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)) {
676 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
683 if (!is_housekeeping_cpu(cpu))
684 cpu = housekeeping_any_cpu();
692 * When add_timer_on() enqueues a timer into the timer wheel of an
693 * idle CPU then this timer might expire before the next timer event
694 * which is scheduled to wake up that CPU. In case of a completely
695 * idle system the next event might even be infinite time into the
696 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
697 * leaves the inner idle loop so the newly added timer is taken into
698 * account when the CPU goes back to idle and evaluates the timer
699 * wheel for the next timer event.
701 static void wake_up_idle_cpu(int cpu)
703 struct rq *rq = cpu_rq(cpu);
705 if (cpu == smp_processor_id())
708 if (set_nr_and_not_polling(rq->idle))
709 smp_send_reschedule(cpu);
711 trace_sched_wake_idle_without_ipi(cpu);
714 static bool wake_up_full_nohz_cpu(int cpu)
717 * We just need the target to call irq_exit() and re-evaluate
718 * the next tick. The nohz full kick at least implies that.
719 * If needed we can still optimize that later with an
722 if (tick_nohz_full_cpu(cpu)) {
723 if (cpu != smp_processor_id() ||
724 tick_nohz_tick_stopped())
725 tick_nohz_full_kick_cpu(cpu);
732 void wake_up_nohz_cpu(int cpu)
734 if (!wake_up_full_nohz_cpu(cpu))
735 wake_up_idle_cpu(cpu);
738 static inline bool got_nohz_idle_kick(void)
740 int cpu = smp_processor_id();
742 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
745 if (idle_cpu(cpu) && !need_resched())
749 * We can't run Idle Load Balance on this CPU for this time so we
750 * cancel it and clear NOHZ_BALANCE_KICK
752 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
756 #else /* CONFIG_NO_HZ_COMMON */
758 static inline bool got_nohz_idle_kick(void)
763 #endif /* CONFIG_NO_HZ_COMMON */
765 #ifdef CONFIG_NO_HZ_FULL
766 bool sched_can_stop_tick(void)
769 * FIFO realtime policy runs the highest priority task. Other runnable
770 * tasks are of a lower priority. The scheduler tick does nothing.
772 if (current->policy == SCHED_FIFO)
776 * Round-robin realtime tasks time slice with other tasks at the same
777 * realtime priority. Is this task the only one at this priority?
779 if (current->policy == SCHED_RR) {
780 struct sched_rt_entity *rt_se = ¤t->rt;
782 return rt_se->run_list.prev == rt_se->run_list.next;
786 * More than one running task need preemption.
787 * nr_running update is assumed to be visible
788 * after IPI is sent from wakers.
790 if (this_rq()->nr_running > 1)
795 #endif /* CONFIG_NO_HZ_FULL */
797 void sched_avg_update(struct rq *rq)
799 s64 period = sched_avg_period();
801 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
803 * Inline assembly required to prevent the compiler
804 * optimising this loop into a divmod call.
805 * See __iter_div_u64_rem() for another example of this.
807 asm("" : "+rm" (rq->age_stamp));
808 rq->age_stamp += period;
813 #endif /* CONFIG_SMP */
815 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
816 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
818 * Iterate task_group tree rooted at *from, calling @down when first entering a
819 * node and @up when leaving it for the final time.
821 * Caller must hold rcu_lock or sufficient equivalent.
823 int walk_tg_tree_from(struct task_group *from,
824 tg_visitor down, tg_visitor up, void *data)
826 struct task_group *parent, *child;
832 ret = (*down)(parent, data);
835 list_for_each_entry_rcu(child, &parent->children, siblings) {
842 ret = (*up)(parent, data);
843 if (ret || parent == from)
847 parent = parent->parent;
854 int tg_nop(struct task_group *tg, void *data)
860 static void set_load_weight(struct task_struct *p)
862 int prio = p->static_prio - MAX_RT_PRIO;
863 struct load_weight *load = &p->se.load;
866 * SCHED_IDLE tasks get minimal weight:
868 if (idle_policy(p->policy)) {
869 load->weight = scale_load(WEIGHT_IDLEPRIO);
870 load->inv_weight = WMULT_IDLEPRIO;
874 load->weight = scale_load(prio_to_weight[prio]);
875 load->inv_weight = prio_to_wmult[prio];
878 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
881 if (!(flags & ENQUEUE_RESTORE))
882 sched_info_queued(rq, p);
883 p->sched_class->enqueue_task(rq, p, flags);
886 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
889 if (!(flags & DEQUEUE_SAVE))
890 sched_info_dequeued(rq, p);
891 p->sched_class->dequeue_task(rq, p, flags);
894 void activate_task(struct rq *rq, struct task_struct *p, int flags)
896 if (task_contributes_to_load(p))
897 rq->nr_uninterruptible--;
899 enqueue_task(rq, p, flags);
902 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
904 if (task_contributes_to_load(p))
905 rq->nr_uninterruptible++;
907 dequeue_task(rq, p, flags);
910 static void update_rq_clock_task(struct rq *rq, s64 delta)
913 * In theory, the compile should just see 0 here, and optimize out the call
914 * to sched_rt_avg_update. But I don't trust it...
916 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
917 s64 steal = 0, irq_delta = 0;
919 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
920 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
923 * Since irq_time is only updated on {soft,}irq_exit, we might run into
924 * this case when a previous update_rq_clock() happened inside a
927 * When this happens, we stop ->clock_task and only update the
928 * prev_irq_time stamp to account for the part that fit, so that a next
929 * update will consume the rest. This ensures ->clock_task is
932 * It does however cause some slight miss-attribution of {soft,}irq
933 * time, a more accurate solution would be to update the irq_time using
934 * the current rq->clock timestamp, except that would require using
937 if (irq_delta > delta)
940 rq->prev_irq_time += irq_delta;
943 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
944 if (static_key_false((¶virt_steal_rq_enabled))) {
945 steal = paravirt_steal_clock(cpu_of(rq));
946 steal -= rq->prev_steal_time_rq;
948 if (unlikely(steal > delta))
951 rq->prev_steal_time_rq += steal;
956 rq->clock_task += delta;
958 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
959 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
960 sched_rt_avg_update(rq, irq_delta + steal);
964 void sched_set_stop_task(int cpu, struct task_struct *stop)
966 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
967 struct task_struct *old_stop = cpu_rq(cpu)->stop;
971 * Make it appear like a SCHED_FIFO task, its something
972 * userspace knows about and won't get confused about.
974 * Also, it will make PI more or less work without too
975 * much confusion -- but then, stop work should not
976 * rely on PI working anyway.
978 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
980 stop->sched_class = &stop_sched_class;
983 cpu_rq(cpu)->stop = stop;
987 * Reset it back to a normal scheduling class so that
988 * it can die in pieces.
990 old_stop->sched_class = &rt_sched_class;
995 * __normal_prio - return the priority that is based on the static prio
997 static inline int __normal_prio(struct task_struct *p)
999 return p->static_prio;
1003 * Calculate the expected normal priority: i.e. priority
1004 * without taking RT-inheritance into account. Might be
1005 * boosted by interactivity modifiers. Changes upon fork,
1006 * setprio syscalls, and whenever the interactivity
1007 * estimator recalculates.
1009 static inline int normal_prio(struct task_struct *p)
1013 if (task_has_dl_policy(p))
1014 prio = MAX_DL_PRIO-1;
1015 else if (task_has_rt_policy(p))
1016 prio = MAX_RT_PRIO-1 - p->rt_priority;
1018 prio = __normal_prio(p);
1023 * Calculate the current priority, i.e. the priority
1024 * taken into account by the scheduler. This value might
1025 * be boosted by RT tasks, or might be boosted by
1026 * interactivity modifiers. Will be RT if the task got
1027 * RT-boosted. If not then it returns p->normal_prio.
1029 static int effective_prio(struct task_struct *p)
1031 p->normal_prio = normal_prio(p);
1033 * If we are RT tasks or we were boosted to RT priority,
1034 * keep the priority unchanged. Otherwise, update priority
1035 * to the normal priority:
1037 if (!rt_prio(p->prio))
1038 return p->normal_prio;
1043 * task_curr - is this task currently executing on a CPU?
1044 * @p: the task in question.
1046 * Return: 1 if the task is currently executing. 0 otherwise.
1048 inline int task_curr(const struct task_struct *p)
1050 return cpu_curr(task_cpu(p)) == p;
1054 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1055 * use the balance_callback list if you want balancing.
1057 * this means any call to check_class_changed() must be followed by a call to
1058 * balance_callback().
1060 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1061 const struct sched_class *prev_class,
1064 if (prev_class != p->sched_class) {
1065 if (prev_class->switched_from)
1066 prev_class->switched_from(rq, p);
1068 p->sched_class->switched_to(rq, p);
1069 } else if (oldprio != p->prio || dl_task(p))
1070 p->sched_class->prio_changed(rq, p, oldprio);
1073 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1075 const struct sched_class *class;
1077 if (p->sched_class == rq->curr->sched_class) {
1078 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1080 for_each_class(class) {
1081 if (class == rq->curr->sched_class)
1083 if (class == p->sched_class) {
1091 * A queue event has occurred, and we're going to schedule. In
1092 * this case, we can save a useless back to back clock update.
1094 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1095 rq_clock_skip_update(rq, true);
1100 * This is how migration works:
1102 * 1) we invoke migration_cpu_stop() on the target CPU using
1104 * 2) stopper starts to run (implicitly forcing the migrated thread
1106 * 3) it checks whether the migrated task is still in the wrong runqueue.
1107 * 4) if it's in the wrong runqueue then the migration thread removes
1108 * it and puts it into the right queue.
1109 * 5) stopper completes and stop_one_cpu() returns and the migration
1114 * move_queued_task - move a queued task to new rq.
1116 * Returns (locked) new rq. Old rq's lock is released.
1118 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1120 lockdep_assert_held(&rq->lock);
1122 dequeue_task(rq, p, 0);
1123 p->on_rq = TASK_ON_RQ_MIGRATING;
1124 set_task_cpu(p, new_cpu);
1125 raw_spin_unlock(&rq->lock);
1127 rq = cpu_rq(new_cpu);
1129 raw_spin_lock(&rq->lock);
1130 BUG_ON(task_cpu(p) != new_cpu);
1131 p->on_rq = TASK_ON_RQ_QUEUED;
1132 enqueue_task(rq, p, 0);
1133 check_preempt_curr(rq, p, 0);
1138 struct migration_arg {
1139 struct task_struct *task;
1144 * Move (not current) task off this cpu, onto dest cpu. We're doing
1145 * this because either it can't run here any more (set_cpus_allowed()
1146 * away from this CPU, or CPU going down), or because we're
1147 * attempting to rebalance this task on exec (sched_exec).
1149 * So we race with normal scheduler movements, but that's OK, as long
1150 * as the task is no longer on this CPU.
1152 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1154 if (unlikely(!cpu_active(dest_cpu)))
1157 /* Affinity changed (again). */
1158 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1161 rq = move_queued_task(rq, p, dest_cpu);
1167 * migration_cpu_stop - this will be executed by a highprio stopper thread
1168 * and performs thread migration by bumping thread off CPU then
1169 * 'pushing' onto another runqueue.
1171 static int migration_cpu_stop(void *data)
1173 struct migration_arg *arg = data;
1174 struct task_struct *p = arg->task;
1175 struct rq *rq = this_rq();
1178 * The original target cpu might have gone down and we might
1179 * be on another cpu but it doesn't matter.
1181 local_irq_disable();
1183 * We need to explicitly wake pending tasks before running
1184 * __migrate_task() such that we will not miss enforcing cpus_allowed
1185 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1187 sched_ttwu_pending();
1189 raw_spin_lock(&p->pi_lock);
1190 raw_spin_lock(&rq->lock);
1192 * If task_rq(p) != rq, it cannot be migrated here, because we're
1193 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1194 * we're holding p->pi_lock.
1196 if (task_rq(p) == rq && task_on_rq_queued(p))
1197 rq = __migrate_task(rq, p, arg->dest_cpu);
1198 raw_spin_unlock(&rq->lock);
1199 raw_spin_unlock(&p->pi_lock);
1206 * sched_class::set_cpus_allowed must do the below, but is not required to
1207 * actually call this function.
1209 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1211 cpumask_copy(&p->cpus_allowed, new_mask);
1212 p->nr_cpus_allowed = cpumask_weight(new_mask);
1215 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1217 struct rq *rq = task_rq(p);
1218 bool queued, running;
1220 lockdep_assert_held(&p->pi_lock);
1222 if (__migrate_disabled(p)) {
1223 cpumask_copy(&p->cpus_allowed, new_mask);
1227 queued = task_on_rq_queued(p);
1228 running = task_current(rq, p);
1232 * Because __kthread_bind() calls this on blocked tasks without
1235 lockdep_assert_held(&rq->lock);
1236 dequeue_task(rq, p, DEQUEUE_SAVE);
1239 put_prev_task(rq, p);
1241 p->sched_class->set_cpus_allowed(p, new_mask);
1244 p->sched_class->set_curr_task(rq);
1246 enqueue_task(rq, p, ENQUEUE_RESTORE);
1249 static DEFINE_PER_CPU(struct cpumask, sched_cpumasks);
1250 static DEFINE_MUTEX(sched_down_mutex);
1251 static cpumask_t sched_down_cpumask;
1253 void tell_sched_cpu_down_begin(int cpu)
1255 mutex_lock(&sched_down_mutex);
1256 cpumask_set_cpu(cpu, &sched_down_cpumask);
1257 mutex_unlock(&sched_down_mutex);
1260 void tell_sched_cpu_down_done(int cpu)
1262 mutex_lock(&sched_down_mutex);
1263 cpumask_clear_cpu(cpu, &sched_down_cpumask);
1264 mutex_unlock(&sched_down_mutex);
1268 * migrate_me - try to move the current task off this cpu
1270 * Used by the pin_current_cpu() code to try to get tasks
1271 * to move off the current CPU as it is going down.
1272 * It will only move the task if the task isn't pinned to
1273 * the CPU (with migrate_disable, affinity or NO_SETAFFINITY)
1274 * and the task has to be in a RUNNING state. Otherwise the
1275 * movement of the task will wake it up (change its state
1276 * to running) when the task did not expect it.
1278 * Returns 1 if it succeeded in moving the current task
1281 int migrate_me(void)
1283 struct task_struct *p = current;
1284 struct migration_arg arg;
1285 struct cpumask *cpumask;
1286 struct cpumask *mask;
1287 unsigned long flags;
1288 unsigned int dest_cpu;
1292 * We can not migrate tasks bounded to a CPU or tasks not
1293 * running. The movement of the task will wake it up.
1295 if (p->flags & PF_NO_SETAFFINITY || p->state)
1298 mutex_lock(&sched_down_mutex);
1299 rq = task_rq_lock(p, &flags);
1301 cpumask = this_cpu_ptr(&sched_cpumasks);
1302 mask = &p->cpus_allowed;
1304 cpumask_andnot(cpumask, mask, &sched_down_cpumask);
1306 if (!cpumask_weight(cpumask)) {
1307 /* It's only on this CPU? */
1308 task_rq_unlock(rq, p, &flags);
1309 mutex_unlock(&sched_down_mutex);
1313 dest_cpu = cpumask_any_and(cpu_active_mask, cpumask);
1316 arg.dest_cpu = dest_cpu;
1318 task_rq_unlock(rq, p, &flags);
1320 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1321 tlb_migrate_finish(p->mm);
1322 mutex_unlock(&sched_down_mutex);
1328 * Change a given task's CPU affinity. Migrate the thread to a
1329 * proper CPU and schedule it away if the CPU it's executing on
1330 * is removed from the allowed bitmask.
1332 * NOTE: the caller must have a valid reference to the task, the
1333 * task must not exit() & deallocate itself prematurely. The
1334 * call is not atomic; no spinlocks may be held.
1336 static int __set_cpus_allowed_ptr(struct task_struct *p,
1337 const struct cpumask *new_mask, bool check)
1339 unsigned long flags;
1341 unsigned int dest_cpu;
1344 rq = task_rq_lock(p, &flags);
1347 * Must re-check here, to close a race against __kthread_bind(),
1348 * sched_setaffinity() is not guaranteed to observe the flag.
1350 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1355 if (cpumask_equal(&p->cpus_allowed, new_mask))
1358 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1363 do_set_cpus_allowed(p, new_mask);
1365 /* Can the task run on the task's current CPU? If so, we're done */
1366 if (cpumask_test_cpu(task_cpu(p), new_mask) || __migrate_disabled(p))
1369 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1370 if (task_running(rq, p) || p->state == TASK_WAKING) {
1371 struct migration_arg arg = { p, dest_cpu };
1372 /* Need help from migration thread: drop lock and wait. */
1373 task_rq_unlock(rq, p, &flags);
1374 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1375 tlb_migrate_finish(p->mm);
1377 } else if (task_on_rq_queued(p)) {
1379 * OK, since we're going to drop the lock immediately
1380 * afterwards anyway.
1382 lockdep_unpin_lock(&rq->lock);
1383 rq = move_queued_task(rq, p, dest_cpu);
1384 lockdep_pin_lock(&rq->lock);
1387 task_rq_unlock(rq, p, &flags);
1392 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1394 return __set_cpus_allowed_ptr(p, new_mask, false);
1396 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1398 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1400 #ifdef CONFIG_SCHED_DEBUG
1402 * We should never call set_task_cpu() on a blocked task,
1403 * ttwu() will sort out the placement.
1405 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1408 #ifdef CONFIG_LOCKDEP
1410 * The caller should hold either p->pi_lock or rq->lock, when changing
1411 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1413 * sched_move_task() holds both and thus holding either pins the cgroup,
1416 * Furthermore, all task_rq users should acquire both locks, see
1419 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1420 lockdep_is_held(&task_rq(p)->lock)));
1424 trace_sched_migrate_task(p, new_cpu);
1426 if (task_cpu(p) != new_cpu) {
1427 if (p->sched_class->migrate_task_rq)
1428 p->sched_class->migrate_task_rq(p);
1429 p->se.nr_migrations++;
1430 perf_event_task_migrate(p);
1433 __set_task_cpu(p, new_cpu);
1436 static void __migrate_swap_task(struct task_struct *p, int cpu)
1438 if (task_on_rq_queued(p)) {
1439 struct rq *src_rq, *dst_rq;
1441 src_rq = task_rq(p);
1442 dst_rq = cpu_rq(cpu);
1444 deactivate_task(src_rq, p, 0);
1445 set_task_cpu(p, cpu);
1446 activate_task(dst_rq, p, 0);
1447 check_preempt_curr(dst_rq, p, 0);
1450 * Task isn't running anymore; make it appear like we migrated
1451 * it before it went to sleep. This means on wakeup we make the
1452 * previous cpu our targer instead of where it really is.
1458 struct migration_swap_arg {
1459 struct task_struct *src_task, *dst_task;
1460 int src_cpu, dst_cpu;
1463 static int migrate_swap_stop(void *data)
1465 struct migration_swap_arg *arg = data;
1466 struct rq *src_rq, *dst_rq;
1469 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1472 src_rq = cpu_rq(arg->src_cpu);
1473 dst_rq = cpu_rq(arg->dst_cpu);
1475 double_raw_lock(&arg->src_task->pi_lock,
1476 &arg->dst_task->pi_lock);
1477 double_rq_lock(src_rq, dst_rq);
1479 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1482 if (task_cpu(arg->src_task) != arg->src_cpu)
1485 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1488 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1491 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1492 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1497 double_rq_unlock(src_rq, dst_rq);
1498 raw_spin_unlock(&arg->dst_task->pi_lock);
1499 raw_spin_unlock(&arg->src_task->pi_lock);
1505 * Cross migrate two tasks
1507 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1509 struct migration_swap_arg arg;
1512 arg = (struct migration_swap_arg){
1514 .src_cpu = task_cpu(cur),
1516 .dst_cpu = task_cpu(p),
1519 if (arg.src_cpu == arg.dst_cpu)
1523 * These three tests are all lockless; this is OK since all of them
1524 * will be re-checked with proper locks held further down the line.
1526 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1529 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1532 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1535 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1536 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1542 static bool check_task_state(struct task_struct *p, long match_state)
1546 raw_spin_lock_irq(&p->pi_lock);
1547 if (p->state == match_state || p->saved_state == match_state)
1549 raw_spin_unlock_irq(&p->pi_lock);
1555 * wait_task_inactive - wait for a thread to unschedule.
1557 * If @match_state is nonzero, it's the @p->state value just checked and
1558 * not expected to change. If it changes, i.e. @p might have woken up,
1559 * then return zero. When we succeed in waiting for @p to be off its CPU,
1560 * we return a positive number (its total switch count). If a second call
1561 * a short while later returns the same number, the caller can be sure that
1562 * @p has remained unscheduled the whole time.
1564 * The caller must ensure that the task *will* unschedule sometime soon,
1565 * else this function might spin for a *long* time. This function can't
1566 * be called with interrupts off, or it may introduce deadlock with
1567 * smp_call_function() if an IPI is sent by the same process we are
1568 * waiting to become inactive.
1570 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1572 unsigned long flags;
1573 int running, queued;
1579 * We do the initial early heuristics without holding
1580 * any task-queue locks at all. We'll only try to get
1581 * the runqueue lock when things look like they will
1587 * If the task is actively running on another CPU
1588 * still, just relax and busy-wait without holding
1591 * NOTE! Since we don't hold any locks, it's not
1592 * even sure that "rq" stays as the right runqueue!
1593 * But we don't care, since "task_running()" will
1594 * return false if the runqueue has changed and p
1595 * is actually now running somewhere else!
1597 while (task_running(rq, p)) {
1598 if (match_state && !check_task_state(p, match_state))
1604 * Ok, time to look more closely! We need the rq
1605 * lock now, to be *sure*. If we're wrong, we'll
1606 * just go back and repeat.
1608 rq = task_rq_lock(p, &flags);
1609 trace_sched_wait_task(p);
1610 running = task_running(rq, p);
1611 queued = task_on_rq_queued(p);
1613 if (!match_state || p->state == match_state ||
1614 p->saved_state == match_state)
1615 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1616 task_rq_unlock(rq, p, &flags);
1619 * If it changed from the expected state, bail out now.
1621 if (unlikely(!ncsw))
1625 * Was it really running after all now that we
1626 * checked with the proper locks actually held?
1628 * Oops. Go back and try again..
1630 if (unlikely(running)) {
1636 * It's not enough that it's not actively running,
1637 * it must be off the runqueue _entirely_, and not
1640 * So if it was still runnable (but just not actively
1641 * running right now), it's preempted, and we should
1642 * yield - it could be a while.
