4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
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 raw_spin_lock_irqsave(&rq->lock, flags);
605 raw_spin_unlock_irqrestore(&rq->lock, flags);
609 #ifdef CONFIG_NO_HZ_COMMON
611 * In the semi idle case, use the nearest busy cpu for migrating timers
612 * from an idle cpu. This is good for power-savings.
614 * We don't do similar optimization for completely idle system, as
615 * selecting an idle cpu will add more delays to the timers than intended
616 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
618 int get_nohz_timer_target(void)
620 int i, cpu = smp_processor_id();
621 struct sched_domain *sd;
623 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
627 for_each_domain(cpu, sd) {
628 for_each_cpu(i, sched_domain_span(sd)) {
632 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
639 if (!is_housekeeping_cpu(cpu))
640 cpu = housekeeping_any_cpu();
646 * When add_timer_on() enqueues a timer into the timer wheel of an
647 * idle CPU then this timer might expire before the next timer event
648 * which is scheduled to wake up that CPU. In case of a completely
649 * idle system the next event might even be infinite time into the
650 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
651 * leaves the inner idle loop so the newly added timer is taken into
652 * account when the CPU goes back to idle and evaluates the timer
653 * wheel for the next timer event.
655 static void wake_up_idle_cpu(int cpu)
657 struct rq *rq = cpu_rq(cpu);
659 if (cpu == smp_processor_id())
662 if (set_nr_and_not_polling(rq->idle))
663 smp_send_reschedule(cpu);
665 trace_sched_wake_idle_without_ipi(cpu);
668 static bool wake_up_full_nohz_cpu(int cpu)
671 * We just need the target to call irq_exit() and re-evaluate
672 * the next tick. The nohz full kick at least implies that.
673 * If needed we can still optimize that later with an
676 if (tick_nohz_full_cpu(cpu)) {
677 if (cpu != smp_processor_id() ||
678 tick_nohz_tick_stopped())
679 tick_nohz_full_kick_cpu(cpu);
686 void wake_up_nohz_cpu(int cpu)
688 if (!wake_up_full_nohz_cpu(cpu))
689 wake_up_idle_cpu(cpu);
692 static inline bool got_nohz_idle_kick(void)
694 int cpu = smp_processor_id();
696 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
699 if (idle_cpu(cpu) && !need_resched())
703 * We can't run Idle Load Balance on this CPU for this time so we
704 * cancel it and clear NOHZ_BALANCE_KICK
706 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
710 #else /* CONFIG_NO_HZ_COMMON */
712 static inline bool got_nohz_idle_kick(void)
717 #endif /* CONFIG_NO_HZ_COMMON */
719 #ifdef CONFIG_NO_HZ_FULL
720 bool sched_can_stop_tick(void)
723 * FIFO realtime policy runs the highest priority task. Other runnable
724 * tasks are of a lower priority. The scheduler tick does nothing.
726 if (current->policy == SCHED_FIFO)
730 * Round-robin realtime tasks time slice with other tasks at the same
731 * realtime priority. Is this task the only one at this priority?
733 if (current->policy == SCHED_RR) {
734 struct sched_rt_entity *rt_se = ¤t->rt;
736 return rt_se->run_list.prev == rt_se->run_list.next;
740 * More than one running task need preemption.
741 * nr_running update is assumed to be visible
742 * after IPI is sent from wakers.
744 if (this_rq()->nr_running > 1)
749 #endif /* CONFIG_NO_HZ_FULL */
751 void sched_avg_update(struct rq *rq)
753 s64 period = sched_avg_period();
755 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
757 * Inline assembly required to prevent the compiler
758 * optimising this loop into a divmod call.
759 * See __iter_div_u64_rem() for another example of this.
761 asm("" : "+rm" (rq->age_stamp));
762 rq->age_stamp += period;
767 #endif /* CONFIG_SMP */
769 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
770 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
772 * Iterate task_group tree rooted at *from, calling @down when first entering a
773 * node and @up when leaving it for the final time.
775 * Caller must hold rcu_lock or sufficient equivalent.
777 int walk_tg_tree_from(struct task_group *from,
778 tg_visitor down, tg_visitor up, void *data)
780 struct task_group *parent, *child;
786 ret = (*down)(parent, data);
789 list_for_each_entry_rcu(child, &parent->children, siblings) {
796 ret = (*up)(parent, data);
797 if (ret || parent == from)
801 parent = parent->parent;
808 int tg_nop(struct task_group *tg, void *data)
814 static void set_load_weight(struct task_struct *p)
816 int prio = p->static_prio - MAX_RT_PRIO;
817 struct load_weight *load = &p->se.load;
820 * SCHED_IDLE tasks get minimal weight:
822 if (idle_policy(p->policy)) {
823 load->weight = scale_load(WEIGHT_IDLEPRIO);
824 load->inv_weight = WMULT_IDLEPRIO;
828 load->weight = scale_load(prio_to_weight[prio]);
829 load->inv_weight = prio_to_wmult[prio];
832 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
835 if (!(flags & ENQUEUE_RESTORE))
836 sched_info_queued(rq, p);
837 p->sched_class->enqueue_task(rq, p, flags);
840 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
843 if (!(flags & DEQUEUE_SAVE))
844 sched_info_dequeued(rq, p);
845 p->sched_class->dequeue_task(rq, p, flags);
848 void activate_task(struct rq *rq, struct task_struct *p, int flags)
850 if (task_contributes_to_load(p))
851 rq->nr_uninterruptible--;
853 enqueue_task(rq, p, flags);
856 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
858 if (task_contributes_to_load(p))
859 rq->nr_uninterruptible++;
861 dequeue_task(rq, p, flags);
864 static void update_rq_clock_task(struct rq *rq, s64 delta)
867 * In theory, the compile should just see 0 here, and optimize out the call
868 * to sched_rt_avg_update. But I don't trust it...
870 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
871 s64 steal = 0, irq_delta = 0;
873 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
874 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
877 * Since irq_time is only updated on {soft,}irq_exit, we might run into
878 * this case when a previous update_rq_clock() happened inside a
881 * When this happens, we stop ->clock_task and only update the
882 * prev_irq_time stamp to account for the part that fit, so that a next
883 * update will consume the rest. This ensures ->clock_task is
886 * It does however cause some slight miss-attribution of {soft,}irq
887 * time, a more accurate solution would be to update the irq_time using
888 * the current rq->clock timestamp, except that would require using
891 if (irq_delta > delta)
894 rq->prev_irq_time += irq_delta;
897 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
898 if (static_key_false((¶virt_steal_rq_enabled))) {
899 steal = paravirt_steal_clock(cpu_of(rq));
900 steal -= rq->prev_steal_time_rq;
902 if (unlikely(steal > delta))
905 rq->prev_steal_time_rq += steal;
910 rq->clock_task += delta;
912 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
913 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
914 sched_rt_avg_update(rq, irq_delta + steal);
918 void sched_set_stop_task(int cpu, struct task_struct *stop)
920 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
921 struct task_struct *old_stop = cpu_rq(cpu)->stop;
925 * Make it appear like a SCHED_FIFO task, its something
926 * userspace knows about and won't get confused about.
928 * Also, it will make PI more or less work without too
929 * much confusion -- but then, stop work should not
930 * rely on PI working anyway.
932 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
934 stop->sched_class = &stop_sched_class;
937 cpu_rq(cpu)->stop = stop;
941 * Reset it back to a normal scheduling class so that
942 * it can die in pieces.
944 old_stop->sched_class = &rt_sched_class;
949 * __normal_prio - return the priority that is based on the static prio
951 static inline int __normal_prio(struct task_struct *p)
953 return p->static_prio;
957 * Calculate the expected normal priority: i.e. priority
958 * without taking RT-inheritance into account. Might be
959 * boosted by interactivity modifiers. Changes upon fork,
960 * setprio syscalls, and whenever the interactivity
961 * estimator recalculates.
963 static inline int normal_prio(struct task_struct *p)
967 if (task_has_dl_policy(p))
968 prio = MAX_DL_PRIO-1;
969 else if (task_has_rt_policy(p))
970 prio = MAX_RT_PRIO-1 - p->rt_priority;
972 prio = __normal_prio(p);
977 * Calculate the current priority, i.e. the priority
978 * taken into account by the scheduler. This value might
979 * be boosted by RT tasks, or might be boosted by
980 * interactivity modifiers. Will be RT if the task got
981 * RT-boosted. If not then it returns p->normal_prio.
983 static int effective_prio(struct task_struct *p)
985 p->normal_prio = normal_prio(p);
987 * If we are RT tasks or we were boosted to RT priority,
988 * keep the priority unchanged. Otherwise, update priority
989 * to the normal priority:
991 if (!rt_prio(p->prio))
992 return p->normal_prio;
997 * task_curr - is this task currently executing on a CPU?
998 * @p: the task in question.
1000 * Return: 1 if the task is currently executing. 0 otherwise.
1002 inline int task_curr(const struct task_struct *p)
1004 return cpu_curr(task_cpu(p)) == p;
1008 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1009 * use the balance_callback list if you want balancing.
1011 * this means any call to check_class_changed() must be followed by a call to
1012 * balance_callback().
1014 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1015 const struct sched_class *prev_class,
1018 if (prev_class != p->sched_class) {
1019 if (prev_class->switched_from)
1020 prev_class->switched_from(rq, p);
1022 p->sched_class->switched_to(rq, p);
1023 } else if (oldprio != p->prio || dl_task(p))
1024 p->sched_class->prio_changed(rq, p, oldprio);
1027 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1029 const struct sched_class *class;
1031 if (p->sched_class == rq->curr->sched_class) {
1032 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1034 for_each_class(class) {
1035 if (class == rq->curr->sched_class)
1037 if (class == p->sched_class) {
1045 * A queue event has occurred, and we're going to schedule. In
1046 * this case, we can save a useless back to back clock update.
1048 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1049 rq_clock_skip_update(rq, true);
1054 * This is how migration works:
1056 * 1) we invoke migration_cpu_stop() on the target CPU using
1058 * 2) stopper starts to run (implicitly forcing the migrated thread
1060 * 3) it checks whether the migrated task is still in the wrong runqueue.
1061 * 4) if it's in the wrong runqueue then the migration thread removes
1062 * it and puts it into the right queue.
1063 * 5) stopper completes and stop_one_cpu() returns and the migration
1068 * move_queued_task - move a queued task to new rq.
1070 * Returns (locked) new rq. Old rq's lock is released.
1072 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1074 lockdep_assert_held(&rq->lock);
1076 dequeue_task(rq, p, 0);
1077 p->on_rq = TASK_ON_RQ_MIGRATING;
1078 set_task_cpu(p, new_cpu);
1079 raw_spin_unlock(&rq->lock);
1081 rq = cpu_rq(new_cpu);
1083 raw_spin_lock(&rq->lock);
1084 BUG_ON(task_cpu(p) != new_cpu);
1085 p->on_rq = TASK_ON_RQ_QUEUED;
1086 enqueue_task(rq, p, 0);
1087 check_preempt_curr(rq, p, 0);
1092 struct migration_arg {
1093 struct task_struct *task;
1098 * Move (not current) task off this cpu, onto dest cpu. We're doing
1099 * this because either it can't run here any more (set_cpus_allowed()
1100 * away from this CPU, or CPU going down), or because we're
1101 * attempting to rebalance this task on exec (sched_exec).
1103 * So we race with normal scheduler movements, but that's OK, as long
1104 * as the task is no longer on this CPU.
1106 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1108 if (unlikely(!cpu_active(dest_cpu)))
1111 /* Affinity changed (again). */
1112 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1115 rq = move_queued_task(rq, p, dest_cpu);
1121 * migration_cpu_stop - this will be executed by a highprio stopper thread
1122 * and performs thread migration by bumping thread off CPU then
1123 * 'pushing' onto another runqueue.
1125 static int migration_cpu_stop(void *data)
1127 struct migration_arg *arg = data;
1128 struct task_struct *p = arg->task;
1129 struct rq *rq = this_rq();
1132 * The original target cpu might have gone down and we might
1133 * be on another cpu but it doesn't matter.
1135 local_irq_disable();
1137 * We need to explicitly wake pending tasks before running
1138 * __migrate_task() such that we will not miss enforcing cpus_allowed
1139 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1141 sched_ttwu_pending();
1143 raw_spin_lock(&p->pi_lock);
1144 raw_spin_lock(&rq->lock);
1146 * If task_rq(p) != rq, it cannot be migrated here, because we're
1147 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1148 * we're holding p->pi_lock.
1150 if (task_rq(p) == rq && task_on_rq_queued(p))
1151 rq = __migrate_task(rq, p, arg->dest_cpu);
1152 raw_spin_unlock(&rq->lock);
1153 raw_spin_unlock(&p->pi_lock);
1160 * sched_class::set_cpus_allowed must do the below, but is not required to
1161 * actually call this function.
1163 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1165 cpumask_copy(&p->cpus_allowed, new_mask);
1166 p->nr_cpus_allowed = cpumask_weight(new_mask);
1169 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1171 struct rq *rq = task_rq(p);
1172 bool queued, running;
1174 lockdep_assert_held(&p->pi_lock);
1176 queued = task_on_rq_queued(p);
1177 running = task_current(rq, p);
1181 * Because __kthread_bind() calls this on blocked tasks without
1184 lockdep_assert_held(&rq->lock);
1185 dequeue_task(rq, p, DEQUEUE_SAVE);
1188 put_prev_task(rq, p);
1190 p->sched_class->set_cpus_allowed(p, new_mask);
1193 p->sched_class->set_curr_task(rq);
1195 enqueue_task(rq, p, ENQUEUE_RESTORE);
1199 * Change a given task's CPU affinity. Migrate the thread to a
1200 * proper CPU and schedule it away if the CPU it's executing on
1201 * is removed from the allowed bitmask.
1203 * NOTE: the caller must have a valid reference to the task, the
1204 * task must not exit() & deallocate itself prematurely. The
1205 * call is not atomic; no spinlocks may be held.
1207 static int __set_cpus_allowed_ptr(struct task_struct *p,
1208 const struct cpumask *new_mask, bool check)
1210 unsigned long flags;
1212 unsigned int dest_cpu;
1215 rq = task_rq_lock(p, &flags);
1218 * Must re-check here, to close a race against __kthread_bind(),
1219 * sched_setaffinity() is not guaranteed to observe the flag.
1221 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1226 if (cpumask_equal(&p->cpus_allowed, new_mask))
1229 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1234 do_set_cpus_allowed(p, new_mask);
1236 /* Can the task run on the task's current CPU? If so, we're done */
1237 if (cpumask_test_cpu(task_cpu(p), new_mask))
1240 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1241 if (task_running(rq, p) || p->state == TASK_WAKING) {
1242 struct migration_arg arg = { p, dest_cpu };
1243 /* Need help from migration thread: drop lock and wait. */
1244 task_rq_unlock(rq, p, &flags);
1245 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1246 tlb_migrate_finish(p->mm);
1248 } else if (task_on_rq_queued(p)) {
1250 * OK, since we're going to drop the lock immediately
1251 * afterwards anyway.
1253 lockdep_unpin_lock(&rq->lock);
1254 rq = move_queued_task(rq, p, dest_cpu);
1255 lockdep_pin_lock(&rq->lock);
1258 task_rq_unlock(rq, p, &flags);
1263 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1265 return __set_cpus_allowed_ptr(p, new_mask, false);
1267 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1269 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1271 #ifdef CONFIG_SCHED_DEBUG
1273 * We should never call set_task_cpu() on a blocked task,
1274 * ttwu() will sort out the placement.
1276 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1279 #ifdef CONFIG_LOCKDEP
1281 * The caller should hold either p->pi_lock or rq->lock, when changing
1282 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1284 * sched_move_task() holds both and thus holding either pins the cgroup,
1287 * Furthermore, all task_rq users should acquire both locks, see
1290 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1291 lockdep_is_held(&task_rq(p)->lock)));
1295 trace_sched_migrate_task(p, new_cpu);
1297 if (task_cpu(p) != new_cpu) {
1298 if (p->sched_class->migrate_task_rq)
1299 p->sched_class->migrate_task_rq(p);
1300 p->se.nr_migrations++;
1301 perf_event_task_migrate(p);
1304 __set_task_cpu(p, new_cpu);
1307 static void __migrate_swap_task(struct task_struct *p, int cpu)
1309 if (task_on_rq_queued(p)) {
1310 struct rq *src_rq, *dst_rq;
1312 src_rq = task_rq(p);
1313 dst_rq = cpu_rq(cpu);
1315 deactivate_task(src_rq, p, 0);
1316 set_task_cpu(p, cpu);
1317 activate_task(dst_rq, p, 0);
1318 check_preempt_curr(dst_rq, p, 0);
1321 * Task isn't running anymore; make it appear like we migrated
1322 * it before it went to sleep. This means on wakeup we make the
1323 * previous cpu our targer instead of where it really is.
1329 struct migration_swap_arg {
1330 struct task_struct *src_task, *dst_task;
1331 int src_cpu, dst_cpu;
1334 static int migrate_swap_stop(void *data)
1336 struct migration_swap_arg *arg = data;
1337 struct rq *src_rq, *dst_rq;
1340 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1343 src_rq = cpu_rq(arg->src_cpu);
1344 dst_rq = cpu_rq(arg->dst_cpu);
1346 double_raw_lock(&arg->src_task->pi_lock,
1347 &arg->dst_task->pi_lock);
1348 double_rq_lock(src_rq, dst_rq);
1350 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1353 if (task_cpu(arg->src_task) != arg->src_cpu)
1356 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1359 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1362 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1363 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1368 double_rq_unlock(src_rq, dst_rq);
1369 raw_spin_unlock(&arg->dst_task->pi_lock);
1370 raw_spin_unlock(&arg->src_task->pi_lock);
1376 * Cross migrate two tasks
1378 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1380 struct migration_swap_arg arg;
1383 arg = (struct migration_swap_arg){
1385 .src_cpu = task_cpu(cur),
1387 .dst_cpu = task_cpu(p),
1390 if (arg.src_cpu == arg.dst_cpu)
1394 * These three tests are all lockless; this is OK since all of them
1395 * will be re-checked with proper locks held further down the line.
1397 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1400 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1403 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1406 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1407 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1414 * wait_task_inactive - wait for a thread to unschedule.
1416 * If @match_state is nonzero, it's the @p->state value just checked and
1417 * not expected to change. If it changes, i.e. @p might have woken up,
1418 * then return zero. When we succeed in waiting for @p to be off its CPU,
1419 * we return a positive number (its total switch count). If a second call
1420 * a short while later returns the same number, the caller can be sure that
1421 * @p has remained unscheduled the whole time.
1423 * The caller must ensure that the task *will* unschedule sometime soon,
1424 * else this function might spin for a *long* time. This function can't
1425 * be called with interrupts off, or it may introduce deadlock with
1426 * smp_call_function() if an IPI is sent by the same process we are
1427 * waiting to become inactive.
1429 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1431 unsigned long flags;
1432 int running, queued;
1438 * We do the initial early heuristics without holding
1439 * any task-queue locks at all. We'll only try to get
1440 * the runqueue lock when things look like they will
1446 * If the task is actively running on another CPU
1447 * still, just relax and busy-wait without holding
1450 * NOTE! Since we don't hold any locks, it's not
1451 * even sure that "rq" stays as the right runqueue!
1452 * But we don't care, since "task_running()" will
1453 * return false if the runqueue has changed and p
1454 * is actually now running somewhere else!
1456 while (task_running(rq, p)) {
1457 if (match_state && unlikely(p->state != match_state))
1463 * Ok, time to look more closely! We need the rq
1464 * lock now, to be *sure*. If we're wrong, we'll
1465 * just go back and repeat.
1467 rq = task_rq_lock(p, &flags);
1468 trace_sched_wait_task(p);
1469 running = task_running(rq, p);
1470 queued = task_on_rq_queued(p);
1472 if (!match_state || p->state == match_state)
1473 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1474 task_rq_unlock(rq, p, &flags);
1477 * If it changed from the expected state, bail out now.
1479 if (unlikely(!ncsw))
1483 * Was it really running after all now that we
1484 * checked with the proper locks actually held?
1486 * Oops. Go back and try again..
1488 if (unlikely(running)) {
1494 * It's not enough that it's not actively running,
1495 * it must be off the runqueue _entirely_, and not
1498 * So if it was still runnable (but just not actively
1499 * running right now), it's preempted, and we should
1500 * yield - it could be a while.
1502 if (unlikely(queued)) {
1503 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1505 set_current_state(TASK_UNINTERRUPTIBLE);
1506 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1511 * Ahh, all good. It wasn't running, and it wasn't
1512 * runnable, which means that it will never become
1513 * running in the future either. We're all done!
1522 * kick_process - kick a running thread to enter/exit the kernel
1523 * @p: the to-be-kicked thread
1525 * Cause a process which is running on another CPU to enter
1526 * kernel-mode, without any delay. (to get signals handled.)
1528 * NOTE: this function doesn't have to take the runqueue lock,
1529 * because all it wants to ensure is that the remote task enters
1530 * the kernel. If the IPI races and the task has been migrated
1531 * to another CPU then no harm is done and the purpose has been
1534 void kick_process(struct task_struct *p)
1540 if ((cpu != smp_processor_id()) && task_curr(p))
1541 smp_send_reschedule(cpu);
1544 EXPORT_SYMBOL_GPL(kick_process);
1547 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1549 static int select_fallback_rq(int cpu, struct task_struct *p)
1551 int nid = cpu_to_node(cpu);
1552 const struct cpumask *nodemask = NULL;
1553 enum { cpuset, possible, fail } state = cpuset;
1557 * If the node that the cpu is on has been offlined, cpu_to_node()
1558 * will return -1. There is no cpu on the node, and we should
1559 * select the cpu on the other node.
