpu_active mask. If cpusets are
7255 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7256 * around partition_sched_domains().
7258 * If we come here as part of a suspend/resume, don't touch cpusets because we
7259 * want to restore it back to its original state upon resume anyway.
7261 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7265 case CPU_ONLINE_FROZEN:
7266 case CPU_DOWN_FAILED_FROZEN:
7269 * num_cpus_frozen tracks how many CPUs are involved in suspend
7270 * resume sequence. As long as this is not the last online
7271 * operation in the resume sequence, just build a single sched
7272 * domain, ignoring cpusets.
7275 if (likely(num_cpus_frozen)) {
7276 partition_sched_domains(1, NULL, NULL);
7281 * This is the last CPU online operation. So fall through and
7282 * restore the original sched domains by considering the
7283 * cpuset configurations.
7287 cpuset_update_active_cpus(true);
7295 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7298 unsigned long flags;
7299 long cpu = (long)hcpu;
7305 case CPU_DOWN_PREPARE:
7306 rcu_read_lock_sched();
7307 dl_b = dl_bw_of(cpu);
7309 raw_spin_lock_irqsave(&dl_b->lock, flags);
7310 cpus = dl_bw_cpus(cpu);
7311 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7312 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7314 rcu_read_unlock_sched();
7317 return notifier_from_errno(-EBUSY);
7318 cpuset_update_active_cpus(false);
7320 case CPU_DOWN_PREPARE_FROZEN:
7322 partition_sched_domains(1, NULL, NULL);
7330 void __init sched_init_smp(void)
7332 cpumask_var_t non_isolated_cpus;
7334 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7335 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7340 * There's no userspace yet to cause hotplug operations; hence all the
7341 * cpu masks are stable and all blatant races in the below code cannot
7344 mutex_lock(&sched_domains_mutex);
7345 init_sched_domains(cpu_active_mask);
7346 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7347 if (cpumask_empty(non_isolated_cpus))
7348 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7349 mutex_unlock(&sched_domains_mutex);
7351 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7352 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7353 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7357 /* Move init over to a non-isolated CPU */
7358 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7360 sched_init_granularity();
7361 free_cpumask_var(non_isolated_cpus);
7363 init_sched_rt_class();
7364 init_sched_dl_class();
7367 void __init sched_init_smp(void)
7369 sched_init_granularity();
7371 #endif /* CONFIG_SMP */
7373 int in_sched_functions(unsigned long addr)
7375 return in_lock_functions(addr) ||
7376 (addr >= (unsigned long)__sched_text_start
7377 && addr < (unsigned long)__sched_text_end);
7380 #ifdef CONFIG_CGROUP_SCHED
7382 * Default task group.
7383 * Every task in system belongs to this group at bootup.
7385 struct task_group root_task_group;
7386 LIST_HEAD(task_groups);
7389 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7391 void __init sched_init(void)
7394 unsigned long alloc_size = 0, ptr;
7396 #ifdef CONFIG_FAIR_GROUP_SCHED
7397 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7399 #ifdef CONFIG_RT_GROUP_SCHED
7400 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7403 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7405 #ifdef CONFIG_FAIR_GROUP_SCHED
7406 root_task_group.se = (struct sched_entity **)ptr;
7407 ptr += nr_cpu_ids * sizeof(void **);
7409 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7410 ptr += nr_cpu_ids * sizeof(void **);
7412 #endif /* CONFIG_FAIR_GROUP_SCHED */
7413 #ifdef CONFIG_RT_GROUP_SCHED
7414 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7415 ptr += nr_cpu_ids * sizeof(void **);
7417 root_task_group.rt_rq = (struct rt_rq **)ptr;
7418 ptr += nr_cpu_ids * sizeof(void **);
7420 #endif /* CONFIG_RT_GROUP_SCHED */
7422 #ifdef CONFIG_CPUMASK_OFFSTACK
7423 for_each_possible_cpu(i) {
7424 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7425 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7427 #endif /* CONFIG_CPUMASK_OFFSTACK */
7429 init_rt_bandwidth(&def_rt_bandwidth,