1644 if (unlikely(queued)) {
1645 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1647 set_current_state(TASK_UNINTERRUPTIBLE);
1648 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1653 * Ahh, all good. It wasn't running, and it wasn't
1654 * runnable, which means that it will never become
1655 * running in the future either. We're all done!
1664 * kick_process - kick a running thread to enter/exit the kernel
1665 * @p: the to-be-kicked thread
1667 * Cause a process which is running on another CPU to enter
1668 * kernel-mode, without any delay. (to get signals handled.)
1670 * NOTE: this function doesn't have to take the runqueue lock,
1671 * because all it wants to ensure is that the remote task enters
1672 * the kernel. If the IPI races and the task has been migrated
1673 * to another CPU then no harm is done and the purpose has been
1676 void kick_process(struct task_struct *p)
1682 if ((cpu != smp_processor_id()) && task_curr(p))
1683 smp_send_reschedule(cpu);
1686 EXPORT_SYMBOL_GPL(kick_process);
1689 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1691 static int select_fallback_rq(int cpu, struct task_struct *p)
1693 int nid = cpu_to_node(cpu);
1694 const struct cpumask *nodemask = NULL;
1695 enum { cpuset, possible, fail } state = cpuset;
1699 * If the node that the cpu is on has been offlined, cpu_to_node()
1700 * will return -1. There is no cpu on the node, and we should
1701 * select the cpu on the other node.
1704 nodemask = cpumask_of_node(nid);
1706 /* Look for allowed, online CPU in same node. */
1707 for_each_cpu(dest_cpu, nodemask) {
1708 if (!cpu_online(dest_cpu))
1710 if (!cpu_active(dest_cpu))
1712 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1718 /* Any allowed, online CPU? */
1719 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1720 if (!cpu_online(dest_cpu))
1722 if (!cpu_active(dest_cpu))
1727 /* No more Mr. Nice Guy. */
1730 if (IS_ENABLED(CONFIG_CPUSETS)) {
1731 cpuset_cpus_allowed_fallback(p);
1737 do_set_cpus_allowed(p, cpu_possible_mask);
1748 if (state != cpuset) {
1750 * Don't tell them about moving exiting tasks or
1751 * kernel threads (both mm NULL), since they never
1754 if (p->mm && printk_ratelimit()) {
1755 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1756 task_pid_nr(p), p->comm, cpu);
1764 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1767 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1769 lockdep_assert_held(&p->pi_lock);
1771 if (tsk_nr_cpus_allowed(p) > 1)
1772 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1775 * In order not to call set_task_cpu() on a blocking task we need
1776 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1779 * Since this is common to all placement strategies, this lives here.
1781 * [ this allows ->select_task() to simply return task_cpu(p) and
1782 * not worry about this generic constraint ]
1784 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1786 cpu = select_fallback_rq(task_cpu(p), p);
1791 static void update_avg(u64 *avg, u64 sample)
1793 s64 diff = sample - *avg;
1799 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1800 const struct cpumask *new_mask, bool check)
1802 return set_cpus_allowed_ptr(p, new_mask);
1805 #endif /* CONFIG_SMP */
1808 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1810 #ifdef CONFIG_SCHEDSTATS
1811 struct rq *rq = this_rq();
1814 int this_cpu = smp_processor_id();
1816 if (cpu == this_cpu) {
1817 schedstat_inc(rq, ttwu_local);
1818 schedstat_inc(p, se.statistics.nr_wakeups_local);
1820 struct sched_domain *sd;
1822 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1824 for_each_domain(this_cpu, sd) {
1825 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1826 schedstat_inc(sd, ttwu_wake_remote);
1833 if (wake_flags & WF_MIGRATED)
1834 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1836 #endif /* CONFIG_SMP */
1838 schedstat_inc(rq, ttwu_count);
1839 schedstat_inc(p, se.statistics.nr_wakeups);
1841 if (wake_flags & WF_SYNC)
1842 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1844 #endif /* CONFIG_SCHEDSTATS */
1847 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1849 activate_task(rq, p, en_flags);
1850 p->on_rq = TASK_ON_RQ_QUEUED;
1854 * Mark the task runnable and perform wakeup-preemption.
1857 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1859 check_preempt_curr(rq, p, wake_flags);
1860 p->state = TASK_RUNNING;
1861 trace_sched_wakeup(p);
1864 if (p->sched_class->task_woken) {
1866 * Our task @p is fully woken up and running; so its safe to
1867 * drop the rq->lock, hereafter rq is only used for statistics.
1869 lockdep_unpin_lock(&rq->lock);
1870 p->sched_class->task_woken(rq, p);
1871 lockdep_pin_lock(&rq->lock);
1874 if (rq->idle_stamp) {
1875 u64 delta = rq_clock(rq) - rq->idle_stamp;
1876 u64 max = 2*rq->max_idle_balance_cost;
1878 update_avg(&rq->avg_idle, delta);
1880 if (rq->avg_idle > max)
1889 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1891 lockdep_assert_held(&rq->lock);
1894 if (p->sched_contributes_to_load)
1895 rq->nr_uninterruptible--;
1898 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1899 ttwu_do_wakeup(rq, p, wake_flags);
1903 * Called in case the task @p isn't fully descheduled from its runqueue,
1904 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1905 * since all we need to do is flip p->state to TASK_RUNNING, since
1906 * the task is still ->on_rq.
1908 static int ttwu_remote(struct task_struct *p, int wake_flags)
1913 rq = __task_rq_lock(p);
1914 if (task_on_rq_queued(p)) {
1915 /* check_preempt_curr() may use rq clock */
1916 update_rq_clock(rq);
1917 ttwu_do_wakeup(rq, p, wake_flags);
1920 __task_rq_unlock(rq);
1926 void sched_ttwu_pending(void)
1928 struct rq *rq = this_rq();
1929 struct llist_node *llist = llist_del_all(&rq->wake_list);
1930 struct task_struct *p;
1931 unsigned long flags;
1936 raw_spin_lock_irqsave(&rq->lock, flags);
1937 lockdep_pin_lock(&rq->lock);
1940 p = llist_entry(llist, struct task_struct, wake_entry);
1941 llist = llist_next(llist);
1942 ttwu_do_activate(rq, p, 0);
1945 lockdep_unpin_lock(&rq->lock);
1946 raw_spin_unlock_irqrestore(&rq->lock, flags);
1949 void scheduler_ipi(void)
1952 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1953 * TIF_NEED_RESCHED remotely (for the first time) will also send
1956 preempt_fold_need_resched();
1958 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1962 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1963 * traditionally all their work was done from the interrupt return
1964 * path. Now that we actually do some work, we need to make sure
1967 * Some archs already do call them, luckily irq_enter/exit nest
1970 * Arguably we should visit all archs and update all handlers,
1971 * however a fair share of IPIs are still resched only so this would
1972 * somewhat pessimize the simple resched case.
1975 sched_ttwu_pending();
1978 * Check if someone kicked us for doing the nohz idle load balance.
1980 if (unlikely(got_nohz_idle_kick())) {
1981 this_rq()->idle_balance = 1;
1982 raise_softirq_irqoff(SCHED_SOFTIRQ);
1987 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1989 struct rq *rq = cpu_rq(cpu);
1991 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1992 if (!set_nr_if_polling(rq->idle))
1993 smp_send_reschedule(cpu);
1995 trace_sched_wake_idle_without_ipi(cpu);
1999 void wake_up_if_idle(int cpu)
2001 struct rq *rq = cpu_rq(cpu);
2002 unsigned long flags;
2006 if (!is_idle_task(rcu_dereference(rq->curr)))
2009 if (set_nr_if_polling(rq->idle)) {
2010 trace_sched_wake_idle_without_ipi(cpu);
2012 raw_spin_lock_irqsave(&rq->lock, flags);
2013 if (is_idle_task(rq->curr))
2014 smp_send_reschedule(cpu);
2015 /* Else cpu is not in idle, do nothing here */
2016 raw_spin_unlock_irqrestore(&rq->lock, flags);
2023 bool cpus_share_cache(int this_cpu, int that_cpu)
2025 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
2027 #endif /* CONFIG_SMP */
2029 static void ttwu_queue(struct task_struct *p, int cpu)
2031 struct rq *rq = cpu_rq(cpu);
2033 #if defined(CONFIG_SMP)
2034 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
2035 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2036 ttwu_queue_remote(p, cpu);
2041 raw_spin_lock(&rq->lock);
2042 lockdep_pin_lock(&rq->lock);
2043 ttwu_do_activate(rq, p, 0);
2044 lockdep_unpin_lock(&rq->lock);
2045 raw_spin_unlock(&rq->lock);
2049 * try_to_wake_up - wake up a thread
2050 * @p: the thread to be awakened
2051 * @state: the mask of task states that can be woken
2052 * @wake_flags: wake modifier flags (WF_*)
2054 * Put it on the run-queue if it's not already there. The "current"
2055 * thread is always on the run-queue (except when the actual
2056 * re-schedule is in progress), and as such you're allowed to do
2057 * the simpler "current->state = TASK_RUNNING" to mark yourself
2058 * runnable without the overhead of this.
2060 * Return: %true if @p was woken up, %false if it was already running.
2061 * or @state didn't match @p's state.
2064 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2066 unsigned long flags;
2067 int cpu, success = 0;
2070 * If we are going to wake up a thread waiting for CONDITION we
2071 * need to ensure that CONDITION=1 done by the caller can not be
2072 * reordered with p->state check below. This pairs with mb() in
2073 * set_current_state() the waiting thread does.
2075 smp_mb__before_spinlock();
2076 raw_spin_lock_irqsave(&p->pi_lock, flags);
2077 if (!(p->state & state)) {
2079 * The task might be running due to a spinlock sleeper
2080 * wakeup. Check the saved state and set it to running
2081 * if the wakeup condition is true.
2083 if (!(wake_flags & WF_LOCK_SLEEPER)) {
2084 if (p->saved_state & state) {
2085 p->saved_state = TASK_RUNNING;
2093 * If this is a regular wakeup, then we can unconditionally
2094 * clear the saved state of a "lock sleeper".
2096 if (!(wake_flags & WF_LOCK_SLEEPER))
2097 p->saved_state = TASK_RUNNING;
2099 trace_sched_waking(p);
2101 success = 1; /* we're going to change ->state */
2105 * Ensure we load p->on_rq _after_ p->state, otherwise it would
2106 * be possible to, falsely, observe p->on_rq == 0 and get stuck
2107 * in smp_cond_load_acquire() below.
2109 * sched_ttwu_pending() try_to_wake_up()
2110 * [S] p->on_rq = 1; [L] P->state
2111 * UNLOCK rq->lock -----.
2115 * LOCK rq->lock -----'
2119 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2121 * Pairs with the UNLOCK+LOCK on rq->lock from the
2122 * last wakeup of our task and the schedule that got our task
2126 if (p->on_rq && ttwu_remote(p, wake_flags))
2131 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2132 * possible to, falsely, observe p->on_cpu == 0.
2134 * One must be running (->on_cpu == 1) in order to remove oneself
2135 * from the runqueue.
2137 * [S] ->on_cpu = 1; [L] ->on_rq
2141 * [S] ->on_rq = 0; [L] ->on_cpu
2143 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2144 * from the consecutive calls to schedule(); the first switching to our
2145 * task, the second putting it to sleep.
2150 * If the owning (remote) cpu is still in the middle of schedule() with
2151 * this task as prev, wait until its done referencing the task.
2156 * Combined with the control dependency above, we have an effective
2157 * smp_load_acquire() without the need for full barriers.
2159 * Pairs with the smp_store_release() in finish_lock_switch().
2161 * This ensures that tasks getting woken will be fully ordered against
2162 * their previous state and preserve Program Order.
2166 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2167 p->state = TASK_WAKING;
2169 if (p->sched_class->task_waking)
2170 p->sched_class->task_waking(p);
2172 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2173 if (task_cpu(p) != cpu) {
2174 wake_flags |= WF_MIGRATED;
2175 set_task_cpu(p, cpu);
2177 #endif /* CONFIG_SMP */
2181 ttwu_stat(p, cpu, wake_flags);
2183 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2189 * wake_up_process - Wake up a specific process
2190 * @p: The process to be woken up.
2192 * Attempt to wake up the nominated process and move it to the set of runnable
2195 * Return: 1 if the process was woken up, 0 if it was already running.
2197 * It may be assumed that this function implies a write memory barrier before
2198 * changing the task state if and only if any tasks are woken up.
2200 int wake_up_process(struct task_struct *p)
2202 return try_to_wake_up(p, TASK_NORMAL, 0);
2204 EXPORT_SYMBOL(wake_up_process);
2207 * wake_up_lock_sleeper - Wake up a specific process blocked on a "sleeping lock"
2208 * @p: The process to be woken up.
2210 * Same as wake_up_process() above, but wake_flags=WF_LOCK_SLEEPER to indicate
2211 * the nature of the wakeup.
2213 int wake_up_lock_sleeper(struct task_struct *p)
2215 return try_to_wake_up(p, TASK_ALL, WF_LOCK_SLEEPER);
2218 int wake_up_state(struct task_struct *p, unsigned int state)
2220 return try_to_wake_up(p, state, 0);
2224 * This function clears the sched_dl_entity static params.
2226 void __dl_clear_params(struct task_struct *p)
2228 struct sched_dl_entity *dl_se = &p->dl;
2230 dl_se->dl_runtime = 0;
2231 dl_se->dl_deadline = 0;
2232 dl_se->dl_period = 0;
2236 dl_se->dl_throttled = 0;
2238 dl_se->dl_yielded = 0;
2242 * Perform scheduler related setup for a newly forked process p.
2243 * p is forked by current.
2245 * __sched_fork() is basic setup used by init_idle() too:
2247 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2252 p->se.exec_start = 0;
2253 p->se.sum_exec_runtime = 0;
2254 p->se.prev_sum_exec_runtime = 0;
2255 p->se.nr_migrations = 0;
2257 INIT_LIST_HEAD(&p->se.group_node);
2259 #ifdef CONFIG_SCHEDSTATS
2260 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2263 RB_CLEAR_NODE(&p->dl.rb_node);
2264 init_dl_task_timer(&p->dl);
2265 __dl_clear_params(p);
2267 INIT_LIST_HEAD(&p->rt.run_list);
2269 #ifdef CONFIG_PREEMPT_NOTIFIERS
2270 INIT_HLIST_HEAD(&p->preempt_notifiers);
2273 #ifdef CONFIG_NUMA_BALANCING
2274 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2275 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2276 p->mm->numa_scan_seq = 0;
2279 if (clone_flags & CLONE_VM)
2280 p->numa_preferred_nid = current->numa_preferred_nid;
2282 p->numa_preferred_nid = -1;
2284 p->node_stamp = 0ULL;
2285 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2286 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2287 p->numa_work.next = &p->numa_work;
2288 p->numa_faults = NULL;
2289 p->last_task_numa_placement = 0;
2290 p->last_sum_exec_runtime = 0;
2292 p->numa_group = NULL;
2293 #endif /* CONFIG_NUMA_BALANCING */
2296 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2298 #ifdef CONFIG_NUMA_BALANCING
2300 void set_numabalancing_state(bool enabled)
2303 static_branch_enable(&sched_numa_balancing);
2305 static_branch_disable(&sched_numa_balancing);
2308 #ifdef CONFIG_PROC_SYSCTL
2309 int sysctl_numa_balancing(struct ctl_table *table, int write,
2310 void __user *buffer, size_t *lenp, loff_t *ppos)
2314 int state = static_branch_likely(&sched_numa_balancing);
2316 if (write && !capable(CAP_SYS_ADMIN))
2321 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2325 set_numabalancing_state(state);
2332 * fork()/clone()-time setup:
2334 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2336 unsigned long flags;
2337 int cpu = get_cpu();
2339 __sched_fork(clone_flags, p);
2341 * We mark the process as running here. This guarantees that
2342 * nobody will actually run it, and a signal or other external
2343 * event cannot wake it up and insert it on the runqueue either.
2345 p->state = TASK_RUNNING;
2348 * Make sure we do not leak PI boosting priority to the child.
2350 p->prio = current->normal_prio;
2353 * Revert to default priority/policy on fork if requested.
2355 if (unlikely(p->sched_reset_on_fork)) {
2356 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2357 p->policy = SCHED_NORMAL;
2358 p->static_prio = NICE_TO_PRIO(0);
2360 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2361 p->static_prio = NICE_TO_PRIO(0);
2363 p->prio = p->normal_prio = __normal_prio(p);
2367 * We don't need the reset flag anymore after the fork. It has
2368 * fulfilled its duty:
2370 p->sched_reset_on_fork = 0;
2373 if (dl_prio(p->prio)) {
2376 } else if (rt_prio(p->prio)) {
2377 p->sched_class = &rt_sched_class;
2379 p->sched_class = &fair_sched_class;
2382 if (p->sched_class->task_fork)
2383 p->sched_class->task_fork(p);
2386 * The child is not yet in the pid-hash so no cgroup attach races,
2387 * and the cgroup is pinned to this child due to cgroup_fork()
2388 * is ran before sched_fork().
2390 * Silence PROVE_RCU.
2392 raw_spin_lock_irqsave(&p->pi_lock, flags);
2393 set_task_cpu(p, cpu);
2394 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2396 #ifdef CONFIG_SCHED_INFO
2397 if (likely(sched_info_on()))
2398 memset(&p->sched_info, 0, sizeof(p->sched_info));
2400 #if defined(CONFIG_SMP)
2403 init_task_preempt_count(p);
2404 #ifdef CONFIG_HAVE_PREEMPT_LAZY
2405 task_thread_info(p)->preempt_lazy_count = 0;
2408 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2409 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2416 unsigned long to_ratio(u64 period, u64 runtime)
2418 if (runtime == RUNTIME_INF)
2422 * Doing this here saves a lot of checks in all
2423 * the calling paths, and returning zero seems
2424 * safe for them anyway.
2429 return div64_u64(runtime << 20, period);
2433 inline struct dl_bw *dl_bw_of(int i)
2435 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2436 "sched RCU must be held");
2437 return &cpu_rq(i)->rd->dl_bw;
2440 static inline int dl_bw_cpus(int i)
2442 struct root_domain *rd = cpu_rq(i)->rd;
2445 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2446 "sched RCU must be held");
2447 for_each_cpu_and(i, rd->span, cpu_active_mask)
2453 inline struct dl_bw *dl_bw_of(int i)
2455 return &cpu_rq(i)->dl.dl_bw;
2458 static inline int dl_bw_cpus(int i)
2465 * We must be sure that accepting a new task (or allowing changing the
2466 * parameters of an existing one) is consistent with the bandwidth
2467 * constraints. If yes, this function also accordingly updates the currently
2468 * allocated bandwidth to reflect the new situation.
2470 * This function is called while holding p's rq->lock.
2472 * XXX we should delay bw change until the task's 0-lag point, see
2475 static int dl_overflow(struct task_struct *p, int policy,
2476 const struct sched_attr *attr)
2479 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2480 u64 period = attr->sched_period ?: attr->sched_deadline;
2481 u64 runtime = attr->sched_runtime;
2482 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2485 if (new_bw == p->dl.dl_bw)
2489 * Either if a task, enters, leave, or stays -deadline but changes
2490 * its parameters, we may need to update accordingly the total
2491 * allocated bandwidth of the container.
2493 raw_spin_lock(&dl_b->lock);
2494 cpus = dl_bw_cpus(task_cpu(p));
2495 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2496 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2497 __dl_add(dl_b, new_bw);
2499 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2500 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2501 __dl_clear(dl_b, p->dl.dl_bw);
2502 __dl_add(dl_b, new_bw);
2504 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2505 __dl_clear(dl_b, p->dl.dl_bw);
2508 raw_spin_unlock(&dl_b->lock);
2513 extern void init_dl_bw(struct dl_bw *dl_b);
2516 * wake_up_new_task - wake up a newly created task for the first time.
2518 * This function will do some initial scheduler statistics housekeeping
2519 * that must be done for every newly created context, then puts the task
2520 * on the runqueue and wakes it.
2522 void wake_up_new_task(struct task_struct *p)
2524 unsigned long flags;
2527 raw_spin_lock_irqsave(&p->pi_lock, flags);
2528 /* Initialize new task's runnable average */
2529 init_entity_runnable_average(&p->se);
2532 * Fork balancing, do it here and not earlier because:
2533 * - cpus_allowed can change in the fork path
2534 * - any previously selected cpu might disappear through hotplug
2536 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2539 rq = __task_rq_lock(p);
2540 activate_task(rq, p, 0);
2541 p->on_rq = TASK_ON_RQ_QUEUED;
2542 trace_sched_wakeup_new(p);
2543 check_preempt_curr(rq, p, WF_FORK);
2545 if (p->sched_class->task_woken) {
2547 * Nothing relies on rq->lock after this, so its fine to
2550 lockdep_unpin_lock(&rq->lock);
2551 p->sched_class->task_woken(rq, p);
2552 lockdep_pin_lock(&rq->lock);
2555 task_rq_unlock(rq, p, &flags);
2558 #ifdef CONFIG_PREEMPT_NOTIFIERS
2560 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2562 void preempt_notifier_inc(void)
2564 static_key_slow_inc(&preempt_notifier_key);
2566 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2568 void preempt_notifier_dec(void)
2570 static_key_slow_dec(&preempt_notifier_key);
2572 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2575 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2576 * @notifier: notifier struct to register
2578 void preempt_notifier_register(struct preempt_notifier *notifier)
2580 if (!static_key_false(&preempt_notifier_key))
2581 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2583 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2585 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2588 * preempt_notifier_unregister - no longer interested in preemption notifications
2589 * @notifier: notifier struct to unregister
2591 * This is *not* safe to call from within a preemption notifier.
2593 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2595 hlist_del(¬ifier->link);
2597 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2599 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2601 struct preempt_notifier *notifier;
2603 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2604 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2607 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2609 if (static_key_false(&preempt_notifier_key))
2610 __fire_sched_in_preempt_notifiers(curr);
2614 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2615 struct task_struct *next)
2617 struct preempt_notifier *notifier;
2619 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2620 notifier->ops->sched_out(notifier, next);
2623 static __always_inline void
2624 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2625 struct task_struct *next)
2627 if (static_key_false(&preempt_notifier_key))
2628 __fire_sched_out_preempt_notifiers(curr, next);
2631 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2633 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2638 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2639 struct task_struct *next)
2643 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2646 * prepare_task_switch - prepare to switch tasks
2647 * @rq: the runqueue preparing to switch
2648 * @prev: the current task that is being switched out
2649 * @next: the task we are going to switch to.
