4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
96 ktime_t soft, hard, now;
99 if (hrtimer_active(period_timer))
102 now = hrtimer_cb_get_time(period_timer);
103 hrtimer_forward(period_timer, now, period);
105 soft = hrtimer_get_softexpires(period_timer);
106 hard = hrtimer_get_expires(period_timer);
107 delta = ktime_to_ns(ktime_sub(hard, soft));
108 __hrtimer_start_range_ns(period_timer, soft, delta,
109 HRTIMER_MODE_ABS_PINNED, 0);
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
118 void update_rq_clock(struct rq *rq)
122 lockdep_assert_held(&rq->lock);
124 if (rq->clock_skip_update & RQCF_ACT_SKIP)
127 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
131 update_rq_clock_task(rq, delta);
135 * Debugging: various feature bits
138 #define SCHED_FEAT(name, enabled) \
139 (1UL << __SCHED_FEAT_##name) * enabled |
141 const_debug unsigned int sysctl_sched_features =
142 #include "features.h"
147 #ifdef CONFIG_SCHED_DEBUG
148 #define SCHED_FEAT(name, enabled) \
151 static const char * const sched_feat_names[] = {
152 #include "features.h"
157 static int sched_feat_show(struct seq_file *m, void *v)
161 for (i = 0; i < __SCHED_FEAT_NR; i++) {
162 if (!(sysctl_sched_features & (1UL << i)))
164 seq_printf(m, "%s ", sched_feat_names[i]);
171 #ifdef HAVE_JUMP_LABEL
173 #define jump_label_key__true STATIC_KEY_INIT_TRUE
174 #define jump_label_key__false STATIC_KEY_INIT_FALSE
176 #define SCHED_FEAT(name, enabled) \
177 jump_label_key__##enabled ,
179 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
180 #include "features.h"
185 static void sched_feat_disable(int i)
187 if (static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_dec(&sched_feat_keys[i]);
191 static void sched_feat_enable(int i)
193 if (!static_key_enabled(&sched_feat_keys[i]))
194 static_key_slow_inc(&sched_feat_keys[i]);
197 static void sched_feat_disable(int i) { };
198 static void sched_feat_enable(int i) { };
199 #endif /* HAVE_JUMP_LABEL */
201 static int sched_feat_set(char *cmp)
206 if (strncmp(cmp, "NO_", 3) == 0) {
211 for (i = 0; i < __SCHED_FEAT_NR; i++) {
212 if (strcmp(cmp, sched_feat_names[i]) == 0) {
214 sysctl_sched_features &= ~(1UL << i);
215 sched_feat_disable(i);
217 sysctl_sched_features |= (1UL << i);
218 sched_feat_enable(i);
228 sched_feat_write(struct file *filp, const char __user *ubuf,
229 size_t cnt, loff_t *ppos)
239 if (copy_from_user(&buf, ubuf, cnt))
245 /* Ensure the static_key remains in a consistent state */
246 inode = file_inode(filp);
247 mutex_lock(&inode->i_mutex);
248 i = sched_feat_set(cmp);
249 mutex_unlock(&inode->i_mutex);
250 if (i == __SCHED_FEAT_NR)
258 static int sched_feat_open(struct inode *inode, struct file *filp)
260 return single_open(filp, sched_feat_show, NULL);
263 static const struct file_operations sched_feat_fops = {
264 .open = sched_feat_open,
265 .write = sched_feat_write,
268 .release = single_release,
271 static __init int sched_init_debug(void)
273 debugfs_create_file("sched_features", 0644, NULL, NULL,
278 late_initcall(sched_init_debug);
279 #endif /* CONFIG_SCHED_DEBUG */
282 * Number of tasks to iterate in a single balance run.
283 * Limited because this is done with IRQs disabled.
285 const_debug unsigned int sysctl_sched_nr_migrate = 32;
288 * period over which we average the RT time consumption, measured
293 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
296 * period over which we measure -rt task cpu usage in us.
299 unsigned int sysctl_sched_rt_period = 1000000;
301 __read_mostly int scheduler_running;
304 * part of the period that we allow rt tasks to run in us.
307 int sysctl_sched_rt_runtime = 950000;
309 /* cpus with isolated domains */
310 cpumask_var_t cpu_isolated_map;
313 * this_rq_lock - lock this runqueue and disable interrupts.
315 static struct rq *this_rq_lock(void)
322 raw_spin_lock(&rq->lock);
327 #ifdef CONFIG_SCHED_HRTICK
329 * Use HR-timers to deliver accurate preemption points.
332 static void hrtick_clear(struct rq *rq)
334 if (hrtimer_active(&rq->hrtick_timer))
335 hrtimer_cancel(&rq->hrtick_timer);
339 * High-resolution timer tick.
340 * Runs from hardirq context with interrupts disabled.
342 static enum hrtimer_restart hrtick(struct hrtimer *timer)
344 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
346 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
348 raw_spin_lock(&rq->lock);
350 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
351 raw_spin_unlock(&rq->lock);
353 return HRTIMER_NORESTART;
358 static int __hrtick_restart(struct rq *rq)
360 struct hrtimer *timer = &rq->hrtick_timer;
361 ktime_t time = hrtimer_get_softexpires(timer);
363 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
367 * called from hardirq (IPI) context
369 static void __hrtick_start(void *arg)
373 raw_spin_lock(&rq->lock);
374 __hrtick_restart(rq);
375 rq->hrtick_csd_pending = 0;
376 raw_spin_unlock(&rq->lock);
380 * Called to set the hrtick timer state.
382 * called with rq->lock held and irqs disabled
384 void hrtick_start(struct rq *rq, u64 delay)
386 struct hrtimer *timer = &rq->hrtick_timer;
391 * Don't schedule slices shorter than 10000ns, that just
392 * doesn't make sense and can cause timer DoS.
394 delta = max_t(s64, delay, 10000LL);
395 time = ktime_add_ns(timer->base->get_time(), delta);
397 hrtimer_set_expires(timer, time);
399 if (rq == this_rq()) {
400 __hrtick_restart(rq);
401 } else if (!rq->hrtick_csd_pending) {
402 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
403 rq->hrtick_csd_pending = 1;
408 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
410 int cpu = (int)(long)hcpu;
413 case CPU_UP_CANCELED:
414 case CPU_UP_CANCELED_FROZEN:
415 case CPU_DOWN_PREPARE:
416 case CPU_DOWN_PREPARE_FROZEN:
418 case CPU_DEAD_FROZEN:
419 hrtick_clear(cpu_rq(cpu));
426 static __init void init_hrtick(void)
428 hotcpu_notifier(hotplug_hrtick, 0);
432 * Called to set the hrtick timer state.
434 * called with rq->lock held and irqs disabled
436 void hrtick_start(struct rq *rq, u64 delay)
439 * Don't schedule slices shorter than 10000ns, that just
440 * doesn't make sense. Rely on vruntime for fairness.
442 delay = max_t(u64, delay, 10000LL);
443 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
444 HRTIMER_MODE_REL_PINNED, 0);
447 static inline void init_hrtick(void)
450 #endif /* CONFIG_SMP */
452 static void init_rq_hrtick(struct rq *rq)
455 rq->hrtick_csd_pending = 0;
457 rq->hrtick_csd.flags = 0;
458 rq->hrtick_csd.func = __hrtick_start;
459 rq->hrtick_csd.info = rq;
462 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
463 rq->hrtick_timer.function = hrtick;
465 #else /* CONFIG_SCHED_HRTICK */
466 static inline void hrtick_clear(struct rq *rq)
470 static inline void init_rq_hrtick(struct rq *rq)
474 static inline void init_hrtick(void)
477 #endif /* CONFIG_SCHED_HRTICK */
480 * cmpxchg based fetch_or, macro so it works for different integer types
482 #define fetch_or(ptr, val) \
483 ({ typeof(*(ptr)) __old, __val = *(ptr); \
485 __old = cmpxchg((ptr), __val, __val | (val)); \
486 if (__old == __val) \
493 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
495 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
496 * this avoids any races wrt polling state changes and thereby avoids
499 static bool set_nr_and_not_polling(struct task_struct *p)
501 struct thread_info *ti = task_thread_info(p);
502 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
506 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
508 * If this returns true, then the idle task promises to call
509 * sched_ttwu_pending() and reschedule soon.
511 static bool set_nr_if_polling(struct task_struct *p)
513 struct thread_info *ti = task_thread_info(p);
514 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
517 if (!(val & _TIF_POLLING_NRFLAG))
519 if (val & _TIF_NEED_RESCHED)
521 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
530 static bool set_nr_and_not_polling(struct task_struct *p)
532 set_tsk_need_resched(p);
537 static bool set_nr_if_polling(struct task_struct *p)
544 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
546 struct wake_q_node *node = &task->wake_q;
549 * Atomically grab the task, if ->wake_q is !nil already it means
550 * its already queued (either by us or someone else) and will get the
551 * wakeup due to that.
553 * This cmpxchg() implies a full barrier, which pairs with the write
554 * barrier implied by the wakeup in wake_up_list().
556 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
559 get_task_struct(task);
562 * The head is context local, there can be no concurrency.
565 head->lastp = &node->next;
568 void wake_up_q(struct wake_q_head *head)
570 struct wake_q_node *node = head->first;
572 while (node != WAKE_Q_TAIL) {
573 struct task_struct *task;
575 task = container_of(node, struct task_struct, wake_q);
577 /* task can safely be re-inserted now */
579 task->wake_q.next = NULL;
582 * wake_up_process() implies a wmb() to pair with the queueing
583 * in wake_q_add() so as not to miss wakeups.
585 wake_up_process(task);
586 put_task_struct(task);
591 * resched_curr - mark rq's current task 'to be rescheduled now'.
593 * On UP this means the setting of the need_resched flag, on SMP it
594 * might also involve a cross-CPU call to trigger the scheduler on
597 void resched_curr(struct rq *rq)
599 struct task_struct *curr = rq->curr;
602 lockdep_assert_held(&rq->lock);
604 if (test_tsk_need_resched(curr))
609 if (cpu == smp_processor_id()) {
610 set_tsk_need_resched(curr);
611 set_preempt_need_resched();
615 if (set_nr_and_not_polling(curr))
616 smp_send_reschedule(cpu);
618 trace_sched_wake_idle_without_ipi(cpu);
621 void resched_cpu(int cpu)
623 struct rq *rq = cpu_rq(cpu);
626 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
629 raw_spin_unlock_irqrestore(&rq->lock, flags);
633 #ifdef CONFIG_NO_HZ_COMMON
635 * In the semi idle case, use the nearest busy cpu for migrating timers
636 * from an idle cpu. This is good for power-savings.
638 * We don't do similar optimization for completely idle system, as
639 * selecting an idle cpu will add more delays to the timers than intended
640 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
642 int get_nohz_timer_target(int pinned)
644 int cpu = smp_processor_id();
646 struct sched_domain *sd;
648 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
652 for_each_domain(cpu, sd) {
653 for_each_cpu(i, sched_domain_span(sd)) {
665 * When add_timer_on() enqueues a timer into the timer wheel of an
666 * idle CPU then this timer might expire before the next timer event
667 * which is scheduled to wake up that CPU. In case of a completely
668 * idle system the next event might even be infinite time into the
669 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
670 * leaves the inner idle loop so the newly added timer is taken into
671 * account when the CPU goes back to idle and evaluates the timer
672 * wheel for the next timer event.
674 static void wake_up_idle_cpu(int cpu)
676 struct rq *rq = cpu_rq(cpu);
678 if (cpu == smp_processor_id())
681 if (set_nr_and_not_polling(rq->idle))
682 smp_send_reschedule(cpu);
684 trace_sched_wake_idle_without_ipi(cpu);
687 static bool wake_up_full_nohz_cpu(int cpu)
690 * We just need the target to call irq_exit() and re-evaluate
691 * the next tick. The nohz full kick at least implies that.
692 * If needed we can still optimize that later with an
695 if (tick_nohz_full_cpu(cpu)) {
696 if (cpu != smp_processor_id() ||
697 tick_nohz_tick_stopped())
698 tick_nohz_full_kick_cpu(cpu);
705 void wake_up_nohz_cpu(int cpu)
707 if (!wake_up_full_nohz_cpu(cpu))
708 wake_up_idle_cpu(cpu);
711 static inline bool got_nohz_idle_kick(void)
713 int cpu = smp_processor_id();
715 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
718 if (idle_cpu(cpu) && !need_resched())
722 * We can't run Idle Load Balance on this CPU for this time so we
723 * cancel it and clear NOHZ_BALANCE_KICK
725 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
729 #else /* CONFIG_NO_HZ_COMMON */
731 static inline bool got_nohz_idle_kick(void)
736 #endif /* CONFIG_NO_HZ_COMMON */
738 #ifdef CONFIG_NO_HZ_FULL
739 bool sched_can_stop_tick(void)
742 * FIFO realtime policy runs the highest priority task. Other runnable
743 * tasks are of a lower priority. The scheduler tick does nothing.
745 if (current->policy == SCHED_FIFO)
749 * Round-robin realtime tasks time slice with other tasks at the same
750 * realtime priority. Is this task the only one at this priority?
752 if (current->policy == SCHED_RR) {
753 struct sched_rt_entity *rt_se = ¤t->rt;
755 return rt_se->run_list.prev == rt_se->run_list.next;
759 * More than one running task need preemption.
760 * nr_running update is assumed to be visible
761 * after IPI is sent from wakers.
763 if (this_rq()->nr_running > 1)
768 #endif /* CONFIG_NO_HZ_FULL */
770 void sched_avg_update(struct rq *rq)
772 s64 period = sched_avg_period();
774 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
776 * Inline assembly required to prevent the compiler
777 * optimising this loop into a divmod call.
778 * See __iter_div_u64_rem() for another example of this.
780 asm("" : "+rm" (rq->age_stamp));
781 rq->age_stamp += period;
786 #endif /* CONFIG_SMP */
788 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
789 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
791 * Iterate task_group tree rooted at *from, calling @down when first entering a
792 * node and @up when leaving it for the final time.
794 * Caller must hold rcu_lock or sufficient equivalent.
796 int walk_tg_tree_from(struct task_group *from,
797 tg_visitor down, tg_visitor up, void *data)
799 struct task_group *parent, *child;
805 ret = (*down)(parent, data);
808 list_for_each_entry_rcu(child, &parent->children, siblings) {
815 ret = (*up)(parent, data);
816 if (ret || parent == from)
820 parent = parent->parent;
827 int tg_nop(struct task_group *tg, void *data)
833 static void set_load_weight(struct task_struct *p)
835 int prio = p->static_prio - MAX_RT_PRIO;
836 struct load_weight *load = &p->se.load;
839 * SCHED_IDLE tasks get minimal weight:
841 if (p->policy == SCHED_IDLE) {
842 load->weight = scale_load(WEIGHT_IDLEPRIO);
843 load->inv_weight = WMULT_IDLEPRIO;
847 load->weight = scale_load(prio_to_weight[prio]);
848 load->inv_weight = prio_to_wmult[prio];
851 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
854 sched_info_queued(rq, p);
855 p->sched_class->enqueue_task(rq, p, flags);
858 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
861 sched_info_dequeued(rq, p);
862 p->sched_class->dequeue_task(rq, p, flags);
865 void activate_task(struct rq *rq, struct task_struct *p, int flags)
867 if (task_contributes_to_load(p))
868 rq->nr_uninterruptible--;
870 enqueue_task(rq, p, flags);
873 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
875 if (task_contributes_to_load(p))
876 rq->nr_uninterruptible++;
878 dequeue_task(rq, p, flags);
881 static void update_rq_clock_task(struct rq *rq, s64 delta)
884 * In theory, the compile should just see 0 here, and optimize out the call
885 * to sched_rt_avg_update. But I don't trust it...
887 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
888 s64 steal = 0, irq_delta = 0;
890 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
891 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
894 * Since irq_time is only updated on {soft,}irq_exit, we might run into
895 * this case when a previous update_rq_clock() happened inside a
898 * When this happens, we stop ->clock_task and only update the
899 * prev_irq_time stamp to account for the part that fit, so that a next
900 * update will consume the rest. This ensures ->clock_task is
903 * It does however cause some slight miss-attribution of {soft,}irq
904 * time, a more accurate solution would be to update the irq_time using
905 * the current rq->clock timestamp, except that would require using
908 if (irq_delta > delta)
911 rq->prev_irq_time += irq_delta;
914 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
915 if (static_key_false((¶virt_steal_rq_enabled))) {
916 steal = paravirt_steal_clock(cpu_of(rq));
917 steal -= rq->prev_steal_time_rq;
919 if (unlikely(steal > delta))
922 rq->prev_steal_time_rq += steal;
927 rq->clock_task += delta;
929 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
930 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
931 sched_rt_avg_update(rq, irq_delta + steal);
935 void sched_set_stop_task(int cpu, struct task_struct *stop)
937 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
938 struct task_struct *old_stop = cpu_rq(cpu)->stop;
942 * Make it appear like a SCHED_FIFO task, its something
943 * userspace knows about and won't get confused about.
945 * Also, it will make PI more or less work without too
946 * much confusion -- but then, stop work should not
947 * rely on PI working anyway.
949 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
951 stop->sched_class = &stop_sched_class;
954 cpu_rq(cpu)->stop = stop;
958 * Reset it back to a normal scheduling class so that
959 * it can die in pieces.
961 old_stop->sched_class = &rt_sched_class;
966 * __normal_prio - return the priority that is based on the static prio
968 static inline int __normal_prio(struct task_struct *p)
970 return p->static_prio;
974 * Calculate the expected normal priority: i.e. priority
975 * without taking RT-inheritance into account. Might be
976 * boosted by interactivity modifiers. Changes upon fork,
977 * setprio syscalls, and whenever the interactivity
978 * estimator recalculates.
980 static inline int normal_prio(struct task_struct *p)
984 if (task_has_dl_policy(p))
985 prio = MAX_DL_PRIO-1;
986 else if (task_has_rt_policy(p))
987 prio = MAX_RT_PRIO-1 - p->rt_priority;
989 prio = __normal_prio(p);
994 * Calculate the current priority, i.e. the priority
995 * taken into account by the scheduler. This value might
996 * be boosted by RT tasks, or might be boosted by
997 * interactivity modifiers. Will be RT if the task got
998 * RT-boosted. If not then it returns p->normal_prio.
1000 static int effective_prio(struct task_struct *p)
1002 p->normal_prio = normal_prio(p);
1004 * If we are RT tasks or we were boosted to RT priority,
1005 * keep the priority unchanged. Otherwise, update priority
1006 * to the normal priority:
1008 if (!rt_prio(p->prio))
1009 return p->normal_prio;
1014 * task_curr - is this task currently executing on a CPU?
1015 * @p: the task in question.
1017 * Return: 1 if the task is currently executing. 0 otherwise.
1019 inline int task_curr(const struct task_struct *p)
1021 return cpu_curr(task_cpu(p)) == p;
1025 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
1027 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1028 const struct sched_class *prev_class,
1031 if (prev_class != p->sched_class) {
1032 if (prev_class->switched_from)
1033 prev_class->switched_from(rq, p);
1034 /* Possble rq->lock 'hole'. */
1035 p->sched_class->switched_to(rq, p);
1036 } else if (oldprio != p->prio || dl_task(p))
1037 p->sched_class->prio_changed(rq, p, oldprio);
1040 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1042 const struct sched_class *class;
1044 if (p->sched_class == rq->curr->sched_class) {
1045 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1047 for_each_class(class) {
1048 if (class == rq->curr->sched_class)
1050 if (class == p->sched_class) {
1058 * A queue event has occurred, and we're going to schedule. In
1059 * this case, we can save a useless back to back clock update.
1061 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1062 rq_clock_skip_update(rq, true);
1066 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1068 #ifdef CONFIG_SCHED_DEBUG
1070 * We should never call set_task_cpu() on a blocked task,
1071 * ttwu() will sort out the placement.
1073 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1076 #ifdef CONFIG_LOCKDEP
1078 * The caller should hold either p->pi_lock or rq->lock, when changing
1079 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1081 * sched_move_task() holds both and thus holding either pins the cgroup,
1084 * Furthermore, all task_rq users should acquire both locks, see
1087 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1088 lockdep_is_held(&task_rq(p)->lock)));
1092 trace_sched_migrate_task(p, new_cpu);
1094 if (task_cpu(p) != new_cpu) {
1095 if (p->sched_class->migrate_task_rq)
1096 p->sched_class->migrate_task_rq(p, new_cpu);
1097 p->se.nr_migrations++;
1098 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1101 __set_task_cpu(p, new_cpu);
1104 static void __migrate_swap_task(struct task_struct *p, int cpu)
1106 if (task_on_rq_queued(p)) {
1107 struct rq *src_rq, *dst_rq;
1109 src_rq = task_rq(p);
1110 dst_rq = cpu_rq(cpu);
1112 deactivate_task(src_rq, p, 0);
1113 set_task_cpu(p, cpu);
1114 activate_task(dst_rq, p, 0);
1115 check_preempt_curr(dst_rq, p, 0);
1118 * Task isn't running anymore; make it appear like we migrated
1119 * it before it went to sleep. This means on wakeup we make the
1120 * previous cpu our targer instead of where it really is.
1126 struct migration_swap_arg {
1127 struct task_struct *src_task, *dst_task;
1128 int src_cpu, dst_cpu;
1131 static int migrate_swap_stop(void *data)
1133 struct migration_swap_arg *arg = data;
1134 struct rq *src_rq, *dst_rq;
1137 src_rq = cpu_rq(arg->src_cpu);
1138 dst_rq = cpu_rq(arg->dst_cpu);
1140 double_raw_lock(&arg->src_task->pi_lock,
1141 &arg->dst_task->pi_lock);
1142 double_rq_lock(src_rq, dst_rq);
1143 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1146 if (task_cpu(arg->src_task) != arg->src_cpu)
1149 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1152 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1155 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1156 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1161 double_rq_unlock(src_rq, dst_rq);
1162 raw_spin_unlock(&arg->dst_task->pi_lock);
1163 raw_spin_unlock(&arg->src_task->pi_lock);
1169 * Cross migrate two tasks
1171 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1173 struct migration_swap_arg arg;
1176 arg = (struct migration_swap_arg){
1178 .src_cpu = task_cpu(cur),
1180 .dst_cpu = task_cpu(p),
1183 if (arg.src_cpu == arg.dst_cpu)
1187 * These three tests are all lockless; this is OK since all of them
1188 * will be re-checked with proper locks held further down the line.