7430 global_rt_period(), global_rt_runtime());
7431 init_dl_bandwidth(&def_dl_bandwidth,
7432 global_rt_period(), global_rt_runtime());
7435 init_defrootdomain();
7438 #ifdef CONFIG_RT_GROUP_SCHED
7439 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7440 global_rt_period(), global_rt_runtime());
7441 #endif /* CONFIG_RT_GROUP_SCHED */
7443 #ifdef CONFIG_CGROUP_SCHED
7444 list_add(&root_task_group.list, &task_groups);
7445 INIT_LIST_HEAD(&root_task_group.children);
7446 INIT_LIST_HEAD(&root_task_group.siblings);
7447 autogroup_init(&init_task);
7449 #endif /* CONFIG_CGROUP_SCHED */
7451 for_each_possible_cpu(i) {
7455 raw_spin_lock_init(&rq->lock);
7457 rq->calc_load_active = 0;
7458 rq->calc_load_update = jiffies + LOAD_FREQ;
7459 init_cfs_rq(&rq->cfs);
7460 init_rt_rq(&rq->rt);
7461 init_dl_rq(&rq->dl);
7462 #ifdef CONFIG_FAIR_GROUP_SCHED
7463 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7464 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7466 * How much cpu bandwidth does root_task_group get?
7468 * In case of task-groups formed thr' the cgroup filesystem, it
7469 * gets 100% of the cpu resources in the system. This overall
7470 * system cpu resource is divided among the tasks of
7471 * root_task_group and its child task-groups in a fair manner,
7472 * based on each entity's (task or task-group's) weight
7473 * (se->load.weight).
7475 * In other words, if root_task_group has 10 tasks of weight
7476 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7477 * then A0's share of the cpu resource is:
7479 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7481 * We achieve this by letting root_task_group's tasks sit
7482 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7484 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7485 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7486 #endif /* CONFIG_FAIR_GROUP_SCHED */
7488 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7489 #ifdef CONFIG_RT_GROUP_SCHED
7490 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7493 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7494 rq->cpu_load[j] = 0;
7496 rq->last_load_update_tick = jiffies;
7501 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7502 rq->balance_callback = NULL;
7503 rq->active_balance = 0;
7504 rq->next_balance = jiffies;
7509 rq->avg_idle = 2*sysctl_sched_migration_cost;
7510 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7512 INIT_LIST_HEAD(&rq->cfs_tasks);
7514 rq_attach_root(rq, &def_root_domain);
7515 #ifdef CONFIG_NO_HZ_COMMON
7518 #ifdef CONFIG_NO_HZ_FULL
7519 rq->last_sched_tick = 0;
7523 atomic_set(&rq->nr_iowait, 0);
7526 set_load_weight(&init_task);
7528 #ifdef CONFIG_PREEMPT_NOTIFIERS
7529 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7533 * The boot idle thread does lazy MMU switching as well:
7535 atomic_inc(&init_mm.mm_count);
7536 enter_lazy_tlb(&init_mm, current);
7539 * During early bootup we pretend to be a normal task:
7541 current->sched_class = &fair_sched_class;
7544 * Make us the idle thread. Technically, schedule() should not be
7545 * called from this thread, however somewhere below it might be,
7546 * but because we are the idle thread, we just pick up running again
7547 * when this runqueue becomes "idle".
7549 init_idle(current, smp_processor_id());
7551 calc_load_update = jiffies + LOAD_FREQ;
7554 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7555 /* May be allocated at isolcpus cmdline parse time */
7556 if (cpu_isolated_map == NULL)
7557 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7558 idle_thread_set_boot_cpu();
7559 set_cpu_rq_start_time();
7561 init_sched_fair_class();
7563 scheduler_running = 1;
7566 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7567 static inline int preempt_count_equals(int preempt_offset)
7569 int nested = preempt_count() + rcu_preempt_depth();
7571 return (nested == preempt_offset);
7574 void __might_sleep(const char *file, int line, int preempt_offset)
7577 * Blocking primitives will set (and therefore destroy) current->state,
7578 * since we will exit with TASK_RUNNING make sure we enter with it,
7579 * otherwise we will destroy state.