1562 nodemask = cpumask_of_node(nid);
1564 /* Look for allowed, online CPU in same node. */
1565 for_each_cpu(dest_cpu, nodemask) {
1566 if (!cpu_online(dest_cpu))
1568 if (!cpu_active(dest_cpu))
1570 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1576 /* Any allowed, online CPU? */
1577 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1578 if (!cpu_online(dest_cpu))
1580 if (!cpu_active(dest_cpu))
1585 /* No more Mr. Nice Guy. */
1588 if (IS_ENABLED(CONFIG_CPUSETS)) {
1589 cpuset_cpus_allowed_fallback(p);
1595 do_set_cpus_allowed(p, cpu_possible_mask);
1606 if (state != cpuset) {
1608 * Don't tell them about moving exiting tasks or
1609 * kernel threads (both mm NULL), since they never
1612 if (p->mm && printk_ratelimit()) {
1613 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1614 task_pid_nr(p), p->comm, cpu);
1622 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1625 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1627 lockdep_assert_held(&p->pi_lock);
1629 if (p->nr_cpus_allowed > 1)
1630 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1633 * In order not to call set_task_cpu() on a blocking task we need
1634 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1637 * Since this is common to all placement strategies, this lives here.
1639 * [ this allows ->select_task() to simply return task_cpu(p) and
1640 * not worry about this generic constraint ]
1642 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1644 cpu = select_fallback_rq(task_cpu(p), p);
1649 static void update_avg(u64 *avg, u64 sample)
1651 s64 diff = sample - *avg;
1657 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1658 const struct cpumask *new_mask, bool check)
1660 return set_cpus_allowed_ptr(p, new_mask);
1663 #endif /* CONFIG_SMP */
1666 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1668 #ifdef CONFIG_SCHEDSTATS
1669 struct rq *rq = this_rq();
1672 int this_cpu = smp_processor_id();
1674 if (cpu == this_cpu) {
1675 schedstat_inc(rq, ttwu_local);
1676 schedstat_inc(p, se.statistics.nr_wakeups_local);
1678 struct sched_domain *sd;
1680 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1682 for_each_domain(this_cpu, sd) {
1683 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1684 schedstat_inc(sd, ttwu_wake_remote);
1691 if (wake_flags & WF_MIGRATED)
1692 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1694 #endif /* CONFIG_SMP */
1696 schedstat_inc(rq, ttwu_count);
1697 schedstat_inc(p, se.statistics.nr_wakeups);
1699 if (wake_flags & WF_SYNC)
1700 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1702 #endif /* CONFIG_SCHEDSTATS */
1705 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1707 activate_task(rq, p, en_flags);
1708 p->on_rq = TASK_ON_RQ_QUEUED;
1710 /* if a worker is waking up, notify workqueue */
1711 if (p->flags & PF_WQ_WORKER)
1712 wq_worker_waking_up(p, cpu_of(rq));
1716 * Mark the task runnable and perform wakeup-preemption.
1719 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1721 check_preempt_curr(rq, p, wake_flags);
1722 p->state = TASK_RUNNING;
1723 trace_sched_wakeup(p);
1726 if (p->sched_class->task_woken) {
1728 * Our task @p is fully woken up and running; so its safe to
1729 * drop the rq->lock, hereafter rq is only used for statistics.
1731 lockdep_unpin_lock(&rq->lock);
1732 p->sched_class->task_woken(rq, p);
1733 lockdep_pin_lock(&rq->lock);
1736 if (rq->idle_stamp) {
1737 u64 delta = rq_clock(rq) - rq->idle_stamp;
1738 u64 max = 2*rq->max_idle_balance_cost;
1740 update_avg(&rq->avg_idle, delta);
1742 if (rq->avg_idle > max)
1751 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1753 lockdep_assert_held(&rq->lock);
1756 if (p->sched_contributes_to_load)
1757 rq->nr_uninterruptible--;
1760 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1761 ttwu_do_wakeup(rq, p, wake_flags);
1765 * Called in case the task @p isn't fully descheduled from its runqueue,
1766 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1767 * since all we need to do is flip p->state to TASK_RUNNING, since
1768 * the task is still ->on_rq.
1770 static int ttwu_remote(struct task_struct *p, int wake_flags)
1775 rq = __task_rq_lock(p);
1776 if (task_on_rq_queued(p)) {
1777 /* check_preempt_curr() may use rq clock */
1778 update_rq_clock(rq);
1779 ttwu_do_wakeup(rq, p, wake_flags);
1782 __task_rq_unlock(rq);
1788 void sched_ttwu_pending(void)
1790 struct rq *rq = this_rq();
1791 struct llist_node *llist = llist_del_all(&rq->wake_list);
1792 struct task_struct *p;
1793 unsigned long flags;
1798 raw_spin_lock_irqsave(&rq->lock, flags);
1799 lockdep_pin_lock(&rq->lock);
1802 p = llist_entry(llist, struct task_struct, wake_entry);
1803 llist = llist_next(llist);
1804 ttwu_do_activate(rq, p, 0);
1807 lockdep_unpin_lock(&rq->lock);
1808 raw_spin_unlock_irqrestore(&rq->lock, flags);
1811 void scheduler_ipi(void)
1814 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1815 * TIF_NEED_RESCHED remotely (for the first time) will also send
1818 preempt_fold_need_resched();
1820 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1824 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1825 * traditionally all their work was done from the interrupt return
1826 * path. Now that we actually do some work, we need to make sure
1829 * Some archs already do call them, luckily irq_enter/exit nest
1832 * Arguably we should visit all archs and update all handlers,
1833 * however a fair share of IPIs are still resched only so this would
1834 * somewhat pessimize the simple resched case.
1837 sched_ttwu_pending();
1840 * Check if someone kicked us for doing the nohz idle load balance.
1842 if (unlikely(got_nohz_idle_kick())) {
1843 this_rq()->idle_balance = 1;
1844 raise_softirq_irqoff(SCHED_SOFTIRQ);
1849 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1851 struct rq *rq = cpu_rq(cpu);
1853 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1854 if (!set_nr_if_polling(rq->idle))
1855 smp_send_reschedule(cpu);
1857 trace_sched_wake_idle_without_ipi(cpu);
1861 void wake_up_if_idle(int cpu)
1863 struct rq *rq = cpu_rq(cpu);
1864 unsigned long flags;
1868 if (!is_idle_task(rcu_dereference(rq->curr)))
1871 if (set_nr_if_polling(rq->idle)) {
1872 trace_sched_wake_idle_without_ipi(cpu);
1874 raw_spin_lock_irqsave(&rq->lock, flags);
1875 if (is_idle_task(rq->curr))
1876 smp_send_reschedule(cpu);
1877 /* Else cpu is not in idle, do nothing here */
1878 raw_spin_unlock_irqrestore(&rq->lock, flags);
1885 bool cpus_share_cache(int this_cpu, int that_cpu)
1887 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1889 #endif /* CONFIG_SMP */
1891 static void ttwu_queue(struct task_struct *p, int cpu)
1893 struct rq *rq = cpu_rq(cpu);
1895 #if defined(CONFIG_SMP)
1896 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1897 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1898 ttwu_queue_remote(p, cpu);
1903 raw_spin_lock(&rq->lock);
1904 lockdep_pin_lock(&rq->lock);
1905 ttwu_do_activate(rq, p, 0);
1906 lockdep_unpin_lock(&rq->lock);
1907 raw_spin_unlock(&rq->lock);
1911 * try_to_wake_up - wake up a thread
1912 * @p: the thread to be awakened
1913 * @state: the mask of task states that can be woken
1914 * @wake_flags: wake modifier flags (WF_*)
1916 * Put it on the run-queue if it's not already there. The "current"
1917 * thread is always on the run-queue (except when the actual
1918 * re-schedule is in progress), and as such you're allowed to do
1919 * the simpler "current->state = TASK_RUNNING" to mark yourself
1920 * runnable without the overhead of this.
1922 * Return: %true if @p was woken up, %false if it was already running.
1923 * or @state didn't match @p's state.
1926 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1928 unsigned long flags;
1929 int cpu, success = 0;
1932 * If we are going to wake up a thread waiting for CONDITION we
1933 * need to ensure that CONDITION=1 done by the caller can not be
1934 * reordered with p->state check below. This pairs with mb() in
1935 * set_current_state() the waiting thread does.
1937 smp_mb__before_spinlock();
1938 raw_spin_lock_irqsave(&p->pi_lock, flags);
1939 if (!(p->state & state))
1942 trace_sched_waking(p);
1944 success = 1; /* we're going to change ->state */
1948 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1949 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1950 * in smp_cond_load_acquire() below.
1952 * sched_ttwu_pending() try_to_wake_up()
1953 * [S] p->on_rq = 1; [L] P->state
1954 * UNLOCK rq->lock -----.
1958 * LOCK rq->lock -----'
1962 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1964 * Pairs with the UNLOCK+LOCK on rq->lock from the
1965 * last wakeup of our task and the schedule that got our task
1969 if (p->on_rq && ttwu_remote(p, wake_flags))
1974 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1975 * possible to, falsely, observe p->on_cpu == 0.
1977 * One must be running (->on_cpu == 1) in order to remove oneself
1978 * from the runqueue.
1980 * [S] ->on_cpu = 1; [L] ->on_rq
1984 * [S] ->on_rq = 0; [L] ->on_cpu
1986 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1987 * from the consecutive calls to schedule(); the first switching to our
1988 * task, the second putting it to sleep.
1993 * If the owning (remote) cpu is still in the middle of schedule() with
1994 * this task as prev, wait until its done referencing the task.
1999 * Combined with the control dependency above, we have an effective
2000 * smp_load_acquire() without the need for full barriers.
2002 * Pairs with the smp_store_release() in finish_lock_switch().
2004 * This ensures that tasks getting woken will be fully ordered against
2005 * their previous state and preserve Program Order.
2009 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2010 p->state = TASK_WAKING;
2012 if (p->sched_class->task_waking)
2013 p->sched_class->task_waking(p);
2015 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2016 if (task_cpu(p) != cpu) {
2017 wake_flags |= WF_MIGRATED;
2018 set_task_cpu(p, cpu);
2020 #endif /* CONFIG_SMP */
2024 ttwu_stat(p, cpu, wake_flags);
2026 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2032 * try_to_wake_up_local - try to wake up a local task with rq lock held
2033 * @p: the thread to be awakened
2035 * Put @p on the run-queue if it's not already there. The caller must
2036 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2039 static void try_to_wake_up_local(struct task_struct *p)
2041 struct rq *rq = task_rq(p);
2043 if (WARN_ON_ONCE(rq != this_rq()) ||
2044 WARN_ON_ONCE(p == current))
2047 lockdep_assert_held(&rq->lock);
2049 if (!raw_spin_trylock(&p->pi_lock)) {
2051 * This is OK, because current is on_cpu, which avoids it being
2052 * picked for load-balance and preemption/IRQs are still
2053 * disabled avoiding further scheduler activity on it and we've
2054 * not yet picked a replacement task.
2056 lockdep_unpin_lock(&rq->lock);
2057 raw_spin_unlock(&rq->lock);
2058 raw_spin_lock(&p->pi_lock);
2059 raw_spin_lock(&rq->lock);
2060 lockdep_pin_lock(&rq->lock);
2063 if (!(p->state & TASK_NORMAL))
2066 trace_sched_waking(p);
2068 if (!task_on_rq_queued(p))
2069 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2071 ttwu_do_wakeup(rq, p, 0);
2072 ttwu_stat(p, smp_processor_id(), 0);
2074 raw_spin_unlock(&p->pi_lock);
2078 * wake_up_process - Wake up a specific process
2079 * @p: The process to be woken up.
2081 * Attempt to wake up the nominated process and move it to the set of runnable
2084 * Return: 1 if the process was woken up, 0 if it was already running.
2086 * It may be assumed that this function implies a write memory barrier before
2087 * changing the task state if and only if any tasks are woken up.
2089 int wake_up_process(struct task_struct *p)
2091 return try_to_wake_up(p, TASK_NORMAL, 0);
2093 EXPORT_SYMBOL(wake_up_process);
2095 int wake_up_state(struct task_struct *p, unsigned int state)
2097 return try_to_wake_up(p, state, 0);
2101 * This function clears the sched_dl_entity static params.
2103 void __dl_clear_params(struct task_struct *p)
2105 struct sched_dl_entity *dl_se = &p->dl;
2107 dl_se->dl_runtime = 0;
2108 dl_se->dl_deadline = 0;
2109 dl_se->dl_period = 0;
2113 dl_se->dl_throttled = 0;
2115 dl_se->dl_yielded = 0;
2119 * Perform scheduler related setup for a newly forked process p.
2120 * p is forked by current.
2122 * __sched_fork() is basic setup used by init_idle() too:
2124 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2129 p->se.exec_start = 0;
2130 p->se.sum_exec_runtime = 0;
2131 p->se.prev_sum_exec_runtime = 0;
2132 p->se.nr_migrations = 0;
2134 INIT_LIST_HEAD(&p->se.group_node);
2136 #ifdef CONFIG_SCHEDSTATS
2137 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2140 RB_CLEAR_NODE(&p->dl.rb_node);
2141 init_dl_task_timer(&p->dl);
2142 __dl_clear_params(p);
2144 INIT_LIST_HEAD(&p->rt.run_list);
2146 #ifdef CONFIG_PREEMPT_NOTIFIERS
2147 INIT_HLIST_HEAD(&p->preempt_notifiers);
2150 #ifdef CONFIG_NUMA_BALANCING
2151 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2152 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2153 p->mm->numa_scan_seq = 0;
2156 if (clone_flags & CLONE_VM)
2157 p->numa_preferred_nid = current->numa_preferred_nid;
2159 p->numa_preferred_nid = -1;
2161 p->node_stamp = 0ULL;
2162 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2163 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2164 p->numa_work.next = &p->numa_work;
2165 p->numa_faults = NULL;
2166 p->last_task_numa_placement = 0;
2167 p->last_sum_exec_runtime = 0;
2169 p->numa_group = NULL;
2170 #endif /* CONFIG_NUMA_BALANCING */
2173 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2175 #ifdef CONFIG_NUMA_BALANCING
2177 void set_numabalancing_state(bool enabled)
2180 static_branch_enable(&sched_numa_balancing);
2182 static_branch_disable(&sched_numa_balancing);
2185 #ifdef CONFIG_PROC_SYSCTL
2186 int sysctl_numa_balancing(struct ctl_table *table, int write,
2187 void __user *buffer, size_t *lenp, loff_t *ppos)
2191 int state = static_branch_likely(&sched_numa_balancing);
2193 if (write && !capable(CAP_SYS_ADMIN))
2198 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2202 set_numabalancing_state(state);
2209 * fork()/clone()-time setup:
2211 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2213 unsigned long flags;
2214 int cpu = get_cpu();
2216 __sched_fork(clone_flags, p);
2218 * We mark the process as running here. This guarantees that
2219 * nobody will actually run it, and a signal or other external
2220 * event cannot wake it up and insert it on the runqueue either.
2222 p->state = TASK_RUNNING;
2225 * Make sure we do not leak PI boosting priority to the child.
2227 p->prio = current->normal_prio;
2230 * Revert to default priority/policy on fork if requested.
2232 if (unlikely(p->sched_reset_on_fork)) {
2233 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2234 p->policy = SCHED_NORMAL;
2235 p->static_prio = NICE_TO_PRIO(0);
2237 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2238 p->static_prio = NICE_TO_PRIO(0);
2240 p->prio = p->normal_prio = __normal_prio(p);
2244 * We don't need the reset flag anymore after the fork. It has
2245 * fulfilled its duty:
2247 p->sched_reset_on_fork = 0;
2250 if (dl_prio(p->prio)) {
2253 } else if (rt_prio(p->prio)) {
2254 p->sched_class = &rt_sched_class;
2256 p->sched_class = &fair_sched_class;
2259 if (p->sched_class->task_fork)
2260 p->sched_class->task_fork(p);
2263 * The child is not yet in the pid-hash so no cgroup attach races,
2264 * and the cgroup is pinned to this child due to cgroup_fork()
2265 * is ran before sched_fork().
2267 * Silence PROVE_RCU.
2269 raw_spin_lock_irqsave(&p->pi_lock, flags);
2270 set_task_cpu(p, cpu);
2271 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2273 #ifdef CONFIG_SCHED_INFO
2274 if (likely(sched_info_on()))
2275 memset(&p->sched_info, 0, sizeof(p->sched_info));
2277 #if defined(CONFIG_SMP)
2280 init_task_preempt_count(p);
2282 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2283 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2290 unsigned long to_ratio(u64 period, u64 runtime)
2292 if (runtime == RUNTIME_INF)
2296 * Doing this here saves a lot of checks in all
2297 * the calling paths, and returning zero seems
2298 * safe for them anyway.
2303 return div64_u64(runtime << 20, period);
2307 inline struct dl_bw *dl_bw_of(int i)
2309 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2310 "sched RCU must be held");
2311 return &cpu_rq(i)->rd->dl_bw;
2314 static inline int dl_bw_cpus(int i)
2316 struct root_domain *rd = cpu_rq(i)->rd;
2319 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2320 "sched RCU must be held");
2321 for_each_cpu_and(i, rd->span, cpu_active_mask)
2327 inline struct dl_bw *dl_bw_of(int i)
2329 return &cpu_rq(i)->dl.dl_bw;
2332 static inline int dl_bw_cpus(int i)
2339 * We must be sure that accepting a new task (or allowing changing the
2340 * parameters of an existing one) is consistent with the bandwidth
2341 * constraints. If yes, this function also accordingly updates the currently
2342 * allocated bandwidth to reflect the new situation.
2344 * This function is called while holding p's rq->lock.
2346 * XXX we should delay bw change until the task's 0-lag point, see
2349 static int dl_overflow(struct task_struct *p, int policy,
2350 const struct sched_attr *attr)
2353 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2354 u64 period = attr->sched_period ?: attr->sched_deadline;
2355 u64 runtime = attr->sched_runtime;
2356 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2359 if (new_bw == p->dl.dl_bw)
2363 * Either if a task, enters, leave, or stays -deadline but changes
2364 * its parameters, we may need to update accordingly the total
2365 * allocated bandwidth of the container.
2367 raw_spin_lock(&dl_b->lock);
2368 cpus = dl_bw_cpus(task_cpu(p));
2369 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2370 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2371 __dl_add(dl_b, new_bw);
2373 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2374 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2375 __dl_clear(dl_b, p->dl.dl_bw);
2376 __dl_add(dl_b, new_bw);
2378 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2379 __dl_clear(dl_b, p->dl.dl_bw);
2382 raw_spin_unlock(&dl_b->lock);
2387 extern void init_dl_bw(struct dl_bw *dl_b);
2390 * wake_up_new_task - wake up a newly created task for the first time.
2392 * This function will do some initial scheduler statistics housekeeping
2393 * that must be done for every newly created context, then puts the task
2394 * on the runqueue and wakes it.
2396 void wake_up_new_task(struct task_struct *p)
2398 unsigned long flags;
2401 raw_spin_lock_irqsave(&p->pi_lock, flags);
2402 /* Initialize new task's runnable average */
2403 init_entity_runnable_average(&p->se);
2406 * Fork balancing, do it here and not earlier because:
2407 * - cpus_allowed can change in the fork path
2408 * - any previously selected cpu might disappear through hotplug
2410 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2413 rq = __task_rq_lock(p);
2414 activate_task(rq, p, 0);
2415 p->on_rq = TASK_ON_RQ_QUEUED;
2416 trace_sched_wakeup_new(p);
2417 check_preempt_curr(rq, p, WF_FORK);
2419 if (p->sched_class->task_woken) {
2421 * Nothing relies on rq->lock after this, so its fine to
2424 lockdep_unpin_lock(&rq->lock);
2425 p->sched_class->task_woken(rq, p);
2426 lockdep_pin_lock(&rq->lock);
2429 task_rq_unlock(rq, p, &flags);
2432 #ifdef CONFIG_PREEMPT_NOTIFIERS
2434 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2436 void preempt_notifier_inc(void)
2438 static_key_slow_inc(&preempt_notifier_key);
2440 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2442 void preempt_notifier_dec(void)
2444 static_key_slow_dec(&preempt_notifier_key);
2446 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2449 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2450 * @notifier: notifier struct to register
2452 void preempt_notifier_register(struct preempt_notifier *notifier)
2454 if (!static_key_false(&preempt_notifier_key))
2455 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2457 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2459 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2462 * preempt_notifier_unregister - no longer interested in preemption notifications
2463 * @notifier: notifier struct to unregister
2465 * This is *not* safe to call from within a preemption notifier.
2467 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2469 hlist_del(¬ifier->link);
2471 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2473 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2475 struct preempt_notifier *notifier;
2477 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2478 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2481 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2483 if (static_key_false(&preempt_notifier_key))
2484 __fire_sched_in_preempt_notifiers(curr);
2488 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2489 struct task_struct *next)
2491 struct preempt_notifier *notifier;
2493 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2494 notifier->ops->sched_out(notifier, next);
2497 static __always_inline void
2498 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2499 struct task_struct *next)
2501 if (static_key_false(&preempt_notifier_key))
2502 __fire_sched_out_preempt_notifiers(curr, next);
2505 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2507 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2512 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2513 struct task_struct *next)
2517 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2520 * prepare_task_switch - prepare to switch tasks
2521 * @rq: the runqueue preparing to switch
2522 * @prev: the current task that is being switched out
2523 * @next: the task we are going to switch to.
2525 * This is called with the rq lock held and interrupts off. It must
2526 * be paired with a subsequent finish_task_switch after the context
2529 * prepare_task_switch sets up locking and calls architecture specific
2533 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2534 struct task_struct *next)
2536 sched_info_switch(rq, prev, next);
2537 perf_event_task_sched_out(prev, next);
2538 fire_sched_out_preempt_notifiers(prev, next);
2539 prepare_lock_switch(rq, next);
2540 prepare_arch_switch(next);
2544 * finish_task_switch - clean up after a task-switch
2545 * @prev: the thread we just switched away from.
2547 * finish_task_switch must be called after the context switch, paired
2548 * with a prepare_task_switch call before the context switch.
2549 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2550 * and do any other architecture-specific cleanup actions.
