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>
94 DEFINE_MUTEX(sched_domains_mutex);
95 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
97 static void update_rq_clock_task(struct rq *rq, s64 delta);
99 void update_rq_clock(struct rq *rq)
103 lockdep_assert_held(&rq->lock);
105 if (rq->clock_skip_update & RQCF_ACT_SKIP)
108 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
112 update_rq_clock_task(rq, delta);
116 * Debugging: various feature bits
119 #define SCHED_FEAT(name, enabled) \
120 (1UL << __SCHED_FEAT_##name) * enabled |
122 const_debug unsigned int sysctl_sched_features =
123 #include "features.h"
128 #ifdef CONFIG_SCHED_DEBUG
129 #define SCHED_FEAT(name, enabled) \
132 static const char * const sched_feat_names[] = {
133 #include "features.h"
138 static int sched_feat_show(struct seq_file *m, void *v)
142 for (i = 0; i < __SCHED_FEAT_NR; i++) {
143 if (!(sysctl_sched_features & (1UL << i)))
145 seq_printf(m, "%s ", sched_feat_names[i]);
152 #ifdef HAVE_JUMP_LABEL
154 #define jump_label_key__true STATIC_KEY_INIT_TRUE
155 #define jump_label_key__false STATIC_KEY_INIT_FALSE
157 #define SCHED_FEAT(name, enabled) \
158 jump_label_key__##enabled ,
160 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
161 #include "features.h"
166 static void sched_feat_disable(int i)
168 static_key_disable(&sched_feat_keys[i]);
171 static void sched_feat_enable(int i)
173 static_key_enable(&sched_feat_keys[i]);
176 static void sched_feat_disable(int i) { };
177 static void sched_feat_enable(int i) { };
178 #endif /* HAVE_JUMP_LABEL */
180 static int sched_feat_set(char *cmp)
185 if (strncmp(cmp, "NO_", 3) == 0) {
190 for (i = 0; i < __SCHED_FEAT_NR; i++) {
191 if (strcmp(cmp, sched_feat_names[i]) == 0) {
193 sysctl_sched_features &= ~(1UL << i);
194 sched_feat_disable(i);
196 sysctl_sched_features |= (1UL << i);
197 sched_feat_enable(i);
207 sched_feat_write(struct file *filp, const char __user *ubuf,
208 size_t cnt, loff_t *ppos)
218 if (copy_from_user(&buf, ubuf, cnt))
224 /* Ensure the static_key remains in a consistent state */
225 inode = file_inode(filp);
226 mutex_lock(&inode->i_mutex);
227 i = sched_feat_set(cmp);
228 mutex_unlock(&inode->i_mutex);
229 if (i == __SCHED_FEAT_NR)
237 static int sched_feat_open(struct inode *inode, struct file *filp)
239 return single_open(filp, sched_feat_show, NULL);
242 static const struct file_operations sched_feat_fops = {
243 .open = sched_feat_open,
244 .write = sched_feat_write,
247 .release = single_release,
250 static __init int sched_init_debug(void)
252 debugfs_create_file("sched_features", 0644, NULL, NULL,
257 late_initcall(sched_init_debug);
258 #endif /* CONFIG_SCHED_DEBUG */
261 * Number of tasks to iterate in a single balance run.
262 * Limited because this is done with IRQs disabled.
264 const_debug unsigned int sysctl_sched_nr_migrate = 32;
267 * period over which we average the RT time consumption, measured
272 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
275 * period over which we measure -rt task cpu usage in us.
278 unsigned int sysctl_sched_rt_period = 1000000;
280 __read_mostly int scheduler_running;
283 * part of the period that we allow rt tasks to run in us.
286 int sysctl_sched_rt_runtime = 950000;
288 /* cpus with isolated domains */
289 cpumask_var_t cpu_isolated_map;
292 lock_rq_of(struct task_struct *p, unsigned long *flags)
294 return task_rq_lock(p, flags);
298 unlock_rq_of(struct rq *rq, struct task_struct *p, unsigned long *flags)
300 task_rq_unlock(rq, p, flags);
304 * this_rq_lock - lock this runqueue and disable interrupts.
306 static struct rq *this_rq_lock(void)
313 raw_spin_lock(&rq->lock);
318 #ifdef CONFIG_SCHED_HRTICK
320 * Use HR-timers to deliver accurate preemption points.