2651 * This is called with the rq lock held and interrupts off. It must
2652 * be paired with a subsequent finish_task_switch after the context
2655 * prepare_task_switch sets up locking and calls architecture specific
2659 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2660 struct task_struct *next)
2662 sched_info_switch(rq, prev, next);
2663 perf_event_task_sched_out(prev, next);
2664 fire_sched_out_preempt_notifiers(prev, next);
2665 prepare_lock_switch(rq, next);
2666 prepare_arch_switch(next);
2670 * finish_task_switch - clean up after a task-switch
2671 * @prev: the thread we just switched away from.
2673 * finish_task_switch must be called after the context switch, paired
2674 * with a prepare_task_switch call before the context switch.
2675 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2676 * and do any other architecture-specific cleanup actions.
2678 * Note that we may have delayed dropping an mm in context_switch(). If
2679 * so, we finish that here outside of the runqueue lock. (Doing it
2680 * with the lock held can cause deadlocks; see schedule() for
2683 * The context switch have flipped the stack from under us and restored the
2684 * local variables which were saved when this task called schedule() in the
2685 * past. prev == current is still correct but we need to recalculate this_rq
2686 * because prev may have moved to another CPU.
2688 static struct rq *finish_task_switch(struct task_struct *prev)
2689 __releases(rq->lock)
2691 struct rq *rq = this_rq();
2692 struct mm_struct *mm = rq->prev_mm;
2696 * The previous task will have left us with a preempt_count of 2
2697 * because it left us after:
2700 * preempt_disable(); // 1
2702 * raw_spin_lock_irq(&rq->lock) // 2
2704 * Also, see FORK_PREEMPT_COUNT.
2706 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2707 "corrupted preempt_count: %s/%d/0x%x\n",
2708 current->comm, current->pid, preempt_count()))
2709 preempt_count_set(FORK_PREEMPT_COUNT);
2714 * A task struct has one reference for the use as "current".
2715 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2716 * schedule one last time. The schedule call will never return, and
2717 * the scheduled task must drop that reference.
2719 * We must observe prev->state before clearing prev->on_cpu (in
2720 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2721 * running on another CPU and we could rave with its RUNNING -> DEAD
2722 * transition, resulting in a double drop.
2724 prev_state = prev->state;
2725 vtime_task_switch(prev);
2726 perf_event_task_sched_in(prev, current);
2727 finish_lock_switch(rq, prev);
2728 finish_arch_post_lock_switch();
2730 fire_sched_in_preempt_notifiers(current);
2732 * We use mmdrop_delayed() here so we don't have to do the
2733 * full __mmdrop() when we are the last user.
2737 if (unlikely(prev_state == TASK_DEAD)) {
2738 if (prev->sched_class->task_dead)
2739 prev->sched_class->task_dead(prev);
2742 * Remove function-return probe instances associated with this
2743 * task and put them back on the free list.
2745 kprobe_flush_task(prev);
2746 put_task_struct(prev);
2749 tick_nohz_task_switch();
2755 /* rq->lock is NOT held, but preemption is disabled */
2756 static void __balance_callback(struct rq *rq)
2758 struct callback_head *head, *next;
2759 void (*func)(struct rq *rq);
2760 unsigned long flags;
2762 raw_spin_lock_irqsave(&rq->lock, flags);
2763 head = rq->balance_callback;
2764 rq->balance_callback = NULL;
2766 func = (void (*)(struct rq *))head->func;
2773 raw_spin_unlock_irqrestore(&rq->lock, flags);
2776 static inline void balance_callback(struct rq *rq)
2778 if (unlikely(rq->balance_callback))
2779 __balance_callback(rq);
2784 static inline void balance_callback(struct rq *rq)
2791 * schedule_tail - first thing a freshly forked thread must call.
2792 * @prev: the thread we just switched away from.
2794 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2795 __releases(rq->lock)
2800 * New tasks start with FORK_PREEMPT_COUNT, see there and
2801 * finish_task_switch() for details.
2803 * finish_task_switch() will drop rq->lock() and lower preempt_count
2804 * and the preempt_enable() will end up enabling preemption (on
2805 * PREEMPT_COUNT kernels).
2808 rq = finish_task_switch(prev);
2809 balance_callback(rq);
2812 if (current->set_child_tid)
2813 put_user(task_pid_vnr(current), current->set_child_tid);
2817 * context_switch - switch to the new MM and the new thread's register state.
2819 static inline struct rq *
2820 context_switch(struct rq *rq, struct task_struct *prev,
2821 struct task_struct *next)
2823 struct mm_struct *mm, *oldmm;
2825 prepare_task_switch(rq, prev, next);
2828 oldmm = prev->active_mm;
2830 * For paravirt, this is coupled with an exit in switch_to to
2831 * combine the page table reload and the switch backend into
2834 arch_start_context_switch(prev);
2837 next->active_mm = oldmm;
2838 atomic_inc(&oldmm->mm_count);
2839 enter_lazy_tlb(oldmm, next);
2841 switch_mm(oldmm, mm, next);
2844 prev->active_mm = NULL;
2845 rq->prev_mm = oldmm;
2848 * Since the runqueue lock will be released by the next
2849 * task (which is an invalid locking op but in the case
2850 * of the scheduler it's an obvious special-case), so we
2851 * do an early lockdep release here:
2853 lockdep_unpin_lock(&rq->lock);
2854 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2856 /* Here we just switch the register state and the stack. */
2857 switch_to(prev, next, prev);
2860 return finish_task_switch(prev);
2864 * nr_running and nr_context_switches:
2866 * externally visible scheduler statistics: current number of runnable
2867 * threads, total number of context switches performed since bootup.
2869 unsigned long nr_running(void)
2871 unsigned long i, sum = 0;
2873 for_each_online_cpu(i)
2874 sum += cpu_rq(i)->nr_running;
2880 * Check if only the current task is running on the cpu.
2882 * Caution: this function does not check that the caller has disabled
2883 * preemption, thus the result might have a time-of-check-to-time-of-use
2884 * race. The caller is responsible to use it correctly, for example:
2886 * - from a non-preemptable section (of course)
2888 * - from a thread that is bound to a single CPU
2890 * - in a loop with very short iterations (e.g. a polling loop)
2892 bool single_task_running(void)
2894 return raw_rq()->nr_running == 1;
2896 EXPORT_SYMBOL(single_task_running);
2898 unsigned long long nr_context_switches(void)
2901 unsigned long long sum = 0;
2903 for_each_possible_cpu(i)
2904 sum += cpu_rq(i)->nr_switches;
2909 unsigned long nr_iowait(void)
2911 unsigned long i, sum = 0;
2913 for_each_possible_cpu(i)
2914 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2919 unsigned long nr_iowait_cpu(int cpu)
2921 struct rq *this = cpu_rq(cpu);
2922 return atomic_read(&this->nr_iowait);
2925 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2927 struct rq *rq = this_rq();
2928 *nr_waiters = atomic_read(&rq->nr_iowait);
2929 *load = rq->load.weight;
2935 * sched_exec - execve() is a valuable balancing opportunity, because at
2936 * this point the task has the smallest effective memory and cache footprint.
2938 void sched_exec(void)
2940 struct task_struct *p = current;
2941 unsigned long flags;
2944 raw_spin_lock_irqsave(&p->pi_lock, flags);
2945 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2946 if (dest_cpu == smp_processor_id())
2949 if (likely(cpu_active(dest_cpu))) {
2950 struct migration_arg arg = { p, dest_cpu };
2952 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2953 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2957 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2962 DEFINE_PER_CPU(struct kernel_stat, kstat);
2963 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2965 EXPORT_PER_CPU_SYMBOL(kstat);
2966 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2969 * Return accounted runtime for the task.
2970 * In case the task is currently running, return the runtime plus current's
2971 * pending runtime that have not been accounted yet.
2973 unsigned long long task_sched_runtime(struct task_struct *p)
2975 unsigned long flags;
2979 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2981 * 64-bit doesn't need locks to atomically read a 64bit value.
2982 * So we have a optimization chance when the task's delta_exec is 0.
2983 * Reading ->on_cpu is racy, but this is ok.
2985 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2986 * If we race with it entering cpu, unaccounted time is 0. This is
2987 * indistinguishable from the read occurring a few cycles earlier.
2988 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2989 * been accounted, so we're correct here as well.
2991 if (!p->on_cpu || !task_on_rq_queued(p))
2992 return p->se.sum_exec_runtime;
2995 rq = task_rq_lock(p, &flags);
2997 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2998 * project cycles that may never be accounted to this
2999 * thread, breaking clock_gettime().
3001 if (task_current(rq, p) && task_on_rq_queued(p)) {
3002 update_rq_clock(rq);
3003 p->sched_class->update_curr(rq);
3005 ns = p->se.sum_exec_runtime;
3006 task_rq_unlock(rq, p, &flags);
3012 * This function gets called by the timer code, with HZ frequency.
3013 * We call it with interrupts disabled.
3015 void scheduler_tick(void)
3017 int cpu = smp_processor_id();
3018 struct rq *rq = cpu_rq(cpu);
3019 struct task_struct *curr = rq->curr;
3023 raw_spin_lock(&rq->lock);
3024 update_rq_clock(rq);
3025 curr->sched_class->task_tick(rq, curr, 0);
3026 update_cpu_load_active(rq);
3027 calc_global_load_tick(rq);
3028 raw_spin_unlock(&rq->lock);
3030 perf_event_task_tick();
3033 rq->idle_balance = idle_cpu(cpu);
3034 trigger_load_balance(rq);
3036 rq_last_tick_reset(rq);
3039 #ifdef CONFIG_NO_HZ_FULL
3041 * scheduler_tick_max_deferment
3043 * Keep at least one tick per second when a single
3044 * active task is running because the scheduler doesn't
3045 * yet completely support full dynticks environment.
3047 * This makes sure that uptime, CFS vruntime, load
3048 * balancing, etc... continue to move forward, even
3049 * with a very low granularity.
3051 * Return: Maximum deferment in nanoseconds.
3053 u64 scheduler_tick_max_deferment(void)
3055 struct rq *rq = this_rq();
3056 unsigned long next, now = READ_ONCE(jiffies);
3058 next = rq->last_sched_tick + HZ;
3060 if (time_before_eq(next, now))
3063 return jiffies_to_nsecs(next - now);
3067 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3068 defined(CONFIG_PREEMPT_TRACER))
3070 void preempt_count_add(int val)
3072 #ifdef CONFIG_DEBUG_PREEMPT
3076 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3079 __preempt_count_add(val);
3080 #ifdef CONFIG_DEBUG_PREEMPT
3082 * Spinlock count overflowing soon?
3084 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3087 if (preempt_count() == val) {
3088 unsigned long ip = get_lock_parent_ip();
3089 #ifdef CONFIG_DEBUG_PREEMPT
3090 current->preempt_disable_ip = ip;
3092 trace_preempt_off(CALLER_ADDR0, ip);
3095 EXPORT_SYMBOL(preempt_count_add);
3096 NOKPROBE_SYMBOL(preempt_count_add);
3098 void preempt_count_sub(int val)
3100 #ifdef CONFIG_DEBUG_PREEMPT
3104 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3107 * Is the spinlock portion underflowing?
3109 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3110 !(preempt_count() & PREEMPT_MASK)))
3114 if (preempt_count() == val)
3115 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
3116 __preempt_count_sub(val);
3118 EXPORT_SYMBOL(preempt_count_sub);
3119 NOKPROBE_SYMBOL(preempt_count_sub);
3124 * Print scheduling while atomic bug:
3126 static noinline void __schedule_bug(struct task_struct *prev)
3128 if (oops_in_progress)
3131 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3132 prev->comm, prev->pid, preempt_count());
3134 debug_show_held_locks(prev);
3136 if (irqs_disabled())
3137 print_irqtrace_events(prev);
3138 #ifdef CONFIG_DEBUG_PREEMPT
3139 if (in_atomic_preempt_off()) {
3140 pr_err("Preemption disabled at:");
3141 print_ip_sym(current->preempt_disable_ip);
3146 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3150 * Various schedule()-time debugging checks and statistics:
3152 static inline void schedule_debug(struct task_struct *prev)
3154 #ifdef CONFIG_SCHED_STACK_END_CHECK
3155 if (task_stack_end_corrupted(prev))
3156 panic("corrupted stack end detected inside scheduler\n");
3159 if (unlikely(in_atomic_preempt_off())) {
3160 __schedule_bug(prev);
3161 preempt_count_set(PREEMPT_DISABLED);
3165 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3167 schedstat_inc(this_rq(), sched_count);
3170 #if defined(CONFIG_PREEMPT_RT_FULL) && defined(CONFIG_SMP)
3172 void migrate_disable(void)
3174 struct task_struct *p = current;
3176 if (in_atomic() || irqs_disabled()) {
3177 #ifdef CONFIG_SCHED_DEBUG
3178 p->migrate_disable_atomic++;
3183 #ifdef CONFIG_SCHED_DEBUG
3184 if (unlikely(p->migrate_disable_atomic)) {
3190 if (p->migrate_disable) {
3191 p->migrate_disable++;
3196 preempt_lazy_disable();
3198 p->migrate_disable = 1;
3201 EXPORT_SYMBOL(migrate_disable);
3203 void migrate_enable(void)
3205 struct task_struct *p = current;
3207 if (in_atomic() || irqs_disabled()) {
3208 #ifdef CONFIG_SCHED_DEBUG
3209 p->migrate_disable_atomic--;
3214 #ifdef CONFIG_SCHED_DEBUG
3215 if (unlikely(p->migrate_disable_atomic)) {
3220 WARN_ON_ONCE(p->migrate_disable <= 0);
3222 if (p->migrate_disable > 1) {
3223 p->migrate_disable--;
3229 * Clearing migrate_disable causes tsk_cpus_allowed to
3230 * show the tasks original cpu affinity.
3232 p->migrate_disable = 0;
3234 unpin_current_cpu();
3236 preempt_lazy_enable();
3238 EXPORT_SYMBOL(migrate_enable);
3242 * Pick up the highest-prio task:
3244 static inline struct task_struct *
3245 pick_next_task(struct rq *rq, struct task_struct *prev)
3247 const struct sched_class *class = &fair_sched_class;
3248 struct task_struct *p;
3251 * Optimization: we know that if all tasks are in
3252 * the fair class we can call that function directly:
3254 if (likely(prev->sched_class == class &&
3255 rq->nr_running == rq->cfs.h_nr_running)) {
3256 p = fair_sched_class.pick_next_task(rq, prev);
3257 if (unlikely(p == RETRY_TASK))
3260 /* assumes fair_sched_class->next == idle_sched_class */
3262 p = idle_sched_class.pick_next_task(rq, prev);
3268 for_each_class(class) {
3269 p = class->pick_next_task(rq, prev);
3271 if (unlikely(p == RETRY_TASK))
3277 BUG(); /* the idle class will always have a runnable task */
3281 * __schedule() is the main scheduler function.
3283 * The main means of driving the scheduler and thus entering this function are:
3285 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3287 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3288 * paths. For example, see arch/x86/entry_64.S.
3290 * To drive preemption between tasks, the scheduler sets the flag in timer
3291 * interrupt handler scheduler_tick().
3293 * 3. Wakeups don't really cause entry into schedule(). They add a
3294 * task to the run-queue and that's it.
3296 * Now, if the new task added to the run-queue preempts the current
3297 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3298 * called on the nearest possible occasion:
3300 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3302 * - in syscall or exception context, at the next outmost
3303 * preempt_enable(). (this might be as soon as the wake_up()'s
3306 * - in IRQ context, return from interrupt-handler to
3307 * preemptible context
3309 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3312 * - cond_resched() call
3313 * - explicit schedule() call
3314 * - return from syscall or exception to user-space
3315 * - return from interrupt-handler to user-space
3317 * WARNING: must be called with preemption disabled!
3319 static void __sched notrace __schedule(bool preempt)
3321 struct task_struct *prev, *next;
3322 unsigned long *switch_count;
3326 cpu = smp_processor_id();
3328 rcu_note_context_switch();
3332 * do_exit() calls schedule() with preemption disabled as an exception;
3333 * however we must fix that up, otherwise the next task will see an
3334 * inconsistent (higher) preempt count.
3336 * It also avoids the below schedule_debug() test from complaining
3339 if (unlikely(prev->state == TASK_DEAD))
3340 preempt_enable_no_resched_notrace();
3342 schedule_debug(prev);
3344 if (sched_feat(HRTICK))
3348 * Make sure that signal_pending_state()->signal_pending() below
3349 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3350 * done by the caller to avoid the race with signal_wake_up().
3352 smp_mb__before_spinlock();
3353 raw_spin_lock_irq(&rq->lock);
3354 lockdep_pin_lock(&rq->lock);
3356 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3358 switch_count = &prev->nivcsw;
3359 if (!preempt && prev->state) {
3360 if (unlikely(signal_pending_state(prev->state, prev))) {
3361 prev->state = TASK_RUNNING;
3363 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3366 switch_count = &prev->nvcsw;
3369 if (task_on_rq_queued(prev))
3370 update_rq_clock(rq);
3372 next = pick_next_task(rq, prev);
3373 clear_tsk_need_resched(prev);
3374 clear_tsk_need_resched_lazy(prev);
3375 clear_preempt_need_resched();
3376 rq->clock_skip_update = 0;
3378 if (likely(prev != next)) {
3383 trace_sched_switch(preempt, prev, next);
3384 rq = context_switch(rq, prev, next); /* unlocks the rq */
3387 lockdep_unpin_lock(&rq->lock);
3388 raw_spin_unlock_irq(&rq->lock);
3391 balance_callback(rq);
3394 static inline void sched_submit_work(struct task_struct *tsk)
3399 * If a worker went to sleep, notify and ask workqueue whether
3400 * it wants to wake up a task to maintain concurrency.
3402 if (tsk->flags & PF_WQ_WORKER)
3403 wq_worker_sleeping(tsk);
3406 if (tsk_is_pi_blocked(tsk))
3410 * If we are going to sleep and we have plugged IO queued,
3411 * make sure to submit it to avoid deadlocks.
3413 if (blk_needs_flush_plug(tsk))
3414 blk_schedule_flush_plug(tsk);
3417 static void sched_update_worker(struct task_struct *tsk)
3419 if (tsk->flags & PF_WQ_WORKER)
3420 wq_worker_running(tsk);
3423 asmlinkage __visible void __sched schedule(void)
3425 struct task_struct *tsk = current;
3427 sched_submit_work(tsk);
3431 sched_preempt_enable_no_resched();
3432 } while (need_resched());
3433 sched_update_worker(tsk);
3435 EXPORT_SYMBOL(schedule);
3437 #ifdef CONFIG_CONTEXT_TRACKING
3438 asmlinkage __visible void __sched schedule_user(void)
3441 * If we come here after a random call to set_need_resched(),
3442 * or we have been woken up remotely but the IPI has not yet arrived,
3443 * we haven't yet exited the RCU idle mode. Do it here manually until
3444 * we find a better solution.
3446 * NB: There are buggy callers of this function. Ideally we
3447 * should warn if prev_state != CONTEXT_USER, but that will trigger
3448 * too frequently to make sense yet.
3450 enum ctx_state prev_state = exception_enter();
3452 exception_exit(prev_state);
3457 * schedule_preempt_disabled - called with preemption disabled
3459 * Returns with preemption disabled. Note: preempt_count must be 1
3461 void __sched schedule_preempt_disabled(void)
3463 sched_preempt_enable_no_resched();
3468 static void __sched notrace preempt_schedule_common(void)
3471 preempt_disable_notrace();
3473 preempt_enable_no_resched_notrace();
3476 * Check again in case we missed a preemption opportunity
3477 * between schedule and now.
3479 } while (need_resched());
3482 #ifdef CONFIG_PREEMPT_LAZY
3484 * If TIF_NEED_RESCHED is then we allow to be scheduled away since this is
3485 * set by a RT task. Oterwise we try to avoid beeing scheduled out as long as
3486 * preempt_lazy_count counter >0.
3488 static __always_inline int preemptible_lazy(void)
3490 if (test_thread_flag(TIF_NEED_RESCHED))
3492 if (current_thread_info()->preempt_lazy_count)
3499 static inline int preemptible_lazy(void)
3506 #ifdef CONFIG_PREEMPT
3508 * this is the entry point to schedule() from in-kernel preemption
3509 * off of preempt_enable. Kernel preemptions off return from interrupt
3510 * occur there and call schedule directly.
3512 asmlinkage __visible void __sched notrace preempt_schedule(void)
3515 * If there is a non-zero preempt_count or interrupts are disabled,
3516 * we do not want to preempt the current task. Just return..
3518 if (likely(!preemptible()))
3520 if (!preemptible_lazy())
3523 preempt_schedule_common();
3525 NOKPROBE_SYMBOL(preempt_schedule);
3526 EXPORT_SYMBOL(preempt_schedule);
3529 * preempt_schedule_notrace - preempt_schedule called by tracing
3531 * The tracing infrastructure uses preempt_enable_notrace to prevent
3532 * recursion and tracing preempt enabling caused by the tracing
3533 * infrastructure itself. But as tracing can happen in areas coming
3534 * from userspace or just about to enter userspace, a preempt enable
3535 * can occur before user_exit() is called. This will cause the scheduler
3536 * to be called when the system is still in usermode.
3538 * To prevent this, the preempt_enable_notrace will use this function
3539 * instead of preempt_schedule() to exit user context if needed before
3540 * calling the scheduler.
3542 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3544 enum ctx_state prev_ctx;
3546 if (likely(!preemptible()))
3548 if (!preemptible_lazy())
3552 preempt_disable_notrace();
3554 * Needs preempt disabled in case user_exit() is traced
3555 * and the tracer calls preempt_enable_notrace() causing
3556 * an infinite recursion.
3558 prev_ctx = exception_enter();
3560 * The add/subtract must not be traced by the function
3561 * tracer. But we still want to account for the
3562 * preempt off latency tracer. Since the _notrace versions
3563 * of add/subtract skip the accounting for latency tracer
3564 * we must force it manually.
3566 start_critical_timings();
3568 stop_critical_timings();
3569 exception_exit(prev_ctx);
3571 preempt_enable_no_resched_notrace();
3572 } while (need_resched());
3574 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3576 #endif /* CONFIG_PREEMPT */
3579 * this is the entry point to schedule() from kernel preemption
3580 * off of irq context.
3581 * Note, that this is called and return with irqs disabled. This will
3582 * protect us against recursive calling from irq.