7581 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7582 "do not call blocking ops when !TASK_RUNNING; "
7583 "state=%lx set at [<%p>] %pS\n",
7585 (void *)current->task_state_change,
7586 (void *)current->task_state_change);
7588 ___might_sleep(file, line, preempt_offset);
7590 EXPORT_SYMBOL(__might_sleep);
7592 void ___might_sleep(const char *file, int line, int preempt_offset)
7594 static unsigned long prev_jiffy; /* ratelimiting */
7596 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7597 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7598 !is_idle_task(current)) ||
7599 system_state != SYSTEM_RUNNING || oops_in_progress)
7601 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7603 prev_jiffy = jiffies;
7606 "BUG: sleeping function called from invalid context at %s:%d\n",
7609 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7610 in_atomic(), irqs_disabled(),
7611 current->pid, current->comm);
7613 if (task_stack_end_corrupted(current))
7614 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7616 debug_show_held_locks(current);
7617 if (irqs_disabled())
7618 print_irqtrace_events(current);
7619 #ifdef CONFIG_DEBUG_PREEMPT
7620 if (!preempt_count_equals(preempt_offset)) {
7621 pr_err("Preemption disabled at:");
7622 print_ip_sym(current->preempt_disable_ip);
7628 EXPORT_SYMBOL(___might_sleep);
7631 #ifdef CONFIG_MAGIC_SYSRQ
7632 void normalize_rt_tasks(void)
7634 struct task_struct *g, *p;
7635 struct sched_attr attr = {
7636 .sched_policy = SCHED_NORMAL,
7639 read_lock(&tasklist_lock);
7640 for_each_process_thread(g, p) {
7642 * Only normalize user tasks:
7644 if (p->flags & PF_KTHREAD)
7647 p->se.exec_start = 0;
7648 #ifdef CONFIG_SCHEDSTATS
7649 p->se.statistics.wait_start = 0;
7650 p->se.statistics.sleep_start = 0;
7651 p->se.statistics.block_start = 0;
7654 if (!dl_task(p) && !rt_task(p)) {
7656 * Renice negative nice level userspace
7659 if (task_nice(p) < 0)
7660 set_user_nice(p, 0);
7664 __sched_setscheduler(p, &attr, false, false);
7666 read_unlock(&tasklist_lock);
7669 #endif /* CONFIG_MAGIC_SYSRQ */
7671 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7673 * These functions are only useful for the IA64 MCA handling, or kdb.
7675 * They can only be called when the whole system has been
7676 * stopped - every CPU needs to be quiescent, and no scheduling
7677 * activity can take place. Using them for anything else would
7678 * be a serious bug, and as a result, they aren't even visible
7679 * under any other configuration.
7683 * curr_task - return the current task for a given cpu.
7684 * @cpu: the processor in question.
7686 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7688 * Return: The current task for @cpu.
7690 struct task_struct *curr_task(int cpu)
7692 return cpu_curr(cpu);
7695 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7699 * set_curr_task - set the current task for a given cpu.
7700 * @cpu: the processor in question.
7701 * @p: the task pointer to set.
7703 * Description: This function must only be used when non-maskable interrupts
7704 * are serviced on a separate stack. It allows the architecture to switch the
7705 * notion of the current task on a cpu in a non-blocking manner. This function
7706 * must be called with all CPU's synchronized, and interrupts disabled, the
7707 * and caller must save the original value of the current task (see
7708 * curr_task() above) and restore that value before reenabling interrupts and
7709 * re-starting the system.