2552 * Note that we may have delayed dropping an mm in context_switch(). If
2553 * so, we finish that here outside of the runqueue lock. (Doing it
2554 * with the lock held can cause deadlocks; see schedule() for
2557 * The context switch have flipped the stack from under us and restored the
2558 * local variables which were saved when this task called schedule() in the
2559 * past. prev == current is still correct but we need to recalculate this_rq
2560 * because prev may have moved to another CPU.
2562 static struct rq *finish_task_switch(struct task_struct *prev)
2563 __releases(rq->lock)
2565 struct rq *rq = this_rq();
2566 struct mm_struct *mm = rq->prev_mm;
2570 * The previous task will have left us with a preempt_count of 2
2571 * because it left us after:
2574 * preempt_disable(); // 1
2576 * raw_spin_lock_irq(&rq->lock) // 2
2578 * Also, see FORK_PREEMPT_COUNT.
2580 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2581 "corrupted preempt_count: %s/%d/0x%x\n",
2582 current->comm, current->pid, preempt_count()))
2583 preempt_count_set(FORK_PREEMPT_COUNT);
2588 * A task struct has one reference for the use as "current".
2589 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2590 * schedule one last time. The schedule call will never return, and
2591 * the scheduled task must drop that reference.
2593 * We must observe prev->state before clearing prev->on_cpu (in
2594 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2595 * running on another CPU and we could rave with its RUNNING -> DEAD
2596 * transition, resulting in a double drop.
2598 prev_state = prev->state;
2599 vtime_task_switch(prev);
2600 perf_event_task_sched_in(prev, current);
2601 finish_lock_switch(rq, prev);
2602 finish_arch_post_lock_switch();
2604 fire_sched_in_preempt_notifiers(current);
2607 if (unlikely(prev_state == TASK_DEAD)) {
2608 if (prev->sched_class->task_dead)
2609 prev->sched_class->task_dead(prev);
2612 * Remove function-return probe instances associated with this
2613 * task and put them back on the free list.
2615 kprobe_flush_task(prev);
2616 put_task_struct(prev);
2619 tick_nohz_task_switch();
2625 /* rq->lock is NOT held, but preemption is disabled */
2626 static void __balance_callback(struct rq *rq)
2628 struct callback_head *head, *next;
2629 void (*func)(struct rq *rq);
2630 unsigned long flags;
2632 raw_spin_lock_irqsave(&rq->lock, flags);
2633 head = rq->balance_callback;
2634 rq->balance_callback = NULL;
2636 func = (void (*)(struct rq *))head->func;
2643 raw_spin_unlock_irqrestore(&rq->lock, flags);
2646 static inline void balance_callback(struct rq *rq)
2648 if (unlikely(rq->balance_callback))
2649 __balance_callback(rq);
2654 static inline void balance_callback(struct rq *rq)
2661 * schedule_tail - first thing a freshly forked thread must call.
2662 * @prev: the thread we just switched away from.
2664 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2665 __releases(rq->lock)
2670 * New tasks start with FORK_PREEMPT_COUNT, see there and
2671 * finish_task_switch() for details.
2673 * finish_task_switch() will drop rq->lock() and lower preempt_count
2674 * and the preempt_enable() will end up enabling preemption (on
2675 * PREEMPT_COUNT kernels).
2678 rq = finish_task_switch(prev);
2679 balance_callback(rq);
2682 if (current->set_child_tid)
2683 put_user(task_pid_vnr(current), current->set_child_tid);
2687 * context_switch - switch to the new MM and the new thread's register state.
2689 static inline struct rq *
2690 context_switch(struct rq *rq, struct task_struct *prev,
2691 struct task_struct *next)
2693 struct mm_struct *mm, *oldmm;
2695 prepare_task_switch(rq, prev, next);
2698 oldmm = prev->active_mm;
2700 * For paravirt, this is coupled with an exit in switch_to to
2701 * combine the page table reload and the switch backend into
2704 arch_start_context_switch(prev);
2707 next->active_mm = oldmm;
2708 atomic_inc(&oldmm->mm_count);
2709 enter_lazy_tlb(oldmm, next);
2711 switch_mm_irqs_off(oldmm, mm, next);
2714 prev->active_mm = NULL;
2715 rq->prev_mm = oldmm;
2718 * Since the runqueue lock will be released by the next
2719 * task (which is an invalid locking op but in the case
2720 * of the scheduler it's an obvious special-case), so we
2721 * do an early lockdep release here:
2723 lockdep_unpin_lock(&rq->lock);
2724 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2726 /* Here we just switch the register state and the stack. */
2727 switch_to(prev, next, prev);
2730 return finish_task_switch(prev);
2734 * nr_running and nr_context_switches:
2736 * externally visible scheduler statistics: current number of runnable
2737 * threads, total number of context switches performed since bootup.
2739 unsigned long nr_running(void)
2741 unsigned long i, sum = 0;
2743 for_each_online_cpu(i)
2744 sum += cpu_rq(i)->nr_running;
2750 * Check if only the current task is running on the cpu.
2752 * Caution: this function does not check that the caller has disabled
2753 * preemption, thus the result might have a time-of-check-to-time-of-use
2754 * race. The caller is responsible to use it correctly, for example:
2756 * - from a non-preemptable section (of course)
2758 * - from a thread that is bound to a single CPU
2760 * - in a loop with very short iterations (e.g. a polling loop)
2762 bool single_task_running(void)
2764 return raw_rq()->nr_running == 1;
2766 EXPORT_SYMBOL(single_task_running);
2768 unsigned long long nr_context_switches(void)
2771 unsigned long long sum = 0;
2773 for_each_possible_cpu(i)
2774 sum += cpu_rq(i)->nr_switches;
2779 unsigned long nr_iowait(void)
2781 unsigned long i, sum = 0;
2783 for_each_possible_cpu(i)
2784 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2789 unsigned long nr_iowait_cpu(int cpu)
2791 struct rq *this = cpu_rq(cpu);
2792 return atomic_read(&this->nr_iowait);
2795 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2797 struct rq *rq = this_rq();
2798 *nr_waiters = atomic_read(&rq->nr_iowait);
2799 *load = rq->load.weight;
2805 * sched_exec - execve() is a valuable balancing opportunity, because at
2806 * this point the task has the smallest effective memory and cache footprint.
2808 void sched_exec(void)
2810 struct task_struct *p = current;
2811 unsigned long flags;
2814 raw_spin_lock_irqsave(&p->pi_lock, flags);
2815 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2816 if (dest_cpu == smp_processor_id())
2819 if (likely(cpu_active(dest_cpu))) {
2820 struct migration_arg arg = { p, dest_cpu };
2822 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2823 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2827 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2832 DEFINE_PER_CPU(struct kernel_stat, kstat);
2833 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2835 EXPORT_PER_CPU_SYMBOL(kstat);
2836 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2839 * Return accounted runtime for the task.
2840 * In case the task is currently running, return the runtime plus current's
2841 * pending runtime that have not been accounted yet.
2843 unsigned long long task_sched_runtime(struct task_struct *p)
2845 unsigned long flags;
2849 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2851 * 64-bit doesn't need locks to atomically read a 64bit value.
2852 * So we have a optimization chance when the task's delta_exec is 0.
2853 * Reading ->on_cpu is racy, but this is ok.
2855 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2856 * If we race with it entering cpu, unaccounted time is 0. This is
2857 * indistinguishable from the read occurring a few cycles earlier.
2858 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2859 * been accounted, so we're correct here as well.
2861 if (!p->on_cpu || !task_on_rq_queued(p))
2862 return p->se.sum_exec_runtime;
2865 rq = task_rq_lock(p, &flags);
2867 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2868 * project cycles that may never be accounted to this
2869 * thread, breaking clock_gettime().
2871 if (task_current(rq, p) && task_on_rq_queued(p)) {
2872 update_rq_clock(rq);
2873 p->sched_class->update_curr(rq);
2875 ns = p->se.sum_exec_runtime;
2876 task_rq_unlock(rq, p, &flags);
2882 * This function gets called by the timer code, with HZ frequency.
2883 * We call it with interrupts disabled.
2885 void scheduler_tick(void)
2887 int cpu = smp_processor_id();
2888 struct rq *rq = cpu_rq(cpu);
2889 struct task_struct *curr = rq->curr;
2893 raw_spin_lock(&rq->lock);
2894 update_rq_clock(rq);
2895 curr->sched_class->task_tick(rq, curr, 0);
2896 update_cpu_load_active(rq);
2897 calc_global_load_tick(rq);
2898 raw_spin_unlock(&rq->lock);
2900 perf_event_task_tick();
2903 rq->idle_balance = idle_cpu(cpu);
2904 trigger_load_balance(rq);
2906 rq_last_tick_reset(rq);
2909 #ifdef CONFIG_NO_HZ_FULL
2911 * scheduler_tick_max_deferment
2913 * Keep at least one tick per second when a single
2914 * active task is running because the scheduler doesn't
2915 * yet completely support full dynticks environment.
2917 * This makes sure that uptime, CFS vruntime, load
2918 * balancing, etc... continue to move forward, even
2919 * with a very low granularity.
2921 * Return: Maximum deferment in nanoseconds.
2923 u64 scheduler_tick_max_deferment(void)
2925 struct rq *rq = this_rq();
2926 unsigned long next, now = READ_ONCE(jiffies);
2928 next = rq->last_sched_tick + HZ;
2930 if (time_before_eq(next, now))
2933 return jiffies_to_nsecs(next - now);
2937 notrace unsigned long get_parent_ip(unsigned long addr)
2939 if (in_lock_functions(addr)) {
2940 addr = CALLER_ADDR2;
2941 if (in_lock_functions(addr))
2942 addr = CALLER_ADDR3;
2947 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2948 defined(CONFIG_PREEMPT_TRACER))
2950 void preempt_count_add(int val)
2952 #ifdef CONFIG_DEBUG_PREEMPT
2956 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2959 __preempt_count_add(val);
2960 #ifdef CONFIG_DEBUG_PREEMPT
2962 * Spinlock count overflowing soon?
2964 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2967 if (preempt_count() == val) {
2968 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2969 #ifdef CONFIG_DEBUG_PREEMPT
2970 current->preempt_disable_ip = ip;
2972 trace_preempt_off(CALLER_ADDR0, ip);
2975 EXPORT_SYMBOL(preempt_count_add);
2976 NOKPROBE_SYMBOL(preempt_count_add);
2978 void preempt_count_sub(int val)
2980 #ifdef CONFIG_DEBUG_PREEMPT
2984 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2987 * Is the spinlock portion underflowing?
2989 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2990 !(preempt_count() & PREEMPT_MASK)))
2994 if (preempt_count() == val)
2995 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2996 __preempt_count_sub(val);
2998 EXPORT_SYMBOL(preempt_count_sub);
2999 NOKPROBE_SYMBOL(preempt_count_sub);
3004 * Print scheduling while atomic bug:
3006 static noinline void __schedule_bug(struct task_struct *prev)
3008 if (oops_in_progress)
3011 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3012 prev->comm, prev->pid, preempt_count());
3014 debug_show_held_locks(prev);
3016 if (irqs_disabled())
3017 print_irqtrace_events(prev);
3018 #ifdef CONFIG_DEBUG_PREEMPT
3019 if (in_atomic_preempt_off()) {
3020 pr_err("Preemption disabled at:");
3021 print_ip_sym(current->preempt_disable_ip);
3026 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3030 * Various schedule()-time debugging checks and statistics:
3032 static inline void schedule_debug(struct task_struct *prev)
3034 #ifdef CONFIG_SCHED_STACK_END_CHECK
3035 if (task_stack_end_corrupted(prev))
3036 panic("corrupted stack end detected inside scheduler\n");
3039 if (unlikely(in_atomic_preempt_off())) {
3040 __schedule_bug(prev);
3041 preempt_count_set(PREEMPT_DISABLED);
3045 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3047 schedstat_inc(this_rq(), sched_count);
3051 * Pick up the highest-prio task:
3053 static inline struct task_struct *
3054 pick_next_task(struct rq *rq, struct task_struct *prev)
3056 const struct sched_class *class = &fair_sched_class;
3057 struct task_struct *p;
3060 * Optimization: we know that if all tasks are in
3061 * the fair class we can call that function directly:
3063 if (likely(prev->sched_class == class &&
3064 rq->nr_running == rq->cfs.h_nr_running)) {
3065 p = fair_sched_class.pick_next_task(rq, prev);
3066 if (unlikely(p == RETRY_TASK))
3069 /* assumes fair_sched_class->next == idle_sched_class */
3071 p = idle_sched_class.pick_next_task(rq, prev);
3077 for_each_class(class) {
3078 p = class->pick_next_task(rq, prev);
3080 if (unlikely(p == RETRY_TASK))
3086 BUG(); /* the idle class will always have a runnable task */
3090 * __schedule() is the main scheduler function.
3092 * The main means of driving the scheduler and thus entering this function are:
3094 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3096 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3097 * paths. For example, see arch/x86/entry_64.S.
3099 * To drive preemption between tasks, the scheduler sets the flag in timer
3100 * interrupt handler scheduler_tick().
3102 * 3. Wakeups don't really cause entry into schedule(). They add a
3103 * task to the run-queue and that's it.
3105 * Now, if the new task added to the run-queue preempts the current
3106 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3107 * called on the nearest possible occasion:
3109 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3111 * - in syscall or exception context, at the next outmost
3112 * preempt_enable(). (this might be as soon as the wake_up()'s
3115 * - in IRQ context, return from interrupt-handler to
3116 * preemptible context
3118 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3121 * - cond_resched() call
3122 * - explicit schedule() call
3123 * - return from syscall or exception to user-space
3124 * - return from interrupt-handler to user-space
3126 * WARNING: must be called with preemption disabled!
3128 static void __sched notrace __schedule(bool preempt)
3130 struct task_struct *prev, *next;
3131 unsigned long *switch_count;
3135 cpu = smp_processor_id();
3137 rcu_note_context_switch();
3141 * do_exit() calls schedule() with preemption disabled as an exception;
3142 * however we must fix that up, otherwise the next task will see an
3143 * inconsistent (higher) preempt count.
3145 * It also avoids the below schedule_debug() test from complaining
3148 if (unlikely(prev->state == TASK_DEAD))
3149 preempt_enable_no_resched_notrace();
3151 schedule_debug(prev);
3153 if (sched_feat(HRTICK))
3157 * Make sure that signal_pending_state()->signal_pending() below
3158 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3159 * done by the caller to avoid the race with signal_wake_up().
3161 smp_mb__before_spinlock();
3162 raw_spin_lock_irq(&rq->lock);
3163 lockdep_pin_lock(&rq->lock);
3165 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3167 switch_count = &prev->nivcsw;
3168 if (!preempt && prev->state) {
3169 if (unlikely(signal_pending_state(prev->state, prev))) {
3170 prev->state = TASK_RUNNING;
3172 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3176 * If a worker went to sleep, notify and ask workqueue
3177 * whether it wants to wake up a task to maintain
3180 if (prev->flags & PF_WQ_WORKER) {
3181 struct task_struct *to_wakeup;
3183 to_wakeup = wq_worker_sleeping(prev, cpu);
3185 try_to_wake_up_local(to_wakeup);
3188 switch_count = &prev->nvcsw;
3191 if (task_on_rq_queued(prev))
3192 update_rq_clock(rq);
3194 next = pick_next_task(rq, prev);
3195 clear_tsk_need_resched(prev);
3196 clear_preempt_need_resched();
3197 rq->clock_skip_update = 0;
3199 if (likely(prev != next)) {
3204 trace_sched_switch(preempt, prev, next);
3205 rq = context_switch(rq, prev, next); /* unlocks the rq */
3208 lockdep_unpin_lock(&rq->lock);
3209 raw_spin_unlock_irq(&rq->lock);
3212 balance_callback(rq);
3215 static inline void sched_submit_work(struct task_struct *tsk)
3217 if (!tsk->state || tsk_is_pi_blocked(tsk))
3220 * If we are going to sleep and we have plugged IO queued,
3221 * make sure to submit it to avoid deadlocks.
3223 if (blk_needs_flush_plug(tsk))
3224 blk_schedule_flush_plug(tsk);
3227 asmlinkage __visible void __sched schedule(void)
3229 struct task_struct *tsk = current;
3231 sched_submit_work(tsk);
3235 sched_preempt_enable_no_resched();
3236 } while (need_resched());
3238 EXPORT_SYMBOL(schedule);
3240 #ifdef CONFIG_CONTEXT_TRACKING
3241 asmlinkage __visible void __sched schedule_user(void)
3244 * If we come here after a random call to set_need_resched(),
3245 * or we have been woken up remotely but the IPI has not yet arrived,
3246 * we haven't yet exited the RCU idle mode. Do it here manually until
3247 * we find a better solution.
3249 * NB: There are buggy callers of this function. Ideally we
3250 * should warn if prev_state != CONTEXT_USER, but that will trigger
3251 * too frequently to make sense yet.
3253 enum ctx_state prev_state = exception_enter();
3255 exception_exit(prev_state);
3260 * schedule_preempt_disabled - called with preemption disabled
3262 * Returns with preemption disabled. Note: preempt_count must be 1
3264 void __sched schedule_preempt_disabled(void)
3266 sched_preempt_enable_no_resched();
3271 static void __sched notrace preempt_schedule_common(void)
3274 preempt_disable_notrace();
3276 preempt_enable_no_resched_notrace();
3279 * Check again in case we missed a preemption opportunity
3280 * between schedule and now.
3282 } while (need_resched());
3285 #ifdef CONFIG_PREEMPT
3287 * this is the entry point to schedule() from in-kernel preemption
3288 * off of preempt_enable. Kernel preemptions off return from interrupt
3289 * occur there and call schedule directly.
3291 asmlinkage __visible void __sched notrace preempt_schedule(void)
3294 * If there is a non-zero preempt_count or interrupts are disabled,
3295 * we do not want to preempt the current task. Just return..
3297 if (likely(!preemptible()))
3300 preempt_schedule_common();
3302 NOKPROBE_SYMBOL(preempt_schedule);
3303 EXPORT_SYMBOL(preempt_schedule);
3306 * preempt_schedule_notrace - preempt_schedule called by tracing
3308 * The tracing infrastructure uses preempt_enable_notrace to prevent
3309 * recursion and tracing preempt enabling caused by the tracing
3310 * infrastructure itself. But as tracing can happen in areas coming
3311 * from userspace or just about to enter userspace, a preempt enable
3312 * can occur before user_exit() is called. This will cause the scheduler
3313 * to be called when the system is still in usermode.
3315 * To prevent this, the preempt_enable_notrace will use this function
3316 * instead of preempt_schedule() to exit user context if needed before
3317 * calling the scheduler.
3319 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3321 enum ctx_state prev_ctx;
3323 if (likely(!preemptible()))
3327 preempt_disable_notrace();
3329 * Needs preempt disabled in case user_exit() is traced
3330 * and the tracer calls preempt_enable_notrace() causing
3331 * an infinite recursion.
3333 prev_ctx = exception_enter();
3335 exception_exit(prev_ctx);
3337 preempt_enable_no_resched_notrace();
3338 } while (need_resched());
3340 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3342 #endif /* CONFIG_PREEMPT */
3345 * this is the entry point to schedule() from kernel preemption
3346 * off of irq context.
3347 * Note, that this is called and return with irqs disabled. This will
3348 * protect us against recursive calling from irq.
3350 asmlinkage __visible void __sched preempt_schedule_irq(void)
3352 enum ctx_state prev_state;
3354 /* Catch callers which need to be fixed */
3355 BUG_ON(preempt_count() || !irqs_disabled());
3357 prev_state = exception_enter();
3363 local_irq_disable();
3364 sched_preempt_enable_no_resched();
3365 } while (need_resched());
3367 exception_exit(prev_state);
3370 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3373 return try_to_wake_up(curr->private, mode, wake_flags);
3375 EXPORT_SYMBOL(default_wake_function);
3377 #ifdef CONFIG_RT_MUTEXES
3380 * rt_mutex_setprio - set the current priority of a task
3382 * @prio: prio value (kernel-internal form)
3384 * This function changes the 'effective' priority of a task. It does
3385 * not touch ->normal_prio like __setscheduler().
3387 * Used by the rt_mutex code to implement priority inheritance
3388 * logic. Call site only calls if the priority of the task changed.
3390 void rt_mutex_setprio(struct task_struct *p, int prio)
3392 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3394 const struct sched_class *prev_class;
3396 BUG_ON(prio > MAX_PRIO);
3398 rq = __task_rq_lock(p);
3401 * Idle task boosting is a nono in general. There is one
3402 * exception, when PREEMPT_RT and NOHZ is active:
3404 * The idle task calls get_next_timer_interrupt() and holds
3405 * the timer wheel base->lock on the CPU and another CPU wants
3406 * to access the timer (probably to cancel it). We can safely
3407 * ignore the boosting request, as the idle CPU runs this code
3408 * with interrupts disabled and will complete the lock
3409 * protected section without being interrupted. So there is no
3410 * real need to boost.
3412 if (unlikely(p == rq->idle)) {
3413 WARN_ON(p != rq->curr);
3414 WARN_ON(p->pi_blocked_on);
3418 trace_sched_pi_setprio(p, prio);
3420 prev_class = p->sched_class;
3421 queued = task_on_rq_queued(p);
3422 running = task_current(rq, p);
3424 dequeue_task(rq, p, DEQUEUE_SAVE);
3426 put_prev_task(rq, p);
3429 * Boosting condition are:
3430 * 1. -rt task is running and holds mutex A
3431 * --> -dl task blocks on mutex A
3433 * 2. -dl task is running and holds mutex A
3434 * --> -dl task blocks on mutex A and could preempt the
3437 if (dl_prio(prio)) {
3438 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3439 if (!dl_prio(p->normal_prio) ||
3440 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3441 p->dl.dl_boosted = 1;
3442 enqueue_flag |= ENQUEUE_REPLENISH;
3444 p->dl.dl_boosted = 0;
3445 p->sched_class = &dl_sched_class;
3446 } else if (rt_prio(prio)) {
3447 if (dl_prio(oldprio))
3448 p->dl.dl_boosted = 0;
3450 enqueue_flag |= ENQUEUE_HEAD;
3451 p->sched_class = &rt_sched_class;
3453 if (dl_prio(oldprio))
3454 p->dl.dl_boosted = 0;
3455 if (rt_prio(oldprio))
3457 p->sched_class = &fair_sched_class;
3463 p->sched_class->set_curr_task(rq);
3465 enqueue_task(rq, p, enqueue_flag);
3467 check_class_changed(rq, p, prev_class, oldprio);
3469 preempt_disable(); /* avoid rq from going away on us */
3470 __task_rq_unlock(rq);
3472 balance_callback(rq);
3477 void set_user_nice(struct task_struct *p, long nice)
3479 int old_prio, delta, queued;
3480 unsigned long flags;
3483 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3486 * We have to be careful, if called from sys_setpriority(),
3487 * the task might be in the middle of scheduling on another CPU.