1190 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1193 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1196 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1199 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1200 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1206 struct migration_arg {
1207 struct task_struct *task;
1211 static int migration_cpu_stop(void *data);
1214 * wait_task_inactive - wait for a thread to unschedule.
1216 * If @match_state is nonzero, it's the @p->state value just checked and
1217 * not expected to change. If it changes, i.e. @p might have woken up,
1218 * then return zero. When we succeed in waiting for @p to be off its CPU,
1219 * we return a positive number (its total switch count). If a second call
1220 * a short while later returns the same number, the caller can be sure that
1221 * @p has remained unscheduled the whole time.
1223 * The caller must ensure that the task *will* unschedule sometime soon,
1224 * else this function might spin for a *long* time. This function can't
1225 * be called with interrupts off, or it may introduce deadlock with
1226 * smp_call_function() if an IPI is sent by the same process we are
1227 * waiting to become inactive.
1229 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1231 unsigned long flags;
1232 int running, queued;
1238 * We do the initial early heuristics without holding
1239 * any task-queue locks at all. We'll only try to get
1240 * the runqueue lock when things look like they will
1246 * If the task is actively running on another CPU
1247 * still, just relax and busy-wait without holding
1250 * NOTE! Since we don't hold any locks, it's not
1251 * even sure that "rq" stays as the right runqueue!
1252 * But we don't care, since "task_running()" will
1253 * return false if the runqueue has changed and p
1254 * is actually now running somewhere else!
1256 while (task_running(rq, p)) {
1257 if (match_state && unlikely(p->state != match_state))
1263 * Ok, time to look more closely! We need the rq
1264 * lock now, to be *sure*. If we're wrong, we'll
1265 * just go back and repeat.
1267 rq = task_rq_lock(p, &flags);
1268 trace_sched_wait_task(p);
1269 running = task_running(rq, p);
1270 queued = task_on_rq_queued(p);
1272 if (!match_state || p->state == match_state)
1273 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1274 task_rq_unlock(rq, p, &flags);
1277 * If it changed from the expected state, bail out now.
1279 if (unlikely(!ncsw))
1283 * Was it really running after all now that we
1284 * checked with the proper locks actually held?
1286 * Oops. Go back and try again..
1288 if (unlikely(running)) {
1294 * It's not enough that it's not actively running,
1295 * it must be off the runqueue _entirely_, and not
1298 * So if it was still runnable (but just not actively
1299 * running right now), it's preempted, and we should
1300 * yield - it could be a while.
1302 if (unlikely(queued)) {
1303 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1305 set_current_state(TASK_UNINTERRUPTIBLE);
1306 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1311 * Ahh, all good. It wasn't running, and it wasn't
1312 * runnable, which means that it will never become
1313 * running in the future either. We're all done!
1322 * kick_process - kick a running thread to enter/exit the kernel
1323 * @p: the to-be-kicked thread
1325 * Cause a process which is running on another CPU to enter
1326 * kernel-mode, without any delay. (to get signals handled.)
1328 * NOTE: this function doesn't have to take the runqueue lock,
1329 * because all it wants to ensure is that the remote task enters
1330 * the kernel. If the IPI races and the task has been migrated
1331 * to another CPU then no harm is done and the purpose has been
1334 void kick_process(struct task_struct *p)
1340 if ((cpu != smp_processor_id()) && task_curr(p))
1341 smp_send_reschedule(cpu);
1344 EXPORT_SYMBOL_GPL(kick_process);
1345 #endif /* CONFIG_SMP */
1349 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1351 static int select_fallback_rq(int cpu, struct task_struct *p)
1353 int nid = cpu_to_node(cpu);
1354 const struct cpumask *nodemask = NULL;
1355 enum { cpuset, possible, fail } state = cpuset;
1359 * If the node that the cpu is on has been offlined, cpu_to_node()
1360 * will return -1. There is no cpu on the node, and we should
1361 * select the cpu on the other node.
1364 nodemask = cpumask_of_node(nid);
1366 /* Look for allowed, online CPU in same node. */
1367 for_each_cpu(dest_cpu, nodemask) {
1368 if (!cpu_online(dest_cpu))
1370 if (!cpu_active(dest_cpu))
1372 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1378 /* Any allowed, online CPU? */
1379 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1380 if (!cpu_online(dest_cpu))
1382 if (!cpu_active(dest_cpu))
1389 /* No more Mr. Nice Guy. */
1390 cpuset_cpus_allowed_fallback(p);
1395 do_set_cpus_allowed(p, cpu_possible_mask);
1406 if (state != cpuset) {
1408 * Don't tell them about moving exiting tasks or
1409 * kernel threads (both mm NULL), since they never
1412 if (p->mm && printk_ratelimit()) {
1413 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1414 task_pid_nr(p), p->comm, cpu);
1422 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1425 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1427 if (p->nr_cpus_allowed > 1)
1428 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1431 * In order not to call set_task_cpu() on a blocking task we need
1432 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1435 * Since this is common to all placement strategies, this lives here.
1437 * [ this allows ->select_task() to simply return task_cpu(p) and
1438 * not worry about this generic constraint ]
1440 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1442 cpu = select_fallback_rq(task_cpu(p), p);
1447 static void update_avg(u64 *avg, u64 sample)
1449 s64 diff = sample - *avg;
1455 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1457 #ifdef CONFIG_SCHEDSTATS
1458 struct rq *rq = this_rq();
1461 int this_cpu = smp_processor_id();
1463 if (cpu == this_cpu) {
1464 schedstat_inc(rq, ttwu_local);
1465 schedstat_inc(p, se.statistics.nr_wakeups_local);
1467 struct sched_domain *sd;
1469 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1471 for_each_domain(this_cpu, sd) {
1472 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1473 schedstat_inc(sd, ttwu_wake_remote);
1480 if (wake_flags & WF_MIGRATED)
1481 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1483 #endif /* CONFIG_SMP */
1485 schedstat_inc(rq, ttwu_count);
1486 schedstat_inc(p, se.statistics.nr_wakeups);
1488 if (wake_flags & WF_SYNC)
1489 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1491 #endif /* CONFIG_SCHEDSTATS */
1494 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1496 activate_task(rq, p, en_flags);
1497 p->on_rq = TASK_ON_RQ_QUEUED;
1499 /* if a worker is waking up, notify workqueue */
1500 if (p->flags & PF_WQ_WORKER)
1501 wq_worker_waking_up(p, cpu_of(rq));
1505 * Mark the task runnable and perform wakeup-preemption.
1508 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1510 check_preempt_curr(rq, p, wake_flags);
1511 trace_sched_wakeup(p, true);
1513 p->state = TASK_RUNNING;
1515 if (p->sched_class->task_woken)
1516 p->sched_class->task_woken(rq, p);
1518 if (rq->idle_stamp) {
1519 u64 delta = rq_clock(rq) - rq->idle_stamp;
1520 u64 max = 2*rq->max_idle_balance_cost;
1522 update_avg(&rq->avg_idle, delta);
1524 if (rq->avg_idle > max)
1533 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1536 if (p->sched_contributes_to_load)
1537 rq->nr_uninterruptible--;
1540 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1541 ttwu_do_wakeup(rq, p, wake_flags);
1545 * Called in case the task @p isn't fully descheduled from its runqueue,
1546 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1547 * since all we need to do is flip p->state to TASK_RUNNING, since
1548 * the task is still ->on_rq.
1550 static int ttwu_remote(struct task_struct *p, int wake_flags)
1555 rq = __task_rq_lock(p);
1556 if (task_on_rq_queued(p)) {
1557 /* check_preempt_curr() may use rq clock */
1558 update_rq_clock(rq);
1559 ttwu_do_wakeup(rq, p, wake_flags);
1562 __task_rq_unlock(rq);
1568 void sched_ttwu_pending(void)
1570 struct rq *rq = this_rq();
1571 struct llist_node *llist = llist_del_all(&rq->wake_list);
1572 struct task_struct *p;
1573 unsigned long flags;
1578 raw_spin_lock_irqsave(&rq->lock, flags);
1581 p = llist_entry(llist, struct task_struct, wake_entry);
1582 llist = llist_next(llist);
1583 ttwu_do_activate(rq, p, 0);
1586 raw_spin_unlock_irqrestore(&rq->lock, flags);
1589 void scheduler_ipi(void)
1592 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1593 * TIF_NEED_RESCHED remotely (for the first time) will also send
1596 preempt_fold_need_resched();
1598 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1602 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1603 * traditionally all their work was done from the interrupt return
1604 * path. Now that we actually do some work, we need to make sure
1607 * Some archs already do call them, luckily irq_enter/exit nest
1610 * Arguably we should visit all archs and update all handlers,
1611 * however a fair share of IPIs are still resched only so this would
1612 * somewhat pessimize the simple resched case.
1615 sched_ttwu_pending();
1618 * Check if someone kicked us for doing the nohz idle load balance.
1620 if (unlikely(got_nohz_idle_kick())) {
1621 this_rq()->idle_balance = 1;
1622 raise_softirq_irqoff(SCHED_SOFTIRQ);
1627 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1629 struct rq *rq = cpu_rq(cpu);
1631 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1632 if (!set_nr_if_polling(rq->idle))
1633 smp_send_reschedule(cpu);
1635 trace_sched_wake_idle_without_ipi(cpu);
1639 void wake_up_if_idle(int cpu)
1641 struct rq *rq = cpu_rq(cpu);
1642 unsigned long flags;
1646 if (!is_idle_task(rcu_dereference(rq->curr)))
1649 if (set_nr_if_polling(rq->idle)) {
1650 trace_sched_wake_idle_without_ipi(cpu);
1652 raw_spin_lock_irqsave(&rq->lock, flags);
1653 if (is_idle_task(rq->curr))
1654 smp_send_reschedule(cpu);
1655 /* Else cpu is not in idle, do nothing here */
1656 raw_spin_unlock_irqrestore(&rq->lock, flags);
1663 bool cpus_share_cache(int this_cpu, int that_cpu)
1665 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1667 #endif /* CONFIG_SMP */
1669 static void ttwu_queue(struct task_struct *p, int cpu)
1671 struct rq *rq = cpu_rq(cpu);
1673 #if defined(CONFIG_SMP)
1674 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1675 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1676 ttwu_queue_remote(p, cpu);
1681 raw_spin_lock(&rq->lock);
1682 ttwu_do_activate(rq, p, 0);
1683 raw_spin_unlock(&rq->lock);
1687 * try_to_wake_up - wake up a thread
1688 * @p: the thread to be awakened
1689 * @state: the mask of task states that can be woken
1690 * @wake_flags: wake modifier flags (WF_*)
1692 * Put it on the run-queue if it's not already there. The "current"
1693 * thread is always on the run-queue (except when the actual
1694 * re-schedule is in progress), and as such you're allowed to do
1695 * the simpler "current->state = TASK_RUNNING" to mark yourself
1696 * runnable without the overhead of this.
1698 * Return: %true if @p was woken up, %false if it was already running.
1699 * or @state didn't match @p's state.
1702 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1704 unsigned long flags;
1705 int cpu, success = 0;
1708 * If we are going to wake up a thread waiting for CONDITION we
1709 * need to ensure that CONDITION=1 done by the caller can not be
1710 * reordered with p->state check below. This pairs with mb() in
1711 * set_current_state() the waiting thread does.
1713 smp_mb__before_spinlock();
1714 raw_spin_lock_irqsave(&p->pi_lock, flags);
1715 if (!(p->state & state))
1718 success = 1; /* we're going to change ->state */
1721 if (p->on_rq && ttwu_remote(p, wake_flags))
1726 * If the owning (remote) cpu is still in the middle of schedule() with
1727 * this task as prev, wait until its done referencing the task.
1732 * Pairs with the smp_wmb() in finish_lock_switch().
1736 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1737 p->state = TASK_WAKING;
1739 if (p->sched_class->task_waking)
1740 p->sched_class->task_waking(p);
1742 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1743 if (task_cpu(p) != cpu) {
1744 wake_flags |= WF_MIGRATED;
1745 set_task_cpu(p, cpu);
1747 #endif /* CONFIG_SMP */
1751 ttwu_stat(p, cpu, wake_flags);
1753 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1759 * try_to_wake_up_local - try to wake up a local task with rq lock held
1760 * @p: the thread to be awakened
1762 * Put @p on the run-queue if it's not already there. The caller must
1763 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1766 static void try_to_wake_up_local(struct task_struct *p)
1768 struct rq *rq = task_rq(p);
1770 if (WARN_ON_ONCE(rq != this_rq()) ||
1771 WARN_ON_ONCE(p == current))
1774 lockdep_assert_held(&rq->lock);
1776 if (!raw_spin_trylock(&p->pi_lock)) {
1777 raw_spin_unlock(&rq->lock);
1778 raw_spin_lock(&p->pi_lock);
1779 raw_spin_lock(&rq->lock);
1782 if (!(p->state & TASK_NORMAL))
1785 if (!task_on_rq_queued(p))
1786 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1788 ttwu_do_wakeup(rq, p, 0);
1789 ttwu_stat(p, smp_processor_id(), 0);
1791 raw_spin_unlock(&p->pi_lock);
1795 * wake_up_process - Wake up a specific process
1796 * @p: The process to be woken up.
1798 * Attempt to wake up the nominated process and move it to the set of runnable
1801 * Return: 1 if the process was woken up, 0 if it was already running.
1803 * It may be assumed that this function implies a write memory barrier before
1804 * changing the task state if and only if any tasks are woken up.
1806 int wake_up_process(struct task_struct *p)
1808 WARN_ON(task_is_stopped_or_traced(p));
1809 return try_to_wake_up(p, TASK_NORMAL, 0);
1811 EXPORT_SYMBOL(wake_up_process);
1813 int wake_up_state(struct task_struct *p, unsigned int state)
1815 return try_to_wake_up(p, state, 0);
1819 * This function clears the sched_dl_entity static params.
1821 void __dl_clear_params(struct task_struct *p)
1823 struct sched_dl_entity *dl_se = &p->dl;
1825 dl_se->dl_runtime = 0;
1826 dl_se->dl_deadline = 0;
1827 dl_se->dl_period = 0;
1831 dl_se->dl_throttled = 0;
1833 dl_se->dl_yielded = 0;
1837 * Perform scheduler related setup for a newly forked process p.
1838 * p is forked by current.
1840 * __sched_fork() is basic setup used by init_idle() too:
1842 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1847 p->se.exec_start = 0;
1848 p->se.sum_exec_runtime = 0;
1849 p->se.prev_sum_exec_runtime = 0;
1850 p->se.nr_migrations = 0;
1853 p->se.avg.decay_count = 0;
1855 INIT_LIST_HEAD(&p->se.group_node);
1857 #ifdef CONFIG_SCHEDSTATS
1858 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1861 RB_CLEAR_NODE(&p->dl.rb_node);
1862 init_dl_task_timer(&p->dl);
1863 __dl_clear_params(p);
1865 INIT_LIST_HEAD(&p->rt.run_list);
1867 #ifdef CONFIG_PREEMPT_NOTIFIERS
1868 INIT_HLIST_HEAD(&p->preempt_notifiers);
1871 #ifdef CONFIG_NUMA_BALANCING
1872 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1873 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1874 p->mm->numa_scan_seq = 0;
1877 if (clone_flags & CLONE_VM)
1878 p->numa_preferred_nid = current->numa_preferred_nid;
1880 p->numa_preferred_nid = -1;
1882 p->node_stamp = 0ULL;
1883 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1884 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1885 p->numa_work.next = &p->numa_work;
1886 p->numa_faults = NULL;
1887 p->last_task_numa_placement = 0;
1888 p->last_sum_exec_runtime = 0;
1890 p->numa_group = NULL;
1891 #endif /* CONFIG_NUMA_BALANCING */
1894 #ifdef CONFIG_NUMA_BALANCING
1895 #ifdef CONFIG_SCHED_DEBUG
1896 void set_numabalancing_state(bool enabled)
1899 sched_feat_set("NUMA");
1901 sched_feat_set("NO_NUMA");
1904 __read_mostly bool numabalancing_enabled;
1906 void set_numabalancing_state(bool enabled)
1908 numabalancing_enabled = enabled;
1910 #endif /* CONFIG_SCHED_DEBUG */
1912 #ifdef CONFIG_PROC_SYSCTL
1913 int sysctl_numa_balancing(struct ctl_table *table, int write,
1914 void __user *buffer, size_t *lenp, loff_t *ppos)
1918 int state = numabalancing_enabled;
1920 if (write && !capable(CAP_SYS_ADMIN))
1925 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1929 set_numabalancing_state(state);
1936 * fork()/clone()-time setup:
1938 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1940 unsigned long flags;
1941 int cpu = get_cpu();
1943 __sched_fork(clone_flags, p);
1945 * We mark the process as running here. This guarantees that
1946 * nobody will actually run it, and a signal or other external
1947 * event cannot wake it up and insert it on the runqueue either.
1949 p->state = TASK_RUNNING;
1952 * Make sure we do not leak PI boosting priority to the child.
1954 p->prio = current->normal_prio;
1957 * Revert to default priority/policy on fork if requested.
1959 if (unlikely(p->sched_reset_on_fork)) {
1960 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1961 p->policy = SCHED_NORMAL;
1962 p->static_prio = NICE_TO_PRIO(0);
1964 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1965 p->static_prio = NICE_TO_PRIO(0);
1967 p->prio = p->normal_prio = __normal_prio(p);
1971 * We don't need the reset flag anymore after the fork. It has
1972 * fulfilled its duty:
1974 p->sched_reset_on_fork = 0;
1977 if (dl_prio(p->prio)) {
1980 } else if (rt_prio(p->prio)) {
1981 p->sched_class = &rt_sched_class;
1983 p->sched_class = &fair_sched_class;
1986 if (p->sched_class->task_fork)
1987 p->sched_class->task_fork(p);
1990 * The child is not yet in the pid-hash so no cgroup attach races,
1991 * and the cgroup is pinned to this child due to cgroup_fork()
1992 * is ran before sched_fork().
1994 * Silence PROVE_RCU.
1996 raw_spin_lock_irqsave(&p->pi_lock, flags);
1997 set_task_cpu(p, cpu);
1998 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2000 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2001 if (likely(sched_info_on()))
2002 memset(&p->sched_info, 0, sizeof(p->sched_info));
2004 #if defined(CONFIG_SMP)
2007 init_task_preempt_count(p);
2009 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2010 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2017 unsigned long to_ratio(u64 period, u64 runtime)
2019 if (runtime == RUNTIME_INF)
2023 * Doing this here saves a lot of checks in all
2024 * the calling paths, and returning zero seems
2025 * safe for them anyway.
2030 return div64_u64(runtime << 20, period);
2034 inline struct dl_bw *dl_bw_of(int i)
2036 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2037 "sched RCU must be held");
2038 return &cpu_rq(i)->rd->dl_bw;
2041 static inline int dl_bw_cpus(int i)
2043 struct root_domain *rd = cpu_rq(i)->rd;
2046 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2047 "sched RCU must be held");
2048 for_each_cpu_and(i, rd->span, cpu_active_mask)
2054 inline struct dl_bw *dl_bw_of(int i)
2056 return &cpu_rq(i)->dl.dl_bw;
2059 static inline int dl_bw_cpus(int i)
2066 * We must be sure that accepting a new task (or allowing changing the
2067 * parameters of an existing one) is consistent with the bandwidth
2068 * constraints. If yes, this function also accordingly updates the currently
2069 * allocated bandwidth to reflect the new situation.
2071 * This function is called while holding p's rq->lock.
2073 * XXX we should delay bw change until the task's 0-lag point, see
2076 static int dl_overflow(struct task_struct *p, int policy,
2077 const struct sched_attr *attr)
2080 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2081 u64 period = attr->sched_period ?: attr->sched_deadline;
2082 u64 runtime = attr->sched_runtime;
2083 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2086 if (new_bw == p->dl.dl_bw)
2090 * Either if a task, enters, leave, or stays -deadline but changes
2091 * its parameters, we may need to update accordingly the total
2092 * allocated bandwidth of the container.
2094 raw_spin_lock(&dl_b->lock);
2095 cpus = dl_bw_cpus(task_cpu(p));
2096 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2097 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2098 __dl_add(dl_b, new_bw);
2100 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2101 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2102 __dl_clear(dl_b, p->dl.dl_bw);
2103 __dl_add(dl_b, new_bw);
2105 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2106 __dl_clear(dl_b, p->dl.dl_bw);
2109 raw_spin_unlock(&dl_b->lock);
2114 extern void init_dl_bw(struct dl_bw *dl_b);
2117 * wake_up_new_task - wake up a newly created task for the first time.
2119 * This function will do some initial scheduler statistics housekeeping
2120 * that must be done for every newly created context, then puts the task
2121 * on the runqueue and wakes it.
2123 void wake_up_new_task(struct task_struct *p)
2125 unsigned long flags;
2128 raw_spin_lock_irqsave(&p->pi_lock, flags);
2131 * Fork balancing, do it here and not earlier because:
2132 * - cpus_allowed can change in the fork path
2133 * - any previously selected cpu might disappear through hotplug
2135 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2138 /* Initialize new task's runnable average */
2139 init_task_runnable_average(p);
2140 rq = __task_rq_lock(p);
2141 activate_task(rq, p, 0);
2142 p->on_rq = TASK_ON_RQ_QUEUED;
2143 trace_sched_wakeup_new(p, true);
2144 check_preempt_curr(rq, p, WF_FORK);
2146 if (p->sched_class->task_woken)
2147 p->sched_class->task_woken(rq, p);
2149 task_rq_unlock(rq, p, &flags);
2152 #ifdef CONFIG_PREEMPT_NOTIFIERS
2155 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2156 * @notifier: notifier struct to register
2158 void preempt_notifier_register(struct preempt_notifier *notifier)
2160 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2162 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2165 * preempt_notifier_unregister - no longer interested in preemption notifications
2166 * @notifier: notifier struct to unregister
2168 * This is safe to call from within a preemption notifier.
2170 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2172 hlist_del(¬ifier->link);
2174 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2176 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2178 struct preempt_notifier *notifier;
2180 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2181 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2185 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2186 struct task_struct *next)
2188 struct preempt_notifier *notifier;
2190 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2191 notifier->ops->sched_out(notifier, next);
2194 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2196 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2201 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2202 struct task_struct *next)
2206 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2209 * prepare_task_switch - prepare to switch tasks
2210 * @rq: the runqueue preparing to switch
2211 * @prev: the current task that is being switched out
2212 * @next: the task we are going to switch to.