7711 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7713 void set_curr_task(int cpu, struct task_struct *p)
7720 #ifdef CONFIG_CGROUP_SCHED
7721 /* task_group_lock serializes the addition/removal of task groups */
7722 static DEFINE_SPINLOCK(task_group_lock);
7724 static void sched_free_group(struct task_group *tg)
7726 free_fair_sched_group(tg);
7727 free_rt_sched_group(tg);
7732 /* allocate runqueue etc for a new task group */
7733 struct task_group *sched_create_group(struct task_group *parent)
7735 struct task_group *tg;
7737 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7739 return ERR_PTR(-ENOMEM);
7741 if (!alloc_fair_sched_group(tg, parent))
7744 if (!alloc_rt_sched_group(tg, parent))
7750 sched_free_group(tg);
7751 return ERR_PTR(-ENOMEM);
7754 void sched_online_group(struct task_group *tg, struct task_group *parent)
7756 unsigned long flags;
7758 spin_lock_irqsave(&task_group_lock, flags);
7759 list_add_rcu(&tg->list, &task_groups);
7761 WARN_ON(!parent); /* root should already exist */
7763 tg->parent = parent;
7764 INIT_LIST_HEAD(&tg->children);
7765 list_add_rcu(&tg->siblings, &parent->children);
7766 spin_unlock_irqrestore(&task_group_lock, flags);
7769 /* rcu callback to free various structures associated with a task group */
7770 static void sched_free_group_rcu(struct rcu_head *rhp)
7772 /* now it should be safe to free those cfs_rqs */
7773 sched_free_group(container_of(rhp, struct task_group, rcu));
7776 void sched_destroy_group(struct task_group *tg)
7778 /* wait for possible concurrent references to cfs_rqs complete */
7779 call_rcu(&tg->rcu, sched_free_group_rcu);
7782 void sched_offline_group(struct task_group *tg)
7784 unsigned long flags;
7787 /* end participation in shares distribution */
7788 for_each_possible_cpu(i)
7789 unregister_fair_sched_group(tg, i);
7791 spin_lock_irqsave(&task_group_lock, flags);
7792 list_del_rcu(&tg->list);
7793 list_del_rcu(&tg->siblings);
7794 spin_unlock_irqrestore(&task_group_lock, flags);
7797 /* change task's runqueue when it moves between groups.
7798 * The caller of this function should have put the task in its new group
7799 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7800 * reflect its new group.
7802 void sched_move_task(struct task_struct *tsk)
7804 struct task_group *tg;
7805 int queued, running;
7806 unsigned long flags;
7809 rq = task_rq_lock(tsk, &flags);
7811 running = task_current(rq, tsk);
7812 queued = task_on_rq_queued(tsk);
7815 dequeue_task(rq, tsk, DEQUEUE_SAVE);
7816 if (unlikely(running))
7817 put_prev_task(rq, tsk);
7820 * All callers are synchronized by task_rq_lock(); we do not use RCU
7821 * which is pointless here. Thus, we pass "true" to task_css_check()
7822 * to prevent lockdep warnings.
7824 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7825 struct task_group, css);
7826 tg = autogroup_task_group(tsk, tg);
7827 tsk->sched_task_group = tg;
7829 #ifdef CONFIG_FAIR_GROUP_SCHED
7830 if (tsk->sched_class->task_move_group)
7831 tsk->sched_class->task_move_group(tsk);
7834 set_task_rq(tsk, task_cpu(tsk));
7836 if (unlikely(running))
7837 tsk->sched_class->set_curr_task(rq);
7839 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
7841 task_rq_unlock(rq, tsk, &flags);
7843 #endif /* CONFIG_CGROUP_SCHED */
7845 #ifdef CONFIG_RT_GROUP_SCHED
7847 * Ensure that the real time constraints are schedulable.
7849 static DEFINE_MUTEX(rt_constraints_mutex);
7851 /* Must be called with tasklist_lock held */
7852 static inline int tg_has_rt_tasks(struct task_group *tg)
7854 struct task_struct *g, *p;
7857 * Autogroups do not have RT tasks; see autogroup_create().
