4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 DEFINE_MUTEX(sched_domains_mutex);
94 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
96 static void update_rq_clock_task(struct rq *rq, s64 delta);
98 void update_rq_clock(struct rq *rq)
102 lockdep_assert_held(&rq->lock);
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
111 update_rq_clock_task(rq, delta);
115 * Debugging: various feature bits
118 #define SCHED_FEAT(name, enabled) \
119 (1UL << __SCHED_FEAT_##name) * enabled |
121 const_debug unsigned int sysctl_sched_features =
122 #include "features.h"
127 #ifdef CONFIG_SCHED_DEBUG
128 #define SCHED_FEAT(name, enabled) \
131 static const char * const sched_feat_names[] = {
132 #include "features.h"
137 static int sched_feat_show(struct seq_file *m, void *v)
141 for (i = 0; i < __SCHED_FEAT_NR; i++) {
142 if (!(sysctl_sched_features & (1UL << i)))
144 seq_printf(m, "%s ", sched_feat_names[i]);
151 #ifdef HAVE_JUMP_LABEL
153 #define jump_label_key__true STATIC_KEY_INIT_TRUE
154 #define jump_label_key__false STATIC_KEY_INIT_FALSE
156 #define SCHED_FEAT(name, enabled) \
157 jump_label_key__##enabled ,
159 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
160 #include "features.h"
165 static void sched_feat_disable(int i)
167 static_key_disable(&sched_feat_keys[i]);
170 static void sched_feat_enable(int i)
172 static_key_enable(&sched_feat_keys[i]);
175 static void sched_feat_disable(int i) { };
176 static void sched_feat_enable(int i) { };
177 #endif /* HAVE_JUMP_LABEL */
179 static int sched_feat_set(char *cmp)
184 if (strncmp(cmp, "NO_", 3) == 0) {
189 for (i = 0; i < __SCHED_FEAT_NR; i++) {
190 if (strcmp(cmp, sched_feat_names[i]) == 0) {
192 sysctl_sched_features &= ~(1UL << i);
193 sched_feat_disable(i);
195 sysctl_sched_features |= (1UL << i);
196 sched_feat_enable(i);
206 sched_feat_write(struct file *filp, const char __user *ubuf,
207 size_t cnt, loff_t *ppos)
217 if (copy_from_user(&buf, ubuf, cnt))
223 /* Ensure the static_key remains in a consistent state */
224 inode = file_inode(filp);
225 mutex_lock(&inode->i_mutex);
226 i = sched_feat_set(cmp);
227 mutex_unlock(&inode->i_mutex);
228 if (i == __SCHED_FEAT_NR)
236 static int sched_feat_open(struct inode *inode, struct file *filp)
238 return single_open(filp, sched_feat_show, NULL);
241 static const struct file_operations sched_feat_fops = {
242 .open = sched_feat_open,
243 .write = sched_feat_write,
246 .release = single_release,
249 static __init int sched_init_debug(void)
251 debugfs_create_file("sched_features", 0644, NULL, NULL,
256 late_initcall(sched_init_debug);
257 #endif /* CONFIG_SCHED_DEBUG */
260 * Number of tasks to iterate in a single balance run.
261 * Limited because this is done with IRQs disabled.
263 #ifndef CONFIG_PREEMPT_RT_FULL
264 const_debug unsigned int sysctl_sched_nr_migrate = 32;
266 const_debug unsigned int sysctl_sched_nr_migrate = 8;
270 * period over which we average the RT time consumption, measured
275 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
278 * period over which we measure -rt task cpu usage in us.
281 unsigned int sysctl_sched_rt_period = 1000000;
283 __read_mostly int scheduler_running;
286 * part of the period that we allow rt tasks to run in us.
289 int sysctl_sched_rt_runtime = 950000;
291 /* cpus with isolated domains */
292 cpumask_var_t cpu_isolated_map;
295 * this_rq_lock - lock this runqueue and disable interrupts.
297 static struct rq *this_rq_lock(void)
304 raw_spin_lock(&rq->lock);
309 #ifdef CONFIG_SCHED_HRTICK
311 * Use HR-timers to deliver accurate preemption points.
314 static void hrtick_clear(struct rq *rq)
316 if (hrtimer_active(&rq->hrtick_timer))
317 hrtimer_cancel(&rq->hrtick_timer);
321 * High-resolution timer tick.
322 * Runs from hardirq context with interrupts disabled.
324 static enum hrtimer_restart hrtick(struct hrtimer *timer)
326 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
328 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
330 raw_spin_lock(&rq->lock);
332 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
333 raw_spin_unlock(&rq->lock);
335 return HRTIMER_NORESTART;
340 static void __hrtick_restart(struct rq *rq)
342 struct hrtimer *timer = &rq->hrtick_timer;
344 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
348 * called from hardirq (IPI) context
350 static void __hrtick_start(void *arg)
354 raw_spin_lock(&rq->lock);
355 __hrtick_restart(rq);
356 rq->hrtick_csd_pending = 0;
357 raw_spin_unlock(&rq->lock);
361 * Called to set the hrtick timer state.
363 * called with rq->lock held and irqs disabled
365 void hrtick_start(struct rq *rq, u64 delay)
367 struct hrtimer *timer = &rq->hrtick_timer;
372 * Don't schedule slices shorter than 10000ns, that just
373 * doesn't make sense and can cause timer DoS.
375 delta = max_t(s64, delay, 10000LL);
376 time = ktime_add_ns(timer->base->get_time(), delta);
378 hrtimer_set_expires(timer, time);
380 if (rq == this_rq()) {
381 __hrtick_restart(rq);
382 } else if (!rq->hrtick_csd_pending) {
383 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
384 rq->hrtick_csd_pending = 1;
389 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
391 int cpu = (int)(long)hcpu;
394 case CPU_UP_CANCELED:
395 case CPU_UP_CANCELED_FROZEN:
396 case CPU_DOWN_PREPARE:
397 case CPU_DOWN_PREPARE_FROZEN:
399 case CPU_DEAD_FROZEN:
400 hrtick_clear(cpu_rq(cpu));
407 static __init void init_hrtick(void)
409 hotcpu_notifier(hotplug_hrtick, 0);
413 * Called to set the hrtick timer state.
415 * called with rq->lock held and irqs disabled
417 void hrtick_start(struct rq *rq, u64 delay)
420 * Don't schedule slices shorter than 10000ns, that just
421 * doesn't make sense. Rely on vruntime for fairness.
423 delay = max_t(u64, delay, 10000LL);
424 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
425 HRTIMER_MODE_REL_PINNED);
428 static inline void init_hrtick(void)
431 #endif /* CONFIG_SMP */
433 static void init_rq_hrtick(struct rq *rq)
436 rq->hrtick_csd_pending = 0;
438 rq->hrtick_csd.flags = 0;
439 rq->hrtick_csd.func = __hrtick_start;
440 rq->hrtick_csd.info = rq;
443 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
444 rq->hrtick_timer.function = hrtick;
445 rq->hrtick_timer.irqsafe = 1;
447 #else /* CONFIG_SCHED_HRTICK */
448 static inline void hrtick_clear(struct rq *rq)
452 static inline void init_rq_hrtick(struct rq *rq)
456 static inline void init_hrtick(void)
459 #endif /* CONFIG_SCHED_HRTICK */
462 * cmpxchg based fetch_or, macro so it works for different integer types
464 #define fetch_or(ptr, val) \
465 ({ typeof(*(ptr)) __old, __val = *(ptr); \
467 __old = cmpxchg((ptr), __val, __val | (val)); \
468 if (__old == __val) \
475 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
477 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
478 * this avoids any races wrt polling state changes and thereby avoids
481 static bool set_nr_and_not_polling(struct task_struct *p)
483 struct thread_info *ti = task_thread_info(p);
484 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
488 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
490 * If this returns true, then the idle task promises to call
491 * sched_ttwu_pending() and reschedule soon.
493 static bool set_nr_if_polling(struct task_struct *p)
495 struct thread_info *ti = task_thread_info(p);
496 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
499 if (!(val & _TIF_POLLING_NRFLAG))
501 if (val & _TIF_NEED_RESCHED)
503 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
512 static bool set_nr_and_not_polling(struct task_struct *p)
514 set_tsk_need_resched(p);
519 static bool set_nr_if_polling(struct task_struct *p)
526 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
528 struct wake_q_node *node = &task->wake_q;
531 * Atomically grab the task, if ->wake_q is !nil already it means
532 * its already queued (either by us or someone else) and will get the
533 * wakeup due to that.
535 * This cmpxchg() implies a full barrier, which pairs with the write
536 * barrier implied by the wakeup in wake_up_list().
538 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
541 get_task_struct(task);
544 * The head is context local, there can be no concurrency.
547 head->lastp = &node->next;
550 void __wake_up_q(struct wake_q_head *head, bool sleeper)
552 struct wake_q_node *node = head->first;
554 while (node != WAKE_Q_TAIL) {
555 struct task_struct *task;
557 task = container_of(node, struct task_struct, wake_q);
559 /* task can safely be re-inserted now */
561 task->wake_q.next = NULL;
564 * wake_up_process() implies a wmb() to pair with the queueing
565 * in wake_q_add() so as not to miss wakeups.
568 wake_up_lock_sleeper(task);
570 wake_up_process(task);
571 put_task_struct(task);
576 * resched_curr - mark rq's current task 'to be rescheduled now'.
578 * On UP this means the setting of the need_resched flag, on SMP it
579 * might also involve a cross-CPU call to trigger the scheduler on
582 void resched_curr(struct rq *rq)
584 struct task_struct *curr = rq->curr;
587 lockdep_assert_held(&rq->lock);
589 if (test_tsk_need_resched(curr))
594 if (cpu == smp_processor_id()) {
595 set_tsk_need_resched(curr);
596 set_preempt_need_resched();
600 if (set_nr_and_not_polling(curr))
601 smp_send_reschedule(cpu);
603 trace_sched_wake_idle_without_ipi(cpu);
606 #ifdef CONFIG_PREEMPT_LAZY
607 void resched_curr_lazy(struct rq *rq)
609 struct task_struct *curr = rq->curr;
612 if (!sched_feat(PREEMPT_LAZY)) {
617 lockdep_assert_held(&rq->lock);
619 if (test_tsk_need_resched(curr))
622 if (test_tsk_need_resched_lazy(curr))
625 set_tsk_need_resched_lazy(curr);
628 if (cpu == smp_processor_id())
631 /* NEED_RESCHED_LAZY must be visible before we test polling */
633 if (!tsk_is_polling(curr))
634 smp_send_reschedule(cpu);
638 void resched_cpu(int cpu)
640 struct rq *rq = cpu_rq(cpu);
643 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
646 raw_spin_unlock_irqrestore(&rq->lock, flags);
650 #ifdef CONFIG_NO_HZ_COMMON
652 * In the semi idle case, use the nearest busy cpu for migrating timers
653 * from an idle cpu. This is good for power-savings.
655 * We don't do similar optimization for completely idle system, as
656 * selecting an idle cpu will add more delays to the timers than intended
657 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
659 int get_nohz_timer_target(void)
662 struct sched_domain *sd;
664 preempt_disable_rt();
665 cpu = smp_processor_id();
667 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
671 for_each_domain(cpu, sd) {
672 for_each_cpu(i, sched_domain_span(sd)) {
673 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
680 if (!is_housekeeping_cpu(cpu))
681 cpu = housekeeping_any_cpu();
689 * When add_timer_on() enqueues a timer into the timer wheel of an
690 * idle CPU then this timer might expire before the next timer event
691 * which is scheduled to wake up that CPU. In case of a completely
692 * idle system the next event might even be infinite time into the
693 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
694 * leaves the inner idle loop so the newly added timer is taken into
695 * account when the CPU goes back to idle and evaluates the timer
696 * wheel for the next timer event.
698 static void wake_up_idle_cpu(int cpu)
700 struct rq *rq = cpu_rq(cpu);
702 if (cpu == smp_processor_id())
705 if (set_nr_and_not_polling(rq->idle))
706 smp_send_reschedule(cpu);
708 trace_sched_wake_idle_without_ipi(cpu);
711 static bool wake_up_full_nohz_cpu(int cpu)
714 * We just need the target to call irq_exit() and re-evaluate
715 * the next tick. The nohz full kick at least implies that.
716 * If needed we can still optimize that later with an
719 if (tick_nohz_full_cpu(cpu)) {
720 if (cpu != smp_processor_id() ||
721 tick_nohz_tick_stopped())
722 tick_nohz_full_kick_cpu(cpu);
729 void wake_up_nohz_cpu(int cpu)
731 if (!wake_up_full_nohz_cpu(cpu))
732 wake_up_idle_cpu(cpu);
735 static inline bool got_nohz_idle_kick(void)
737 int cpu = smp_processor_id();
739 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
742 if (idle_cpu(cpu) && !need_resched())
746 * We can't run Idle Load Balance on this CPU for this time so we
747 * cancel it and clear NOHZ_BALANCE_KICK
749 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
753 #else /* CONFIG_NO_HZ_COMMON */
755 static inline bool got_nohz_idle_kick(void)
760 #endif /* CONFIG_NO_HZ_COMMON */
762 #ifdef CONFIG_NO_HZ_FULL
763 bool sched_can_stop_tick(void)
766 * FIFO realtime policy runs the highest priority task. Other runnable
767 * tasks are of a lower priority. The scheduler tick does nothing.
769 if (current->policy == SCHED_FIFO)
773 * Round-robin realtime tasks time slice with other tasks at the same
774 * realtime priority. Is this task the only one at this priority?
776 if (current->policy == SCHED_RR) {
777 struct sched_rt_entity *rt_se = ¤t->rt;
779 return rt_se->run_list.prev == rt_se->run_list.next;
783 * More than one running task need preemption.
784 * nr_running update is assumed to be visible
785 * after IPI is sent from wakers.
787 if (this_rq()->nr_running > 1)
792 #endif /* CONFIG_NO_HZ_FULL */
794 void sched_avg_update(struct rq *rq)
796 s64 period = sched_avg_period();
798 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
800 * Inline assembly required to prevent the compiler
801 * optimising this loop into a divmod call.
802 * See __iter_div_u64_rem() for another example of this.
804 asm("" : "+rm" (rq->age_stamp));
805 rq->age_stamp += period;
810 #endif /* CONFIG_SMP */
812 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
813 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
815 * Iterate task_group tree rooted at *from, calling @down when first entering a
816 * node and @up when leaving it for the final time.
818 * Caller must hold rcu_lock or sufficient equivalent.
820 int walk_tg_tree_from(struct task_group *from,
821 tg_visitor down, tg_visitor up, void *data)
823 struct task_group *parent, *child;
829 ret = (*down)(parent, data);
832 list_for_each_entry_rcu(child, &parent->children, siblings) {
839 ret = (*up)(parent, data);
840 if (ret || parent == from)
844 parent = parent->parent;
851 int tg_nop(struct task_group *tg, void *data)
857 static void set_load_weight(struct task_struct *p)
859 int prio = p->static_prio - MAX_RT_PRIO;
860 struct load_weight *load = &p->se.load;
863 * SCHED_IDLE tasks get minimal weight:
865 if (idle_policy(p->policy)) {
866 load->weight = scale_load(WEIGHT_IDLEPRIO);
867 load->inv_weight = WMULT_IDLEPRIO;
871 load->weight = scale_load(prio_to_weight[prio]);
872 load->inv_weight = prio_to_wmult[prio];
875 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
878 if (!(flags & ENQUEUE_RESTORE))
879 sched_info_queued(rq, p);
880 p->sched_class->enqueue_task(rq, p, flags);
883 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
886 if (!(flags & DEQUEUE_SAVE))
887 sched_info_dequeued(rq, p);
888 p->sched_class->dequeue_task(rq, p, flags);
891 void activate_task(struct rq *rq, struct task_struct *p, int flags)
893 if (task_contributes_to_load(p))
894 rq->nr_uninterruptible--;
896 enqueue_task(rq, p, flags);
899 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
901 if (task_contributes_to_load(p))
902 rq->nr_uninterruptible++;
904 dequeue_task(rq, p, flags);
907 static void update_rq_clock_task(struct rq *rq, s64 delta)
910 * In theory, the compile should just see 0 here, and optimize out the call
911 * to sched_rt_avg_update. But I don't trust it...
913 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
914 s64 steal = 0, irq_delta = 0;
916 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
917 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
920 * Since irq_time is only updated on {soft,}irq_exit, we might run into
921 * this case when a previous update_rq_clock() happened inside a
924 * When this happens, we stop ->clock_task and only update the
925 * prev_irq_time stamp to account for the part that fit, so that a next
926 * update will consume the rest. This ensures ->clock_task is
929 * It does however cause some slight miss-attribution of {soft,}irq
930 * time, a more accurate solution would be to update the irq_time using
931 * the current rq->clock timestamp, except that would require using
934 if (irq_delta > delta)
937 rq->prev_irq_time += irq_delta;
940 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
941 if (static_key_false((¶virt_steal_rq_enabled))) {
942 steal = paravirt_steal_clock(cpu_of(rq));
943 steal -= rq->prev_steal_time_rq;
945 if (unlikely(steal > delta))
948 rq->prev_steal_time_rq += steal;
953 rq->clock_task += delta;
955 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
956 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
957 sched_rt_avg_update(rq, irq_delta + steal);
961 void sched_set_stop_task(int cpu, struct task_struct *stop)
963 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
964 struct task_struct *old_stop = cpu_rq(cpu)->stop;
968 * Make it appear like a SCHED_FIFO task, its something
969 * userspace knows about and won't get confused about.
971 * Also, it will make PI more or less work without too
972 * much confusion -- but then, stop work should not
973 * rely on PI working anyway.
975 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
977 stop->sched_class = &stop_sched_class;
980 cpu_rq(cpu)->stop = stop;
984 * Reset it back to a normal scheduling class so that
985 * it can die in pieces.
987 old_stop->sched_class = &rt_sched_class;
992 * __normal_prio - return the priority that is based on the static prio
994 static inline int __normal_prio(struct task_struct *p)
996 return p->static_prio;
1000 * Calculate the expected normal priority: i.e. priority
1001 * without taking RT-inheritance into account. Might be
1002 * boosted by interactivity modifiers. Changes upon fork,
1003 * setprio syscalls, and whenever the interactivity
1004 * estimator recalculates.
1006 static inline int normal_prio(struct task_struct *p)
1010 if (task_has_dl_policy(p))
1011 prio = MAX_DL_PRIO-1;
1012 else if (task_has_rt_policy(p))
1013 prio = MAX_RT_PRIO-1 - p->rt_priority;
1015 prio = __normal_prio(p);
1020 * Calculate the current priority, i.e. the priority
1021 * taken into account by the scheduler. This value might
1022 * be boosted by RT tasks, or might be boosted by
1023 * interactivity modifiers. Will be RT if the task got
1024 * RT-boosted. If not then it returns p->normal_prio.
1026 static int effective_prio(struct task_struct *p)
1028 p->normal_prio = normal_prio(p);
1030 * If we are RT tasks or we were boosted to RT priority,
1031 * keep the priority unchanged. Otherwise, update priority
1032 * to the normal priority:
1034 if (!rt_prio(p->prio))
1035 return p->normal_prio;
1040 * task_curr - is this task currently executing on a CPU?
1041 * @p: the task in question.
1043 * Return: 1 if the task is currently executing. 0 otherwise.
1045 inline int task_curr(const struct task_struct *p)
1047 return cpu_curr(task_cpu(p)) == p;
1051 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1052 * use the balance_callback list if you want balancing.
1054 * this means any call to check_class_changed() must be followed by a call to
1055 * balance_callback().
1057 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1058 const struct sched_class *prev_class,
1061 if (prev_class != p->sched_class) {
1062 if (prev_class->switched_from)
1063 prev_class->switched_from(rq, p);
1065 p->sched_class->switched_to(rq, p);
1066 } else if (oldprio != p->prio || dl_task(p))
1067 p->sched_class->prio_changed(rq, p, oldprio);
1070 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1072 const struct sched_class *class;
1074 if (p->sched_class == rq->curr->sched_class) {
1075 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1077 for_each_class(class) {
1078 if (class == rq->curr->sched_class)
1080 if (class == p->sched_class) {
1088 * A queue event has occurred, and we're going to schedule. In
1089 * this case, we can save a useless back to back clock update.
1091 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1092 rq_clock_skip_update(rq, true);
1097 * This is how migration works:
1099 * 1) we invoke migration_cpu_stop() on the target CPU using
1101 * 2) stopper starts to run (implicitly forcing the migrated thread
1103 * 3) it checks whether the migrated task is still in the wrong runqueue.
1104 * 4) if it's in the wrong runqueue then the migration thread removes
1105 * it and puts it into the right queue.
1106 * 5) stopper completes and stop_one_cpu() returns and the migration
1111 * move_queued_task - move a queued task to new rq.
1113 * Returns (locked) new rq. Old rq's lock is released.
1115 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1117 lockdep_assert_held(&rq->lock);
1119 dequeue_task(rq, p, 0);
1120 p->on_rq = TASK_ON_RQ_MIGRATING;
1121 set_task_cpu(p, new_cpu);
1122 raw_spin_unlock(&rq->lock);
1124 rq = cpu_rq(new_cpu);
1126 raw_spin_lock(&rq->lock);
1127 BUG_ON(task_cpu(p) != new_cpu);
1128 p->on_rq = TASK_ON_RQ_QUEUED;
1129 enqueue_task(rq, p, 0);
1130 check_preempt_curr(rq, p, 0);
1135 struct migration_arg {
1136 struct task_struct *task;
1141 * Move (not current) task off this cpu, onto dest cpu. We're doing
1142 * this because either it can't run here any more (set_cpus_allowed()
1143 * away from this CPU, or CPU going down), or because we're
1144 * attempting to rebalance this task on exec (sched_exec).
1146 * So we race with normal scheduler movements, but that's OK, as long
1147 * as the task is no longer on this CPU.
1149 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1151 if (unlikely(!cpu_active(dest_cpu)))
1154 /* Affinity changed (again). */
1155 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1158 rq = move_queued_task(rq, p, dest_cpu);
1164 * migration_cpu_stop - this will be executed by a highprio stopper thread
1165 * and performs thread migration by bumping thread off CPU then
1166 * 'pushing' onto another runqueue.
1168 static int migration_cpu_stop(void *data)
1170 struct migration_arg *arg = data;
1171 struct task_struct *p = arg->task;
1172 struct rq *rq = this_rq();
1175 * The original target cpu might have gone down and we might
1176 * be on another cpu but it doesn't matter.
1178 local_irq_disable();
1180 * We need to explicitly wake pending tasks before running
1181 * __migrate_task() such that we will not miss enforcing cpus_allowed
1182 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1184 sched_ttwu_pending();
1186 raw_spin_lock(&p->pi_lock);
1187 raw_spin_lock(&rq->lock);
1189 * If task_rq(p) != rq, it cannot be migrated here, because we're
1190 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1191 * we're holding p->pi_lock.
1193 if (task_rq(p) == rq && task_on_rq_queued(p))
1194 rq = __migrate_task(rq, p, arg->dest_cpu);
1195 raw_spin_unlock(&rq->lock);
1196 raw_spin_unlock(&p->pi_lock);
1203 * sched_class::set_cpus_allowed must do the below, but is not required to
1204 * actually call this function.