323 static void hrtick_clear(struct rq *rq)
325 if (hrtimer_active(&rq->hrtick_timer))
326 hrtimer_cancel(&rq->hrtick_timer);
330 * High-resolution timer tick.
331 * Runs from hardirq context with interrupts disabled.
333 static enum hrtimer_restart hrtick(struct hrtimer *timer)
335 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
337 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
339 raw_spin_lock(&rq->lock);
341 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
342 raw_spin_unlock(&rq->lock);
344 return HRTIMER_NORESTART;
349 static void __hrtick_restart(struct rq *rq)
351 struct hrtimer *timer = &rq->hrtick_timer;
353 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
357 * called from hardirq (IPI) context
359 static void __hrtick_start(void *arg)
363 raw_spin_lock(&rq->lock);
364 __hrtick_restart(rq);
365 rq->hrtick_csd_pending = 0;
366 raw_spin_unlock(&rq->lock);
370 * Called to set the hrtick timer state.
372 * called with rq->lock held and irqs disabled
374 void hrtick_start(struct rq *rq, u64 delay)
376 struct hrtimer *timer = &rq->hrtick_timer;
381 * Don't schedule slices shorter than 10000ns, that just
382 * doesn't make sense and can cause timer DoS.
384 delta = max_t(s64, delay, 10000LL);
385 time = ktime_add_ns(timer->base->get_time(), delta);
387 hrtimer_set_expires(timer, time);
389 if (rq == this_rq()) {
390 __hrtick_restart(rq);
391 } else if (!rq->hrtick_csd_pending) {
392 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
393 rq->hrtick_csd_pending = 1;
398 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
400 int cpu = (int)(long)hcpu;
403 case CPU_UP_CANCELED:
404 case CPU_UP_CANCELED_FROZEN:
405 case CPU_DOWN_PREPARE:
406 case CPU_DOWN_PREPARE_FROZEN:
408 case CPU_DEAD_FROZEN:
409 hrtick_clear(cpu_rq(cpu));
416 static __init void init_hrtick(void)
418 hotcpu_notifier(hotplug_hrtick, 0);
422 * Called to set the hrtick timer state.
424 * called with rq->lock held and irqs disabled
426 void hrtick_start(struct rq *rq, u64 delay)
429 * Don't schedule slices shorter than 10000ns, that just
430 * doesn't make sense. Rely on vruntime for fairness.
432 delay = max_t(u64, delay, 10000LL);
433 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
434 HRTIMER_MODE_REL_PINNED);
437 static inline void init_hrtick(void)
440 #endif /* CONFIG_SMP */
442 static void init_rq_hrtick(struct rq *rq)
445 rq->hrtick_csd_pending = 0;
447 rq->hrtick_csd.flags = 0;
448 rq->hrtick_csd.func = __hrtick_start;
449 rq->hrtick_csd.info = rq;
452 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
453 rq->hrtick_timer.function = hrtick;
455 #else /* CONFIG_SCHED_HRTICK */
456 static inline void hrtick_clear(struct rq *rq)
460 static inline void init_rq_hrtick(struct rq *rq)
464 static inline void init_hrtick(void)
467 #endif /* CONFIG_SCHED_HRTICK */
470 * cmpxchg based fetch_or, macro so it works for different integer types
472 #define fetch_or(ptr, val) \
473 ({ typeof(*(ptr)) __old, __val = *(ptr); \
475 __old = cmpxchg((ptr), __val, __val | (val)); \
476 if (__old == __val) \
483 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
485 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
486 * this avoids any races wrt polling state changes and thereby avoids
489 static bool set_nr_and_not_polling(struct task_struct *p)
491 struct thread_info *ti = task_thread_info(p);
492 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
496 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
498 * If this returns true, then the idle task promises to call
499 * sched_ttwu_pending() and reschedule soon.
501 static bool set_nr_if_polling(struct task_struct *p)
503 struct thread_info *ti = task_thread_info(p);
504 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
507 if (!(val & _TIF_POLLING_NRFLAG))
509 if (val & _TIF_NEED_RESCHED)
511 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
520 static bool set_nr_and_not_polling(struct task_struct *p)
522 set_tsk_need_resched(p);
527 static bool set_nr_if_polling(struct task_struct *p)
534 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
536 struct wake_q_node *node = &task->wake_q;
539 * Atomically grab the task, if ->wake_q is !nil already it means
540 * its already queued (either by us or someone else) and will get the
541 * wakeup due to that.