2214 * This is called with the rq lock held and interrupts off. It must
2215 * be paired with a subsequent finish_task_switch after the context
2218 * prepare_task_switch sets up locking and calls architecture specific
2222 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2223 struct task_struct *next)
2225 trace_sched_switch(prev, next);
2226 sched_info_switch(rq, prev, next);
2227 perf_event_task_sched_out(prev, next);
2228 fire_sched_out_preempt_notifiers(prev, next);
2229 prepare_lock_switch(rq, next);
2230 prepare_arch_switch(next);
2234 * finish_task_switch - clean up after a task-switch
2235 * @prev: the thread we just switched away from.
2237 * finish_task_switch must be called after the context switch, paired
2238 * with a prepare_task_switch call before the context switch.
2239 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2240 * and do any other architecture-specific cleanup actions.
2242 * Note that we may have delayed dropping an mm in context_switch(). If
2243 * so, we finish that here outside of the runqueue lock. (Doing it
2244 * with the lock held can cause deadlocks; see schedule() for
2247 * The context switch have flipped the stack from under us and restored the
2248 * local variables which were saved when this task called schedule() in the
2249 * past. prev == current is still correct but we need to recalculate this_rq
2250 * because prev may have moved to another CPU.
2252 static struct rq *finish_task_switch(struct task_struct *prev)
2253 __releases(rq->lock)
2255 struct rq *rq = this_rq();
2256 struct mm_struct *mm = rq->prev_mm;
2262 * A task struct has one reference for the use as "current".
2263 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2264 * schedule one last time. The schedule call will never return, and
2265 * the scheduled task must drop that reference.
2266 * The test for TASK_DEAD must occur while the runqueue locks are
2267 * still held, otherwise prev could be scheduled on another cpu, die
2268 * there before we look at prev->state, and then the reference would
2270 * Manfred Spraul <manfred@colorfullife.com>
2272 prev_state = prev->state;
2273 vtime_task_switch(prev);
2274 finish_arch_switch(prev);
2275 perf_event_task_sched_in(prev, current);
2276 finish_lock_switch(rq, prev);
2277 finish_arch_post_lock_switch();
2279 fire_sched_in_preempt_notifiers(current);
2282 if (unlikely(prev_state == TASK_DEAD)) {
2283 if (prev->sched_class->task_dead)
2284 prev->sched_class->task_dead(prev);
2287 * Remove function-return probe instances associated with this
2288 * task and put them back on the free list.
2290 kprobe_flush_task(prev);
2291 put_task_struct(prev);
2294 tick_nohz_task_switch(current);
2300 /* rq->lock is NOT held, but preemption is disabled */
2301 static inline void post_schedule(struct rq *rq)
2303 if (rq->post_schedule) {
2304 unsigned long flags;
2306 raw_spin_lock_irqsave(&rq->lock, flags);
2307 if (rq->curr->sched_class->post_schedule)
2308 rq->curr->sched_class->post_schedule(rq);
2309 raw_spin_unlock_irqrestore(&rq->lock, flags);
2311 rq->post_schedule = 0;
2317 static inline void post_schedule(struct rq *rq)
2324 * schedule_tail - first thing a freshly forked thread must call.
2325 * @prev: the thread we just switched away from.
2327 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2328 __releases(rq->lock)
2332 /* finish_task_switch() drops rq->lock and enables preemtion */
2334 rq = finish_task_switch(prev);
2338 if (current->set_child_tid)
2339 put_user(task_pid_vnr(current), current->set_child_tid);
2343 * context_switch - switch to the new MM and the new thread's register state.
2345 static inline struct rq *
2346 context_switch(struct rq *rq, struct task_struct *prev,
2347 struct task_struct *next)
2349 struct mm_struct *mm, *oldmm;
2351 prepare_task_switch(rq, prev, next);
2354 oldmm = prev->active_mm;
2356 * For paravirt, this is coupled with an exit in switch_to to
2357 * combine the page table reload and the switch backend into
2360 arch_start_context_switch(prev);
2363 next->active_mm = oldmm;
2364 atomic_inc(&oldmm->mm_count);
2365 enter_lazy_tlb(oldmm, next);
2367 switch_mm(oldmm, mm, next);
2370 prev->active_mm = NULL;
2371 rq->prev_mm = oldmm;
2374 * Since the runqueue lock will be released by the next
2375 * task (which is an invalid locking op but in the case
2376 * of the scheduler it's an obvious special-case), so we
2377 * do an early lockdep release here:
2379 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2381 context_tracking_task_switch(prev, next);
2382 /* Here we just switch the register state and the stack. */
2383 switch_to(prev, next, prev);
2386 return finish_task_switch(prev);
2390 * nr_running and nr_context_switches:
2392 * externally visible scheduler statistics: current number of runnable
2393 * threads, total number of context switches performed since bootup.
2395 unsigned long nr_running(void)
2397 unsigned long i, sum = 0;
2399 for_each_online_cpu(i)
2400 sum += cpu_rq(i)->nr_running;
2406 * Check if only the current task is running on the cpu.
2408 bool single_task_running(void)
2410 if (cpu_rq(smp_processor_id())->nr_running == 1)
2415 EXPORT_SYMBOL(single_task_running);
2417 unsigned long long nr_context_switches(void)
2420 unsigned long long sum = 0;
2422 for_each_possible_cpu(i)
2423 sum += cpu_rq(i)->nr_switches;
2428 unsigned long nr_iowait(void)
2430 unsigned long i, sum = 0;
2432 for_each_possible_cpu(i)
2433 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2438 unsigned long nr_iowait_cpu(int cpu)
2440 struct rq *this = cpu_rq(cpu);
2441 return atomic_read(&this->nr_iowait);
2444 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2446 struct rq *rq = this_rq();
2447 *nr_waiters = atomic_read(&rq->nr_iowait);
2448 *load = rq->load.weight;
2454 * sched_exec - execve() is a valuable balancing opportunity, because at
2455 * this point the task has the smallest effective memory and cache footprint.
2457 void sched_exec(void)
2459 struct task_struct *p = current;
2460 unsigned long flags;
2463 raw_spin_lock_irqsave(&p->pi_lock, flags);
2464 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2465 if (dest_cpu == smp_processor_id())
2468 if (likely(cpu_active(dest_cpu))) {
2469 struct migration_arg arg = { p, dest_cpu };
2471 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2472 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2476 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2481 DEFINE_PER_CPU(struct kernel_stat, kstat);
2482 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2484 EXPORT_PER_CPU_SYMBOL(kstat);
2485 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2488 * Return accounted runtime for the task.
2489 * In case the task is currently running, return the runtime plus current's
2490 * pending runtime that have not been accounted yet.
2492 unsigned long long task_sched_runtime(struct task_struct *p)
2494 unsigned long flags;
2498 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2500 * 64-bit doesn't need locks to atomically read a 64bit value.
2501 * So we have a optimization chance when the task's delta_exec is 0.
2502 * Reading ->on_cpu is racy, but this is ok.
2504 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2505 * If we race with it entering cpu, unaccounted time is 0. This is
2506 * indistinguishable from the read occurring a few cycles earlier.
2507 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2508 * been accounted, so we're correct here as well.
2510 if (!p->on_cpu || !task_on_rq_queued(p))
2511 return p->se.sum_exec_runtime;
2514 rq = task_rq_lock(p, &flags);
2516 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2517 * project cycles that may never be accounted to this
2518 * thread, breaking clock_gettime().
2520 if (task_current(rq, p) && task_on_rq_queued(p)) {
2521 update_rq_clock(rq);
2522 p->sched_class->update_curr(rq);
2524 ns = p->se.sum_exec_runtime;
2525 task_rq_unlock(rq, p, &flags);
2531 * This function gets called by the timer code, with HZ frequency.
2532 * We call it with interrupts disabled.
2534 void scheduler_tick(void)
2536 int cpu = smp_processor_id();
2537 struct rq *rq = cpu_rq(cpu);
2538 struct task_struct *curr = rq->curr;
2542 raw_spin_lock(&rq->lock);
2543 update_rq_clock(rq);
2544 curr->sched_class->task_tick(rq, curr, 0);
2545 update_cpu_load_active(rq);
2546 calc_global_load_tick(rq);
2547 raw_spin_unlock(&rq->lock);
2549 perf_event_task_tick();
2552 rq->idle_balance = idle_cpu(cpu);
2553 trigger_load_balance(rq);
2555 rq_last_tick_reset(rq);
2558 #ifdef CONFIG_NO_HZ_FULL
2560 * scheduler_tick_max_deferment
2562 * Keep at least one tick per second when a single
2563 * active task is running because the scheduler doesn't
2564 * yet completely support full dynticks environment.
2566 * This makes sure that uptime, CFS vruntime, load
2567 * balancing, etc... continue to move forward, even
2568 * with a very low granularity.
2570 * Return: Maximum deferment in nanoseconds.
2572 u64 scheduler_tick_max_deferment(void)
2574 struct rq *rq = this_rq();
2575 unsigned long next, now = READ_ONCE(jiffies);
2577 next = rq->last_sched_tick + HZ;
2579 if (time_before_eq(next, now))
2582 return jiffies_to_nsecs(next - now);
2586 notrace unsigned long get_parent_ip(unsigned long addr)
2588 if (in_lock_functions(addr)) {
2589 addr = CALLER_ADDR2;
2590 if (in_lock_functions(addr))
2591 addr = CALLER_ADDR3;
2596 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2597 defined(CONFIG_PREEMPT_TRACER))
2599 void preempt_count_add(int val)
2601 #ifdef CONFIG_DEBUG_PREEMPT
2605 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2608 __preempt_count_add(val);
2609 #ifdef CONFIG_DEBUG_PREEMPT
2611 * Spinlock count overflowing soon?
2613 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2616 if (preempt_count() == val) {
2617 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2618 #ifdef CONFIG_DEBUG_PREEMPT
2619 current->preempt_disable_ip = ip;
2621 trace_preempt_off(CALLER_ADDR0, ip);
2624 EXPORT_SYMBOL(preempt_count_add);
2625 NOKPROBE_SYMBOL(preempt_count_add);
2627 void preempt_count_sub(int val)
2629 #ifdef CONFIG_DEBUG_PREEMPT
2633 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2636 * Is the spinlock portion underflowing?
2638 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2639 !(preempt_count() & PREEMPT_MASK)))
2643 if (preempt_count() == val)
2644 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2645 __preempt_count_sub(val);
2647 EXPORT_SYMBOL(preempt_count_sub);
2648 NOKPROBE_SYMBOL(preempt_count_sub);
2653 * Print scheduling while atomic bug:
2655 static noinline void __schedule_bug(struct task_struct *prev)
2657 if (oops_in_progress)
2660 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2661 prev->comm, prev->pid, preempt_count());
2663 debug_show_held_locks(prev);
2665 if (irqs_disabled())
2666 print_irqtrace_events(prev);
2667 #ifdef CONFIG_DEBUG_PREEMPT
2668 if (in_atomic_preempt_off()) {
2669 pr_err("Preemption disabled at:");
2670 print_ip_sym(current->preempt_disable_ip);
2675 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2679 * Various schedule()-time debugging checks and statistics:
2681 static inline void schedule_debug(struct task_struct *prev)
2683 #ifdef CONFIG_SCHED_STACK_END_CHECK
2684 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2687 * Test if we are atomic. Since do_exit() needs to call into
2688 * schedule() atomically, we ignore that path. Otherwise whine
2689 * if we are scheduling when we should not.
2691 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2692 __schedule_bug(prev);
2695 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2697 schedstat_inc(this_rq(), sched_count);
2701 * Pick up the highest-prio task:
2703 static inline struct task_struct *
2704 pick_next_task(struct rq *rq, struct task_struct *prev)
2706 const struct sched_class *class = &fair_sched_class;
2707 struct task_struct *p;
2710 * Optimization: we know that if all tasks are in
2711 * the fair class we can call that function directly:
2713 if (likely(prev->sched_class == class &&
2714 rq->nr_running == rq->cfs.h_nr_running)) {
2715 p = fair_sched_class.pick_next_task(rq, prev);
2716 if (unlikely(p == RETRY_TASK))
2719 /* assumes fair_sched_class->next == idle_sched_class */
2721 p = idle_sched_class.pick_next_task(rq, prev);
2727 for_each_class(class) {
2728 p = class->pick_next_task(rq, prev);
2730 if (unlikely(p == RETRY_TASK))
2736 BUG(); /* the idle class will always have a runnable task */
2740 * __schedule() is the main scheduler function.
2742 * The main means of driving the scheduler and thus entering this function are:
2744 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2746 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2747 * paths. For example, see arch/x86/entry_64.S.
2749 * To drive preemption between tasks, the scheduler sets the flag in timer
2750 * interrupt handler scheduler_tick().
2752 * 3. Wakeups don't really cause entry into schedule(). They add a
2753 * task to the run-queue and that's it.
2755 * Now, if the new task added to the run-queue preempts the current
2756 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2757 * called on the nearest possible occasion:
2759 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2761 * - in syscall or exception context, at the next outmost
2762 * preempt_enable(). (this might be as soon as the wake_up()'s
2765 * - in IRQ context, return from interrupt-handler to
2766 * preemptible context
2768 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2771 * - cond_resched() call
2772 * - explicit schedule() call
2773 * - return from syscall or exception to user-space
2774 * - return from interrupt-handler to user-space
2776 * WARNING: must be called with preemption disabled!
2778 static void __sched __schedule(void)
2780 struct task_struct *prev, *next;
2781 unsigned long *switch_count;
2785 cpu = smp_processor_id();
2787 rcu_note_context_switch();
2790 schedule_debug(prev);
2792 if (sched_feat(HRTICK))
2796 * Make sure that signal_pending_state()->signal_pending() below
2797 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2798 * done by the caller to avoid the race with signal_wake_up().
2800 smp_mb__before_spinlock();
2801 raw_spin_lock_irq(&rq->lock);
2803 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2805 switch_count = &prev->nivcsw;
2806 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2807 if (unlikely(signal_pending_state(prev->state, prev))) {
2808 prev->state = TASK_RUNNING;
2810 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2814 * If a worker went to sleep, notify and ask workqueue
2815 * whether it wants to wake up a task to maintain
2818 if (prev->flags & PF_WQ_WORKER) {
2819 struct task_struct *to_wakeup;
2821 to_wakeup = wq_worker_sleeping(prev, cpu);
2823 try_to_wake_up_local(to_wakeup);
2826 switch_count = &prev->nvcsw;
2829 if (task_on_rq_queued(prev))
2830 update_rq_clock(rq);
2832 next = pick_next_task(rq, prev);
2833 clear_tsk_need_resched(prev);
2834 clear_preempt_need_resched();
2835 rq->clock_skip_update = 0;
2837 if (likely(prev != next)) {
2842 rq = context_switch(rq, prev, next); /* unlocks the rq */
2845 raw_spin_unlock_irq(&rq->lock);
2850 static inline void sched_submit_work(struct task_struct *tsk)
2852 if (!tsk->state || tsk_is_pi_blocked(tsk))
2855 * If we are going to sleep and we have plugged IO queued,
2856 * make sure to submit it to avoid deadlocks.
2858 if (blk_needs_flush_plug(tsk))
2859 blk_schedule_flush_plug(tsk);
2862 asmlinkage __visible void __sched schedule(void)
2864 struct task_struct *tsk = current;
2866 sched_submit_work(tsk);
2870 sched_preempt_enable_no_resched();
2871 } while (need_resched());
2873 EXPORT_SYMBOL(schedule);
2875 #ifdef CONFIG_CONTEXT_TRACKING
2876 asmlinkage __visible void __sched schedule_user(void)
2879 * If we come here after a random call to set_need_resched(),
2880 * or we have been woken up remotely but the IPI has not yet arrived,
2881 * we haven't yet exited the RCU idle mode. Do it here manually until
2882 * we find a better solution.
2884 * NB: There are buggy callers of this function. Ideally we
2885 * should warn if prev_state != CONTEXT_USER, but that will trigger
2886 * too frequently to make sense yet.
2888 enum ctx_state prev_state = exception_enter();
2890 exception_exit(prev_state);
2895 * schedule_preempt_disabled - called with preemption disabled
2897 * Returns with preemption disabled. Note: preempt_count must be 1
2899 void __sched schedule_preempt_disabled(void)
2901 sched_preempt_enable_no_resched();
2906 static void __sched notrace preempt_schedule_common(void)
2909 preempt_active_enter();
2911 preempt_active_exit();
2914 * Check again in case we missed a preemption opportunity
2915 * between schedule and now.
2917 } while (need_resched());
2920 #ifdef CONFIG_PREEMPT
2922 * this is the entry point to schedule() from in-kernel preemption
2923 * off of preempt_enable. Kernel preemptions off return from interrupt
2924 * occur there and call schedule directly.
2926 asmlinkage __visible void __sched notrace preempt_schedule(void)
2929 * If there is a non-zero preempt_count or interrupts are disabled,
2930 * we do not want to preempt the current task. Just return..
2932 if (likely(!preemptible()))
2935 preempt_schedule_common();
2937 NOKPROBE_SYMBOL(preempt_schedule);
2938 EXPORT_SYMBOL(preempt_schedule);
2940 #ifdef CONFIG_CONTEXT_TRACKING
2942 * preempt_schedule_context - preempt_schedule called by tracing
2944 * The tracing infrastructure uses preempt_enable_notrace to prevent
2945 * recursion and tracing preempt enabling caused by the tracing
2946 * infrastructure itself. But as tracing can happen in areas coming
2947 * from userspace or just about to enter userspace, a preempt enable
2948 * can occur before user_exit() is called. This will cause the scheduler
2949 * to be called when the system is still in usermode.
2951 * To prevent this, the preempt_enable_notrace will use this function
2952 * instead of preempt_schedule() to exit user context if needed before
2953 * calling the scheduler.
2955 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2957 enum ctx_state prev_ctx;
2959 if (likely(!preemptible()))
2963 preempt_active_enter();
2965 * Needs preempt disabled in case user_exit() is traced
2966 * and the tracer calls preempt_enable_notrace() causing
2967 * an infinite recursion.
2969 prev_ctx = exception_enter();
2971 exception_exit(prev_ctx);
2973 preempt_active_exit();
2974 } while (need_resched());
2976 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2977 #endif /* CONFIG_CONTEXT_TRACKING */
2979 #endif /* CONFIG_PREEMPT */
2982 * this is the entry point to schedule() from kernel preemption
2983 * off of irq context.
2984 * Note, that this is called and return with irqs disabled. This will
2985 * protect us against recursive calling from irq.
2987 asmlinkage __visible void __sched preempt_schedule_irq(void)
2989 enum ctx_state prev_state;
2991 /* Catch callers which need to be fixed */
2992 BUG_ON(preempt_count() || !irqs_disabled());
2994 prev_state = exception_enter();
2997 preempt_active_enter();
3000 local_irq_disable();
3001 preempt_active_exit();
3002 } while (need_resched());
3004 exception_exit(prev_state);
3007 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3010 return try_to_wake_up(curr->private, mode, wake_flags);
3012 EXPORT_SYMBOL(default_wake_function);
3014 #ifdef CONFIG_RT_MUTEXES
3017 * rt_mutex_setprio - set the current priority of a task
3019 * @prio: prio value (kernel-internal form)
3021 * This function changes the 'effective' priority of a task. It does
3022 * not touch ->normal_prio like __setscheduler().
3024 * Used by the rt_mutex code to implement priority inheritance
3025 * logic. Call site only calls if the priority of the task changed.
3027 void rt_mutex_setprio(struct task_struct *p, int prio)
3029 int oldprio, queued, running, enqueue_flag = 0;
3031 const struct sched_class *prev_class;
3033 BUG_ON(prio > MAX_PRIO);
3035 rq = __task_rq_lock(p);
3038 * Idle task boosting is a nono in general. There is one
3039 * exception, when PREEMPT_RT and NOHZ is active:
3041 * The idle task calls get_next_timer_interrupt() and holds
3042 * the timer wheel base->lock on the CPU and another CPU wants
3043 * to access the timer (probably to cancel it). We can safely
3044 * ignore the boosting request, as the idle CPU runs this code
3045 * with interrupts disabled and will complete the lock
3046 * protected section without being interrupted. So there is no
3047 * real need to boost.
3049 if (unlikely(p == rq->idle)) {
3050 WARN_ON(p != rq->curr);
3051 WARN_ON(p->pi_blocked_on);
3055 trace_sched_pi_setprio(p, prio);
3057 prev_class = p->sched_class;
3058 queued = task_on_rq_queued(p);
3059 running = task_current(rq, p);
3061 dequeue_task(rq, p, 0);
3063 put_prev_task(rq, p);
3066 * Boosting condition are:
3067 * 1. -rt task is running and holds mutex A
3068 * --> -dl task blocks on mutex A
3070 * 2. -dl task is running and holds mutex A
3071 * --> -dl task blocks on mutex A and could preempt the
3074 if (dl_prio(prio)) {
3075 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3076 if (!dl_prio(p->normal_prio) ||
3077 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3078 p->dl.dl_boosted = 1;
3079 p->dl.dl_throttled = 0;
3080 enqueue_flag = ENQUEUE_REPLENISH;
3082 p->dl.dl_boosted = 0;
3083 p->sched_class = &dl_sched_class;
3084 } else if (rt_prio(prio)) {
3085 if (dl_prio(oldprio))
3086 p->dl.dl_boosted = 0;
3088 enqueue_flag = ENQUEUE_HEAD;
3089 p->sched_class = &rt_sched_class;
3091 if (dl_prio(oldprio))
3092 p->dl.dl_boosted = 0;
3093 if (rt_prio(oldprio))
3095 p->sched_class = &fair_sched_class;
3101 p->sched_class->set_curr_task(rq);
3103 enqueue_task(rq, p, enqueue_flag);
3105 check_class_changed(rq, p, prev_class, oldprio);
3107 __task_rq_unlock(rq);
3111 void set_user_nice(struct task_struct *p, long nice)
3113 int old_prio, delta, queued;
3114 unsigned long flags;
3117 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3120 * We have to be careful, if called from sys_setpriority(),
3121 * the task might be in the middle of scheduling on another CPU.