3584 asmlinkage __visible void __sched preempt_schedule_irq(void)
3586 enum ctx_state prev_state;
3588 /* Catch callers which need to be fixed */
3589 BUG_ON(preempt_count() || !irqs_disabled());
3591 prev_state = exception_enter();
3597 local_irq_disable();
3598 sched_preempt_enable_no_resched();
3599 } while (need_resched());
3601 exception_exit(prev_state);
3604 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3607 return try_to_wake_up(curr->private, mode, wake_flags);
3609 EXPORT_SYMBOL(default_wake_function);
3611 #ifdef CONFIG_RT_MUTEXES
3614 * rt_mutex_setprio - set the current priority of a task
3616 * @prio: prio value (kernel-internal form)
3618 * This function changes the 'effective' priority of a task. It does
3619 * not touch ->normal_prio like __setscheduler().
3621 * Used by the rt_mutex code to implement priority inheritance
3622 * logic. Call site only calls if the priority of the task changed.
3624 void rt_mutex_setprio(struct task_struct *p, int prio)
3626 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3628 const struct sched_class *prev_class;
3630 BUG_ON(prio > MAX_PRIO);
3632 rq = __task_rq_lock(p);
3635 * Idle task boosting is a nono in general. There is one
3636 * exception, when PREEMPT_RT and NOHZ is active:
3638 * The idle task calls get_next_timer_interrupt() and holds
3639 * the timer wheel base->lock on the CPU and another CPU wants
3640 * to access the timer (probably to cancel it). We can safely
3641 * ignore the boosting request, as the idle CPU runs this code
3642 * with interrupts disabled and will complete the lock
3643 * protected section without being interrupted. So there is no
3644 * real need to boost.
3646 if (unlikely(p == rq->idle)) {
3647 WARN_ON(p != rq->curr);
3648 WARN_ON(p->pi_blocked_on);
3652 trace_sched_pi_setprio(p, prio);
3654 prev_class = p->sched_class;
3655 queued = task_on_rq_queued(p);
3656 running = task_current(rq, p);
3658 dequeue_task(rq, p, DEQUEUE_SAVE);
3660 put_prev_task(rq, p);
3663 * Boosting condition are:
3664 * 1. -rt task is running and holds mutex A
3665 * --> -dl task blocks on mutex A
3667 * 2. -dl task is running and holds mutex A
3668 * --> -dl task blocks on mutex A and could preempt the
3671 if (dl_prio(prio)) {
3672 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3673 if (!dl_prio(p->normal_prio) ||
3674 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3675 p->dl.dl_boosted = 1;
3676 enqueue_flag |= ENQUEUE_REPLENISH;
3678 p->dl.dl_boosted = 0;
3679 p->sched_class = &dl_sched_class;
3680 } else if (rt_prio(prio)) {
3681 if (dl_prio(oldprio))
3682 p->dl.dl_boosted = 0;
3684 enqueue_flag |= ENQUEUE_HEAD;
3685 p->sched_class = &rt_sched_class;
3687 if (dl_prio(oldprio))
3688 p->dl.dl_boosted = 0;
3689 if (rt_prio(oldprio))
3691 p->sched_class = &fair_sched_class;
3697 p->sched_class->set_curr_task(rq);
3699 enqueue_task(rq, p, enqueue_flag);
3701 check_class_changed(rq, p, prev_class, oldprio);
3703 preempt_disable(); /* avoid rq from going away on us */
3704 __task_rq_unlock(rq);
3706 balance_callback(rq);
3711 void set_user_nice(struct task_struct *p, long nice)
3713 int old_prio, delta, queued;
3714 unsigned long flags;
3717 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3720 * We have to be careful, if called from sys_setpriority(),
3721 * the task might be in the middle of scheduling on another CPU.
3723 rq = task_rq_lock(p, &flags);
3725 * The RT priorities are set via sched_setscheduler(), but we still
3726 * allow the 'normal' nice value to be set - but as expected
3727 * it wont have any effect on scheduling until the task is
3728 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3730 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3731 p->static_prio = NICE_TO_PRIO(nice);
3734 queued = task_on_rq_queued(p);
3736 dequeue_task(rq, p, DEQUEUE_SAVE);
3738 p->static_prio = NICE_TO_PRIO(nice);
3741 p->prio = effective_prio(p);
3742 delta = p->prio - old_prio;
3745 enqueue_task(rq, p, ENQUEUE_RESTORE);
3747 * If the task increased its priority or is running and
3748 * lowered its priority, then reschedule its CPU:
3750 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3754 task_rq_unlock(rq, p, &flags);
3756 EXPORT_SYMBOL(set_user_nice);
3759 * can_nice - check if a task can reduce its nice value
3763 int can_nice(const struct task_struct *p, const int nice)
3765 /* convert nice value [19,-20] to rlimit style value [1,40] */
3766 int nice_rlim = nice_to_rlimit(nice);
3768 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3769 capable(CAP_SYS_NICE));
3772 #ifdef __ARCH_WANT_SYS_NICE
3775 * sys_nice - change the priority of the current process.
3776 * @increment: priority increment
3778 * sys_setpriority is a more generic, but much slower function that
3779 * does similar things.
3781 SYSCALL_DEFINE1(nice, int, increment)
3786 * Setpriority might change our priority at the same moment.
3787 * We don't have to worry. Conceptually one call occurs first
3788 * and we have a single winner.
3790 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3791 nice = task_nice(current) + increment;
3793 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3794 if (increment < 0 && !can_nice(current, nice))
3797 retval = security_task_setnice(current, nice);
3801 set_user_nice(current, nice);
3808 * task_prio - return the priority value of a given task.
3809 * @p: the task in question.
3811 * Return: The priority value as seen by users in /proc.
3812 * RT tasks are offset by -200. Normal tasks are centered
3813 * around 0, value goes from -16 to +15.
3815 int task_prio(const struct task_struct *p)
3817 return p->prio - MAX_RT_PRIO;
3821 * idle_cpu - is a given cpu idle currently?
3822 * @cpu: the processor in question.
3824 * Return: 1 if the CPU is currently idle. 0 otherwise.
3826 int idle_cpu(int cpu)
3828 struct rq *rq = cpu_rq(cpu);
3830 if (rq->curr != rq->idle)
3837 if (!llist_empty(&rq->wake_list))
3845 * idle_task - return the idle task for a given cpu.
3846 * @cpu: the processor in question.
3848 * Return: The idle task for the cpu @cpu.
3850 struct task_struct *idle_task(int cpu)
3852 return cpu_rq(cpu)->idle;
3856 * find_process_by_pid - find a process with a matching PID value.
3857 * @pid: the pid in question.
3859 * The task of @pid, if found. %NULL otherwise.
3861 static struct task_struct *find_process_by_pid(pid_t pid)
3863 return pid ? find_task_by_vpid(pid) : current;
3867 * This function initializes the sched_dl_entity of a newly becoming
3868 * SCHED_DEADLINE task.
3870 * Only the static values are considered here, the actual runtime and the
3871 * absolute deadline will be properly calculated when the task is enqueued
3872 * for the first time with its new policy.
3875 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3877 struct sched_dl_entity *dl_se = &p->dl;
3879 dl_se->dl_runtime = attr->sched_runtime;
3880 dl_se->dl_deadline = attr->sched_deadline;
3881 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3882 dl_se->flags = attr->sched_flags;
3883 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3886 * Changing the parameters of a task is 'tricky' and we're not doing
3887 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3889 * What we SHOULD do is delay the bandwidth release until the 0-lag
3890 * point. This would include retaining the task_struct until that time
3891 * and change dl_overflow() to not immediately decrement the current
3894 * Instead we retain the current runtime/deadline and let the new
3895 * parameters take effect after the current reservation period lapses.
3896 * This is safe (albeit pessimistic) because the 0-lag point is always
3897 * before the current scheduling deadline.
3899 * We can still have temporary overloads because we do not delay the
3900 * change in bandwidth until that time; so admission control is
3901 * not on the safe side. It does however guarantee tasks will never
3902 * consume more than promised.
3907 * sched_setparam() passes in -1 for its policy, to let the functions
3908 * it calls know not to change it.
3910 #define SETPARAM_POLICY -1
3912 static void __setscheduler_params(struct task_struct *p,
3913 const struct sched_attr *attr)
3915 int policy = attr->sched_policy;
3917 if (policy == SETPARAM_POLICY)
3922 if (dl_policy(policy))
3923 __setparam_dl(p, attr);
3924 else if (fair_policy(policy))
3925 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3928 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3929 * !rt_policy. Always setting this ensures that things like
3930 * getparam()/getattr() don't report silly values for !rt tasks.
3932 p->rt_priority = attr->sched_priority;
3933 p->normal_prio = normal_prio(p);
3937 /* Actually do priority change: must hold pi & rq lock. */
3938 static void __setscheduler(struct rq *rq, struct task_struct *p,
3939 const struct sched_attr *attr, bool keep_boost)
3941 __setscheduler_params(p, attr);
3944 * Keep a potential priority boosting if called from
3945 * sched_setscheduler().
3948 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3950 p->prio = normal_prio(p);
3952 if (dl_prio(p->prio))
3953 p->sched_class = &dl_sched_class;
3954 else if (rt_prio(p->prio))
3955 p->sched_class = &rt_sched_class;
3957 p->sched_class = &fair_sched_class;
3961 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3963 struct sched_dl_entity *dl_se = &p->dl;
3965 attr->sched_priority = p->rt_priority;
3966 attr->sched_runtime = dl_se->dl_runtime;
3967 attr->sched_deadline = dl_se->dl_deadline;
3968 attr->sched_period = dl_se->dl_period;
3969 attr->sched_flags = dl_se->flags;
3973 * This function validates the new parameters of a -deadline task.
3974 * We ask for the deadline not being zero, and greater or equal
3975 * than the runtime, as well as the period of being zero or
3976 * greater than deadline. Furthermore, we have to be sure that
3977 * user parameters are above the internal resolution of 1us (we
3978 * check sched_runtime only since it is always the smaller one) and
3979 * below 2^63 ns (we have to check both sched_deadline and
3980 * sched_period, as the latter can be zero).
3983 __checkparam_dl(const struct sched_attr *attr)
3986 if (attr->sched_deadline == 0)
3990 * Since we truncate DL_SCALE bits, make sure we're at least
3993 if (attr->sched_runtime < (1ULL << DL_SCALE))
3997 * Since we use the MSB for wrap-around and sign issues, make
3998 * sure it's not set (mind that period can be equal to zero).
4000 if (attr->sched_deadline & (1ULL << 63) ||
4001 attr->sched_period & (1ULL << 63))
4004 /* runtime <= deadline <= period (if period != 0) */
4005 if ((attr->sched_period != 0 &&
4006 attr->sched_period < attr->sched_deadline) ||
4007 attr->sched_deadline < attr->sched_runtime)
4014 * check the target process has a UID that matches the current process's
4016 static bool check_same_owner(struct task_struct *p)
4018 const struct cred *cred = current_cred(), *pcred;
4022 pcred = __task_cred(p);
4023 match = (uid_eq(cred->euid, pcred->euid) ||
4024 uid_eq(cred->euid, pcred->uid));
4029 static bool dl_param_changed(struct task_struct *p,
4030 const struct sched_attr *attr)
4032 struct sched_dl_entity *dl_se = &p->dl;
4034 if (dl_se->dl_runtime != attr->sched_runtime ||
4035 dl_se->dl_deadline != attr->sched_deadline ||
4036 dl_se->dl_period != attr->sched_period ||
4037 dl_se->flags != attr->sched_flags)
4043 static int __sched_setscheduler(struct task_struct *p,
4044 const struct sched_attr *attr,
4047 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4048 MAX_RT_PRIO - 1 - attr->sched_priority;
4049 int retval, oldprio, oldpolicy = -1, queued, running;
4050 int new_effective_prio, policy = attr->sched_policy;
4051 unsigned long flags;
4052 const struct sched_class *prev_class;
4056 /* may grab non-irq protected spin_locks */
4057 BUG_ON(in_interrupt());
4059 /* double check policy once rq lock held */
4061 reset_on_fork = p->sched_reset_on_fork;
4062 policy = oldpolicy = p->policy;
4064 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4066 if (!valid_policy(policy))
4070 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4074 * Valid priorities for SCHED_FIFO and SCHED_RR are
4075 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4076 * SCHED_BATCH and SCHED_IDLE is 0.
4078 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4079 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4081 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4082 (rt_policy(policy) != (attr->sched_priority != 0)))
4086 * Allow unprivileged RT tasks to decrease priority:
4088 if (user && !capable(CAP_SYS_NICE)) {
4089 if (fair_policy(policy)) {
4090 if (attr->sched_nice < task_nice(p) &&
4091 !can_nice(p, attr->sched_nice))
4095 if (rt_policy(policy)) {
4096 unsigned long rlim_rtprio =
4097 task_rlimit(p, RLIMIT_RTPRIO);
4099 /* can't set/change the rt policy */
4100 if (policy != p->policy && !rlim_rtprio)
4103 /* can't increase priority */
4104 if (attr->sched_priority > p->rt_priority &&
4105 attr->sched_priority > rlim_rtprio)
4110 * Can't set/change SCHED_DEADLINE policy at all for now
4111 * (safest behavior); in the future we would like to allow
4112 * unprivileged DL tasks to increase their relative deadline
4113 * or reduce their runtime (both ways reducing utilization)
4115 if (dl_policy(policy))
4119 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4120 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4122 if (idle_policy(p->policy) && !idle_policy(policy)) {
4123 if (!can_nice(p, task_nice(p)))
4127 /* can't change other user's priorities */
4128 if (!check_same_owner(p))
4131 /* Normal users shall not reset the sched_reset_on_fork flag */
4132 if (p->sched_reset_on_fork && !reset_on_fork)
4137 retval = security_task_setscheduler(p);
4143 * make sure no PI-waiters arrive (or leave) while we are
4144 * changing the priority of the task:
4146 * To be able to change p->policy safely, the appropriate
4147 * runqueue lock must be held.
4149 rq = task_rq_lock(p, &flags);
4152 * Changing the policy of the stop threads its a very bad idea
4154 if (p == rq->stop) {
4155 task_rq_unlock(rq, p, &flags);
4160 * If not changing anything there's no need to proceed further,
4161 * but store a possible modification of reset_on_fork.
4163 if (unlikely(policy == p->policy)) {
4164 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4166 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4168 if (dl_policy(policy) && dl_param_changed(p, attr))
4171 p->sched_reset_on_fork = reset_on_fork;
4172 task_rq_unlock(rq, p, &flags);
4178 #ifdef CONFIG_RT_GROUP_SCHED
4180 * Do not allow realtime tasks into groups that have no runtime
4183 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4184 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4185 !task_group_is_autogroup(task_group(p))) {
4186 task_rq_unlock(rq, p, &flags);
4191 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4192 cpumask_t *span = rq->rd->span;
4195 * Don't allow tasks with an affinity mask smaller than
4196 * the entire root_domain to become SCHED_DEADLINE. We
4197 * will also fail if there's no bandwidth available.
4199 if (!cpumask_subset(span, &p->cpus_allowed) ||
4200 rq->rd->dl_bw.bw == 0) {
4201 task_rq_unlock(rq, p, &flags);
4208 /* recheck policy now with rq lock held */
4209 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4210 policy = oldpolicy = -1;
4211 task_rq_unlock(rq, p, &flags);
4216 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4217 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4220 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4221 task_rq_unlock(rq, p, &flags);
4225 p->sched_reset_on_fork = reset_on_fork;
4230 * Take priority boosted tasks into account. If the new
4231 * effective priority is unchanged, we just store the new
4232 * normal parameters and do not touch the scheduler class and
4233 * the runqueue. This will be done when the task deboost
4236 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4237 if (new_effective_prio == oldprio) {
4238 __setscheduler_params(p, attr);
4239 task_rq_unlock(rq, p, &flags);
4244 queued = task_on_rq_queued(p);
4245 running = task_current(rq, p);
4247 dequeue_task(rq, p, DEQUEUE_SAVE);
4249 put_prev_task(rq, p);
4251 prev_class = p->sched_class;
4252 __setscheduler(rq, p, attr, pi);
4255 p->sched_class->set_curr_task(rq);
4257 int enqueue_flags = ENQUEUE_RESTORE;
4259 * We enqueue to tail when the priority of a task is
4260 * increased (user space view).
4262 if (oldprio <= p->prio)
4263 enqueue_flags |= ENQUEUE_HEAD;
4265 enqueue_task(rq, p, enqueue_flags);
4268 check_class_changed(rq, p, prev_class, oldprio);
4269 preempt_disable(); /* avoid rq from going away on us */
4270 task_rq_unlock(rq, p, &flags);
4273 rt_mutex_adjust_pi(p);
4276 * Run balance callbacks after we've adjusted the PI chain.
4278 balance_callback(rq);
4284 static int _sched_setscheduler(struct task_struct *p, int policy,
4285 const struct sched_param *param, bool check)
4287 struct sched_attr attr = {
4288 .sched_policy = policy,
4289 .sched_priority = param->sched_priority,
4290 .sched_nice = PRIO_TO_NICE(p->static_prio),
4293 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4294 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4295 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4296 policy &= ~SCHED_RESET_ON_FORK;
4297 attr.sched_policy = policy;
4300 return __sched_setscheduler(p, &attr, check, true);
4303 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4304 * @p: the task in question.
4305 * @policy: new policy.
4306 * @param: structure containing the new RT priority.
4308 * Return: 0 on success. An error code otherwise.
4310 * NOTE that the task may be already dead.
4312 int sched_setscheduler(struct task_struct *p, int policy,
4313 const struct sched_param *param)
4315 return _sched_setscheduler(p, policy, param, true);
4317 EXPORT_SYMBOL_GPL(sched_setscheduler);
4319 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4321 return __sched_setscheduler(p, attr, true, true);
4323 EXPORT_SYMBOL_GPL(sched_setattr);
4326 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4327 * @p: the task in question.
4328 * @policy: new policy.
4329 * @param: structure containing the new RT priority.
4331 * Just like sched_setscheduler, only don't bother checking if the
4332 * current context has permission. For example, this is needed in
4333 * stop_machine(): we create temporary high priority worker threads,
4334 * but our caller might not have that capability.
4336 * Return: 0 on success. An error code otherwise.
4338 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4339 const struct sched_param *param)
4341 return _sched_setscheduler(p, policy, param, false);
4343 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4346 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4348 struct sched_param lparam;
4349 struct task_struct *p;
4352 if (!param || pid < 0)
4354 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4359 p = find_process_by_pid(pid);
4361 retval = sched_setscheduler(p, policy, &lparam);
4368 * Mimics kernel/events/core.c perf_copy_attr().
4370 static int sched_copy_attr(struct sched_attr __user *uattr,
4371 struct sched_attr *attr)
4376 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4380 * zero the full structure, so that a short copy will be nice.
4382 memset(attr, 0, sizeof(*attr));
4384 ret = get_user(size, &uattr->size);
4388 if (size > PAGE_SIZE) /* silly large */
4391 if (!size) /* abi compat */
4392 size = SCHED_ATTR_SIZE_VER0;
4394 if (size < SCHED_ATTR_SIZE_VER0)
4398 * If we're handed a bigger struct than we know of,
4399 * ensure all the unknown bits are 0 - i.e. new
4400 * user-space does not rely on any kernel feature
4401 * extensions we dont know about yet.
4403 if (size > sizeof(*attr)) {
4404 unsigned char __user *addr;
4405 unsigned char __user *end;
4408 addr = (void __user *)uattr + sizeof(*attr);
4409 end = (void __user *)uattr + size;
4411 for (; addr < end; addr++) {
4412 ret = get_user(val, addr);
4418 size = sizeof(*attr);
4421 ret = copy_from_user(attr, uattr, size);
4426 * XXX: do we want to be lenient like existing syscalls; or do we want
4427 * to be strict and return an error on out-of-bounds values?
4429 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4434 put_user(sizeof(*attr), &uattr->size);
4439 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4440 * @pid: the pid in question.
4441 * @policy: new policy.
4442 * @param: structure containing the new RT priority.
4444 * Return: 0 on success. An error code otherwise.
4446 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4447 struct sched_param __user *, param)
4449 /* negative values for policy are not valid */
4453 return do_sched_setscheduler(pid, policy, param);
4457 * sys_sched_setparam - set/change the RT priority of a thread
4458 * @pid: the pid in question.
4459 * @param: structure containing the new RT priority.
4461 * Return: 0 on success. An error code otherwise.
4463 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4465 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4469 * sys_sched_setattr - same as above, but with extended sched_attr
4470 * @pid: the pid in question.
4471 * @uattr: structure containing the extended parameters.
4472 * @flags: for future extension.
4474 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4475 unsigned int, flags)
4477 struct sched_attr attr;
4478 struct task_struct *p;
4481 if (!uattr || pid < 0 || flags)
4484 retval = sched_copy_attr(uattr, &attr);
4488 if ((int)attr.sched_policy < 0)
4493 p = find_process_by_pid(pid);
4495 retval = sched_setattr(p, &attr);
4502 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4503 * @pid: the pid in question.
4505 * Return: On success, the policy of the thread. Otherwise, a negative error
4508 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4510 struct task_struct *p;
4518 p = find_process_by_pid(pid);
4520 retval = security_task_getscheduler(p);
4523 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4530 * sys_sched_getparam - get the RT priority of a thread
4531 * @pid: the pid in question.
4532 * @param: structure containing the RT priority.
4534 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4537 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4539 struct sched_param lp = { .sched_priority = 0 };
4540 struct task_struct *p;
4543 if (!param || pid < 0)
4547 p = find_process_by_pid(pid);
4552 retval = security_task_getscheduler(p);
4556 if (task_has_rt_policy(p))
4557 lp.sched_priority = p->rt_priority;
4561 * This one might sleep, we cannot do it with a spinlock held ...
4563 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4572 static int sched_read_attr(struct sched_attr __user *uattr,
4573 struct sched_attr *attr,
4578 if (!access_ok(VERIFY_WRITE, uattr, usize))
4582 * If we're handed a smaller struct than we know of,
4583 * ensure all the unknown bits are 0 - i.e. old
4584 * user-space does not get uncomplete information.
4586 if (usize < sizeof(*attr)) {
4587 unsigned char *addr;
4590 addr = (void *)attr + usize;
4591 end = (void *)attr + sizeof(*attr);
4593 for (; addr < end; addr++) {
4601 ret = copy_to_user(uattr, attr, attr->size);
4609 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4610 * @pid: the pid in question.
4611 * @uattr: structure containing the extended parameters.
4612 * @size: sizeof(attr) for fwd/bwd comp.
4613 * @flags: for future extension.