7859 if (task_group_is_autogroup(tg))
7862 for_each_process_thread(g, p) {
7863 if (rt_task(p) && task_group(p) == tg)
7870 struct rt_schedulable_data {
7871 struct task_group *tg;
7876 static int tg_rt_schedulable(struct task_group *tg, void *data)
7878 struct rt_schedulable_data *d = data;
7879 struct task_group *child;
7880 unsigned long total, sum = 0;
7881 u64 period, runtime;
7883 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7884 runtime = tg->rt_bandwidth.rt_runtime;
7887 period = d->rt_period;
7888 runtime = d->rt_runtime;
7892 * Cannot have more runtime than the period.
7894 if (runtime > period && runtime != RUNTIME_INF)
7898 * Ensure we don't starve existing RT tasks.
7900 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7903 total = to_ratio(period, runtime);
7906 * Nobody can have more than the global setting allows.
7908 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7912 * The sum of our children's runtime should not exceed our own.
7914 list_for_each_entry_rcu(child, &tg->children, siblings) {
7915 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7916 runtime = child->rt_bandwidth.rt_runtime;
7918 if (child == d->tg) {
7919 period = d->rt_period;
7920 runtime = d->rt_runtime;
7923 sum += to_ratio(period, runtime);
7932 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7936 struct rt_schedulable_data data = {
7938 .rt_period = period,
7939 .rt_runtime = runtime,
7943 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7949 static int tg_set_rt_bandwidth(struct task_group *tg,
7950 u64 rt_period, u64 rt_runtime)
7955 * Disallowing the root group RT runtime is BAD, it would disallow the
7956 * kernel creating (and or operating) RT threads.
7958 if (tg == &root_task_group && rt_runtime == 0)
7961 /* No period doesn't make any sense. */
7965 mutex_lock(&rt_constraints_mutex);
7966 read_lock(&tasklist_lock);
7967 err = __rt_schedulable(tg, rt_period, rt_runtime);
7971 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7972 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7973 tg->rt_bandwidth.rt_runtime = rt_runtime;
7975 for_each_possible_cpu(i) {
7976 struct rt_rq *rt_rq = tg->rt_rq[i];
7978 raw_spin_lock(&rt_rq->rt_runtime_lock);
7979 rt_rq->rt_runtime = rt_runtime;
7980 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7982 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7984 read_unlock(&tasklist_lock);
7985 mutex_unlock(&rt_constraints_mutex);
7990 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7992 u64 rt_runtime, rt_period;
7994 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7995 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7996 if (rt_runtime_us < 0)
7997 rt_runtime = RUNTIME_INF;
7999 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8002 static long sched_group_rt_runtime(struct task_group *tg)
8006 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8009 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8010 do_div(rt_runtime_us, NSEC_PER_USEC);
8011 return rt_runtime_us;
8014 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8016 u64 rt_runtime, rt_period;
8018 rt_period = rt_period_us * NSEC_PER_USEC;
8019 rt_runtime = tg->rt_bandwidth.rt_runtime;
8021 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8024 static long sched_group_rt_period(struct task_group *tg)
8028 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8029 do_div(rt_period_us, NSEC_PER_USEC);
8030 return rt_period_us;
8032 #endif /* CONFIG_RT_GROUP_SCHED */
8034 #ifdef CONFIG_RT_GROUP_SCHED
8035 static int sched_rt_global_constraints(void)
8039 mutex_lock(&rt_constraints_mutex);
8040 read_lock(&tasklist_lock);
8041 ret = __rt_schedulable(NULL, 0, 0);
8042 read_unlock(&tasklist_lock);
8043 mutex_unlock(&rt_constraints_mutex);
8048 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8050 /* Don't accept realtime tasks when there is no way for them to run */
8051 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8057 #else /* !CONFIG_RT_GROUP_SCHED */
8058 static int sched_rt_global_constraints(void)
8060 unsigned long flags;
8063 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8064 for_each_possible_cpu(i) {
8065 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8067 raw_spin_lock(&rt_rq->rt_runtime_lock);
8068 rt_rq->rt_runtime = global_rt_runtime();
8069 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8071 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8075 #endif /* CONFIG_RT_GROUP_SCHED */
8077 static int sched_dl_global_validate(void)
8079 u64 runtime = global_rt_runtime();
8080 u64 period = global_rt_period();
8081 u64 new_bw = to_ratio(period, runtime);
8084 unsigned long flags;
8087 * Here we want to check the bandwidth not being set to some
8088 * value smaller than the currently allocated bandwidth in
8089 * any of the root_domains.