3489 rq = task_rq_lock(p, &flags);
3491 * The RT priorities are set via sched_setscheduler(), but we still
3492 * allow the 'normal' nice value to be set - but as expected
3493 * it wont have any effect on scheduling until the task is
3494 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3496 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3497 p->static_prio = NICE_TO_PRIO(nice);
3500 queued = task_on_rq_queued(p);
3502 dequeue_task(rq, p, DEQUEUE_SAVE);
3504 p->static_prio = NICE_TO_PRIO(nice);
3507 p->prio = effective_prio(p);
3508 delta = p->prio - old_prio;
3511 enqueue_task(rq, p, ENQUEUE_RESTORE);
3513 * If the task increased its priority or is running and
3514 * lowered its priority, then reschedule its CPU:
3516 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3520 task_rq_unlock(rq, p, &flags);
3522 EXPORT_SYMBOL(set_user_nice);
3525 * can_nice - check if a task can reduce its nice value
3529 int can_nice(const struct task_struct *p, const int nice)
3531 /* convert nice value [19,-20] to rlimit style value [1,40] */
3532 int nice_rlim = nice_to_rlimit(nice);
3534 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3535 capable(CAP_SYS_NICE));
3538 #ifdef __ARCH_WANT_SYS_NICE
3541 * sys_nice - change the priority of the current process.
3542 * @increment: priority increment
3544 * sys_setpriority is a more generic, but much slower function that
3545 * does similar things.
3547 SYSCALL_DEFINE1(nice, int, increment)
3552 * Setpriority might change our priority at the same moment.
3553 * We don't have to worry. Conceptually one call occurs first
3554 * and we have a single winner.
3556 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3557 nice = task_nice(current) + increment;
3559 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3560 if (increment < 0 && !can_nice(current, nice))
3563 retval = security_task_setnice(current, nice);
3567 set_user_nice(current, nice);
3574 * task_prio - return the priority value of a given task.
3575 * @p: the task in question.
3577 * Return: The priority value as seen by users in /proc.
3578 * RT tasks are offset by -200. Normal tasks are centered
3579 * around 0, value goes from -16 to +15.
3581 int task_prio(const struct task_struct *p)
3583 return p->prio - MAX_RT_PRIO;
3587 * idle_cpu - is a given cpu idle currently?
3588 * @cpu: the processor in question.
3590 * Return: 1 if the CPU is currently idle. 0 otherwise.
3592 int idle_cpu(int cpu)
3594 struct rq *rq = cpu_rq(cpu);
3596 if (rq->curr != rq->idle)
3603 if (!llist_empty(&rq->wake_list))
3611 * idle_task - return the idle task for a given cpu.
3612 * @cpu: the processor in question.
3614 * Return: The idle task for the cpu @cpu.
3616 struct task_struct *idle_task(int cpu)
3618 return cpu_rq(cpu)->idle;
3622 * find_process_by_pid - find a process with a matching PID value.
3623 * @pid: the pid in question.
3625 * The task of @pid, if found. %NULL otherwise.
3627 static struct task_struct *find_process_by_pid(pid_t pid)
3629 return pid ? find_task_by_vpid(pid) : current;
3633 * This function initializes the sched_dl_entity of a newly becoming
3634 * SCHED_DEADLINE task.
3636 * Only the static values are considered here, the actual runtime and the
3637 * absolute deadline will be properly calculated when the task is enqueued
3638 * for the first time with its new policy.
3641 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3643 struct sched_dl_entity *dl_se = &p->dl;
3645 dl_se->dl_runtime = attr->sched_runtime;
3646 dl_se->dl_deadline = attr->sched_deadline;
3647 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3648 dl_se->flags = attr->sched_flags;
3649 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3652 * Changing the parameters of a task is 'tricky' and we're not doing
3653 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3655 * What we SHOULD do is delay the bandwidth release until the 0-lag
3656 * point. This would include retaining the task_struct until that time
3657 * and change dl_overflow() to not immediately decrement the current
3660 * Instead we retain the current runtime/deadline and let the new
3661 * parameters take effect after the current reservation period lapses.
3662 * This is safe (albeit pessimistic) because the 0-lag point is always
3663 * before the current scheduling deadline.
3665 * We can still have temporary overloads because we do not delay the
3666 * change in bandwidth until that time; so admission control is
3667 * not on the safe side. It does however guarantee tasks will never
3668 * consume more than promised.
3673 * sched_setparam() passes in -1 for its policy, to let the functions
3674 * it calls know not to change it.
3676 #define SETPARAM_POLICY -1
3678 static void __setscheduler_params(struct task_struct *p,
3679 const struct sched_attr *attr)
3681 int policy = attr->sched_policy;
3683 if (policy == SETPARAM_POLICY)
3688 if (dl_policy(policy))
3689 __setparam_dl(p, attr);
3690 else if (fair_policy(policy))
3691 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3694 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3695 * !rt_policy. Always setting this ensures that things like
3696 * getparam()/getattr() don't report silly values for !rt tasks.
3698 p->rt_priority = attr->sched_priority;
3699 p->normal_prio = normal_prio(p);
3703 /* Actually do priority change: must hold pi & rq lock. */
3704 static void __setscheduler(struct rq *rq, struct task_struct *p,
3705 const struct sched_attr *attr, bool keep_boost)
3707 __setscheduler_params(p, attr);
3710 * Keep a potential priority boosting if called from
3711 * sched_setscheduler().
3714 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3716 p->prio = normal_prio(p);
3718 if (dl_prio(p->prio))
3719 p->sched_class = &dl_sched_class;
3720 else if (rt_prio(p->prio))
3721 p->sched_class = &rt_sched_class;
3723 p->sched_class = &fair_sched_class;
3727 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3729 struct sched_dl_entity *dl_se = &p->dl;
3731 attr->sched_priority = p->rt_priority;
3732 attr->sched_runtime = dl_se->dl_runtime;
3733 attr->sched_deadline = dl_se->dl_deadline;
3734 attr->sched_period = dl_se->dl_period;
3735 attr->sched_flags = dl_se->flags;
3739 * This function validates the new parameters of a -deadline task.
3740 * We ask for the deadline not being zero, and greater or equal
3741 * than the runtime, as well as the period of being zero or
3742 * greater than deadline. Furthermore, we have to be sure that
3743 * user parameters are above the internal resolution of 1us (we
3744 * check sched_runtime only since it is always the smaller one) and
3745 * below 2^63 ns (we have to check both sched_deadline and
3746 * sched_period, as the latter can be zero).
3749 __checkparam_dl(const struct sched_attr *attr)
3752 if (attr->sched_deadline == 0)
3756 * Since we truncate DL_SCALE bits, make sure we're at least
3759 if (attr->sched_runtime < (1ULL << DL_SCALE))
3763 * Since we use the MSB for wrap-around and sign issues, make
3764 * sure it's not set (mind that period can be equal to zero).
3766 if (attr->sched_deadline & (1ULL << 63) ||
3767 attr->sched_period & (1ULL << 63))
3770 /* runtime <= deadline <= period (if period != 0) */
3771 if ((attr->sched_period != 0 &&
3772 attr->sched_period < attr->sched_deadline) ||
3773 attr->sched_deadline < attr->sched_runtime)
3780 * check the target process has a UID that matches the current process's
3782 static bool check_same_owner(struct task_struct *p)
3784 const struct cred *cred = current_cred(), *pcred;
3788 pcred = __task_cred(p);
3789 match = (uid_eq(cred->euid, pcred->euid) ||
3790 uid_eq(cred->euid, pcred->uid));
3795 static bool dl_param_changed(struct task_struct *p,
3796 const struct sched_attr *attr)
3798 struct sched_dl_entity *dl_se = &p->dl;
3800 if (dl_se->dl_runtime != attr->sched_runtime ||
3801 dl_se->dl_deadline != attr->sched_deadline ||
3802 dl_se->dl_period != attr->sched_period ||
3803 dl_se->flags != attr->sched_flags)
3809 static int __sched_setscheduler(struct task_struct *p,
3810 const struct sched_attr *attr,
3813 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3814 MAX_RT_PRIO - 1 - attr->sched_priority;
3815 int retval, oldprio, oldpolicy = -1, queued, running;
3816 int new_effective_prio, policy = attr->sched_policy;
3817 unsigned long flags;
3818 const struct sched_class *prev_class;
3822 /* may grab non-irq protected spin_locks */
3823 BUG_ON(in_interrupt());
3825 /* double check policy once rq lock held */
3827 reset_on_fork = p->sched_reset_on_fork;
3828 policy = oldpolicy = p->policy;
3830 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3832 if (!valid_policy(policy))
3836 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3840 * Valid priorities for SCHED_FIFO and SCHED_RR are
3841 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3842 * SCHED_BATCH and SCHED_IDLE is 0.
3844 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3845 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3847 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3848 (rt_policy(policy) != (attr->sched_priority != 0)))
3852 * Allow unprivileged RT tasks to decrease priority:
3854 if (user && !capable(CAP_SYS_NICE)) {
3855 if (fair_policy(policy)) {
3856 if (attr->sched_nice < task_nice(p) &&
3857 !can_nice(p, attr->sched_nice))
3861 if (rt_policy(policy)) {
3862 unsigned long rlim_rtprio =
3863 task_rlimit(p, RLIMIT_RTPRIO);
3865 /* can't set/change the rt policy */
3866 if (policy != p->policy && !rlim_rtprio)
3869 /* can't increase priority */
3870 if (attr->sched_priority > p->rt_priority &&
3871 attr->sched_priority > rlim_rtprio)
3876 * Can't set/change SCHED_DEADLINE policy at all for now
3877 * (safest behavior); in the future we would like to allow
3878 * unprivileged DL tasks to increase their relative deadline
3879 * or reduce their runtime (both ways reducing utilization)
3881 if (dl_policy(policy))
3885 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3886 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3888 if (idle_policy(p->policy) && !idle_policy(policy)) {
3889 if (!can_nice(p, task_nice(p)))
3893 /* can't change other user's priorities */
3894 if (!check_same_owner(p))
3897 /* Normal users shall not reset the sched_reset_on_fork flag */
3898 if (p->sched_reset_on_fork && !reset_on_fork)
3903 retval = security_task_setscheduler(p);
3909 * make sure no PI-waiters arrive (or leave) while we are
3910 * changing the priority of the task:
3912 * To be able to change p->policy safely, the appropriate
3913 * runqueue lock must be held.
3915 rq = task_rq_lock(p, &flags);
3918 * Changing the policy of the stop threads its a very bad idea
3920 if (p == rq->stop) {
3921 task_rq_unlock(rq, p, &flags);
3926 * If not changing anything there's no need to proceed further,
3927 * but store a possible modification of reset_on_fork.
3929 if (unlikely(policy == p->policy)) {
3930 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3932 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3934 if (dl_policy(policy) && dl_param_changed(p, attr))
3937 p->sched_reset_on_fork = reset_on_fork;
3938 task_rq_unlock(rq, p, &flags);
3944 #ifdef CONFIG_RT_GROUP_SCHED
3946 * Do not allow realtime tasks into groups that have no runtime
3949 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3950 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3951 !task_group_is_autogroup(task_group(p))) {
3952 task_rq_unlock(rq, p, &flags);
3957 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3958 cpumask_t *span = rq->rd->span;
3961 * Don't allow tasks with an affinity mask smaller than
3962 * the entire root_domain to become SCHED_DEADLINE. We
3963 * will also fail if there's no bandwidth available.
3965 if (!cpumask_subset(span, &p->cpus_allowed) ||
3966 rq->rd->dl_bw.bw == 0) {
3967 task_rq_unlock(rq, p, &flags);
3974 /* recheck policy now with rq lock held */
3975 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3976 policy = oldpolicy = -1;
3977 task_rq_unlock(rq, p, &flags);
3982 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3983 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3986 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3987 task_rq_unlock(rq, p, &flags);
3991 p->sched_reset_on_fork = reset_on_fork;
3996 * Take priority boosted tasks into account. If the new
3997 * effective priority is unchanged, we just store the new
3998 * normal parameters and do not touch the scheduler class and
3999 * the runqueue. This will be done when the task deboost
4002 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4003 if (new_effective_prio == oldprio) {
4004 __setscheduler_params(p, attr);
4005 task_rq_unlock(rq, p, &flags);
4010 queued = task_on_rq_queued(p);
4011 running = task_current(rq, p);
4013 dequeue_task(rq, p, DEQUEUE_SAVE);
4015 put_prev_task(rq, p);
4017 prev_class = p->sched_class;
4018 __setscheduler(rq, p, attr, pi);
4021 p->sched_class->set_curr_task(rq);
4023 int enqueue_flags = ENQUEUE_RESTORE;
4025 * We enqueue to tail when the priority of a task is
4026 * increased (user space view).
4028 if (oldprio <= p->prio)
4029 enqueue_flags |= ENQUEUE_HEAD;
4031 enqueue_task(rq, p, enqueue_flags);
4034 check_class_changed(rq, p, prev_class, oldprio);
4035 preempt_disable(); /* avoid rq from going away on us */
4036 task_rq_unlock(rq, p, &flags);
4039 rt_mutex_adjust_pi(p);
4042 * Run balance callbacks after we've adjusted the PI chain.
4044 balance_callback(rq);
4050 static int _sched_setscheduler(struct task_struct *p, int policy,
4051 const struct sched_param *param, bool check)
4053 struct sched_attr attr = {
4054 .sched_policy = policy,
4055 .sched_priority = param->sched_priority,
4056 .sched_nice = PRIO_TO_NICE(p->static_prio),
4059 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4060 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4061 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4062 policy &= ~SCHED_RESET_ON_FORK;
4063 attr.sched_policy = policy;
4066 return __sched_setscheduler(p, &attr, check, true);
4069 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4070 * @p: the task in question.
4071 * @policy: new policy.
4072 * @param: structure containing the new RT priority.
4074 * Return: 0 on success. An error code otherwise.
4076 * NOTE that the task may be already dead.
4078 int sched_setscheduler(struct task_struct *p, int policy,
4079 const struct sched_param *param)
4081 return _sched_setscheduler(p, policy, param, true);
4083 EXPORT_SYMBOL_GPL(sched_setscheduler);
4085 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4087 return __sched_setscheduler(p, attr, true, true);
4089 EXPORT_SYMBOL_GPL(sched_setattr);
4092 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4093 * @p: the task in question.
4094 * @policy: new policy.
4095 * @param: structure containing the new RT priority.
4097 * Just like sched_setscheduler, only don't bother checking if the
4098 * current context has permission. For example, this is needed in
4099 * stop_machine(): we create temporary high priority worker threads,
4100 * but our caller might not have that capability.
4102 * Return: 0 on success. An error code otherwise.
4104 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4105 const struct sched_param *param)
4107 return _sched_setscheduler(p, policy, param, false);
4109 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4112 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4114 struct sched_param lparam;
4115 struct task_struct *p;
4118 if (!param || pid < 0)
4120 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4125 p = find_process_by_pid(pid);
4127 retval = sched_setscheduler(p, policy, &lparam);
4134 * Mimics kernel/events/core.c perf_copy_attr().
4136 static int sched_copy_attr(struct sched_attr __user *uattr,
4137 struct sched_attr *attr)
4142 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4146 * zero the full structure, so that a short copy will be nice.
4148 memset(attr, 0, sizeof(*attr));
4150 ret = get_user(size, &uattr->size);
4154 if (size > PAGE_SIZE) /* silly large */
4157 if (!size) /* abi compat */
4158 size = SCHED_ATTR_SIZE_VER0;
4160 if (size < SCHED_ATTR_SIZE_VER0)
4164 * If we're handed a bigger struct than we know of,
4165 * ensure all the unknown bits are 0 - i.e. new
4166 * user-space does not rely on any kernel feature
4167 * extensions we dont know about yet.
4169 if (size > sizeof(*attr)) {
4170 unsigned char __user *addr;
4171 unsigned char __user *end;
4174 addr = (void __user *)uattr + sizeof(*attr);
4175 end = (void __user *)uattr + size;
4177 for (; addr < end; addr++) {
4178 ret = get_user(val, addr);
4184 size = sizeof(*attr);
4187 ret = copy_from_user(attr, uattr, size);
4192 * XXX: do we want to be lenient like existing syscalls; or do we want
4193 * to be strict and return an error on out-of-bounds values?
4195 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4200 put_user(sizeof(*attr), &uattr->size);
4205 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4206 * @pid: the pid in question.
4207 * @policy: new policy.
4208 * @param: structure containing the new RT priority.
4210 * Return: 0 on success. An error code otherwise.
4212 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4213 struct sched_param __user *, param)
4215 /* negative values for policy are not valid */
4219 return do_sched_setscheduler(pid, policy, param);
4223 * sys_sched_setparam - set/change the RT priority of a thread
4224 * @pid: the pid in question.
4225 * @param: structure containing the new RT priority.
4227 * Return: 0 on success. An error code otherwise.
4229 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4231 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4235 * sys_sched_setattr - same as above, but with extended sched_attr
4236 * @pid: the pid in question.
4237 * @uattr: structure containing the extended parameters.
4238 * @flags: for future extension.
4240 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4241 unsigned int, flags)
4243 struct sched_attr attr;
4244 struct task_struct *p;
4247 if (!uattr || pid < 0 || flags)
4250 retval = sched_copy_attr(uattr, &attr);
4254 if ((int)attr.sched_policy < 0)
4259 p = find_process_by_pid(pid);
4261 retval = sched_setattr(p, &attr);
4268 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4269 * @pid: the pid in question.
4271 * Return: On success, the policy of the thread. Otherwise, a negative error
4274 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4276 struct task_struct *p;
4284 p = find_process_by_pid(pid);
4286 retval = security_task_getscheduler(p);
4289 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4296 * sys_sched_getparam - get the RT priority of a thread
4297 * @pid: the pid in question.
4298 * @param: structure containing the RT priority.
4300 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4303 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4305 struct sched_param lp = { .sched_priority = 0 };
4306 struct task_struct *p;
4309 if (!param || pid < 0)
4313 p = find_process_by_pid(pid);
4318 retval = security_task_getscheduler(p);
4322 if (task_has_rt_policy(p))
4323 lp.sched_priority = p->rt_priority;
4327 * This one might sleep, we cannot do it with a spinlock held ...
4329 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4338 static int sched_read_attr(struct sched_attr __user *uattr,
4339 struct sched_attr *attr,
4344 if (!access_ok(VERIFY_WRITE, uattr, usize))
4348 * If we're handed a smaller struct than we know of,
4349 * ensure all the unknown bits are 0 - i.e. old
4350 * user-space does not get uncomplete information.
4352 if (usize < sizeof(*attr)) {
4353 unsigned char *addr;
4356 addr = (void *)attr + usize;
4357 end = (void *)attr + sizeof(*attr);
4359 for (; addr < end; addr++) {
4367 ret = copy_to_user(uattr, attr, attr->size);
4375 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4376 * @pid: the pid in question.
4377 * @uattr: structure containing the extended parameters.
4378 * @size: sizeof(attr) for fwd/bwd comp.
4379 * @flags: for future extension.
4381 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4382 unsigned int, size, unsigned int, flags)
4384 struct sched_attr attr = {
4385 .size = sizeof(struct sched_attr),
4387 struct task_struct *p;
4390 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4391 size < SCHED_ATTR_SIZE_VER0 || flags)
4395 p = find_process_by_pid(pid);
4400 retval = security_task_getscheduler(p);
4404 attr.sched_policy = p->policy;
4405 if (p->sched_reset_on_fork)
4406 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4407 if (task_has_dl_policy(p))
4408 __getparam_dl(p, &attr);
4409 else if (task_has_rt_policy(p))
4410 attr.sched_priority = p->rt_priority;
4412 attr.sched_nice = task_nice(p);
4416 retval = sched_read_attr(uattr, &attr, size);
4424 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4426 cpumask_var_t cpus_allowed, new_mask;
4427 struct task_struct *p;
4432 p = find_process_by_pid(pid);
4438 /* Prevent p going away */
4442 if (p->flags & PF_NO_SETAFFINITY) {
4446 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4450 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4452 goto out_free_cpus_allowed;
4455 if (!check_same_owner(p)) {
4457 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4459 goto out_free_new_mask;
4464 retval = security_task_setscheduler(p);
4466 goto out_free_new_mask;
4469 cpuset_cpus_allowed(p, cpus_allowed);
4470 cpumask_and(new_mask, in_mask, cpus_allowed);
4473 * Since bandwidth control happens on root_domain basis,
4474 * if admission test is enabled, we only admit -deadline
4475 * tasks allowed to run on all the CPUs in the task's
4479 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4481 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4484 goto out_free_new_mask;
4490 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4493 cpuset_cpus_allowed(p, cpus_allowed);
4494 if (!cpumask_subset(new_mask, cpus_allowed)) {
4496 * We must have raced with a concurrent cpuset
4497 * update. Just reset the cpus_allowed to the
4498 * cpuset's cpus_allowed
4500 cpumask_copy(new_mask, cpus_allowed);
4505 free_cpumask_var(new_mask);
4506 out_free_cpus_allowed:
4507 free_cpumask_var(cpus_allowed);
4513 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4514 struct cpumask *new_mask)
4516 if (len < cpumask_size())
4517 cpumask_clear(new_mask);
4518 else if (len > cpumask_size())
4519 len = cpumask_size();
4521 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4525 * sys_sched_setaffinity - set the cpu affinity of a process
4526 * @pid: pid of the process
4527 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4528 * @user_mask_ptr: user-space pointer to the new cpu mask
4530 * Return: 0 on success. An error code otherwise.