3123 rq = task_rq_lock(p, &flags);
3125 * The RT priorities are set via sched_setscheduler(), but we still
3126 * allow the 'normal' nice value to be set - but as expected
3127 * it wont have any effect on scheduling until the task is
3128 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3130 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3131 p->static_prio = NICE_TO_PRIO(nice);
3134 queued = task_on_rq_queued(p);
3136 dequeue_task(rq, p, 0);
3138 p->static_prio = NICE_TO_PRIO(nice);
3141 p->prio = effective_prio(p);
3142 delta = p->prio - old_prio;
3145 enqueue_task(rq, p, 0);
3147 * If the task increased its priority or is running and
3148 * lowered its priority, then reschedule its CPU:
3150 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3154 task_rq_unlock(rq, p, &flags);
3156 EXPORT_SYMBOL(set_user_nice);
3159 * can_nice - check if a task can reduce its nice value
3163 int can_nice(const struct task_struct *p, const int nice)
3165 /* convert nice value [19,-20] to rlimit style value [1,40] */
3166 int nice_rlim = nice_to_rlimit(nice);
3168 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3169 capable(CAP_SYS_NICE));
3172 #ifdef __ARCH_WANT_SYS_NICE
3175 * sys_nice - change the priority of the current process.
3176 * @increment: priority increment
3178 * sys_setpriority is a more generic, but much slower function that
3179 * does similar things.
3181 SYSCALL_DEFINE1(nice, int, increment)
3186 * Setpriority might change our priority at the same moment.
3187 * We don't have to worry. Conceptually one call occurs first
3188 * and we have a single winner.
3190 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3191 nice = task_nice(current) + increment;
3193 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3194 if (increment < 0 && !can_nice(current, nice))
3197 retval = security_task_setnice(current, nice);
3201 set_user_nice(current, nice);
3208 * task_prio - return the priority value of a given task.
3209 * @p: the task in question.
3211 * Return: The priority value as seen by users in /proc.
3212 * RT tasks are offset by -200. Normal tasks are centered
3213 * around 0, value goes from -16 to +15.
3215 int task_prio(const struct task_struct *p)
3217 return p->prio - MAX_RT_PRIO;
3221 * idle_cpu - is a given cpu idle currently?
3222 * @cpu: the processor in question.
3224 * Return: 1 if the CPU is currently idle. 0 otherwise.
3226 int idle_cpu(int cpu)
3228 struct rq *rq = cpu_rq(cpu);
3230 if (rq->curr != rq->idle)
3237 if (!llist_empty(&rq->wake_list))
3245 * idle_task - return the idle task for a given cpu.
3246 * @cpu: the processor in question.
3248 * Return: The idle task for the cpu @cpu.
3250 struct task_struct *idle_task(int cpu)
3252 return cpu_rq(cpu)->idle;
3256 * find_process_by_pid - find a process with a matching PID value.
3257 * @pid: the pid in question.
3259 * The task of @pid, if found. %NULL otherwise.
3261 static struct task_struct *find_process_by_pid(pid_t pid)
3263 return pid ? find_task_by_vpid(pid) : current;
3267 * This function initializes the sched_dl_entity of a newly becoming
3268 * SCHED_DEADLINE task.
3270 * Only the static values are considered here, the actual runtime and the
3271 * absolute deadline will be properly calculated when the task is enqueued
3272 * for the first time with its new policy.
3275 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3277 struct sched_dl_entity *dl_se = &p->dl;
3279 dl_se->dl_runtime = attr->sched_runtime;
3280 dl_se->dl_deadline = attr->sched_deadline;
3281 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3282 dl_se->flags = attr->sched_flags;
3283 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3286 * Changing the parameters of a task is 'tricky' and we're not doing
3287 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3289 * What we SHOULD do is delay the bandwidth release until the 0-lag
3290 * point. This would include retaining the task_struct until that time
3291 * and change dl_overflow() to not immediately decrement the current
3294 * Instead we retain the current runtime/deadline and let the new
3295 * parameters take effect after the current reservation period lapses.
3296 * This is safe (albeit pessimistic) because the 0-lag point is always
3297 * before the current scheduling deadline.
3299 * We can still have temporary overloads because we do not delay the
3300 * change in bandwidth until that time; so admission control is
3301 * not on the safe side. It does however guarantee tasks will never
3302 * consume more than promised.
3307 * sched_setparam() passes in -1 for its policy, to let the functions
3308 * it calls know not to change it.
3310 #define SETPARAM_POLICY -1
3312 static void __setscheduler_params(struct task_struct *p,
3313 const struct sched_attr *attr)
3315 int policy = attr->sched_policy;
3317 if (policy == SETPARAM_POLICY)
3322 if (dl_policy(policy))
3323 __setparam_dl(p, attr);
3324 else if (fair_policy(policy))
3325 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3328 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3329 * !rt_policy. Always setting this ensures that things like
3330 * getparam()/getattr() don't report silly values for !rt tasks.
3332 p->rt_priority = attr->sched_priority;
3333 p->normal_prio = normal_prio(p);
3337 /* Actually do priority change: must hold pi & rq lock. */
3338 static void __setscheduler(struct rq *rq, struct task_struct *p,
3339 const struct sched_attr *attr, bool keep_boost)
3341 __setscheduler_params(p, attr);
3344 * Keep a potential priority boosting if called from
3345 * sched_setscheduler().
3348 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3350 p->prio = normal_prio(p);
3352 if (dl_prio(p->prio))
3353 p->sched_class = &dl_sched_class;
3354 else if (rt_prio(p->prio))
3355 p->sched_class = &rt_sched_class;
3357 p->sched_class = &fair_sched_class;
3361 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3363 struct sched_dl_entity *dl_se = &p->dl;
3365 attr->sched_priority = p->rt_priority;
3366 attr->sched_runtime = dl_se->dl_runtime;
3367 attr->sched_deadline = dl_se->dl_deadline;
3368 attr->sched_period = dl_se->dl_period;
3369 attr->sched_flags = dl_se->flags;
3373 * This function validates the new parameters of a -deadline task.
3374 * We ask for the deadline not being zero, and greater or equal
3375 * than the runtime, as well as the period of being zero or
3376 * greater than deadline. Furthermore, we have to be sure that
3377 * user parameters are above the internal resolution of 1us (we
3378 * check sched_runtime only since it is always the smaller one) and
3379 * below 2^63 ns (we have to check both sched_deadline and
3380 * sched_period, as the latter can be zero).
3383 __checkparam_dl(const struct sched_attr *attr)
3386 if (attr->sched_deadline == 0)
3390 * Since we truncate DL_SCALE bits, make sure we're at least
3393 if (attr->sched_runtime < (1ULL << DL_SCALE))
3397 * Since we use the MSB for wrap-around and sign issues, make
3398 * sure it's not set (mind that period can be equal to zero).
3400 if (attr->sched_deadline & (1ULL << 63) ||
3401 attr->sched_period & (1ULL << 63))
3404 /* runtime <= deadline <= period (if period != 0) */
3405 if ((attr->sched_period != 0 &&
3406 attr->sched_period < attr->sched_deadline) ||
3407 attr->sched_deadline < attr->sched_runtime)
3414 * check the target process has a UID that matches the current process's
3416 static bool check_same_owner(struct task_struct *p)
3418 const struct cred *cred = current_cred(), *pcred;
3422 pcred = __task_cred(p);
3423 match = (uid_eq(cred->euid, pcred->euid) ||
3424 uid_eq(cred->euid, pcred->uid));
3429 static bool dl_param_changed(struct task_struct *p,
3430 const struct sched_attr *attr)
3432 struct sched_dl_entity *dl_se = &p->dl;
3434 if (dl_se->dl_runtime != attr->sched_runtime ||
3435 dl_se->dl_deadline != attr->sched_deadline ||
3436 dl_se->dl_period != attr->sched_period ||
3437 dl_se->flags != attr->sched_flags)
3443 static int __sched_setscheduler(struct task_struct *p,
3444 const struct sched_attr *attr,
3447 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3448 MAX_RT_PRIO - 1 - attr->sched_priority;
3449 int retval, oldprio, oldpolicy = -1, queued, running;
3450 int new_effective_prio, policy = attr->sched_policy;
3451 unsigned long flags;
3452 const struct sched_class *prev_class;
3456 /* may grab non-irq protected spin_locks */
3457 BUG_ON(in_interrupt());
3459 /* double check policy once rq lock held */
3461 reset_on_fork = p->sched_reset_on_fork;
3462 policy = oldpolicy = p->policy;
3464 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3466 if (policy != SCHED_DEADLINE &&
3467 policy != SCHED_FIFO && policy != SCHED_RR &&
3468 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3469 policy != SCHED_IDLE)
3473 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3477 * Valid priorities for SCHED_FIFO and SCHED_RR are
3478 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3479 * SCHED_BATCH and SCHED_IDLE is 0.
3481 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3482 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3484 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3485 (rt_policy(policy) != (attr->sched_priority != 0)))
3489 * Allow unprivileged RT tasks to decrease priority:
3491 if (user && !capable(CAP_SYS_NICE)) {
3492 if (fair_policy(policy)) {
3493 if (attr->sched_nice < task_nice(p) &&
3494 !can_nice(p, attr->sched_nice))
3498 if (rt_policy(policy)) {
3499 unsigned long rlim_rtprio =
3500 task_rlimit(p, RLIMIT_RTPRIO);
3502 /* can't set/change the rt policy */
3503 if (policy != p->policy && !rlim_rtprio)
3506 /* can't increase priority */
3507 if (attr->sched_priority > p->rt_priority &&
3508 attr->sched_priority > rlim_rtprio)
3513 * Can't set/change SCHED_DEADLINE policy at all for now
3514 * (safest behavior); in the future we would like to allow
3515 * unprivileged DL tasks to increase their relative deadline
3516 * or reduce their runtime (both ways reducing utilization)
3518 if (dl_policy(policy))
3522 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3523 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3525 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3526 if (!can_nice(p, task_nice(p)))
3530 /* can't change other user's priorities */
3531 if (!check_same_owner(p))
3534 /* Normal users shall not reset the sched_reset_on_fork flag */
3535 if (p->sched_reset_on_fork && !reset_on_fork)
3540 retval = security_task_setscheduler(p);
3546 * make sure no PI-waiters arrive (or leave) while we are
3547 * changing the priority of the task:
3549 * To be able to change p->policy safely, the appropriate
3550 * runqueue lock must be held.
3552 rq = task_rq_lock(p, &flags);
3555 * Changing the policy of the stop threads its a very bad idea
3557 if (p == rq->stop) {
3558 task_rq_unlock(rq, p, &flags);
3563 * If not changing anything there's no need to proceed further,
3564 * but store a possible modification of reset_on_fork.
3566 if (unlikely(policy == p->policy)) {
3567 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3569 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3571 if (dl_policy(policy) && dl_param_changed(p, attr))
3574 p->sched_reset_on_fork = reset_on_fork;
3575 task_rq_unlock(rq, p, &flags);
3581 #ifdef CONFIG_RT_GROUP_SCHED
3583 * Do not allow realtime tasks into groups that have no runtime
3586 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3587 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3588 !task_group_is_autogroup(task_group(p))) {
3589 task_rq_unlock(rq, p, &flags);
3594 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3595 cpumask_t *span = rq->rd->span;
3598 * Don't allow tasks with an affinity mask smaller than
3599 * the entire root_domain to become SCHED_DEADLINE. We
3600 * will also fail if there's no bandwidth available.
3602 if (!cpumask_subset(span, &p->cpus_allowed) ||
3603 rq->rd->dl_bw.bw == 0) {
3604 task_rq_unlock(rq, p, &flags);
3611 /* recheck policy now with rq lock held */
3612 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3613 policy = oldpolicy = -1;
3614 task_rq_unlock(rq, p, &flags);
3619 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3620 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3623 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3624 task_rq_unlock(rq, p, &flags);
3628 p->sched_reset_on_fork = reset_on_fork;
3632 * Take priority boosted tasks into account. If the new
3633 * effective priority is unchanged, we just store the new
3634 * normal parameters and do not touch the scheduler class and
3635 * the runqueue. This will be done when the task deboost
3638 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3639 if (new_effective_prio == oldprio) {
3640 __setscheduler_params(p, attr);
3641 task_rq_unlock(rq, p, &flags);
3645 queued = task_on_rq_queued(p);
3646 running = task_current(rq, p);
3648 dequeue_task(rq, p, 0);
3650 put_prev_task(rq, p);
3652 prev_class = p->sched_class;
3653 __setscheduler(rq, p, attr, true);
3656 p->sched_class->set_curr_task(rq);
3659 * We enqueue to tail when the priority of a task is
3660 * increased (user space view).
3662 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3665 check_class_changed(rq, p, prev_class, oldprio);
3666 task_rq_unlock(rq, p, &flags);
3668 rt_mutex_adjust_pi(p);
3673 static int _sched_setscheduler(struct task_struct *p, int policy,
3674 const struct sched_param *param, bool check)
3676 struct sched_attr attr = {
3677 .sched_policy = policy,
3678 .sched_priority = param->sched_priority,
3679 .sched_nice = PRIO_TO_NICE(p->static_prio),
3682 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3683 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3684 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3685 policy &= ~SCHED_RESET_ON_FORK;
3686 attr.sched_policy = policy;
3689 return __sched_setscheduler(p, &attr, check);
3692 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3693 * @p: the task in question.
3694 * @policy: new policy.
3695 * @param: structure containing the new RT priority.
3697 * Return: 0 on success. An error code otherwise.
3699 * NOTE that the task may be already dead.
3701 int sched_setscheduler(struct task_struct *p, int policy,
3702 const struct sched_param *param)
3704 return _sched_setscheduler(p, policy, param, true);
3706 EXPORT_SYMBOL_GPL(sched_setscheduler);
3708 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3710 return __sched_setscheduler(p, attr, true);
3712 EXPORT_SYMBOL_GPL(sched_setattr);
3715 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3716 * @p: the task in question.
3717 * @policy: new policy.
3718 * @param: structure containing the new RT priority.
3720 * Just like sched_setscheduler, only don't bother checking if the
3721 * current context has permission. For example, this is needed in
3722 * stop_machine(): we create temporary high priority worker threads,
3723 * but our caller might not have that capability.
3725 * Return: 0 on success. An error code otherwise.
3727 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3728 const struct sched_param *param)
3730 return _sched_setscheduler(p, policy, param, false);
3734 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3736 struct sched_param lparam;
3737 struct task_struct *p;
3740 if (!param || pid < 0)
3742 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3747 p = find_process_by_pid(pid);
3749 retval = sched_setscheduler(p, policy, &lparam);
3756 * Mimics kernel/events/core.c perf_copy_attr().
3758 static int sched_copy_attr(struct sched_attr __user *uattr,
3759 struct sched_attr *attr)
3764 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3768 * zero the full structure, so that a short copy will be nice.
3770 memset(attr, 0, sizeof(*attr));
3772 ret = get_user(size, &uattr->size);
3776 if (size > PAGE_SIZE) /* silly large */
3779 if (!size) /* abi compat */
3780 size = SCHED_ATTR_SIZE_VER0;
3782 if (size < SCHED_ATTR_SIZE_VER0)
3786 * If we're handed a bigger struct than we know of,
3787 * ensure all the unknown bits are 0 - i.e. new
3788 * user-space does not rely on any kernel feature
3789 * extensions we dont know about yet.
3791 if (size > sizeof(*attr)) {
3792 unsigned char __user *addr;
3793 unsigned char __user *end;
3796 addr = (void __user *)uattr + sizeof(*attr);
3797 end = (void __user *)uattr + size;
3799 for (; addr < end; addr++) {
3800 ret = get_user(val, addr);
3806 size = sizeof(*attr);
3809 ret = copy_from_user(attr, uattr, size);
3814 * XXX: do we want to be lenient like existing syscalls; or do we want
3815 * to be strict and return an error on out-of-bounds values?
3817 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3822 put_user(sizeof(*attr), &uattr->size);
3827 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3828 * @pid: the pid in question.
3829 * @policy: new policy.
3830 * @param: structure containing the new RT priority.
3832 * Return: 0 on success. An error code otherwise.
3834 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3835 struct sched_param __user *, param)
3837 /* negative values for policy are not valid */
3841 return do_sched_setscheduler(pid, policy, param);
3845 * sys_sched_setparam - set/change the RT priority of a thread
3846 * @pid: the pid in question.
3847 * @param: structure containing the new RT priority.
3849 * Return: 0 on success. An error code otherwise.
3851 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3853 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3857 * sys_sched_setattr - same as above, but with extended sched_attr
3858 * @pid: the pid in question.
3859 * @uattr: structure containing the extended parameters.
3860 * @flags: for future extension.
3862 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3863 unsigned int, flags)
3865 struct sched_attr attr;
3866 struct task_struct *p;
3869 if (!uattr || pid < 0 || flags)
3872 retval = sched_copy_attr(uattr, &attr);
3876 if ((int)attr.sched_policy < 0)
3881 p = find_process_by_pid(pid);
3883 retval = sched_setattr(p, &attr);
3890 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3891 * @pid: the pid in question.
3893 * Return: On success, the policy of the thread. Otherwise, a negative error
3896 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3898 struct task_struct *p;
3906 p = find_process_by_pid(pid);
3908 retval = security_task_getscheduler(p);
3911 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3918 * sys_sched_getparam - get the RT priority of a thread
3919 * @pid: the pid in question.
3920 * @param: structure containing the RT priority.
3922 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3925 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3927 struct sched_param lp = { .sched_priority = 0 };
3928 struct task_struct *p;
3931 if (!param || pid < 0)
3935 p = find_process_by_pid(pid);
3940 retval = security_task_getscheduler(p);
3944 if (task_has_rt_policy(p))
3945 lp.sched_priority = p->rt_priority;
3949 * This one might sleep, we cannot do it with a spinlock held ...
3951 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3960 static int sched_read_attr(struct sched_attr __user *uattr,
3961 struct sched_attr *attr,
3966 if (!access_ok(VERIFY_WRITE, uattr, usize))
3970 * If we're handed a smaller struct than we know of,
3971 * ensure all the unknown bits are 0 - i.e. old
3972 * user-space does not get uncomplete information.
3974 if (usize < sizeof(*attr)) {
3975 unsigned char *addr;
3978 addr = (void *)attr + usize;
3979 end = (void *)attr + sizeof(*attr);
3981 for (; addr < end; addr++) {
3989 ret = copy_to_user(uattr, attr, attr->size);
3997 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3998 * @pid: the pid in question.
3999 * @uattr: structure containing the extended parameters.
4000 * @size: sizeof(attr) for fwd/bwd comp.
4001 * @flags: for future extension.
4003 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4004 unsigned int, size, unsigned int, flags)
4006 struct sched_attr attr = {
4007 .size = sizeof(struct sched_attr),
4009 struct task_struct *p;
4012 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4013 size < SCHED_ATTR_SIZE_VER0 || flags)
4017 p = find_process_by_pid(pid);
4022 retval = security_task_getscheduler(p);
4026 attr.sched_policy = p->policy;
4027 if (p->sched_reset_on_fork)
4028 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4029 if (task_has_dl_policy(p))
4030 __getparam_dl(p, &attr);
4031 else if (task_has_rt_policy(p))
4032 attr.sched_priority = p->rt_priority;
4034 attr.sched_nice = task_nice(p);
4038 retval = sched_read_attr(uattr, &attr, size);
4046 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4048 cpumask_var_t cpus_allowed, new_mask;
4049 struct task_struct *p;
4054 p = find_process_by_pid(pid);
4060 /* Prevent p going away */
4064 if (p->flags & PF_NO_SETAFFINITY) {
4068 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4072 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4074 goto out_free_cpus_allowed;
4077 if (!check_same_owner(p)) {
4079 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4081 goto out_free_new_mask;
4086 retval = security_task_setscheduler(p);
4088 goto out_free_new_mask;
4091 cpuset_cpus_allowed(p, cpus_allowed);
4092 cpumask_and(new_mask, in_mask, cpus_allowed);
4095 * Since bandwidth control happens on root_domain basis,
4096 * if admission test is enabled, we only admit -deadline
4097 * tasks allowed to run on all the CPUs in the task's
4101 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4103 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4106 goto out_free_new_mask;
4112 retval = set_cpus_allowed_ptr(p, new_mask);
4115 cpuset_cpus_allowed(p, cpus_allowed);
4116 if (!cpumask_subset(new_mask, cpus_allowed)) {
4118 * We must have raced with a concurrent cpuset
4119 * update. Just reset the cpus_allowed to the
4120 * cpuset's cpus_allowed
4122 cpumask_copy(new_mask, cpus_allowed);
4127 free_cpumask_var(new_mask);
4128 out_free_cpus_allowed:
4129 free_cpumask_var(cpus_allowed);
4135 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4136 struct cpumask *new_mask)
4138 if (len < cpumask_size())
4139 cpumask_clear(new_mask);
4140 else if (len > cpumask_size())
4141 len = cpumask_size();
4143 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4147 * sys_sched_setaffinity - set the cpu affinity of a process
4148 * @pid: pid of the process
4149 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4150 * @user_mask_ptr: user-space pointer to the new cpu mask
4152 * Return: 0 on success. An error code otherwise.
4154 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4155 unsigned long __user *, user_mask_ptr)
4157 cpumask_var_t new_mask;
4160 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4163 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4165 retval = sched_setaffinity(pid, new_mask);
4166 free_cpumask_var(new_mask);
4170 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4172 struct task_struct *p;
4173 unsigned long flags;
4179 p = find_process_by_pid(pid);
4183 retval = security_task_getscheduler(p);
4187 raw_spin_lock_irqsave(&p->pi_lock, flags);
4188 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4189 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4198 * sys_sched_getaffinity - get the cpu affinity of a process
4199 * @pid: pid of the process
4200 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4201 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4203 * Return: 0 on success. An error code otherwise.
4205 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4206 unsigned long __user *, user_mask_ptr)
4211 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4213 if (len & (sizeof(unsigned long)-1))
4216 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4219 ret = sched_getaffinity(pid, mask);
4221 size_t retlen = min_t(size_t, len, cpumask_size());
4223 if (copy_to_user(user_mask_ptr, mask, retlen))
4228 free_cpumask_var(mask);
4234 * sys_sched_yield - yield the current processor to other threads.
4236 * This function yields the current CPU to other tasks. If there are no
4237 * other threads running on this CPU then this function will return.