543 * This cmpxchg() implies a full barrier, which pairs with the write
544 * barrier implied by the wakeup in wake_up_list().
546 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
549 get_task_struct(task);
552 * The head is context local, there can be no concurrency.
555 head->lastp = &node->next;
558 void wake_up_q(struct wake_q_head *head)
560 struct wake_q_node *node = head->first;
562 while (node != WAKE_Q_TAIL) {
563 struct task_struct *task;
565 task = container_of(node, struct task_struct, wake_q);
567 /* task can safely be re-inserted now */
569 task->wake_q.next = NULL;
572 * wake_up_process() implies a wmb() to pair with the queueing
573 * in wake_q_add() so as not to miss wakeups.
575 wake_up_process(task);
576 put_task_struct(task);
581 * resched_curr - mark rq's current task 'to be rescheduled now'.
583 * On UP this means the setting of the need_resched flag, on SMP it
584 * might also involve a cross-CPU call to trigger the scheduler on
587 void resched_curr(struct rq *rq)
589 struct task_struct *curr = rq->curr;
592 lockdep_assert_held(&rq->lock);
594 if (test_tsk_need_resched(curr))
599 if (cpu == smp_processor_id()) {
600 set_tsk_need_resched(curr);
601 set_preempt_need_resched();
605 if (set_nr_and_not_polling(curr))
606 smp_send_reschedule(cpu);
608 trace_sched_wake_idle_without_ipi(cpu);
611 void resched_cpu(int cpu)
613 struct rq *rq = cpu_rq(cpu);
616 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
619 raw_spin_unlock_irqrestore(&rq->lock, flags);
623 #ifdef CONFIG_NO_HZ_COMMON
625 * In the semi idle case, use the nearest busy cpu for migrating timers
626 * from an idle cpu. This is good for power-savings.
628 * We don't do similar optimization for completely idle system, as
629 * selecting an idle cpu will add more delays to the timers than intended
630 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
632 int get_nohz_timer_target(void)
634 int i, cpu = smp_processor_id();
635 struct sched_domain *sd;
637 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
641 for_each_domain(cpu, sd) {
642 for_each_cpu(i, sched_domain_span(sd)) {
643 if (!idle_cpu(i) && is_housekeeping_cpu(cpu)) {
650 if (!is_housekeeping_cpu(cpu))
651 cpu = housekeeping_any_cpu();
657 * When add_timer_on() enqueues a timer into the timer wheel of an
658 * idle CPU then this timer might expire before the next timer event
659 * which is scheduled to wake up that CPU. In case of a completely
660 * idle system the next event might even be infinite time into the
661 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
662 * leaves the inner idle loop so the newly added timer is taken into
663 * account when the CPU goes back to idle and evaluates the timer
664 * wheel for the next timer event.
666 static void wake_up_idle_cpu(int cpu)
668 struct rq *rq = cpu_rq(cpu);
670 if (cpu == smp_processor_id())
673 if (set_nr_and_not_polling(rq->idle))
674 smp_send_reschedule(cpu);
676 trace_sched_wake_idle_without_ipi(cpu);
679 static bool wake_up_full_nohz_cpu(int cpu)
682 * We just need the target to call irq_exit() and re-evaluate
683 * the next tick. The nohz full kick at least implies that.
684 * If needed we can still optimize that later with an
687 if (tick_nohz_full_cpu(cpu)) {
688 if (cpu != smp_processor_id() ||
689 tick_nohz_tick_stopped())
690 tick_nohz_full_kick_cpu(cpu);
697 void wake_up_nohz_cpu(int cpu)
699 if (!wake_up_full_nohz_cpu(cpu))
700 wake_up_idle_cpu(cpu);
703 static inline bool got_nohz_idle_kick(void)
705 int cpu = smp_processor_id();
707 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
710 if (idle_cpu(cpu) && !need_resched())
714 * We can't run Idle Load Balance on this CPU for this time so we
715 * cancel it and clear NOHZ_BALANCE_KICK
717 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
721 #else /* CONFIG_NO_HZ_COMMON */
723 static inline bool got_nohz_idle_kick(void)
728 #endif /* CONFIG_NO_HZ_COMMON */
730 #ifdef CONFIG_NO_HZ_FULL
731 bool sched_can_stop_tick(void)
734 * FIFO realtime policy runs the highest priority task. Other runnable
735 * tasks are of a lower priority. The scheduler tick does nothing.