1206 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1208 cpumask_copy(&p->cpus_allowed, new_mask);
1209 p->nr_cpus_allowed = cpumask_weight(new_mask);
1212 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1214 struct rq *rq = task_rq(p);
1215 bool queued, running;
1217 lockdep_assert_held(&p->pi_lock);
1219 if (__migrate_disabled(p)) {
1220 cpumask_copy(&p->cpus_allowed, new_mask);
1224 queued = task_on_rq_queued(p);
1225 running = task_current(rq, p);
1229 * Because __kthread_bind() calls this on blocked tasks without
1232 lockdep_assert_held(&rq->lock);
1233 dequeue_task(rq, p, DEQUEUE_SAVE);
1236 put_prev_task(rq, p);
1238 p->sched_class->set_cpus_allowed(p, new_mask);
1241 p->sched_class->set_curr_task(rq);
1243 enqueue_task(rq, p, ENQUEUE_RESTORE);
1246 static DEFINE_PER_CPU(struct cpumask, sched_cpumasks);
1247 static DEFINE_MUTEX(sched_down_mutex);
1248 static cpumask_t sched_down_cpumask;
1250 void tell_sched_cpu_down_begin(int cpu)
1252 mutex_lock(&sched_down_mutex);
1253 cpumask_set_cpu(cpu, &sched_down_cpumask);
1254 mutex_unlock(&sched_down_mutex);
1257 void tell_sched_cpu_down_done(int cpu)
1259 mutex_lock(&sched_down_mutex);
1260 cpumask_clear_cpu(cpu, &sched_down_cpumask);
1261 mutex_unlock(&sched_down_mutex);
1265 * migrate_me - try to move the current task off this cpu
1267 * Used by the pin_current_cpu() code to try to get tasks
1268 * to move off the current CPU as it is going down.
1269 * It will only move the task if the task isn't pinned to
1270 * the CPU (with migrate_disable, affinity or NO_SETAFFINITY)
1271 * and the task has to be in a RUNNING state. Otherwise the
1272 * movement of the task will wake it up (change its state
1273 * to running) when the task did not expect it.
1275 * Returns 1 if it succeeded in moving the current task
1278 int migrate_me(void)
1280 struct task_struct *p = current;
1281 struct migration_arg arg;
1282 struct cpumask *cpumask;
1283 struct cpumask *mask;
1284 unsigned long flags;
1285 unsigned int dest_cpu;
1289 * We can not migrate tasks bounded to a CPU or tasks not
1290 * running. The movement of the task will wake it up.
1292 if (p->flags & PF_NO_SETAFFINITY || p->state)
1295 mutex_lock(&sched_down_mutex);
1296 rq = task_rq_lock(p, &flags);
1298 cpumask = this_cpu_ptr(&sched_cpumasks);
1299 mask = &p->cpus_allowed;
1301 cpumask_andnot(cpumask, mask, &sched_down_cpumask);
1303 if (!cpumask_weight(cpumask)) {
1304 /* It's only on this CPU? */
1305 task_rq_unlock(rq, p, &flags);
1306 mutex_unlock(&sched_down_mutex);
1310 dest_cpu = cpumask_any_and(cpu_active_mask, cpumask);
1313 arg.dest_cpu = dest_cpu;
1315 task_rq_unlock(rq, p, &flags);
1317 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1318 tlb_migrate_finish(p->mm);
1319 mutex_unlock(&sched_down_mutex);
1325 * Change a given task's CPU affinity. Migrate the thread to a
1326 * proper CPU and schedule it away if the CPU it's executing on
1327 * is removed from the allowed bitmask.
1329 * NOTE: the caller must have a valid reference to the task, the
1330 * task must not exit() & deallocate itself prematurely. The
1331 * call is not atomic; no spinlocks may be held.
1333 static int __set_cpus_allowed_ptr(struct task_struct *p,
1334 const struct cpumask *new_mask, bool check)
1336 unsigned long flags;
1338 unsigned int dest_cpu;
1341 rq = task_rq_lock(p, &flags);
1344 * Must re-check here, to close a race against __kthread_bind(),
1345 * sched_setaffinity() is not guaranteed to observe the flag.
1347 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1352 if (cpumask_equal(&p->cpus_allowed, new_mask))
1355 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1360 do_set_cpus_allowed(p, new_mask);
1362 /* Can the task run on the task's current CPU? If so, we're done */
1363 if (cpumask_test_cpu(task_cpu(p), new_mask) || __migrate_disabled(p))
1366 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1367 if (task_running(rq, p) || p->state == TASK_WAKING) {
1368 struct migration_arg arg = { p, dest_cpu };
1369 /* Need help from migration thread: drop lock and wait. */
1370 task_rq_unlock(rq, p, &flags);
1371 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1372 tlb_migrate_finish(p->mm);
1374 } else if (task_on_rq_queued(p)) {
1376 * OK, since we're going to drop the lock immediately
1377 * afterwards anyway.
1379 lockdep_unpin_lock(&rq->lock);
1380 rq = move_queued_task(rq, p, dest_cpu);
1381 lockdep_pin_lock(&rq->lock);
1384 task_rq_unlock(rq, p, &flags);
1389 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1391 return __set_cpus_allowed_ptr(p, new_mask, false);
1393 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1395 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1397 #ifdef CONFIG_SCHED_DEBUG
1399 * We should never call set_task_cpu() on a blocked task,
1400 * ttwu() will sort out the placement.
1402 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1405 #ifdef CONFIG_LOCKDEP
1407 * The caller should hold either p->pi_lock or rq->lock, when changing
1408 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1410 * sched_move_task() holds both and thus holding either pins the cgroup,
1413 * Furthermore, all task_rq users should acquire both locks, see
1416 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1417 lockdep_is_held(&task_rq(p)->lock)));
1421 trace_sched_migrate_task(p, new_cpu);
1423 if (task_cpu(p) != new_cpu) {
1424 if (p->sched_class->migrate_task_rq)
1425 p->sched_class->migrate_task_rq(p);
1426 p->se.nr_migrations++;
1427 perf_event_task_migrate(p);
1430 __set_task_cpu(p, new_cpu);
1433 static void __migrate_swap_task(struct task_struct *p, int cpu)
1435 if (task_on_rq_queued(p)) {
1436 struct rq *src_rq, *dst_rq;
1438 src_rq = task_rq(p);
1439 dst_rq = cpu_rq(cpu);
1441 deactivate_task(src_rq, p, 0);
1442 set_task_cpu(p, cpu);
1443 activate_task(dst_rq, p, 0);
1444 check_preempt_curr(dst_rq, p, 0);
1447 * Task isn't running anymore; make it appear like we migrated
1448 * it before it went to sleep. This means on wakeup we make the
1449 * previous cpu our targer instead of where it really is.
1455 struct migration_swap_arg {
1456 struct task_struct *src_task, *dst_task;
1457 int src_cpu, dst_cpu;
1460 static int migrate_swap_stop(void *data)
1462 struct migration_swap_arg *arg = data;
1463 struct rq *src_rq, *dst_rq;
1466 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1469 src_rq = cpu_rq(arg->src_cpu);
1470 dst_rq = cpu_rq(arg->dst_cpu);
1472 double_raw_lock(&arg->src_task->pi_lock,
1473 &arg->dst_task->pi_lock);
1474 double_rq_lock(src_rq, dst_rq);
1476 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1479 if (task_cpu(arg->src_task) != arg->src_cpu)
1482 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1485 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1488 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1489 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1494 double_rq_unlock(src_rq, dst_rq);
1495 raw_spin_unlock(&arg->dst_task->pi_lock);
1496 raw_spin_unlock(&arg->src_task->pi_lock);
1502 * Cross migrate two tasks
1504 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1506 struct migration_swap_arg arg;
1509 arg = (struct migration_swap_arg){
1511 .src_cpu = task_cpu(cur),
1513 .dst_cpu = task_cpu(p),
1516 if (arg.src_cpu == arg.dst_cpu)
1520 * These three tests are all lockless; this is OK since all of them
1521 * will be re-checked with proper locks held further down the line.
1523 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1526 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1529 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1532 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1533 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1539 static bool check_task_state(struct task_struct *p, long match_state)
1543 raw_spin_lock_irq(&p->pi_lock);
1544 if (p->state == match_state || p->saved_state == match_state)
1546 raw_spin_unlock_irq(&p->pi_lock);
1552 * wait_task_inactive - wait for a thread to unschedule.
1554 * If @match_state is nonzero, it's the @p->state value just checked and
1555 * not expected to change. If it changes, i.e. @p might have woken up,
1556 * then return zero. When we succeed in waiting for @p to be off its CPU,
1557 * we return a positive number (its total switch count). If a second call
1558 * a short while later returns the same number, the caller can be sure that
1559 * @p has remained unscheduled the whole time.
1561 * The caller must ensure that the task *will* unschedule sometime soon,
1562 * else this function might spin for a *long* time. This function can't
1563 * be called with interrupts off, or it may introduce deadlock with
1564 * smp_call_function() if an IPI is sent by the same process we are
1565 * waiting to become inactive.
1567 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1569 unsigned long flags;
1570 int running, queued;
1576 * We do the initial early heuristics without holding
1577 * any task-queue locks at all. We'll only try to get
1578 * the runqueue lock when things look like they will
1584 * If the task is actively running on another CPU
1585 * still, just relax and busy-wait without holding
1588 * NOTE! Since we don't hold any locks, it's not
1589 * even sure that "rq" stays as the right runqueue!
1590 * But we don't care, since "task_running()" will
1591 * return false if the runqueue has changed and p
1592 * is actually now running somewhere else!
1594 while (task_running(rq, p)) {
1595 if (match_state && !check_task_state(p, match_state))
1601 * Ok, time to look more closely! We need the rq
1602 * lock now, to be *sure*. If we're wrong, we'll
1603 * just go back and repeat.
1605 rq = task_rq_lock(p, &flags);
1606 trace_sched_wait_task(p);
1607 running = task_running(rq, p);
1608 queued = task_on_rq_queued(p);
1610 if (!match_state || p->state == match_state ||
1611 p->saved_state == match_state)
1612 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1613 task_rq_unlock(rq, p, &flags);
1616 * If it changed from the expected state, bail out now.
1618 if (unlikely(!ncsw))
1622 * Was it really running after all now that we
1623 * checked with the proper locks actually held?
1625 * Oops. Go back and try again..
1627 if (unlikely(running)) {
1633 * It's not enough that it's not actively running,
1634 * it must be off the runqueue _entirely_, and not
1637 * So if it was still runnable (but just not actively
1638 * running right now), it's preempted, and we should
1639 * yield - it could be a while.
1641 if (unlikely(queued)) {
1642 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1644 set_current_state(TASK_UNINTERRUPTIBLE);
1645 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1650 * Ahh, all good. It wasn't running, and it wasn't
1651 * runnable, which means that it will never become
1652 * running in the future either. We're all done!
1661 * kick_process - kick a running thread to enter/exit the kernel
1662 * @p: the to-be-kicked thread
1664 * Cause a process which is running on another CPU to enter
1665 * kernel-mode, without any delay. (to get signals handled.)
1667 * NOTE: this function doesn't have to take the runqueue lock,
1668 * because all it wants to ensure is that the remote task enters
1669 * the kernel. If the IPI races and the task has been migrated
1670 * to another CPU then no harm is done and the purpose has been
1673 void kick_process(struct task_struct *p)
1679 if ((cpu != smp_processor_id()) && task_curr(p))
1680 smp_send_reschedule(cpu);
1683 EXPORT_SYMBOL_GPL(kick_process);
1686 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1688 static int select_fallback_rq(int cpu, struct task_struct *p)
1690 int nid = cpu_to_node(cpu);
1691 const struct cpumask *nodemask = NULL;
1692 enum { cpuset, possible, fail } state = cpuset;
1696 * If the node that the cpu is on has been offlined, cpu_to_node()
1697 * will return -1. There is no cpu on the node, and we should
1698 * select the cpu on the other node.
1701 nodemask = cpumask_of_node(nid);
1703 /* Look for allowed, online CPU in same node. */
1704 for_each_cpu(dest_cpu, nodemask) {
1705 if (!cpu_online(dest_cpu))
1707 if (!cpu_active(dest_cpu))
1709 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1715 /* Any allowed, online CPU? */
1716 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1717 if (!cpu_online(dest_cpu))
1719 if (!cpu_active(dest_cpu))
1724 /* No more Mr. Nice Guy. */
1727 if (IS_ENABLED(CONFIG_CPUSETS)) {
1728 cpuset_cpus_allowed_fallback(p);
1734 do_set_cpus_allowed(p, cpu_possible_mask);
1745 if (state != cpuset) {
1747 * Don't tell them about moving exiting tasks or
1748 * kernel threads (both mm NULL), since they never
1751 if (p->mm && printk_ratelimit()) {
1752 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1753 task_pid_nr(p), p->comm, cpu);
1761 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1764 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1766 lockdep_assert_held(&p->pi_lock);
1768 if (tsk_nr_cpus_allowed(p) > 1)
1769 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1772 * In order not to call set_task_cpu() on a blocking task we need
1773 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1776 * Since this is common to all placement strategies, this lives here.
1778 * [ this allows ->select_task() to simply return task_cpu(p) and
1779 * not worry about this generic constraint ]
1781 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1783 cpu = select_fallback_rq(task_cpu(p), p);
1788 static void update_avg(u64 *avg, u64 sample)
1790 s64 diff = sample - *avg;
1796 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1797 const struct cpumask *new_mask, bool check)
1799 return set_cpus_allowed_ptr(p, new_mask);
1802 #endif /* CONFIG_SMP */
1805 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1807 #ifdef CONFIG_SCHEDSTATS
1808 struct rq *rq = this_rq();
1811 int this_cpu = smp_processor_id();
1813 if (cpu == this_cpu) {
1814 schedstat_inc(rq, ttwu_local);
1815 schedstat_inc(p, se.statistics.nr_wakeups_local);
1817 struct sched_domain *sd;
1819 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1821 for_each_domain(this_cpu, sd) {
1822 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1823 schedstat_inc(sd, ttwu_wake_remote);
1830 if (wake_flags & WF_MIGRATED)
1831 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1833 #endif /* CONFIG_SMP */
1835 schedstat_inc(rq, ttwu_count);
1836 schedstat_inc(p, se.statistics.nr_wakeups);
1838 if (wake_flags & WF_SYNC)
1839 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1841 #endif /* CONFIG_SCHEDSTATS */
1844 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1846 activate_task(rq, p, en_flags);
1847 p->on_rq = TASK_ON_RQ_QUEUED;
1851 * Mark the task runnable and perform wakeup-preemption.
1854 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1856 check_preempt_curr(rq, p, wake_flags);
1857 p->state = TASK_RUNNING;
1858 trace_sched_wakeup(p);
1861 if (p->sched_class->task_woken) {
1863 * Our task @p is fully woken up and running; so its safe to
1864 * drop the rq->lock, hereafter rq is only used for statistics.
1866 lockdep_unpin_lock(&rq->lock);
1867 p->sched_class->task_woken(rq, p);
1868 lockdep_pin_lock(&rq->lock);
1871 if (rq->idle_stamp) {
1872 u64 delta = rq_clock(rq) - rq->idle_stamp;
1873 u64 max = 2*rq->max_idle_balance_cost;
1875 update_avg(&rq->avg_idle, delta);
1877 if (rq->avg_idle > max)
1886 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1888 lockdep_assert_held(&rq->lock);
1891 if (p->sched_contributes_to_load)
1892 rq->nr_uninterruptible--;
1895 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1896 ttwu_do_wakeup(rq, p, wake_flags);
1900 * Called in case the task @p isn't fully descheduled from its runqueue,
1901 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1902 * since all we need to do is flip p->state to TASK_RUNNING, since
1903 * the task is still ->on_rq.
1905 static int ttwu_remote(struct task_struct *p, int wake_flags)
1910 rq = __task_rq_lock(p);
1911 if (task_on_rq_queued(p)) {
1912 /* check_preempt_curr() may use rq clock */
1913 update_rq_clock(rq);
1914 ttwu_do_wakeup(rq, p, wake_flags);
1917 __task_rq_unlock(rq);
1923 void sched_ttwu_pending(void)
1925 struct rq *rq = this_rq();
1926 struct llist_node *llist = llist_del_all(&rq->wake_list);
1927 struct task_struct *p;
1928 unsigned long flags;
1933 raw_spin_lock_irqsave(&rq->lock, flags);
1934 lockdep_pin_lock(&rq->lock);
1937 p = llist_entry(llist, struct task_struct, wake_entry);
1938 llist = llist_next(llist);
1939 ttwu_do_activate(rq, p, 0);
1942 lockdep_unpin_lock(&rq->lock);
1943 raw_spin_unlock_irqrestore(&rq->lock, flags);
1946 void scheduler_ipi(void)
1949 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1950 * TIF_NEED_RESCHED remotely (for the first time) will also send
1953 preempt_fold_need_resched();
1955 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1959 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1960 * traditionally all their work was done from the interrupt return
1961 * path. Now that we actually do some work, we need to make sure
1964 * Some archs already do call them, luckily irq_enter/exit nest
1967 * Arguably we should visit all archs and update all handlers,
1968 * however a fair share of IPIs are still resched only so this would
1969 * somewhat pessimize the simple resched case.
1972 sched_ttwu_pending();
1975 * Check if someone kicked us for doing the nohz idle load balance.
1977 if (unlikely(got_nohz_idle_kick())) {
1978 this_rq()->idle_balance = 1;
1979 raise_softirq_irqoff(SCHED_SOFTIRQ);
1984 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1986 struct rq *rq = cpu_rq(cpu);
1988 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1989 if (!set_nr_if_polling(rq->idle))
1990 smp_send_reschedule(cpu);
1992 trace_sched_wake_idle_without_ipi(cpu);
1996 void wake_up_if_idle(int cpu)
1998 struct rq *rq = cpu_rq(cpu);
1999 unsigned long flags;
2003 if (!is_idle_task(rcu_dereference(rq->curr)))
2006 if (set_nr_if_polling(rq->idle)) {
2007 trace_sched_wake_idle_without_ipi(cpu);
2009 raw_spin_lock_irqsave(&rq->lock, flags);
2010 if (is_idle_task(rq->curr))
2011 smp_send_reschedule(cpu);
2012 /* Else cpu is not in idle, do nothing here */
2013 raw_spin_unlock_irqrestore(&rq->lock, flags);
2020 bool cpus_share_cache(int this_cpu, int that_cpu)
2022 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2024 #endif /* CONFIG_SMP */
2026 static void ttwu_queue(struct task_struct *p, int cpu)
2028 struct rq *rq = cpu_rq(cpu);
2030 #if defined(CONFIG_SMP)
2031 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
2032 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2033 ttwu_queue_remote(p, cpu);
2038 raw_spin_lock(&rq->lock);
2039 lockdep_pin_lock(&rq->lock);
2040 ttwu_do_activate(rq, p, 0);
2041 lockdep_unpin_lock(&rq->lock);
2042 raw_spin_unlock(&rq->lock);
2046 * try_to_wake_up - wake up a thread
2047 * @p: the thread to be awakened
2048 * @state: the mask of task states that can be woken
2049 * @wake_flags: wake modifier flags (WF_*)
2051 * Put it on the run-queue if it's not already there. The "current"
2052 * thread is always on the run-queue (except when the actual
2053 * re-schedule is in progress), and as such you're allowed to do
2054 * the simpler "current->state = TASK_RUNNING" to mark yourself
2055 * runnable without the overhead of this.
2057 * Return: %true if @p was woken up, %false if it was already running.
2058 * or @state didn't match @p's state.
2061 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2063 unsigned long flags;
2064 int cpu, success = 0;
2067 * If we are going to wake up a thread waiting for CONDITION we
2068 * need to ensure that CONDITION=1 done by the caller can not be
2069 * reordered with p->state check below. This pairs with mb() in
2070 * set_current_state() the waiting thread does.
2072 smp_mb__before_spinlock();
2073 raw_spin_lock_irqsave(&p->pi_lock, flags);
2074 if (!(p->state & state)) {
2076 * The task might be running due to a spinlock sleeper
2077 * wakeup. Check the saved state and set it to running
2078 * if the wakeup condition is true.
2080 if (!(wake_flags & WF_LOCK_SLEEPER)) {
2081 if (p->saved_state & state) {
2082 p->saved_state = TASK_RUNNING;
2090 * If this is a regular wakeup, then we can unconditionally
2091 * clear the saved state of a "lock sleeper".
2093 if (!(wake_flags & WF_LOCK_SLEEPER))
2094 p->saved_state = TASK_RUNNING;
2096 trace_sched_waking(p);
2098 success = 1; /* we're going to change ->state */
2101 if (p->on_rq && ttwu_remote(p, wake_flags))
2106 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2107 * possible to, falsely, observe p->on_cpu == 0.
2109 * One must be running (->on_cpu == 1) in order to remove oneself
2110 * from the runqueue.
2112 * [S] ->on_cpu = 1; [L] ->on_rq
2116 * [S] ->on_rq = 0; [L] ->on_cpu
2118 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2119 * from the consecutive calls to schedule(); the first switching to our
2120 * task, the second putting it to sleep.
2125 * If the owning (remote) cpu is still in the middle of schedule() with
2126 * this task as prev, wait until its done referencing the task.
2131 * Combined with the control dependency above, we have an effective
2132 * smp_load_acquire() without the need for full barriers.
2134 * Pairs with the smp_store_release() in finish_lock_switch().
2136 * This ensures that tasks getting woken will be fully ordered against
2137 * their previous state and preserve Program Order.
2141 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2142 p->state = TASK_WAKING;
2144 if (p->sched_class->task_waking)
2145 p->sched_class->task_waking(p);
2147 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2148 if (task_cpu(p) != cpu) {
2149 wake_flags |= WF_MIGRATED;
2150 set_task_cpu(p, cpu);
2152 #endif /* CONFIG_SMP */
2156 ttwu_stat(p, cpu, wake_flags);
2158 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2164 * wake_up_process - Wake up a specific process
2165 * @p: The process to be woken up.
2167 * Attempt to wake up the nominated process and move it to the set of runnable
2170 * Return: 1 if the process was woken up, 0 if it was already running.
2172 * It may be assumed that this function implies a write memory barrier before
2173 * changing the task state if and only if any tasks are woken up.
2175 int wake_up_process(struct task_struct *p)
2177 return try_to_wake_up(p, TASK_NORMAL, 0);
2179 EXPORT_SYMBOL(wake_up_process);
2182 * wake_up_lock_sleeper - Wake up a specific process blocked on a "sleeping lock"
2183 * @p: The process to be woken up.
2185 * Same as wake_up_process() above, but wake_flags=WF_LOCK_SLEEPER to indicate
2186 * the nature of the wakeup.
2188 int wake_up_lock_sleeper(struct task_struct *p)
2190 return try_to_wake_up(p, TASK_ALL, WF_LOCK_SLEEPER);
2193 int wake_up_state(struct task_struct *p, unsigned int state)
2195 return try_to_wake_up(p, state, 0);
2199 * This function clears the sched_dl_entity static params.
2201 void __dl_clear_params(struct task_struct *p)
2203 struct sched_dl_entity *dl_se = &p->dl;
2205 dl_se->dl_runtime = 0;
2206 dl_se->dl_deadline = 0;
2207 dl_se->dl_period = 0;
2211 dl_se->dl_throttled = 0;
2213 dl_se->dl_yielded = 0;
2217 * Perform scheduler related setup for a newly forked process p.
2218 * p is forked by current.