4241 SYSCALL_DEFINE0(sched_yield)
4243 struct rq *rq = this_rq_lock();
4245 schedstat_inc(rq, yld_count);
4246 current->sched_class->yield_task(rq);
4249 * Since we are going to call schedule() anyway, there's
4250 * no need to preempt or enable interrupts:
4252 __release(rq->lock);
4253 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4254 do_raw_spin_unlock(&rq->lock);
4255 sched_preempt_enable_no_resched();
4262 int __sched _cond_resched(void)
4264 if (should_resched()) {
4265 preempt_schedule_common();
4270 EXPORT_SYMBOL(_cond_resched);
4273 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4274 * call schedule, and on return reacquire the lock.
4276 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4277 * operations here to prevent schedule() from being called twice (once via
4278 * spin_unlock(), once by hand).
4280 int __cond_resched_lock(spinlock_t *lock)
4282 int resched = should_resched();
4285 lockdep_assert_held(lock);
4287 if (spin_needbreak(lock) || resched) {
4290 preempt_schedule_common();
4298 EXPORT_SYMBOL(__cond_resched_lock);
4300 int __sched __cond_resched_softirq(void)
4302 BUG_ON(!in_softirq());
4304 if (should_resched()) {
4306 preempt_schedule_common();
4312 EXPORT_SYMBOL(__cond_resched_softirq);
4315 * yield - yield the current processor to other threads.
4317 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4319 * The scheduler is at all times free to pick the calling task as the most
4320 * eligible task to run, if removing the yield() call from your code breaks
4321 * it, its already broken.
4323 * Typical broken usage is:
4328 * where one assumes that yield() will let 'the other' process run that will
4329 * make event true. If the current task is a SCHED_FIFO task that will never
4330 * happen. Never use yield() as a progress guarantee!!
4332 * If you want to use yield() to wait for something, use wait_event().
4333 * If you want to use yield() to be 'nice' for others, use cond_resched().
4334 * If you still want to use yield(), do not!
4336 void __sched yield(void)
4338 set_current_state(TASK_RUNNING);
4341 EXPORT_SYMBOL(yield);
4344 * yield_to - yield the current processor to another thread in
4345 * your thread group, or accelerate that thread toward the
4346 * processor it's on.
4348 * @preempt: whether task preemption is allowed or not
4350 * It's the caller's job to ensure that the target task struct
4351 * can't go away on us before we can do any checks.
4354 * true (>0) if we indeed boosted the target task.
4355 * false (0) if we failed to boost the target.
4356 * -ESRCH if there's no task to yield to.
4358 int __sched yield_to(struct task_struct *p, bool preempt)
4360 struct task_struct *curr = current;
4361 struct rq *rq, *p_rq;
4362 unsigned long flags;
4365 local_irq_save(flags);
4371 * If we're the only runnable task on the rq and target rq also
4372 * has only one task, there's absolutely no point in yielding.
4374 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4379 double_rq_lock(rq, p_rq);
4380 if (task_rq(p) != p_rq) {
4381 double_rq_unlock(rq, p_rq);
4385 if (!curr->sched_class->yield_to_task)
4388 if (curr->sched_class != p->sched_class)
4391 if (task_running(p_rq, p) || p->state)
4394 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4396 schedstat_inc(rq, yld_count);
4398 * Make p's CPU reschedule; pick_next_entity takes care of
4401 if (preempt && rq != p_rq)
4406 double_rq_unlock(rq, p_rq);
4408 local_irq_restore(flags);
4415 EXPORT_SYMBOL_GPL(yield_to);
4418 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4419 * that process accounting knows that this is a task in IO wait state.
4421 long __sched io_schedule_timeout(long timeout)
4423 int old_iowait = current->in_iowait;
4427 current->in_iowait = 1;
4428 blk_schedule_flush_plug(current);
4430 delayacct_blkio_start();
4432 atomic_inc(&rq->nr_iowait);
4433 ret = schedule_timeout(timeout);
4434 current->in_iowait = old_iowait;
4435 atomic_dec(&rq->nr_iowait);
4436 delayacct_blkio_end();
4440 EXPORT_SYMBOL(io_schedule_timeout);
4443 * sys_sched_get_priority_max - return maximum RT priority.
4444 * @policy: scheduling class.
4446 * Return: On success, this syscall returns the maximum
4447 * rt_priority that can be used by a given scheduling class.
4448 * On failure, a negative error code is returned.
4450 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4457 ret = MAX_USER_RT_PRIO-1;
4459 case SCHED_DEADLINE:
4470 * sys_sched_get_priority_min - return minimum RT priority.
4471 * @policy: scheduling class.
4473 * Return: On success, this syscall returns the minimum
4474 * rt_priority that can be used by a given scheduling class.
4475 * On failure, a negative error code is returned.
4477 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4486 case SCHED_DEADLINE:
4496 * sys_sched_rr_get_interval - return the default timeslice of a process.
4497 * @pid: pid of the process.
4498 * @interval: userspace pointer to the timeslice value.
4500 * this syscall writes the default timeslice value of a given process
4501 * into the user-space timespec buffer. A value of '0' means infinity.
4503 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4506 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4507 struct timespec __user *, interval)
4509 struct task_struct *p;
4510 unsigned int time_slice;
4511 unsigned long flags;
4521 p = find_process_by_pid(pid);
4525 retval = security_task_getscheduler(p);
4529 rq = task_rq_lock(p, &flags);
4531 if (p->sched_class->get_rr_interval)
4532 time_slice = p->sched_class->get_rr_interval(rq, p);
4533 task_rq_unlock(rq, p, &flags);
4536 jiffies_to_timespec(time_slice, &t);
4537 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4545 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4547 void sched_show_task(struct task_struct *p)
4549 unsigned long free = 0;
4551 unsigned long state = p->state;
4554 state = __ffs(state) + 1;
4555 printk(KERN_INFO "%-15.15s %c", p->comm,
4556 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4557 #if BITS_PER_LONG == 32
4558 if (state == TASK_RUNNING)
4559 printk(KERN_CONT " running ");
4561 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4563 if (state == TASK_RUNNING)
4564 printk(KERN_CONT " running task ");
4566 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4568 #ifdef CONFIG_DEBUG_STACK_USAGE
4569 free = stack_not_used(p);
4574 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4576 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4577 task_pid_nr(p), ppid,
4578 (unsigned long)task_thread_info(p)->flags);
4580 print_worker_info(KERN_INFO, p);
4581 show_stack(p, NULL);
4584 void show_state_filter(unsigned long state_filter)
4586 struct task_struct *g, *p;
4588 #if BITS_PER_LONG == 32
4590 " task PC stack pid father\n");
4593 " task PC stack pid father\n");
4596 for_each_process_thread(g, p) {
4598 * reset the NMI-timeout, listing all files on a slow
4599 * console might take a lot of time:
4601 touch_nmi_watchdog();
4602 if (!state_filter || (p->state & state_filter))
4606 touch_all_softlockup_watchdogs();
4608 #ifdef CONFIG_SCHED_DEBUG
4609 sysrq_sched_debug_show();
4613 * Only show locks if all tasks are dumped:
4616 debug_show_all_locks();
4619 void init_idle_bootup_task(struct task_struct *idle)
4621 idle->sched_class = &idle_sched_class;
4625 * init_idle - set up an idle thread for a given CPU
4626 * @idle: task in question
4627 * @cpu: cpu the idle task belongs to
4629 * NOTE: this function does not set the idle thread's NEED_RESCHED
4630 * flag, to make booting more robust.
4632 void init_idle(struct task_struct *idle, int cpu)
4634 struct rq *rq = cpu_rq(cpu);
4635 unsigned long flags;
4637 raw_spin_lock_irqsave(&rq->lock, flags);
4639 __sched_fork(0, idle);
4640 idle->state = TASK_RUNNING;
4641 idle->se.exec_start = sched_clock();
4643 do_set_cpus_allowed(idle, cpumask_of(cpu));
4645 * We're having a chicken and egg problem, even though we are
4646 * holding rq->lock, the cpu isn't yet set to this cpu so the
4647 * lockdep check in task_group() will fail.
4649 * Similar case to sched_fork(). / Alternatively we could
4650 * use task_rq_lock() here and obtain the other rq->lock.
4655 __set_task_cpu(idle, cpu);
4658 rq->curr = rq->idle = idle;
4659 idle->on_rq = TASK_ON_RQ_QUEUED;
4660 #if defined(CONFIG_SMP)
4663 raw_spin_unlock_irqrestore(&rq->lock, flags);
4665 /* Set the preempt count _outside_ the spinlocks! */
4666 init_idle_preempt_count(idle, cpu);
4669 * The idle tasks have their own, simple scheduling class:
4671 idle->sched_class = &idle_sched_class;
4672 ftrace_graph_init_idle_task(idle, cpu);
4673 vtime_init_idle(idle, cpu);
4674 #if defined(CONFIG_SMP)
4675 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4679 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4680 const struct cpumask *trial)
4682 int ret = 1, trial_cpus;
4683 struct dl_bw *cur_dl_b;
4684 unsigned long flags;
4686 if (!cpumask_weight(cur))
4689 rcu_read_lock_sched();
4690 cur_dl_b = dl_bw_of(cpumask_any(cur));
4691 trial_cpus = cpumask_weight(trial);
4693 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4694 if (cur_dl_b->bw != -1 &&
4695 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4697 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4698 rcu_read_unlock_sched();
4703 int task_can_attach(struct task_struct *p,
4704 const struct cpumask *cs_cpus_allowed)
4709 * Kthreads which disallow setaffinity shouldn't be moved
4710 * to a new cpuset; we don't want to change their cpu
4711 * affinity and isolating such threads by their set of
4712 * allowed nodes is unnecessary. Thus, cpusets are not
4713 * applicable for such threads. This prevents checking for
4714 * success of set_cpus_allowed_ptr() on all attached tasks
4715 * before cpus_allowed may be changed.
4717 if (p->flags & PF_NO_SETAFFINITY) {
4723 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4725 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4730 unsigned long flags;
4732 rcu_read_lock_sched();
4733 dl_b = dl_bw_of(dest_cpu);
4734 raw_spin_lock_irqsave(&dl_b->lock, flags);
4735 cpus = dl_bw_cpus(dest_cpu);
4736 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4741 * We reserve space for this task in the destination
4742 * root_domain, as we can't fail after this point.
4743 * We will free resources in the source root_domain
4744 * later on (see set_cpus_allowed_dl()).
4746 __dl_add(dl_b, p->dl.dl_bw);
4748 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4749 rcu_read_unlock_sched();
4759 * move_queued_task - move a queued task to new rq.
4761 * Returns (locked) new rq. Old rq's lock is released.
4763 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4765 struct rq *rq = task_rq(p);
4767 lockdep_assert_held(&rq->lock);
4769 dequeue_task(rq, p, 0);
4770 p->on_rq = TASK_ON_RQ_MIGRATING;
4771 set_task_cpu(p, new_cpu);
4772 raw_spin_unlock(&rq->lock);
4774 rq = cpu_rq(new_cpu);
4776 raw_spin_lock(&rq->lock);
4777 BUG_ON(task_cpu(p) != new_cpu);
4778 p->on_rq = TASK_ON_RQ_QUEUED;
4779 enqueue_task(rq, p, 0);
4780 check_preempt_curr(rq, p, 0);
4785 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4787 if (p->sched_class->set_cpus_allowed)
4788 p->sched_class->set_cpus_allowed(p, new_mask);
4790 cpumask_copy(&p->cpus_allowed, new_mask);
4791 p->nr_cpus_allowed = cpumask_weight(new_mask);
4795 * This is how migration works:
4797 * 1) we invoke migration_cpu_stop() on the target CPU using
4799 * 2) stopper starts to run (implicitly forcing the migrated thread
4801 * 3) it checks whether the migrated task is still in the wrong runqueue.
4802 * 4) if it's in the wrong runqueue then the migration thread removes
4803 * it and puts it into the right queue.
4804 * 5) stopper completes and stop_one_cpu() returns and the migration
4809 * Change a given task's CPU affinity. Migrate the thread to a
4810 * proper CPU and schedule it away if the CPU it's executing on
4811 * is removed from the allowed bitmask.
4813 * NOTE: the caller must have a valid reference to the task, the
4814 * task must not exit() & deallocate itself prematurely. The
4815 * call is not atomic; no spinlocks may be held.
4817 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4819 unsigned long flags;
4821 unsigned int dest_cpu;
4824 rq = task_rq_lock(p, &flags);
4826 if (cpumask_equal(&p->cpus_allowed, new_mask))
4829 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4834 do_set_cpus_allowed(p, new_mask);
4836 /* Can the task run on the task's current CPU? If so, we're done */
4837 if (cpumask_test_cpu(task_cpu(p), new_mask))
4840 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4841 if (task_running(rq, p) || p->state == TASK_WAKING) {
4842 struct migration_arg arg = { p, dest_cpu };
4843 /* Need help from migration thread: drop lock and wait. */
4844 task_rq_unlock(rq, p, &flags);
4845 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4846 tlb_migrate_finish(p->mm);
4848 } else if (task_on_rq_queued(p))
4849 rq = move_queued_task(p, dest_cpu);
4851 task_rq_unlock(rq, p, &flags);
4855 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4858 * Move (not current) task off this cpu, onto dest cpu. We're doing
4859 * this because either it can't run here any more (set_cpus_allowed()
4860 * away from this CPU, or CPU going down), or because we're
4861 * attempting to rebalance this task on exec (sched_exec).
4863 * So we race with normal scheduler movements, but that's OK, as long
4864 * as the task is no longer on this CPU.
4866 * Returns non-zero if task was successfully migrated.
4868 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4873 if (unlikely(!cpu_active(dest_cpu)))
4876 rq = cpu_rq(src_cpu);
4878 raw_spin_lock(&p->pi_lock);
4879 raw_spin_lock(&rq->lock);
4880 /* Already moved. */
4881 if (task_cpu(p) != src_cpu)
4884 /* Affinity changed (again). */
4885 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4889 * If we're not on a rq, the next wake-up will ensure we're
4892 if (task_on_rq_queued(p))
4893 rq = move_queued_task(p, dest_cpu);
4897 raw_spin_unlock(&rq->lock);
4898 raw_spin_unlock(&p->pi_lock);
4902 #ifdef CONFIG_NUMA_BALANCING
4903 /* Migrate current task p to target_cpu */
4904 int migrate_task_to(struct task_struct *p, int target_cpu)
4906 struct migration_arg arg = { p, target_cpu };
4907 int curr_cpu = task_cpu(p);
4909 if (curr_cpu == target_cpu)
4912 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4915 /* TODO: This is not properly updating schedstats */
4917 trace_sched_move_numa(p, curr_cpu, target_cpu);
4918 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4922 * Requeue a task on a given node and accurately track the number of NUMA
4923 * tasks on the runqueues
4925 void sched_setnuma(struct task_struct *p, int nid)
4928 unsigned long flags;
4929 bool queued, running;
4931 rq = task_rq_lock(p, &flags);
4932 queued = task_on_rq_queued(p);
4933 running = task_current(rq, p);
4936 dequeue_task(rq, p, 0);
4938 put_prev_task(rq, p);
4940 p->numa_preferred_nid = nid;
4943 p->sched_class->set_curr_task(rq);
4945 enqueue_task(rq, p, 0);
4946 task_rq_unlock(rq, p, &flags);
4951 * migration_cpu_stop - this will be executed by a highprio stopper thread
4952 * and performs thread migration by bumping thread off CPU then
4953 * 'pushing' onto another runqueue.
4955 static int migration_cpu_stop(void *data)
4957 struct migration_arg *arg = data;
4960 * The original target cpu might have gone down and we might
4961 * be on another cpu but it doesn't matter.
4963 local_irq_disable();
4965 * We need to explicitly wake pending tasks before running
4966 * __migrate_task() such that we will not miss enforcing cpus_allowed
4967 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4969 sched_ttwu_pending();
4970 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4975 #ifdef CONFIG_HOTPLUG_CPU
4978 * Ensures that the idle task is using init_mm right before its cpu goes
4981 void idle_task_exit(void)
4983 struct mm_struct *mm = current->active_mm;
4985 BUG_ON(cpu_online(smp_processor_id()));
4987 if (mm != &init_mm) {
4988 switch_mm(mm, &init_mm, current);
4989 finish_arch_post_lock_switch();
4995 * Since this CPU is going 'away' for a while, fold any nr_active delta
4996 * we might have. Assumes we're called after migrate_tasks() so that the
4997 * nr_active count is stable.
4999 * Also see the comment "Global load-average calculations".
5001 static void calc_load_migrate(struct rq *rq)
5003 long delta = calc_load_fold_active(rq);
5005 atomic_long_add(delta, &calc_load_tasks);
5008 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5012 static const struct sched_class fake_sched_class = {
5013 .put_prev_task = put_prev_task_fake,
5016 static struct task_struct fake_task = {
5018 * Avoid pull_{rt,dl}_task()
5020 .prio = MAX_PRIO + 1,
5021 .sched_class = &fake_sched_class,
5025 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5026 * try_to_wake_up()->select_task_rq().
5028 * Called with rq->lock held even though we'er in stop_machine() and
5029 * there's no concurrency possible, we hold the required locks anyway
5030 * because of lock validation efforts.
5032 static void migrate_tasks(unsigned int dead_cpu)
5034 struct rq *rq = cpu_rq(dead_cpu);
5035 struct task_struct *next, *stop = rq->stop;
5039 * Fudge the rq selection such that the below task selection loop
5040 * doesn't get stuck on the currently eligible stop task.
5042 * We're currently inside stop_machine() and the rq is either stuck
5043 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5044 * either way we should never end up calling schedule() until we're
5050 * put_prev_task() and pick_next_task() sched
5051 * class method both need to have an up-to-date
5052 * value of rq->clock[_task]
5054 update_rq_clock(rq);
5058 * There's this thread running, bail when that's the only
5061 if (rq->nr_running == 1)
5064 next = pick_next_task(rq, &fake_task);
5066 next->sched_class->put_prev_task(rq, next);
5068 /* Find suitable destination for @next, with force if needed. */
5069 dest_cpu = select_fallback_rq(dead_cpu, next);
5070 raw_spin_unlock(&rq->lock);
5072 __migrate_task(next, dead_cpu, dest_cpu);
5074 raw_spin_lock(&rq->lock);
5080 #endif /* CONFIG_HOTPLUG_CPU */
5082 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5084 static struct ctl_table sd_ctl_dir[] = {
5086 .procname = "sched_domain",
5092 static struct ctl_table sd_ctl_root[] = {
5094 .procname = "kernel",
5096 .child = sd_ctl_dir,
5101 static struct ctl_table *sd_alloc_ctl_entry(int n)
5103 struct ctl_table *entry =
5104 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5109 static void sd_free_ctl_entry(struct ctl_table **tablep)
5111 struct ctl_table *entry;
5114 * In the intermediate directories, both the child directory and
5115 * procname are dynamically allocated and could fail but the mode
5116 * will always be set. In the lowest directory the names are
5117 * static strings and all have proc handlers.
5119 for (entry = *tablep; entry->mode; entry++) {
5121 sd_free_ctl_entry(&entry->child);
5122 if (entry->proc_handler == NULL)
5123 kfree(entry->procname);
5130 static int min_load_idx = 0;
5131 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5134 set_table_entry(struct ctl_table *entry,
5135 const char *procname, void *data, int maxlen,
5136 umode_t mode, proc_handler *proc_handler,
5139 entry->procname = procname;
5141 entry->maxlen = maxlen;
5143 entry->proc_handler = proc_handler;
5146 entry->extra1 = &min_load_idx;
5147 entry->extra2 = &max_load_idx;
5151 static struct ctl_table *
5152 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5154 struct ctl_table *table = sd_alloc_ctl_entry(14);
5159 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5160 sizeof(long), 0644, proc_doulongvec_minmax, false);
5161 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5162 sizeof(long), 0644, proc_doulongvec_minmax, false);
5163 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5164 sizeof(int), 0644, proc_dointvec_minmax, true);
5165 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5166 sizeof(int), 0644, proc_dointvec_minmax, true);
5167 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5168 sizeof(int), 0644, proc_dointvec_minmax, true);
5169 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5170 sizeof(int), 0644, proc_dointvec_minmax, true);
5171 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5172 sizeof(int), 0644, proc_dointvec_minmax, true);
5173 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5174 sizeof(int), 0644, proc_dointvec_minmax, false);
5175 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5176 sizeof(int), 0644, proc_dointvec_minmax, false);
5177 set_table_entry(&table[9], "cache_nice_tries",
5178 &sd->cache_nice_tries,
5179 sizeof(int), 0644, proc_dointvec_minmax, false);
5180 set_table_entry(&table[10], "flags", &sd->flags,
5181 sizeof(int), 0644, proc_dointvec_minmax, false);
5182 set_table_entry(&table[11], "max_newidle_lb_cost",
5183 &sd->max_newidle_lb_cost,
5184 sizeof(long), 0644, proc_doulongvec_minmax, false);
5185 set_table_entry(&table[12], "name", sd->name,
5186 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5187 /* &table[13] is terminator */
5192 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5194 struct ctl_table *entry, *table;
5195 struct sched_domain *sd;
5196 int domain_num = 0, i;
5199 for_each_domain(cpu, sd)
5201 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5206 for_each_domain(cpu, sd) {
5207 snprintf(buf, 32, "domain%d", i);
5208 entry->procname = kstrdup(buf, GFP_KERNEL);
5210 entry->child = sd_alloc_ctl_domain_table(sd);
5217 static struct ctl_table_header *sd_sysctl_header;
5218 static void register_sched_domain_sysctl(void)
5220 int i, cpu_num = num_possible_cpus();
5221 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5224 WARN_ON(sd_ctl_dir[0].child);
5225 sd_ctl_dir[0].child = entry;
5230 for_each_possible_cpu(i) {
5231 snprintf(buf, 32, "cpu%d", i);
5232 entry->procname = kstrdup(buf, GFP_KERNEL);
5234 entry->child = sd_alloc_ctl_cpu_table(i);
5238 WARN_ON(sd_sysctl_header);
5239 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5242 /* may be called multiple times per register */
5243 static void unregister_sched_domain_sysctl(void)
5245 if (sd_sysctl_header)
5246 unregister_sysctl_table(sd_sysctl_header);
5247 sd_sysctl_header = NULL;
5248 if (sd_ctl_dir[0].child)
5249 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5252 static void register_sched_domain_sysctl(void)
5255 static void unregister_sched_domain_sysctl(void)
5260 static void set_rq_online(struct rq *rq)
5263 const struct sched_class *class;
5265 cpumask_set_cpu(rq->cpu, rq->rd->online);
5268 for_each_class(class) {
5269 if (class->rq_online)
5270 class->rq_online(rq);
5275 static void set_rq_offline(struct rq *rq)
5278 const struct sched_class *class;
5280 for_each_class(class) {
5281 if (class->rq_offline)
5282 class->rq_offline(rq);
5285 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5291 * migration_call - callback that gets triggered when a CPU is added.