4532 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4533 unsigned long __user *, user_mask_ptr)
4535 cpumask_var_t new_mask;
4538 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4541 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4543 retval = sched_setaffinity(pid, new_mask);
4544 free_cpumask_var(new_mask);
4548 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4550 struct task_struct *p;
4551 unsigned long flags;
4557 p = find_process_by_pid(pid);
4561 retval = security_task_getscheduler(p);
4565 raw_spin_lock_irqsave(&p->pi_lock, flags);
4566 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4567 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4576 * sys_sched_getaffinity - get the cpu affinity of a process
4577 * @pid: pid of the process
4578 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4579 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4581 * Return: 0 on success. An error code otherwise.
4583 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4584 unsigned long __user *, user_mask_ptr)
4589 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4591 if (len & (sizeof(unsigned long)-1))
4594 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4597 ret = sched_getaffinity(pid, mask);
4599 size_t retlen = min_t(size_t, len, cpumask_size());
4601 if (copy_to_user(user_mask_ptr, mask, retlen))
4606 free_cpumask_var(mask);
4612 * sys_sched_yield - yield the current processor to other threads.
4614 * This function yields the current CPU to other tasks. If there are no
4615 * other threads running on this CPU then this function will return.
4619 SYSCALL_DEFINE0(sched_yield)
4621 struct rq *rq = this_rq_lock();
4623 schedstat_inc(rq, yld_count);
4624 current->sched_class->yield_task(rq);
4627 * Since we are going to call schedule() anyway, there's
4628 * no need to preempt or enable interrupts:
4630 __release(rq->lock);
4631 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4632 do_raw_spin_unlock(&rq->lock);
4633 sched_preempt_enable_no_resched();
4640 int __sched _cond_resched(void)
4642 if (should_resched(0)) {
4643 preempt_schedule_common();
4648 EXPORT_SYMBOL(_cond_resched);
4651 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4652 * call schedule, and on return reacquire the lock.
4654 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4655 * operations here to prevent schedule() from being called twice (once via
4656 * spin_unlock(), once by hand).
4658 int __cond_resched_lock(spinlock_t *lock)
4660 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4663 lockdep_assert_held(lock);
4665 if (spin_needbreak(lock) || resched) {
4668 preempt_schedule_common();
4676 EXPORT_SYMBOL(__cond_resched_lock);
4678 int __sched __cond_resched_softirq(void)
4680 BUG_ON(!in_softirq());
4682 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4684 preempt_schedule_common();
4690 EXPORT_SYMBOL(__cond_resched_softirq);
4693 * yield - yield the current processor to other threads.
4695 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4697 * The scheduler is at all times free to pick the calling task as the most
4698 * eligible task to run, if removing the yield() call from your code breaks
4699 * it, its already broken.
4701 * Typical broken usage is:
4706 * where one assumes that yield() will let 'the other' process run that will
4707 * make event true. If the current task is a SCHED_FIFO task that will never
4708 * happen. Never use yield() as a progress guarantee!!
4710 * If you want to use yield() to wait for something, use wait_event().
4711 * If you want to use yield() to be 'nice' for others, use cond_resched().
4712 * If you still want to use yield(), do not!
4714 void __sched yield(void)
4716 set_current_state(TASK_RUNNING);
4719 EXPORT_SYMBOL(yield);
4722 * yield_to - yield the current processor to another thread in
4723 * your thread group, or accelerate that thread toward the
4724 * processor it's on.
4726 * @preempt: whether task preemption is allowed or not
4728 * It's the caller's job to ensure that the target task struct
4729 * can't go away on us before we can do any checks.
4732 * true (>0) if we indeed boosted the target task.
4733 * false (0) if we failed to boost the target.
4734 * -ESRCH if there's no task to yield to.
4736 int __sched yield_to(struct task_struct *p, bool preempt)
4738 struct task_struct *curr = current;
4739 struct rq *rq, *p_rq;
4740 unsigned long flags;
4743 local_irq_save(flags);
4749 * If we're the only runnable task on the rq and target rq also
4750 * has only one task, there's absolutely no point in yielding.
4752 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4757 double_rq_lock(rq, p_rq);
4758 if (task_rq(p) != p_rq) {
4759 double_rq_unlock(rq, p_rq);
4763 if (!curr->sched_class->yield_to_task)
4766 if (curr->sched_class != p->sched_class)
4769 if (task_running(p_rq, p) || p->state)
4772 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4774 schedstat_inc(rq, yld_count);
4776 * Make p's CPU reschedule; pick_next_entity takes care of
4779 if (preempt && rq != p_rq)
4784 double_rq_unlock(rq, p_rq);
4786 local_irq_restore(flags);
4793 EXPORT_SYMBOL_GPL(yield_to);
4796 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4797 * that process accounting knows that this is a task in IO wait state.
4799 long __sched io_schedule_timeout(long timeout)
4801 int old_iowait = current->in_iowait;
4805 current->in_iowait = 1;
4806 blk_schedule_flush_plug(current);
4808 delayacct_blkio_start();
4810 atomic_inc(&rq->nr_iowait);
4811 ret = schedule_timeout(timeout);
4812 current->in_iowait = old_iowait;
4813 atomic_dec(&rq->nr_iowait);
4814 delayacct_blkio_end();
4818 EXPORT_SYMBOL(io_schedule_timeout);
4821 * sys_sched_get_priority_max - return maximum RT priority.
4822 * @policy: scheduling class.
4824 * Return: On success, this syscall returns the maximum
4825 * rt_priority that can be used by a given scheduling class.
4826 * On failure, a negative error code is returned.
4828 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4835 ret = MAX_USER_RT_PRIO-1;
4837 case SCHED_DEADLINE:
4848 * sys_sched_get_priority_min - return minimum RT priority.
4849 * @policy: scheduling class.
4851 * Return: On success, this syscall returns the minimum
4852 * rt_priority that can be used by a given scheduling class.
4853 * On failure, a negative error code is returned.
4855 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4864 case SCHED_DEADLINE:
4874 * sys_sched_rr_get_interval - return the default timeslice of a process.
4875 * @pid: pid of the process.
4876 * @interval: userspace pointer to the timeslice value.
4878 * this syscall writes the default timeslice value of a given process
4879 * into the user-space timespec buffer. A value of '0' means infinity.
4881 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4884 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4885 struct timespec __user *, interval)
4887 struct task_struct *p;
4888 unsigned int time_slice;
4889 unsigned long flags;
4899 p = find_process_by_pid(pid);
4903 retval = security_task_getscheduler(p);
4907 rq = task_rq_lock(p, &flags);
4909 if (p->sched_class->get_rr_interval)
4910 time_slice = p->sched_class->get_rr_interval(rq, p);
4911 task_rq_unlock(rq, p, &flags);
4914 jiffies_to_timespec(time_slice, &t);
4915 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4923 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4925 void sched_show_task(struct task_struct *p)
4927 unsigned long free = 0;
4929 unsigned long state = p->state;
4932 state = __ffs(state) + 1;
4933 printk(KERN_INFO "%-15.15s %c", p->comm,
4934 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4935 #if BITS_PER_LONG == 32
4936 if (state == TASK_RUNNING)
4937 printk(KERN_CONT " running ");
4939 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4941 if (state == TASK_RUNNING)
4942 printk(KERN_CONT " running task ");
4944 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4946 #ifdef CONFIG_DEBUG_STACK_USAGE
4947 free = stack_not_used(p);
4952 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4954 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4955 task_pid_nr(p), ppid,
4956 (unsigned long)task_thread_info(p)->flags);
4958 print_worker_info(KERN_INFO, p);
4959 show_stack(p, NULL);
4962 void show_state_filter(unsigned long state_filter)
4964 struct task_struct *g, *p;
4966 #if BITS_PER_LONG == 32
4968 " task PC stack pid father\n");
4971 " task PC stack pid father\n");
4974 for_each_process_thread(g, p) {
4976 * reset the NMI-timeout, listing all files on a slow
4977 * console might take a lot of time:
4978 * Also, reset softlockup watchdogs on all CPUs, because
4979 * another CPU might be blocked waiting for us to process
4982 touch_nmi_watchdog();
4983 touch_all_softlockup_watchdogs();
4984 if (!state_filter || (p->state & state_filter))
4988 #ifdef CONFIG_SCHED_DEBUG
4989 sysrq_sched_debug_show();
4993 * Only show locks if all tasks are dumped:
4996 debug_show_all_locks();
4999 void init_idle_bootup_task(struct task_struct *idle)
5001 idle->sched_class = &idle_sched_class;
5005 * init_idle - set up an idle thread for a given CPU
5006 * @idle: task in question
5007 * @cpu: cpu the idle task belongs to
5009 * NOTE: this function does not set the idle thread's NEED_RESCHED
5010 * flag, to make booting more robust.
5012 void init_idle(struct task_struct *idle, int cpu)
5014 struct rq *rq = cpu_rq(cpu);
5015 unsigned long flags;
5017 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5018 raw_spin_lock(&rq->lock);
5020 __sched_fork(0, idle);
5021 idle->state = TASK_RUNNING;
5022 idle->se.exec_start = sched_clock();
5026 * Its possible that init_idle() gets called multiple times on a task,
5027 * in that case do_set_cpus_allowed() will not do the right thing.
5029 * And since this is boot we can forgo the serialization.
5031 set_cpus_allowed_common(idle, cpumask_of(cpu));
5034 * We're having a chicken and egg problem, even though we are
5035 * holding rq->lock, the cpu isn't yet set to this cpu so the
5036 * lockdep check in task_group() will fail.
5038 * Similar case to sched_fork(). / Alternatively we could
5039 * use task_rq_lock() here and obtain the other rq->lock.
5044 __set_task_cpu(idle, cpu);
5047 rq->curr = rq->idle = idle;
5048 idle->on_rq = TASK_ON_RQ_QUEUED;
5052 raw_spin_unlock(&rq->lock);
5053 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5055 /* Set the preempt count _outside_ the spinlocks! */
5056 init_idle_preempt_count(idle, cpu);
5059 * The idle tasks have their own, simple scheduling class:
5061 idle->sched_class = &idle_sched_class;
5062 ftrace_graph_init_idle_task(idle, cpu);
5063 vtime_init_idle(idle, cpu);
5065 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5069 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5070 const struct cpumask *trial)
5072 int ret = 1, trial_cpus;
5073 struct dl_bw *cur_dl_b;
5074 unsigned long flags;
5076 if (!cpumask_weight(cur))
5079 rcu_read_lock_sched();
5080 cur_dl_b = dl_bw_of(cpumask_any(cur));
5081 trial_cpus = cpumask_weight(trial);
5083 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5084 if (cur_dl_b->bw != -1 &&
5085 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5087 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5088 rcu_read_unlock_sched();
5093 int task_can_attach(struct task_struct *p,
5094 const struct cpumask *cs_cpus_allowed)
5099 * Kthreads which disallow setaffinity shouldn't be moved
5100 * to a new cpuset; we don't want to change their cpu
5101 * affinity and isolating such threads by their set of
5102 * allowed nodes is unnecessary. Thus, cpusets are not
5103 * applicable for such threads. This prevents checking for
5104 * success of set_cpus_allowed_ptr() on all attached tasks
5105 * before cpus_allowed may be changed.
5107 if (p->flags & PF_NO_SETAFFINITY) {
5113 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5115 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5120 unsigned long flags;
5122 rcu_read_lock_sched();
5123 dl_b = dl_bw_of(dest_cpu);
5124 raw_spin_lock_irqsave(&dl_b->lock, flags);
5125 cpus = dl_bw_cpus(dest_cpu);
5126 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5131 * We reserve space for this task in the destination
5132 * root_domain, as we can't fail after this point.
5133 * We will free resources in the source root_domain
5134 * later on (see set_cpus_allowed_dl()).
5136 __dl_add(dl_b, p->dl.dl_bw);
5138 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5139 rcu_read_unlock_sched();
5149 #ifdef CONFIG_NUMA_BALANCING
5150 /* Migrate current task p to target_cpu */
5151 int migrate_task_to(struct task_struct *p, int target_cpu)
5153 struct migration_arg arg = { p, target_cpu };
5154 int curr_cpu = task_cpu(p);
5156 if (curr_cpu == target_cpu)
5159 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5162 /* TODO: This is not properly updating schedstats */
5164 trace_sched_move_numa(p, curr_cpu, target_cpu);
5165 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5169 * Requeue a task on a given node and accurately track the number of NUMA
5170 * tasks on the runqueues
5172 void sched_setnuma(struct task_struct *p, int nid)
5175 unsigned long flags;
5176 bool queued, running;
5178 rq = task_rq_lock(p, &flags);
5179 queued = task_on_rq_queued(p);
5180 running = task_current(rq, p);
5183 dequeue_task(rq, p, DEQUEUE_SAVE);
5185 put_prev_task(rq, p);
5187 p->numa_preferred_nid = nid;
5190 p->sched_class->set_curr_task(rq);
5192 enqueue_task(rq, p, ENQUEUE_RESTORE);
5193 task_rq_unlock(rq, p, &flags);
5195 #endif /* CONFIG_NUMA_BALANCING */
5197 #ifdef CONFIG_HOTPLUG_CPU
5199 * Ensures that the idle task is using init_mm right before its cpu goes
5202 void idle_task_exit(void)
5204 struct mm_struct *mm = current->active_mm;
5206 BUG_ON(cpu_online(smp_processor_id()));
5208 if (mm != &init_mm) {
5209 switch_mm_irqs_off(mm, &init_mm, current);
5210 finish_arch_post_lock_switch();
5216 * Since this CPU is going 'away' for a while, fold any nr_active delta
5217 * we might have. Assumes we're called after migrate_tasks() so that the
5218 * nr_active count is stable.
5220 * Also see the comment "Global load-average calculations".
5222 static void calc_load_migrate(struct rq *rq)
5224 long delta = calc_load_fold_active(rq);
5226 atomic_long_add(delta, &calc_load_tasks);
5229 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5233 static const struct sched_class fake_sched_class = {
5234 .put_prev_task = put_prev_task_fake,
5237 static struct task_struct fake_task = {
5239 * Avoid pull_{rt,dl}_task()
5241 .prio = MAX_PRIO + 1,
5242 .sched_class = &fake_sched_class,
5246 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5247 * try_to_wake_up()->select_task_rq().
5249 * Called with rq->lock held even though we'er in stop_machine() and
5250 * there's no concurrency possible, we hold the required locks anyway
5251 * because of lock validation efforts.
5253 static void migrate_tasks(struct rq *dead_rq)
5255 struct rq *rq = dead_rq;
5256 struct task_struct *next, *stop = rq->stop;
5260 * Fudge the rq selection such that the below task selection loop
5261 * doesn't get stuck on the currently eligible stop task.
5263 * We're currently inside stop_machine() and the rq is either stuck
5264 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5265 * either way we should never end up calling schedule() until we're
5271 * put_prev_task() and pick_next_task() sched
5272 * class method both need to have an up-to-date
5273 * value of rq->clock[_task]
5275 update_rq_clock(rq);
5279 * There's this thread running, bail when that's the only
5282 if (rq->nr_running == 1)
5286 * pick_next_task assumes pinned rq->lock.
5288 lockdep_pin_lock(&rq->lock);
5289 next = pick_next_task(rq, &fake_task);
5291 next->sched_class->put_prev_task(rq, next);
5294 * Rules for changing task_struct::cpus_allowed are holding
5295 * both pi_lock and rq->lock, such that holding either
5296 * stabilizes the mask.
5298 * Drop rq->lock is not quite as disastrous as it usually is
5299 * because !cpu_active at this point, which means load-balance
5300 * will not interfere. Also, stop-machine.
5302 lockdep_unpin_lock(&rq->lock);
5303 raw_spin_unlock(&rq->lock);
5304 raw_spin_lock(&next->pi_lock);
5305 raw_spin_lock(&rq->lock);
5308 * Since we're inside stop-machine, _nothing_ should have
5309 * changed the task, WARN if weird stuff happened, because in
5310 * that case the above rq->lock drop is a fail too.
5312 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5313 raw_spin_unlock(&next->pi_lock);
5317 /* Find suitable destination for @next, with force if needed. */
5318 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5320 rq = __migrate_task(rq, next, dest_cpu);
5321 if (rq != dead_rq) {
5322 raw_spin_unlock(&rq->lock);
5324 raw_spin_lock(&rq->lock);
5326 raw_spin_unlock(&next->pi_lock);
5331 #endif /* CONFIG_HOTPLUG_CPU */
5333 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5335 static struct ctl_table sd_ctl_dir[] = {
5337 .procname = "sched_domain",
5343 static struct ctl_table sd_ctl_root[] = {
5345 .procname = "kernel",
5347 .child = sd_ctl_dir,
5352 static struct ctl_table *sd_alloc_ctl_entry(int n)
5354 struct ctl_table *entry =
5355 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5360 static void sd_free_ctl_entry(struct ctl_table **tablep)
5362 struct ctl_table *entry;
5365 * In the intermediate directories, both the child directory and
5366 * procname are dynamically allocated and could fail but the mode
5367 * will always be set. In the lowest directory the names are
5368 * static strings and all have proc handlers.
5370 for (entry = *tablep; entry->mode; entry++) {
5372 sd_free_ctl_entry(&entry->child);
5373 if (entry->proc_handler == NULL)
5374 kfree(entry->procname);
5381 static int min_load_idx = 0;
5382 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5385 set_table_entry(struct ctl_table *entry,
5386 const char *procname, void *data, int maxlen,
5387 umode_t mode, proc_handler *proc_handler,
5390 entry->procname = procname;
5392 entry->maxlen = maxlen;
5394 entry->proc_handler = proc_handler;
5397 entry->extra1 = &min_load_idx;
5398 entry->extra2 = &max_load_idx;
5402 static struct ctl_table *
5403 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5405 struct ctl_table *table = sd_alloc_ctl_entry(14);
5410 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5411 sizeof(long), 0644, proc_doulongvec_minmax, false);
5412 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5413 sizeof(long), 0644, proc_doulongvec_minmax, false);
5414 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5415 sizeof(int), 0644, proc_dointvec_minmax, true);
5416 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5417 sizeof(int), 0644, proc_dointvec_minmax, true);
5418 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5419 sizeof(int), 0644, proc_dointvec_minmax, true);
5420 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5421 sizeof(int), 0644, proc_dointvec_minmax, true);
5422 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5423 sizeof(int), 0644, proc_dointvec_minmax, true);
5424 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5425 sizeof(int), 0644, proc_dointvec_minmax, false);
5426 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5427 sizeof(int), 0644, proc_dointvec_minmax, false);
5428 set_table_entry(&table[9], "cache_nice_tries",
5429 &sd->cache_nice_tries,
5430 sizeof(int), 0644, proc_dointvec_minmax, false);
5431 set_table_entry(&table[10], "flags", &sd->flags,
5432 sizeof(int), 0644, proc_dointvec_minmax, false);
5433 set_table_entry(&table[11], "max_newidle_lb_cost",
5434 &sd->max_newidle_lb_cost,
5435 sizeof(long), 0644, proc_doulongvec_minmax, false);
5436 set_table_entry(&table[12], "name", sd->name,
5437 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5438 /* &table[13] is terminator */
5443 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5445 struct ctl_table *entry, *table;
5446 struct sched_domain *sd;
5447 int domain_num = 0, i;
5450 for_each_domain(cpu, sd)
5452 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5457 for_each_domain(cpu, sd) {
5458 snprintf(buf, 32, "domain%d", i);
5459 entry->procname = kstrdup(buf, GFP_KERNEL);
5461 entry->child = sd_alloc_ctl_domain_table(sd);
5468 static struct ctl_table_header *sd_sysctl_header;
5469 static void register_sched_domain_sysctl(void)
5471 int i, cpu_num = num_possible_cpus();
5472 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5475 WARN_ON(sd_ctl_dir[0].child);
5476 sd_ctl_dir[0].child = entry;
5481 for_each_possible_cpu(i) {
5482 snprintf(buf, 32, "cpu%d", i);
5483 entry->procname = kstrdup(buf, GFP_KERNEL);
5485 entry->child = sd_alloc_ctl_cpu_table(i);
5489 WARN_ON(sd_sysctl_header);
5490 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5493 /* may be called multiple times per register */
5494 static void unregister_sched_domain_sysctl(void)
5496 unregister_sysctl_table(sd_sysctl_header);
5497 sd_sysctl_header = NULL;
5498 if (sd_ctl_dir[0].child)
5499 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5502 static void register_sched_domain_sysctl(void)
5505 static void unregister_sched_domain_sysctl(void)
5508 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5510 static void set_rq_online(struct rq *rq)
5513 const struct sched_class *class;
5515 cpumask_set_cpu(rq->cpu, rq->rd->online);
5518 for_each_class(class) {
5519 if (class->rq_online)
5520 class->rq_online(rq);
5525 static void set_rq_offline(struct rq *rq)
5528 const struct sched_class *class;
5530 for_each_class(class) {
5531 if (class->rq_offline)
5532 class->rq_offline(rq);
5535 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5541 * migration_call - callback that gets triggered when a CPU is added.
5542 * Here we can start up the necessary migration thread for the new CPU.
5545 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5547 int cpu = (long)hcpu;
5548 unsigned long flags;
5549 struct rq *rq = cpu_rq(cpu);
5551 switch (action & ~CPU_TASKS_FROZEN) {
5553 case CPU_UP_PREPARE:
5554 rq->calc_load_update = calc_load_update;
5558 /* Update our root-domain */
5559 raw_spin_lock_irqsave(&rq->lock, flags);
5561 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5565 raw_spin_unlock_irqrestore(&rq->lock, flags);
5568 #ifdef CONFIG_HOTPLUG_CPU
5570 sched_ttwu_pending();
5571 /* Update our root-domain */
5572 raw_spin_lock_irqsave(&rq->lock, flags);
5574 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5578 BUG_ON(rq->nr_running != 1); /* the migration thread */
5579 raw_spin_unlock_irqrestore(&rq->lock, flags);
5583 calc_load_migrate(rq);
5588 update_max_interval();
5594 * Register at high priority so that task migration (migrate_all_tasks)
5595 * happens before everything else. This has to be lower priority than
5596 * the notifier in the perf_event subsystem, though.