2220 * __sched_fork() is basic setup used by init_idle() too:
2222 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2227 p->se.exec_start = 0;
2228 p->se.sum_exec_runtime = 0;
2229 p->se.prev_sum_exec_runtime = 0;
2230 p->se.nr_migrations = 0;
2232 INIT_LIST_HEAD(&p->se.group_node);
2234 #ifdef CONFIG_SCHEDSTATS
2235 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2238 RB_CLEAR_NODE(&p->dl.rb_node);
2239 init_dl_task_timer(&p->dl);
2240 __dl_clear_params(p);
2242 INIT_LIST_HEAD(&p->rt.run_list);
2244 #ifdef CONFIG_PREEMPT_NOTIFIERS
2245 INIT_HLIST_HEAD(&p->preempt_notifiers);
2248 #ifdef CONFIG_NUMA_BALANCING
2249 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2250 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2251 p->mm->numa_scan_seq = 0;
2254 if (clone_flags & CLONE_VM)
2255 p->numa_preferred_nid = current->numa_preferred_nid;
2257 p->numa_preferred_nid = -1;
2259 p->node_stamp = 0ULL;
2260 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2261 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2262 p->numa_work.next = &p->numa_work;
2263 p->numa_faults = NULL;
2264 p->last_task_numa_placement = 0;
2265 p->last_sum_exec_runtime = 0;
2267 p->numa_group = NULL;
2268 #endif /* CONFIG_NUMA_BALANCING */
2271 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2273 #ifdef CONFIG_NUMA_BALANCING
2275 void set_numabalancing_state(bool enabled)
2278 static_branch_enable(&sched_numa_balancing);
2280 static_branch_disable(&sched_numa_balancing);
2283 #ifdef CONFIG_PROC_SYSCTL
2284 int sysctl_numa_balancing(struct ctl_table *table, int write,
2285 void __user *buffer, size_t *lenp, loff_t *ppos)
2289 int state = static_branch_likely(&sched_numa_balancing);
2291 if (write && !capable(CAP_SYS_ADMIN))
2296 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2300 set_numabalancing_state(state);
2307 * fork()/clone()-time setup:
2309 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2311 unsigned long flags;
2312 int cpu = get_cpu();
2314 __sched_fork(clone_flags, p);
2316 * We mark the process as running here. This guarantees that
2317 * nobody will actually run it, and a signal or other external
2318 * event cannot wake it up and insert it on the runqueue either.
2320 p->state = TASK_RUNNING;
2323 * Make sure we do not leak PI boosting priority to the child.
2325 p->prio = current->normal_prio;
2328 * Revert to default priority/policy on fork if requested.
2330 if (unlikely(p->sched_reset_on_fork)) {
2331 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2332 p->policy = SCHED_NORMAL;
2333 p->static_prio = NICE_TO_PRIO(0);
2335 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2336 p->static_prio = NICE_TO_PRIO(0);
2338 p->prio = p->normal_prio = __normal_prio(p);
2342 * We don't need the reset flag anymore after the fork. It has
2343 * fulfilled its duty:
2345 p->sched_reset_on_fork = 0;
2348 if (dl_prio(p->prio)) {
2351 } else if (rt_prio(p->prio)) {
2352 p->sched_class = &rt_sched_class;
2354 p->sched_class = &fair_sched_class;
2357 if (p->sched_class->task_fork)
2358 p->sched_class->task_fork(p);
2361 * The child is not yet in the pid-hash so no cgroup attach races,
2362 * and the cgroup is pinned to this child due to cgroup_fork()
2363 * is ran before sched_fork().
2365 * Silence PROVE_RCU.
2367 raw_spin_lock_irqsave(&p->pi_lock, flags);
2368 set_task_cpu(p, cpu);
2369 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2371 #ifdef CONFIG_SCHED_INFO
2372 if (likely(sched_info_on()))
2373 memset(&p->sched_info, 0, sizeof(p->sched_info));
2375 #if defined(CONFIG_SMP)
2378 init_task_preempt_count(p);
2379 #ifdef CONFIG_HAVE_PREEMPT_LAZY
2380 task_thread_info(p)->preempt_lazy_count = 0;
2383 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2384 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2391 unsigned long to_ratio(u64 period, u64 runtime)
2393 if (runtime == RUNTIME_INF)
2397 * Doing this here saves a lot of checks in all
2398 * the calling paths, and returning zero seems
2399 * safe for them anyway.
2404 return div64_u64(runtime << 20, period);
2408 inline struct dl_bw *dl_bw_of(int i)
2410 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2411 "sched RCU must be held");
2412 return &cpu_rq(i)->rd->dl_bw;
2415 static inline int dl_bw_cpus(int i)
2417 struct root_domain *rd = cpu_rq(i)->rd;
2420 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2421 "sched RCU must be held");
2422 for_each_cpu_and(i, rd->span, cpu_active_mask)
2428 inline struct dl_bw *dl_bw_of(int i)
2430 return &cpu_rq(i)->dl.dl_bw;
2433 static inline int dl_bw_cpus(int i)
2440 * We must be sure that accepting a new task (or allowing changing the
2441 * parameters of an existing one) is consistent with the bandwidth
2442 * constraints. If yes, this function also accordingly updates the currently
2443 * allocated bandwidth to reflect the new situation.
2445 * This function is called while holding p's rq->lock.
2447 * XXX we should delay bw change until the task's 0-lag point, see
2450 static int dl_overflow(struct task_struct *p, int policy,
2451 const struct sched_attr *attr)
2454 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2455 u64 period = attr->sched_period ?: attr->sched_deadline;
2456 u64 runtime = attr->sched_runtime;
2457 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2460 if (new_bw == p->dl.dl_bw)
2464 * Either if a task, enters, leave, or stays -deadline but changes
2465 * its parameters, we may need to update accordingly the total
2466 * allocated bandwidth of the container.
2468 raw_spin_lock(&dl_b->lock);
2469 cpus = dl_bw_cpus(task_cpu(p));
2470 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2471 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2472 __dl_add(dl_b, new_bw);
2474 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2475 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2476 __dl_clear(dl_b, p->dl.dl_bw);
2477 __dl_add(dl_b, new_bw);
2479 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2480 __dl_clear(dl_b, p->dl.dl_bw);
2483 raw_spin_unlock(&dl_b->lock);
2488 extern void init_dl_bw(struct dl_bw *dl_b);
2491 * wake_up_new_task - wake up a newly created task for the first time.
2493 * This function will do some initial scheduler statistics housekeeping
2494 * that must be done for every newly created context, then puts the task
2495 * on the runqueue and wakes it.
2497 void wake_up_new_task(struct task_struct *p)
2499 unsigned long flags;
2502 raw_spin_lock_irqsave(&p->pi_lock, flags);
2503 /* Initialize new task's runnable average */
2504 init_entity_runnable_average(&p->se);
2507 * Fork balancing, do it here and not earlier because:
2508 * - cpus_allowed can change in the fork path
2509 * - any previously selected cpu might disappear through hotplug
2511 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2514 rq = __task_rq_lock(p);
2515 activate_task(rq, p, 0);
2516 p->on_rq = TASK_ON_RQ_QUEUED;
2517 trace_sched_wakeup_new(p);
2518 check_preempt_curr(rq, p, WF_FORK);
2520 if (p->sched_class->task_woken) {
2522 * Nothing relies on rq->lock after this, so its fine to
2525 lockdep_unpin_lock(&rq->lock);
2526 p->sched_class->task_woken(rq, p);
2527 lockdep_pin_lock(&rq->lock);
2530 task_rq_unlock(rq, p, &flags);
2533 #ifdef CONFIG_PREEMPT_NOTIFIERS
2535 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2537 void preempt_notifier_inc(void)
2539 static_key_slow_inc(&preempt_notifier_key);
2541 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2543 void preempt_notifier_dec(void)
2545 static_key_slow_dec(&preempt_notifier_key);
2547 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2550 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2551 * @notifier: notifier struct to register
2553 void preempt_notifier_register(struct preempt_notifier *notifier)
2555 if (!static_key_false(&preempt_notifier_key))
2556 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2558 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2560 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2563 * preempt_notifier_unregister - no longer interested in preemption notifications
2564 * @notifier: notifier struct to unregister
2566 * This is *not* safe to call from within a preemption notifier.
2568 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2570 hlist_del(¬ifier->link);
2572 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2574 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2576 struct preempt_notifier *notifier;
2578 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2579 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2582 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2584 if (static_key_false(&preempt_notifier_key))
2585 __fire_sched_in_preempt_notifiers(curr);
2589 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2590 struct task_struct *next)
2592 struct preempt_notifier *notifier;
2594 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2595 notifier->ops->sched_out(notifier, next);
2598 static __always_inline void
2599 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2600 struct task_struct *next)
2602 if (static_key_false(&preempt_notifier_key))
2603 __fire_sched_out_preempt_notifiers(curr, next);
2606 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2608 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2613 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2614 struct task_struct *next)
2618 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2621 * prepare_task_switch - prepare to switch tasks
2622 * @rq: the runqueue preparing to switch
2623 * @prev: the current task that is being switched out
2624 * @next: the task we are going to switch to.
2626 * This is called with the rq lock held and interrupts off. It must
2627 * be paired with a subsequent finish_task_switch after the context
2630 * prepare_task_switch sets up locking and calls architecture specific
2634 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2635 struct task_struct *next)
2637 sched_info_switch(rq, prev, next);
2638 perf_event_task_sched_out(prev, next);
2639 fire_sched_out_preempt_notifiers(prev, next);
2640 prepare_lock_switch(rq, next);
2641 prepare_arch_switch(next);
2645 * finish_task_switch - clean up after a task-switch
2646 * @prev: the thread we just switched away from.
2648 * finish_task_switch must be called after the context switch, paired
2649 * with a prepare_task_switch call before the context switch.
2650 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2651 * and do any other architecture-specific cleanup actions.
2653 * Note that we may have delayed dropping an mm in context_switch(). If
2654 * so, we finish that here outside of the runqueue lock. (Doing it
2655 * with the lock held can cause deadlocks; see schedule() for
2658 * The context switch have flipped the stack from under us and restored the
2659 * local variables which were saved when this task called schedule() in the
2660 * past. prev == current is still correct but we need to recalculate this_rq
2661 * because prev may have moved to another CPU.
2663 static struct rq *finish_task_switch(struct task_struct *prev)
2664 __releases(rq->lock)
2666 struct rq *rq = this_rq();
2667 struct mm_struct *mm = rq->prev_mm;
2671 * The previous task will have left us with a preempt_count of 2
2672 * because it left us after:
2675 * preempt_disable(); // 1
2677 * raw_spin_lock_irq(&rq->lock) // 2
2679 * Also, see FORK_PREEMPT_COUNT.
2681 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2682 "corrupted preempt_count: %s/%d/0x%x\n",
2683 current->comm, current->pid, preempt_count()))
2684 preempt_count_set(FORK_PREEMPT_COUNT);
2689 * A task struct has one reference for the use as "current".
2690 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2691 * schedule one last time. The schedule call will never return, and
2692 * the scheduled task must drop that reference.
2694 * We must observe prev->state before clearing prev->on_cpu (in
2695 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2696 * running on another CPU and we could rave with its RUNNING -> DEAD
2697 * transition, resulting in a double drop.
2699 prev_state = prev->state;
2700 vtime_task_switch(prev);
2701 perf_event_task_sched_in(prev, current);
2702 finish_lock_switch(rq, prev);
2703 finish_arch_post_lock_switch();
2705 fire_sched_in_preempt_notifiers(current);
2707 * We use mmdrop_delayed() here so we don't have to do the
2708 * full __mmdrop() when we are the last user.
2712 if (unlikely(prev_state == TASK_DEAD)) {
2713 if (prev->sched_class->task_dead)
2714 prev->sched_class->task_dead(prev);
2717 * Remove function-return probe instances associated with this
2718 * task and put them back on the free list.
2720 kprobe_flush_task(prev);
2721 put_task_struct(prev);
2724 tick_nohz_task_switch();
2730 /* rq->lock is NOT held, but preemption is disabled */
2731 static void __balance_callback(struct rq *rq)
2733 struct callback_head *head, *next;
2734 void (*func)(struct rq *rq);
2735 unsigned long flags;
2737 raw_spin_lock_irqsave(&rq->lock, flags);
2738 head = rq->balance_callback;
2739 rq->balance_callback = NULL;
2741 func = (void (*)(struct rq *))head->func;
2748 raw_spin_unlock_irqrestore(&rq->lock, flags);
2751 static inline void balance_callback(struct rq *rq)
2753 if (unlikely(rq->balance_callback))
2754 __balance_callback(rq);
2759 static inline void balance_callback(struct rq *rq)
2766 * schedule_tail - first thing a freshly forked thread must call.
2767 * @prev: the thread we just switched away from.
2769 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2770 __releases(rq->lock)
2775 * New tasks start with FORK_PREEMPT_COUNT, see there and
2776 * finish_task_switch() for details.
2778 * finish_task_switch() will drop rq->lock() and lower preempt_count
2779 * and the preempt_enable() will end up enabling preemption (on
2780 * PREEMPT_COUNT kernels).
2783 rq = finish_task_switch(prev);
2784 balance_callback(rq);
2787 if (current->set_child_tid)
2788 put_user(task_pid_vnr(current), current->set_child_tid);
2792 * context_switch - switch to the new MM and the new thread's register state.
2794 static inline struct rq *
2795 context_switch(struct rq *rq, struct task_struct *prev,
2796 struct task_struct *next)
2798 struct mm_struct *mm, *oldmm;
2800 prepare_task_switch(rq, prev, next);
2803 oldmm = prev->active_mm;
2805 * For paravirt, this is coupled with an exit in switch_to to
2806 * combine the page table reload and the switch backend into
2809 arch_start_context_switch(prev);
2812 next->active_mm = oldmm;
2813 atomic_inc(&oldmm->mm_count);
2814 enter_lazy_tlb(oldmm, next);
2816 switch_mm(oldmm, mm, next);
2819 prev->active_mm = NULL;
2820 rq->prev_mm = oldmm;
2823 * Since the runqueue lock will be released by the next
2824 * task (which is an invalid locking op but in the case
2825 * of the scheduler it's an obvious special-case), so we
2826 * do an early lockdep release here:
2828 lockdep_unpin_lock(&rq->lock);
2829 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2831 /* Here we just switch the register state and the stack. */
2832 switch_to(prev, next, prev);
2835 return finish_task_switch(prev);
2839 * nr_running and nr_context_switches:
2841 * externally visible scheduler statistics: current number of runnable
2842 * threads, total number of context switches performed since bootup.
2844 unsigned long nr_running(void)
2846 unsigned long i, sum = 0;
2848 for_each_online_cpu(i)
2849 sum += cpu_rq(i)->nr_running;
2855 * Check if only the current task is running on the cpu.
2857 * Caution: this function does not check that the caller has disabled
2858 * preemption, thus the result might have a time-of-check-to-time-of-use
2859 * race. The caller is responsible to use it correctly, for example:
2861 * - from a non-preemptable section (of course)
2863 * - from a thread that is bound to a single CPU
2865 * - in a loop with very short iterations (e.g. a polling loop)
2867 bool single_task_running(void)
2869 return raw_rq()->nr_running == 1;
2871 EXPORT_SYMBOL(single_task_running);
2873 unsigned long long nr_context_switches(void)
2876 unsigned long long sum = 0;
2878 for_each_possible_cpu(i)
2879 sum += cpu_rq(i)->nr_switches;
2884 unsigned long nr_iowait(void)
2886 unsigned long i, sum = 0;
2888 for_each_possible_cpu(i)
2889 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2894 unsigned long nr_iowait_cpu(int cpu)
2896 struct rq *this = cpu_rq(cpu);
2897 return atomic_read(&this->nr_iowait);
2900 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2902 struct rq *rq = this_rq();
2903 *nr_waiters = atomic_read(&rq->nr_iowait);
2904 *load = rq->load.weight;
2910 * sched_exec - execve() is a valuable balancing opportunity, because at
2911 * this point the task has the smallest effective memory and cache footprint.
2913 void sched_exec(void)
2915 struct task_struct *p = current;
2916 unsigned long flags;
2919 raw_spin_lock_irqsave(&p->pi_lock, flags);
2920 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2921 if (dest_cpu == smp_processor_id())
2924 if (likely(cpu_active(dest_cpu))) {
2925 struct migration_arg arg = { p, dest_cpu };
2927 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2928 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2932 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2937 DEFINE_PER_CPU(struct kernel_stat, kstat);
2938 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2940 EXPORT_PER_CPU_SYMBOL(kstat);
2941 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2944 * Return accounted runtime for the task.
2945 * In case the task is currently running, return the runtime plus current's
2946 * pending runtime that have not been accounted yet.
2948 unsigned long long task_sched_runtime(struct task_struct *p)
2950 unsigned long flags;
2954 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2956 * 64-bit doesn't need locks to atomically read a 64bit value.
2957 * So we have a optimization chance when the task's delta_exec is 0.
2958 * Reading ->on_cpu is racy, but this is ok.
2960 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2961 * If we race with it entering cpu, unaccounted time is 0. This is
2962 * indistinguishable from the read occurring a few cycles earlier.
2963 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2964 * been accounted, so we're correct here as well.
2966 if (!p->on_cpu || !task_on_rq_queued(p))
2967 return p->se.sum_exec_runtime;
2970 rq = task_rq_lock(p, &flags);
2972 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2973 * project cycles that may never be accounted to this
2974 * thread, breaking clock_gettime().
2976 if (task_current(rq, p) && task_on_rq_queued(p)) {
2977 update_rq_clock(rq);
2978 p->sched_class->update_curr(rq);
2980 ns = p->se.sum_exec_runtime;
2981 task_rq_unlock(rq, p, &flags);
2987 * This function gets called by the timer code, with HZ frequency.
2988 * We call it with interrupts disabled.
2990 void scheduler_tick(void)
2992 int cpu = smp_processor_id();
2993 struct rq *rq = cpu_rq(cpu);
2994 struct task_struct *curr = rq->curr;
2998 raw_spin_lock(&rq->lock);
2999 update_rq_clock(rq);
3000 curr->sched_class->task_tick(rq, curr, 0);
3001 update_cpu_load_active(rq);
3002 calc_global_load_tick(rq);
3003 raw_spin_unlock(&rq->lock);
3005 perf_event_task_tick();
3008 rq->idle_balance = idle_cpu(cpu);
3009 trigger_load_balance(rq);
3011 rq_last_tick_reset(rq);
3014 #ifdef CONFIG_NO_HZ_FULL
3016 * scheduler_tick_max_deferment
3018 * Keep at least one tick per second when a single
3019 * active task is running because the scheduler doesn't
3020 * yet completely support full dynticks environment.
3022 * This makes sure that uptime, CFS vruntime, load
3023 * balancing, etc... continue to move forward, even
3024 * with a very low granularity.
3026 * Return: Maximum deferment in nanoseconds.
3028 u64 scheduler_tick_max_deferment(void)
3030 struct rq *rq = this_rq();
3031 unsigned long next, now = READ_ONCE(jiffies);
3033 next = rq->last_sched_tick + HZ;
3035 if (time_before_eq(next, now))
3038 return jiffies_to_nsecs(next - now);
3042 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3043 defined(CONFIG_PREEMPT_TRACER))
3045 void preempt_count_add(int val)
3047 #ifdef CONFIG_DEBUG_PREEMPT
3051 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3054 __preempt_count_add(val);
3055 #ifdef CONFIG_DEBUG_PREEMPT
3057 * Spinlock count overflowing soon?
3059 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3062 if (preempt_count() == val) {
3063 unsigned long ip = get_lock_parent_ip();
3064 #ifdef CONFIG_DEBUG_PREEMPT
3065 current->preempt_disable_ip = ip;
3067 trace_preempt_off(CALLER_ADDR0, ip);
3070 EXPORT_SYMBOL(preempt_count_add);
3071 NOKPROBE_SYMBOL(preempt_count_add);
3073 void preempt_count_sub(int val)
3075 #ifdef CONFIG_DEBUG_PREEMPT
3079 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3082 * Is the spinlock portion underflowing?
3084 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3085 !(preempt_count() & PREEMPT_MASK)))
3089 if (preempt_count() == val)
3090 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3091 __preempt_count_sub(val);
3093 EXPORT_SYMBOL(preempt_count_sub);
3094 NOKPROBE_SYMBOL(preempt_count_sub);
3099 * Print scheduling while atomic bug:
3101 static noinline void __schedule_bug(struct task_struct *prev)
3103 if (oops_in_progress)
3106 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3107 prev->comm, prev->pid, preempt_count());
3109 debug_show_held_locks(prev);
3111 if (irqs_disabled())
3112 print_irqtrace_events(prev);
3113 #ifdef CONFIG_DEBUG_PREEMPT
3114 if (in_atomic_preempt_off()) {
3115 pr_err("Preemption disabled at:");
3116 print_ip_sym(current->preempt_disable_ip);
3121 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3125 * Various schedule()-time debugging checks and statistics:
3127 static inline void schedule_debug(struct task_struct *prev)
3129 #ifdef CONFIG_SCHED_STACK_END_CHECK
3130 BUG_ON(task_stack_end_corrupted(prev));
3133 if (unlikely(in_atomic_preempt_off())) {
3134 __schedule_bug(prev);
3135 preempt_count_set(PREEMPT_DISABLED);
3139 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3141 schedstat_inc(this_rq(), sched_count);
3144 #if defined(CONFIG_PREEMPT_RT_FULL) && defined(CONFIG_SMP)
3146 void migrate_disable(void)
3148 struct task_struct *p = current;
3150 if (in_atomic() || irqs_disabled()) {
3151 #ifdef CONFIG_SCHED_DEBUG
3152 p->migrate_disable_atomic++;
3157 #ifdef CONFIG_SCHED_DEBUG
3158 if (unlikely(p->migrate_disable_atomic)) {
3164 if (p->migrate_disable) {
3165 p->migrate_disable++;
3170 preempt_lazy_disable();
3172 p->migrate_disable = 1;
3175 EXPORT_SYMBOL(migrate_disable);
3177 void migrate_enable(void)
3179 struct task_struct *p = current;
3181 if (in_atomic() || irqs_disabled()) {
3182 #ifdef CONFIG_SCHED_DEBUG
3183 p->migrate_disable_atomic--;
3188 #ifdef CONFIG_SCHED_DEBUG
3189 if (unlikely(p->migrate_disable_atomic)) {
3194 WARN_ON_ONCE(p->migrate_disable <= 0);
3196 if (p->migrate_disable > 1) {
3197 p->migrate_disable--;
3203 * Clearing migrate_disable causes tsk_cpus_allowed to
3204 * show the tasks original cpu affinity.
3206 p->migrate_disable = 0;
3208 unpin_current_cpu();
3210 preempt_lazy_enable();
3212 EXPORT_SYMBOL(migrate_enable);
3216 * Pick up the highest-prio task:
3218 static inline struct task_struct *
3219 pick_next_task(struct rq *rq, struct task_struct *prev)
3221 const struct sched_class *class = &fair_sched_class;
3222 struct task_struct *p;
3225 * Optimization: we know that if all tasks are in
3226 * the fair class we can call that function directly:
3228 if (likely(prev->sched_class == class &&
3229 rq->nr_running == rq->cfs.h_nr_running)) {
3230 p = fair_sched_class.pick_next_task(rq, prev);
3231 if (unlikely(p == RETRY_TASK))
3234 /* assumes fair_sched_class->next == idle_sched_class */
3236 p = idle_sched_class.pick_next_task(rq, prev);
3242 for_each_class(class) {
3243 p = class->pick_next_task(rq, prev);
3245 if (unlikely(p == RETRY_TASK))
3251 BUG(); /* the idle class will always have a runnable task */
3255 * __schedule() is the main scheduler function.
3257 * The main means of driving the scheduler and thus entering this function are:
3259 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3261 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3262 * paths. For example, see arch/x86/entry_64.S.
3264 * To drive preemption between tasks, the scheduler sets the flag in timer
3265 * interrupt handler scheduler_tick().
3267 * 3. Wakeups don't really cause entry into schedule(). They add a
3268 * task to the run-queue and that's it.
3270 * Now, if the new task added to the run-queue preempts the current
3271 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3272 * called on the nearest possible occasion:
3274 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3276 * - in syscall or exception context, at the next outmost
3277 * preempt_enable(). (this might be as soon as the wake_up()'s
3280 * - in IRQ context, return from interrupt-handler to
3281 * preemptible context
3283 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3286 * - cond_resched() call
3287 * - explicit schedule() call
3288 * - return from syscall or exception to user-space
3289 * - return from interrupt-handler to user-space
3291 * WARNING: must be called with preemption disabled!
3293 static void __sched notrace __schedule(bool preempt)
3295 struct task_struct *prev, *next;
3296 unsigned long *switch_count;
3300 cpu = smp_processor_id();
3302 rcu_note_context_switch();
3306 * do_exit() calls schedule() with preemption disabled as an exception;
3307 * however we must fix that up, otherwise the next task will see an
3308 * inconsistent (higher) preempt count.