4615 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4616 unsigned int, size, unsigned int, flags)
4618 struct sched_attr attr = {
4619 .size = sizeof(struct sched_attr),
4621 struct task_struct *p;
4624 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4625 size < SCHED_ATTR_SIZE_VER0 || flags)
4629 p = find_process_by_pid(pid);
4634 retval = security_task_getscheduler(p);
4638 attr.sched_policy = p->policy;
4639 if (p->sched_reset_on_fork)
4640 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4641 if (task_has_dl_policy(p))
4642 __getparam_dl(p, &attr);
4643 else if (task_has_rt_policy(p))
4644 attr.sched_priority = p->rt_priority;
4646 attr.sched_nice = task_nice(p);
4650 retval = sched_read_attr(uattr, &attr, size);
4658 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4660 cpumask_var_t cpus_allowed, new_mask;
4661 struct task_struct *p;
4666 p = find_process_by_pid(pid);
4672 /* Prevent p going away */
4676 if (p->flags & PF_NO_SETAFFINITY) {
4680 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4684 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4686 goto out_free_cpus_allowed;
4689 if (!check_same_owner(p)) {
4691 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4693 goto out_free_new_mask;
4698 retval = security_task_setscheduler(p);
4700 goto out_free_new_mask;
4703 cpuset_cpus_allowed(p, cpus_allowed);
4704 cpumask_and(new_mask, in_mask, cpus_allowed);
4707 * Since bandwidth control happens on root_domain basis,
4708 * if admission test is enabled, we only admit -deadline
4709 * tasks allowed to run on all the CPUs in the task's
4713 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4715 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4718 goto out_free_new_mask;
4724 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4727 cpuset_cpus_allowed(p, cpus_allowed);
4728 if (!cpumask_subset(new_mask, cpus_allowed)) {
4730 * We must have raced with a concurrent cpuset
4731 * update. Just reset the cpus_allowed to the
4732 * cpuset's cpus_allowed
4734 cpumask_copy(new_mask, cpus_allowed);
4739 free_cpumask_var(new_mask);
4740 out_free_cpus_allowed:
4741 free_cpumask_var(cpus_allowed);
4747 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4748 struct cpumask *new_mask)
4750 if (len < cpumask_size())
4751 cpumask_clear(new_mask);
4752 else if (len > cpumask_size())
4753 len = cpumask_size();
4755 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4759 * sys_sched_setaffinity - set the cpu affinity of a process
4760 * @pid: pid of the process
4761 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4762 * @user_mask_ptr: user-space pointer to the new cpu mask
4764 * Return: 0 on success. An error code otherwise.
4766 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4767 unsigned long __user *, user_mask_ptr)
4769 cpumask_var_t new_mask;
4772 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4775 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4777 retval = sched_setaffinity(pid, new_mask);
4778 free_cpumask_var(new_mask);
4782 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4784 struct task_struct *p;
4785 unsigned long flags;
4791 p = find_process_by_pid(pid);
4795 retval = security_task_getscheduler(p);
4799 raw_spin_lock_irqsave(&p->pi_lock, flags);
4800 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4801 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4810 * sys_sched_getaffinity - get the cpu affinity of a process
4811 * @pid: pid of the process
4812 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4813 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4815 * Return: 0 on success. An error code otherwise.
4817 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4818 unsigned long __user *, user_mask_ptr)
4823 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4825 if (len & (sizeof(unsigned long)-1))
4828 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4831 ret = sched_getaffinity(pid, mask);
4833 size_t retlen = min_t(size_t, len, cpumask_size());
4835 if (copy_to_user(user_mask_ptr, mask, retlen))
4840 free_cpumask_var(mask);
4846 * sys_sched_yield - yield the current processor to other threads.
4848 * This function yields the current CPU to other tasks. If there are no
4849 * other threads running on this CPU then this function will return.
4853 SYSCALL_DEFINE0(sched_yield)
4855 struct rq *rq = this_rq_lock();
4857 schedstat_inc(rq, yld_count);
4858 current->sched_class->yield_task(rq);
4861 * Since we are going to call schedule() anyway, there's
4862 * no need to preempt or enable interrupts:
4864 __release(rq->lock);
4865 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4866 do_raw_spin_unlock(&rq->lock);
4867 sched_preempt_enable_no_resched();
4874 int __sched _cond_resched(void)
4876 if (should_resched(0)) {
4877 preempt_schedule_common();
4882 EXPORT_SYMBOL(_cond_resched);
4885 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4886 * call schedule, and on return reacquire the lock.
4888 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4889 * operations here to prevent schedule() from being called twice (once via
4890 * spin_unlock(), once by hand).
4892 int __cond_resched_lock(spinlock_t *lock)
4894 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4897 lockdep_assert_held(lock);
4899 if (spin_needbreak(lock) || resched) {
4902 preempt_schedule_common();
4910 EXPORT_SYMBOL(__cond_resched_lock);
4912 #ifndef CONFIG_PREEMPT_RT_FULL
4913 int __sched __cond_resched_softirq(void)
4915 BUG_ON(!in_softirq());
4917 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4919 preempt_schedule_common();
4925 EXPORT_SYMBOL(__cond_resched_softirq);
4929 * yield - yield the current processor to other threads.
4931 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4933 * The scheduler is at all times free to pick the calling task as the most
4934 * eligible task to run, if removing the yield() call from your code breaks
4935 * it, its already broken.
4937 * Typical broken usage is:
4942 * where one assumes that yield() will let 'the other' process run that will
4943 * make event true. If the current task is a SCHED_FIFO task that will never
4944 * happen. Never use yield() as a progress guarantee!!
4946 * If you want to use yield() to wait for something, use wait_event().
4947 * If you want to use yield() to be 'nice' for others, use cond_resched().
4948 * If you still want to use yield(), do not!
4950 void __sched yield(void)
4952 set_current_state(TASK_RUNNING);
4955 EXPORT_SYMBOL(yield);
4958 * yield_to - yield the current processor to another thread in
4959 * your thread group, or accelerate that thread toward the
4960 * processor it's on.
4962 * @preempt: whether task preemption is allowed or not
4964 * It's the caller's job to ensure that the target task struct
4965 * can't go away on us before we can do any checks.
4968 * true (>0) if we indeed boosted the target task.
4969 * false (0) if we failed to boost the target.
4970 * -ESRCH if there's no task to yield to.
4972 int __sched yield_to(struct task_struct *p, bool preempt)
4974 struct task_struct *curr = current;
4975 struct rq *rq, *p_rq;
4976 unsigned long flags;
4979 local_irq_save(flags);
4985 * If we're the only runnable task on the rq and target rq also
4986 * has only one task, there's absolutely no point in yielding.
4988 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4993 double_rq_lock(rq, p_rq);
4994 if (task_rq(p) != p_rq) {
4995 double_rq_unlock(rq, p_rq);
4999 if (!curr->sched_class->yield_to_task)
5002 if (curr->sched_class != p->sched_class)
5005 if (task_running(p_rq, p) || p->state)
5008 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5010 schedstat_inc(rq, yld_count);
5012 * Make p's CPU reschedule; pick_next_entity takes care of
5015 if (preempt && rq != p_rq)
5020 double_rq_unlock(rq, p_rq);
5022 local_irq_restore(flags);
5029 EXPORT_SYMBOL_GPL(yield_to);
5032 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5033 * that process accounting knows that this is a task in IO wait state.
5035 long __sched io_schedule_timeout(long timeout)
5037 int old_iowait = current->in_iowait;
5041 current->in_iowait = 1;
5042 blk_schedule_flush_plug(current);
5044 delayacct_blkio_start();
5046 atomic_inc(&rq->nr_iowait);
5047 ret = schedule_timeout(timeout);
5048 current->in_iowait = old_iowait;
5049 atomic_dec(&rq->nr_iowait);
5050 delayacct_blkio_end();
5054 EXPORT_SYMBOL(io_schedule_timeout);
5057 * sys_sched_get_priority_max - return maximum RT priority.
5058 * @policy: scheduling class.
5060 * Return: On success, this syscall returns the maximum
5061 * rt_priority that can be used by a given scheduling class.
5062 * On failure, a negative error code is returned.
5064 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5071 ret = MAX_USER_RT_PRIO-1;
5073 case SCHED_DEADLINE:
5084 * sys_sched_get_priority_min - return minimum RT priority.
5085 * @policy: scheduling class.
5087 * Return: On success, this syscall returns the minimum
5088 * rt_priority that can be used by a given scheduling class.
5089 * On failure, a negative error code is returned.
5091 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5100 case SCHED_DEADLINE:
5110 * sys_sched_rr_get_interval - return the default timeslice of a process.
5111 * @pid: pid of the process.
5112 * @interval: userspace pointer to the timeslice value.
5114 * this syscall writes the default timeslice value of a given process
5115 * into the user-space timespec buffer. A value of '0' means infinity.
5117 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5120 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5121 struct timespec __user *, interval)
5123 struct task_struct *p;
5124 unsigned int time_slice;
5125 unsigned long flags;
5135 p = find_process_by_pid(pid);
5139 retval = security_task_getscheduler(p);
5143 rq = task_rq_lock(p, &flags);
5145 if (p->sched_class->get_rr_interval)
5146 time_slice = p->sched_class->get_rr_interval(rq, p);
5147 task_rq_unlock(rq, p, &flags);
5150 jiffies_to_timespec(time_slice, &t);
5151 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5159 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5161 void sched_show_task(struct task_struct *p)
5163 unsigned long free = 0;
5165 unsigned long state = p->state;
5168 state = __ffs(state) + 1;
5169 printk(KERN_INFO "%-15.15s %c", p->comm,
5170 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5171 #if BITS_PER_LONG == 32
5172 if (state == TASK_RUNNING)
5173 printk(KERN_CONT " running ");
5175 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5177 if (state == TASK_RUNNING)
5178 printk(KERN_CONT " running task ");
5180 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5182 #ifdef CONFIG_DEBUG_STACK_USAGE
5183 free = stack_not_used(p);
5188 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5190 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5191 task_pid_nr(p), ppid,
5192 (unsigned long)task_thread_info(p)->flags);
5194 print_worker_info(KERN_INFO, p);
5195 show_stack(p, NULL);
5198 void show_state_filter(unsigned long state_filter)
5200 struct task_struct *g, *p;
5202 #if BITS_PER_LONG == 32
5204 " task PC stack pid father\n");
5207 " task PC stack pid father\n");
5210 for_each_process_thread(g, p) {
5212 * reset the NMI-timeout, listing all files on a slow
5213 * console might take a lot of time:
5214 * Also, reset softlockup watchdogs on all CPUs, because
5215 * another CPU might be blocked waiting for us to process
5218 touch_nmi_watchdog();
5219 touch_all_softlockup_watchdogs();
5220 if (!state_filter || (p->state & state_filter))
5224 #ifdef CONFIG_SCHED_DEBUG
5225 sysrq_sched_debug_show();
5229 * Only show locks if all tasks are dumped:
5232 debug_show_all_locks();
5235 void init_idle_bootup_task(struct task_struct *idle)
5237 idle->sched_class = &idle_sched_class;
5241 * init_idle - set up an idle thread for a given CPU
5242 * @idle: task in question
5243 * @cpu: cpu the idle task belongs to
5245 * NOTE: this function does not set the idle thread's NEED_RESCHED
5246 * flag, to make booting more robust.
5248 void init_idle(struct task_struct *idle, int cpu)
5250 struct rq *rq = cpu_rq(cpu);
5251 unsigned long flags;
5253 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5254 raw_spin_lock(&rq->lock);
5256 __sched_fork(0, idle);
5257 idle->state = TASK_RUNNING;
5258 idle->se.exec_start = sched_clock();
5262 * Its possible that init_idle() gets called multiple times on a task,
5263 * in that case do_set_cpus_allowed() will not do the right thing.
5265 * And since this is boot we can forgo the serialization.
5267 set_cpus_allowed_common(idle, cpumask_of(cpu));
5270 * We're having a chicken and egg problem, even though we are
5271 * holding rq->lock, the cpu isn't yet set to this cpu so the
5272 * lockdep check in task_group() will fail.
5274 * Similar case to sched_fork(). / Alternatively we could
5275 * use task_rq_lock() here and obtain the other rq->lock.
5280 __set_task_cpu(idle, cpu);
5283 rq->curr = rq->idle = idle;
5284 idle->on_rq = TASK_ON_RQ_QUEUED;
5288 raw_spin_unlock(&rq->lock);
5289 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5291 /* Set the preempt count _outside_ the spinlocks! */
5292 init_idle_preempt_count(idle, cpu);
5293 #ifdef CONFIG_HAVE_PREEMPT_LAZY
5294 task_thread_info(idle)->preempt_lazy_count = 0;
5297 * The idle tasks have their own, simple scheduling class:
5299 idle->sched_class = &idle_sched_class;
5300 ftrace_graph_init_idle_task(idle, cpu);
5301 vtime_init_idle(idle, cpu);
5303 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5307 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5308 const struct cpumask *trial)
5310 int ret = 1, trial_cpus;
5311 struct dl_bw *cur_dl_b;
5312 unsigned long flags;
5314 if (!cpumask_weight(cur))
5317 rcu_read_lock_sched();
5318 cur_dl_b = dl_bw_of(cpumask_any(cur));
5319 trial_cpus = cpumask_weight(trial);
5321 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5322 if (cur_dl_b->bw != -1 &&
5323 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5325 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5326 rcu_read_unlock_sched();
5331 int task_can_attach(struct task_struct *p,
5332 const struct cpumask *cs_cpus_allowed)
5337 * Kthreads which disallow setaffinity shouldn't be moved
5338 * to a new cpuset; we don't want to change their cpu
5339 * affinity and isolating such threads by their set of
5340 * allowed nodes is unnecessary. Thus, cpusets are not
5341 * applicable for such threads. This prevents checking for
5342 * success of set_cpus_allowed_ptr() on all attached tasks
5343 * before cpus_allowed may be changed.
5345 if (p->flags & PF_NO_SETAFFINITY) {
5351 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5353 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5358 unsigned long flags;
5360 rcu_read_lock_sched();
5361 dl_b = dl_bw_of(dest_cpu);
5362 raw_spin_lock_irqsave(&dl_b->lock, flags);
5363 cpus = dl_bw_cpus(dest_cpu);
5364 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5369 * We reserve space for this task in the destination
5370 * root_domain, as we can't fail after this point.
5371 * We will free resources in the source root_domain
5372 * later on (see set_cpus_allowed_dl()).
5374 __dl_add(dl_b, p->dl.dl_bw);
5376 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5377 rcu_read_unlock_sched();
5387 #ifdef CONFIG_NUMA_BALANCING
5388 /* Migrate current task p to target_cpu */
5389 int migrate_task_to(struct task_struct *p, int target_cpu)
5391 struct migration_arg arg = { p, target_cpu };
5392 int curr_cpu = task_cpu(p);
5394 if (curr_cpu == target_cpu)
5397 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5400 /* TODO: This is not properly updating schedstats */
5402 trace_sched_move_numa(p, curr_cpu, target_cpu);
5403 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5407 * Requeue a task on a given node and accurately track the number of NUMA
5408 * tasks on the runqueues
5410 void sched_setnuma(struct task_struct *p, int nid)
5413 unsigned long flags;
5414 bool queued, running;
5416 rq = task_rq_lock(p, &flags);
5417 queued = task_on_rq_queued(p);
5418 running = task_current(rq, p);
5421 dequeue_task(rq, p, DEQUEUE_SAVE);
5423 put_prev_task(rq, p);
5425 p->numa_preferred_nid = nid;
5428 p->sched_class->set_curr_task(rq);
5430 enqueue_task(rq, p, ENQUEUE_RESTORE);
5431 task_rq_unlock(rq, p, &flags);
5433 #endif /* CONFIG_NUMA_BALANCING */
5435 #ifdef CONFIG_HOTPLUG_CPU
5436 static DEFINE_PER_CPU(struct mm_struct *, idle_last_mm);
5439 * Ensures that the idle task is using init_mm right before its cpu goes
5442 void idle_task_exit(void)
5444 struct mm_struct *mm = current->active_mm;
5446 BUG_ON(cpu_online(smp_processor_id()));
5448 if (mm != &init_mm) {
5449 switch_mm(mm, &init_mm, current);
5450 finish_arch_post_lock_switch();
5453 * Defer the cleanup to an alive cpu. On RT we can neither
5454 * call mmdrop() nor mmdrop_delayed() from here.
5456 per_cpu(idle_last_mm, smp_processor_id()) = mm;
5460 * Since this CPU is going 'away' for a while, fold any nr_active delta
5461 * we might have. Assumes we're called after migrate_tasks() so that the
5462 * nr_active count is stable.
5464 * Also see the comment "Global load-average calculations".
5466 static void calc_load_migrate(struct rq *rq)
5468 long delta = calc_load_fold_active(rq);
5470 atomic_long_add(delta, &calc_load_tasks);
5473 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5477 static const struct sched_class fake_sched_class = {
5478 .put_prev_task = put_prev_task_fake,
5481 static struct task_struct fake_task = {
5483 * Avoid pull_{rt,dl}_task()
5485 .prio = MAX_PRIO + 1,
5486 .sched_class = &fake_sched_class,
5490 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5491 * try_to_wake_up()->select_task_rq().
5493 * Called with rq->lock held even though we'er in stop_machine() and
5494 * there's no concurrency possible, we hold the required locks anyway
5495 * because of lock validation efforts.
5497 static void migrate_tasks(struct rq *dead_rq)
5499 struct rq *rq = dead_rq;
5500 struct task_struct *next, *stop = rq->stop;
5504 * Fudge the rq selection such that the below task selection loop
5505 * doesn't get stuck on the currently eligible stop task.
5507 * We're currently inside stop_machine() and the rq is either stuck
5508 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5509 * either way we should never end up calling schedule() until we're
5515 * put_prev_task() and pick_next_task() sched
5516 * class method both need to have an up-to-date
5517 * value of rq->clock[_task]
5519 update_rq_clock(rq);
5523 * There's this thread running, bail when that's the only
5526 if (rq->nr_running == 1)
5530 * pick_next_task assumes pinned rq->lock.
5532 lockdep_pin_lock(&rq->lock);
5533 next = pick_next_task(rq, &fake_task);
5535 next->sched_class->put_prev_task(rq, next);
5538 * Rules for changing task_struct::cpus_allowed are holding
5539 * both pi_lock and rq->lock, such that holding either
5540 * stabilizes the mask.
5542 * Drop rq->lock is not quite as disastrous as it usually is
5543 * because !cpu_active at this point, which means load-balance
5544 * will not interfere. Also, stop-machine.
5546 lockdep_unpin_lock(&rq->lock);
5547 raw_spin_unlock(&rq->lock);
5548 raw_spin_lock(&next->pi_lock);
5549 raw_spin_lock(&rq->lock);
5552 * Since we're inside stop-machine, _nothing_ should have
5553 * changed the task, WARN if weird stuff happened, because in
5554 * that case the above rq->lock drop is a fail too.
5556 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5557 raw_spin_unlock(&next->pi_lock);
5561 /* Find suitable destination for @next, with force if needed. */
5562 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5564 rq = __migrate_task(rq, next, dest_cpu);
5565 if (rq != dead_rq) {
5566 raw_spin_unlock(&rq->lock);
5568 raw_spin_lock(&rq->lock);
5570 raw_spin_unlock(&next->pi_lock);
5575 #endif /* CONFIG_HOTPLUG_CPU */
5577 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5579 static struct ctl_table sd_ctl_dir[] = {
5581 .procname = "sched_domain",
5587 static struct ctl_table sd_ctl_root[] = {
5589 .procname = "kernel",
5591 .child = sd_ctl_dir,
5596 static struct ctl_table *sd_alloc_ctl_entry(int n)
5598 struct ctl_table *entry =
5599 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5604 static void sd_free_ctl_entry(struct ctl_table **tablep)
5606 struct ctl_table *entry;
5609 * In the intermediate directories, both the child directory and
5610 * procname are dynamically allocated and could fail but the mode
5611 * will always be set. In the lowest directory the names are
5612 * static strings and all have proc handlers.
5614 for (entry = *tablep; entry->mode; entry++) {
5616 sd_free_ctl_entry(&entry->child);
5617 if (entry->proc_handler == NULL)
5618 kfree(entry->procname);
5625 static int min_load_idx = 0;
5626 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5629 set_table_entry(struct ctl_table *entry,
5630 const char *procname, void *data, int maxlen,
5631 umode_t mode, proc_handler *proc_handler,
5634 entry->procname = procname;
5636 entry->maxlen = maxlen;
5638 entry->proc_handler = proc_handler;
5641 entry->extra1 = &min_load_idx;
5642 entry->extra2 = &max_load_idx;
5646 static struct ctl_table *
5647 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5649 struct ctl_table *table = sd_alloc_ctl_entry(14);
5654 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5655 sizeof(long), 0644, proc_doulongvec_minmax, false);
5656 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5657 sizeof(long), 0644, proc_doulongvec_minmax, false);
5658 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5659 sizeof(int), 0644, proc_dointvec_minmax, true);
5660 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5661 sizeof(int), 0644, proc_dointvec_minmax, true);
5662 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5663 sizeof(int), 0644, proc_dointvec_minmax, true);
5664 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5665 sizeof(int), 0644, proc_dointvec_minmax, true);
5666 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5667 sizeof(int), 0644, proc_dointvec_minmax, true);
5668 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5669 sizeof(int), 0644, proc_dointvec_minmax, false);
5670 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5671 sizeof(int), 0644, proc_dointvec_minmax, false);
5672 set_table_entry(&table[9], "cache_nice_tries",
5673 &sd->cache_nice_tries,
5674 sizeof(int), 0644, proc_dointvec_minmax, false);
5675 set_table_entry(&table[10], "flags", &sd->flags,
5676 sizeof(int), 0644, proc_dointvec_minmax, false);
5677 set_table_entry(&table[11], "max_newidle_lb_cost",
5678 &sd->max_newidle_lb_cost,
5679 sizeof(long), 0644, proc_doulongvec_minmax, false);
5680 set_table_entry(&table[12], "name", sd->name,
5681 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5682 /* &table[13] is terminator */
5687 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5689 struct ctl_table *entry, *table;
5690 struct sched_domain *sd;
5691 int domain_num = 0, i;
5694 for_each_domain(cpu, sd)
5696 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5701 for_each_domain(cpu, sd) {
5702 snprintf(buf, 32, "domain%d", i);
5703 entry->procname = kstrdup(buf, GFP_KERNEL);
5705 entry->child = sd_alloc_ctl_domain_table(sd);
5712 static struct ctl_table_header *sd_sysctl_header;
5713 static void register_sched_domain_sysctl(void)
5715 int i, cpu_num = num_possible_cpus();
5716 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5719 WARN_ON(sd_ctl_dir[0].child);
5720 sd_ctl_dir[0].child = entry;
5725 for_each_possible_cpu(i) {
5726 snprintf(buf, 32, "cpu%d", i);
5727 entry->procname = kstrdup(buf, GFP_KERNEL);
5729 entry->child = sd_alloc_ctl_cpu_table(i);
5733 WARN_ON(sd_sysctl_header);
5734 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5737 /* may be called multiple times per register */
5738 static void unregister_sched_domain_sysctl(void)
5740 unregister_sysctl_table(sd_sysctl_header);
5741 sd_sysctl_header = NULL;
5742 if (sd_ctl_dir[0].child)
5743 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5746 static void register_sched_domain_sysctl(void)
5749 static void unregister_sched_domain_sysctl(void)
5752 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5754 static void set_rq_online(struct rq *rq)
5757 const struct sched_class *class;
5759 cpumask_set_cpu(rq->cpu, rq->rd->online);
5762 for_each_class(class) {
5763 if (class->rq_online)
5764 class->rq_online(rq);
5769 static void set_rq_offline(struct rq *rq)
5772 const struct sched_class *class;
5774 for_each_class(class) {
5775 if (class->rq_offline)
5776 class->rq_offline(rq);
5779 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5785 * migration_call - callback that gets triggered when a CPU is added.