737 if (current->policy == SCHED_FIFO)
741 * Round-robin realtime tasks time slice with other tasks at the same
742 * realtime priority. Is this task the only one at this priority?
744 if (current->policy == SCHED_RR) {
745 struct sched_rt_entity *rt_se = ¤t->rt;
747 return rt_se->run_list.prev == rt_se->run_list.next;
751 * More than one running task need preemption.
752 * nr_running update is assumed to be visible
753 * after IPI is sent from wakers.
755 if (this_rq()->nr_running > 1)
760 #endif /* CONFIG_NO_HZ_FULL */
762 void sched_avg_update(struct rq *rq)
764 s64 period = sched_avg_period();
766 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
768 * Inline assembly required to prevent the compiler
769 * optimising this loop into a divmod call.
770 * See __iter_div_u64_rem() for another example of this.
772 asm("" : "+rm" (rq->age_stamp));
773 rq->age_stamp += period;
778 #endif /* CONFIG_SMP */
780 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
781 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
783 * Iterate task_group tree rooted at *from, calling @down when first entering a
784 * node and @up when leaving it for the final time.
786 * Caller must hold rcu_lock or sufficient equivalent.
788 int walk_tg_tree_from(struct task_group *from,
789 tg_visitor down, tg_visitor up, void *data)
791 struct task_group *parent, *child;
797 ret = (*down)(parent, data);
800 list_for_each_entry_rcu(child, &parent->children, siblings) {
807 ret = (*up)(parent, data);
808 if (ret || parent == from)
812 parent = parent->parent;
819 int tg_nop(struct task_group *tg, void *data)
825 static void set_load_weight(struct task_struct *p)
827 int prio = p->static_prio - MAX_RT_PRIO;
828 struct load_weight *load = &p->se.load;
831 * SCHED_IDLE tasks get minimal weight:
833 if (idle_policy(p->policy)) {
834 load->weight = scale_load(WEIGHT_IDLEPRIO);
835 load->inv_weight = WMULT_IDLEPRIO;
839 load->weight = scale_load(prio_to_weight[prio]);
840 load->inv_weight = prio_to_wmult[prio];
843 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
846 if (!(flags & ENQUEUE_RESTORE))
847 sched_info_queued(rq, p);
848 p->sched_class->enqueue_task(rq, p, flags);
851 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
854 if (!(flags & DEQUEUE_SAVE))
855 sched_info_dequeued(rq, p);
856 p->sched_class->dequeue_task(rq, p, flags);
859 void activate_task(struct rq *rq, struct task_struct *p, int flags)
861 if (task_contributes_to_load(p))
862 rq->nr_uninterruptible--;
864 enqueue_task(rq, p, flags);
867 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
869 if (task_contributes_to_load(p))
870 rq->nr_uninterruptible++;
872 dequeue_task(rq, p, flags);
875 static void update_rq_clock_task(struct rq *rq, s64 delta)
878 * In theory, the compile should just see 0 here, and optimize out the call
879 * to sched_rt_avg_update. But I don't trust it...
881 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
882 s64 steal = 0, irq_delta = 0;
884 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
885 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
888 * Since irq_time is only updated on {soft,}irq_exit, we might run into
889 * this case when a previous update_rq_clock() happened inside a
892 * When this happens, we stop ->clock_task and only update the
893 * prev_irq_time stamp to account for the part that fit, so that a next
894 * update will consume the rest. This ensures ->clock_task is
897 * It does however cause some slight miss-attribution of {soft,}irq
898 * time, a more accurate solution would be to update the irq_time using
899 * the current rq->clock timestamp, except that would require using
902 if (irq_delta > delta)
905 rq->prev_irq_time += irq_delta;
908 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
909 if (static_key_false((¶virt_steal_rq_enabled))) {
910 steal = paravirt_steal_clock(cpu_of(rq));
911 steal -= rq->prev_steal_time_rq;
913 if (unlikely(steal > delta))
916 rq->prev_steal_time_rq += steal;
921 rq->clock_task += delta;
923 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
924 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
925 sched_rt_avg_update(rq, irq_delta + steal);
929 void sched_set_stop_task(int cpu, struct task_struct *stop)
931 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
932 struct task_struct *old_stop = cpu_rq(cpu)->stop;
936 * Make it appear like a SCHED_FIFO task, its something
937 * userspace knows about and won't get confused about.