5598 static struct notifier_block migration_notifier = {
5599 .notifier_call = migration_call,
5600 .priority = CPU_PRI_MIGRATION,
5603 static void set_cpu_rq_start_time(void)
5605 int cpu = smp_processor_id();
5606 struct rq *rq = cpu_rq(cpu);
5607 rq->age_stamp = sched_clock_cpu(cpu);
5610 static int sched_cpu_active(struct notifier_block *nfb,
5611 unsigned long action, void *hcpu)
5613 int cpu = (long)hcpu;
5615 switch (action & ~CPU_TASKS_FROZEN) {
5617 set_cpu_rq_start_time();
5622 * At this point a starting CPU has marked itself as online via
5623 * set_cpu_online(). But it might not yet have marked itself
5624 * as active, which is essential from here on.
5626 set_cpu_active(cpu, true);
5627 stop_machine_unpark(cpu);
5630 case CPU_DOWN_FAILED:
5631 set_cpu_active(cpu, true);
5639 static int sched_cpu_inactive(struct notifier_block *nfb,
5640 unsigned long action, void *hcpu)
5642 switch (action & ~CPU_TASKS_FROZEN) {
5643 case CPU_DOWN_PREPARE:
5644 set_cpu_active((long)hcpu, false);
5651 static int __init migration_init(void)
5653 void *cpu = (void *)(long)smp_processor_id();
5656 /* Initialize migration for the boot CPU */
5657 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5658 BUG_ON(err == NOTIFY_BAD);
5659 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5660 register_cpu_notifier(&migration_notifier);
5662 /* Register cpu active notifiers */
5663 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5664 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5668 early_initcall(migration_init);
5670 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5672 #ifdef CONFIG_SCHED_DEBUG
5674 static __read_mostly int sched_debug_enabled;
5676 static int __init sched_debug_setup(char *str)
5678 sched_debug_enabled = 1;
5682 early_param("sched_debug", sched_debug_setup);
5684 static inline bool sched_debug(void)
5686 return sched_debug_enabled;
5689 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5690 struct cpumask *groupmask)
5692 struct sched_group *group = sd->groups;
5694 cpumask_clear(groupmask);
5696 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5698 if (!(sd->flags & SD_LOAD_BALANCE)) {
5699 printk("does not load-balance\n");
5701 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5706 printk(KERN_CONT "span %*pbl level %s\n",
5707 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5709 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5710 printk(KERN_ERR "ERROR: domain->span does not contain "
5713 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5714 printk(KERN_ERR "ERROR: domain->groups does not contain"
5718 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5722 printk(KERN_ERR "ERROR: group is NULL\n");
5726 if (!cpumask_weight(sched_group_cpus(group))) {
5727 printk(KERN_CONT "\n");
5728 printk(KERN_ERR "ERROR: empty group\n");
5732 if (!(sd->flags & SD_OVERLAP) &&
5733 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5734 printk(KERN_CONT "\n");
5735 printk(KERN_ERR "ERROR: repeated CPUs\n");
5739 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5741 printk(KERN_CONT " %*pbl",
5742 cpumask_pr_args(sched_group_cpus(group)));
5743 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5744 printk(KERN_CONT " (cpu_capacity = %d)",
5745 group->sgc->capacity);
5748 group = group->next;
5749 } while (group != sd->groups);
5750 printk(KERN_CONT "\n");
5752 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5753 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5756 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5757 printk(KERN_ERR "ERROR: parent span is not a superset "
5758 "of domain->span\n");
5762 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5766 if (!sched_debug_enabled)
5770 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5774 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5777 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5785 #else /* !CONFIG_SCHED_DEBUG */
5786 # define sched_domain_debug(sd, cpu) do { } while (0)
5787 static inline bool sched_debug(void)
5791 #endif /* CONFIG_SCHED_DEBUG */
5793 static int sd_degenerate(struct sched_domain *sd)
5795 if (cpumask_weight(sched_domain_span(sd)) == 1)
5798 /* Following flags need at least 2 groups */
5799 if (sd->flags & (SD_LOAD_BALANCE |
5800 SD_BALANCE_NEWIDLE |
5803 SD_SHARE_CPUCAPACITY |
5804 SD_SHARE_PKG_RESOURCES |
5805 SD_SHARE_POWERDOMAIN)) {
5806 if (sd->groups != sd->groups->next)
5810 /* Following flags don't use groups */
5811 if (sd->flags & (SD_WAKE_AFFINE))
5818 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5820 unsigned long cflags = sd->flags, pflags = parent->flags;
5822 if (sd_degenerate(parent))
5825 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5828 /* Flags needing groups don't count if only 1 group in parent */
5829 if (parent->groups == parent->groups->next) {
5830 pflags &= ~(SD_LOAD_BALANCE |
5831 SD_BALANCE_NEWIDLE |
5834 SD_SHARE_CPUCAPACITY |
5835 SD_SHARE_PKG_RESOURCES |
5837 SD_SHARE_POWERDOMAIN);
5838 if (nr_node_ids == 1)
5839 pflags &= ~SD_SERIALIZE;
5841 if (~cflags & pflags)
5847 static void free_rootdomain(struct rcu_head *rcu)
5849 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5851 cpupri_cleanup(&rd->cpupri);
5852 cpudl_cleanup(&rd->cpudl);
5853 free_cpumask_var(rd->dlo_mask);
5854 free_cpumask_var(rd->rto_mask);
5855 free_cpumask_var(rd->online);
5856 free_cpumask_var(rd->span);
5860 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5862 struct root_domain *old_rd = NULL;
5863 unsigned long flags;
5865 raw_spin_lock_irqsave(&rq->lock, flags);
5870 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5873 cpumask_clear_cpu(rq->cpu, old_rd->span);
5876 * If we dont want to free the old_rd yet then
5877 * set old_rd to NULL to skip the freeing later
5880 if (!atomic_dec_and_test(&old_rd->refcount))
5884 atomic_inc(&rd->refcount);
5887 cpumask_set_cpu(rq->cpu, rd->span);
5888 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5891 raw_spin_unlock_irqrestore(&rq->lock, flags);
5894 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5897 static int init_rootdomain(struct root_domain *rd)
5899 memset(rd, 0, sizeof(*rd));
5901 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
5903 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
5905 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5907 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5910 #ifdef HAVE_RT_PUSH_IPI
5912 raw_spin_lock_init(&rd->rto_lock);
5913 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
5916 init_dl_bw(&rd->dl_bw);
5917 if (cpudl_init(&rd->cpudl) != 0)
5920 if (cpupri_init(&rd->cpupri) != 0)
5925 free_cpumask_var(rd->rto_mask);
5927 free_cpumask_var(rd->dlo_mask);
5929 free_cpumask_var(rd->online);
5931 free_cpumask_var(rd->span);
5937 * By default the system creates a single root-domain with all cpus as
5938 * members (mimicking the global state we have today).
5940 struct root_domain def_root_domain;
5942 static void init_defrootdomain(void)
5944 init_rootdomain(&def_root_domain);
5946 atomic_set(&def_root_domain.refcount, 1);
5949 static struct root_domain *alloc_rootdomain(void)
5951 struct root_domain *rd;
5953 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5957 if (init_rootdomain(rd) != 0) {
5965 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5967 struct sched_group *tmp, *first;
5976 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5981 } while (sg != first);
5984 static void free_sched_domain(struct rcu_head *rcu)
5986 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5989 * If its an overlapping domain it has private groups, iterate and
5992 if (sd->flags & SD_OVERLAP) {
5993 free_sched_groups(sd->groups, 1);
5994 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5995 kfree(sd->groups->sgc);
6001 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6003 call_rcu(&sd->rcu, free_sched_domain);
6006 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6008 for (; sd; sd = sd->parent)
6009 destroy_sched_domain(sd, cpu);
6013 * Keep a special pointer to the highest sched_domain that has
6014 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6015 * allows us to avoid some pointer chasing select_idle_sibling().
6017 * Also keep a unique ID per domain (we use the first cpu number in
6018 * the cpumask of the domain), this allows us to quickly tell if
6019 * two cpus are in the same cache domain, see cpus_share_cache().
6021 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6022 DEFINE_PER_CPU(int, sd_llc_size);
6023 DEFINE_PER_CPU(int, sd_llc_id);
6024 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6025 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6026 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6028 static void update_top_cache_domain(int cpu)
6030 struct sched_domain *sd;
6031 struct sched_domain *busy_sd = NULL;
6035 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6037 id = cpumask_first(sched_domain_span(sd));
6038 size = cpumask_weight(sched_domain_span(sd));
6039 busy_sd = sd->parent; /* sd_busy */
6041 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6043 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6044 per_cpu(sd_llc_size, cpu) = size;
6045 per_cpu(sd_llc_id, cpu) = id;
6047 sd = lowest_flag_domain(cpu, SD_NUMA);
6048 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6050 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6051 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6055 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6056 * hold the hotplug lock.
6059 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6061 struct rq *rq = cpu_rq(cpu);
6062 struct sched_domain *tmp;
6064 /* Remove the sched domains which do not contribute to scheduling. */
6065 for (tmp = sd; tmp; ) {
6066 struct sched_domain *parent = tmp->parent;
6070 if (sd_parent_degenerate(tmp, parent)) {
6071 tmp->parent = parent->parent;
6073 parent->parent->child = tmp;
6075 * Transfer SD_PREFER_SIBLING down in case of a
6076 * degenerate parent; the spans match for this
6077 * so the property transfers.
6079 if (parent->flags & SD_PREFER_SIBLING)
6080 tmp->flags |= SD_PREFER_SIBLING;
6081 destroy_sched_domain(parent, cpu);
6086 if (sd && sd_degenerate(sd)) {
6089 destroy_sched_domain(tmp, cpu);
6094 sched_domain_debug(sd, cpu);
6096 rq_attach_root(rq, rd);
6098 rcu_assign_pointer(rq->sd, sd);
6099 destroy_sched_domains(tmp, cpu);
6101 update_top_cache_domain(cpu);
6104 /* Setup the mask of cpus configured for isolated domains */
6105 static int __init isolated_cpu_setup(char *str)
6107 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6108 cpulist_parse(str, cpu_isolated_map);
6112 __setup("isolcpus=", isolated_cpu_setup);
6115 struct sched_domain ** __percpu sd;
6116 struct root_domain *rd;
6127 * Build an iteration mask that can exclude certain CPUs from the upwards
6130 * Only CPUs that can arrive at this group should be considered to continue
6133 * Asymmetric node setups can result in situations where the domain tree is of
6134 * unequal depth, make sure to skip domains that already cover the entire
6137 * In that case build_sched_domains() will have terminated the iteration early
6138 * and our sibling sd spans will be empty. Domains should always include the
6139 * cpu they're built on, so check that.
6142 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6144 const struct cpumask *sg_span = sched_group_cpus(sg);
6145 struct sd_data *sdd = sd->private;
6146 struct sched_domain *sibling;
6149 for_each_cpu(i, sg_span) {
6150 sibling = *per_cpu_ptr(sdd->sd, i);
6153 * Can happen in the asymmetric case, where these siblings are
6154 * unused. The mask will not be empty because those CPUs that
6155 * do have the top domain _should_ span the domain.
6157 if (!sibling->child)
6160 /* If we would not end up here, we can't continue from here */
6161 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
6164 cpumask_set_cpu(i, sched_group_mask(sg));
6167 /* We must not have empty masks here */
6168 WARN_ON_ONCE(cpumask_empty(sched_group_mask(sg)));
6172 * Return the canonical balance cpu for this group, this is the first cpu
6173 * of this group that's also in the iteration mask.
6175 int group_balance_cpu(struct sched_group *sg)
6177 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6181 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6183 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6184 const struct cpumask *span = sched_domain_span(sd);
6185 struct cpumask *covered = sched_domains_tmpmask;
6186 struct sd_data *sdd = sd->private;
6187 struct sched_domain *sibling;
6190 cpumask_clear(covered);
6192 for_each_cpu(i, span) {
6193 struct cpumask *sg_span;
6195 if (cpumask_test_cpu(i, covered))
6198 sibling = *per_cpu_ptr(sdd->sd, i);
6200 /* See the comment near build_group_mask(). */
6201 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6204 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6205 GFP_KERNEL, cpu_to_node(cpu));
6210 sg_span = sched_group_cpus(sg);
6212 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6214 cpumask_set_cpu(i, sg_span);
6216 cpumask_or(covered, covered, sg_span);
6218 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6219 if (atomic_inc_return(&sg->sgc->ref) == 1)
6220 build_group_mask(sd, sg);
6223 * Initialize sgc->capacity such that even if we mess up the
6224 * domains and no possible iteration will get us here, we won't
6227 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6230 * Make sure the first group of this domain contains the
6231 * canonical balance cpu. Otherwise the sched_domain iteration
6232 * breaks. See update_sg_lb_stats().
6234 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6235 group_balance_cpu(sg) == cpu)
6245 sd->groups = groups;
6250 free_sched_groups(first, 0);
6255 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6257 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6258 struct sched_domain *child = sd->child;
6261 cpu = cpumask_first(sched_domain_span(child));
6264 *sg = *per_cpu_ptr(sdd->sg, cpu);
6265 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6266 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6273 * build_sched_groups will build a circular linked list of the groups
6274 * covered by the given span, and will set each group's ->cpumask correctly,
6275 * and ->cpu_capacity to 0.
6277 * Assumes the sched_domain tree is fully constructed
6280 build_sched_groups(struct sched_domain *sd, int cpu)
6282 struct sched_group *first = NULL, *last = NULL;
6283 struct sd_data *sdd = sd->private;
6284 const struct cpumask *span = sched_domain_span(sd);
6285 struct cpumask *covered;
6288 get_group(cpu, sdd, &sd->groups);
6289 atomic_inc(&sd->groups->ref);
6291 if (cpu != cpumask_first(span))
6294 lockdep_assert_held(&sched_domains_mutex);
6295 covered = sched_domains_tmpmask;
6297 cpumask_clear(covered);
6299 for_each_cpu(i, span) {
6300 struct sched_group *sg;
6303 if (cpumask_test_cpu(i, covered))
6306 group = get_group(i, sdd, &sg);
6307 cpumask_setall(sched_group_mask(sg));
6309 for_each_cpu(j, span) {
6310 if (get_group(j, sdd, NULL) != group)
6313 cpumask_set_cpu(j, covered);
6314 cpumask_set_cpu(j, sched_group_cpus(sg));
6329 * Initialize sched groups cpu_capacity.
6331 * cpu_capacity indicates the capacity of sched group, which is used while
6332 * distributing the load between different sched groups in a sched domain.
6333 * Typically cpu_capacity for all the groups in a sched domain will be same
6334 * unless there are asymmetries in the topology. If there are asymmetries,
6335 * group having more cpu_capacity will pickup more load compared to the
6336 * group having less cpu_capacity.
6338 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6340 struct sched_group *sg = sd->groups;
6345 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6347 } while (sg != sd->groups);
6349 if (cpu != group_balance_cpu(sg))
6352 update_group_capacity(sd, cpu);
6353 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6357 * Initializers for schedule domains
6358 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6361 static int default_relax_domain_level = -1;
6362 int sched_domain_level_max;
6364 static int __init setup_relax_domain_level(char *str)
6366 if (kstrtoint(str, 0, &default_relax_domain_level))
6367 pr_warn("Unable to set relax_domain_level\n");
6371 __setup("relax_domain_level=", setup_relax_domain_level);
6373 static void set_domain_attribute(struct sched_domain *sd,
6374 struct sched_domain_attr *attr)
6378 if (!attr || attr->relax_domain_level < 0) {
6379 if (default_relax_domain_level < 0)
6382 request = default_relax_domain_level;
6384 request = attr->relax_domain_level;
6385 if (request < sd->level) {
6386 /* turn off idle balance on this domain */
6387 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6389 /* turn on idle balance on this domain */
6390 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6394 static void __sdt_free(const struct cpumask *cpu_map);
6395 static int __sdt_alloc(const struct cpumask *cpu_map);
6397 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6398 const struct cpumask *cpu_map)
6402 if (!atomic_read(&d->rd->refcount))
6403 free_rootdomain(&d->rd->rcu); /* fall through */
6405 free_percpu(d->sd); /* fall through */
6407 __sdt_free(cpu_map); /* fall through */
6413 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6414 const struct cpumask *cpu_map)
6416 memset(d, 0, sizeof(*d));
6418 if (__sdt_alloc(cpu_map))
6419 return sa_sd_storage;
6420 d->sd = alloc_percpu(struct sched_domain *);
6422 return sa_sd_storage;
6423 d->rd = alloc_rootdomain();
6426 return sa_rootdomain;
6430 * NULL the sd_data elements we've used to build the sched_domain and
6431 * sched_group structure so that the subsequent __free_domain_allocs()
6432 * will not free the data we're using.
6434 static void claim_allocations(int cpu, struct sched_domain *sd)
6436 struct sd_data *sdd = sd->private;
6438 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6439 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6441 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6442 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6444 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6445 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6449 static int sched_domains_numa_levels;
6450 enum numa_topology_type sched_numa_topology_type;
6451 static int *sched_domains_numa_distance;
6452 int sched_max_numa_distance;
6453 static struct cpumask ***sched_domains_numa_masks;
6454 static int sched_domains_curr_level;
6458 * SD_flags allowed in topology descriptions.
6460 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6461 * SD_SHARE_PKG_RESOURCES - describes shared caches
6462 * SD_NUMA - describes NUMA topologies
6463 * SD_SHARE_POWERDOMAIN - describes shared power domain
6466 * SD_ASYM_PACKING - describes SMT quirks
6468 #define TOPOLOGY_SD_FLAGS \
6469 (SD_SHARE_CPUCAPACITY | \
6470 SD_SHARE_PKG_RESOURCES | \
6473 SD_SHARE_POWERDOMAIN)
6475 static struct sched_domain *
6476 sd_init(struct sched_domain_topology_level *tl, int cpu)
6478 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6479 int sd_weight, sd_flags = 0;
6483 * Ugly hack to pass state to sd_numa_mask()...
6485 sched_domains_curr_level = tl->numa_level;
6488 sd_weight = cpumask_weight(tl->mask(cpu));
6491 sd_flags = (*tl->sd_flags)();
6492 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6493 "wrong sd_flags in topology description\n"))
6494 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6496 *sd = (struct sched_domain){
6497 .min_interval = sd_weight,
6498 .max_interval = 2*sd_weight,
6500 .imbalance_pct = 125,
6502 .cache_nice_tries = 0,
6509 .flags = 1*SD_LOAD_BALANCE
6510 | 1*SD_BALANCE_NEWIDLE
6515 | 0*SD_SHARE_CPUCAPACITY
6516 | 0*SD_SHARE_PKG_RESOURCES
6518 | 0*SD_PREFER_SIBLING
6523 .last_balance = jiffies,
6524 .balance_interval = sd_weight,
6526 .max_newidle_lb_cost = 0,
6527 .next_decay_max_lb_cost = jiffies,
6528 #ifdef CONFIG_SCHED_DEBUG
6534 * Convert topological properties into behaviour.
6537 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6538 sd->flags |= SD_PREFER_SIBLING;
6539 sd->imbalance_pct = 110;
6540 sd->smt_gain = 1178; /* ~15% */
6542 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6543 sd->imbalance_pct = 117;
6544 sd->cache_nice_tries = 1;
6548 } else if (sd->flags & SD_NUMA) {
6549 sd->cache_nice_tries = 2;
6553 sd->flags |= SD_SERIALIZE;
6554 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6555 sd->flags &= ~(SD_BALANCE_EXEC |
6562 sd->flags |= SD_PREFER_SIBLING;
6563 sd->cache_nice_tries = 1;
6568 sd->private = &tl->data;
6574 * Topology list, bottom-up.
6576 static struct sched_domain_topology_level default_topology[] = {
6577 #ifdef CONFIG_SCHED_SMT
6578 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6580 #ifdef CONFIG_SCHED_MC
6581 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6583 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6587 static struct sched_domain_topology_level *sched_domain_topology =
6590 #define for_each_sd_topology(tl) \
6591 for (tl = sched_domain_topology; tl->mask; tl++)
6593 void set_sched_topology(struct sched_domain_topology_level *tl)
6595 sched_domain_topology = tl;
6600 static const struct cpumask *sd_numa_mask(int cpu)
6602 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6605 static void sched_numa_warn(const char *str)
6607 static int done = false;
6615 printk(KERN_WARNING "ERROR: %s\n\n", str);
6617 for (i = 0; i < nr_node_ids; i++) {
6618 printk(KERN_WARNING " ");
6619 for (j = 0; j < nr_node_ids; j++)
6620 printk(KERN_CONT "%02d ", node_distance(i,j));
6621 printk(KERN_CONT "\n");
6623 printk(KERN_WARNING "\n");
6626 bool find_numa_distance(int distance)
6630 if (distance == node_distance(0, 0))
6633 for (i = 0; i < sched_domains_numa_levels; i++) {
6634 if (sched_domains_numa_distance[i] == distance)
6642 * A system can have three types of NUMA topology:
6643 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6644 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6645 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6647 * The difference between a glueless mesh topology and a backplane
6648 * topology lies in whether communication between not directly
6649 * connected nodes goes through intermediary nodes (where programs
6650 * could run), or through backplane controllers. This affects
6651 * placement of programs.
6653 * The type of topology can be discerned with the following tests:
6654 * - If the maximum distance between any nodes is 1 hop, the system
6655 * is directly connected.
6656 * - If for two nodes A and B, located N > 1 hops away from each other,
6657 * there is an intermediary node C, which is < N hops away from both
6658 * nodes A and B, the system is a glueless mesh.