5786 * Here we can start up the necessary migration thread for the new CPU.
5789 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5791 int cpu = (long)hcpu;
5792 unsigned long flags;
5793 struct rq *rq = cpu_rq(cpu);
5795 switch (action & ~CPU_TASKS_FROZEN) {
5797 case CPU_UP_PREPARE:
5798 rq->calc_load_update = calc_load_update;
5799 account_reset_rq(rq);
5803 /* Update our root-domain */
5804 raw_spin_lock_irqsave(&rq->lock, flags);
5806 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5810 raw_spin_unlock_irqrestore(&rq->lock, flags);
5813 #ifdef CONFIG_HOTPLUG_CPU
5815 sched_ttwu_pending();
5816 /* Update our root-domain */
5817 raw_spin_lock_irqsave(&rq->lock, flags);
5819 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5823 BUG_ON(rq->nr_running != 1); /* the migration thread */
5824 raw_spin_unlock_irqrestore(&rq->lock, flags);
5828 calc_load_migrate(rq);
5829 if (per_cpu(idle_last_mm, cpu)) {
5830 mmdrop(per_cpu(idle_last_mm, cpu));
5831 per_cpu(idle_last_mm, cpu) = NULL;
5837 update_max_interval();
5843 * Register at high priority so that task migration (migrate_all_tasks)
5844 * happens before everything else. This has to be lower priority than
5845 * the notifier in the perf_event subsystem, though.
5847 static struct notifier_block migration_notifier = {
5848 .notifier_call = migration_call,
5849 .priority = CPU_PRI_MIGRATION,
5852 static void set_cpu_rq_start_time(void)
5854 int cpu = smp_processor_id();
5855 struct rq *rq = cpu_rq(cpu);
5856 rq->age_stamp = sched_clock_cpu(cpu);
5859 static int sched_cpu_active(struct notifier_block *nfb,
5860 unsigned long action, void *hcpu)
5862 int cpu = (long)hcpu;
5864 switch (action & ~CPU_TASKS_FROZEN) {
5866 set_cpu_rq_start_time();
5871 * At this point a starting CPU has marked itself as online via
5872 * set_cpu_online(). But it might not yet have marked itself
5873 * as active, which is essential from here on.
5875 set_cpu_active(cpu, true);
5876 stop_machine_unpark(cpu);
5879 case CPU_DOWN_FAILED:
5880 set_cpu_active(cpu, true);
5888 static int sched_cpu_inactive(struct notifier_block *nfb,
5889 unsigned long action, void *hcpu)
5891 switch (action & ~CPU_TASKS_FROZEN) {
5892 case CPU_DOWN_PREPARE:
5893 set_cpu_active((long)hcpu, false);
5900 static int __init migration_init(void)
5902 void *cpu = (void *)(long)smp_processor_id();
5905 /* Initialize migration for the boot CPU */
5906 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5907 BUG_ON(err == NOTIFY_BAD);
5908 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5909 register_cpu_notifier(&migration_notifier);
5911 /* Register cpu active notifiers */
5912 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5913 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5917 early_initcall(migration_init);
5919 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5921 #ifdef CONFIG_SCHED_DEBUG
5923 static __read_mostly int sched_debug_enabled;
5925 static int __init sched_debug_setup(char *str)
5927 sched_debug_enabled = 1;
5931 early_param("sched_debug", sched_debug_setup);
5933 static inline bool sched_debug(void)
5935 return sched_debug_enabled;
5938 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5939 struct cpumask *groupmask)
5941 struct sched_group *group = sd->groups;
5943 cpumask_clear(groupmask);
5945 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5947 if (!(sd->flags & SD_LOAD_BALANCE)) {
5948 printk("does not load-balance\n");
5950 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5955 printk(KERN_CONT "span %*pbl level %s\n",
5956 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5958 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5959 printk(KERN_ERR "ERROR: domain->span does not contain "
5962 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5963 printk(KERN_ERR "ERROR: domain->groups does not contain"
5967 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5971 printk(KERN_ERR "ERROR: group is NULL\n");
5975 if (!cpumask_weight(sched_group_cpus(group))) {
5976 printk(KERN_CONT "\n");
5977 printk(KERN_ERR "ERROR: empty group\n");
5981 if (!(sd->flags & SD_OVERLAP) &&
5982 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5983 printk(KERN_CONT "\n");
5984 printk(KERN_ERR "ERROR: repeated CPUs\n");
5988 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5990 printk(KERN_CONT " %*pbl",
5991 cpumask_pr_args(sched_group_cpus(group)));
5992 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5993 printk(KERN_CONT " (cpu_capacity = %d)",
5994 group->sgc->capacity);
5997 group = group->next;
5998 } while (group != sd->groups);
5999 printk(KERN_CONT "\n");
6001 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6002 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6005 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6006 printk(KERN_ERR "ERROR: parent span is not a superset "
6007 "of domain->span\n");
6011 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6015 if (!sched_debug_enabled)
6019 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6023 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6026 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6034 #else /* !CONFIG_SCHED_DEBUG */
6035 # define sched_domain_debug(sd, cpu) do { } while (0)
6036 static inline bool sched_debug(void)
6040 #endif /* CONFIG_SCHED_DEBUG */
6042 static int sd_degenerate(struct sched_domain *sd)
6044 if (cpumask_weight(sched_domain_span(sd)) == 1)
6047 /* Following flags need at least 2 groups */
6048 if (sd->flags & (SD_LOAD_BALANCE |
6049 SD_BALANCE_NEWIDLE |
6052 SD_SHARE_CPUCAPACITY |
6053 SD_SHARE_PKG_RESOURCES |
6054 SD_SHARE_POWERDOMAIN)) {
6055 if (sd->groups != sd->groups->next)
6059 /* Following flags don't use groups */
6060 if (sd->flags & (SD_WAKE_AFFINE))
6067 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6069 unsigned long cflags = sd->flags, pflags = parent->flags;
6071 if (sd_degenerate(parent))
6074 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6077 /* Flags needing groups don't count if only 1 group in parent */
6078 if (parent->groups == parent->groups->next) {
6079 pflags &= ~(SD_LOAD_BALANCE |
6080 SD_BALANCE_NEWIDLE |
6083 SD_SHARE_CPUCAPACITY |
6084 SD_SHARE_PKG_RESOURCES |
6086 SD_SHARE_POWERDOMAIN);
6087 if (nr_node_ids == 1)
6088 pflags &= ~SD_SERIALIZE;
6090 if (~cflags & pflags)
6096 static void free_rootdomain(struct rcu_head *rcu)
6098 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6100 cpupri_cleanup(&rd->cpupri);
6101 cpudl_cleanup(&rd->cpudl);
6102 free_cpumask_var(rd->dlo_mask);
6103 free_cpumask_var(rd->rto_mask);
6104 free_cpumask_var(rd->online);
6105 free_cpumask_var(rd->span);
6109 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6111 struct root_domain *old_rd = NULL;
6112 unsigned long flags;
6114 raw_spin_lock_irqsave(&rq->lock, flags);
6119 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6122 cpumask_clear_cpu(rq->cpu, old_rd->span);
6125 * If we dont want to free the old_rd yet then
6126 * set old_rd to NULL to skip the freeing later
6129 if (!atomic_dec_and_test(&old_rd->refcount))
6133 atomic_inc(&rd->refcount);
6136 cpumask_set_cpu(rq->cpu, rd->span);
6137 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6140 raw_spin_unlock_irqrestore(&rq->lock, flags);
6143 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6146 static int init_rootdomain(struct root_domain *rd)
6148 memset(rd, 0, sizeof(*rd));
6150 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6152 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6154 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6156 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6159 init_dl_bw(&rd->dl_bw);
6160 if (cpudl_init(&rd->cpudl) != 0)
6163 if (cpupri_init(&rd->cpupri) != 0)
6168 free_cpumask_var(rd->rto_mask);
6170 free_cpumask_var(rd->dlo_mask);
6172 free_cpumask_var(rd->online);
6174 free_cpumask_var(rd->span);
6180 * By default the system creates a single root-domain with all cpus as
6181 * members (mimicking the global state we have today).
6183 struct root_domain def_root_domain;
6185 static void init_defrootdomain(void)
6187 init_rootdomain(&def_root_domain);
6189 atomic_set(&def_root_domain.refcount, 1);
6192 static struct root_domain *alloc_rootdomain(void)
6194 struct root_domain *rd;
6196 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6200 if (init_rootdomain(rd) != 0) {
6208 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6210 struct sched_group *tmp, *first;
6219 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6224 } while (sg != first);
6227 static void free_sched_domain(struct rcu_head *rcu)
6229 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6232 * If its an overlapping domain it has private groups, iterate and
6235 if (sd->flags & SD_OVERLAP) {
6236 free_sched_groups(sd->groups, 1);
6237 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6238 kfree(sd->groups->sgc);
6244 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6246 call_rcu(&sd->rcu, free_sched_domain);
6249 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6251 for (; sd; sd = sd->parent)
6252 destroy_sched_domain(sd, cpu);
6256 * Keep a special pointer to the highest sched_domain that has
6257 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6258 * allows us to avoid some pointer chasing select_idle_sibling().
6260 * Also keep a unique ID per domain (we use the first cpu number in
6261 * the cpumask of the domain), this allows us to quickly tell if
6262 * two cpus are in the same cache domain, see cpus_share_cache().
6264 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6265 DEFINE_PER_CPU(int, sd_llc_size);
6266 DEFINE_PER_CPU(int, sd_llc_id);
6267 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6268 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6269 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6271 static void update_top_cache_domain(int cpu)
6273 struct sched_domain *sd;
6274 struct sched_domain *busy_sd = NULL;
6278 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6280 id = cpumask_first(sched_domain_span(sd));
6281 size = cpumask_weight(sched_domain_span(sd));
6282 busy_sd = sd->parent; /* sd_busy */
6284 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6286 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6287 per_cpu(sd_llc_size, cpu) = size;
6288 per_cpu(sd_llc_id, cpu) = id;
6290 sd = lowest_flag_domain(cpu, SD_NUMA);
6291 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6293 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6294 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6298 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6299 * hold the hotplug lock.
6302 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6304 struct rq *rq = cpu_rq(cpu);
6305 struct sched_domain *tmp;
6307 /* Remove the sched domains which do not contribute to scheduling. */
6308 for (tmp = sd; tmp; ) {
6309 struct sched_domain *parent = tmp->parent;
6313 if (sd_parent_degenerate(tmp, parent)) {
6314 tmp->parent = parent->parent;
6316 parent->parent->child = tmp;
6318 * Transfer SD_PREFER_SIBLING down in case of a
6319 * degenerate parent; the spans match for this
6320 * so the property transfers.
6322 if (parent->flags & SD_PREFER_SIBLING)
6323 tmp->flags |= SD_PREFER_SIBLING;
6324 destroy_sched_domain(parent, cpu);
6329 if (sd && sd_degenerate(sd)) {
6332 destroy_sched_domain(tmp, cpu);
6337 sched_domain_debug(sd, cpu);
6339 rq_attach_root(rq, rd);
6341 rcu_assign_pointer(rq->sd, sd);
6342 destroy_sched_domains(tmp, cpu);
6344 update_top_cache_domain(cpu);
6347 /* Setup the mask of cpus configured for isolated domains */
6348 static int __init isolated_cpu_setup(char *str)
6350 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6351 cpulist_parse(str, cpu_isolated_map);
6355 __setup("isolcpus=", isolated_cpu_setup);
6358 struct sched_domain ** __percpu sd;
6359 struct root_domain *rd;
6370 * Build an iteration mask that can exclude certain CPUs from the upwards
6373 * Asymmetric node setups can result in situations where the domain tree is of
6374 * unequal depth, make sure to skip domains that already cover the entire
6377 * In that case build_sched_domains() will have terminated the iteration early
6378 * and our sibling sd spans will be empty. Domains should always include the
6379 * cpu they're built on, so check that.
6382 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6384 const struct cpumask *span = sched_domain_span(sd);
6385 struct sd_data *sdd = sd->private;
6386 struct sched_domain *sibling;
6389 for_each_cpu(i, span) {
6390 sibling = *per_cpu_ptr(sdd->sd, i);
6391 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6394 cpumask_set_cpu(i, sched_group_mask(sg));
6399 * Return the canonical balance cpu for this group, this is the first cpu
6400 * of this group that's also in the iteration mask.
6402 int group_balance_cpu(struct sched_group *sg)
6404 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6408 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6410 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6411 const struct cpumask *span = sched_domain_span(sd);
6412 struct cpumask *covered = sched_domains_tmpmask;
6413 struct sd_data *sdd = sd->private;
6414 struct sched_domain *sibling;
6417 cpumask_clear(covered);
6419 for_each_cpu(i, span) {
6420 struct cpumask *sg_span;
6422 if (cpumask_test_cpu(i, covered))
6425 sibling = *per_cpu_ptr(sdd->sd, i);
6427 /* See the comment near build_group_mask(). */
6428 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6431 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6432 GFP_KERNEL, cpu_to_node(cpu));
6437 sg_span = sched_group_cpus(sg);
6439 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6441 cpumask_set_cpu(i, sg_span);
6443 cpumask_or(covered, covered, sg_span);
6445 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6446 if (atomic_inc_return(&sg->sgc->ref) == 1)
6447 build_group_mask(sd, sg);
6450 * Initialize sgc->capacity such that even if we mess up the
6451 * domains and no possible iteration will get us here, we won't
6454 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6457 * Make sure the first group of this domain contains the
6458 * canonical balance cpu. Otherwise the sched_domain iteration
6459 * breaks. See update_sg_lb_stats().
6461 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6462 group_balance_cpu(sg) == cpu)
6472 sd->groups = groups;
6477 free_sched_groups(first, 0);
6482 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6484 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6485 struct sched_domain *child = sd->child;
6488 cpu = cpumask_first(sched_domain_span(child));
6491 *sg = *per_cpu_ptr(sdd->sg, cpu);
6492 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6493 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6500 * build_sched_groups will build a circular linked list of the groups
6501 * covered by the given span, and will set each group's ->cpumask correctly,
6502 * and ->cpu_capacity to 0.
6504 * Assumes the sched_domain tree is fully constructed
6507 build_sched_groups(struct sched_domain *sd, int cpu)
6509 struct sched_group *first = NULL, *last = NULL;
6510 struct sd_data *sdd = sd->private;
6511 const struct cpumask *span = sched_domain_span(sd);
6512 struct cpumask *covered;
6515 get_group(cpu, sdd, &sd->groups);
6516 atomic_inc(&sd->groups->ref);
6518 if (cpu != cpumask_first(span))
6521 lockdep_assert_held(&sched_domains_mutex);
6522 covered = sched_domains_tmpmask;
6524 cpumask_clear(covered);
6526 for_each_cpu(i, span) {
6527 struct sched_group *sg;
6530 if (cpumask_test_cpu(i, covered))
6533 group = get_group(i, sdd, &sg);
6534 cpumask_setall(sched_group_mask(sg));
6536 for_each_cpu(j, span) {
6537 if (get_group(j, sdd, NULL) != group)
6540 cpumask_set_cpu(j, covered);
6541 cpumask_set_cpu(j, sched_group_cpus(sg));
6556 * Initialize sched groups cpu_capacity.
6558 * cpu_capacity indicates the capacity of sched group, which is used while
6559 * distributing the load between different sched groups in a sched domain.
6560 * Typically cpu_capacity for all the groups in a sched domain will be same
6561 * unless there are asymmetries in the topology. If there are asymmetries,
6562 * group having more cpu_capacity will pickup more load compared to the
6563 * group having less cpu_capacity.
6565 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6567 struct sched_group *sg = sd->groups;
6572 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6574 } while (sg != sd->groups);
6576 if (cpu != group_balance_cpu(sg))
6579 update_group_capacity(sd, cpu);
6580 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6584 * Initializers for schedule domains
6585 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6588 static int default_relax_domain_level = -1;
6589 int sched_domain_level_max;
6591 static int __init setup_relax_domain_level(char *str)
6593 if (kstrtoint(str, 0, &default_relax_domain_level))
6594 pr_warn("Unable to set relax_domain_level\n");
6598 __setup("relax_domain_level=", setup_relax_domain_level);
6600 static void set_domain_attribute(struct sched_domain *sd,
6601 struct sched_domain_attr *attr)
6605 if (!attr || attr->relax_domain_level < 0) {
6606 if (default_relax_domain_level < 0)
6609 request = default_relax_domain_level;
6611 request = attr->relax_domain_level;
6612 if (request < sd->level) {
6613 /* turn off idle balance on this domain */
6614 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6616 /* turn on idle balance on this domain */
6617 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6621 static void __sdt_free(const struct cpumask *cpu_map);
6622 static int __sdt_alloc(const struct cpumask *cpu_map);
6624 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6625 const struct cpumask *cpu_map)
6629 if (!atomic_read(&d->rd->refcount))
6630 free_rootdomain(&d->rd->rcu); /* fall through */
6632 free_percpu(d->sd); /* fall through */
6634 __sdt_free(cpu_map); /* fall through */
6640 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6641 const struct cpumask *cpu_map)
6643 memset(d, 0, sizeof(*d));
6645 if (__sdt_alloc(cpu_map))
6646 return sa_sd_storage;
6647 d->sd = alloc_percpu(struct sched_domain *);
6649 return sa_sd_storage;
6650 d->rd = alloc_rootdomain();
6653 return sa_rootdomain;
6657 * NULL the sd_data elements we've used to build the sched_domain and
6658 * sched_group structure so that the subsequent __free_domain_allocs()
6659 * will not free the data we're using.
6661 static void claim_allocations(int cpu, struct sched_domain *sd)
6663 struct sd_data *sdd = sd->private;
6665 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6666 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6668 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6669 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6671 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6672 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6676 static int sched_domains_numa_levels;
6677 enum numa_topology_type sched_numa_topology_type;
6678 static int *sched_domains_numa_distance;
6679 int sched_max_numa_distance;
6680 static struct cpumask ***sched_domains_numa_masks;
6681 static int sched_domains_curr_level;
6685 * SD_flags allowed in topology descriptions.
6687 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6688 * SD_SHARE_PKG_RESOURCES - describes shared caches
6689 * SD_NUMA - describes NUMA topologies
6690 * SD_SHARE_POWERDOMAIN - describes shared power domain
6693 * SD_ASYM_PACKING - describes SMT quirks
6695 #define TOPOLOGY_SD_FLAGS \
6696 (SD_SHARE_CPUCAPACITY | \
6697 SD_SHARE_PKG_RESOURCES | \
6700 SD_SHARE_POWERDOMAIN)
6702 static struct sched_domain *
6703 sd_init(struct sched_domain_topology_level *tl, int cpu)
6705 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6706 int sd_weight, sd_flags = 0;
6710 * Ugly hack to pass state to sd_numa_mask()...
6712 sched_domains_curr_level = tl->numa_level;
6715 sd_weight = cpumask_weight(tl->mask(cpu));
6718 sd_flags = (*tl->sd_flags)();
6719 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6720 "wrong sd_flags in topology description\n"))
6721 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6723 *sd = (struct sched_domain){
6724 .min_interval = sd_weight,
6725 .max_interval = 2*sd_weight,
6727 .imbalance_pct = 125,
6729 .cache_nice_tries = 0,
6736 .flags = 1*SD_LOAD_BALANCE
6737 | 1*SD_BALANCE_NEWIDLE
6742 | 0*SD_SHARE_CPUCAPACITY
6743 | 0*SD_SHARE_PKG_RESOURCES
6745 | 0*SD_PREFER_SIBLING
6750 .last_balance = jiffies,
6751 .balance_interval = sd_weight,
6753 .max_newidle_lb_cost = 0,
6754 .next_decay_max_lb_cost = jiffies,
6755 #ifdef CONFIG_SCHED_DEBUG
6761 * Convert topological properties into behaviour.
6764 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6765 sd->flags |= SD_PREFER_SIBLING;
6766 sd->imbalance_pct = 110;
6767 sd->smt_gain = 1178; /* ~15% */
6769 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6770 sd->imbalance_pct = 117;
6771 sd->cache_nice_tries = 1;
6775 } else if (sd->flags & SD_NUMA) {
6776 sd->cache_nice_tries = 2;
6780 sd->flags |= SD_SERIALIZE;
6781 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6782 sd->flags &= ~(SD_BALANCE_EXEC |
6789 sd->flags |= SD_PREFER_SIBLING;
6790 sd->cache_nice_tries = 1;
6795 sd->private = &tl->data;
6801 * Topology list, bottom-up.