3310 * It also avoids the below schedule_debug() test from complaining
3313 if (unlikely(prev->state == TASK_DEAD))
3314 preempt_enable_no_resched_notrace();
3316 schedule_debug(prev);
3318 if (sched_feat(HRTICK))
3322 * Make sure that signal_pending_state()->signal_pending() below
3323 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3324 * done by the caller to avoid the race with signal_wake_up().
3326 smp_mb__before_spinlock();
3327 raw_spin_lock_irq(&rq->lock);
3328 lockdep_pin_lock(&rq->lock);
3330 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3332 switch_count = &prev->nivcsw;
3333 if (!preempt && prev->state) {
3334 if (unlikely(signal_pending_state(prev->state, prev))) {
3335 prev->state = TASK_RUNNING;
3337 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3340 switch_count = &prev->nvcsw;
3343 if (task_on_rq_queued(prev))
3344 update_rq_clock(rq);
3346 next = pick_next_task(rq, prev);
3347 clear_tsk_need_resched(prev);
3348 clear_tsk_need_resched_lazy(prev);
3349 clear_preempt_need_resched();
3350 rq->clock_skip_update = 0;
3352 if (likely(prev != next)) {
3357 trace_sched_switch(preempt, prev, next);
3358 rq = context_switch(rq, prev, next); /* unlocks the rq */
3361 lockdep_unpin_lock(&rq->lock);
3362 raw_spin_unlock_irq(&rq->lock);
3365 balance_callback(rq);
3368 static inline void sched_submit_work(struct task_struct *tsk)
3373 * If a worker went to sleep, notify and ask workqueue whether
3374 * it wants to wake up a task to maintain concurrency.
3376 if (tsk->flags & PF_WQ_WORKER)
3377 wq_worker_sleeping(tsk);
3380 if (tsk_is_pi_blocked(tsk))
3384 * If we are going to sleep and we have plugged IO queued,
3385 * make sure to submit it to avoid deadlocks.
3387 if (blk_needs_flush_plug(tsk))
3388 blk_schedule_flush_plug(tsk);
3391 static void sched_update_worker(struct task_struct *tsk)
3393 if (tsk->flags & PF_WQ_WORKER)
3394 wq_worker_running(tsk);
3397 asmlinkage __visible void __sched schedule(void)
3399 struct task_struct *tsk = current;
3401 sched_submit_work(tsk);
3405 sched_preempt_enable_no_resched();
3406 } while (need_resched());
3407 sched_update_worker(tsk);
3409 EXPORT_SYMBOL(schedule);
3411 #ifdef CONFIG_CONTEXT_TRACKING
3412 asmlinkage __visible void __sched schedule_user(void)
3415 * If we come here after a random call to set_need_resched(),
3416 * or we have been woken up remotely but the IPI has not yet arrived,
3417 * we haven't yet exited the RCU idle mode. Do it here manually until
3418 * we find a better solution.
3420 * NB: There are buggy callers of this function. Ideally we
3421 * should warn if prev_state != CONTEXT_USER, but that will trigger
3422 * too frequently to make sense yet.
3424 enum ctx_state prev_state = exception_enter();
3426 exception_exit(prev_state);
3431 * schedule_preempt_disabled - called with preemption disabled
3433 * Returns with preemption disabled. Note: preempt_count must be 1
3435 void __sched schedule_preempt_disabled(void)
3437 sched_preempt_enable_no_resched();
3442 static void __sched notrace preempt_schedule_common(void)
3445 preempt_disable_notrace();
3447 preempt_enable_no_resched_notrace();
3450 * Check again in case we missed a preemption opportunity
3451 * between schedule and now.
3453 } while (need_resched());
3456 #ifdef CONFIG_PREEMPT_LAZY
3458 * If TIF_NEED_RESCHED is then we allow to be scheduled away since this is
3459 * set by a RT task. Oterwise we try to avoid beeing scheduled out as long as
3460 * preempt_lazy_count counter >0.
3462 static __always_inline int preemptible_lazy(void)
3464 if (test_thread_flag(TIF_NEED_RESCHED))
3466 if (current_thread_info()->preempt_lazy_count)
3473 static int preemptible_lazy(void)
3480 #ifdef CONFIG_PREEMPT
3482 * this is the entry point to schedule() from in-kernel preemption
3483 * off of preempt_enable. Kernel preemptions off return from interrupt
3484 * occur there and call schedule directly.
3486 asmlinkage __visible void __sched notrace preempt_schedule(void)
3489 * If there is a non-zero preempt_count or interrupts are disabled,
3490 * we do not want to preempt the current task. Just return..
3492 if (likely(!preemptible()))
3494 if (!preemptible_lazy())
3497 preempt_schedule_common();
3499 NOKPROBE_SYMBOL(preempt_schedule);
3500 EXPORT_SYMBOL(preempt_schedule);
3503 * preempt_schedule_notrace - preempt_schedule called by tracing
3505 * The tracing infrastructure uses preempt_enable_notrace to prevent
3506 * recursion and tracing preempt enabling caused by the tracing
3507 * infrastructure itself. But as tracing can happen in areas coming
3508 * from userspace or just about to enter userspace, a preempt enable
3509 * can occur before user_exit() is called. This will cause the scheduler
3510 * to be called when the system is still in usermode.
3512 * To prevent this, the preempt_enable_notrace will use this function
3513 * instead of preempt_schedule() to exit user context if needed before
3514 * calling the scheduler.
3516 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3518 enum ctx_state prev_ctx;
3520 if (likely(!preemptible()))
3522 if (!preemptible_lazy())
3526 preempt_disable_notrace();
3528 * Needs preempt disabled in case user_exit() is traced
3529 * and the tracer calls preempt_enable_notrace() causing
3530 * an infinite recursion.
3532 prev_ctx = exception_enter();
3534 * The add/subtract must not be traced by the function
3535 * tracer. But we still want to account for the
3536 * preempt off latency tracer. Since the _notrace versions
3537 * of add/subtract skip the accounting for latency tracer
3538 * we must force it manually.
3540 start_critical_timings();
3542 stop_critical_timings();
3543 exception_exit(prev_ctx);
3545 preempt_enable_no_resched_notrace();
3546 } while (need_resched());
3548 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3550 #endif /* CONFIG_PREEMPT */
3553 * this is the entry point to schedule() from kernel preemption
3554 * off of irq context.
3555 * Note, that this is called and return with irqs disabled. This will
3556 * protect us against recursive calling from irq.
3558 asmlinkage __visible void __sched preempt_schedule_irq(void)
3560 enum ctx_state prev_state;
3562 /* Catch callers which need to be fixed */
3563 BUG_ON(preempt_count() || !irqs_disabled());
3565 prev_state = exception_enter();
3571 local_irq_disable();
3572 sched_preempt_enable_no_resched();
3573 } while (need_resched());
3575 exception_exit(prev_state);
3578 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3581 return try_to_wake_up(curr->private, mode, wake_flags);
3583 EXPORT_SYMBOL(default_wake_function);
3585 #ifdef CONFIG_RT_MUTEXES
3588 * rt_mutex_setprio - set the current priority of a task
3590 * @prio: prio value (kernel-internal form)
3592 * This function changes the 'effective' priority of a task. It does
3593 * not touch ->normal_prio like __setscheduler().
3595 * Used by the rt_mutex code to implement priority inheritance
3596 * logic. Call site only calls if the priority of the task changed.
3598 void rt_mutex_setprio(struct task_struct *p, int prio)
3600 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3602 const struct sched_class *prev_class;
3604 BUG_ON(prio > MAX_PRIO);
3606 rq = __task_rq_lock(p);
3609 * Idle task boosting is a nono in general. There is one
3610 * exception, when PREEMPT_RT and NOHZ is active:
3612 * The idle task calls get_next_timer_interrupt() and holds
3613 * the timer wheel base->lock on the CPU and another CPU wants
3614 * to access the timer (probably to cancel it). We can safely
3615 * ignore the boosting request, as the idle CPU runs this code
3616 * with interrupts disabled and will complete the lock
3617 * protected section without being interrupted. So there is no
3618 * real need to boost.
3620 if (unlikely(p == rq->idle)) {
3621 WARN_ON(p != rq->curr);
3622 WARN_ON(p->pi_blocked_on);
3626 trace_sched_pi_setprio(p, prio);
3628 prev_class = p->sched_class;
3629 queued = task_on_rq_queued(p);
3630 running = task_current(rq, p);
3632 dequeue_task(rq, p, DEQUEUE_SAVE);
3634 put_prev_task(rq, p);
3637 * Boosting condition are:
3638 * 1. -rt task is running and holds mutex A
3639 * --> -dl task blocks on mutex A
3641 * 2. -dl task is running and holds mutex A
3642 * --> -dl task blocks on mutex A and could preempt the
3645 if (dl_prio(prio)) {
3646 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3647 if (!dl_prio(p->normal_prio) ||
3648 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3649 p->dl.dl_boosted = 1;
3650 enqueue_flag |= ENQUEUE_REPLENISH;
3652 p->dl.dl_boosted = 0;
3653 p->sched_class = &dl_sched_class;
3654 } else if (rt_prio(prio)) {
3655 if (dl_prio(oldprio))
3656 p->dl.dl_boosted = 0;
3658 enqueue_flag |= ENQUEUE_HEAD;
3659 p->sched_class = &rt_sched_class;
3661 if (dl_prio(oldprio))
3662 p->dl.dl_boosted = 0;
3663 if (rt_prio(oldprio))
3665 p->sched_class = &fair_sched_class;
3671 p->sched_class->set_curr_task(rq);
3673 enqueue_task(rq, p, enqueue_flag);
3675 check_class_changed(rq, p, prev_class, oldprio);
3677 preempt_disable(); /* avoid rq from going away on us */
3678 __task_rq_unlock(rq);
3680 balance_callback(rq);
3685 void set_user_nice(struct task_struct *p, long nice)
3687 int old_prio, delta, queued;
3688 unsigned long flags;
3691 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3694 * We have to be careful, if called from sys_setpriority(),
3695 * the task might be in the middle of scheduling on another CPU.
3697 rq = task_rq_lock(p, &flags);
3699 * The RT priorities are set via sched_setscheduler(), but we still
3700 * allow the 'normal' nice value to be set - but as expected
3701 * it wont have any effect on scheduling until the task is
3702 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3704 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3705 p->static_prio = NICE_TO_PRIO(nice);
3708 queued = task_on_rq_queued(p);
3710 dequeue_task(rq, p, DEQUEUE_SAVE);
3712 p->static_prio = NICE_TO_PRIO(nice);
3715 p->prio = effective_prio(p);
3716 delta = p->prio - old_prio;
3719 enqueue_task(rq, p, ENQUEUE_RESTORE);
3721 * If the task increased its priority or is running and
3722 * lowered its priority, then reschedule its CPU:
3724 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3728 task_rq_unlock(rq, p, &flags);
3730 EXPORT_SYMBOL(set_user_nice);
3733 * can_nice - check if a task can reduce its nice value
3737 int can_nice(const struct task_struct *p, const int nice)
3739 /* convert nice value [19,-20] to rlimit style value [1,40] */
3740 int nice_rlim = nice_to_rlimit(nice);
3742 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3743 capable(CAP_SYS_NICE));
3746 #ifdef __ARCH_WANT_SYS_NICE
3749 * sys_nice - change the priority of the current process.
3750 * @increment: priority increment
3752 * sys_setpriority is a more generic, but much slower function that
3753 * does similar things.
3755 SYSCALL_DEFINE1(nice, int, increment)
3760 * Setpriority might change our priority at the same moment.
3761 * We don't have to worry. Conceptually one call occurs first
3762 * and we have a single winner.
3764 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3765 nice = task_nice(current) + increment;
3767 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3768 if (increment < 0 && !can_nice(current, nice))
3771 retval = security_task_setnice(current, nice);
3775 set_user_nice(current, nice);
3782 * task_prio - return the priority value of a given task.
3783 * @p: the task in question.
3785 * Return: The priority value as seen by users in /proc.
3786 * RT tasks are offset by -200. Normal tasks are centered
3787 * around 0, value goes from -16 to +15.
3789 int task_prio(const struct task_struct *p)
3791 return p->prio - MAX_RT_PRIO;
3795 * idle_cpu - is a given cpu idle currently?
3796 * @cpu: the processor in question.
3798 * Return: 1 if the CPU is currently idle. 0 otherwise.
3800 int idle_cpu(int cpu)
3802 struct rq *rq = cpu_rq(cpu);
3804 if (rq->curr != rq->idle)
3811 if (!llist_empty(&rq->wake_list))
3819 * idle_task - return the idle task for a given cpu.
3820 * @cpu: the processor in question.
3822 * Return: The idle task for the cpu @cpu.
3824 struct task_struct *idle_task(int cpu)
3826 return cpu_rq(cpu)->idle;
3830 * find_process_by_pid - find a process with a matching PID value.
3831 * @pid: the pid in question.
3833 * The task of @pid, if found. %NULL otherwise.
3835 static struct task_struct *find_process_by_pid(pid_t pid)
3837 return pid ? find_task_by_vpid(pid) : current;
3841 * This function initializes the sched_dl_entity of a newly becoming
3842 * SCHED_DEADLINE task.
3844 * Only the static values are considered here, the actual runtime and the
3845 * absolute deadline will be properly calculated when the task is enqueued
3846 * for the first time with its new policy.
3849 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3851 struct sched_dl_entity *dl_se = &p->dl;
3853 dl_se->dl_runtime = attr->sched_runtime;
3854 dl_se->dl_deadline = attr->sched_deadline;
3855 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3856 dl_se->flags = attr->sched_flags;
3857 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3860 * Changing the parameters of a task is 'tricky' and we're not doing
3861 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3863 * What we SHOULD do is delay the bandwidth release until the 0-lag
3864 * point. This would include retaining the task_struct until that time
3865 * and change dl_overflow() to not immediately decrement the current
3868 * Instead we retain the current runtime/deadline and let the new
3869 * parameters take effect after the current reservation period lapses.
3870 * This is safe (albeit pessimistic) because the 0-lag point is always
3871 * before the current scheduling deadline.
3873 * We can still have temporary overloads because we do not delay the
3874 * change in bandwidth until that time; so admission control is
3875 * not on the safe side. It does however guarantee tasks will never
3876 * consume more than promised.
3881 * sched_setparam() passes in -1 for its policy, to let the functions
3882 * it calls know not to change it.
3884 #define SETPARAM_POLICY -1
3886 static void __setscheduler_params(struct task_struct *p,
3887 const struct sched_attr *attr)
3889 int policy = attr->sched_policy;
3891 if (policy == SETPARAM_POLICY)
3896 if (dl_policy(policy))
3897 __setparam_dl(p, attr);
3898 else if (fair_policy(policy))
3899 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3902 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3903 * !rt_policy. Always setting this ensures that things like
3904 * getparam()/getattr() don't report silly values for !rt tasks.
3906 p->rt_priority = attr->sched_priority;
3907 p->normal_prio = normal_prio(p);
3911 /* Actually do priority change: must hold pi & rq lock. */
3912 static void __setscheduler(struct rq *rq, struct task_struct *p,
3913 const struct sched_attr *attr, bool keep_boost)
3915 __setscheduler_params(p, attr);
3918 * Keep a potential priority boosting if called from
3919 * sched_setscheduler().
3922 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3924 p->prio = normal_prio(p);
3926 if (dl_prio(p->prio))
3927 p->sched_class = &dl_sched_class;
3928 else if (rt_prio(p->prio))
3929 p->sched_class = &rt_sched_class;
3931 p->sched_class = &fair_sched_class;
3935 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3937 struct sched_dl_entity *dl_se = &p->dl;
3939 attr->sched_priority = p->rt_priority;
3940 attr->sched_runtime = dl_se->dl_runtime;
3941 attr->sched_deadline = dl_se->dl_deadline;
3942 attr->sched_period = dl_se->dl_period;
3943 attr->sched_flags = dl_se->flags;
3947 * This function validates the new parameters of a -deadline task.
3948 * We ask for the deadline not being zero, and greater or equal
3949 * than the runtime, as well as the period of being zero or
3950 * greater than deadline. Furthermore, we have to be sure that
3951 * user parameters are above the internal resolution of 1us (we
3952 * check sched_runtime only since it is always the smaller one) and
3953 * below 2^63 ns (we have to check both sched_deadline and
3954 * sched_period, as the latter can be zero).
3957 __checkparam_dl(const struct sched_attr *attr)
3960 if (attr->sched_deadline == 0)
3964 * Since we truncate DL_SCALE bits, make sure we're at least
3967 if (attr->sched_runtime < (1ULL << DL_SCALE))
3971 * Since we use the MSB for wrap-around and sign issues, make
3972 * sure it's not set (mind that period can be equal to zero).
3974 if (attr->sched_deadline & (1ULL << 63) ||
3975 attr->sched_period & (1ULL << 63))
3978 /* runtime <= deadline <= period (if period != 0) */
3979 if ((attr->sched_period != 0 &&
3980 attr->sched_period < attr->sched_deadline) ||
3981 attr->sched_deadline < attr->sched_runtime)
3988 * check the target process has a UID that matches the current process's
3990 static bool check_same_owner(struct task_struct *p)
3992 const struct cred *cred = current_cred(), *pcred;
3996 pcred = __task_cred(p);
3997 match = (uid_eq(cred->euid, pcred->euid) ||
3998 uid_eq(cred->euid, pcred->uid));
4003 static bool dl_param_changed(struct task_struct *p,
4004 const struct sched_attr *attr)
4006 struct sched_dl_entity *dl_se = &p->dl;
4008 if (dl_se->dl_runtime != attr->sched_runtime ||
4009 dl_se->dl_deadline != attr->sched_deadline ||
4010 dl_se->dl_period != attr->sched_period ||
4011 dl_se->flags != attr->sched_flags)
4017 static int __sched_setscheduler(struct task_struct *p,
4018 const struct sched_attr *attr,
4021 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4022 MAX_RT_PRIO - 1 - attr->sched_priority;
4023 int retval, oldprio, oldpolicy = -1, queued, running;
4024 int new_effective_prio, policy = attr->sched_policy;
4025 unsigned long flags;
4026 const struct sched_class *prev_class;
4030 /* may grab non-irq protected spin_locks */
4031 BUG_ON(in_interrupt());
4033 /* double check policy once rq lock held */
4035 reset_on_fork = p->sched_reset_on_fork;
4036 policy = oldpolicy = p->policy;
4038 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4040 if (!valid_policy(policy))
4044 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4048 * Valid priorities for SCHED_FIFO and SCHED_RR are
4049 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4050 * SCHED_BATCH and SCHED_IDLE is 0.
4052 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4053 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4055 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4056 (rt_policy(policy) != (attr->sched_priority != 0)))
4060 * Allow unprivileged RT tasks to decrease priority:
4062 if (user && !capable(CAP_SYS_NICE)) {
4063 if (fair_policy(policy)) {
4064 if (attr->sched_nice < task_nice(p) &&
4065 !can_nice(p, attr->sched_nice))
4069 if (rt_policy(policy)) {
4070 unsigned long rlim_rtprio =
4071 task_rlimit(p, RLIMIT_RTPRIO);
4073 /* can't set/change the rt policy */
4074 if (policy != p->policy && !rlim_rtprio)
4077 /* can't increase priority */
4078 if (attr->sched_priority > p->rt_priority &&
4079 attr->sched_priority > rlim_rtprio)
4084 * Can't set/change SCHED_DEADLINE policy at all for now
4085 * (safest behavior); in the future we would like to allow
4086 * unprivileged DL tasks to increase their relative deadline
4087 * or reduce their runtime (both ways reducing utilization)
4089 if (dl_policy(policy))
4093 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4094 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4096 if (idle_policy(p->policy) && !idle_policy(policy)) {
4097 if (!can_nice(p, task_nice(p)))
4101 /* can't change other user's priorities */
4102 if (!check_same_owner(p))
4105 /* Normal users shall not reset the sched_reset_on_fork flag */
4106 if (p->sched_reset_on_fork && !reset_on_fork)
4111 retval = security_task_setscheduler(p);
4117 * make sure no PI-waiters arrive (or leave) while we are
4118 * changing the priority of the task:
4120 * To be able to change p->policy safely, the appropriate
4121 * runqueue lock must be held.
4123 rq = task_rq_lock(p, &flags);
4126 * Changing the policy of the stop threads its a very bad idea
4128 if (p == rq->stop) {
4129 task_rq_unlock(rq, p, &flags);
4134 * If not changing anything there's no need to proceed further,
4135 * but store a possible modification of reset_on_fork.
4137 if (unlikely(policy == p->policy)) {
4138 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4140 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4142 if (dl_policy(policy) && dl_param_changed(p, attr))
4145 p->sched_reset_on_fork = reset_on_fork;
4146 task_rq_unlock(rq, p, &flags);
4152 #ifdef CONFIG_RT_GROUP_SCHED
4154 * Do not allow realtime tasks into groups that have no runtime
4157 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4158 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4159 !task_group_is_autogroup(task_group(p))) {
4160 task_rq_unlock(rq, p, &flags);
4165 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4166 cpumask_t *span = rq->rd->span;
4169 * Don't allow tasks with an affinity mask smaller than
4170 * the entire root_domain to become SCHED_DEADLINE. We
4171 * will also fail if there's no bandwidth available.
4173 if (!cpumask_subset(span, &p->cpus_allowed) ||
4174 rq->rd->dl_bw.bw == 0) {
4175 task_rq_unlock(rq, p, &flags);
4182 /* recheck policy now with rq lock held */
4183 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4184 policy = oldpolicy = -1;
4185 task_rq_unlock(rq, p, &flags);
4190 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4191 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4194 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4195 task_rq_unlock(rq, p, &flags);
4199 p->sched_reset_on_fork = reset_on_fork;
4204 * Take priority boosted tasks into account. If the new
4205 * effective priority is unchanged, we just store the new
4206 * normal parameters and do not touch the scheduler class and
4207 * the runqueue. This will be done when the task deboost
4210 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4211 if (new_effective_prio == oldprio) {
4212 __setscheduler_params(p, attr);
4213 task_rq_unlock(rq, p, &flags);
4218 queued = task_on_rq_queued(p);
4219 running = task_current(rq, p);
4221 dequeue_task(rq, p, DEQUEUE_SAVE);
4223 put_prev_task(rq, p);
4225 prev_class = p->sched_class;
4226 __setscheduler(rq, p, attr, pi);
4229 p->sched_class->set_curr_task(rq);
4231 int enqueue_flags = ENQUEUE_RESTORE;
4233 * We enqueue to tail when the priority of a task is
4234 * increased (user space view).
4236 if (oldprio <= p->prio)
4237 enqueue_flags |= ENQUEUE_HEAD;
4239 enqueue_task(rq, p, enqueue_flags);
4242 check_class_changed(rq, p, prev_class, oldprio);
4243 preempt_disable(); /* avoid rq from going away on us */
4244 task_rq_unlock(rq, p, &flags);
4247 rt_mutex_adjust_pi(p);
4250 * Run balance callbacks after we've adjusted the PI chain.
4252 balance_callback(rq);
4258 static int _sched_setscheduler(struct task_struct *p, int policy,
4259 const struct sched_param *param, bool check)
4261 struct sched_attr attr = {
4262 .sched_policy = policy,
4263 .sched_priority = param->sched_priority,
4264 .sched_nice = PRIO_TO_NICE(p->static_prio),
4267 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4268 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4269 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4270 policy &= ~SCHED_RESET_ON_FORK;
4271 attr.sched_policy = policy;
4274 return __sched_setscheduler(p, &attr, check, true);
4277 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4278 * @p: the task in question.
4279 * @policy: new policy.
4280 * @param: structure containing the new RT priority.