5292 * Here we can start up the necessary migration thread for the new CPU.
5295 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5297 int cpu = (long)hcpu;
5298 unsigned long flags;
5299 struct rq *rq = cpu_rq(cpu);
5301 switch (action & ~CPU_TASKS_FROZEN) {
5303 case CPU_UP_PREPARE:
5304 rq->calc_load_update = calc_load_update;
5308 /* Update our root-domain */
5309 raw_spin_lock_irqsave(&rq->lock, flags);
5311 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5315 raw_spin_unlock_irqrestore(&rq->lock, flags);
5318 #ifdef CONFIG_HOTPLUG_CPU
5320 sched_ttwu_pending();
5321 /* Update our root-domain */
5322 raw_spin_lock_irqsave(&rq->lock, flags);
5324 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5328 BUG_ON(rq->nr_running != 1); /* the migration thread */
5329 raw_spin_unlock_irqrestore(&rq->lock, flags);
5333 calc_load_migrate(rq);
5338 update_max_interval();
5344 * Register at high priority so that task migration (migrate_all_tasks)
5345 * happens before everything else. This has to be lower priority than
5346 * the notifier in the perf_event subsystem, though.
5348 static struct notifier_block migration_notifier = {
5349 .notifier_call = migration_call,
5350 .priority = CPU_PRI_MIGRATION,
5353 static void set_cpu_rq_start_time(void)
5355 int cpu = smp_processor_id();
5356 struct rq *rq = cpu_rq(cpu);
5357 rq->age_stamp = sched_clock_cpu(cpu);
5360 static int sched_cpu_active(struct notifier_block *nfb,
5361 unsigned long action, void *hcpu)
5363 switch (action & ~CPU_TASKS_FROZEN) {
5365 set_cpu_rq_start_time();
5367 case CPU_DOWN_FAILED:
5368 set_cpu_active((long)hcpu, true);
5375 static int sched_cpu_inactive(struct notifier_block *nfb,
5376 unsigned long action, void *hcpu)
5378 switch (action & ~CPU_TASKS_FROZEN) {
5379 case CPU_DOWN_PREPARE:
5380 set_cpu_active((long)hcpu, false);
5387 static int __init migration_init(void)
5389 void *cpu = (void *)(long)smp_processor_id();
5392 /* Initialize migration for the boot CPU */
5393 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5394 BUG_ON(err == NOTIFY_BAD);
5395 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5396 register_cpu_notifier(&migration_notifier);
5398 /* Register cpu active notifiers */
5399 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5400 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5404 early_initcall(migration_init);
5409 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5411 #ifdef CONFIG_SCHED_DEBUG
5413 static __read_mostly int sched_debug_enabled;
5415 static int __init sched_debug_setup(char *str)
5417 sched_debug_enabled = 1;
5421 early_param("sched_debug", sched_debug_setup);
5423 static inline bool sched_debug(void)
5425 return sched_debug_enabled;
5428 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5429 struct cpumask *groupmask)
5431 struct sched_group *group = sd->groups;
5433 cpumask_clear(groupmask);
5435 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5437 if (!(sd->flags & SD_LOAD_BALANCE)) {
5438 printk("does not load-balance\n");
5440 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5445 printk(KERN_CONT "span %*pbl level %s\n",
5446 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5448 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5449 printk(KERN_ERR "ERROR: domain->span does not contain "
5452 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5453 printk(KERN_ERR "ERROR: domain->groups does not contain"
5457 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5461 printk(KERN_ERR "ERROR: group is NULL\n");
5465 if (!cpumask_weight(sched_group_cpus(group))) {
5466 printk(KERN_CONT "\n");
5467 printk(KERN_ERR "ERROR: empty group\n");
5471 if (!(sd->flags & SD_OVERLAP) &&
5472 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5473 printk(KERN_CONT "\n");
5474 printk(KERN_ERR "ERROR: repeated CPUs\n");
5478 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5480 printk(KERN_CONT " %*pbl",
5481 cpumask_pr_args(sched_group_cpus(group)));
5482 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5483 printk(KERN_CONT " (cpu_capacity = %d)",
5484 group->sgc->capacity);
5487 group = group->next;
5488 } while (group != sd->groups);
5489 printk(KERN_CONT "\n");
5491 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5492 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5495 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5496 printk(KERN_ERR "ERROR: parent span is not a superset "
5497 "of domain->span\n");
5501 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5505 if (!sched_debug_enabled)
5509 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5513 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5516 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5524 #else /* !CONFIG_SCHED_DEBUG */
5525 # define sched_domain_debug(sd, cpu) do { } while (0)
5526 static inline bool sched_debug(void)
5530 #endif /* CONFIG_SCHED_DEBUG */
5532 static int sd_degenerate(struct sched_domain *sd)
5534 if (cpumask_weight(sched_domain_span(sd)) == 1)
5537 /* Following flags need at least 2 groups */
5538 if (sd->flags & (SD_LOAD_BALANCE |
5539 SD_BALANCE_NEWIDLE |
5542 SD_SHARE_CPUCAPACITY |
5543 SD_SHARE_PKG_RESOURCES |
5544 SD_SHARE_POWERDOMAIN)) {
5545 if (sd->groups != sd->groups->next)
5549 /* Following flags don't use groups */
5550 if (sd->flags & (SD_WAKE_AFFINE))
5557 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5559 unsigned long cflags = sd->flags, pflags = parent->flags;
5561 if (sd_degenerate(parent))
5564 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5567 /* Flags needing groups don't count if only 1 group in parent */
5568 if (parent->groups == parent->groups->next) {
5569 pflags &= ~(SD_LOAD_BALANCE |
5570 SD_BALANCE_NEWIDLE |
5573 SD_SHARE_CPUCAPACITY |
5574 SD_SHARE_PKG_RESOURCES |
5576 SD_SHARE_POWERDOMAIN);
5577 if (nr_node_ids == 1)
5578 pflags &= ~SD_SERIALIZE;
5580 if (~cflags & pflags)
5586 static void free_rootdomain(struct rcu_head *rcu)
5588 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5590 cpupri_cleanup(&rd->cpupri);
5591 cpudl_cleanup(&rd->cpudl);
5592 free_cpumask_var(rd->dlo_mask);
5593 free_cpumask_var(rd->rto_mask);
5594 free_cpumask_var(rd->online);
5595 free_cpumask_var(rd->span);
5599 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5601 struct root_domain *old_rd = NULL;
5602 unsigned long flags;
5604 raw_spin_lock_irqsave(&rq->lock, flags);
5609 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5612 cpumask_clear_cpu(rq->cpu, old_rd->span);
5615 * If we dont want to free the old_rd yet then
5616 * set old_rd to NULL to skip the freeing later
5619 if (!atomic_dec_and_test(&old_rd->refcount))
5623 atomic_inc(&rd->refcount);
5626 cpumask_set_cpu(rq->cpu, rd->span);
5627 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5630 raw_spin_unlock_irqrestore(&rq->lock, flags);
5633 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5636 static int init_rootdomain(struct root_domain *rd)
5638 memset(rd, 0, sizeof(*rd));
5640 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5642 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5644 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5646 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5649 init_dl_bw(&rd->dl_bw);
5650 if (cpudl_init(&rd->cpudl) != 0)
5653 if (cpupri_init(&rd->cpupri) != 0)
5658 free_cpumask_var(rd->rto_mask);
5660 free_cpumask_var(rd->dlo_mask);
5662 free_cpumask_var(rd->online);
5664 free_cpumask_var(rd->span);
5670 * By default the system creates a single root-domain with all cpus as
5671 * members (mimicking the global state we have today).
5673 struct root_domain def_root_domain;
5675 static void init_defrootdomain(void)
5677 init_rootdomain(&def_root_domain);
5679 atomic_set(&def_root_domain.refcount, 1);
5682 static struct root_domain *alloc_rootdomain(void)
5684 struct root_domain *rd;
5686 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5690 if (init_rootdomain(rd) != 0) {
5698 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5700 struct sched_group *tmp, *first;
5709 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5714 } while (sg != first);
5717 static void free_sched_domain(struct rcu_head *rcu)
5719 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5722 * If its an overlapping domain it has private groups, iterate and
5725 if (sd->flags & SD_OVERLAP) {
5726 free_sched_groups(sd->groups, 1);
5727 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5728 kfree(sd->groups->sgc);
5734 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5736 call_rcu(&sd->rcu, free_sched_domain);
5739 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5741 for (; sd; sd = sd->parent)
5742 destroy_sched_domain(sd, cpu);
5746 * Keep a special pointer to the highest sched_domain that has
5747 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5748 * allows us to avoid some pointer chasing select_idle_sibling().
5750 * Also keep a unique ID per domain (we use the first cpu number in
5751 * the cpumask of the domain), this allows us to quickly tell if
5752 * two cpus are in the same cache domain, see cpus_share_cache().
5754 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5755 DEFINE_PER_CPU(int, sd_llc_size);
5756 DEFINE_PER_CPU(int, sd_llc_id);
5757 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5758 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5759 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5761 static void update_top_cache_domain(int cpu)
5763 struct sched_domain *sd;
5764 struct sched_domain *busy_sd = NULL;
5768 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5770 id = cpumask_first(sched_domain_span(sd));
5771 size = cpumask_weight(sched_domain_span(sd));
5772 busy_sd = sd->parent; /* sd_busy */
5774 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5776 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5777 per_cpu(sd_llc_size, cpu) = size;
5778 per_cpu(sd_llc_id, cpu) = id;
5780 sd = lowest_flag_domain(cpu, SD_NUMA);
5781 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5783 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5784 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5788 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5789 * hold the hotplug lock.
5792 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5794 struct rq *rq = cpu_rq(cpu);
5795 struct sched_domain *tmp;
5797 /* Remove the sched domains which do not contribute to scheduling. */
5798 for (tmp = sd; tmp; ) {
5799 struct sched_domain *parent = tmp->parent;
5803 if (sd_parent_degenerate(tmp, parent)) {
5804 tmp->parent = parent->parent;
5806 parent->parent->child = tmp;
5808 * Transfer SD_PREFER_SIBLING down in case of a
5809 * degenerate parent; the spans match for this
5810 * so the property transfers.
5812 if (parent->flags & SD_PREFER_SIBLING)
5813 tmp->flags |= SD_PREFER_SIBLING;
5814 destroy_sched_domain(parent, cpu);
5819 if (sd && sd_degenerate(sd)) {
5822 destroy_sched_domain(tmp, cpu);
5827 sched_domain_debug(sd, cpu);
5829 rq_attach_root(rq, rd);
5831 rcu_assign_pointer(rq->sd, sd);
5832 destroy_sched_domains(tmp, cpu);
5834 update_top_cache_domain(cpu);
5837 /* Setup the mask of cpus configured for isolated domains */
5838 static int __init isolated_cpu_setup(char *str)
5840 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5841 cpulist_parse(str, cpu_isolated_map);
5845 __setup("isolcpus=", isolated_cpu_setup);
5848 struct sched_domain ** __percpu sd;
5849 struct root_domain *rd;
5860 * Build an iteration mask that can exclude certain CPUs from the upwards
5863 * Asymmetric node setups can result in situations where the domain tree is of
5864 * unequal depth, make sure to skip domains that already cover the entire
5867 * In that case build_sched_domains() will have terminated the iteration early
5868 * and our sibling sd spans will be empty. Domains should always include the
5869 * cpu they're built on, so check that.
5872 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5874 const struct cpumask *span = sched_domain_span(sd);
5875 struct sd_data *sdd = sd->private;
5876 struct sched_domain *sibling;
5879 for_each_cpu(i, span) {
5880 sibling = *per_cpu_ptr(sdd->sd, i);
5881 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5884 cpumask_set_cpu(i, sched_group_mask(sg));
5889 * Return the canonical balance cpu for this group, this is the first cpu
5890 * of this group that's also in the iteration mask.
5892 int group_balance_cpu(struct sched_group *sg)
5894 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5898 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5900 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5901 const struct cpumask *span = sched_domain_span(sd);
5902 struct cpumask *covered = sched_domains_tmpmask;
5903 struct sd_data *sdd = sd->private;
5904 struct sched_domain *sibling;
5907 cpumask_clear(covered);
5909 for_each_cpu(i, span) {
5910 struct cpumask *sg_span;
5912 if (cpumask_test_cpu(i, covered))
5915 sibling = *per_cpu_ptr(sdd->sd, i);
5917 /* See the comment near build_group_mask(). */
5918 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5921 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5922 GFP_KERNEL, cpu_to_node(cpu));
5927 sg_span = sched_group_cpus(sg);
5929 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5931 cpumask_set_cpu(i, sg_span);
5933 cpumask_or(covered, covered, sg_span);
5935 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5936 if (atomic_inc_return(&sg->sgc->ref) == 1)
5937 build_group_mask(sd, sg);
5940 * Initialize sgc->capacity such that even if we mess up the
5941 * domains and no possible iteration will get us here, we won't
5944 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5947 * Make sure the first group of this domain contains the
5948 * canonical balance cpu. Otherwise the sched_domain iteration
5949 * breaks. See update_sg_lb_stats().
5951 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5952 group_balance_cpu(sg) == cpu)
5962 sd->groups = groups;
5967 free_sched_groups(first, 0);
5972 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5974 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5975 struct sched_domain *child = sd->child;
5978 cpu = cpumask_first(sched_domain_span(child));
5981 *sg = *per_cpu_ptr(sdd->sg, cpu);
5982 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5983 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5990 * build_sched_groups will build a circular linked list of the groups
5991 * covered by the given span, and will set each group's ->cpumask correctly,
5992 * and ->cpu_capacity to 0.
5994 * Assumes the sched_domain tree is fully constructed
5997 build_sched_groups(struct sched_domain *sd, int cpu)
5999 struct sched_group *first = NULL, *last = NULL;
6000 struct sd_data *sdd = sd->private;
6001 const struct cpumask *span = sched_domain_span(sd);
6002 struct cpumask *covered;
6005 get_group(cpu, sdd, &sd->groups);
6006 atomic_inc(&sd->groups->ref);
6008 if (cpu != cpumask_first(span))
6011 lockdep_assert_held(&sched_domains_mutex);
6012 covered = sched_domains_tmpmask;
6014 cpumask_clear(covered);
6016 for_each_cpu(i, span) {
6017 struct sched_group *sg;
6020 if (cpumask_test_cpu(i, covered))
6023 group = get_group(i, sdd, &sg);
6024 cpumask_setall(sched_group_mask(sg));
6026 for_each_cpu(j, span) {
6027 if (get_group(j, sdd, NULL) != group)
6030 cpumask_set_cpu(j, covered);
6031 cpumask_set_cpu(j, sched_group_cpus(sg));
6046 * Initialize sched groups cpu_capacity.
6048 * cpu_capacity indicates the capacity of sched group, which is used while
6049 * distributing the load between different sched groups in a sched domain.
6050 * Typically cpu_capacity for all the groups in a sched domain will be same
6051 * unless there are asymmetries in the topology. If there are asymmetries,
6052 * group having more cpu_capacity will pickup more load compared to the
6053 * group having less cpu_capacity.
6055 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6057 struct sched_group *sg = sd->groups;
6062 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6064 } while (sg != sd->groups);
6066 if (cpu != group_balance_cpu(sg))
6069 update_group_capacity(sd, cpu);
6070 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6074 * Initializers for schedule domains
6075 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6078 static int default_relax_domain_level = -1;
6079 int sched_domain_level_max;
6081 static int __init setup_relax_domain_level(char *str)
6083 if (kstrtoint(str, 0, &default_relax_domain_level))
6084 pr_warn("Unable to set relax_domain_level\n");
6088 __setup("relax_domain_level=", setup_relax_domain_level);
6090 static void set_domain_attribute(struct sched_domain *sd,
6091 struct sched_domain_attr *attr)
6095 if (!attr || attr->relax_domain_level < 0) {
6096 if (default_relax_domain_level < 0)
6099 request = default_relax_domain_level;
6101 request = attr->relax_domain_level;
6102 if (request < sd->level) {
6103 /* turn off idle balance on this domain */
6104 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6106 /* turn on idle balance on this domain */
6107 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6111 static void __sdt_free(const struct cpumask *cpu_map);
6112 static int __sdt_alloc(const struct cpumask *cpu_map);
6114 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6115 const struct cpumask *cpu_map)
6119 if (!atomic_read(&d->rd->refcount))
6120 free_rootdomain(&d->rd->rcu); /* fall through */
6122 free_percpu(d->sd); /* fall through */
6124 __sdt_free(cpu_map); /* fall through */
6130 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6131 const struct cpumask *cpu_map)
6133 memset(d, 0, sizeof(*d));
6135 if (__sdt_alloc(cpu_map))
6136 return sa_sd_storage;
6137 d->sd = alloc_percpu(struct sched_domain *);
6139 return sa_sd_storage;
6140 d->rd = alloc_rootdomain();
6143 return sa_rootdomain;
6147 * NULL the sd_data elements we've used to build the sched_domain and
6148 * sched_group structure so that the subsequent __free_domain_allocs()
6149 * will not free the data we're using.
6151 static void claim_allocations(int cpu, struct sched_domain *sd)
6153 struct sd_data *sdd = sd->private;
6155 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6156 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6158 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6159 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6161 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6162 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6166 static int sched_domains_numa_levels;
6167 enum numa_topology_type sched_numa_topology_type;
6168 static int *sched_domains_numa_distance;
6169 int sched_max_numa_distance;
6170 static struct cpumask ***sched_domains_numa_masks;
6171 static int sched_domains_curr_level;
6175 * SD_flags allowed in topology descriptions.
6177 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6178 * SD_SHARE_PKG_RESOURCES - describes shared caches
6179 * SD_NUMA - describes NUMA topologies
6180 * SD_SHARE_POWERDOMAIN - describes shared power domain
6183 * SD_ASYM_PACKING - describes SMT quirks
6185 #define TOPOLOGY_SD_FLAGS \
6186 (SD_SHARE_CPUCAPACITY | \
6187 SD_SHARE_PKG_RESOURCES | \
6190 SD_SHARE_POWERDOMAIN)
6192 static struct sched_domain *
6193 sd_init(struct sched_domain_topology_level *tl, int cpu)
6195 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6196 int sd_weight, sd_flags = 0;
6200 * Ugly hack to pass state to sd_numa_mask()...
6202 sched_domains_curr_level = tl->numa_level;
6205 sd_weight = cpumask_weight(tl->mask(cpu));
6208 sd_flags = (*tl->sd_flags)();
6209 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6210 "wrong sd_flags in topology description\n"))
6211 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6213 *sd = (struct sched_domain){
6214 .min_interval = sd_weight,
6215 .max_interval = 2*sd_weight,
6217 .imbalance_pct = 125,
6219 .cache_nice_tries = 0,
6226 .flags = 1*SD_LOAD_BALANCE
6227 | 1*SD_BALANCE_NEWIDLE
6232 | 0*SD_SHARE_CPUCAPACITY
6233 | 0*SD_SHARE_PKG_RESOURCES
6235 | 0*SD_PREFER_SIBLING
6240 .last_balance = jiffies,
6241 .balance_interval = sd_weight,
6243 .max_newidle_lb_cost = 0,
6244 .next_decay_max_lb_cost = jiffies,
6245 #ifdef CONFIG_SCHED_DEBUG
6251 * Convert topological properties into behaviour.
6254 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6255 sd->flags |= SD_PREFER_SIBLING;
6256 sd->imbalance_pct = 110;
6257 sd->smt_gain = 1178; /* ~15% */
6259 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6260 sd->imbalance_pct = 117;
6261 sd->cache_nice_tries = 1;
6265 } else if (sd->flags & SD_NUMA) {
6266 sd->cache_nice_tries = 2;
6270 sd->flags |= SD_SERIALIZE;
6271 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6272 sd->flags &= ~(SD_BALANCE_EXEC |
6279 sd->flags |= SD_PREFER_SIBLING;
6280 sd->cache_nice_tries = 1;
6285 sd->private = &tl->data;
6291 * Topology list, bottom-up.
6293 static struct sched_domain_topology_level default_topology[] = {
6294 #ifdef CONFIG_SCHED_SMT
6295 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6297 #ifdef CONFIG_SCHED_MC
6298 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6300 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6304 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6306 #define for_each_sd_topology(tl) \
6307 for (tl = sched_domain_topology; tl->mask; tl++)
6309 void set_sched_topology(struct sched_domain_topology_level *tl)
6311 sched_domain_topology = tl;
6316 static const struct cpumask *sd_numa_mask(int cpu)
6318 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6321 static void sched_numa_warn(const char *str)
6323 static int done = false;
6331 printk(KERN_WARNING "ERROR: %s\n\n", str);
6333 for (i = 0; i < nr_node_ids; i++) {
6334 printk(KERN_WARNING " ");
6335 for (j = 0; j < nr_node_ids; j++)
6336 printk(KERN_CONT "%02d ", node_distance(i,j));
6337 printk(KERN_CONT "\n");
6339 printk(KERN_WARNING "\n");
6342 bool find_numa_distance(int distance)
6346 if (distance == node_distance(0, 0))
6349 for (i = 0; i < sched_domains_numa_levels; i++) {
6350 if (sched_domains_numa_distance[i] == distance)
6358 * A system can have three types of NUMA topology:
6359 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6360 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6361 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6363 * The difference between a glueless mesh topology and a backplane
6364 * topology lies in whether communication between not directly
6365 * connected nodes goes through intermediary nodes (where programs
6366 * could run), or through backplane controllers. This affects
6367 * placement of programs.
6369 * The type of topology can be discerned with the following tests:
6370 * - If the maximum distance between any nodes is 1 hop, the system
6371 * is directly connected.
6372 * - If for two nodes A and B, located N > 1 hops away from each other,
6373 * there is an intermediary node C, which is < N hops away from both
6374 * nodes A and B, the system is a glueless mesh.