8091 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8092 * cycling on root_domains... Discussion on different/better
8093 * solutions is welcome!
8095 for_each_possible_cpu(cpu) {
8096 rcu_read_lock_sched();
8097 dl_b = dl_bw_of(cpu);
8099 raw_spin_lock_irqsave(&dl_b->lock, flags);
8100 if (new_bw < dl_b->total_bw)
8102 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8104 rcu_read_unlock_sched();
8113 static void sched_dl_do_global(void)
8118 unsigned long flags;
8120 def_dl_bandwidth.dl_period = global_rt_period();
8121 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8123 if (global_rt_runtime() != RUNTIME_INF)
8124 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8127 * FIXME: As above...
8129 for_each_possible_cpu(cpu) {
8130 rcu_read_lock_sched();
8131 dl_b = dl_bw_of(cpu);
8133 raw_spin_lock_irqsave(&dl_b->lock, flags);
8135 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8137 rcu_read_unlock_sched();
8141 static int sched_rt_global_validate(void)
8143 if (sysctl_sched_rt_period <= 0)
8146 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8147 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8153 static void sched_rt_do_global(void)
8155 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8156 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8159 int sched_rt_handler(struct ctl_table *table, int write,
8160 void __user *buffer, size_t *lenp,
8163 int old_period, old_runtime;
8164 static DEFINE_MUTEX(mutex);
8168 old_period = sysctl_sched_rt_period;
8169 old_runtime = sysctl_sched_rt_runtime;
8171 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8173 if (!ret && write) {
8174 ret = sched_rt_global_validate();
8178 ret = sched_dl_global_validate();
8182 ret = sched_rt_global_constraints();
8186 sched_rt_do_global();
8187 sched_dl_do_global();
8191 sysctl_sched_rt_period = old_period;
8192 sysctl_sched_rt_runtime = old_runtime;
8194 mutex_unlock(&mutex);
8199 int sched_rr_handler(struct ctl_table *table, int write,
8200 void __user *buffer, size_t *lenp,
8204 static DEFINE_MUTEX(mutex);
8207 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8208 /* make sure that internally we keep jiffies */
8209 /* also, writing zero resets timeslice to default */
8210 if (!ret && write) {
8211 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8212 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8214 mutex_unlock(&mutex);
8218 #ifdef CONFIG_CGROUP_SCHED
8220 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8222 return css ? container_of(css, struct task_group, css) : NULL;
8225 static struct cgroup_subsys_state *
8226 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8228 struct task_group *parent = css_tg(parent_css);
8229 struct task_group *tg;
8232 /* This is early initialization for the top cgroup */
8233 return &root_task_group.css;
8236 tg = sched_create_group(parent);
8238 return ERR_PTR(-ENOMEM);
8240 sched_online_group(tg, parent);
8245 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8247 struct task_group *tg = css_tg(css);
8249 sched_offline_group(tg);
8252 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8254 struct task_group *tg = css_tg(css);
8257 * Relies on the RCU grace period between css_released() and this.