4282 * Return: 0 on success. An error code otherwise.
4284 * NOTE that the task may be already dead.
4286 int sched_setscheduler(struct task_struct *p, int policy,
4287 const struct sched_param *param)
4289 return _sched_setscheduler(p, policy, param, true);
4291 EXPORT_SYMBOL_GPL(sched_setscheduler);
4293 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4295 return __sched_setscheduler(p, attr, true, true);
4297 EXPORT_SYMBOL_GPL(sched_setattr);
4300 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4301 * @p: the task in question.
4302 * @policy: new policy.
4303 * @param: structure containing the new RT priority.
4305 * Just like sched_setscheduler, only don't bother checking if the
4306 * current context has permission. For example, this is needed in
4307 * stop_machine(): we create temporary high priority worker threads,
4308 * but our caller might not have that capability.
4310 * Return: 0 on success. An error code otherwise.
4312 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4313 const struct sched_param *param)
4315 return _sched_setscheduler(p, policy, param, false);
4317 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4320 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4322 struct sched_param lparam;
4323 struct task_struct *p;
4326 if (!param || pid < 0)
4328 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4333 p = find_process_by_pid(pid);
4335 retval = sched_setscheduler(p, policy, &lparam);
4342 * Mimics kernel/events/core.c perf_copy_attr().
4344 static int sched_copy_attr(struct sched_attr __user *uattr,
4345 struct sched_attr *attr)
4350 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4354 * zero the full structure, so that a short copy will be nice.
4356 memset(attr, 0, sizeof(*attr));
4358 ret = get_user(size, &uattr->size);
4362 if (size > PAGE_SIZE) /* silly large */
4365 if (!size) /* abi compat */
4366 size = SCHED_ATTR_SIZE_VER0;
4368 if (size < SCHED_ATTR_SIZE_VER0)
4372 * If we're handed a bigger struct than we know of,
4373 * ensure all the unknown bits are 0 - i.e. new
4374 * user-space does not rely on any kernel feature
4375 * extensions we dont know about yet.
4377 if (size > sizeof(*attr)) {
4378 unsigned char __user *addr;
4379 unsigned char __user *end;
4382 addr = (void __user *)uattr + sizeof(*attr);
4383 end = (void __user *)uattr + size;
4385 for (; addr < end; addr++) {
4386 ret = get_user(val, addr);
4392 size = sizeof(*attr);
4395 ret = copy_from_user(attr, uattr, size);
4400 * XXX: do we want to be lenient like existing syscalls; or do we want
4401 * to be strict and return an error on out-of-bounds values?
4403 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4408 put_user(sizeof(*attr), &uattr->size);
4413 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4414 * @pid: the pid in question.
4415 * @policy: new policy.
4416 * @param: structure containing the new RT priority.
4418 * Return: 0 on success. An error code otherwise.
4420 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4421 struct sched_param __user *, param)
4423 /* negative values for policy are not valid */
4427 return do_sched_setscheduler(pid, policy, param);
4431 * sys_sched_setparam - set/change the RT priority of a thread
4432 * @pid: the pid in question.
4433 * @param: structure containing the new RT priority.
4435 * Return: 0 on success. An error code otherwise.
4437 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4439 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4443 * sys_sched_setattr - same as above, but with extended sched_attr
4444 * @pid: the pid in question.
4445 * @uattr: structure containing the extended parameters.
4446 * @flags: for future extension.
4448 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4449 unsigned int, flags)
4451 struct sched_attr attr;
4452 struct task_struct *p;
4455 if (!uattr || pid < 0 || flags)
4458 retval = sched_copy_attr(uattr, &attr);
4462 if ((int)attr.sched_policy < 0)
4467 p = find_process_by_pid(pid);
4469 retval = sched_setattr(p, &attr);
4476 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4477 * @pid: the pid in question.
4479 * Return: On success, the policy of the thread. Otherwise, a negative error
4482 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4484 struct task_struct *p;
4492 p = find_process_by_pid(pid);
4494 retval = security_task_getscheduler(p);
4497 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4504 * sys_sched_getparam - get the RT priority of a thread
4505 * @pid: the pid in question.
4506 * @param: structure containing the RT priority.
4508 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4511 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4513 struct sched_param lp = { .sched_priority = 0 };
4514 struct task_struct *p;
4517 if (!param || pid < 0)
4521 p = find_process_by_pid(pid);
4526 retval = security_task_getscheduler(p);
4530 if (task_has_rt_policy(p))
4531 lp.sched_priority = p->rt_priority;
4535 * This one might sleep, we cannot do it with a spinlock held ...
4537 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4546 static int sched_read_attr(struct sched_attr __user *uattr,
4547 struct sched_attr *attr,
4552 if (!access_ok(VERIFY_WRITE, uattr, usize))
4556 * If we're handed a smaller struct than we know of,
4557 * ensure all the unknown bits are 0 - i.e. old
4558 * user-space does not get uncomplete information.
4560 if (usize < sizeof(*attr)) {
4561 unsigned char *addr;
4564 addr = (void *)attr + usize;
4565 end = (void *)attr + sizeof(*attr);
4567 for (; addr < end; addr++) {
4575 ret = copy_to_user(uattr, attr, attr->size);
4583 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4584 * @pid: the pid in question.
4585 * @uattr: structure containing the extended parameters.
4586 * @size: sizeof(attr) for fwd/bwd comp.
4587 * @flags: for future extension.
4589 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4590 unsigned int, size, unsigned int, flags)
4592 struct sched_attr attr = {
4593 .size = sizeof(struct sched_attr),
4595 struct task_struct *p;
4598 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4599 size < SCHED_ATTR_SIZE_VER0 || flags)
4603 p = find_process_by_pid(pid);
4608 retval = security_task_getscheduler(p);
4612 attr.sched_policy = p->policy;
4613 if (p->sched_reset_on_fork)
4614 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4615 if (task_has_dl_policy(p))
4616 __getparam_dl(p, &attr);
4617 else if (task_has_rt_policy(p))
4618 attr.sched_priority = p->rt_priority;
4620 attr.sched_nice = task_nice(p);
4624 retval = sched_read_attr(uattr, &attr, size);
4632 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4634 cpumask_var_t cpus_allowed, new_mask;
4635 struct task_struct *p;
4640 p = find_process_by_pid(pid);
4646 /* Prevent p going away */
4650 if (p->flags & PF_NO_SETAFFINITY) {
4654 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4658 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4660 goto out_free_cpus_allowed;
4663 if (!check_same_owner(p)) {
4665 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4667 goto out_free_new_mask;
4672 retval = security_task_setscheduler(p);
4674 goto out_free_new_mask;
4677 cpuset_cpus_allowed(p, cpus_allowed);
4678 cpumask_and(new_mask, in_mask, cpus_allowed);
4681 * Since bandwidth control happens on root_domain basis,
4682 * if admission test is enabled, we only admit -deadline
4683 * tasks allowed to run on all the CPUs in the task's
4687 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4689 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4692 goto out_free_new_mask;
4698 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4701 cpuset_cpus_allowed(p, cpus_allowed);
4702 if (!cpumask_subset(new_mask, cpus_allowed)) {
4704 * We must have raced with a concurrent cpuset
4705 * update. Just reset the cpus_allowed to the
4706 * cpuset's cpus_allowed
4708 cpumask_copy(new_mask, cpus_allowed);
4713 free_cpumask_var(new_mask);
4714 out_free_cpus_allowed:
4715 free_cpumask_var(cpus_allowed);
4721 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4722 struct cpumask *new_mask)
4724 if (len < cpumask_size())
4725 cpumask_clear(new_mask);
4726 else if (len > cpumask_size())
4727 len = cpumask_size();
4729 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4733 * sys_sched_setaffinity - set the cpu affinity of a process
4734 * @pid: pid of the process
4735 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4736 * @user_mask_ptr: user-space pointer to the new cpu mask
4738 * Return: 0 on success. An error code otherwise.
4740 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4741 unsigned long __user *, user_mask_ptr)
4743 cpumask_var_t new_mask;
4746 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4749 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4751 retval = sched_setaffinity(pid, new_mask);
4752 free_cpumask_var(new_mask);
4756 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4758 struct task_struct *p;
4759 unsigned long flags;
4765 p = find_process_by_pid(pid);
4769 retval = security_task_getscheduler(p);
4773 raw_spin_lock_irqsave(&p->pi_lock, flags);
4774 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4775 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4784 * sys_sched_getaffinity - get the cpu affinity of a process
4785 * @pid: pid of the process
4786 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4787 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4789 * Return: 0 on success. An error code otherwise.
4791 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4792 unsigned long __user *, user_mask_ptr)
4797 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4799 if (len & (sizeof(unsigned long)-1))
4802 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4805 ret = sched_getaffinity(pid, mask);
4807 size_t retlen = min_t(size_t, len, cpumask_size());
4809 if (copy_to_user(user_mask_ptr, mask, retlen))
4814 free_cpumask_var(mask);
4820 * sys_sched_yield - yield the current processor to other threads.
4822 * This function yields the current CPU to other tasks. If there are no
4823 * other threads running on this CPU then this function will return.
4827 SYSCALL_DEFINE0(sched_yield)
4829 struct rq *rq = this_rq_lock();
4831 schedstat_inc(rq, yld_count);
4832 current->sched_class->yield_task(rq);
4835 * Since we are going to call schedule() anyway, there's
4836 * no need to preempt or enable interrupts:
4838 __release(rq->lock);
4839 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4840 do_raw_spin_unlock(&rq->lock);
4841 sched_preempt_enable_no_resched();
4848 int __sched _cond_resched(void)
4850 if (should_resched(0)) {
4851 preempt_schedule_common();
4856 EXPORT_SYMBOL(_cond_resched);
4859 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4860 * call schedule, and on return reacquire the lock.
4862 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4863 * operations here to prevent schedule() from being called twice (once via
4864 * spin_unlock(), once by hand).
4866 int __cond_resched_lock(spinlock_t *lock)
4868 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4871 lockdep_assert_held(lock);
4873 if (spin_needbreak(lock) || resched) {
4876 preempt_schedule_common();
4884 EXPORT_SYMBOL(__cond_resched_lock);
4886 #ifndef CONFIG_PREEMPT_RT_FULL
4887 int __sched __cond_resched_softirq(void)
4889 BUG_ON(!in_softirq());
4891 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4893 preempt_schedule_common();
4899 EXPORT_SYMBOL(__cond_resched_softirq);
4903 * yield - yield the current processor to other threads.
4905 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4907 * The scheduler is at all times free to pick the calling task as the most
4908 * eligible task to run, if removing the yield() call from your code breaks
4909 * it, its already broken.
4911 * Typical broken usage is:
4916 * where one assumes that yield() will let 'the other' process run that will
4917 * make event true. If the current task is a SCHED_FIFO task that will never
4918 * happen. Never use yield() as a progress guarantee!!
4920 * If you want to use yield() to wait for something, use wait_event().
4921 * If you want to use yield() to be 'nice' for others, use cond_resched().
4922 * If you still want to use yield(), do not!
4924 void __sched yield(void)
4926 set_current_state(TASK_RUNNING);
4929 EXPORT_SYMBOL(yield);
4932 * yield_to - yield the current processor to another thread in
4933 * your thread group, or accelerate that thread toward the
4934 * processor it's on.
4936 * @preempt: whether task preemption is allowed or not
4938 * It's the caller's job to ensure that the target task struct
4939 * can't go away on us before we can do any checks.
4942 * true (>0) if we indeed boosted the target task.
4943 * false (0) if we failed to boost the target.
4944 * -ESRCH if there's no task to yield to.
4946 int __sched yield_to(struct task_struct *p, bool preempt)
4948 struct task_struct *curr = current;
4949 struct rq *rq, *p_rq;
4950 unsigned long flags;
4953 local_irq_save(flags);
4959 * If we're the only runnable task on the rq and target rq also
4960 * has only one task, there's absolutely no point in yielding.
4962 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4967 double_rq_lock(rq, p_rq);
4968 if (task_rq(p) != p_rq) {
4969 double_rq_unlock(rq, p_rq);
4973 if (!curr->sched_class->yield_to_task)
4976 if (curr->sched_class != p->sched_class)
4979 if (task_running(p_rq, p) || p->state)
4982 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4984 schedstat_inc(rq, yld_count);
4986 * Make p's CPU reschedule; pick_next_entity takes care of
4989 if (preempt && rq != p_rq)
4994 double_rq_unlock(rq, p_rq);
4996 local_irq_restore(flags);
5003 EXPORT_SYMBOL_GPL(yield_to);
5006 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5007 * that process accounting knows that this is a task in IO wait state.
5009 long __sched io_schedule_timeout(long timeout)
5011 int old_iowait = current->in_iowait;
5015 current->in_iowait = 1;
5016 blk_schedule_flush_plug(current);
5018 delayacct_blkio_start();
5020 atomic_inc(&rq->nr_iowait);
5021 ret = schedule_timeout(timeout);
5022 current->in_iowait = old_iowait;
5023 atomic_dec(&rq->nr_iowait);
5024 delayacct_blkio_end();
5028 EXPORT_SYMBOL(io_schedule_timeout);
5031 * sys_sched_get_priority_max - return maximum RT priority.
5032 * @policy: scheduling class.
5034 * Return: On success, this syscall returns the maximum
5035 * rt_priority that can be used by a given scheduling class.
5036 * On failure, a negative error code is returned.
5038 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5045 ret = MAX_USER_RT_PRIO-1;
5047 case SCHED_DEADLINE:
5058 * sys_sched_get_priority_min - return minimum RT priority.
5059 * @policy: scheduling class.
5061 * Return: On success, this syscall returns the minimum
5062 * rt_priority that can be used by a given scheduling class.
5063 * On failure, a negative error code is returned.
5065 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5074 case SCHED_DEADLINE:
5084 * sys_sched_rr_get_interval - return the default timeslice of a process.
5085 * @pid: pid of the process.
5086 * @interval: userspace pointer to the timeslice value.
5088 * this syscall writes the default timeslice value of a given process
5089 * into the user-space timespec buffer. A value of '0' means infinity.
5091 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5094 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5095 struct timespec __user *, interval)
5097 struct task_struct *p;
5098 unsigned int time_slice;
5099 unsigned long flags;
5109 p = find_process_by_pid(pid);
5113 retval = security_task_getscheduler(p);
5117 rq = task_rq_lock(p, &flags);
5119 if (p->sched_class->get_rr_interval)
5120 time_slice = p->sched_class->get_rr_interval(rq, p);
5121 task_rq_unlock(rq, p, &flags);
5124 jiffies_to_timespec(time_slice, &t);
5125 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5133 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5135 void sched_show_task(struct task_struct *p)
5137 unsigned long free = 0;
5139 unsigned long state = p->state;
5142 state = __ffs(state) + 1;
5143 printk(KERN_INFO "%-15.15s %c", p->comm,
5144 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5145 #if BITS_PER_LONG == 32
5146 if (state == TASK_RUNNING)
5147 printk(KERN_CONT " running ");
5149 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5151 if (state == TASK_RUNNING)
5152 printk(KERN_CONT " running task ");
5154 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5156 #ifdef CONFIG_DEBUG_STACK_USAGE
5157 free = stack_not_used(p);
5162 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5164 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5165 task_pid_nr(p), ppid,
5166 (unsigned long)task_thread_info(p)->flags);
5168 print_worker_info(KERN_INFO, p);
5169 show_stack(p, NULL);
5172 void show_state_filter(unsigned long state_filter)
5174 struct task_struct *g, *p;
5176 #if BITS_PER_LONG == 32
5178 " task PC stack pid father\n");
5181 " task PC stack pid father\n");
5184 for_each_process_thread(g, p) {
5186 * reset the NMI-timeout, listing all files on a slow
5187 * console might take a lot of time:
5189 touch_nmi_watchdog();
5190 if (!state_filter || (p->state & state_filter))
5194 touch_all_softlockup_watchdogs();
5196 #ifdef CONFIG_SCHED_DEBUG
5197 sysrq_sched_debug_show();
5201 * Only show locks if all tasks are dumped:
5204 debug_show_all_locks();
5207 void init_idle_bootup_task(struct task_struct *idle)
5209 idle->sched_class = &idle_sched_class;
5213 * init_idle - set up an idle thread for a given CPU
5214 * @idle: task in question
5215 * @cpu: cpu the idle task belongs to
5217 * NOTE: this function does not set the idle thread's NEED_RESCHED
5218 * flag, to make booting more robust.
5220 void init_idle(struct task_struct *idle, int cpu)
5222 struct rq *rq = cpu_rq(cpu);
5223 unsigned long flags;
5225 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5226 raw_spin_lock(&rq->lock);
5228 __sched_fork(0, idle);
5229 idle->state = TASK_RUNNING;
5230 idle->se.exec_start = sched_clock();
5234 * Its possible that init_idle() gets called multiple times on a task,
5235 * in that case do_set_cpus_allowed() will not do the right thing.
5237 * And since this is boot we can forgo the serialization.
5239 set_cpus_allowed_common(idle, cpumask_of(cpu));
5242 * We're having a chicken and egg problem, even though we are
5243 * holding rq->lock, the cpu isn't yet set to this cpu so the
5244 * lockdep check in task_group() will fail.
5246 * Similar case to sched_fork(). / Alternatively we could
5247 * use task_rq_lock() here and obtain the other rq->lock.
5252 __set_task_cpu(idle, cpu);
5255 rq->curr = rq->idle = idle;
5256 idle->on_rq = TASK_ON_RQ_QUEUED;
5260 raw_spin_unlock(&rq->lock);
5261 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5263 /* Set the preempt count _outside_ the spinlocks! */
5264 init_idle_preempt_count(idle, cpu);
5265 #ifdef CONFIG_HAVE_PREEMPT_LAZY
5266 task_thread_info(idle)->preempt_lazy_count = 0;
5269 * The idle tasks have their own, simple scheduling class:
5271 idle->sched_class = &idle_sched_class;
5272 ftrace_graph_init_idle_task(idle, cpu);
5273 vtime_init_idle(idle, cpu);
5275 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5279 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5280 const struct cpumask *trial)
5282 int ret = 1, trial_cpus;
5283 struct dl_bw *cur_dl_b;
5284 unsigned long flags;
5286 if (!cpumask_weight(cur))
5289 rcu_read_lock_sched();
5290 cur_dl_b = dl_bw_of(cpumask_any(cur));
5291 trial_cpus = cpumask_weight(trial);
5293 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5294 if (cur_dl_b->bw != -1 &&
5295 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5297 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5298 rcu_read_unlock_sched();
5303 int task_can_attach(struct task_struct *p,
5304 const struct cpumask *cs_cpus_allowed)
5309 * Kthreads which disallow setaffinity shouldn't be moved
5310 * to a new cpuset; we don't want to change their cpu
5311 * affinity and isolating such threads by their set of
5312 * allowed nodes is unnecessary. Thus, cpusets are not
5313 * applicable for such threads. This prevents checking for
5314 * success of set_cpus_allowed_ptr() on all attached tasks
5315 * before cpus_allowed may be changed.
5317 if (p->flags & PF_NO_SETAFFINITY) {
5323 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5325 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5330 unsigned long flags;
5332 rcu_read_lock_sched();
5333 dl_b = dl_bw_of(dest_cpu);
5334 raw_spin_lock_irqsave(&dl_b->lock, flags);
5335 cpus = dl_bw_cpus(dest_cpu);
5336 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5341 * We reserve space for this task in the destination
5342 * root_domain, as we can't fail after this point.
5343 * We will free resources in the source root_domain
5344 * later on (see set_cpus_allowed_dl()).
5346 __dl_add(dl_b, p->dl.dl_bw);
5348 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5349 rcu_read_unlock_sched();
5359 #ifdef CONFIG_NUMA_BALANCING
5360 /* Migrate current task p to target_cpu */
5361 int migrate_task_to(struct task_struct *p, int target_cpu)
5363 struct migration_arg arg = { p, target_cpu };
5364 int curr_cpu = task_cpu(p);
5366 if (curr_cpu == target_cpu)
5369 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5372 /* TODO: This is not properly updating schedstats */
5374 trace_sched_move_numa(p, curr_cpu, target_cpu);
5375 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5379 * Requeue a task on a given node and accurately track the number of NUMA
5380 * tasks on the runqueues
5382 void sched_setnuma(struct task_struct *p, int nid)
5385 unsigned long flags;
5386 bool queued, running;
5388 rq = task_rq_lock(p, &flags);
5389 queued = task_on_rq_queued(p);
5390 running = task_current(rq, p);
5393 dequeue_task(rq, p, DEQUEUE_SAVE);
5395 put_prev_task(rq, p);
5397 p->numa_preferred_nid = nid;
5400 p->sched_class->set_curr_task(rq);
5402 enqueue_task(rq, p, ENQUEUE_RESTORE);
5403 task_rq_unlock(rq, p, &flags);
5405 #endif /* CONFIG_NUMA_BALANCING */
5407 #ifdef CONFIG_HOTPLUG_CPU
5408 static DEFINE_PER_CPU(struct mm_struct *, idle_last_mm);
5411 * Ensures that the idle task is using init_mm right before its cpu goes
5414 void idle_task_exit(void)
5416 struct mm_struct *mm = current->active_mm;
5418 BUG_ON(cpu_online(smp_processor_id()));
5420 if (mm != &init_mm) {
5421 switch_mm(mm, &init_mm, current);
5422 finish_arch_post_lock_switch();
5425 * Defer the cleanup to an alive cpu. On RT we can neither
5426 * call mmdrop() nor mmdrop_delayed() from here.
5428 per_cpu(idle_last_mm, smp_processor_id()) = mm;
5432 * Since this CPU is going 'away' for a while, fold any nr_active delta
5433 * we might have. Assumes we're called after migrate_tasks() so that the
5434 * nr_active count is stable.
5436 * Also see the comment "Global load-average calculations".
5438 static void calc_load_migrate(struct rq *rq)
5440 long delta = calc_load_fold_active(rq);
5442 atomic_long_add(delta, &calc_load_tasks);
5445 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5449 static const struct sched_class fake_sched_class = {
5450 .put_prev_task = put_prev_task_fake,
5453 static struct task_struct fake_task = {
5455 * Avoid pull_{rt,dl}_task()
5457 .prio = MAX_PRIO + 1,
5458 .sched_class = &fake_sched_class,
5462 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5463 * try_to_wake_up()->select_task_rq().
5465 * Called with rq->lock held even though we'er in stop_machine() and
5466 * there's no concurrency possible, we hold the required locks anyway
5467 * because of lock validation efforts.
5469 static void migrate_tasks(struct rq *dead_rq)
5471 struct rq *rq = dead_rq;
5472 struct task_struct *next, *stop = rq->stop;
5476 * Fudge the rq selection such that the below task selection loop
5477 * doesn't get stuck on the currently eligible stop task.
5479 * We're currently inside stop_machine() and the rq is either stuck
5480 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5481 * either way we should never end up calling schedule() until we're
5487 * put_prev_task() and pick_next_task() sched
5488 * class method both need to have an up-to-date
5489 * value of rq->clock[_task]
5491 update_rq_clock(rq);
5495 * There's this thread running, bail when that's the only
5498 if (rq->nr_running == 1)
5502 * pick_next_task assumes pinned rq->lock.
5504 lockdep_pin_lock(&rq->lock);
5505 next = pick_next_task(rq, &fake_task);
5507 next->sched_class->put_prev_task(rq, next);
5510 * Rules for changing task_struct::cpus_allowed are holding
5511 * both pi_lock and rq->lock, such that holding either
5512 * stabilizes the mask.
5514 * Drop rq->lock is not quite as disastrous as it usually is
5515 * because !cpu_active at this point, which means load-balance
5516 * will not interfere. Also, stop-machine.
5518 lockdep_unpin_lock(&rq->lock);
5519 raw_spin_unlock(&rq->lock);
5520 raw_spin_lock(&next->pi_lock);
5521 raw_spin_lock(&rq->lock);
5524 * Since we're inside stop-machine, _nothing_ should have
5525 * changed the task, WARN if weird stuff happened, because in
5526 * that case the above rq->lock drop is a fail too.