6803 static struct sched_domain_topology_level default_topology[] = {
6804 #ifdef CONFIG_SCHED_SMT
6805 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6807 #ifdef CONFIG_SCHED_MC
6808 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6810 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6814 static struct sched_domain_topology_level *sched_domain_topology =
6817 #define for_each_sd_topology(tl) \
6818 for (tl = sched_domain_topology; tl->mask; tl++)
6820 void set_sched_topology(struct sched_domain_topology_level *tl)
6822 sched_domain_topology = tl;
6827 static const struct cpumask *sd_numa_mask(int cpu)
6829 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6832 static void sched_numa_warn(const char *str)
6834 static int done = false;
6842 printk(KERN_WARNING "ERROR: %s\n\n", str);
6844 for (i = 0; i < nr_node_ids; i++) {
6845 printk(KERN_WARNING " ");
6846 for (j = 0; j < nr_node_ids; j++)
6847 printk(KERN_CONT "%02d ", node_distance(i,j));
6848 printk(KERN_CONT "\n");
6850 printk(KERN_WARNING "\n");
6853 bool find_numa_distance(int distance)
6857 if (distance == node_distance(0, 0))
6860 for (i = 0; i < sched_domains_numa_levels; i++) {
6861 if (sched_domains_numa_distance[i] == distance)
6869 * A system can have three types of NUMA topology:
6870 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6871 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6872 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6874 * The difference between a glueless mesh topology and a backplane
6875 * topology lies in whether communication between not directly
6876 * connected nodes goes through intermediary nodes (where programs
6877 * could run), or through backplane controllers. This affects
6878 * placement of programs.
6880 * The type of topology can be discerned with the following tests:
6881 * - If the maximum distance between any nodes is 1 hop, the system
6882 * is directly connected.
6883 * - If for two nodes A and B, located N > 1 hops away from each other,
6884 * there is an intermediary node C, which is < N hops away from both
6885 * nodes A and B, the system is a glueless mesh.
6887 static void init_numa_topology_type(void)
6891 n = sched_max_numa_distance;
6893 if (sched_domains_numa_levels <= 1) {
6894 sched_numa_topology_type = NUMA_DIRECT;
6898 for_each_online_node(a) {
6899 for_each_online_node(b) {
6900 /* Find two nodes furthest removed from each other. */
6901 if (node_distance(a, b) < n)
6904 /* Is there an intermediary node between a and b? */
6905 for_each_online_node(c) {
6906 if (node_distance(a, c) < n &&
6907 node_distance(b, c) < n) {
6908 sched_numa_topology_type =
6914 sched_numa_topology_type = NUMA_BACKPLANE;
6920 static void sched_init_numa(void)
6922 int next_distance, curr_distance = node_distance(0, 0);
6923 struct sched_domain_topology_level *tl;
6927 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6928 if (!sched_domains_numa_distance)
6932 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6933 * unique distances in the node_distance() table.
6935 * Assumes node_distance(0,j) includes all distances in
6936 * node_distance(i,j) in order to avoid cubic time.
6938 next_distance = curr_distance;
6939 for (i = 0; i < nr_node_ids; i++) {
6940 for (j = 0; j < nr_node_ids; j++) {
6941 for (k = 0; k < nr_node_ids; k++) {
6942 int distance = node_distance(i, k);
6944 if (distance > curr_distance &&
6945 (distance < next_distance ||
6946 next_distance == curr_distance))
6947 next_distance = distance;
6950 * While not a strong assumption it would be nice to know
6951 * about cases where if node A is connected to B, B is not
6952 * equally connected to A.
6954 if (sched_debug() && node_distance(k, i) != distance)
6955 sched_numa_warn("Node-distance not symmetric");
6957 if (sched_debug() && i && !find_numa_distance(distance))
6958 sched_numa_warn("Node-0 not representative");
6960 if (next_distance != curr_distance) {
6961 sched_domains_numa_distance[level++] = next_distance;
6962 sched_domains_numa_levels = level;
6963 curr_distance = next_distance;
6968 * In case of sched_debug() we verify the above assumption.
6978 * 'level' contains the number of unique distances, excluding the
6979 * identity distance node_distance(i,i).
6981 * The sched_domains_numa_distance[] array includes the actual distance
6986 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6987 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6988 * the array will contain less then 'level' members. This could be
6989 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6990 * in other functions.
6992 * We reset it to 'level' at the end of this function.
6994 sched_domains_numa_levels = 0;
6996 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6997 if (!sched_domains_numa_masks)
7001 * Now for each level, construct a mask per node which contains all
7002 * cpus of nodes that are that many hops away from us.
7004 for (i = 0; i < level; i++) {
7005 sched_domains_numa_masks[i] =
7006 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
7007 if (!sched_domains_numa_masks[i])
7010 for (j = 0; j < nr_node_ids; j++) {
7011 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
7015 sched_domains_numa_masks[i][j] = mask;
7018 if (node_distance(j, k) > sched_domains_numa_distance[i])
7021 cpumask_or(mask, mask, cpumask_of_node(k));
7026 /* Compute default topology size */
7027 for (i = 0; sched_domain_topology[i].mask; i++);
7029 tl = kzalloc((i + level + 1) *
7030 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7035 * Copy the default topology bits..
7037 for (i = 0; sched_domain_topology[i].mask; i++)
7038 tl[i] = sched_domain_topology[i];
7041 * .. and append 'j' levels of NUMA goodness.
7043 for (j = 0; j < level; i++, j++) {
7044 tl[i] = (struct sched_domain_topology_level){
7045 .mask = sd_numa_mask,
7046 .sd_flags = cpu_numa_flags,
7047 .flags = SDTL_OVERLAP,
7053 sched_domain_topology = tl;
7055 sched_domains_numa_levels = level;
7056 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7058 init_numa_topology_type();
7061 static void sched_domains_numa_masks_set(int cpu)
7064 int node = cpu_to_node(cpu);
7066 for (i = 0; i < sched_domains_numa_levels; i++) {
7067 for (j = 0; j < nr_node_ids; j++) {
7068 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7069 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7074 static void sched_domains_numa_masks_clear(int cpu)
7077 for (i = 0; i < sched_domains_numa_levels; i++) {
7078 for (j = 0; j < nr_node_ids; j++)
7079 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7084 * Update sched_domains_numa_masks[level][node] array when new cpus
7087 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7088 unsigned long action,
7091 int cpu = (long)hcpu;
7093 switch (action & ~CPU_TASKS_FROZEN) {
7095 sched_domains_numa_masks_set(cpu);
7099 sched_domains_numa_masks_clear(cpu);
7109 static inline void sched_init_numa(void)
7113 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7114 unsigned long action,
7119 #endif /* CONFIG_NUMA */
7121 static int __sdt_alloc(const struct cpumask *cpu_map)
7123 struct sched_domain_topology_level *tl;
7126 for_each_sd_topology(tl) {
7127 struct sd_data *sdd = &tl->data;
7129 sdd->sd = alloc_percpu(struct sched_domain *);
7133 sdd->sg = alloc_percpu(struct sched_group *);
7137 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7141 for_each_cpu(j, cpu_map) {
7142 struct sched_domain *sd;
7143 struct sched_group *sg;
7144 struct sched_group_capacity *sgc;
7146 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7147 GFP_KERNEL, cpu_to_node(j));
7151 *per_cpu_ptr(sdd->sd, j) = sd;
7153 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7154 GFP_KERNEL, cpu_to_node(j));
7160 *per_cpu_ptr(sdd->sg, j) = sg;
7162 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7163 GFP_KERNEL, cpu_to_node(j));
7167 *per_cpu_ptr(sdd->sgc, j) = sgc;
7174 static void __sdt_free(const struct cpumask *cpu_map)
7176 struct sched_domain_topology_level *tl;
7179 for_each_sd_topology(tl) {
7180 struct sd_data *sdd = &tl->data;
7182 for_each_cpu(j, cpu_map) {
7183 struct sched_domain *sd;
7186 sd = *per_cpu_ptr(sdd->sd, j);
7187 if (sd && (sd->flags & SD_OVERLAP))
7188 free_sched_groups(sd->groups, 0);
7189 kfree(*per_cpu_ptr(sdd->sd, j));
7193 kfree(*per_cpu_ptr(sdd->sg, j));
7195 kfree(*per_cpu_ptr(sdd->sgc, j));
7197 free_percpu(sdd->sd);
7199 free_percpu(sdd->sg);
7201 free_percpu(sdd->sgc);
7206 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7207 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7208 struct sched_domain *child, int cpu)
7210 struct sched_domain *sd = sd_init(tl, cpu);
7214 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7216 sd->level = child->level + 1;
7217 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7221 if (!cpumask_subset(sched_domain_span(child),
7222 sched_domain_span(sd))) {
7223 pr_err("BUG: arch topology borken\n");
7224 #ifdef CONFIG_SCHED_DEBUG
7225 pr_err(" the %s domain not a subset of the %s domain\n",
7226 child->name, sd->name);
7228 /* Fixup, ensure @sd has at least @child cpus. */
7229 cpumask_or(sched_domain_span(sd),
7230 sched_domain_span(sd),
7231 sched_domain_span(child));
7235 set_domain_attribute(sd, attr);
7241 * Build sched domains for a given set of cpus and attach the sched domains
7242 * to the individual cpus
7244 static int build_sched_domains(const struct cpumask *cpu_map,
7245 struct sched_domain_attr *attr)
7247 enum s_alloc alloc_state;
7248 struct sched_domain *sd;
7250 int i, ret = -ENOMEM;
7252 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7253 if (alloc_state != sa_rootdomain)
7256 /* Set up domains for cpus specified by the cpu_map. */
7257 for_each_cpu(i, cpu_map) {
7258 struct sched_domain_topology_level *tl;
7261 for_each_sd_topology(tl) {
7262 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7263 if (tl == sched_domain_topology)
7264 *per_cpu_ptr(d.sd, i) = sd;
7265 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7266 sd->flags |= SD_OVERLAP;
7267 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7272 /* Build the groups for the domains */
7273 for_each_cpu(i, cpu_map) {
7274 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7275 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7276 if (sd->flags & SD_OVERLAP) {
7277 if (build_overlap_sched_groups(sd, i))
7280 if (build_sched_groups(sd, i))
7286 /* Calculate CPU capacity for physical packages and nodes */
7287 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7288 if (!cpumask_test_cpu(i, cpu_map))
7291 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7292 claim_allocations(i, sd);
7293 init_sched_groups_capacity(i, sd);
7297 /* Attach the domains */
7299 for_each_cpu(i, cpu_map) {
7300 sd = *per_cpu_ptr(d.sd, i);
7301 cpu_attach_domain(sd, d.rd, i);
7307 __free_domain_allocs(&d, alloc_state, cpu_map);
7311 static cpumask_var_t *doms_cur; /* current sched domains */
7312 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7313 static struct sched_domain_attr *dattr_cur;
7314 /* attribues of custom domains in 'doms_cur' */
7317 * Special case: If a kmalloc of a doms_cur partition (array of
7318 * cpumask) fails, then fallback to a single sched domain,
7319 * as determined by the single cpumask fallback_doms.
7321 static cpumask_var_t fallback_doms;
7324 * arch_update_cpu_topology lets virtualized architectures update the
7325 * cpu core maps. It is supposed to return 1 if the topology changed
7326 * or 0 if it stayed the same.
7328 int __weak arch_update_cpu_topology(void)
7333 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7336 cpumask_var_t *doms;
7338 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7341 for (i = 0; i < ndoms; i++) {
7342 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7343 free_sched_domains(doms, i);
7350 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7353 for (i = 0; i < ndoms; i++)
7354 free_cpumask_var(doms[i]);
7359 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7360 * For now this just excludes isolated cpus, but could be used to
7361 * exclude other special cases in the future.
7363 static int init_sched_domains(const struct cpumask *cpu_map)
7367 arch_update_cpu_topology();
7369 doms_cur = alloc_sched_domains(ndoms_cur);
7371 doms_cur = &fallback_doms;
7372 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7373 err = build_sched_domains(doms_cur[0], NULL);
7374 register_sched_domain_sysctl();
7380 * Detach sched domains from a group of cpus specified in cpu_map
7381 * These cpus will now be attached to the NULL domain
7383 static void detach_destroy_domains(const struct cpumask *cpu_map)
7388 for_each_cpu(i, cpu_map)
7389 cpu_attach_domain(NULL, &def_root_domain, i);
7393 /* handle null as "default" */
7394 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7395 struct sched_domain_attr *new, int idx_new)
7397 struct sched_domain_attr tmp;
7404 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7405 new ? (new + idx_new) : &tmp,
7406 sizeof(struct sched_domain_attr));
7410 * Partition sched domains as specified by the 'ndoms_new'
7411 * cpumasks in the array doms_new[] of cpumasks. This compares
7412 * doms_new[] to the current sched domain partitioning, doms_cur[].
7413 * It destroys each deleted domain and builds each new domain.
7415 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7416 * The masks don't intersect (don't overlap.) We should setup one
7417 * sched domain for each mask. CPUs not in any of the cpumasks will
7418 * not be load balanced. If the same cpumask appears both in the
7419 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7422 * The passed in 'doms_new' should be allocated using
7423 * alloc_sched_domains. This routine takes ownership of it and will
7424 * free_sched_domains it when done with it. If the caller failed the
7425 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7426 * and partition_sched_domains() will fallback to the single partition
7427 * 'fallback_doms', it also forces the domains to be rebuilt.
7429 * If doms_new == NULL it will be replaced with cpu_online_mask.
7430 * ndoms_new == 0 is a special case for destroying existing domains,
7431 * and it will not create the default domain.
7433 * Call with hotplug lock held
7435 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7436 struct sched_domain_attr *dattr_new)
7441 mutex_lock(&sched_domains_mutex);
7443 /* always unregister in case we don't destroy any domains */
7444 unregister_sched_domain_sysctl();
7446 /* Let architecture update cpu core mappings. */
7447 new_topology = arch_update_cpu_topology();
7449 n = doms_new ? ndoms_new : 0;
7451 /* Destroy deleted domains */
7452 for (i = 0; i < ndoms_cur; i++) {
7453 for (j = 0; j < n && !new_topology; j++) {
7454 if (cpumask_equal(doms_cur[i], doms_new[j])
7455 && dattrs_equal(dattr_cur, i, dattr_new, j))
7458 /* no match - a current sched domain not in new doms_new[] */
7459 detach_destroy_domains(doms_cur[i]);
7465 if (doms_new == NULL) {
7467 doms_new = &fallback_doms;
7468 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7469 WARN_ON_ONCE(dattr_new);
7472 /* Build new domains */
7473 for (i = 0; i < ndoms_new; i++) {
7474 for (j = 0; j < n && !new_topology; j++) {
7475 if (cpumask_equal(doms_new[i], doms_cur[j])
7476 && dattrs_equal(dattr_new, i, dattr_cur, j))
7479 /* no match - add a new doms_new */
7480 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7485 /* Remember the new sched domains */
7486 if (doms_cur != &fallback_doms)
7487 free_sched_domains(doms_cur, ndoms_cur);
7488 kfree(dattr_cur); /* kfree(NULL) is safe */
7489 doms_cur = doms_new;
7490 dattr_cur = dattr_new;
7491 ndoms_cur = ndoms_new;
7493 register_sched_domain_sysctl();
7495 mutex_unlock(&sched_domains_mutex);
7498 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7501 * Update cpusets according to cpu_active mask. If cpusets are
7502 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7503 * around partition_sched_domains().
7505 * If we come here as part of a suspend/resume, don't touch cpusets because we
7506 * want to restore it back to its original state upon resume anyway.
7508 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7512 case CPU_ONLINE_FROZEN:
7513 case CPU_DOWN_FAILED_FROZEN:
7516 * num_cpus_frozen tracks how many CPUs are involved in suspend
7517 * resume sequence. As long as this is not the last online
7518 * operation in the resume sequence, just build a single sched
7519 * domain, ignoring cpusets.
7522 if (likely(num_cpus_frozen)) {
7523 partition_sched_domains(1, NULL, NULL);
7528 * This is the last CPU online operation. So fall through and
7529 * restore the original sched domains by considering the
7530 * cpuset configurations.
7534 cpuset_update_active_cpus(true);
7542 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7545 unsigned long flags;
7546 long cpu = (long)hcpu;
7552 case CPU_DOWN_PREPARE:
7553 rcu_read_lock_sched();
7554 dl_b = dl_bw_of(cpu);
7556 raw_spin_lock_irqsave(&dl_b->lock, flags);
7557 cpus = dl_bw_cpus(cpu);
7558 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7559 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7561 rcu_read_unlock_sched();
7564 return notifier_from_errno(-EBUSY);
7565 cpuset_update_active_cpus(false);
7567 case CPU_DOWN_PREPARE_FROZEN:
7569 partition_sched_domains(1, NULL, NULL);
7577 void __init sched_init_smp(void)
7579 cpumask_var_t non_isolated_cpus;
7581 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7582 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7587 * There's no userspace yet to cause hotplug operations; hence all the
7588 * cpu masks are stable and all blatant races in the below code cannot
7591 mutex_lock(&sched_domains_mutex);
7592 init_sched_domains(cpu_active_mask);
7593 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7594 if (cpumask_empty(non_isolated_cpus))
7595 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7596 mutex_unlock(&sched_domains_mutex);
7598 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7599 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7600 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7604 /* Move init over to a non-isolated CPU */
7605 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7607 sched_init_granularity();
7608 free_cpumask_var(non_isolated_cpus);
7610 init_sched_rt_class();
7611 init_sched_dl_class();
7614 void __init sched_init_smp(void)
7616 sched_init_granularity();
7618 #endif /* CONFIG_SMP */
7620 int in_sched_functions(unsigned long addr)
7622 return in_lock_functions(addr) ||
7623 (addr >= (unsigned long)__sched_text_start
7624 && addr < (unsigned long)__sched_text_end);
7627 #ifdef CONFIG_CGROUP_SCHED
7629 * Default task group.
7630 * Every task in system belongs to this group at bootup.
7632 struct task_group root_task_group;
7633 LIST_HEAD(task_groups);
7636 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7638 void __init sched_init(void)
7641 unsigned long alloc_size = 0, ptr;
7643 #ifdef CONFIG_FAIR_GROUP_SCHED
7644 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7646 #ifdef CONFIG_RT_GROUP_SCHED
7647 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7650 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7652 #ifdef CONFIG_FAIR_GROUP_SCHED
7653 root_task_group.se = (struct sched_entity **)ptr;
7654 ptr += nr_cpu_ids * sizeof(void **);
7656 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7657 ptr += nr_cpu_ids * sizeof(void **);
7659 #endif /* CONFIG_FAIR_GROUP_SCHED */
7660 #ifdef CONFIG_RT_GROUP_SCHED
7661 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7662 ptr += nr_cpu_ids * sizeof(void **);
7664 root_task_group.rt_rq = (struct rt_rq **)ptr;
7665 ptr += nr_cpu_ids * sizeof(void **);
7667 #endif /* CONFIG_RT_GROUP_SCHED */
7669 #ifdef CONFIG_CPUMASK_OFFSTACK
7670 for_each_possible_cpu(i) {
7671 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7672 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7674 #endif /* CONFIG_CPUMASK_OFFSTACK */
7676 init_rt_bandwidth(&def_rt_bandwidth,
7677 global_rt_period(), global_rt_runtime());
7678 init_dl_bandwidth(&def_dl_bandwidth,
7679 global_rt_period(), global_rt_runtime());
7682 init_defrootdomain();
7685 #ifdef CONFIG_RT_GROUP_SCHED
7686 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7687 global_rt_period(), global_rt_runtime());
7688 #endif /* CONFIG_RT_GROUP_SCHED */
7690 #ifdef CONFIG_CGROUP_SCHED
7691 list_add(&root_task_group.list, &task_groups);
7692 INIT_LIST_HEAD(&root_task_group.children);
7693 INIT_LIST_HEAD(&root_task_group.siblings);
7694 autogroup_init(&init_task);
7696 #endif /* CONFIG_CGROUP_SCHED */
7698 for_each_possible_cpu(i) {
7702 raw_spin_lock_init(&rq->lock);
7704 rq->calc_load_active = 0;
7705 rq->calc_load_update = jiffies + LOAD_FREQ;
7706 init_cfs_rq(&rq->cfs);
7707 init_rt_rq(&rq->rt);
7708 init_dl_rq(&rq->dl);
7709 #ifdef CONFIG_FAIR_GROUP_SCHED
7710 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7711 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7713 * How much cpu bandwidth does root_task_group get?
7715 * In case of task-groups formed thr' the cgroup filesystem, it
7716 * gets 100% of the cpu resources in the system. This overall
7717 * system cpu resource is divided among the tasks of
7718 * root_task_group and its child task-groups in a fair manner,
7719 * based on each entity's (task or task-group's) weight
7720 * (se->load.weight).
7722 * In other words, if root_task_group has 10 tasks of weight
7723 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7724 * then A0's share of the cpu resource is:
7726 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7728 * We achieve this by letting root_task_group's tasks sit
7729 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7731 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7732 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7733 #endif /* CONFIG_FAIR_GROUP_SCHED */
7735 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7736 #ifdef CONFIG_RT_GROUP_SCHED
7737 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7740 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7741 rq->cpu_load[j] = 0;
7743 rq->last_load_update_tick = jiffies;
7748 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7749 rq->balance_callback = NULL;
7750 rq->active_balance = 0;
7751 rq->next_balance = jiffies;
7756 rq->avg_idle = 2*sysctl_sched_migration_cost;
7757 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7759 INIT_LIST_HEAD(&rq->cfs_tasks);
7761 rq_attach_root(rq, &def_root_domain);
7762 #ifdef CONFIG_NO_HZ_COMMON
7765 #ifdef CONFIG_NO_HZ_FULL
7766 rq->last_sched_tick = 0;
7770 atomic_set(&rq->nr_iowait, 0);
7773 set_load_weight(&init_task);
7775 #ifdef CONFIG_PREEMPT_NOTIFIERS
7776 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7780 * The boot idle thread does lazy MMU switching as well:
7782 atomic_inc(&init_mm.mm_count);
7783 enter_lazy_tlb(&init_mm, current);
7786 * During early bootup we pretend to be a normal task:
7788 current->sched_class = &fair_sched_class;
7791 * Make us the idle thread. Technically, schedule() should not be
7792 * called from this thread, however somewhere below it might be,
7793 * but because we are the idle thread, we just pick up running again
7794 * when this runqueue becomes "idle".
7796 init_idle(current, smp_processor_id());
7798 calc_load_update = jiffies + LOAD_FREQ;
7801 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7802 /* May be allocated at isolcpus cmdline parse time */
7803 if (cpu_isolated_map == NULL)
7804 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7805 idle_thread_set_boot_cpu();
7806 set_cpu_rq_start_time();
7808 init_sched_fair_class();
7810 scheduler_running = 1;
7813 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7814 static inline int preempt_count_equals(int preempt_offset)
7816 int nested = preempt_count() + sched_rcu_preempt_depth();
7818 return (nested == preempt_offset);
7821 void __might_sleep(const char *file, int line, int preempt_offset)
7824 * Blocking primitives will set (and therefore destroy) current->state,
7825 * since we will exit with TASK_RUNNING make sure we enter with it,
7826 * otherwise we will destroy state.