8259 sched_free_group(tg);
8262 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8264 sched_move_task(task);
8267 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8269 struct task_struct *task;
8270 struct cgroup_subsys_state *css;
8272 cgroup_taskset_for_each(task, css, tset) {
8273 #ifdef CONFIG_RT_GROUP_SCHED
8274 if (!sched_rt_can_attach(css_tg(css), task))
8277 /* We don't support RT-tasks being in separate groups */
8278 if (task->sched_class != &fair_sched_class)
8285 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8287 struct task_struct *task;
8288 struct cgroup_subsys_state *css;
8290 cgroup_taskset_for_each(task, css, tset)
8291 sched_move_task(task);
8294 #ifdef CONFIG_FAIR_GROUP_SCHED
8295 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8296 struct cftype *cftype, u64 shareval)
8298 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8301 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8304 struct task_group *tg = css_tg(css);
8306 return (u64) scale_load_down(tg->shares);
8309 #ifdef CONFIG_CFS_BANDWIDTH
8310 static DEFINE_MUTEX(cfs_constraints_mutex);
8312 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8313 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8315 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8317 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8319 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8320 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8322 if (tg == &root_task_group)
8326 * Ensure we have at some amount of bandwidth every period. This is
8327 * to prevent reaching a state of large arrears when throttled via
8328 * entity_tick() resulting in prolonged exit starvation.
8330 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8334 * Likewise, bound things on the otherside by preventing insane quota
8335 * periods. This also allows us to normalize in computing quota
8338 if (period > max_cfs_quota_period)
8342 * Prevent race between setting of cfs_rq->runtime_enabled and
8343 * unthrottle_offline_cfs_rqs().
8346 mutex_lock(&cfs_constraints_mutex);
8347 ret = __cfs_schedulable(tg, period, quota);
8351 runtime_enabled = quota != RUNTIME_INF;
8352 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8354 * If we need to toggle cfs_bandwidth_used, off->on must occur
8355 * before making related changes, and on->off must occur afterwards
8357 if (runtime_enabled && !runtime_was_enabled)
8358 cfs_bandwidth_usage_inc();
8359 raw_spin_lock_irq(&cfs_b->lock);
8360 cfs_b->period = ns_to_ktime(period);
8361 cfs_b->quota = quota;
8363 __refill_cfs_bandwidth_runtime(cfs_b);
8364 /* restart the period timer (if active) to handle new period expiry */
8365 if (runtime_enabled)
8366 start_cfs_bandwidth(cfs_b);
8367 raw_spin_unlock_irq(&cfs_b->lock);
8369 for_each_online_cpu(i) {
8370 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8371 struct rq *rq = cfs_rq->rq;
8373 raw_spin_lock_irq(&rq->lock);
8374 cfs_rq->runtime_enabled = runtime_enabled;
8375 cfs_rq->runtime_remaining = 0;
8377 if (cfs_rq->throttled)
8378 unthrottle_cfs_rq(cfs_rq);
8379 raw_spin_unlock_irq(&rq->lock);
8381 if (runtime_was_enabled && !runtime_enabled)
8382 cfs_bandwidth_usage_dec();
8384 mutex_unlock(&cfs_constraints_mutex);
8390 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8394 period = ktime_to_ns(tg->cfs_bandwidth.period);
8395 if (cfs_quota_us < 0)
8396 quota = RUNTIME_INF;
8398 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8400 return tg_set_cfs_bandwidth(tg, period, quota);
8403 long tg_get_cfs_quota(struct task_group *tg)
8407 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8410 quota_us = tg->cfs_bandwidth.quota;
8411 do_div(quota_us, NSEC_PER_USEC);
8416 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8420 period = (u64)cfs_period_us * NSEC_PER_USEC;
8421 quota = tg->cfs_bandwidth.quota;
8423 return tg_set_cfs_bandwidth(tg, period, quota);
8426 long tg_get_cfs_period(struct task_group *tg)
8430 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8431 do_div(cfs_period_us, NSEC_PER_USEC);