5528 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5529 raw_spin_unlock(&next->pi_lock);
5533 /* Find suitable destination for @next, with force if needed. */
5534 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5536 rq = __migrate_task(rq, next, dest_cpu);
5537 if (rq != dead_rq) {
5538 raw_spin_unlock(&rq->lock);
5540 raw_spin_lock(&rq->lock);
5542 raw_spin_unlock(&next->pi_lock);
5547 #endif /* CONFIG_HOTPLUG_CPU */
5549 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5551 static struct ctl_table sd_ctl_dir[] = {
5553 .procname = "sched_domain",
5559 static struct ctl_table sd_ctl_root[] = {
5561 .procname = "kernel",
5563 .child = sd_ctl_dir,
5568 static struct ctl_table *sd_alloc_ctl_entry(int n)
5570 struct ctl_table *entry =
5571 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5576 static void sd_free_ctl_entry(struct ctl_table **tablep)
5578 struct ctl_table *entry;
5581 * In the intermediate directories, both the child directory and
5582 * procname are dynamically allocated and could fail but the mode
5583 * will always be set. In the lowest directory the names are
5584 * static strings and all have proc handlers.
5586 for (entry = *tablep; entry->mode; entry++) {
5588 sd_free_ctl_entry(&entry->child);
5589 if (entry->proc_handler == NULL)
5590 kfree(entry->procname);
5597 static int min_load_idx = 0;
5598 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5601 set_table_entry(struct ctl_table *entry,
5602 const char *procname, void *data, int maxlen,
5603 umode_t mode, proc_handler *proc_handler,
5606 entry->procname = procname;
5608 entry->maxlen = maxlen;
5610 entry->proc_handler = proc_handler;
5613 entry->extra1 = &min_load_idx;
5614 entry->extra2 = &max_load_idx;
5618 static struct ctl_table *
5619 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5621 struct ctl_table *table = sd_alloc_ctl_entry(14);
5626 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5627 sizeof(long), 0644, proc_doulongvec_minmax, false);
5628 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5629 sizeof(long), 0644, proc_doulongvec_minmax, false);
5630 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5631 sizeof(int), 0644, proc_dointvec_minmax, true);
5632 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5633 sizeof(int), 0644, proc_dointvec_minmax, true);
5634 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5635 sizeof(int), 0644, proc_dointvec_minmax, true);
5636 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5637 sizeof(int), 0644, proc_dointvec_minmax, true);
5638 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5639 sizeof(int), 0644, proc_dointvec_minmax, true);
5640 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5641 sizeof(int), 0644, proc_dointvec_minmax, false);
5642 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5643 sizeof(int), 0644, proc_dointvec_minmax, false);
5644 set_table_entry(&table[9], "cache_nice_tries",
5645 &sd->cache_nice_tries,
5646 sizeof(int), 0644, proc_dointvec_minmax, false);
5647 set_table_entry(&table[10], "flags", &sd->flags,
5648 sizeof(int), 0644, proc_dointvec_minmax, false);
5649 set_table_entry(&table[11], "max_newidle_lb_cost",
5650 &sd->max_newidle_lb_cost,
5651 sizeof(long), 0644, proc_doulongvec_minmax, false);
5652 set_table_entry(&table[12], "name", sd->name,
5653 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5654 /* &table[13] is terminator */
5659 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5661 struct ctl_table *entry, *table;
5662 struct sched_domain *sd;
5663 int domain_num = 0, i;
5666 for_each_domain(cpu, sd)
5668 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5673 for_each_domain(cpu, sd) {
5674 snprintf(buf, 32, "domain%d", i);
5675 entry->procname = kstrdup(buf, GFP_KERNEL);
5677 entry->child = sd_alloc_ctl_domain_table(sd);
5684 static struct ctl_table_header *sd_sysctl_header;
5685 static void register_sched_domain_sysctl(void)
5687 int i, cpu_num = num_possible_cpus();
5688 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5691 WARN_ON(sd_ctl_dir[0].child);
5692 sd_ctl_dir[0].child = entry;
5697 for_each_possible_cpu(i) {
5698 snprintf(buf, 32, "cpu%d", i);
5699 entry->procname = kstrdup(buf, GFP_KERNEL);
5701 entry->child = sd_alloc_ctl_cpu_table(i);
5705 WARN_ON(sd_sysctl_header);
5706 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5709 /* may be called multiple times per register */
5710 static void unregister_sched_domain_sysctl(void)
5712 unregister_sysctl_table(sd_sysctl_header);
5713 sd_sysctl_header = NULL;
5714 if (sd_ctl_dir[0].child)
5715 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5718 static void register_sched_domain_sysctl(void)
5721 static void unregister_sched_domain_sysctl(void)
5724 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5726 static void set_rq_online(struct rq *rq)
5729 const struct sched_class *class;
5731 cpumask_set_cpu(rq->cpu, rq->rd->online);
5734 for_each_class(class) {
5735 if (class->rq_online)
5736 class->rq_online(rq);
5741 static void set_rq_offline(struct rq *rq)
5744 const struct sched_class *class;
5746 for_each_class(class) {
5747 if (class->rq_offline)
5748 class->rq_offline(rq);
5751 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5757 * migration_call - callback that gets triggered when a CPU is added.
5758 * Here we can start up the necessary migration thread for the new CPU.
5761 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5763 int cpu = (long)hcpu;
5764 unsigned long flags;
5765 struct rq *rq = cpu_rq(cpu);
5767 switch (action & ~CPU_TASKS_FROZEN) {
5769 case CPU_UP_PREPARE:
5770 rq->calc_load_update = calc_load_update;
5774 /* Update our root-domain */
5775 raw_spin_lock_irqsave(&rq->lock, flags);
5777 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5781 raw_spin_unlock_irqrestore(&rq->lock, flags);
5784 #ifdef CONFIG_HOTPLUG_CPU
5786 sched_ttwu_pending();
5787 /* Update our root-domain */
5788 raw_spin_lock_irqsave(&rq->lock, flags);
5790 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5794 BUG_ON(rq->nr_running != 1); /* the migration thread */
5795 raw_spin_unlock_irqrestore(&rq->lock, flags);
5799 calc_load_migrate(rq);
5800 if (per_cpu(idle_last_mm, cpu)) {
5801 mmdrop(per_cpu(idle_last_mm, cpu));
5802 per_cpu(idle_last_mm, cpu) = NULL;
5808 update_max_interval();
5814 * Register at high priority so that task migration (migrate_all_tasks)
5815 * happens before everything else. This has to be lower priority than
5816 * the notifier in the perf_event subsystem, though.
5818 static struct notifier_block migration_notifier = {
5819 .notifier_call = migration_call,
5820 .priority = CPU_PRI_MIGRATION,
5823 static void set_cpu_rq_start_time(void)
5825 int cpu = smp_processor_id();
5826 struct rq *rq = cpu_rq(cpu);
5827 rq->age_stamp = sched_clock_cpu(cpu);
5830 static int sched_cpu_active(struct notifier_block *nfb,
5831 unsigned long action, void *hcpu)
5833 int cpu = (long)hcpu;
5835 switch (action & ~CPU_TASKS_FROZEN) {
5837 set_cpu_rq_start_time();
5842 * At this point a starting CPU has marked itself as online via
5843 * set_cpu_online(). But it might not yet have marked itself
5844 * as active, which is essential from here on.
5846 set_cpu_active(cpu, true);
5847 stop_machine_unpark(cpu);
5850 case CPU_DOWN_FAILED:
5851 set_cpu_active(cpu, true);
5859 static int sched_cpu_inactive(struct notifier_block *nfb,
5860 unsigned long action, void *hcpu)
5862 switch (action & ~CPU_TASKS_FROZEN) {
5863 case CPU_DOWN_PREPARE:
5864 set_cpu_active((long)hcpu, false);
5871 static int __init migration_init(void)
5873 void *cpu = (void *)(long)smp_processor_id();
5876 /* Initialize migration for the boot CPU */
5877 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5878 BUG_ON(err == NOTIFY_BAD);
5879 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5880 register_cpu_notifier(&migration_notifier);
5882 /* Register cpu active notifiers */
5883 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5884 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5888 early_initcall(migration_init);
5890 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5892 #ifdef CONFIG_SCHED_DEBUG
5894 static __read_mostly int sched_debug_enabled;
5896 static int __init sched_debug_setup(char *str)
5898 sched_debug_enabled = 1;
5902 early_param("sched_debug", sched_debug_setup);
5904 static inline bool sched_debug(void)
5906 return sched_debug_enabled;
5909 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5910 struct cpumask *groupmask)
5912 struct sched_group *group = sd->groups;
5914 cpumask_clear(groupmask);
5916 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5918 if (!(sd->flags & SD_LOAD_BALANCE)) {
5919 printk("does not load-balance\n");
5921 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5926 printk(KERN_CONT "span %*pbl level %s\n",
5927 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5929 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5930 printk(KERN_ERR "ERROR: domain->span does not contain "
5933 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5934 printk(KERN_ERR "ERROR: domain->groups does not contain"
5938 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5942 printk(KERN_ERR "ERROR: group is NULL\n");
5946 if (!cpumask_weight(sched_group_cpus(group))) {
5947 printk(KERN_CONT "\n");
5948 printk(KERN_ERR "ERROR: empty group\n");
5952 if (!(sd->flags & SD_OVERLAP) &&
5953 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5954 printk(KERN_CONT "\n");
5955 printk(KERN_ERR "ERROR: repeated CPUs\n");
5959 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5961 printk(KERN_CONT " %*pbl",
5962 cpumask_pr_args(sched_group_cpus(group)));
5963 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5964 printk(KERN_CONT " (cpu_capacity = %d)",
5965 group->sgc->capacity);
5968 group = group->next;
5969 } while (group != sd->groups);
5970 printk(KERN_CONT "\n");
5972 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5973 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5976 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5977 printk(KERN_ERR "ERROR: parent span is not a superset "
5978 "of domain->span\n");
5982 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5986 if (!sched_debug_enabled)
5990 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5994 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5997 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6005 #else /* !CONFIG_SCHED_DEBUG */
6006 # define sched_domain_debug(sd, cpu) do { } while (0)
6007 static inline bool sched_debug(void)
6011 #endif /* CONFIG_SCHED_DEBUG */
6013 static int sd_degenerate(struct sched_domain *sd)
6015 if (cpumask_weight(sched_domain_span(sd)) == 1)
6018 /* Following flags need at least 2 groups */
6019 if (sd->flags & (SD_LOAD_BALANCE |
6020 SD_BALANCE_NEWIDLE |
6023 SD_SHARE_CPUCAPACITY |
6024 SD_SHARE_PKG_RESOURCES |
6025 SD_SHARE_POWERDOMAIN)) {
6026 if (sd->groups != sd->groups->next)
6030 /* Following flags don't use groups */
6031 if (sd->flags & (SD_WAKE_AFFINE))
6038 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6040 unsigned long cflags = sd->flags, pflags = parent->flags;
6042 if (sd_degenerate(parent))
6045 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6048 /* Flags needing groups don't count if only 1 group in parent */
6049 if (parent->groups == parent->groups->next) {
6050 pflags &= ~(SD_LOAD_BALANCE |
6051 SD_BALANCE_NEWIDLE |
6054 SD_SHARE_CPUCAPACITY |
6055 SD_SHARE_PKG_RESOURCES |
6057 SD_SHARE_POWERDOMAIN);
6058 if (nr_node_ids == 1)
6059 pflags &= ~SD_SERIALIZE;
6061 if (~cflags & pflags)
6067 static void free_rootdomain(struct rcu_head *rcu)
6069 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6071 cpupri_cleanup(&rd->cpupri);
6072 cpudl_cleanup(&rd->cpudl);
6073 free_cpumask_var(rd->dlo_mask);
6074 free_cpumask_var(rd->rto_mask);
6075 free_cpumask_var(rd->online);
6076 free_cpumask_var(rd->span);
6080 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6082 struct root_domain *old_rd = NULL;
6083 unsigned long flags;
6085 raw_spin_lock_irqsave(&rq->lock, flags);
6090 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6093 cpumask_clear_cpu(rq->cpu, old_rd->span);
6096 * If we dont want to free the old_rd yet then
6097 * set old_rd to NULL to skip the freeing later
6100 if (!atomic_dec_and_test(&old_rd->refcount))
6104 atomic_inc(&rd->refcount);
6107 cpumask_set_cpu(rq->cpu, rd->span);
6108 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6111 raw_spin_unlock_irqrestore(&rq->lock, flags);
6114 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6117 static int init_rootdomain(struct root_domain *rd)
6119 memset(rd, 0, sizeof(*rd));
6121 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6123 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6125 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6127 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6130 init_dl_bw(&rd->dl_bw);
6131 if (cpudl_init(&rd->cpudl) != 0)
6134 if (cpupri_init(&rd->cpupri) != 0)
6139 free_cpumask_var(rd->rto_mask);
6141 free_cpumask_var(rd->dlo_mask);
6143 free_cpumask_var(rd->online);
6145 free_cpumask_var(rd->span);
6151 * By default the system creates a single root-domain with all cpus as
6152 * members (mimicking the global state we have today).
6154 struct root_domain def_root_domain;
6156 static void init_defrootdomain(void)
6158 init_rootdomain(&def_root_domain);
6160 atomic_set(&def_root_domain.refcount, 1);
6163 static struct root_domain *alloc_rootdomain(void)
6165 struct root_domain *rd;
6167 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6171 if (init_rootdomain(rd) != 0) {
6179 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6181 struct sched_group *tmp, *first;
6190 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6195 } while (sg != first);
6198 static void free_sched_domain(struct rcu_head *rcu)
6200 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6203 * If its an overlapping domain it has private groups, iterate and
6206 if (sd->flags & SD_OVERLAP) {
6207 free_sched_groups(sd->groups, 1);
6208 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6209 kfree(sd->groups->sgc);
6215 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6217 call_rcu(&sd->rcu, free_sched_domain);
6220 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6222 for (; sd; sd = sd->parent)
6223 destroy_sched_domain(sd, cpu);
6227 * Keep a special pointer to the highest sched_domain that has
6228 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6229 * allows us to avoid some pointer chasing select_idle_sibling().
6231 * Also keep a unique ID per domain (we use the first cpu number in
6232 * the cpumask of the domain), this allows us to quickly tell if
6233 * two cpus are in the same cache domain, see cpus_share_cache().
6235 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6236 DEFINE_PER_CPU(int, sd_llc_size);
6237 DEFINE_PER_CPU(int, sd_llc_id);
6238 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6239 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6240 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6242 static void update_top_cache_domain(int cpu)
6244 struct sched_domain *sd;
6245 struct sched_domain *busy_sd = NULL;
6249 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6251 id = cpumask_first(sched_domain_span(sd));
6252 size = cpumask_weight(sched_domain_span(sd));
6253 busy_sd = sd->parent; /* sd_busy */
6255 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6257 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6258 per_cpu(sd_llc_size, cpu) = size;
6259 per_cpu(sd_llc_id, cpu) = id;
6261 sd = lowest_flag_domain(cpu, SD_NUMA);
6262 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6264 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6265 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6269 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6270 * hold the hotplug lock.
6273 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6275 struct rq *rq = cpu_rq(cpu);
6276 struct sched_domain *tmp;
6278 /* Remove the sched domains which do not contribute to scheduling. */
6279 for (tmp = sd; tmp; ) {
6280 struct sched_domain *parent = tmp->parent;
6284 if (sd_parent_degenerate(tmp, parent)) {
6285 tmp->parent = parent->parent;
6287 parent->parent->child = tmp;
6289 * Transfer SD_PREFER_SIBLING down in case of a
6290 * degenerate parent; the spans match for this
6291 * so the property transfers.
6293 if (parent->flags & SD_PREFER_SIBLING)
6294 tmp->flags |= SD_PREFER_SIBLING;
6295 destroy_sched_domain(parent, cpu);
6300 if (sd && sd_degenerate(sd)) {
6303 destroy_sched_domain(tmp, cpu);
6308 sched_domain_debug(sd, cpu);
6310 rq_attach_root(rq, rd);
6312 rcu_assign_pointer(rq->sd, sd);
6313 destroy_sched_domains(tmp, cpu);
6315 update_top_cache_domain(cpu);
6318 /* Setup the mask of cpus configured for isolated domains */
6319 static int __init isolated_cpu_setup(char *str)
6321 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6322 cpulist_parse(str, cpu_isolated_map);
6326 __setup("isolcpus=", isolated_cpu_setup);
6329 struct sched_domain ** __percpu sd;
6330 struct root_domain *rd;
6341 * Build an iteration mask that can exclude certain CPUs from the upwards
6344 * Asymmetric node setups can result in situations where the domain tree is of
6345 * unequal depth, make sure to skip domains that already cover the entire
6348 * In that case build_sched_domains() will have terminated the iteration early
6349 * and our sibling sd spans will be empty. Domains should always include the
6350 * cpu they're built on, so check that.
6353 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6355 const struct cpumask *span = sched_domain_span(sd);
6356 struct sd_data *sdd = sd->private;
6357 struct sched_domain *sibling;
6360 for_each_cpu(i, span) {
6361 sibling = *per_cpu_ptr(sdd->sd, i);
6362 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6365 cpumask_set_cpu(i, sched_group_mask(sg));
6370 * Return the canonical balance cpu for this group, this is the first cpu
6371 * of this group that's also in the iteration mask.
6373 int group_balance_cpu(struct sched_group *sg)
6375 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6379 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6381 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6382 const struct cpumask *span = sched_domain_span(sd);
6383 struct cpumask *covered = sched_domains_tmpmask;
6384 struct sd_data *sdd = sd->private;
6385 struct sched_domain *sibling;
6388 cpumask_clear(covered);
6390 for_each_cpu(i, span) {
6391 struct cpumask *sg_span;
6393 if (cpumask_test_cpu(i, covered))
6396 sibling = *per_cpu_ptr(sdd->sd, i);
6398 /* See the comment near build_group_mask(). */
6399 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6402 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6403 GFP_KERNEL, cpu_to_node(cpu));
6408 sg_span = sched_group_cpus(sg);
6410 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6412 cpumask_set_cpu(i, sg_span);
6414 cpumask_or(covered, covered, sg_span);
6416 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6417 if (atomic_inc_return(&sg->sgc->ref) == 1)
6418 build_group_mask(sd, sg);
6421 * Initialize sgc->capacity such that even if we mess up the
6422 * domains and no possible iteration will get us here, we won't
6425 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6428 * Make sure the first group of this domain contains the
6429 * canonical balance cpu. Otherwise the sched_domain iteration
6430 * breaks. See update_sg_lb_stats().
6432 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6433 group_balance_cpu(sg) == cpu)
6443 sd->groups = groups;
6448 free_sched_groups(first, 0);
6453 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6455 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6456 struct sched_domain *child = sd->child;
6459 cpu = cpumask_first(sched_domain_span(child));
6462 *sg = *per_cpu_ptr(sdd->sg, cpu);
6463 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6464 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6471 * build_sched_groups will build a circular linked list of the groups
6472 * covered by the given span, and will set each group's ->cpumask correctly,
6473 * and ->cpu_capacity to 0.
6475 * Assumes the sched_domain tree is fully constructed
6478 build_sched_groups(struct sched_domain *sd, int cpu)
6480 struct sched_group *first = NULL, *last = NULL;
6481 struct sd_data *sdd = sd->private;
6482 const struct cpumask *span = sched_domain_span(sd);
6483 struct cpumask *covered;
6486 get_group(cpu, sdd, &sd->groups);
6487 atomic_inc(&sd->groups->ref);
6489 if (cpu != cpumask_first(span))
6492 lockdep_assert_held(&sched_domains_mutex);
6493 covered = sched_domains_tmpmask;
6495 cpumask_clear(covered);
6497 for_each_cpu(i, span) {
6498 struct sched_group *sg;
6501 if (cpumask_test_cpu(i, covered))
6504 group = get_group(i, sdd, &sg);
6505 cpumask_setall(sched_group_mask(sg));
6507 for_each_cpu(j, span) {
6508 if (get_group(j, sdd, NULL) != group)
6511 cpumask_set_cpu(j, covered);
6512 cpumask_set_cpu(j, sched_group_cpus(sg));
6527 * Initialize sched groups cpu_capacity.
6529 * cpu_capacity indicates the capacity of sched group, which is used while
6530 * distributing the load between different sched groups in a sched domain.
6531 * Typically cpu_capacity for all the groups in a sched domain will be same
6532 * unless there are asymmetries in the topology. If there are asymmetries,
6533 * group having more cpu_capacity will pickup more load compared to the
6534 * group having less cpu_capacity.
6536 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6538 struct sched_group *sg = sd->groups;
6543 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6545 } while (sg != sd->groups);
6547 if (cpu != group_balance_cpu(sg))
6550 update_group_capacity(sd, cpu);
6551 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6555 * Initializers for schedule domains
6556 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6559 static int default_relax_domain_level = -1;
6560 int sched_domain_level_max;
6562 static int __init setup_relax_domain_level(char *str)
6564 if (kstrtoint(str, 0, &default_relax_domain_level))
6565 pr_warn("Unable to set relax_domain_level\n");
6569 __setup("relax_domain_level=", setup_relax_domain_level);
6571 static void set_domain_attribute(struct sched_domain *sd,
6572 struct sched_domain_attr *attr)
6576 if (!attr || attr->relax_domain_level < 0) {
6577 if (default_relax_domain_level < 0)
6580 request = default_relax_domain_level;
6582 request = attr->relax_domain_level;
6583 if (request < sd->level) {
6584 /* turn off idle balance on this domain */
6585 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6587 /* turn on idle balance on this domain */
6588 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6592 static void __sdt_free(const struct cpumask *cpu_map);
6593 static int __sdt_alloc(const struct cpumask *cpu_map);
6595 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6596 const struct cpumask *cpu_map)
6600 if (!atomic_read(&d->rd->refcount))
6601 free_rootdomain(&d->rd->rcu); /* fall through */
6603 free_percpu(d->sd); /* fall through */
6605 __sdt_free(cpu_map); /* fall through */
6611 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6612 const struct cpumask *cpu_map)
6614 memset(d, 0, sizeof(*d));
6616 if (__sdt_alloc(cpu_map))
6617 return sa_sd_storage;
6618 d->sd = alloc_percpu(struct sched_domain *);
6620 return sa_sd_storage;
6621 d->rd = alloc_rootdomain();
6624 return sa_rootdomain;
6628 * NULL the sd_data elements we've used to build the sched_domain and
6629 * sched_group structure so that the subsequent __free_domain_allocs()
6630 * will not free the data we're using.
6632 static void claim_allocations(int cpu, struct sched_domain *sd)
6634 struct sd_data *sdd = sd->private;
6636 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6637 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6639 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6640 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6642 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6643 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6647 static int sched_domains_numa_levels;
6648 enum numa_topology_type sched_numa_topology_type;
6649 static int *sched_domains_numa_distance;
6650 int sched_max_numa_distance;
6651 static struct cpumask ***sched_domains_numa_masks;
6652 static int sched_domains_curr_level;
6656 * SD_flags allowed in topology descriptions.
6658 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6659 * SD_SHARE_PKG_RESOURCES - describes shared caches
6660 * SD_NUMA - describes NUMA topologies
6661 * SD_SHARE_POWERDOMAIN - describes shared power domain
6664 * SD_ASYM_PACKING - describes SMT quirks
6666 #define TOPOLOGY_SD_FLAGS \
6667 (SD_SHARE_CPUCAPACITY | \
6668 SD_SHARE_PKG_RESOURCES | \
6671 SD_SHARE_POWERDOMAIN)
6673 static struct sched_domain *
6674 sd_init(struct sched_domain_topology_level *tl, int cpu)
6676 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6677 int sd_weight, sd_flags = 0;
6681 * Ugly hack to pass state to sd_numa_mask()...