6660 static void init_numa_topology_type(void)
6664 n = sched_max_numa_distance;
6666 if (sched_domains_numa_levels <= 1) {
6667 sched_numa_topology_type = NUMA_DIRECT;
6671 for_each_online_node(a) {
6672 for_each_online_node(b) {
6673 /* Find two nodes furthest removed from each other. */
6674 if (node_distance(a, b) < n)
6677 /* Is there an intermediary node between a and b? */
6678 for_each_online_node(c) {
6679 if (node_distance(a, c) < n &&
6680 node_distance(b, c) < n) {
6681 sched_numa_topology_type =
6687 sched_numa_topology_type = NUMA_BACKPLANE;
6693 static void sched_init_numa(void)
6695 int next_distance, curr_distance = node_distance(0, 0);
6696 struct sched_domain_topology_level *tl;
6700 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6701 if (!sched_domains_numa_distance)
6705 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6706 * unique distances in the node_distance() table.
6708 * Assumes node_distance(0,j) includes all distances in
6709 * node_distance(i,j) in order to avoid cubic time.
6711 next_distance = curr_distance;
6712 for (i = 0; i < nr_node_ids; i++) {
6713 for (j = 0; j < nr_node_ids; j++) {
6714 for (k = 0; k < nr_node_ids; k++) {
6715 int distance = node_distance(i, k);
6717 if (distance > curr_distance &&
6718 (distance < next_distance ||
6719 next_distance == curr_distance))
6720 next_distance = distance;
6723 * While not a strong assumption it would be nice to know
6724 * about cases where if node A is connected to B, B is not
6725 * equally connected to A.
6727 if (sched_debug() && node_distance(k, i) != distance)
6728 sched_numa_warn("Node-distance not symmetric");
6730 if (sched_debug() && i && !find_numa_distance(distance))
6731 sched_numa_warn("Node-0 not representative");
6733 if (next_distance != curr_distance) {
6734 sched_domains_numa_distance[level++] = next_distance;
6735 sched_domains_numa_levels = level;
6736 curr_distance = next_distance;
6741 * In case of sched_debug() we verify the above assumption.
6751 * 'level' contains the number of unique distances, excluding the
6752 * identity distance node_distance(i,i).
6754 * The sched_domains_numa_distance[] array includes the actual distance
6759 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6760 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6761 * the array will contain less then 'level' members. This could be
6762 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6763 * in other functions.
6765 * We reset it to 'level' at the end of this function.
6767 sched_domains_numa_levels = 0;
6769 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6770 if (!sched_domains_numa_masks)
6774 * Now for each level, construct a mask per node which contains all
6775 * cpus of nodes that are that many hops away from us.
6777 for (i = 0; i < level; i++) {
6778 sched_domains_numa_masks[i] =
6779 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6780 if (!sched_domains_numa_masks[i])
6783 for (j = 0; j < nr_node_ids; j++) {
6784 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6788 sched_domains_numa_masks[i][j] = mask;
6791 if (node_distance(j, k) > sched_domains_numa_distance[i])
6794 cpumask_or(mask, mask, cpumask_of_node(k));
6799 /* Compute default topology size */
6800 for (i = 0; sched_domain_topology[i].mask; i++);
6802 tl = kzalloc((i + level + 1) *
6803 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6808 * Copy the default topology bits..
6810 for (i = 0; sched_domain_topology[i].mask; i++)
6811 tl[i] = sched_domain_topology[i];
6814 * .. and append 'j' levels of NUMA goodness.
6816 for (j = 0; j < level; i++, j++) {
6817 tl[i] = (struct sched_domain_topology_level){
6818 .mask = sd_numa_mask,
6819 .sd_flags = cpu_numa_flags,
6820 .flags = SDTL_OVERLAP,
6826 sched_domain_topology = tl;
6828 sched_domains_numa_levels = level;
6829 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6831 init_numa_topology_type();
6834 static void sched_domains_numa_masks_set(int cpu)
6837 int node = cpu_to_node(cpu);
6839 for (i = 0; i < sched_domains_numa_levels; i++) {
6840 for (j = 0; j < nr_node_ids; j++) {
6841 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6842 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6847 static void sched_domains_numa_masks_clear(int cpu)
6850 for (i = 0; i < sched_domains_numa_levels; i++) {
6851 for (j = 0; j < nr_node_ids; j++)
6852 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6857 * Update sched_domains_numa_masks[level][node] array when new cpus
6860 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6861 unsigned long action,
6864 int cpu = (long)hcpu;
6866 switch (action & ~CPU_TASKS_FROZEN) {
6868 sched_domains_numa_masks_set(cpu);
6872 sched_domains_numa_masks_clear(cpu);
6882 static inline void sched_init_numa(void)
6886 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6887 unsigned long action,
6892 #endif /* CONFIG_NUMA */
6894 static int __sdt_alloc(const struct cpumask *cpu_map)
6896 struct sched_domain_topology_level *tl;
6899 for_each_sd_topology(tl) {
6900 struct sd_data *sdd = &tl->data;
6902 sdd->sd = alloc_percpu(struct sched_domain *);
6906 sdd->sg = alloc_percpu(struct sched_group *);
6910 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6914 for_each_cpu(j, cpu_map) {
6915 struct sched_domain *sd;
6916 struct sched_group *sg;
6917 struct sched_group_capacity *sgc;
6919 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6920 GFP_KERNEL, cpu_to_node(j));
6924 *per_cpu_ptr(sdd->sd, j) = sd;
6926 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6927 GFP_KERNEL, cpu_to_node(j));
6933 *per_cpu_ptr(sdd->sg, j) = sg;
6935 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6936 GFP_KERNEL, cpu_to_node(j));
6940 *per_cpu_ptr(sdd->sgc, j) = sgc;
6947 static void __sdt_free(const struct cpumask *cpu_map)
6949 struct sched_domain_topology_level *tl;
6952 for_each_sd_topology(tl) {
6953 struct sd_data *sdd = &tl->data;
6955 for_each_cpu(j, cpu_map) {
6956 struct sched_domain *sd;
6959 sd = *per_cpu_ptr(sdd->sd, j);
6960 if (sd && (sd->flags & SD_OVERLAP))
6961 free_sched_groups(sd->groups, 0);
6962 kfree(*per_cpu_ptr(sdd->sd, j));
6966 kfree(*per_cpu_ptr(sdd->sg, j));
6968 kfree(*per_cpu_ptr(sdd->sgc, j));
6970 free_percpu(sdd->sd);
6972 free_percpu(sdd->sg);
6974 free_percpu(sdd->sgc);
6979 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6980 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6981 struct sched_domain *child, int cpu)
6983 struct sched_domain *sd = sd_init(tl, cpu);
6987 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6989 sd->level = child->level + 1;
6990 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6994 if (!cpumask_subset(sched_domain_span(child),
6995 sched_domain_span(sd))) {
6996 pr_err("BUG: arch topology borken\n");
6997 #ifdef CONFIG_SCHED_DEBUG
6998 pr_err(" the %s domain not a subset of the %s domain\n",
6999 child->name, sd->name);
7001 /* Fixup, ensure @sd has at least @child cpus. */
7002 cpumask_or(sched_domain_span(sd),
7003 sched_domain_span(sd),
7004 sched_domain_span(child));
7008 set_domain_attribute(sd, attr);
7014 * Build sched domains for a given set of cpus and attach the sched domains
7015 * to the individual cpus
7017 static int build_sched_domains(const struct cpumask *cpu_map,
7018 struct sched_domain_attr *attr)
7020 enum s_alloc alloc_state;
7021 struct sched_domain *sd;
7023 int i, ret = -ENOMEM;
7025 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7026 if (alloc_state != sa_rootdomain)
7029 /* Set up domains for cpus specified by the cpu_map. */
7030 for_each_cpu(i, cpu_map) {
7031 struct sched_domain_topology_level *tl;
7034 for_each_sd_topology(tl) {
7035 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7036 if (tl == sched_domain_topology)
7037 *per_cpu_ptr(d.sd, i) = sd;
7038 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7039 sd->flags |= SD_OVERLAP;
7040 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7045 /* Build the groups for the domains */
7046 for_each_cpu(i, cpu_map) {
7047 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7048 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7049 if (sd->flags & SD_OVERLAP) {
7050 if (build_overlap_sched_groups(sd, i))
7053 if (build_sched_groups(sd, i))
7059 /* Calculate CPU capacity for physical packages and nodes */
7060 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7061 if (!cpumask_test_cpu(i, cpu_map))
7064 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7065 claim_allocations(i, sd);
7066 init_sched_groups_capacity(i, sd);
7070 /* Attach the domains */
7072 for_each_cpu(i, cpu_map) {
7073 sd = *per_cpu_ptr(d.sd, i);
7074 cpu_attach_domain(sd, d.rd, i);
7080 __free_domain_allocs(&d, alloc_state, cpu_map);
7084 static cpumask_var_t *doms_cur; /* current sched domains */
7085 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7086 static struct sched_domain_attr *dattr_cur;
7087 /* attribues of custom domains in 'doms_cur' */
7090 * Special case: If a kmalloc of a doms_cur partition (array of
7091 * cpumask) fails, then fallback to a single sched domain,
7092 * as determined by the single cpumask fallback_doms.
7094 static cpumask_var_t fallback_doms;
7097 * arch_update_cpu_topology lets virtualized architectures update the
7098 * cpu core maps. It is supposed to return 1 if the topology changed
7099 * or 0 if it stayed the same.
7101 int __weak arch_update_cpu_topology(void)
7106 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7109 cpumask_var_t *doms;
7111 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7114 for (i = 0; i < ndoms; i++) {
7115 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7116 free_sched_domains(doms, i);
7123 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7126 for (i = 0; i < ndoms; i++)
7127 free_cpumask_var(doms[i]);
7132 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7133 * For now this just excludes isolated cpus, but could be used to
7134 * exclude other special cases in the future.
7136 static int init_sched_domains(const struct cpumask *cpu_map)
7140 arch_update_cpu_topology();
7142 doms_cur = alloc_sched_domains(ndoms_cur);
7144 doms_cur = &fallback_doms;
7145 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7146 err = build_sched_domains(doms_cur[0], NULL);
7147 register_sched_domain_sysctl();
7153 * Detach sched domains from a group of cpus specified in cpu_map
7154 * These cpus will now be attached to the NULL domain
7156 static void detach_destroy_domains(const struct cpumask *cpu_map)
7161 for_each_cpu(i, cpu_map)
7162 cpu_attach_domain(NULL, &def_root_domain, i);
7166 /* handle null as "default" */
7167 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7168 struct sched_domain_attr *new, int idx_new)
7170 struct sched_domain_attr tmp;
7177 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7178 new ? (new + idx_new) : &tmp,
7179 sizeof(struct sched_domain_attr));
7183 * Partition sched domains as specified by the 'ndoms_new'
7184 * cpumasks in the array doms_new[] of cpumasks. This compares
7185 * doms_new[] to the current sched domain partitioning, doms_cur[].
7186 * It destroys each deleted domain and builds each new domain.
7188 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7189 * The masks don't intersect (don't overlap.) We should setup one
7190 * sched domain for each mask. CPUs not in any of the cpumasks will
7191 * not be load balanced. If the same cpumask appears both in the
7192 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7195 * The passed in 'doms_new' should be allocated using
7196 * alloc_sched_domains. This routine takes ownership of it and will
7197 * free_sched_domains it when done with it. If the caller failed the
7198 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7199 * and partition_sched_domains() will fallback to the single partition
7200 * 'fallback_doms', it also forces the domains to be rebuilt.
7202 * If doms_new == NULL it will be replaced with cpu_online_mask.
7203 * ndoms_new == 0 is a special case for destroying existing domains,
7204 * and it will not create the default domain.
7206 * Call with hotplug lock held
7208 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7209 struct sched_domain_attr *dattr_new)
7214 mutex_lock(&sched_domains_mutex);
7216 /* always unregister in case we don't destroy any domains */
7217 unregister_sched_domain_sysctl();
7219 /* Let architecture update cpu core mappings. */
7220 new_topology = arch_update_cpu_topology();
7222 n = doms_new ? ndoms_new : 0;
7224 /* Destroy deleted domains */
7225 for (i = 0; i < ndoms_cur; i++) {
7226 for (j = 0; j < n && !new_topology; j++) {
7227 if (cpumask_equal(doms_cur[i], doms_new[j])
7228 && dattrs_equal(dattr_cur, i, dattr_new, j))
7231 /* no match - a current sched domain not in new doms_new[] */
7232 detach_destroy_domains(doms_cur[i]);
7238 if (doms_new == NULL) {
7240 doms_new = &fallback_doms;
7241 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7242 WARN_ON_ONCE(dattr_new);
7245 /* Build new domains */
7246 for (i = 0; i < ndoms_new; i++) {
7247 for (j = 0; j < n && !new_topology; j++) {
7248 if (cpumask_equal(doms_new[i], doms_cur[j])
7249 && dattrs_equal(dattr_new, i, dattr_cur, j))
7252 /* no match - add a new doms_new */
7253 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7258 /* Remember the new sched domains */
7259 if (doms_cur != &fallback_doms)
7260 free_sched_domains(doms_cur, ndoms_cur);
7261 kfree(dattr_cur); /* kfree(NULL) is safe */
7262 doms_cur = doms_new;
7263 dattr_cur = dattr_new;
7264 ndoms_cur = ndoms_new;
7266 register_sched_domain_sysctl();
7268 mutex_unlock(&sched_domains_mutex);
7271 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7274 * Update cpusets according to cpu_active mask. If cpusets are
7275 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7276 * around partition_sched_domains().
7278 * If we come here as part of a suspend/resume, don't touch cpusets because we
7279 * want to restore it back to its original state upon resume anyway.
7281 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7285 case CPU_ONLINE_FROZEN:
7286 case CPU_DOWN_FAILED_FROZEN:
7289 * num_cpus_frozen tracks how many CPUs are involved in suspend
7290 * resume sequence. As long as this is not the last online
7291 * operation in the resume sequence, just build a single sched
7292 * domain, ignoring cpusets.
7294 partition_sched_domains(1, NULL, NULL);
7295 if (--num_cpus_frozen)
7299 * This is the last CPU online operation. So fall through and
7300 * restore the original sched domains by considering the
7301 * cpuset configurations.
7303 cpuset_force_rebuild();
7306 cpuset_update_active_cpus(true);
7314 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7317 unsigned long flags;
7318 long cpu = (long)hcpu;
7324 case CPU_DOWN_PREPARE:
7325 rcu_read_lock_sched();
7326 dl_b = dl_bw_of(cpu);
7328 raw_spin_lock_irqsave(&dl_b->lock, flags);
7329 cpus = dl_bw_cpus(cpu);
7330 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7331 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7333 rcu_read_unlock_sched();
7336 return notifier_from_errno(-EBUSY);
7337 cpuset_update_active_cpus(false);
7339 case CPU_DOWN_PREPARE_FROZEN:
7341 partition_sched_domains(1, NULL, NULL);
7349 void __init sched_init_smp(void)
7351 cpumask_var_t non_isolated_cpus;
7353 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7354 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7359 * There's no userspace yet to cause hotplug operations; hence all the
7360 * cpu masks are stable and all blatant races in the below code cannot
7363 mutex_lock(&sched_domains_mutex);
7364 init_sched_domains(cpu_active_mask);
7365 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7366 if (cpumask_empty(non_isolated_cpus))
7367 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7368 mutex_unlock(&sched_domains_mutex);
7370 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7371 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7372 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7376 /* Move init over to a non-isolated CPU */
7377 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7379 sched_init_granularity();
7380 free_cpumask_var(non_isolated_cpus);
7382 init_sched_rt_class();
7383 init_sched_dl_class();
7386 void __init sched_init_smp(void)
7388 sched_init_granularity();
7390 #endif /* CONFIG_SMP */
7392 int in_sched_functions(unsigned long addr)
7394 return in_lock_functions(addr) ||
7395 (addr >= (unsigned long)__sched_text_start
7396 && addr < (unsigned long)__sched_text_end);
7399 #ifdef CONFIG_CGROUP_SCHED
7401 * Default task group.
7402 * Every task in system belongs to this group at bootup.
7404 struct task_group root_task_group;
7405 LIST_HEAD(task_groups);
7408 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7410 void __init sched_init(void)
7413 unsigned long alloc_size = 0, ptr;
7415 #ifdef CONFIG_FAIR_GROUP_SCHED
7416 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7418 #ifdef CONFIG_RT_GROUP_SCHED
7419 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7422 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7424 #ifdef CONFIG_FAIR_GROUP_SCHED
7425 root_task_group.se = (struct sched_entity **)ptr;
7426 ptr += nr_cpu_ids * sizeof(void **);
7428 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7429 ptr += nr_cpu_ids * sizeof(void **);
7431 #endif /* CONFIG_FAIR_GROUP_SCHED */
7432 #ifdef CONFIG_RT_GROUP_SCHED
7433 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7434 ptr += nr_cpu_ids * sizeof(void **);
7436 root_task_group.rt_rq = (struct rt_rq **)ptr;
7437 ptr += nr_cpu_ids * sizeof(void **);
7439 #endif /* CONFIG_RT_GROUP_SCHED */
7441 #ifdef CONFIG_CPUMASK_OFFSTACK
7442 for_each_possible_cpu(i) {
7443 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7444 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7446 #endif /* CONFIG_CPUMASK_OFFSTACK */
7448 init_rt_bandwidth(&def_rt_bandwidth,
7449 global_rt_period(), global_rt_runtime());
7450 init_dl_bandwidth(&def_dl_bandwidth,
7451 global_rt_period(), global_rt_runtime());
7454 init_defrootdomain();
7457 #ifdef CONFIG_RT_GROUP_SCHED
7458 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7459 global_rt_period(), global_rt_runtime());
7460 #endif /* CONFIG_RT_GROUP_SCHED */
7462 #ifdef CONFIG_CGROUP_SCHED
7463 list_add(&root_task_group.list, &task_groups);
7464 INIT_LIST_HEAD(&root_task_group.children);
7465 INIT_LIST_HEAD(&root_task_group.siblings);
7466 autogroup_init(&init_task);
7468 #endif /* CONFIG_CGROUP_SCHED */
7470 for_each_possible_cpu(i) {
7474 raw_spin_lock_init(&rq->lock);
7476 rq->calc_load_active = 0;
7477 rq->calc_load_update = jiffies + LOAD_FREQ;
7478 init_cfs_rq(&rq->cfs);
7479 init_rt_rq(&rq->rt);
7480 init_dl_rq(&rq->dl);
7481 #ifdef CONFIG_FAIR_GROUP_SCHED
7482 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7483 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7485 * How much cpu bandwidth does root_task_group get?
7487 * In case of task-groups formed thr' the cgroup filesystem, it
7488 * gets 100% of the cpu resources in the system. This overall
7489 * system cpu resource is divided among the tasks of
7490 * root_task_group and its child task-groups in a fair manner,
7491 * based on each entity's (task or task-group's) weight
7492 * (se->load.weight).
7494 * In other words, if root_task_group has 10 tasks of weight
7495 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7496 * then A0's share of the cpu resource is:
7498 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7500 * We achieve this by letting root_task_group's tasks sit
7501 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7503 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7504 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7505 #endif /* CONFIG_FAIR_GROUP_SCHED */
7507 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7508 #ifdef CONFIG_RT_GROUP_SCHED
7509 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7512 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7513 rq->cpu_load[j] = 0;
7515 rq->last_load_update_tick = jiffies;
7520 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7521 rq->balance_callback = NULL;
7522 rq->active_balance = 0;
7523 rq->next_balance = jiffies;
7528 rq->avg_idle = 2*sysctl_sched_migration_cost;
7529 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7531 INIT_LIST_HEAD(&rq->cfs_tasks);
7533 rq_attach_root(rq, &def_root_domain);
7534 #ifdef CONFIG_NO_HZ_COMMON
7537 #ifdef CONFIG_NO_HZ_FULL
7538 rq->last_sched_tick = 0;
7542 atomic_set(&rq->nr_iowait, 0);
7545 set_load_weight(&init_task);
7547 #ifdef CONFIG_PREEMPT_NOTIFIERS
7548 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7552 * The boot idle thread does lazy MMU switching as well:
7554 atomic_inc(&init_mm.mm_count);
7555 enter_lazy_tlb(&init_mm, current);
7558 * During early bootup we pretend to be a normal task:
7560 current->sched_class = &fair_sched_class;
7563 * Make us the idle thread. Technically, schedule() should not be
7564 * called from this thread, however somewhere below it might be,
7565 * but because we are the idle thread, we just pick up running again
7566 * when this runqueue becomes "idle".
7568 init_idle(current, smp_processor_id());
7570 calc_load_update = jiffies + LOAD_FREQ;
7573 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7574 /* May be allocated at isolcpus cmdline parse time */
7575 if (cpu_isolated_map == NULL)
7576 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7577 idle_thread_set_boot_cpu();
7578 set_cpu_rq_start_time();
7580 init_sched_fair_class();
7582 scheduler_running = 1;
7585 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7586 static inline int preempt_count_equals(int preempt_offset)
7588 int nested = preempt_count() + rcu_preempt_depth();
7590 return (nested == preempt_offset);
7593 void __might_sleep(const char *file, int line, int preempt_offset)
7596 * Blocking primitives will set (and therefore destroy) current->state,
7597 * since we will exit with TASK_RUNNING make sure we enter with it,
7598 * otherwise we will destroy state.