6376 static void init_numa_topology_type(void)
6380 n = sched_max_numa_distance;
6383 sched_numa_topology_type = NUMA_DIRECT;
6385 for_each_online_node(a) {
6386 for_each_online_node(b) {
6387 /* Find two nodes furthest removed from each other. */
6388 if (node_distance(a, b) < n)
6391 /* Is there an intermediary node between a and b? */
6392 for_each_online_node(c) {
6393 if (node_distance(a, c) < n &&
6394 node_distance(b, c) < n) {
6395 sched_numa_topology_type =
6401 sched_numa_topology_type = NUMA_BACKPLANE;
6407 static void sched_init_numa(void)
6409 int next_distance, curr_distance = node_distance(0, 0);
6410 struct sched_domain_topology_level *tl;
6414 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6415 if (!sched_domains_numa_distance)
6419 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6420 * unique distances in the node_distance() table.
6422 * Assumes node_distance(0,j) includes all distances in
6423 * node_distance(i,j) in order to avoid cubic time.
6425 next_distance = curr_distance;
6426 for (i = 0; i < nr_node_ids; i++) {
6427 for (j = 0; j < nr_node_ids; j++) {
6428 for (k = 0; k < nr_node_ids; k++) {
6429 int distance = node_distance(i, k);
6431 if (distance > curr_distance &&
6432 (distance < next_distance ||
6433 next_distance == curr_distance))
6434 next_distance = distance;
6437 * While not a strong assumption it would be nice to know
6438 * about cases where if node A is connected to B, B is not
6439 * equally connected to A.
6441 if (sched_debug() && node_distance(k, i) != distance)
6442 sched_numa_warn("Node-distance not symmetric");
6444 if (sched_debug() && i && !find_numa_distance(distance))
6445 sched_numa_warn("Node-0 not representative");
6447 if (next_distance != curr_distance) {
6448 sched_domains_numa_distance[level++] = next_distance;
6449 sched_domains_numa_levels = level;
6450 curr_distance = next_distance;
6455 * In case of sched_debug() we verify the above assumption.
6465 * 'level' contains the number of unique distances, excluding the
6466 * identity distance node_distance(i,i).
6468 * The sched_domains_numa_distance[] array includes the actual distance
6473 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6474 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6475 * the array will contain less then 'level' members. This could be
6476 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6477 * in other functions.
6479 * We reset it to 'level' at the end of this function.
6481 sched_domains_numa_levels = 0;
6483 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6484 if (!sched_domains_numa_masks)
6488 * Now for each level, construct a mask per node which contains all
6489 * cpus of nodes that are that many hops away from us.
6491 for (i = 0; i < level; i++) {
6492 sched_domains_numa_masks[i] =
6493 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6494 if (!sched_domains_numa_masks[i])
6497 for (j = 0; j < nr_node_ids; j++) {
6498 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6502 sched_domains_numa_masks[i][j] = mask;
6504 for (k = 0; k < nr_node_ids; k++) {
6505 if (node_distance(j, k) > sched_domains_numa_distance[i])
6508 cpumask_or(mask, mask, cpumask_of_node(k));
6513 /* Compute default topology size */
6514 for (i = 0; sched_domain_topology[i].mask; i++);
6516 tl = kzalloc((i + level + 1) *
6517 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6522 * Copy the default topology bits..
6524 for (i = 0; sched_domain_topology[i].mask; i++)
6525 tl[i] = sched_domain_topology[i];
6528 * .. and append 'j' levels of NUMA goodness.
6530 for (j = 0; j < level; i++, j++) {
6531 tl[i] = (struct sched_domain_topology_level){
6532 .mask = sd_numa_mask,
6533 .sd_flags = cpu_numa_flags,
6534 .flags = SDTL_OVERLAP,
6540 sched_domain_topology = tl;
6542 sched_domains_numa_levels = level;
6543 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6545 init_numa_topology_type();
6548 static void sched_domains_numa_masks_set(int cpu)
6551 int node = cpu_to_node(cpu);
6553 for (i = 0; i < sched_domains_numa_levels; i++) {
6554 for (j = 0; j < nr_node_ids; j++) {
6555 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6556 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6561 static void sched_domains_numa_masks_clear(int cpu)
6564 for (i = 0; i < sched_domains_numa_levels; i++) {
6565 for (j = 0; j < nr_node_ids; j++)
6566 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6571 * Update sched_domains_numa_masks[level][node] array when new cpus
6574 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6575 unsigned long action,
6578 int cpu = (long)hcpu;
6580 switch (action & ~CPU_TASKS_FROZEN) {
6582 sched_domains_numa_masks_set(cpu);
6586 sched_domains_numa_masks_clear(cpu);
6596 static inline void sched_init_numa(void)
6600 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6601 unsigned long action,
6606 #endif /* CONFIG_NUMA */
6608 static int __sdt_alloc(const struct cpumask *cpu_map)
6610 struct sched_domain_topology_level *tl;
6613 for_each_sd_topology(tl) {
6614 struct sd_data *sdd = &tl->data;
6616 sdd->sd = alloc_percpu(struct sched_domain *);
6620 sdd->sg = alloc_percpu(struct sched_group *);
6624 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6628 for_each_cpu(j, cpu_map) {
6629 struct sched_domain *sd;
6630 struct sched_group *sg;
6631 struct sched_group_capacity *sgc;
6633 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6634 GFP_KERNEL, cpu_to_node(j));
6638 *per_cpu_ptr(sdd->sd, j) = sd;
6640 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6641 GFP_KERNEL, cpu_to_node(j));
6647 *per_cpu_ptr(sdd->sg, j) = sg;
6649 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6650 GFP_KERNEL, cpu_to_node(j));
6654 *per_cpu_ptr(sdd->sgc, j) = sgc;
6661 static void __sdt_free(const struct cpumask *cpu_map)
6663 struct sched_domain_topology_level *tl;
6666 for_each_sd_topology(tl) {
6667 struct sd_data *sdd = &tl->data;
6669 for_each_cpu(j, cpu_map) {
6670 struct sched_domain *sd;
6673 sd = *per_cpu_ptr(sdd->sd, j);
6674 if (sd && (sd->flags & SD_OVERLAP))
6675 free_sched_groups(sd->groups, 0);
6676 kfree(*per_cpu_ptr(sdd->sd, j));
6680 kfree(*per_cpu_ptr(sdd->sg, j));
6682 kfree(*per_cpu_ptr(sdd->sgc, j));
6684 free_percpu(sdd->sd);
6686 free_percpu(sdd->sg);
6688 free_percpu(sdd->sgc);
6693 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6694 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6695 struct sched_domain *child, int cpu)
6697 struct sched_domain *sd = sd_init(tl, cpu);
6701 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6703 sd->level = child->level + 1;
6704 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6708 if (!cpumask_subset(sched_domain_span(child),
6709 sched_domain_span(sd))) {
6710 pr_err("BUG: arch topology borken\n");
6711 #ifdef CONFIG_SCHED_DEBUG
6712 pr_err(" the %s domain not a subset of the %s domain\n",
6713 child->name, sd->name);
6715 /* Fixup, ensure @sd has at least @child cpus. */
6716 cpumask_or(sched_domain_span(sd),
6717 sched_domain_span(sd),
6718 sched_domain_span(child));
6722 set_domain_attribute(sd, attr);
6728 * Build sched domains for a given set of cpus and attach the sched domains
6729 * to the individual cpus
6731 static int build_sched_domains(const struct cpumask *cpu_map,
6732 struct sched_domain_attr *attr)
6734 enum s_alloc alloc_state;
6735 struct sched_domain *sd;
6737 int i, ret = -ENOMEM;
6739 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6740 if (alloc_state != sa_rootdomain)
6743 /* Set up domains for cpus specified by the cpu_map. */
6744 for_each_cpu(i, cpu_map) {
6745 struct sched_domain_topology_level *tl;
6748 for_each_sd_topology(tl) {
6749 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6750 if (tl == sched_domain_topology)
6751 *per_cpu_ptr(d.sd, i) = sd;
6752 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6753 sd->flags |= SD_OVERLAP;
6754 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6759 /* Build the groups for the domains */
6760 for_each_cpu(i, cpu_map) {
6761 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6762 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6763 if (sd->flags & SD_OVERLAP) {
6764 if (build_overlap_sched_groups(sd, i))
6767 if (build_sched_groups(sd, i))
6773 /* Calculate CPU capacity for physical packages and nodes */
6774 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6775 if (!cpumask_test_cpu(i, cpu_map))
6778 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6779 claim_allocations(i, sd);
6780 init_sched_groups_capacity(i, sd);
6784 /* Attach the domains */
6786 for_each_cpu(i, cpu_map) {
6787 sd = *per_cpu_ptr(d.sd, i);
6788 cpu_attach_domain(sd, d.rd, i);
6794 __free_domain_allocs(&d, alloc_state, cpu_map);
6798 static cpumask_var_t *doms_cur; /* current sched domains */
6799 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6800 static struct sched_domain_attr *dattr_cur;
6801 /* attribues of custom domains in 'doms_cur' */
6804 * Special case: If a kmalloc of a doms_cur partition (array of
6805 * cpumask) fails, then fallback to a single sched domain,
6806 * as determined by the single cpumask fallback_doms.
6808 static cpumask_var_t fallback_doms;
6811 * arch_update_cpu_topology lets virtualized architectures update the
6812 * cpu core maps. It is supposed to return 1 if the topology changed
6813 * or 0 if it stayed the same.
6815 int __weak arch_update_cpu_topology(void)
6820 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6823 cpumask_var_t *doms;
6825 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6828 for (i = 0; i < ndoms; i++) {
6829 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6830 free_sched_domains(doms, i);
6837 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6840 for (i = 0; i < ndoms; i++)
6841 free_cpumask_var(doms[i]);
6846 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6847 * For now this just excludes isolated cpus, but could be used to
6848 * exclude other special cases in the future.
6850 static int init_sched_domains(const struct cpumask *cpu_map)
6854 arch_update_cpu_topology();
6856 doms_cur = alloc_sched_domains(ndoms_cur);
6858 doms_cur = &fallback_doms;
6859 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6860 err = build_sched_domains(doms_cur[0], NULL);
6861 register_sched_domain_sysctl();
6867 * Detach sched domains from a group of cpus specified in cpu_map
6868 * These cpus will now be attached to the NULL domain
6870 static void detach_destroy_domains(const struct cpumask *cpu_map)
6875 for_each_cpu(i, cpu_map)
6876 cpu_attach_domain(NULL, &def_root_domain, i);
6880 /* handle null as "default" */
6881 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6882 struct sched_domain_attr *new, int idx_new)
6884 struct sched_domain_attr tmp;
6891 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6892 new ? (new + idx_new) : &tmp,
6893 sizeof(struct sched_domain_attr));
6897 * Partition sched domains as specified by the 'ndoms_new'
6898 * cpumasks in the array doms_new[] of cpumasks. This compares
6899 * doms_new[] to the current sched domain partitioning, doms_cur[].
6900 * It destroys each deleted domain and builds each new domain.
6902 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6903 * The masks don't intersect (don't overlap.) We should setup one
6904 * sched domain for each mask. CPUs not in any of the cpumasks will
6905 * not be load balanced. If the same cpumask appears both in the
6906 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6909 * The passed in 'doms_new' should be allocated using
6910 * alloc_sched_domains. This routine takes ownership of it and will
6911 * free_sched_domains it when done with it. If the caller failed the
6912 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6913 * and partition_sched_domains() will fallback to the single partition
6914 * 'fallback_doms', it also forces the domains to be rebuilt.
6916 * If doms_new == NULL it will be replaced with cpu_online_mask.
6917 * ndoms_new == 0 is a special case for destroying existing domains,
6918 * and it will not create the default domain.
6920 * Call with hotplug lock held
6922 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6923 struct sched_domain_attr *dattr_new)
6928 mutex_lock(&sched_domains_mutex);
6930 /* always unregister in case we don't destroy any domains */
6931 unregister_sched_domain_sysctl();
6933 /* Let architecture update cpu core mappings. */
6934 new_topology = arch_update_cpu_topology();
6936 n = doms_new ? ndoms_new : 0;
6938 /* Destroy deleted domains */
6939 for (i = 0; i < ndoms_cur; i++) {
6940 for (j = 0; j < n && !new_topology; j++) {
6941 if (cpumask_equal(doms_cur[i], doms_new[j])
6942 && dattrs_equal(dattr_cur, i, dattr_new, j))
6945 /* no match - a current sched domain not in new doms_new[] */
6946 detach_destroy_domains(doms_cur[i]);
6952 if (doms_new == NULL) {
6954 doms_new = &fallback_doms;
6955 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6956 WARN_ON_ONCE(dattr_new);
6959 /* Build new domains */
6960 for (i = 0; i < ndoms_new; i++) {
6961 for (j = 0; j < n && !new_topology; j++) {
6962 if (cpumask_equal(doms_new[i], doms_cur[j])
6963 && dattrs_equal(dattr_new, i, dattr_cur, j))
6966 /* no match - add a new doms_new */
6967 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6972 /* Remember the new sched domains */
6973 if (doms_cur != &fallback_doms)
6974 free_sched_domains(doms_cur, ndoms_cur);
6975 kfree(dattr_cur); /* kfree(NULL) is safe */
6976 doms_cur = doms_new;
6977 dattr_cur = dattr_new;
6978 ndoms_cur = ndoms_new;
6980 register_sched_domain_sysctl();
6982 mutex_unlock(&sched_domains_mutex);
6985 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6988 * Update cpusets according to cpu_active mask. If cpusets are
6989 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6990 * around partition_sched_domains().
6992 * If we come here as part of a suspend/resume, don't touch cpusets because we
6993 * want to restore it back to its original state upon resume anyway.
6995 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6999 case CPU_ONLINE_FROZEN:
7000 case CPU_DOWN_FAILED_FROZEN:
7003 * num_cpus_frozen tracks how many CPUs are involved in suspend
7004 * resume sequence. As long as this is not the last online
7005 * operation in the resume sequence, just build a single sched
7006 * domain, ignoring cpusets.
7009 if (likely(num_cpus_frozen)) {
7010 partition_sched_domains(1, NULL, NULL);
7015 * This is the last CPU online operation. So fall through and
7016 * restore the original sched domains by considering the
7017 * cpuset configurations.
7021 cpuset_update_active_cpus(true);
7029 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7032 unsigned long flags;
7033 long cpu = (long)hcpu;
7039 case CPU_DOWN_PREPARE:
7040 rcu_read_lock_sched();
7041 dl_b = dl_bw_of(cpu);
7043 raw_spin_lock_irqsave(&dl_b->lock, flags);
7044 cpus = dl_bw_cpus(cpu);
7045 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7046 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7048 rcu_read_unlock_sched();
7051 return notifier_from_errno(-EBUSY);
7052 cpuset_update_active_cpus(false);
7054 case CPU_DOWN_PREPARE_FROZEN:
7056 partition_sched_domains(1, NULL, NULL);
7064 void __init sched_init_smp(void)
7066 cpumask_var_t non_isolated_cpus;
7068 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7069 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7074 * There's no userspace yet to cause hotplug operations; hence all the
7075 * cpu masks are stable and all blatant races in the below code cannot
7078 mutex_lock(&sched_domains_mutex);
7079 init_sched_domains(cpu_active_mask);
7080 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7081 if (cpumask_empty(non_isolated_cpus))
7082 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7083 mutex_unlock(&sched_domains_mutex);
7085 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7086 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7087 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7091 /* Move init over to a non-isolated CPU */
7092 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7094 sched_init_granularity();
7095 free_cpumask_var(non_isolated_cpus);
7097 init_sched_rt_class();
7098 init_sched_dl_class();
7101 void __init sched_init_smp(void)
7103 sched_init_granularity();
7105 #endif /* CONFIG_SMP */
7107 const_debug unsigned int sysctl_timer_migration = 1;
7109 int in_sched_functions(unsigned long addr)
7111 return in_lock_functions(addr) ||
7112 (addr >= (unsigned long)__sched_text_start
7113 && addr < (unsigned long)__sched_text_end);
7116 #ifdef CONFIG_CGROUP_SCHED
7118 * Default task group.
7119 * Every task in system belongs to this group at bootup.
7121 struct task_group root_task_group;
7122 LIST_HEAD(task_groups);
7125 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7127 void __init sched_init(void)
7130 unsigned long alloc_size = 0, ptr;
7132 #ifdef CONFIG_FAIR_GROUP_SCHED
7133 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7135 #ifdef CONFIG_RT_GROUP_SCHED
7136 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7139 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7141 #ifdef CONFIG_FAIR_GROUP_SCHED
7142 root_task_group.se = (struct sched_entity **)ptr;
7143 ptr += nr_cpu_ids * sizeof(void **);
7145 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7146 ptr += nr_cpu_ids * sizeof(void **);
7148 #endif /* CONFIG_FAIR_GROUP_SCHED */
7149 #ifdef CONFIG_RT_GROUP_SCHED
7150 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7151 ptr += nr_cpu_ids * sizeof(void **);
7153 root_task_group.rt_rq = (struct rt_rq **)ptr;
7154 ptr += nr_cpu_ids * sizeof(void **);
7156 #endif /* CONFIG_RT_GROUP_SCHED */
7158 #ifdef CONFIG_CPUMASK_OFFSTACK
7159 for_each_possible_cpu(i) {
7160 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7161 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7163 #endif /* CONFIG_CPUMASK_OFFSTACK */
7165 init_rt_bandwidth(&def_rt_bandwidth,
7166 global_rt_period(), global_rt_runtime());
7167 init_dl_bandwidth(&def_dl_bandwidth,
7168 global_rt_period(), global_rt_runtime());
7171 init_defrootdomain();
7174 #ifdef CONFIG_RT_GROUP_SCHED
7175 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7176 global_rt_period(), global_rt_runtime());
7177 #endif /* CONFIG_RT_GROUP_SCHED */
7179 #ifdef CONFIG_CGROUP_SCHED
7180 list_add(&root_task_group.list, &task_groups);
7181 INIT_LIST_HEAD(&root_task_group.children);
7182 INIT_LIST_HEAD(&root_task_group.siblings);
7183 autogroup_init(&init_task);
7185 #endif /* CONFIG_CGROUP_SCHED */
7187 for_each_possible_cpu(i) {
7191 raw_spin_lock_init(&rq->lock);
7193 rq->calc_load_active = 0;
7194 rq->calc_load_update = jiffies + LOAD_FREQ;
7195 init_cfs_rq(&rq->cfs);
7196 init_rt_rq(&rq->rt);
7197 init_dl_rq(&rq->dl);
7198 #ifdef CONFIG_FAIR_GROUP_SCHED
7199 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7200 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7202 * How much cpu bandwidth does root_task_group get?
7204 * In case of task-groups formed thr' the cgroup filesystem, it
7205 * gets 100% of the cpu resources in the system. This overall
7206 * system cpu resource is divided among the tasks of
7207 * root_task_group and its child task-groups in a fair manner,
7208 * based on each entity's (task or task-group's) weight
7209 * (se->load.weight).
7211 * In other words, if root_task_group has 10 tasks of weight
7212 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7213 * then A0's share of the cpu resource is:
7215 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7217 * We achieve this by letting root_task_group's tasks sit
7218 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7220 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7221 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7222 #endif /* CONFIG_FAIR_GROUP_SCHED */
7224 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7225 #ifdef CONFIG_RT_GROUP_SCHED
7226 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7229 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7230 rq->cpu_load[j] = 0;
7232 rq->last_load_update_tick = jiffies;
7237 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7238 rq->post_schedule = 0;
7239 rq->active_balance = 0;
7240 rq->next_balance = jiffies;
7245 rq->avg_idle = 2*sysctl_sched_migration_cost;
7246 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7248 INIT_LIST_HEAD(&rq->cfs_tasks);
7250 rq_attach_root(rq, &def_root_domain);
7251 #ifdef CONFIG_NO_HZ_COMMON
7254 #ifdef CONFIG_NO_HZ_FULL
7255 rq->last_sched_tick = 0;
7259 atomic_set(&rq->nr_iowait, 0);
7262 set_load_weight(&init_task);
7264 #ifdef CONFIG_PREEMPT_NOTIFIERS
7265 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7269 * The boot idle thread does lazy MMU switching as well:
7271 atomic_inc(&init_mm.mm_count);
7272 enter_lazy_tlb(&init_mm, current);
7275 * During early bootup we pretend to be a normal task:
7277 current->sched_class = &fair_sched_class;
7280 * Make us the idle thread. Technically, schedule() should not be
7281 * called from this thread, however somewhere below it might be,
7282 * but because we are the idle thread, we just pick up running again
7283 * when this runqueue becomes "idle".
7285 init_idle(current, smp_processor_id());
7287 calc_load_update = jiffies + LOAD_FREQ;
7290 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7291 /* May be allocated at isolcpus cmdline parse time */
7292 if (cpu_isolated_map == NULL)
7293 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7294 idle_thread_set_boot_cpu();
7295 set_cpu_rq_start_time();
7297 init_sched_fair_class();
7299 scheduler_running = 1;
7302 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7303 static inline int preempt_count_equals(int preempt_offset)
7305 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7307 return (nested == preempt_offset);
7310 void __might_sleep(const char *file, int line, int preempt_offset)
7313 * Blocking primitives will set (and therefore destroy) current->state,
7314 * since we will exit with TASK_RUNNING make sure we enter with it,
7315 * otherwise we will destroy state.