6683 sched_domains_curr_level = tl->numa_level;
6686 sd_weight = cpumask_weight(tl->mask(cpu));
6689 sd_flags = (*tl->sd_flags)();
6690 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6691 "wrong sd_flags in topology description\n"))
6692 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6694 *sd = (struct sched_domain){
6695 .min_interval = sd_weight,
6696 .max_interval = 2*sd_weight,
6698 .imbalance_pct = 125,
6700 .cache_nice_tries = 0,
6707 .flags = 1*SD_LOAD_BALANCE
6708 | 1*SD_BALANCE_NEWIDLE
6713 | 0*SD_SHARE_CPUCAPACITY
6714 | 0*SD_SHARE_PKG_RESOURCES
6716 | 0*SD_PREFER_SIBLING
6721 .last_balance = jiffies,
6722 .balance_interval = sd_weight,
6724 .max_newidle_lb_cost = 0,
6725 .next_decay_max_lb_cost = jiffies,
6726 #ifdef CONFIG_SCHED_DEBUG
6732 * Convert topological properties into behaviour.
6735 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6736 sd->flags |= SD_PREFER_SIBLING;
6737 sd->imbalance_pct = 110;
6738 sd->smt_gain = 1178; /* ~15% */
6740 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6741 sd->imbalance_pct = 117;
6742 sd->cache_nice_tries = 1;
6746 } else if (sd->flags & SD_NUMA) {
6747 sd->cache_nice_tries = 2;
6751 sd->flags |= SD_SERIALIZE;
6752 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6753 sd->flags &= ~(SD_BALANCE_EXEC |
6760 sd->flags |= SD_PREFER_SIBLING;
6761 sd->cache_nice_tries = 1;
6766 sd->private = &tl->data;
6772 * Topology list, bottom-up.
6774 static struct sched_domain_topology_level default_topology[] = {
6775 #ifdef CONFIG_SCHED_SMT
6776 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6778 #ifdef CONFIG_SCHED_MC
6779 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6781 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6785 static struct sched_domain_topology_level *sched_domain_topology =
6788 #define for_each_sd_topology(tl) \
6789 for (tl = sched_domain_topology; tl->mask; tl++)
6791 void set_sched_topology(struct sched_domain_topology_level *tl)
6793 sched_domain_topology = tl;
6798 static const struct cpumask *sd_numa_mask(int cpu)
6800 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6803 static void sched_numa_warn(const char *str)
6805 static int done = false;
6813 printk(KERN_WARNING "ERROR: %s\n\n", str);
6815 for (i = 0; i < nr_node_ids; i++) {
6816 printk(KERN_WARNING " ");
6817 for (j = 0; j < nr_node_ids; j++)
6818 printk(KERN_CONT "%02d ", node_distance(i,j));
6819 printk(KERN_CONT "\n");
6821 printk(KERN_WARNING "\n");
6824 bool find_numa_distance(int distance)
6828 if (distance == node_distance(0, 0))
6831 for (i = 0; i < sched_domains_numa_levels; i++) {
6832 if (sched_domains_numa_distance[i] == distance)
6840 * A system can have three types of NUMA topology:
6841 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6842 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6843 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6845 * The difference between a glueless mesh topology and a backplane
6846 * topology lies in whether communication between not directly
6847 * connected nodes goes through intermediary nodes (where programs
6848 * could run), or through backplane controllers. This affects
6849 * placement of programs.
6851 * The type of topology can be discerned with the following tests:
6852 * - If the maximum distance between any nodes is 1 hop, the system
6853 * is directly connected.
6854 * - If for two nodes A and B, located N > 1 hops away from each other,
6855 * there is an intermediary node C, which is < N hops away from both
6856 * nodes A and B, the system is a glueless mesh.
6858 static void init_numa_topology_type(void)
6862 n = sched_max_numa_distance;
6864 if (sched_domains_numa_levels <= 1) {
6865 sched_numa_topology_type = NUMA_DIRECT;
6869 for_each_online_node(a) {
6870 for_each_online_node(b) {
6871 /* Find two nodes furthest removed from each other. */
6872 if (node_distance(a, b) < n)
6875 /* Is there an intermediary node between a and b? */
6876 for_each_online_node(c) {
6877 if (node_distance(a, c) < n &&
6878 node_distance(b, c) < n) {
6879 sched_numa_topology_type =
6885 sched_numa_topology_type = NUMA_BACKPLANE;
6891 static void sched_init_numa(void)
6893 int next_distance, curr_distance = node_distance(0, 0);
6894 struct sched_domain_topology_level *tl;
6898 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6899 if (!sched_domains_numa_distance)
6903 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6904 * unique distances in the node_distance() table.
6906 * Assumes node_distance(0,j) includes all distances in
6907 * node_distance(i,j) in order to avoid cubic time.
6909 next_distance = curr_distance;
6910 for (i = 0; i < nr_node_ids; i++) {
6911 for (j = 0; j < nr_node_ids; j++) {
6912 for (k = 0; k < nr_node_ids; k++) {
6913 int distance = node_distance(i, k);
6915 if (distance > curr_distance &&
6916 (distance < next_distance ||
6917 next_distance == curr_distance))
6918 next_distance = distance;
6921 * While not a strong assumption it would be nice to know
6922 * about cases where if node A is connected to B, B is not
6923 * equally connected to A.
6925 if (sched_debug() && node_distance(k, i) != distance)
6926 sched_numa_warn("Node-distance not symmetric");
6928 if (sched_debug() && i && !find_numa_distance(distance))
6929 sched_numa_warn("Node-0 not representative");
6931 if (next_distance != curr_distance) {
6932 sched_domains_numa_distance[level++] = next_distance;
6933 sched_domains_numa_levels = level;
6934 curr_distance = next_distance;
6939 * In case of sched_debug() we verify the above assumption.
6949 * 'level' contains the number of unique distances, excluding the
6950 * identity distance node_distance(i,i).
6952 * The sched_domains_numa_distance[] array includes the actual distance
6957 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6958 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6959 * the array will contain less then 'level' members. This could be
6960 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6961 * in other functions.
6963 * We reset it to 'level' at the end of this function.
6965 sched_domains_numa_levels = 0;
6967 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6968 if (!sched_domains_numa_masks)
6972 * Now for each level, construct a mask per node which contains all
6973 * cpus of nodes that are that many hops away from us.
6975 for (i = 0; i < level; i++) {
6976 sched_domains_numa_masks[i] =
6977 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6978 if (!sched_domains_numa_masks[i])
6981 for (j = 0; j < nr_node_ids; j++) {
6982 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6986 sched_domains_numa_masks[i][j] = mask;
6989 if (node_distance(j, k) > sched_domains_numa_distance[i])
6992 cpumask_or(mask, mask, cpumask_of_node(k));
6997 /* Compute default topology size */
6998 for (i = 0; sched_domain_topology[i].mask; i++);
7000 tl = kzalloc((i + level + 1) *
7001 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7006 * Copy the default topology bits..
7008 for (i = 0; sched_domain_topology[i].mask; i++)
7009 tl[i] = sched_domain_topology[i];
7012 * .. and append 'j' levels of NUMA goodness.
7014 for (j = 0; j < level; i++, j++) {
7015 tl[i] = (struct sched_domain_topology_level){
7016 .mask = sd_numa_mask,
7017 .sd_flags = cpu_numa_flags,
7018 .flags = SDTL_OVERLAP,
7024 sched_domain_topology = tl;
7026 sched_domains_numa_levels = level;
7027 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7029 init_numa_topology_type();
7032 static void sched_domains_numa_masks_set(int cpu)
7035 int node = cpu_to_node(cpu);
7037 for (i = 0; i < sched_domains_numa_levels; i++) {
7038 for (j = 0; j < nr_node_ids; j++) {
7039 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7040 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7045 static void sched_domains_numa_masks_clear(int cpu)
7048 for (i = 0; i < sched_domains_numa_levels; i++) {
7049 for (j = 0; j < nr_node_ids; j++)
7050 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7055 * Update sched_domains_numa_masks[level][node] array when new cpus
7058 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7059 unsigned long action,
7062 int cpu = (long)hcpu;
7064 switch (action & ~CPU_TASKS_FROZEN) {
7066 sched_domains_numa_masks_set(cpu);
7070 sched_domains_numa_masks_clear(cpu);
7080 static inline void sched_init_numa(void)
7084 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7085 unsigned long action,
7090 #endif /* CONFIG_NUMA */
7092 static int __sdt_alloc(const struct cpumask *cpu_map)
7094 struct sched_domain_topology_level *tl;
7097 for_each_sd_topology(tl) {
7098 struct sd_data *sdd = &tl->data;
7100 sdd->sd = alloc_percpu(struct sched_domain *);
7104 sdd->sg = alloc_percpu(struct sched_group *);
7108 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7112 for_each_cpu(j, cpu_map) {
7113 struct sched_domain *sd;
7114 struct sched_group *sg;
7115 struct sched_group_capacity *sgc;
7117 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7118 GFP_KERNEL, cpu_to_node(j));
7122 *per_cpu_ptr(sdd->sd, j) = sd;
7124 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7125 GFP_KERNEL, cpu_to_node(j));
7131 *per_cpu_ptr(sdd->sg, j) = sg;
7133 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7134 GFP_KERNEL, cpu_to_node(j));
7138 *per_cpu_ptr(sdd->sgc, j) = sgc;
7145 static void __sdt_free(const struct cpumask *cpu_map)
7147 struct sched_domain_topology_level *tl;
7150 for_each_sd_topology(tl) {
7151 struct sd_data *sdd = &tl->data;
7153 for_each_cpu(j, cpu_map) {
7154 struct sched_domain *sd;
7157 sd = *per_cpu_ptr(sdd->sd, j);
7158 if (sd && (sd->flags & SD_OVERLAP))
7159 free_sched_groups(sd->groups, 0);
7160 kfree(*per_cpu_ptr(sdd->sd, j));
7164 kfree(*per_cpu_ptr(sdd->sg, j));
7166 kfree(*per_cpu_ptr(sdd->sgc, j));
7168 free_percpu(sdd->sd);
7170 free_percpu(sdd->sg);
7172 free_percpu(sdd->sgc);
7177 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7178 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7179 struct sched_domain *child, int cpu)
7181 struct sched_domain *sd = sd_init(tl, cpu);
7185 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7187 sd->level = child->level + 1;
7188 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7192 if (!cpumask_subset(sched_domain_span(child),
7193 sched_domain_span(sd))) {
7194 pr_err("BUG: arch topology borken\n");
7195 #ifdef CONFIG_SCHED_DEBUG
7196 pr_err(" the %s domain not a subset of the %s domain\n",
7197 child->name, sd->name);
7199 /* Fixup, ensure @sd has at least @child cpus. */
7200 cpumask_or(sched_domain_span(sd),
7201 sched_domain_span(sd),
7202 sched_domain_span(child));
7206 set_domain_attribute(sd, attr);
7212 * Build sched domains for a given set of cpus and attach the sched domains
7213 * to the individual cpus
7215 static int build_sched_domains(const struct cpumask *cpu_map,
7216 struct sched_domain_attr *attr)
7218 enum s_alloc alloc_state;
7219 struct sched_domain *sd;
7221 int i, ret = -ENOMEM;
7223 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7224 if (alloc_state != sa_rootdomain)
7227 /* Set up domains for cpus specified by the cpu_map. */
7228 for_each_cpu(i, cpu_map) {
7229 struct sched_domain_topology_level *tl;
7232 for_each_sd_topology(tl) {
7233 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7234 if (tl == sched_domain_topology)
7235 *per_cpu_ptr(d.sd, i) = sd;
7236 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7237 sd->flags |= SD_OVERLAP;
7238 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7243 /* Build the groups for the domains */
7244 for_each_cpu(i, cpu_map) {
7245 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7246 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7247 if (sd->flags & SD_OVERLAP) {
7248 if (build_overlap_sched_groups(sd, i))
7251 if (build_sched_groups(sd, i))
7257 /* Calculate CPU capacity for physical packages and nodes */
7258 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7259 if (!cpumask_test_cpu(i, cpu_map))
7262 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7263 claim_allocations(i, sd);
7264 init_sched_groups_capacity(i, sd);
7268 /* Attach the domains */
7270 for_each_cpu(i, cpu_map) {
7271 sd = *per_cpu_ptr(d.sd, i);
7272 cpu_attach_domain(sd, d.rd, i);
7278 __free_domain_allocs(&d, alloc_state, cpu_map);
7282 static cpumask_var_t *doms_cur; /* current sched domains */
7283 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7284 static struct sched_domain_attr *dattr_cur;
7285 /* attribues of custom domains in 'doms_cur' */
7288 * Special case: If a kmalloc of a doms_cur partition (array of
7289 * cpumask) fails, then fallback to a single sched domain,
7290 * as determined by the single cpumask fallback_doms.
7292 static cpumask_var_t fallback_doms;
7295 * arch_update_cpu_topology lets virtualized architectures update the
7296 * cpu core maps. It is supposed to return 1 if the topology changed
7297 * or 0 if it stayed the same.
7299 int __weak arch_update_cpu_topology(void)
7304 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7307 cpumask_var_t *doms;
7309 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7312 for (i = 0; i < ndoms; i++) {
7313 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7314 free_sched_domains(doms, i);
7321 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7324 for (i = 0; i < ndoms; i++)
7325 free_cpumask_var(doms[i]);
7330 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7331 * For now this just excludes isolated cpus, but could be used to
7332 * exclude other special cases in the future.
7334 static int init_sched_domains(const struct cpumask *cpu_map)
7338 arch_update_cpu_topology();
7340 doms_cur = alloc_sched_domains(ndoms_cur);
7342 doms_cur = &fallback_doms;
7343 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7344 err = build_sched_domains(doms_cur[0], NULL);
7345 register_sched_domain_sysctl();
7351 * Detach sched domains from a group of cpus specified in cpu_map
7352 * These cpus will now be attached to the NULL domain
7354 static void detach_destroy_domains(const struct cpumask *cpu_map)
7359 for_each_cpu(i, cpu_map)
7360 cpu_attach_domain(NULL, &def_root_domain, i);
7364 /* handle null as "default" */
7365 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7366 struct sched_domain_attr *new, int idx_new)
7368 struct sched_domain_attr tmp;
7375 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7376 new ? (new + idx_new) : &tmp,
7377 sizeof(struct sched_domain_attr));
7381 * Partition sched domains as specified by the 'ndoms_new'
7382 * cpumasks in the array doms_new[] of cpumasks. This compares
7383 * doms_new[] to the current sched domain partitioning, doms_cur[].
7384 * It destroys each deleted domain and builds each new domain.
7386 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7387 * The masks don't intersect (don't overlap.) We should setup one
7388 * sched domain for each mask. CPUs not in any of the cpumasks will
7389 * not be load balanced. If the same cpumask appears both in the
7390 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7393 * The passed in 'doms_new' should be allocated using
7394 * alloc_sched_domains. This routine takes ownership of it and will
7395 * free_sched_domains it when done with it. If the caller failed the
7396 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7397 * and partition_sched_domains() will fallback to the single partition
7398 * 'fallback_doms', it also forces the domains to be rebuilt.
7400 * If doms_new == NULL it will be replaced with cpu_online_mask.
7401 * ndoms_new == 0 is a special case for destroying existing domains,
7402 * and it will not create the default domain.
7404 * Call with hotplug lock held
7406 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7407 struct sched_domain_attr *dattr_new)
7412 mutex_lock(&sched_domains_mutex);
7414 /* always unregister in case we don't destroy any domains */
7415 unregister_sched_domain_sysctl();
7417 /* Let architecture update cpu core mappings. */
7418 new_topology = arch_update_cpu_topology();
7420 n = doms_new ? ndoms_new : 0;
7422 /* Destroy deleted domains */
7423 for (i = 0; i < ndoms_cur; i++) {
7424 for (j = 0; j < n && !new_topology; j++) {
7425 if (cpumask_equal(doms_cur[i], doms_new[j])
7426 && dattrs_equal(dattr_cur, i, dattr_new, j))
7429 /* no match - a current sched domain not in new doms_new[] */
7430 detach_destroy_domains(doms_cur[i]);
7436 if (doms_new == NULL) {
7438 doms_new = &fallback_doms;
7439 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7440 WARN_ON_ONCE(dattr_new);
7443 /* Build new domains */
7444 for (i = 0; i < ndoms_new; i++) {
7445 for (j = 0; j < n && !new_topology; j++) {
7446 if (cpumask_equal(doms_new[i], doms_cur[j])
7447 && dattrs_equal(dattr_new, i, dattr_cur, j))
7450 /* no match - add a new doms_new */
7451 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7456 /* Remember the new sched domains */
7457 if (doms_cur != &fallback_doms)
7458 free_sched_domains(doms_cur, ndoms_cur);
7459 kfree(dattr_cur); /* kfree(NULL) is safe */
7460 doms_cur = doms_new;
7461 dattr_cur = dattr_new;
7462 ndoms_cur = ndoms_new;
7464 register_sched_domain_sysctl();
7466 mutex_unlock(&sched_domains_mutex);
7469 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7472 * Update cpusets according to cpu_active mask. If cpusets are
7473 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7474 * around partition_sched_domains().
7476 * If we come here as part of a suspend/resume, don't touch cpusets because we
7477 * want to restore it back to its original state upon resume anyway.
7479 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7483 case CPU_ONLINE_FROZEN:
7484 case CPU_DOWN_FAILED_FROZEN:
7487 * num_cpus_frozen tracks how many CPUs are involved in suspend
7488 * resume sequence. As long as this is not the last online
7489 * operation in the resume sequence, just build a single sched
7490 * domain, ignoring cpusets.
7493 if (likely(num_cpus_frozen)) {
7494 partition_sched_domains(1, NULL, NULL);
7499 * This is the last CPU online operation. So fall through and
7500 * restore the original sched domains by considering the
7501 * cpuset configurations.
7505 cpuset_update_active_cpus(true);
7513 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7516 unsigned long flags;
7517 long cpu = (long)hcpu;
7523 case CPU_DOWN_PREPARE:
7524 rcu_read_lock_sched();
7525 dl_b = dl_bw_of(cpu);
7527 raw_spin_lock_irqsave(&dl_b->lock, flags);
7528 cpus = dl_bw_cpus(cpu);
7529 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7530 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7532 rcu_read_unlock_sched();
7535 return notifier_from_errno(-EBUSY);
7536 cpuset_update_active_cpus(false);
7538 case CPU_DOWN_PREPARE_FROZEN:
7540 partition_sched_domains(1, NULL, NULL);
7548 void __init sched_init_smp(void)
7550 cpumask_var_t non_isolated_cpus;
7552 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7553 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7558 * There's no userspace yet to cause hotplug operations; hence all the
7559 * cpu masks are stable and all blatant races in the below code cannot
7562 mutex_lock(&sched_domains_mutex);
7563 init_sched_domains(cpu_active_mask);
7564 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7565 if (cpumask_empty(non_isolated_cpus))
7566 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7567 mutex_unlock(&sched_domains_mutex);
7569 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7570 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7571 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7575 /* Move init over to a non-isolated CPU */
7576 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7578 sched_init_granularity();
7579 free_cpumask_var(non_isolated_cpus);
7581 init_sched_rt_class();
7582 init_sched_dl_class();
7585 void __init sched_init_smp(void)
7587 sched_init_granularity();
7589 #endif /* CONFIG_SMP */
7591 int in_sched_functions(unsigned long addr)
7593 return in_lock_functions(addr) ||
7594 (addr >= (unsigned long)__sched_text_start
7595 && addr < (unsigned long)__sched_text_end);
7598 #ifdef CONFIG_CGROUP_SCHED
7600 * Default task group.
7601 * Every task in system belongs to this group at bootup.
7603 struct task_group root_task_group;
7604 LIST_HEAD(task_groups);
7607 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7609 void __init sched_init(void)
7612 unsigned long alloc_size = 0, ptr;
7614 #ifdef CONFIG_FAIR_GROUP_SCHED
7615 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7617 #ifdef CONFIG_RT_GROUP_SCHED
7618 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7621 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7623 #ifdef CONFIG_FAIR_GROUP_SCHED
7624 root_task_group.se = (struct sched_entity **)ptr;
7625 ptr += nr_cpu_ids * sizeof(void **);
7627 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7628 ptr += nr_cpu_ids * sizeof(void **);
7630 #endif /* CONFIG_FAIR_GROUP_SCHED */
7631 #ifdef CONFIG_RT_GROUP_SCHED
7632 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7633 ptr += nr_cpu_ids * sizeof(void **);
7635 root_task_group.rt_rq = (struct rt_rq **)ptr;
7636 ptr += nr_cpu_ids * sizeof(void **);
7638 #endif /* CONFIG_RT_GROUP_SCHED */
7640 #ifdef CONFIG_CPUMASK_OFFSTACK
7641 for_each_possible_cpu(i) {
7642 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7643 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7645 #endif /* CONFIG_CPUMASK_OFFSTACK */
7647 init_rt_bandwidth(&def_rt_bandwidth,
7648 global_rt_period(), global_rt_runtime());
7649 init_dl_bandwidth(&def_dl_bandwidth,
7650 global_rt_period(), global_rt_runtime());
7653 init_defrootdomain();
7656 #ifdef CONFIG_RT_GROUP_SCHED
7657 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7658 global_rt_period(), global_rt_runtime());
7659 #endif /* CONFIG_RT_GROUP_SCHED */
7661 #ifdef CONFIG_CGROUP_SCHED
7662 list_add(&root_task_group.list, &task_groups);
7663 INIT_LIST_HEAD(&root_task_group.children);
7664 INIT_LIST_HEAD(&root_task_group.siblings);
7665 autogroup_init(&init_task);
7667 #endif /* CONFIG_CGROUP_SCHED */
7669 for_each_possible_cpu(i) {
7673 raw_spin_lock_init(&rq->lock);
7675 rq->calc_load_active = 0;
7676 rq->calc_load_update = jiffies + LOAD_FREQ;
7677 init_cfs_rq(&rq->cfs);
7678 init_rt_rq(&rq->rt);
7679 init_dl_rq(&rq->dl);
7680 #ifdef CONFIG_FAIR_GROUP_SCHED
7681 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7682 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7684 * How much cpu bandwidth does root_task_group get?
7686 * In case of task-groups formed thr' the cgroup filesystem, it
7687 * gets 100% of the cpu resources in the system. This overall
7688 * system cpu resource is divided among the tasks of
7689 * root_task_group and its child task-groups in a fair manner,
7690 * based on each entity's (task or task-group's) weight
7691 * (se->load.weight).
7693 * In other words, if root_task_group has 10 tasks of weight
7694 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7695 * then A0's share of the cpu resource is:
7697 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7699 * We achieve this by letting root_task_group's tasks sit
7700 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7702 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7703 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7704 #endif /* CONFIG_FAIR_GROUP_SCHED */
7706 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7707 #ifdef CONFIG_RT_GROUP_SCHED
7708 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7711 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7712 rq->cpu_load[j] = 0;
7714 rq->last_load_update_tick = jiffies;
7719 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7720 rq->balance_callback = NULL;
7721 rq->active_balance = 0;
7722 rq->next_balance = jiffies;
7727 rq->avg_idle = 2*sysctl_sched_migration_cost;
7728 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7730 INIT_LIST_HEAD(&rq->cfs_tasks);
7732 rq_attach_root(rq, &def_root_domain);
7733 #ifdef CONFIG_NO_HZ_COMMON
7736 #ifdef CONFIG_NO_HZ_FULL
7737 rq->last_sched_tick = 0;
7741 atomic_set(&rq->nr_iowait, 0);
7744 set_load_weight(&init_task);
7746 #ifdef CONFIG_PREEMPT_NOTIFIERS
7747 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7751 * The boot idle thread does lazy MMU switching as well:
7753 atomic_inc(&init_mm.mm_count);
7754 enter_lazy_tlb(&init_mm, current);
7757 * During early bootup we pretend to be a normal task:
7759 current->sched_class = &fair_sched_class;
7762 * Make us the idle thread. Technically, schedule() should not be
7763 * called from this thread, however somewhere below it might be,
7764 * but because we are the idle thread, we just pick up running again
7765 * when this runqueue becomes "idle".