8433 return cfs_period_us;
8436 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8439 return tg_get_cfs_quota(css_tg(css));
8442 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8443 struct cftype *cftype, s64 cfs_quota_us)
8445 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8448 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8451 return tg_get_cfs_period(css_tg(css));
8454 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8455 struct cftype *cftype, u64 cfs_period_us)
8457 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8460 struct cfs_schedulable_data {
8461 struct task_group *tg;
8466 * normalize group quota/period to be quota/max_period
8467 * note: units are usecs
8469 static u64 normalize_cfs_quota(struct task_group *tg,
8470 struct cfs_schedulable_data *d)
8478 period = tg_get_cfs_period(tg);
8479 quota = tg_get_cfs_quota(tg);
8482 /* note: these should typically be equivalent */
8483 if (quota == RUNTIME_INF || quota == -1)
8486 return to_ratio(period, quota);
8489 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8491 struct cfs_schedulable_data *d = data;
8492 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8493 s64 quota = 0, parent_quota = -1;
8496 quota = RUNTIME_INF;
8498 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8500 quota = normalize_cfs_quota(tg, d);
8501 parent_quota = parent_b->hierarchical_quota;
8504 * ensure max(child_quota) <= parent_quota, inherit when no
8507 if (quota == RUNTIME_INF)
8508 quota = parent_quota;
8509 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8512 cfs_b->hierarchical_quota = quota;
8517 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8520 struct cfs_schedulable_data data = {
8526 if (quota != RUNTIME_INF) {
8527 do_div(data.period, NSEC_PER_USEC);
8528 do_div(data.quota, NSEC_PER_USEC);
8532 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8538 static int cpu_stats_show(struct seq_file *sf, void *v)
8540 struct task_group *tg = css_tg(seq_css(sf));
8541 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8543 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8544 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8545 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8549 #endif /* CONFIG_CFS_BANDWIDTH */
8550 #endif /* CONFIG_FAIR_GROUP_SCHED */
8552 #ifdef CONFIG_RT_GROUP_SCHED
8553 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8554 struct cftype *cft, s64 val)
8556 return sched_group_set_rt_runtime(css_tg(css), val);
8559 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8562 return sched_group_rt_runtime(css_tg(css));
8565 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8566 struct cftype *cftype, u64 rt_period_us)
8568 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8571 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8574 return sched_group_rt_period(css_tg(css));
8576 #endif /* CONFIG_RT_GROUP_SCHED */
8578 static struct cftype cpu_files[] = {
8579 #ifdef CONFIG_FAIR_GROUP_SCHED
8582 .read_u64 = cpu_shares_read_u64,
8583 .write_u64 = cpu_shares_write_u64,
8586 #ifdef CONFIG_CFS_BANDWIDTH
8588 .name = "cfs_quota_us",
8589 .read_s64 = cpu_cfs_quota_read_s64,
8590 .write_s64 = cpu_cfs_quota_write_s64,
8593 .name = "cfs_period_us",
8594 .read_u64 = cpu_cfs_period_read_u64,
8595 .write_u64 = cpu_cfs_period_write_u64,
8599 .seq_show = cpu_stats_show,
8602 #ifdef CONFIG_RT_GROUP_SCHED
8604 .name = "rt_runtime_us",
8605 .read_s64 = cpu_rt_runtime_read,
8606 .write_s64 = cpu_rt_runtime_write,
8609 .name = "rt_period_us",
8610 .read_u64 = cpu_rt_period_read_uint,
8611 .write_u64 = cpu_rt_period_write_uint,
8617 struct cgroup_subsys cpu_cgrp_subsys = {
8618 .css_alloc = cpu_cgroup_css_alloc,
8619 .css_released = cpu_cgroup_css_released,
8620 .css_free = cpu_cgroup_css_free,
8621 .fork = cpu_cgroup_fork,
8622 .can_attach = cpu_cgroup_can_attach,
8623 .attach = cpu_cgroup_attach,
8624 .legacy_cftypes = cpu_files,
8628 #endif /* CONFIG_CGROUP_SCHED */
8630 void dump_cpu_task(int cpu)
8632 pr_info("Task dump for CPU %d:\n", cpu);
8633 sched_show_task(cpu_curr(cpu));