7600 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7601 "do not call blocking ops when !TASK_RUNNING; "
7602 "state=%lx set at [<%p>] %pS\n",
7604 (void *)current->task_state_change,
7605 (void *)current->task_state_change);
7607 ___might_sleep(file, line, preempt_offset);
7609 EXPORT_SYMBOL(__might_sleep);
7611 void ___might_sleep(const char *file, int line, int preempt_offset)
7613 static unsigned long prev_jiffy; /* ratelimiting */
7615 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7616 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7617 !is_idle_task(current)) ||
7618 system_state != SYSTEM_RUNNING || oops_in_progress)
7620 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7622 prev_jiffy = jiffies;
7625 "BUG: sleeping function called from invalid context at %s:%d\n",
7628 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7629 in_atomic(), irqs_disabled(),
7630 current->pid, current->comm);
7632 if (task_stack_end_corrupted(current))
7633 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7635 debug_show_held_locks(current);
7636 if (irqs_disabled())
7637 print_irqtrace_events(current);
7638 #ifdef CONFIG_DEBUG_PREEMPT
7639 if (!preempt_count_equals(preempt_offset)) {
7640 pr_err("Preemption disabled at:");
7641 print_ip_sym(current->preempt_disable_ip);
7647 EXPORT_SYMBOL(___might_sleep);
7650 #ifdef CONFIG_MAGIC_SYSRQ
7651 void normalize_rt_tasks(void)
7653 struct task_struct *g, *p;
7654 struct sched_attr attr = {
7655 .sched_policy = SCHED_NORMAL,
7658 read_lock(&tasklist_lock);
7659 for_each_process_thread(g, p) {
7661 * Only normalize user tasks:
7663 if (p->flags & PF_KTHREAD)
7666 p->se.exec_start = 0;
7667 #ifdef CONFIG_SCHEDSTATS
7668 p->se.statistics.wait_start = 0;
7669 p->se.statistics.sleep_start = 0;
7670 p->se.statistics.block_start = 0;
7673 if (!dl_task(p) && !rt_task(p)) {
7675 * Renice negative nice level userspace
7678 if (task_nice(p) < 0)
7679 set_user_nice(p, 0);
7683 __sched_setscheduler(p, &attr, false, false);
7685 read_unlock(&tasklist_lock);
7688 #endif /* CONFIG_MAGIC_SYSRQ */
7690 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7692 * These functions are only useful for the IA64 MCA handling, or kdb.
7694 * They can only be called when the whole system has been
7695 * stopped - every CPU needs to be quiescent, and no scheduling
7696 * activity can take place. Using them for anything else would
7697 * be a serious bug, and as a result, they aren't even visible
7698 * under any other configuration.
7702 * curr_task - return the current task for a given cpu.
7703 * @cpu: the processor in question.
7705 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7707 * Return: The current task for @cpu.
7709 struct task_struct *curr_task(int cpu)
7711 return cpu_curr(cpu);
7714 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7718 * set_curr_task - set the current task for a given cpu.
7719 * @cpu: the processor in question.
7720 * @p: the task pointer to set.
7722 * Description: This function must only be used when non-maskable interrupts
7723 * are serviced on a separate stack. It allows the architecture to switch the
7724 * notion of the current task on a cpu in a non-blocking manner. This function
7725 * must be called with all CPU's synchronized, and interrupts disabled, the
7726 * and caller must save the original value of the current task (see
7727 * curr_task() above) and restore that value before reenabling interrupts and
7728 * re-starting the system.
7730 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7732 void set_curr_task(int cpu, struct task_struct *p)
7739 #ifdef CONFIG_CGROUP_SCHED
7740 /* task_group_lock serializes the addition/removal of task groups */
7741 static DEFINE_SPINLOCK(task_group_lock);
7743 static void sched_free_group(struct task_group *tg)
7745 free_fair_sched_group(tg);
7746 free_rt_sched_group(tg);
7751 /* allocate runqueue etc for a new task group */
7752 struct task_group *sched_create_group(struct task_group *parent)
7754 struct task_group *tg;
7756 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7758 return ERR_PTR(-ENOMEM);
7760 if (!alloc_fair_sched_group(tg, parent))
7763 if (!alloc_rt_sched_group(tg, parent))
7769 sched_free_group(tg);
7770 return ERR_PTR(-ENOMEM);
7773 void sched_online_group(struct task_group *tg, struct task_group *parent)
7775 unsigned long flags;
7777 spin_lock_irqsave(&task_group_lock, flags);
7778 list_add_rcu(&tg->list, &task_groups);
7780 WARN_ON(!parent); /* root should already exist */
7782 tg->parent = parent;
7783 INIT_LIST_HEAD(&tg->children);
7784 list_add_rcu(&tg->siblings, &parent->children);
7785 spin_unlock_irqrestore(&task_group_lock, flags);
7788 /* rcu callback to free various structures associated with a task group */
7789 static void sched_free_group_rcu(struct rcu_head *rhp)
7791 /* now it should be safe to free those cfs_rqs */
7792 sched_free_group(container_of(rhp, struct task_group, rcu));
7795 void sched_destroy_group(struct task_group *tg)
7797 /* wait for possible concurrent references to cfs_rqs complete */
7798 call_rcu(&tg->rcu, sched_free_group_rcu);
7801 void sched_offline_group(struct task_group *tg)
7803 unsigned long flags;
7806 /* end participation in shares distribution */
7807 for_each_possible_cpu(i)
7808 unregister_fair_sched_group(tg, i);
7810 spin_lock_irqsave(&task_group_lock, flags);
7811 list_del_rcu(&tg->list);
7812 list_del_rcu(&tg->siblings);
7813 spin_unlock_irqrestore(&task_group_lock, flags);
7816 /* change task's runqueue when it moves between groups.
7817 * The caller of this function should have put the task in its new group
7818 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7819 * reflect its new group.
7821 void sched_move_task(struct task_struct *tsk)
7823 struct task_group *tg;
7824 int queued, running;
7825 unsigned long flags;
7828 rq = task_rq_lock(tsk, &flags);
7830 running = task_current(rq, tsk);
7831 queued = task_on_rq_queued(tsk);
7834 dequeue_task(rq, tsk, DEQUEUE_SAVE);
7835 if (unlikely(running))
7836 put_prev_task(rq, tsk);
7839 * All callers are synchronized by task_rq_lock(); we do not use RCU
7840 * which is pointless here. Thus, we pass "true" to task_css_check()
7841 * to prevent lockdep warnings.
7843 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7844 struct task_group, css);
7845 tg = autogroup_task_group(tsk, tg);
7846 tsk->sched_task_group = tg;
7848 #ifdef CONFIG_FAIR_GROUP_SCHED
7849 if (tsk->sched_class->task_move_group)
7850 tsk->sched_class->task_move_group(tsk);
7853 set_task_rq(tsk, task_cpu(tsk));
7855 if (unlikely(running))
7856 tsk->sched_class->set_curr_task(rq);
7858 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
7860 task_rq_unlock(rq, tsk, &flags);
7862 #endif /* CONFIG_CGROUP_SCHED */
7864 #ifdef CONFIG_RT_GROUP_SCHED
7866 * Ensure that the real time constraints are schedulable.
7868 static DEFINE_MUTEX(rt_constraints_mutex);
7870 /* Must be called with tasklist_lock held */
7871 static inline int tg_has_rt_tasks(struct task_group *tg)
7873 struct task_struct *g, *p;
7876 * Autogroups do not have RT tasks; see autogroup_create().
7878 if (task_group_is_autogroup(tg))
7881 for_each_process_thread(g, p) {
7882 if (rt_task(p) && task_group(p) == tg)
7889 struct rt_schedulable_data {
7890 struct task_group *tg;
7895 static int tg_rt_schedulable(struct task_group *tg, void *data)
7897 struct rt_schedulable_data *d = data;
7898 struct task_group *child;
7899 unsigned long total, sum = 0;
7900 u64 period, runtime;
7902 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7903 runtime = tg->rt_bandwidth.rt_runtime;
7906 period = d->rt_period;
7907 runtime = d->rt_runtime;
7911 * Cannot have more runtime than the period.
7913 if (runtime > period && runtime != RUNTIME_INF)
7917 * Ensure we don't starve existing RT tasks.
7919 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7922 total = to_ratio(period, runtime);
7925 * Nobody can have more than the global setting allows.
7927 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7931 * The sum of our children's runtime should not exceed our own.
7933 list_for_each_entry_rcu(child, &tg->children, siblings) {
7934 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7935 runtime = child->rt_bandwidth.rt_runtime;
7937 if (child == d->tg) {
7938 period = d->rt_period;
7939 runtime = d->rt_runtime;
7942 sum += to_ratio(period, runtime);
7951 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7955 struct rt_schedulable_data data = {
7957 .rt_period = period,
7958 .rt_runtime = runtime,
7962 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7968 static int tg_set_rt_bandwidth(struct task_group *tg,
7969 u64 rt_period, u64 rt_runtime)
7974 * Disallowing the root group RT runtime is BAD, it would disallow the
7975 * kernel creating (and or operating) RT threads.
7977 if (tg == &root_task_group && rt_runtime == 0)
7980 /* No period doesn't make any sense. */
7984 mutex_lock(&rt_constraints_mutex);
7985 read_lock(&tasklist_lock);
7986 err = __rt_schedulable(tg, rt_period, rt_runtime);
7990 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7991 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7992 tg->rt_bandwidth.rt_runtime = rt_runtime;
7994 for_each_possible_cpu(i) {
7995 struct rt_rq *rt_rq = tg->rt_rq[i];
7997 raw_spin_lock(&rt_rq->rt_runtime_lock);
7998 rt_rq->rt_runtime = rt_runtime;
7999 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8001 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8003 read_unlock(&tasklist_lock);
8004 mutex_unlock(&rt_constraints_mutex);
8009 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8011 u64 rt_runtime, rt_period;
8013 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8014 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8015 if (rt_runtime_us < 0)
8016 rt_runtime = RUNTIME_INF;
8018 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8021 static long sched_group_rt_runtime(struct task_group *tg)
8025 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8028 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8029 do_div(rt_runtime_us, NSEC_PER_USEC);
8030 return rt_runtime_us;
8033 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8035 u64 rt_runtime, rt_period;
8037 rt_period = rt_period_us * NSEC_PER_USEC;
8038 rt_runtime = tg->rt_bandwidth.rt_runtime;
8040 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8043 static long sched_group_rt_period(struct task_group *tg)
8047 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8048 do_div(rt_period_us, NSEC_PER_USEC);
8049 return rt_period_us;
8051 #endif /* CONFIG_RT_GROUP_SCHED */
8053 #ifdef CONFIG_RT_GROUP_SCHED
8054 static int sched_rt_global_constraints(void)
8058 mutex_lock(&rt_constraints_mutex);
8059 read_lock(&tasklist_lock);
8060 ret = __rt_schedulable(NULL, 0, 0);
8061 read_unlock(&tasklist_lock);
8062 mutex_unlock(&rt_constraints_mutex);
8067 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8069 /* Don't accept realtime tasks when there is no way for them to run */
8070 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8076 #else /* !CONFIG_RT_GROUP_SCHED */
8077 static int sched_rt_global_constraints(void)
8079 unsigned long flags;
8082 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8083 for_each_possible_cpu(i) {
8084 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8086 raw_spin_lock(&rt_rq->rt_runtime_lock);
8087 rt_rq->rt_runtime = global_rt_runtime();
8088 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8090 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8094 #endif /* CONFIG_RT_GROUP_SCHED */
8096 static int sched_dl_global_validate(void)
8098 u64 runtime = global_rt_runtime();
8099 u64 period = global_rt_period();
8100 u64 new_bw = to_ratio(period, runtime);
8103 unsigned long flags;
8106 * Here we want to check the bandwidth not being set to some
8107 * value smaller than the currently allocated bandwidth in
8108 * any of the root_domains.
8110 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8111 * cycling on root_domains... Discussion on different/better
8112 * solutions is welcome!
8114 for_each_possible_cpu(cpu) {
8115 rcu_read_lock_sched();
8116 dl_b = dl_bw_of(cpu);
8118 raw_spin_lock_irqsave(&dl_b->lock, flags);
8119 if (new_bw < dl_b->total_bw)
8121 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8123 rcu_read_unlock_sched();
8132 static void sched_dl_do_global(void)
8137 unsigned long flags;
8139 def_dl_bandwidth.dl_period = global_rt_period();
8140 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8142 if (global_rt_runtime() != RUNTIME_INF)
8143 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8146 * FIXME: As above...
8148 for_each_possible_cpu(cpu) {
8149 rcu_read_lock_sched();
8150 dl_b = dl_bw_of(cpu);
8152 raw_spin_lock_irqsave(&dl_b->lock, flags);
8154 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8156 rcu_read_unlock_sched();
8160 static int sched_rt_global_validate(void)
8162 if (sysctl_sched_rt_period <= 0)
8165 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8166 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8172 static void sched_rt_do_global(void)
8174 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8175 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8178 int sched_rt_handler(struct ctl_table *table, int write,
8179 void __user *buffer, size_t *lenp,
8182 int old_period, old_runtime;
8183 static DEFINE_MUTEX(mutex);
8187 old_period = sysctl_sched_rt_period;
8188 old_runtime = sysctl_sched_rt_runtime;
8190 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8192 if (!ret && write) {
8193 ret = sched_rt_global_validate();
8197 ret = sched_dl_global_validate();
8201 ret = sched_rt_global_constraints();
8205 sched_rt_do_global();
8206 sched_dl_do_global();
8210 sysctl_sched_rt_period = old_period;
8211 sysctl_sched_rt_runtime = old_runtime;
8213 mutex_unlock(&mutex);
8218 int sched_rr_handler(struct ctl_table *table, int write,
8219 void __user *buffer, size_t *lenp,
8223 static DEFINE_MUTEX(mutex);
8226 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8227 /* make sure that internally we keep jiffies */
8228 /* also, writing zero resets timeslice to default */
8229 if (!ret && write) {
8230 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8231 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8233 mutex_unlock(&mutex);
8237 #ifdef CONFIG_CGROUP_SCHED
8239 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8241 return css ? container_of(css, struct task_group, css) : NULL;
8244 static struct cgroup_subsys_state *
8245 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8247 struct task_group *parent = css_tg(parent_css);
8248 struct task_group *tg;
8251 /* This is early initialization for the top cgroup */
8252 return &root_task_group.css;
8255 tg = sched_create_group(parent);
8257 return ERR_PTR(-ENOMEM);
8262 /* Expose task group only after completing cgroup initialization */
8263 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8265 struct task_group *tg = css_tg(css);
8266 struct task_group *parent = css_tg(css->parent);
8269 sched_online_group(tg, parent);
8273 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8275 struct task_group *tg = css_tg(css);
8277 sched_offline_group(tg);
8280 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8282 struct task_group *tg = css_tg(css);
8285 * Relies on the RCU grace period between css_released() and this.
8287 sched_free_group(tg);
8290 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8292 sched_move_task(task);
8295 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8297 struct task_struct *task;
8298 struct cgroup_subsys_state *css;
8300 cgroup_taskset_for_each(task, css, tset) {
8301 #ifdef CONFIG_RT_GROUP_SCHED
8302 if (!sched_rt_can_attach(css_tg(css), task))
8305 /* We don't support RT-tasks being in separate groups */
8306 if (task->sched_class != &fair_sched_class)
8313 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8315 struct task_struct *task;
8316 struct cgroup_subsys_state *css;
8318 cgroup_taskset_for_each(task, css, tset)
8319 sched_move_task(task);
8322 #ifdef CONFIG_FAIR_GROUP_SCHED
8323 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8324 struct cftype *cftype, u64 shareval)
8326 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8329 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8332 struct task_group *tg = css_tg(css);
8334 return (u64) scale_load_down(tg->shares);
8337 #ifdef CONFIG_CFS_BANDWIDTH
8338 static DEFINE_MUTEX(cfs_constraints_mutex);
8340 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8341 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8343 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8345 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8347 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8348 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8350 if (tg == &root_task_group)
8354 * Ensure we have at some amount of bandwidth every period. This is
8355 * to prevent reaching a state of large arrears when throttled via
8356 * entity_tick() resulting in prolonged exit starvation.
8358 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8362 * Likewise, bound things on the otherside by preventing insane quota
8363 * periods. This also allows us to normalize in computing quota
8366 if (period > max_cfs_quota_period)
8370 * Prevent race between setting of cfs_rq->runtime_enabled and
8371 * unthrottle_offline_cfs_rqs().
8374 mutex_lock(&cfs_constraints_mutex);
8375 ret = __cfs_schedulable(tg, period, quota);
8379 runtime_enabled = quota != RUNTIME_INF;
8380 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8382 * If we need to toggle cfs_bandwidth_used, off->on must occur
8383 * before making related changes, and on->off must occur afterwards
8385 if (runtime_enabled && !runtime_was_enabled)
8386 cfs_bandwidth_usage_inc();
8387 raw_spin_lock_irq(&cfs_b->lock);
8388 cfs_b->period = ns_to_ktime(period);
8389 cfs_b->quota = quota;
8391 __refill_cfs_bandwidth_runtime(cfs_b);
8392 /* restart the period timer (if active) to handle new period expiry */
8393 if (runtime_enabled)
8394 start_cfs_bandwidth(cfs_b);
8395 raw_spin_unlock_irq(&cfs_b->lock);
8397 for_each_online_cpu(i) {
8398 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8399 struct rq *rq = cfs_rq->rq;
8401 raw_spin_lock_irq(&rq->lock);
8402 cfs_rq->runtime_enabled = runtime_enabled;
8403 cfs_rq->runtime_remaining = 0;
8405 if (cfs_rq->throttled)
8406 unthrottle_cfs_rq(cfs_rq);
8407 raw_spin_unlock_irq(&rq->lock);
8409 if (runtime_was_enabled && !runtime_enabled)
8410 cfs_bandwidth_usage_dec();
8412 mutex_unlock(&cfs_constraints_mutex);
8418 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8422 period = ktime_to_ns(tg->cfs_bandwidth.period);
8423 if (cfs_quota_us < 0)
8424 quota = RUNTIME_INF;
8426 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8428 return tg_set_cfs_bandwidth(tg, period, quota);
8431 long tg_get_cfs_quota(struct task_group *tg)
8435 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8438 quota_us = tg->cfs_bandwidth.quota;
8439 do_div(quota_us, NSEC_PER_USEC);
8444 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8448 period = (u64)cfs_period_us * NSEC_PER_USEC;
8449 quota = tg->cfs_bandwidth.quota;
8451 return tg_set_cfs_bandwidth(tg, period, quota);
8454 long tg_get_cfs_period(struct task_group *tg)
8458 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8459 do_div(cfs_period_us, NSEC_PER_USEC);
8461 return cfs_period_us;
8464 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8467 return tg_get_cfs_quota(css_tg(css));
8470 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8471 struct cftype *cftype, s64 cfs_quota_us)
8473 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8476 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8479 return tg_get_cfs_period(css_tg(css));
8482 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8483 struct cftype *cftype, u64 cfs_period_us)
8485 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8488 struct cfs_schedulable_data {
8489 struct task_group *tg;
8494 * normalize group quota/period to be quota/max_period
8495 * note: units are usecs
8497 static u64 normalize_cfs_quota(struct task_group *tg,
8498 struct cfs_schedulable_data *d)
8506 period = tg_get_cfs_period(tg);
8507 quota = tg_get_cfs_quota(tg);
8510 /* note: these should typically be equivalent */
8511 if (quota == RUNTIME_INF || quota == -1)
8514 return to_ratio(period, quota);
8517 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8519 struct cfs_schedulable_data *d = data;
8520 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8521 s64 quota = 0, parent_quota = -1;
8524 quota = RUNTIME_INF;
8526 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8528 quota = normalize_cfs_quota(tg, d);
8529 parent_quota = parent_b->hierarchical_quota;
8532 * ensure max(child_quota) <= parent_quota, inherit when no
8535 if (quota == RUNTIME_INF)
8536 quota = parent_quota;
8537 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8540 cfs_b->hierarchical_quota = quota;
8545 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8548 struct cfs_schedulable_data data = {
8554 if (quota != RUNTIME_INF) {
8555 do_div(data.period, NSEC_PER_USEC);
8556 do_div(data.quota, NSEC_PER_USEC);
8560 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8566 static int cpu_stats_show(struct seq_file *sf, void *v)
8568 struct task_group *tg = css_tg(seq_css(sf));
8569 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8571 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8572 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8573 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8577 #endif /* CONFIG_CFS_BANDWIDTH */
8578 #endif /* CONFIG_FAIR_GROUP_SCHED */
8580 #ifdef CONFIG_RT_GROUP_SCHED
8581 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8582 struct cftype *cft, s64 val)
8584 return sched_group_set_rt_runtime(css_tg(css), val);
8587 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8590 return sched_group_rt_runtime(css_tg(css));
8593 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8594 struct cftype *cftype, u64 rt_period_us)
8596 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8599 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8602 return sched_group_rt_period(css_tg(css));
8604 #endif /* CONFIG_RT_GROUP_SCHED */
8606 static struct cftype cpu_files[] = {
8607 #ifdef CONFIG_FAIR_GROUP_SCHED
8610 .read_u64 = cpu_shares_read_u64,
8611 .write_u64 = cpu_shares_write_u64,
8614 #ifdef CONFIG_CFS_BANDWIDTH
8616 .name = "cfs_quota_us",
8617 .read_s64 = cpu_cfs_quota_read_s64,
8618 .write_s64 = cpu_cfs_quota_write_s64,
8621 .name = "cfs_period_us",
8622 .read_u64 = cpu_cfs_period_read_u64,
8623 .write_u64 = cpu_cfs_period_write_u64,
8627 .seq_show = cpu_stats_show,
8630 #ifdef CONFIG_RT_GROUP_SCHED
8632 .name = "rt_runtime_us",
8633 .read_s64 = cpu_rt_runtime_read,
8634 .write_s64 = cpu_rt_runtime_write,
8637 .name = "rt_period_us",
8638 .read_u64 = cpu_rt_period_read_uint,
8639 .write_u64 = cpu_rt_period_write_uint,
8645 struct cgroup_subsys cpu_cgrp_subsys = {
8646 .css_alloc = cpu_cgroup_css_alloc,
8647 .css_online = cpu_cgroup_css_online,
8648 .css_released = cpu_cgroup_css_released,
8649 .css_free = cpu_cgroup_css_free,
8650 .fork = cpu_cgroup_fork,
8651 .can_attach = cpu_cgroup_can_attach,
8652 .attach = cpu_cgroup_attach,
8653 .legacy_cftypes = cpu_files,
8657 #endif /* CONFIG_CGROUP_SCHED */
8659 void dump_cpu_task(int cpu)
8661 pr_info("Task dump for CPU %d:\n", cpu);
8662 sched_show_task(cpu_curr(cpu));