7767 init_idle(current, smp_processor_id());
7769 calc_load_update = jiffies + LOAD_FREQ;
7772 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7773 /* May be allocated at isolcpus cmdline parse time */
7774 if (cpu_isolated_map == NULL)
7775 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7776 idle_thread_set_boot_cpu();
7777 set_cpu_rq_start_time();
7779 init_sched_fair_class();
7781 scheduler_running = 1;
7784 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7785 static inline int preempt_count_equals(int preempt_offset)
7787 int nested = preempt_count() + sched_rcu_preempt_depth();
7789 return (nested == preempt_offset);
7792 void __might_sleep(const char *file, int line, int preempt_offset)
7795 * Blocking primitives will set (and therefore destroy) current->state,
7796 * since we will exit with TASK_RUNNING make sure we enter with it,
7797 * otherwise we will destroy state.
7799 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7800 "do not call blocking ops when !TASK_RUNNING; "
7801 "state=%lx set at [<%p>] %pS\n",
7803 (void *)current->task_state_change,
7804 (void *)current->task_state_change);
7806 ___might_sleep(file, line, preempt_offset);
7808 EXPORT_SYMBOL(__might_sleep);
7810 void ___might_sleep(const char *file, int line, int preempt_offset)
7812 static unsigned long prev_jiffy; /* ratelimiting */
7814 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7815 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7816 !is_idle_task(current)) ||
7817 system_state != SYSTEM_RUNNING || oops_in_progress)
7819 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7821 prev_jiffy = jiffies;
7824 "BUG: sleeping function called from invalid context at %s:%d\n",
7827 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7828 in_atomic(), irqs_disabled(),
7829 current->pid, current->comm);
7831 if (task_stack_end_corrupted(current))
7832 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7834 debug_show_held_locks(current);
7835 if (irqs_disabled())
7836 print_irqtrace_events(current);
7837 #ifdef CONFIG_DEBUG_PREEMPT
7838 if (!preempt_count_equals(preempt_offset)) {
7839 pr_err("Preemption disabled at:");
7840 print_ip_sym(current->preempt_disable_ip);
7846 EXPORT_SYMBOL(___might_sleep);
7849 #ifdef CONFIG_MAGIC_SYSRQ
7850 void normalize_rt_tasks(void)
7852 struct task_struct *g, *p;
7853 struct sched_attr attr = {
7854 .sched_policy = SCHED_NORMAL,
7857 read_lock(&tasklist_lock);
7858 for_each_process_thread(g, p) {
7860 * Only normalize user tasks:
7862 if (p->flags & PF_KTHREAD)
7865 p->se.exec_start = 0;
7866 #ifdef CONFIG_SCHEDSTATS
7867 p->se.statistics.wait_start = 0;
7868 p->se.statistics.sleep_start = 0;
7869 p->se.statistics.block_start = 0;
7872 if (!dl_task(p) && !rt_task(p)) {
7874 * Renice negative nice level userspace
7877 if (task_nice(p) < 0)
7878 set_user_nice(p, 0);
7882 __sched_setscheduler(p, &attr, false, false);
7884 read_unlock(&tasklist_lock);
7887 #endif /* CONFIG_MAGIC_SYSRQ */
7889 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7891 * These functions are only useful for the IA64 MCA handling, or kdb.
7893 * They can only be called when the whole system has been
7894 * stopped - every CPU needs to be quiescent, and no scheduling
7895 * activity can take place. Using them for anything else would
7896 * be a serious bug, and as a result, they aren't even visible
7897 * under any other configuration.
7901 * curr_task - return the current task for a given cpu.
7902 * @cpu: the processor in question.
7904 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7906 * Return: The current task for @cpu.
7908 struct task_struct *curr_task(int cpu)
7910 return cpu_curr(cpu);
7913 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7917 * set_curr_task - set the current task for a given cpu.
7918 * @cpu: the processor in question.
7919 * @p: the task pointer to set.
7921 * Description: This function must only be used when non-maskable interrupts
7922 * are serviced on a separate stack. It allows the architecture to switch the
7923 * notion of the current task on a cpu in a non-blocking manner. This function
7924 * must be called with all CPU's synchronized, and interrupts disabled, the
7925 * and caller must save the original value of the current task (see
7926 * curr_task() above) and restore that value before reenabling interrupts and
7927 * re-starting the system.
7929 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7931 void set_curr_task(int cpu, struct task_struct *p)
7938 #ifdef CONFIG_CGROUP_SCHED
7939 /* task_group_lock serializes the addition/removal of task groups */
7940 static DEFINE_SPINLOCK(task_group_lock);
7942 static void free_sched_group(struct task_group *tg)
7944 free_fair_sched_group(tg);
7945 free_rt_sched_group(tg);
7950 /* allocate runqueue etc for a new task group */
7951 struct task_group *sched_create_group(struct task_group *parent)
7953 struct task_group *tg;
7955 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7957 return ERR_PTR(-ENOMEM);
7959 if (!alloc_fair_sched_group(tg, parent))
7962 if (!alloc_rt_sched_group(tg, parent))
7968 free_sched_group(tg);
7969 return ERR_PTR(-ENOMEM);
7972 void sched_online_group(struct task_group *tg, struct task_group *parent)
7974 unsigned long flags;
7976 spin_lock_irqsave(&task_group_lock, flags);
7977 list_add_rcu(&tg->list, &task_groups);
7979 WARN_ON(!parent); /* root should already exist */
7981 tg->parent = parent;
7982 INIT_LIST_HEAD(&tg->children);
7983 list_add_rcu(&tg->siblings, &parent->children);
7984 spin_unlock_irqrestore(&task_group_lock, flags);
7987 /* rcu callback to free various structures associated with a task group */
7988 static void free_sched_group_rcu(struct rcu_head *rhp)
7990 /* now it should be safe to free those cfs_rqs */
7991 free_sched_group(container_of(rhp, struct task_group, rcu));
7994 /* Destroy runqueue etc associated with a task group */
7995 void sched_destroy_group(struct task_group *tg)
7997 /* wait for possible concurrent references to cfs_rqs complete */
7998 call_rcu(&tg->rcu, free_sched_group_rcu);
8001 void sched_offline_group(struct task_group *tg)
8003 unsigned long flags;
8006 /* end participation in shares distribution */
8007 for_each_possible_cpu(i)
8008 unregister_fair_sched_group(tg, i);
8010 spin_lock_irqsave(&task_group_lock, flags);
8011 list_del_rcu(&tg->list);
8012 list_del_rcu(&tg->siblings);
8013 spin_unlock_irqrestore(&task_group_lock, flags);
8016 /* change task's runqueue when it moves between groups.
8017 * The caller of this function should have put the task in its new group
8018 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8019 * reflect its new group.
8021 void sched_move_task(struct task_struct *tsk)
8023 struct task_group *tg;
8024 int queued, running;
8025 unsigned long flags;
8028 rq = task_rq_lock(tsk, &flags);
8030 running = task_current(rq, tsk);
8031 queued = task_on_rq_queued(tsk);
8034 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8035 if (unlikely(running))
8036 put_prev_task(rq, tsk);
8039 * All callers are synchronized by task_rq_lock(); we do not use RCU
8040 * which is pointless here. Thus, we pass "true" to task_css_check()
8041 * to prevent lockdep warnings.
8043 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8044 struct task_group, css);
8045 tg = autogroup_task_group(tsk, tg);
8046 tsk->sched_task_group = tg;
8048 #ifdef CONFIG_FAIR_GROUP_SCHED
8049 if (tsk->sched_class->task_move_group)
8050 tsk->sched_class->task_move_group(tsk);
8053 set_task_rq(tsk, task_cpu(tsk));
8055 if (unlikely(running))
8056 tsk->sched_class->set_curr_task(rq);
8058 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8060 task_rq_unlock(rq, tsk, &flags);
8062 #endif /* CONFIG_CGROUP_SCHED */
8064 #ifdef CONFIG_RT_GROUP_SCHED
8066 * Ensure that the real time constraints are schedulable.
8068 static DEFINE_MUTEX(rt_constraints_mutex);
8070 /* Must be called with tasklist_lock held */
8071 static inline int tg_has_rt_tasks(struct task_group *tg)
8073 struct task_struct *g, *p;
8076 * Autogroups do not have RT tasks; see autogroup_create().
8078 if (task_group_is_autogroup(tg))
8081 for_each_process_thread(g, p) {
8082 if (rt_task(p) && task_group(p) == tg)
8089 struct rt_schedulable_data {
8090 struct task_group *tg;
8095 static int tg_rt_schedulable(struct task_group *tg, void *data)
8097 struct rt_schedulable_data *d = data;
8098 struct task_group *child;
8099 unsigned long total, sum = 0;
8100 u64 period, runtime;
8102 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8103 runtime = tg->rt_bandwidth.rt_runtime;
8106 period = d->rt_period;
8107 runtime = d->rt_runtime;
8111 * Cannot have more runtime than the period.
8113 if (runtime > period && runtime != RUNTIME_INF)
8117 * Ensure we don't starve existing RT tasks.
8119 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8122 total = to_ratio(period, runtime);
8125 * Nobody can have more than the global setting allows.
8127 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8131 * The sum of our children's runtime should not exceed our own.
8133 list_for_each_entry_rcu(child, &tg->children, siblings) {
8134 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8135 runtime = child->rt_bandwidth.rt_runtime;
8137 if (child == d->tg) {
8138 period = d->rt_period;
8139 runtime = d->rt_runtime;
8142 sum += to_ratio(period, runtime);
8151 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8155 struct rt_schedulable_data data = {
8157 .rt_period = period,
8158 .rt_runtime = runtime,
8162 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8168 static int tg_set_rt_bandwidth(struct task_group *tg,
8169 u64 rt_period, u64 rt_runtime)
8174 * Disallowing the root group RT runtime is BAD, it would disallow the
8175 * kernel creating (and or operating) RT threads.
8177 if (tg == &root_task_group && rt_runtime == 0)
8180 /* No period doesn't make any sense. */
8184 mutex_lock(&rt_constraints_mutex);
8185 read_lock(&tasklist_lock);
8186 err = __rt_schedulable(tg, rt_period, rt_runtime);
8190 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8191 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8192 tg->rt_bandwidth.rt_runtime = rt_runtime;
8194 for_each_possible_cpu(i) {
8195 struct rt_rq *rt_rq = tg->rt_rq[i];
8197 raw_spin_lock(&rt_rq->rt_runtime_lock);
8198 rt_rq->rt_runtime = rt_runtime;
8199 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8201 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8203 read_unlock(&tasklist_lock);
8204 mutex_unlock(&rt_constraints_mutex);
8209 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8211 u64 rt_runtime, rt_period;
8213 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8214 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8215 if (rt_runtime_us < 0)
8216 rt_runtime = RUNTIME_INF;
8218 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8221 static long sched_group_rt_runtime(struct task_group *tg)
8225 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8228 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8229 do_div(rt_runtime_us, NSEC_PER_USEC);
8230 return rt_runtime_us;
8233 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8235 u64 rt_runtime, rt_period;
8237 rt_period = rt_period_us * NSEC_PER_USEC;
8238 rt_runtime = tg->rt_bandwidth.rt_runtime;
8240 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8243 static long sched_group_rt_period(struct task_group *tg)
8247 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8248 do_div(rt_period_us, NSEC_PER_USEC);
8249 return rt_period_us;
8251 #endif /* CONFIG_RT_GROUP_SCHED */
8253 #ifdef CONFIG_RT_GROUP_SCHED
8254 static int sched_rt_global_constraints(void)
8258 mutex_lock(&rt_constraints_mutex);
8259 read_lock(&tasklist_lock);
8260 ret = __rt_schedulable(NULL, 0, 0);
8261 read_unlock(&tasklist_lock);
8262 mutex_unlock(&rt_constraints_mutex);
8267 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8269 /* Don't accept realtime tasks when there is no way for them to run */
8270 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8276 #else /* !CONFIG_RT_GROUP_SCHED */
8277 static int sched_rt_global_constraints(void)
8279 unsigned long flags;
8282 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8283 for_each_possible_cpu(i) {
8284 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8286 raw_spin_lock(&rt_rq->rt_runtime_lock);
8287 rt_rq->rt_runtime = global_rt_runtime();
8288 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8290 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8294 #endif /* CONFIG_RT_GROUP_SCHED */
8296 static int sched_dl_global_validate(void)
8298 u64 runtime = global_rt_runtime();
8299 u64 period = global_rt_period();
8300 u64 new_bw = to_ratio(period, runtime);
8303 unsigned long flags;
8306 * Here we want to check the bandwidth not being set to some
8307 * value smaller than the currently allocated bandwidth in
8308 * any of the root_domains.
8310 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8311 * cycling on root_domains... Discussion on different/better
8312 * solutions is welcome!
8314 for_each_possible_cpu(cpu) {
8315 rcu_read_lock_sched();
8316 dl_b = dl_bw_of(cpu);
8318 raw_spin_lock_irqsave(&dl_b->lock, flags);
8319 if (new_bw < dl_b->total_bw)
8321 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8323 rcu_read_unlock_sched();
8332 static void sched_dl_do_global(void)
8337 unsigned long flags;
8339 def_dl_bandwidth.dl_period = global_rt_period();
8340 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8342 if (global_rt_runtime() != RUNTIME_INF)
8343 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8346 * FIXME: As above...
8348 for_each_possible_cpu(cpu) {
8349 rcu_read_lock_sched();
8350 dl_b = dl_bw_of(cpu);
8352 raw_spin_lock_irqsave(&dl_b->lock, flags);
8354 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8356 rcu_read_unlock_sched();
8360 static int sched_rt_global_validate(void)
8362 if (sysctl_sched_rt_period <= 0)
8365 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8366 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8372 static void sched_rt_do_global(void)
8374 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8375 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8378 int sched_rt_handler(struct ctl_table *table, int write,
8379 void __user *buffer, size_t *lenp,
8382 int old_period, old_runtime;
8383 static DEFINE_MUTEX(mutex);
8387 old_period = sysctl_sched_rt_period;
8388 old_runtime = sysctl_sched_rt_runtime;
8390 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8392 if (!ret && write) {
8393 ret = sched_rt_global_validate();
8397 ret = sched_dl_global_validate();
8401 ret = sched_rt_global_constraints();
8405 sched_rt_do_global();
8406 sched_dl_do_global();
8410 sysctl_sched_rt_period = old_period;
8411 sysctl_sched_rt_runtime = old_runtime;
8413 mutex_unlock(&mutex);
8418 int sched_rr_handler(struct ctl_table *table, int write,
8419 void __user *buffer, size_t *lenp,
8423 static DEFINE_MUTEX(mutex);
8426 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8427 /* make sure that internally we keep jiffies */
8428 /* also, writing zero resets timeslice to default */
8429 if (!ret && write) {
8430 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8431 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8433 mutex_unlock(&mutex);
8437 #ifdef CONFIG_CGROUP_SCHED
8439 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8441 return css ? container_of(css, struct task_group, css) : NULL;
8444 static struct cgroup_subsys_state *
8445 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8447 struct task_group *parent = css_tg(parent_css);
8448 struct task_group *tg;
8451 /* This is early initialization for the top cgroup */
8452 return &root_task_group.css;
8455 tg = sched_create_group(parent);
8457 return ERR_PTR(-ENOMEM);
8462 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8464 struct task_group *tg = css_tg(css);
8465 struct task_group *parent = css_tg(css->parent);
8468 sched_online_group(tg, parent);
8472 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8474 struct task_group *tg = css_tg(css);
8476 sched_destroy_group(tg);
8479 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8481 struct task_group *tg = css_tg(css);
8483 sched_offline_group(tg);
8486 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8488 sched_move_task(task);
8491 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8493 struct task_struct *task;
8494 struct cgroup_subsys_state *css;
8496 cgroup_taskset_for_each(task, css, tset) {
8497 #ifdef CONFIG_RT_GROUP_SCHED
8498 if (!sched_rt_can_attach(css_tg(css), task))
8501 /* We don't support RT-tasks being in separate groups */
8502 if (task->sched_class != &fair_sched_class)
8509 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8511 struct task_struct *task;
8512 struct cgroup_subsys_state *css;
8514 cgroup_taskset_for_each(task, css, tset)
8515 sched_move_task(task);
8518 #ifdef CONFIG_FAIR_GROUP_SCHED
8519 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8520 struct cftype *cftype, u64 shareval)
8522 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8525 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8528 struct task_group *tg = css_tg(css);
8530 return (u64) scale_load_down(tg->shares);
8533 #ifdef CONFIG_CFS_BANDWIDTH
8534 static DEFINE_MUTEX(cfs_constraints_mutex);
8536 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8537 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8539 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8541 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8543 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8544 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8546 if (tg == &root_task_group)
8550 * Ensure we have at some amount of bandwidth every period. This is
8551 * to prevent reaching a state of large arrears when throttled via
8552 * entity_tick() resulting in prolonged exit starvation.
8554 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8558 * Likewise, bound things on the otherside by preventing insane quota
8559 * periods. This also allows us to normalize in computing quota
8562 if (period > max_cfs_quota_period)
8566 * Prevent race between setting of cfs_rq->runtime_enabled and
8567 * unthrottle_offline_cfs_rqs().
8570 mutex_lock(&cfs_constraints_mutex);
8571 ret = __cfs_schedulable(tg, period, quota);
8575 runtime_enabled = quota != RUNTIME_INF;
8576 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8578 * If we need to toggle cfs_bandwidth_used, off->on must occur
8579 * before making related changes, and on->off must occur afterwards
8581 if (runtime_enabled && !runtime_was_enabled)
8582 cfs_bandwidth_usage_inc();
8583 raw_spin_lock_irq(&cfs_b->lock);
8584 cfs_b->period = ns_to_ktime(period);
8585 cfs_b->quota = quota;
8587 __refill_cfs_bandwidth_runtime(cfs_b);
8588 /* restart the period timer (if active) to handle new period expiry */
8589 if (runtime_enabled)
8590 start_cfs_bandwidth(cfs_b);
8591 raw_spin_unlock_irq(&cfs_b->lock);
8593 for_each_online_cpu(i) {
8594 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8595 struct rq *rq = cfs_rq->rq;
8597 raw_spin_lock_irq(&rq->lock);
8598 cfs_rq->runtime_enabled = runtime_enabled;
8599 cfs_rq->runtime_remaining = 0;
8601 if (cfs_rq->throttled)
8602 unthrottle_cfs_rq(cfs_rq);
8603 raw_spin_unlock_irq(&rq->lock);
8605 if (runtime_was_enabled && !runtime_enabled)
8606 cfs_bandwidth_usage_dec();
8608 mutex_unlock(&cfs_constraints_mutex);
8614 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8618 period = ktime_to_ns(tg->cfs_bandwidth.period);
8619 if (cfs_quota_us < 0)
8620 quota = RUNTIME_INF;
8622 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8624 return tg_set_cfs_bandwidth(tg, period, quota);
8627 long tg_get_cfs_quota(struct task_group *tg)
8631 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8634 quota_us = tg->cfs_bandwidth.quota;
8635 do_div(quota_us, NSEC_PER_USEC);
8640 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8644 period = (u64)cfs_period_us * NSEC_PER_USEC;
8645 quota = tg->cfs_bandwidth.quota;
8647 return tg_set_cfs_bandwidth(tg, period, quota);
8650 long tg_get_cfs_period(struct task_group *tg)
8654 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8655 do_div(cfs_period_us, NSEC_PER_USEC);
8657 return cfs_period_us;
8660 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8663 return tg_get_cfs_quota(css_tg(css));
8666 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8667 struct cftype *cftype, s64 cfs_quota_us)
8669 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8672 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8675 return tg_get_cfs_period(css_tg(css));
8678 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8679 struct cftype *cftype, u64 cfs_period_us)
8681 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8684 struct cfs_schedulable_data {
8685 struct task_group *tg;
8690 * normalize group quota/period to be quota/max_period
8691 * note: units are usecs
8693 static u64 normalize_cfs_quota(struct task_group *tg,
8694 struct cfs_schedulable_data *d)
8702 period = tg_get_cfs_period(tg);
8703 quota = tg_get_cfs_quota(tg);
8706 /* note: these should typically be equivalent */
8707 if (quota == RUNTIME_INF || quota == -1)
8710 return to_ratio(period, quota);
8713 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8715 struct cfs_schedulable_data *d = data;
8716 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8717 s64 quota = 0, parent_quota = -1;
8720 quota = RUNTIME_INF;
8722 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8724 quota = normalize_cfs_quota(tg, d);
8725 parent_quota = parent_b->hierarchical_quota;
8728 * ensure max(child_quota) <= parent_quota, inherit when no
8731 if (quota == RUNTIME_INF)
8732 quota = parent_quota;
8733 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8736 cfs_b->hierarchical_quota = quota;
8741 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8744 struct cfs_schedulable_data data = {
8750 if (quota != RUNTIME_INF) {
8751 do_div(data.period, NSEC_PER_USEC);
8752 do_div(data.quota, NSEC_PER_USEC);
8756 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8762 static int cpu_stats_show(struct seq_file *sf, void *v)
8764 struct task_group *tg = css_tg(seq_css(sf));
8765 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8767 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8768 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8769 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8773 #endif /* CONFIG_CFS_BANDWIDTH */
8774 #endif /* CONFIG_FAIR_GROUP_SCHED */
8776 #ifdef CONFIG_RT_GROUP_SCHED
8777 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8778 struct cftype *cft, s64 val)
8780 return sched_group_set_rt_runtime(css_tg(css), val);
8783 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8786 return sched_group_rt_runtime(css_tg(css));
8789 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8790 struct cftype *cftype, u64 rt_period_us)
8792 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8795 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8798 return sched_group_rt_period(css_tg(css));
8800 #endif /* CONFIG_RT_GROUP_SCHED */
8802 static struct cftype cpu_files[] = {
8803 #ifdef CONFIG_FAIR_GROUP_SCHED
8806 .read_u64 = cpu_shares_read_u64,
8807 .write_u64 = cpu_shares_write_u64,
8810 #ifdef CONFIG_CFS_BANDWIDTH
8812 .name = "cfs_quota_us",
8813 .read_s64 = cpu_cfs_quota_read_s64,
8814 .write_s64 = cpu_cfs_quota_write_s64,
8817 .name = "cfs_period_us",
8818 .read_u64 = cpu_cfs_period_read_u64,
8819 .write_u64 = cpu_cfs_period_write_u64,
8823 .seq_show = cpu_stats_show,
8826 #ifdef CONFIG_RT_GROUP_SCHED
8828 .name = "rt_runtime_us",
8829 .read_s64 = cpu_rt_runtime_read,
8830 .write_s64 = cpu_rt_runtime_write,
8833 .name = "rt_period_us",
8834 .read_u64 = cpu_rt_period_read_uint,
8835 .write_u64 = cpu_rt_period_write_uint,
8841 struct cgroup_subsys cpu_cgrp_subsys = {
8842 .css_alloc = cpu_cgroup_css_alloc,
8843 .css_free = cpu_cgroup_css_free,
8844 .css_online = cpu_cgroup_css_online,
8845 .css_offline = cpu_cgroup_css_offline,
8846 .fork = cpu_cgroup_fork,
8847 .can_attach = cpu_cgroup_can_attach,
8848 .attach = cpu_cgroup_attach,
8849 .legacy_cftypes = cpu_files,
8853 #endif /* CONFIG_CGROUP_SCHED */
8855 void dump_cpu_task(int cpu)
8857 pr_info("Task dump for CPU %d:\n", cpu);
8858 sched_show_task(cpu_curr(cpu));