7317 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7318 "do not call blocking ops when !TASK_RUNNING; "
7319 "state=%lx set at [<%p>] %pS\n",
7321 (void *)current->task_state_change,
7322 (void *)current->task_state_change);
7324 ___might_sleep(file, line, preempt_offset);
7326 EXPORT_SYMBOL(__might_sleep);
7328 void ___might_sleep(const char *file, int line, int preempt_offset)
7330 static unsigned long prev_jiffy; /* ratelimiting */
7332 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7333 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7334 !is_idle_task(current)) ||
7335 system_state != SYSTEM_RUNNING || oops_in_progress)
7337 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7339 prev_jiffy = jiffies;
7342 "BUG: sleeping function called from invalid context at %s:%d\n",
7345 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7346 in_atomic(), irqs_disabled(),
7347 current->pid, current->comm);
7349 if (task_stack_end_corrupted(current))
7350 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7352 debug_show_held_locks(current);
7353 if (irqs_disabled())
7354 print_irqtrace_events(current);
7355 #ifdef CONFIG_DEBUG_PREEMPT
7356 if (!preempt_count_equals(preempt_offset)) {
7357 pr_err("Preemption disabled at:");
7358 print_ip_sym(current->preempt_disable_ip);
7364 EXPORT_SYMBOL(___might_sleep);
7367 #ifdef CONFIG_MAGIC_SYSRQ
7368 static void normalize_task(struct rq *rq, struct task_struct *p)
7370 const struct sched_class *prev_class = p->sched_class;
7371 struct sched_attr attr = {
7372 .sched_policy = SCHED_NORMAL,
7374 int old_prio = p->prio;
7377 queued = task_on_rq_queued(p);
7379 dequeue_task(rq, p, 0);
7380 __setscheduler(rq, p, &attr, false);
7382 enqueue_task(rq, p, 0);
7386 check_class_changed(rq, p, prev_class, old_prio);
7389 void normalize_rt_tasks(void)
7391 struct task_struct *g, *p;
7392 unsigned long flags;
7395 read_lock(&tasklist_lock);
7396 for_each_process_thread(g, p) {
7398 * Only normalize user tasks:
7400 if (p->flags & PF_KTHREAD)
7403 p->se.exec_start = 0;
7404 #ifdef CONFIG_SCHEDSTATS
7405 p->se.statistics.wait_start = 0;
7406 p->se.statistics.sleep_start = 0;
7407 p->se.statistics.block_start = 0;
7410 if (!dl_task(p) && !rt_task(p)) {
7412 * Renice negative nice level userspace
7415 if (task_nice(p) < 0)
7416 set_user_nice(p, 0);
7420 rq = task_rq_lock(p, &flags);
7421 normalize_task(rq, p);
7422 task_rq_unlock(rq, p, &flags);
7424 read_unlock(&tasklist_lock);
7427 #endif /* CONFIG_MAGIC_SYSRQ */
7429 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7431 * These functions are only useful for the IA64 MCA handling, or kdb.
7433 * They can only be called when the whole system has been
7434 * stopped - every CPU needs to be quiescent, and no scheduling
7435 * activity can take place. Using them for anything else would
7436 * be a serious bug, and as a result, they aren't even visible
7437 * under any other configuration.
7441 * curr_task - return the current task for a given cpu.
7442 * @cpu: the processor in question.
7444 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7446 * Return: The current task for @cpu.
7448 struct task_struct *curr_task(int cpu)
7450 return cpu_curr(cpu);
7453 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7457 * set_curr_task - set the current task for a given cpu.
7458 * @cpu: the processor in question.
7459 * @p: the task pointer to set.
7461 * Description: This function must only be used when non-maskable interrupts
7462 * are serviced on a separate stack. It allows the architecture to switch the
7463 * notion of the current task on a cpu in a non-blocking manner. This function
7464 * must be called with all CPU's synchronized, and interrupts disabled, the
7465 * and caller must save the original value of the current task (see
7466 * curr_task() above) and restore that value before reenabling interrupts and
7467 * re-starting the system.
7469 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7471 void set_curr_task(int cpu, struct task_struct *p)
7478 #ifdef CONFIG_CGROUP_SCHED
7479 /* task_group_lock serializes the addition/removal of task groups */
7480 static DEFINE_SPINLOCK(task_group_lock);
7482 static void free_sched_group(struct task_group *tg)
7484 free_fair_sched_group(tg);
7485 free_rt_sched_group(tg);
7490 /* allocate runqueue etc for a new task group */
7491 struct task_group *sched_create_group(struct task_group *parent)
7493 struct task_group *tg;
7495 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7497 return ERR_PTR(-ENOMEM);
7499 if (!alloc_fair_sched_group(tg, parent))
7502 if (!alloc_rt_sched_group(tg, parent))
7508 free_sched_group(tg);
7509 return ERR_PTR(-ENOMEM);
7512 void sched_online_group(struct task_group *tg, struct task_group *parent)
7514 unsigned long flags;
7516 spin_lock_irqsave(&task_group_lock, flags);
7517 list_add_rcu(&tg->list, &task_groups);
7519 WARN_ON(!parent); /* root should already exist */
7521 tg->parent = parent;
7522 INIT_LIST_HEAD(&tg->children);
7523 list_add_rcu(&tg->siblings, &parent->children);
7524 spin_unlock_irqrestore(&task_group_lock, flags);
7527 /* rcu callback to free various structures associated with a task group */
7528 static void free_sched_group_rcu(struct rcu_head *rhp)
7530 /* now it should be safe to free those cfs_rqs */
7531 free_sched_group(container_of(rhp, struct task_group, rcu));
7534 /* Destroy runqueue etc associated with a task group */
7535 void sched_destroy_group(struct task_group *tg)
7537 /* wait for possible concurrent references to cfs_rqs complete */
7538 call_rcu(&tg->rcu, free_sched_group_rcu);
7541 void sched_offline_group(struct task_group *tg)
7543 unsigned long flags;
7546 /* end participation in shares distribution */
7547 for_each_possible_cpu(i)
7548 unregister_fair_sched_group(tg, i);
7550 spin_lock_irqsave(&task_group_lock, flags);
7551 list_del_rcu(&tg->list);
7552 list_del_rcu(&tg->siblings);
7553 spin_unlock_irqrestore(&task_group_lock, flags);
7556 /* change task's runqueue when it moves between groups.
7557 * The caller of this function should have put the task in its new group
7558 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7559 * reflect its new group.
7561 void sched_move_task(struct task_struct *tsk)
7563 struct task_group *tg;
7564 int queued, running;
7565 unsigned long flags;
7568 rq = task_rq_lock(tsk, &flags);
7570 running = task_current(rq, tsk);
7571 queued = task_on_rq_queued(tsk);
7574 dequeue_task(rq, tsk, 0);
7575 if (unlikely(running))
7576 put_prev_task(rq, tsk);
7579 * All callers are synchronized by task_rq_lock(); we do not use RCU
7580 * which is pointless here. Thus, we pass "true" to task_css_check()
7581 * to prevent lockdep warnings.
7583 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7584 struct task_group, css);
7585 tg = autogroup_task_group(tsk, tg);
7586 tsk->sched_task_group = tg;
7588 #ifdef CONFIG_FAIR_GROUP_SCHED
7589 if (tsk->sched_class->task_move_group)
7590 tsk->sched_class->task_move_group(tsk, queued);
7593 set_task_rq(tsk, task_cpu(tsk));
7595 if (unlikely(running))
7596 tsk->sched_class->set_curr_task(rq);
7598 enqueue_task(rq, tsk, 0);
7600 task_rq_unlock(rq, tsk, &flags);
7602 #endif /* CONFIG_CGROUP_SCHED */
7604 #ifdef CONFIG_RT_GROUP_SCHED
7606 * Ensure that the real time constraints are schedulable.
7608 static DEFINE_MUTEX(rt_constraints_mutex);
7610 /* Must be called with tasklist_lock held */
7611 static inline int tg_has_rt_tasks(struct task_group *tg)
7613 struct task_struct *g, *p;
7616 * Autogroups do not have RT tasks; see autogroup_create().
7618 if (task_group_is_autogroup(tg))
7621 for_each_process_thread(g, p) {
7622 if (rt_task(p) && task_group(p) == tg)
7629 struct rt_schedulable_data {
7630 struct task_group *tg;
7635 static int tg_rt_schedulable(struct task_group *tg, void *data)
7637 struct rt_schedulable_data *d = data;
7638 struct task_group *child;
7639 unsigned long total, sum = 0;
7640 u64 period, runtime;
7642 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7643 runtime = tg->rt_bandwidth.rt_runtime;
7646 period = d->rt_period;
7647 runtime = d->rt_runtime;
7651 * Cannot have more runtime than the period.
7653 if (runtime > period && runtime != RUNTIME_INF)
7657 * Ensure we don't starve existing RT tasks.
7659 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7662 total = to_ratio(period, runtime);
7665 * Nobody can have more than the global setting allows.
7667 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7671 * The sum of our children's runtime should not exceed our own.
7673 list_for_each_entry_rcu(child, &tg->children, siblings) {
7674 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7675 runtime = child->rt_bandwidth.rt_runtime;
7677 if (child == d->tg) {
7678 period = d->rt_period;
7679 runtime = d->rt_runtime;
7682 sum += to_ratio(period, runtime);
7691 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7695 struct rt_schedulable_data data = {
7697 .rt_period = period,
7698 .rt_runtime = runtime,
7702 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7708 static int tg_set_rt_bandwidth(struct task_group *tg,
7709 u64 rt_period, u64 rt_runtime)
7714 * Disallowing the root group RT runtime is BAD, it would disallow the
7715 * kernel creating (and or operating) RT threads.
7717 if (tg == &root_task_group && rt_runtime == 0)
7720 /* No period doesn't make any sense. */
7724 mutex_lock(&rt_constraints_mutex);
7725 read_lock(&tasklist_lock);
7726 err = __rt_schedulable(tg, rt_period, rt_runtime);
7730 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7731 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7732 tg->rt_bandwidth.rt_runtime = rt_runtime;
7734 for_each_possible_cpu(i) {
7735 struct rt_rq *rt_rq = tg->rt_rq[i];
7737 raw_spin_lock(&rt_rq->rt_runtime_lock);
7738 rt_rq->rt_runtime = rt_runtime;
7739 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7741 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7743 read_unlock(&tasklist_lock);
7744 mutex_unlock(&rt_constraints_mutex);
7749 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7751 u64 rt_runtime, rt_period;
7753 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7754 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7755 if (rt_runtime_us < 0)
7756 rt_runtime = RUNTIME_INF;
7758 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7761 static long sched_group_rt_runtime(struct task_group *tg)
7765 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7768 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7769 do_div(rt_runtime_us, NSEC_PER_USEC);
7770 return rt_runtime_us;
7773 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
7775 u64 rt_runtime, rt_period;
7777 rt_period = rt_period_us * NSEC_PER_USEC;
7778 rt_runtime = tg->rt_bandwidth.rt_runtime;
7780 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7783 static long sched_group_rt_period(struct task_group *tg)
7787 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7788 do_div(rt_period_us, NSEC_PER_USEC);
7789 return rt_period_us;
7791 #endif /* CONFIG_RT_GROUP_SCHED */
7793 #ifdef CONFIG_RT_GROUP_SCHED
7794 static int sched_rt_global_constraints(void)
7798 mutex_lock(&rt_constraints_mutex);
7799 read_lock(&tasklist_lock);
7800 ret = __rt_schedulable(NULL, 0, 0);
7801 read_unlock(&tasklist_lock);
7802 mutex_unlock(&rt_constraints_mutex);
7807 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7809 /* Don't accept realtime tasks when there is no way for them to run */
7810 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7816 #else /* !CONFIG_RT_GROUP_SCHED */
7817 static int sched_rt_global_constraints(void)
7819 unsigned long flags;
7822 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7823 for_each_possible_cpu(i) {
7824 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7826 raw_spin_lock(&rt_rq->rt_runtime_lock);
7827 rt_rq->rt_runtime = global_rt_runtime();
7828 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7830 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7834 #endif /* CONFIG_RT_GROUP_SCHED */
7836 static int sched_dl_global_validate(void)
7838 u64 runtime = global_rt_runtime();
7839 u64 period = global_rt_period();
7840 u64 new_bw = to_ratio(period, runtime);
7843 unsigned long flags;
7846 * Here we want to check the bandwidth not being set to some
7847 * value smaller than the currently allocated bandwidth in
7848 * any of the root_domains.
7850 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7851 * cycling on root_domains... Discussion on different/better
7852 * solutions is welcome!
7854 for_each_possible_cpu(cpu) {
7855 rcu_read_lock_sched();
7856 dl_b = dl_bw_of(cpu);
7858 raw_spin_lock_irqsave(&dl_b->lock, flags);
7859 if (new_bw < dl_b->total_bw)
7861 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7863 rcu_read_unlock_sched();
7872 static void sched_dl_do_global(void)
7877 unsigned long flags;
7879 def_dl_bandwidth.dl_period = global_rt_period();
7880 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7882 if (global_rt_runtime() != RUNTIME_INF)
7883 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7886 * FIXME: As above...
7888 for_each_possible_cpu(cpu) {
7889 rcu_read_lock_sched();
7890 dl_b = dl_bw_of(cpu);
7892 raw_spin_lock_irqsave(&dl_b->lock, flags);
7894 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7896 rcu_read_unlock_sched();
7900 static int sched_rt_global_validate(void)
7902 if (sysctl_sched_rt_period <= 0)
7905 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7906 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7912 static void sched_rt_do_global(void)
7914 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7915 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7918 int sched_rt_handler(struct ctl_table *table, int write,
7919 void __user *buffer, size_t *lenp,
7922 int old_period, old_runtime;
7923 static DEFINE_MUTEX(mutex);
7927 old_period = sysctl_sched_rt_period;
7928 old_runtime = sysctl_sched_rt_runtime;
7930 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7932 if (!ret && write) {
7933 ret = sched_rt_global_validate();
7937 ret = sched_dl_global_validate();
7941 ret = sched_rt_global_constraints();
7945 sched_rt_do_global();
7946 sched_dl_do_global();
7950 sysctl_sched_rt_period = old_period;
7951 sysctl_sched_rt_runtime = old_runtime;
7953 mutex_unlock(&mutex);
7958 int sched_rr_handler(struct ctl_table *table, int write,
7959 void __user *buffer, size_t *lenp,
7963 static DEFINE_MUTEX(mutex);
7966 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7967 /* make sure that internally we keep jiffies */
7968 /* also, writing zero resets timeslice to default */
7969 if (!ret && write) {
7970 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7971 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7973 mutex_unlock(&mutex);
7977 #ifdef CONFIG_CGROUP_SCHED
7979 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7981 return css ? container_of(css, struct task_group, css) : NULL;
7984 static struct cgroup_subsys_state *
7985 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7987 struct task_group *parent = css_tg(parent_css);
7988 struct task_group *tg;
7991 /* This is early initialization for the top cgroup */
7992 return &root_task_group.css;
7995 tg = sched_create_group(parent);
7997 return ERR_PTR(-ENOMEM);
8002 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8004 struct task_group *tg = css_tg(css);
8005 struct task_group *parent = css_tg(css->parent);
8008 sched_online_group(tg, parent);
8012 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8014 struct task_group *tg = css_tg(css);
8016 sched_destroy_group(tg);
8019 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8021 struct task_group *tg = css_tg(css);
8023 sched_offline_group(tg);
8026 static void cpu_cgroup_fork(struct task_struct *task)
8028 sched_move_task(task);
8031 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8032 struct cgroup_taskset *tset)
8034 struct task_struct *task;
8036 cgroup_taskset_for_each(task, tset) {
8037 #ifdef CONFIG_RT_GROUP_SCHED
8038 if (!sched_rt_can_attach(css_tg(css), task))
8041 /* We don't support RT-tasks being in separate groups */
8042 if (task->sched_class != &fair_sched_class)
8049 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8050 struct cgroup_taskset *tset)
8052 struct task_struct *task;
8054 cgroup_taskset_for_each(task, tset)
8055 sched_move_task(task);
8058 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8059 struct cgroup_subsys_state *old_css,
8060 struct task_struct *task)
8063 * cgroup_exit() is called in the copy_process() failure path.
8064 * Ignore this case since the task hasn't ran yet, this avoids
8065 * trying to poke a half freed task state from generic code.
8067 if (!(task->flags & PF_EXITING))
8070 sched_move_task(task);
8073 #ifdef CONFIG_FAIR_GROUP_SCHED
8074 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8075 struct cftype *cftype, u64 shareval)
8077 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8080 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8083 struct task_group *tg = css_tg(css);
8085 return (u64) scale_load_down(tg->shares);
8088 #ifdef CONFIG_CFS_BANDWIDTH
8089 static DEFINE_MUTEX(cfs_constraints_mutex);
8091 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8092 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8094 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8096 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8098 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8099 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8101 if (tg == &root_task_group)
8105 * Ensure we have at some amount of bandwidth every period. This is
8106 * to prevent reaching a state of large arrears when throttled via
8107 * entity_tick() resulting in prolonged exit starvation.
8109 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8113 * Likewise, bound things on the otherside by preventing insane quota
8114 * periods. This also allows us to normalize in computing quota
8117 if (period > max_cfs_quota_period)
8121 * Prevent race between setting of cfs_rq->runtime_enabled and
8122 * unthrottle_offline_cfs_rqs().
8125 mutex_lock(&cfs_constraints_mutex);
8126 ret = __cfs_schedulable(tg, period, quota);
8130 runtime_enabled = quota != RUNTIME_INF;
8131 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8133 * If we need to toggle cfs_bandwidth_used, off->on must occur
8134 * before making related changes, and on->off must occur afterwards
8136 if (runtime_enabled && !runtime_was_enabled)
8137 cfs_bandwidth_usage_inc();
8138 raw_spin_lock_irq(&cfs_b->lock);
8139 cfs_b->period = ns_to_ktime(period);
8140 cfs_b->quota = quota;
8142 __refill_cfs_bandwidth_runtime(cfs_b);
8143 /* restart the period timer (if active) to handle new period expiry */
8144 if (runtime_enabled && cfs_b->timer_active) {
8145 /* force a reprogram */
8146 __start_cfs_bandwidth(cfs_b, true);
8148 raw_spin_unlock_irq(&cfs_b->lock);
8150 for_each_online_cpu(i) {
8151 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8152 struct rq *rq = cfs_rq->rq;
8154 raw_spin_lock_irq(&rq->lock);
8155 cfs_rq->runtime_enabled = runtime_enabled;
8156 cfs_rq->runtime_remaining = 0;
8158 if (cfs_rq->throttled)
8159 unthrottle_cfs_rq(cfs_rq);
8160 raw_spin_unlock_irq(&rq->lock);
8162 if (runtime_was_enabled && !runtime_enabled)
8163 cfs_bandwidth_usage_dec();
8165 mutex_unlock(&cfs_constraints_mutex);
8171 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8175 period = ktime_to_ns(tg->cfs_bandwidth.period);
8176 if (cfs_quota_us < 0)
8177 quota = RUNTIME_INF;
8179 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8181 return tg_set_cfs_bandwidth(tg, period, quota);
8184 long tg_get_cfs_quota(struct task_group *tg)
8188 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8191 quota_us = tg->cfs_bandwidth.quota;
8192 do_div(quota_us, NSEC_PER_USEC);
8197 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8201 period = (u64)cfs_period_us * NSEC_PER_USEC;
8202 quota = tg->cfs_bandwidth.quota;
8204 return tg_set_cfs_bandwidth(tg, period, quota);
8207 long tg_get_cfs_period(struct task_group *tg)
8211 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8212 do_div(cfs_period_us, NSEC_PER_USEC);
8214 return cfs_period_us;
8217 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8220 return tg_get_cfs_quota(css_tg(css));
8223 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8224 struct cftype *cftype, s64 cfs_quota_us)
8226 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8229 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8232 return tg_get_cfs_period(css_tg(css));
8235 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8236 struct cftype *cftype, u64 cfs_period_us)
8238 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8241 struct cfs_schedulable_data {
8242 struct task_group *tg;
8247 * normalize group quota/period to be quota/max_period
8248 * note: units are usecs
8250 static u64 normalize_cfs_quota(struct task_group *tg,
8251 struct cfs_schedulable_data *d)
8259 period = tg_get_cfs_period(tg);
8260 quota = tg_get_cfs_quota(tg);
8263 /* note: these should typically be equivalent */
8264 if (quota == RUNTIME_INF || quota == -1)
8267 return to_ratio(period, quota);
8270 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8272 struct cfs_schedulable_data *d = data;
8273 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8274 s64 quota = 0, parent_quota = -1;
8277 quota = RUNTIME_INF;
8279 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8281 quota = normalize_cfs_quota(tg, d);
8282 parent_quota = parent_b->hierarchical_quota;
8285 * ensure max(child_quota) <= parent_quota, inherit when no
8288 if (quota == RUNTIME_INF)
8289 quota = parent_quota;
8290 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8293 cfs_b->hierarchical_quota = quota;
8298 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8301 struct cfs_schedulable_data data = {
8307 if (quota != RUNTIME_INF) {
8308 do_div(data.period, NSEC_PER_USEC);
8309 do_div(data.quota, NSEC_PER_USEC);
8313 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8319 static int cpu_stats_show(struct seq_file *sf, void *v)
8321 struct task_group *tg = css_tg(seq_css(sf));
8322 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8324 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8325 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8326 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8330 #endif /* CONFIG_CFS_BANDWIDTH */
8331 #endif /* CONFIG_FAIR_GROUP_SCHED */
8333 #ifdef CONFIG_RT_GROUP_SCHED
8334 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8335 struct cftype *cft, s64 val)
8337 return sched_group_set_rt_runtime(css_tg(css), val);
8340 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8343 return sched_group_rt_runtime(css_tg(css));
8346 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8347 struct cftype *cftype, u64 rt_period_us)
8349 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8352 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8355 return sched_group_rt_period(css_tg(css));
8357 #endif /* CONFIG_RT_GROUP_SCHED */
8359 static struct cftype cpu_files[] = {
8360 #ifdef CONFIG_FAIR_GROUP_SCHED
8363 .read_u64 = cpu_shares_read_u64,
8364 .write_u64 = cpu_shares_write_u64,
8367 #ifdef CONFIG_CFS_BANDWIDTH
8369 .name = "cfs_quota_us",
8370 .read_s64 = cpu_cfs_quota_read_s64,
8371 .write_s64 = cpu_cfs_quota_write_s64,
8374 .name = "cfs_period_us",
8375 .read_u64 = cpu_cfs_period_read_u64,
8376 .write_u64 = cpu_cfs_period_write_u64,
8380 .seq_show = cpu_stats_show,
8383 #ifdef CONFIG_RT_GROUP_SCHED
8385 .name = "rt_runtime_us",
8386 .read_s64 = cpu_rt_runtime_read,
8387 .write_s64 = cpu_rt_runtime_write,
8390 .name = "rt_period_us",
8391 .read_u64 = cpu_rt_period_read_uint,
8392 .write_u64 = cpu_rt_period_write_uint,
8398 struct cgroup_subsys cpu_cgrp_subsys = {
8399 .css_alloc = cpu_cgroup_css_alloc,
8400 .css_free = cpu_cgroup_css_free,
8401 .css_online = cpu_cgroup_css_online,
8402 .css_offline = cpu_cgroup_css_offline,
8403 .fork = cpu_cgroup_fork,
8404 .can_attach = cpu_cgroup_can_attach,
8405 .attach = cpu_cgroup_attach,
8406 .exit = cpu_cgroup_exit,
8407 .legacy_cftypes = cpu_files,
8411 #endif /* CONFIG_CGROUP_SCHED */
8413 void dump_cpu_task(int cpu)
8415 pr_info("Task dump for CPU %d:\n", cpu);
8416 sched_show_task(cpu_curr(cpu));