7828 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7829 "do not call blocking ops when !TASK_RUNNING; "
7830 "state=%lx set at [<%p>] %pS\n",
7832 (void *)current->task_state_change,
7833 (void *)current->task_state_change);
7835 ___might_sleep(file, line, preempt_offset);
7837 EXPORT_SYMBOL(__might_sleep);
7839 void ___might_sleep(const char *file, int line, int preempt_offset)
7841 static unsigned long prev_jiffy; /* ratelimiting */
7843 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7844 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7845 !is_idle_task(current)) ||
7846 system_state != SYSTEM_RUNNING || oops_in_progress)
7848 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7850 prev_jiffy = jiffies;
7853 "BUG: sleeping function called from invalid context at %s:%d\n",
7856 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7857 in_atomic(), irqs_disabled(),
7858 current->pid, current->comm);
7860 if (task_stack_end_corrupted(current))
7861 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7863 debug_show_held_locks(current);
7864 if (irqs_disabled())
7865 print_irqtrace_events(current);
7866 #ifdef CONFIG_DEBUG_PREEMPT
7867 if (!preempt_count_equals(preempt_offset)) {
7868 pr_err("Preemption disabled at:");
7869 print_ip_sym(current->preempt_disable_ip);
7875 EXPORT_SYMBOL(___might_sleep);
7878 #ifdef CONFIG_MAGIC_SYSRQ
7879 void normalize_rt_tasks(void)
7881 struct task_struct *g, *p;
7882 struct sched_attr attr = {
7883 .sched_policy = SCHED_NORMAL,
7886 read_lock(&tasklist_lock);
7887 for_each_process_thread(g, p) {
7889 * Only normalize user tasks:
7891 if (p->flags & PF_KTHREAD)
7894 p->se.exec_start = 0;
7895 #ifdef CONFIG_SCHEDSTATS
7896 p->se.statistics.wait_start = 0;
7897 p->se.statistics.sleep_start = 0;
7898 p->se.statistics.block_start = 0;
7901 if (!dl_task(p) && !rt_task(p)) {
7903 * Renice negative nice level userspace
7906 if (task_nice(p) < 0)
7907 set_user_nice(p, 0);
7911 __sched_setscheduler(p, &attr, false, false);
7913 read_unlock(&tasklist_lock);
7916 #endif /* CONFIG_MAGIC_SYSRQ */
7918 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7920 * These functions are only useful for the IA64 MCA handling, or kdb.
7922 * They can only be called when the whole system has been
7923 * stopped - every CPU needs to be quiescent, and no scheduling
7924 * activity can take place. Using them for anything else would
7925 * be a serious bug, and as a result, they aren't even visible
7926 * under any other configuration.
7930 * curr_task - return the current task for a given cpu.
7931 * @cpu: the processor in question.
7933 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7935 * Return: The current task for @cpu.
7937 struct task_struct *curr_task(int cpu)
7939 return cpu_curr(cpu);
7942 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7946 * set_curr_task - set the current task for a given cpu.
7947 * @cpu: the processor in question.
7948 * @p: the task pointer to set.
7950 * Description: This function must only be used when non-maskable interrupts
7951 * are serviced on a separate stack. It allows the architecture to switch the
7952 * notion of the current task on a cpu in a non-blocking manner. This function
7953 * must be called with all CPU's synchronized, and interrupts disabled, the
7954 * and caller must save the original value of the current task (see
7955 * curr_task() above) and restore that value before reenabling interrupts and
7956 * re-starting the system.
7958 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7960 void set_curr_task(int cpu, struct task_struct *p)
7967 #ifdef CONFIG_CGROUP_SCHED
7968 /* task_group_lock serializes the addition/removal of task groups */
7969 static DEFINE_SPINLOCK(task_group_lock);
7971 static void sched_free_group(struct task_group *tg)
7973 free_fair_sched_group(tg);
7974 free_rt_sched_group(tg);
7979 /* allocate runqueue etc for a new task group */
7980 struct task_group *sched_create_group(struct task_group *parent)
7982 struct task_group *tg;
7984 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7986 return ERR_PTR(-ENOMEM);
7988 if (!alloc_fair_sched_group(tg, parent))
7991 if (!alloc_rt_sched_group(tg, parent))
7997 sched_free_group(tg);
7998 return ERR_PTR(-ENOMEM);
8001 void sched_online_group(struct task_group *tg, struct task_group *parent)
8003 unsigned long flags;
8005 spin_lock_irqsave(&task_group_lock, flags);
8006 list_add_rcu(&tg->list, &task_groups);
8008 WARN_ON(!parent); /* root should already exist */
8010 tg->parent = parent;
8011 INIT_LIST_HEAD(&tg->children);
8012 list_add_rcu(&tg->siblings, &parent->children);
8013 spin_unlock_irqrestore(&task_group_lock, flags);
8016 /* rcu callback to free various structures associated with a task group */
8017 static void sched_free_group_rcu(struct rcu_head *rhp)
8019 /* now it should be safe to free those cfs_rqs */
8020 sched_free_group(container_of(rhp, struct task_group, rcu));
8023 void sched_destroy_group(struct task_group *tg)
8025 /* wait for possible concurrent references to cfs_rqs complete */
8026 call_rcu(&tg->rcu, sched_free_group_rcu);
8029 void sched_offline_group(struct task_group *tg)
8031 unsigned long flags;
8034 /* end participation in shares distribution */
8035 for_each_possible_cpu(i)
8036 unregister_fair_sched_group(tg, i);
8038 spin_lock_irqsave(&task_group_lock, flags);
8039 list_del_rcu(&tg->list);
8040 list_del_rcu(&tg->siblings);
8041 spin_unlock_irqrestore(&task_group_lock, flags);
8044 /* change task's runqueue when it moves between groups.
8045 * The caller of this function should have put the task in its new group
8046 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8047 * reflect its new group.
8049 void sched_move_task(struct task_struct *tsk)
8051 struct task_group *tg;
8052 int queued, running;
8053 unsigned long flags;
8056 rq = task_rq_lock(tsk, &flags);
8058 running = task_current(rq, tsk);
8059 queued = task_on_rq_queued(tsk);
8062 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8063 if (unlikely(running))
8064 put_prev_task(rq, tsk);
8067 * All callers are synchronized by task_rq_lock(); we do not use RCU
8068 * which is pointless here. Thus, we pass "true" to task_css_check()
8069 * to prevent lockdep warnings.
8071 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8072 struct task_group, css);
8073 tg = autogroup_task_group(tsk, tg);
8074 tsk->sched_task_group = tg;
8076 #ifdef CONFIG_FAIR_GROUP_SCHED
8077 if (tsk->sched_class->task_move_group)
8078 tsk->sched_class->task_move_group(tsk);
8081 set_task_rq(tsk, task_cpu(tsk));
8083 if (unlikely(running))
8084 tsk->sched_class->set_curr_task(rq);
8086 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8088 task_rq_unlock(rq, tsk, &flags);
8090 #endif /* CONFIG_CGROUP_SCHED */
8092 #ifdef CONFIG_RT_GROUP_SCHED
8094 * Ensure that the real time constraints are schedulable.
8096 static DEFINE_MUTEX(rt_constraints_mutex);
8098 /* Must be called with tasklist_lock held */
8099 static inline int tg_has_rt_tasks(struct task_group *tg)
8101 struct task_struct *g, *p;
8104 * Autogroups do not have RT tasks; see autogroup_create().
8106 if (task_group_is_autogroup(tg))
8109 for_each_process_thread(g, p) {
8110 if (rt_task(p) && task_group(p) == tg)
8117 struct rt_schedulable_data {
8118 struct task_group *tg;
8123 static int tg_rt_schedulable(struct task_group *tg, void *data)
8125 struct rt_schedulable_data *d = data;
8126 struct task_group *child;
8127 unsigned long total, sum = 0;
8128 u64 period, runtime;
8130 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8131 runtime = tg->rt_bandwidth.rt_runtime;
8134 period = d->rt_period;
8135 runtime = d->rt_runtime;
8139 * Cannot have more runtime than the period.
8141 if (runtime > period && runtime != RUNTIME_INF)
8145 * Ensure we don't starve existing RT tasks.
8147 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8150 total = to_ratio(period, runtime);
8153 * Nobody can have more than the global setting allows.
8155 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8159 * The sum of our children's runtime should not exceed our own.
8161 list_for_each_entry_rcu(child, &tg->children, siblings) {
8162 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8163 runtime = child->rt_bandwidth.rt_runtime;
8165 if (child == d->tg) {
8166 period = d->rt_period;
8167 runtime = d->rt_runtime;
8170 sum += to_ratio(period, runtime);
8179 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8183 struct rt_schedulable_data data = {
8185 .rt_period = period,
8186 .rt_runtime = runtime,
8190 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8196 static int tg_set_rt_bandwidth(struct task_group *tg,
8197 u64 rt_period, u64 rt_runtime)
8202 * Disallowing the root group RT runtime is BAD, it would disallow the
8203 * kernel creating (and or operating) RT threads.
8205 if (tg == &root_task_group && rt_runtime == 0)
8208 /* No period doesn't make any sense. */
8212 mutex_lock(&rt_constraints_mutex);
8213 read_lock(&tasklist_lock);
8214 err = __rt_schedulable(tg, rt_period, rt_runtime);
8218 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8219 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8220 tg->rt_bandwidth.rt_runtime = rt_runtime;
8222 for_each_possible_cpu(i) {
8223 struct rt_rq *rt_rq = tg->rt_rq[i];
8225 raw_spin_lock(&rt_rq->rt_runtime_lock);
8226 rt_rq->rt_runtime = rt_runtime;
8227 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8229 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8231 read_unlock(&tasklist_lock);
8232 mutex_unlock(&rt_constraints_mutex);
8237 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8239 u64 rt_runtime, rt_period;
8241 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8242 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8243 if (rt_runtime_us < 0)
8244 rt_runtime = RUNTIME_INF;
8246 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8249 static long sched_group_rt_runtime(struct task_group *tg)
8253 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8256 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8257 do_div(rt_runtime_us, NSEC_PER_USEC);
8258 return rt_runtime_us;
8261 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8263 u64 rt_runtime, rt_period;
8265 rt_period = rt_period_us * NSEC_PER_USEC;
8266 rt_runtime = tg->rt_bandwidth.rt_runtime;
8268 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8271 static long sched_group_rt_period(struct task_group *tg)
8275 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8276 do_div(rt_period_us, NSEC_PER_USEC);
8277 return rt_period_us;
8279 #endif /* CONFIG_RT_GROUP_SCHED */
8281 #ifdef CONFIG_RT_GROUP_SCHED
8282 static int sched_rt_global_constraints(void)
8286 mutex_lock(&rt_constraints_mutex);
8287 read_lock(&tasklist_lock);
8288 ret = __rt_schedulable(NULL, 0, 0);
8289 read_unlock(&tasklist_lock);
8290 mutex_unlock(&rt_constraints_mutex);
8295 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8297 /* Don't accept realtime tasks when there is no way for them to run */
8298 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8304 #else /* !CONFIG_RT_GROUP_SCHED */
8305 static int sched_rt_global_constraints(void)
8307 unsigned long flags;
8310 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8311 for_each_possible_cpu(i) {
8312 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8314 raw_spin_lock(&rt_rq->rt_runtime_lock);
8315 rt_rq->rt_runtime = global_rt_runtime();
8316 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8318 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8322 #endif /* CONFIG_RT_GROUP_SCHED */
8324 static int sched_dl_global_validate(void)
8326 u64 runtime = global_rt_runtime();
8327 u64 period = global_rt_period();
8328 u64 new_bw = to_ratio(period, runtime);
8331 unsigned long flags;
8334 * Here we want to check the bandwidth not being set to some
8335 * value smaller than the currently allocated bandwidth in
8336 * any of the root_domains.
8338 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8339 * cycling on root_domains... Discussion on different/better
8340 * solutions is welcome!
8342 for_each_possible_cpu(cpu) {
8343 rcu_read_lock_sched();
8344 dl_b = dl_bw_of(cpu);
8346 raw_spin_lock_irqsave(&dl_b->lock, flags);
8347 if (new_bw < dl_b->total_bw)
8349 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8351 rcu_read_unlock_sched();
8360 static void sched_dl_do_global(void)
8365 unsigned long flags;
8367 def_dl_bandwidth.dl_period = global_rt_period();
8368 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8370 if (global_rt_runtime() != RUNTIME_INF)
8371 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8374 * FIXME: As above...
8376 for_each_possible_cpu(cpu) {
8377 rcu_read_lock_sched();
8378 dl_b = dl_bw_of(cpu);
8380 raw_spin_lock_irqsave(&dl_b->lock, flags);
8382 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8384 rcu_read_unlock_sched();
8388 static int sched_rt_global_validate(void)
8390 if (sysctl_sched_rt_period <= 0)
8393 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8394 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8400 static void sched_rt_do_global(void)
8402 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8403 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8406 int sched_rt_handler(struct ctl_table *table, int write,
8407 void __user *buffer, size_t *lenp,
8410 int old_period, old_runtime;
8411 static DEFINE_MUTEX(mutex);
8415 old_period = sysctl_sched_rt_period;
8416 old_runtime = sysctl_sched_rt_runtime;
8418 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8420 if (!ret && write) {
8421 ret = sched_rt_global_validate();
8425 ret = sched_dl_global_validate();
8429 ret = sched_rt_global_constraints();
8433 sched_rt_do_global();
8434 sched_dl_do_global();
8438 sysctl_sched_rt_period = old_period;
8439 sysctl_sched_rt_runtime = old_runtime;
8441 mutex_unlock(&mutex);
8446 int sched_rr_handler(struct ctl_table *table, int write,
8447 void __user *buffer, size_t *lenp,
8451 static DEFINE_MUTEX(mutex);
8454 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8455 /* make sure that internally we keep jiffies */
8456 /* also, writing zero resets timeslice to default */
8457 if (!ret && write) {
8458 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8459 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8461 mutex_unlock(&mutex);
8465 #ifdef CONFIG_CGROUP_SCHED
8467 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8469 return css ? container_of(css, struct task_group, css) : NULL;
8472 static struct cgroup_subsys_state *
8473 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8475 struct task_group *parent = css_tg(parent_css);
8476 struct task_group *tg;
8479 /* This is early initialization for the top cgroup */
8480 return &root_task_group.css;
8483 tg = sched_create_group(parent);
8485 return ERR_PTR(-ENOMEM);
8487 sched_online_group(tg, parent);
8492 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8494 struct task_group *tg = css_tg(css);
8496 sched_offline_group(tg);
8499 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8501 struct task_group *tg = css_tg(css);
8504 * Relies on the RCU grace period between css_released() and this.
8506 sched_free_group(tg);
8509 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8511 sched_move_task(task);
8514 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8516 struct task_struct *task;
8517 struct cgroup_subsys_state *css;
8519 cgroup_taskset_for_each(task, css, tset) {
8520 #ifdef CONFIG_RT_GROUP_SCHED
8521 if (!sched_rt_can_attach(css_tg(css), task))
8524 /* We don't support RT-tasks being in separate groups */
8525 if (task->sched_class != &fair_sched_class)
8532 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8534 struct task_struct *task;
8535 struct cgroup_subsys_state *css;
8537 cgroup_taskset_for_each(task, css, tset)
8538 sched_move_task(task);
8541 #ifdef CONFIG_FAIR_GROUP_SCHED
8542 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8543 struct cftype *cftype, u64 shareval)
8545 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8548 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8551 struct task_group *tg = css_tg(css);
8553 return (u64) scale_load_down(tg->shares);
8556 #ifdef CONFIG_CFS_BANDWIDTH
8557 static DEFINE_MUTEX(cfs_constraints_mutex);
8559 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8560 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8562 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8564 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8566 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8567 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8569 if (tg == &root_task_group)
8573 * Ensure we have at some amount of bandwidth every period. This is
8574 * to prevent reaching a state of large arrears when throttled via
8575 * entity_tick() resulting in prolonged exit starvation.
8577 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8581 * Likewise, bound things on the otherside by preventing insane quota
8582 * periods. This also allows us to normalize in computing quota
8585 if (period > max_cfs_quota_period)
8589 * Prevent race between setting of cfs_rq->runtime_enabled and
8590 * unthrottle_offline_cfs_rqs().
8593 mutex_lock(&cfs_constraints_mutex);
8594 ret = __cfs_schedulable(tg, period, quota);
8598 runtime_enabled = quota != RUNTIME_INF;
8599 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8601 * If we need to toggle cfs_bandwidth_used, off->on must occur
8602 * before making related changes, and on->off must occur afterwards
8604 if (runtime_enabled && !runtime_was_enabled)
8605 cfs_bandwidth_usage_inc();
8606 raw_spin_lock_irq(&cfs_b->lock);
8607 cfs_b->period = ns_to_ktime(period);
8608 cfs_b->quota = quota;
8610 __refill_cfs_bandwidth_runtime(cfs_b);
8611 /* restart the period timer (if active) to handle new period expiry */
8612 if (runtime_enabled)
8613 start_cfs_bandwidth(cfs_b);
8614 raw_spin_unlock_irq(&cfs_b->lock);
8616 for_each_online_cpu(i) {
8617 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8618 struct rq *rq = cfs_rq->rq;
8620 raw_spin_lock_irq(&rq->lock);
8621 cfs_rq->runtime_enabled = runtime_enabled;
8622 cfs_rq->runtime_remaining = 0;
8624 if (cfs_rq->throttled)
8625 unthrottle_cfs_rq(cfs_rq);
8626 raw_spin_unlock_irq(&rq->lock);
8628 if (runtime_was_enabled && !runtime_enabled)
8629 cfs_bandwidth_usage_dec();
8631 mutex_unlock(&cfs_constraints_mutex);
8637 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8641 period = ktime_to_ns(tg->cfs_bandwidth.period);
8642 if (cfs_quota_us < 0)
8643 quota = RUNTIME_INF;
8645 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8647 return tg_set_cfs_bandwidth(tg, period, quota);
8650 long tg_get_cfs_quota(struct task_group *tg)
8654 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8657 quota_us = tg->cfs_bandwidth.quota;
8658 do_div(quota_us, NSEC_PER_USEC);
8663 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8667 period = (u64)cfs_period_us * NSEC_PER_USEC;
8668 quota = tg->cfs_bandwidth.quota;
8670 return tg_set_cfs_bandwidth(tg, period, quota);
8673 long tg_get_cfs_period(struct task_group *tg)
8677 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8678 do_div(cfs_period_us, NSEC_PER_USEC);
8680 return cfs_period_us;
8683 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8686 return tg_get_cfs_quota(css_tg(css));
8689 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8690 struct cftype *cftype, s64 cfs_quota_us)
8692 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8695 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8698 return tg_get_cfs_period(css_tg(css));
8701 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8702 struct cftype *cftype, u64 cfs_period_us)
8704 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8707 struct cfs_schedulable_data {
8708 struct task_group *tg;
8713 * normalize group quota/period to be quota/max_period
8714 * note: units are usecs
8716 static u64 normalize_cfs_quota(struct task_group *tg,
8717 struct cfs_schedulable_data *d)
8725 period = tg_get_cfs_period(tg);
8726 quota = tg_get_cfs_quota(tg);
8729 /* note: these should typically be equivalent */
8730 if (quota == RUNTIME_INF || quota == -1)
8733 return to_ratio(period, quota);
8736 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8738 struct cfs_schedulable_data *d = data;
8739 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8740 s64 quota = 0, parent_quota = -1;
8743 quota = RUNTIME_INF;
8745 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8747 quota = normalize_cfs_quota(tg, d);
8748 parent_quota = parent_b->hierarchical_quota;
8751 * ensure max(child_quota) <= parent_quota, inherit when no
8754 if (quota == RUNTIME_INF)
8755 quota = parent_quota;
8756 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8759 cfs_b->hierarchical_quota = quota;
8764 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8767 struct cfs_schedulable_data data = {
8773 if (quota != RUNTIME_INF) {
8774 do_div(data.period, NSEC_PER_USEC);
8775 do_div(data.quota, NSEC_PER_USEC);
8779 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8785 static int cpu_stats_show(struct seq_file *sf, void *v)
8787 struct task_group *tg = css_tg(seq_css(sf));
8788 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8790 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8791 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8792 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8796 #endif /* CONFIG_CFS_BANDWIDTH */
8797 #endif /* CONFIG_FAIR_GROUP_SCHED */
8799 #ifdef CONFIG_RT_GROUP_SCHED
8800 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8801 struct cftype *cft, s64 val)
8803 return sched_group_set_rt_runtime(css_tg(css), val);
8806 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8809 return sched_group_rt_runtime(css_tg(css));
8812 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8813 struct cftype *cftype, u64 rt_period_us)
8815 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8818 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8821 return sched_group_rt_period(css_tg(css));
8823 #endif /* CONFIG_RT_GROUP_SCHED */
8825 static struct cftype cpu_files[] = {
8826 #ifdef CONFIG_FAIR_GROUP_SCHED
8829 .read_u64 = cpu_shares_read_u64,
8830 .write_u64 = cpu_shares_write_u64,
8833 #ifdef CONFIG_CFS_BANDWIDTH
8835 .name = "cfs_quota_us",
8836 .read_s64 = cpu_cfs_quota_read_s64,
8837 .write_s64 = cpu_cfs_quota_write_s64,
8840 .name = "cfs_period_us",
8841 .read_u64 = cpu_cfs_period_read_u64,
8842 .write_u64 = cpu_cfs_period_write_u64,
8846 .seq_show = cpu_stats_show,
8849 #ifdef CONFIG_RT_GROUP_SCHED
8851 .name = "rt_runtime_us",
8852 .read_s64 = cpu_rt_runtime_read,
8853 .write_s64 = cpu_rt_runtime_write,
8856 .name = "rt_period_us",
8857 .read_u64 = cpu_rt_period_read_uint,
8858 .write_u64 = cpu_rt_period_write_uint,
8864 struct cgroup_subsys cpu_cgrp_subsys = {
8865 .css_alloc = cpu_cgroup_css_alloc,
8866 .css_released = cpu_cgroup_css_released,
8867 .css_free = cpu_cgroup_css_free,
8868 .fork = cpu_cgroup_fork,
8869 .can_attach = cpu_cgroup_can_attach,
8870 .attach = cpu_cgroup_attach,
8871 .legacy_cftypes = cpu_files,
8875 #endif /* CONFIG_CGROUP_SCHED */
8877 void dump_cpu_task(int cpu)
8879 pr_info("Task dump for CPU %d:\n", cpu);
8880 sched_show_task(cpu_curr(cpu));