939 * Also, it will make PI more or less work without too
940 * much confusion -- but then, stop work should not
941 * rely on PI working anyway.
943 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
945 stop->sched_class = &stop_sched_class;
948 cpu_rq(cpu)->stop = stop;
952 * Reset it back to a normal scheduling class so that
953 * it can die in pieces.
955 old_stop->sched_class = &rt_sched_class;
960 * __normal_prio - return the priority that is based on the static prio
962 static inline int __normal_prio(struct task_struct *p)
964 return p->static_prio;
968 * Calculate the expected normal priority: i.e. priority
969 * without taking RT-inheritance into account. Might be
970 * boosted by interactivity modifiers. Changes upon fork,
971 * setprio syscalls, and whenever the interactivity
972 * estimator recalculates.
974 static inline int normal_prio(struct task_struct *p)
978 if (task_has_dl_policy(p))
979 prio = MAX_DL_PRIO-1;
980 else if (task_has_rt_policy(p))
981 prio = MAX_RT_PRIO-1 - p->rt_priority;
983 prio = __normal_prio(p);
988 * Calculate the current priority, i.e. the priority
989 * taken into account by the scheduler. This value might
990 * be boosted by RT tasks, or might be boosted by
991 * interactivity modifiers. Will be RT if the task got
992 * RT-boosted. If not then it returns p->normal_prio.
994 static int effective_prio(struct task_struct *p)
996 p->normal_prio = normal_prio(p);
998 * If we are RT tasks or we were boosted to RT priority,
999 * keep the priority unchanged. Otherwise, update priority
1000 * to the normal priority:
1002 if (!rt_prio(p->prio))
1003 return p->normal_prio;
1008 * task_curr - is this task currently executing on a CPU?
1009 * @p: the task in question.
1011 * Return: 1 if the task is currently executing. 0 otherwise.
1013 inline int task_curr(const struct task_struct *p)
1015 return cpu_curr(task_cpu(p)) == p;
1019 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1020 * use the balance_callback list if you want balancing.
1022 * this means any call to check_class_changed() must be followed by a call to
1023 * balance_callback().
1025 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1026 const struct sched_class *prev_class,
1029 if (prev_class != p->sched_class) {
1030 if (prev_class->switched_from)
1031 prev_class->switched_from(rq, p);
1033 p->sched_class->switched_to(rq, p);
1034 } else if (oldprio != p->prio || dl_task(p))
1035 p->sched_class->prio_changed(rq, p, oldprio);
1038 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1040 const struct sched_class *class;
1042 if (p->sched_class == rq->curr->sched_class) {
1043 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1045 for_each_class(class) {
1046 if (class == rq->curr->sched_class)
1048 if (class == p->sched_class) {
1056 * A queue event has occurred, and we're going to schedule. In
1057 * this case, we can save a useless back to back clock update.
1059 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1060 rq_clock_skip_update(rq, true);
1065 * This is how migration works:
1067 * 1) we invoke migration_cpu_stop() on the target CPU using
1069 * 2) stopper starts to run (implicitly forcing the migrated thread
1071 * 3) it checks whether the migrated task is still in the wrong runqueue.
1072 * 4) if it's in the wrong runqueue then the migration thread removes
1073 * it and puts it into the right queue.
1074 * 5) stopper completes and stop_one_cpu() returns and the migration
1079 * move_queued_task - move a queued task to new rq.
1081 * Returns (locked) new rq. Old rq's lock is released.
1083 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1085 lockdep_assert_held(&rq->lock);
1087 dequeue_task(rq, p, 0);
1088 p->on_rq = TASK_ON_RQ_MIGRATING;
1089 double_lock_balance(rq, cpu_rq(new_cpu));
1090 set_task_cpu(p, new_cpu);
1091 double_unlock_balance(rq, cpu_rq(new_cpu));
1092 raw_spin_unlock(&rq->lock);
1094 rq = cpu_rq(new_cpu);
1096 raw_spin_lock(&rq->lock);
1097 BUG_ON(task_cpu(p) != new_cpu);
1098 p->on_rq = TASK_ON_RQ_QUEUED;
1099 enqueue_task(rq, p, 0);
1100 check_preempt_curr(rq, p, 0);
1105 struct migration_arg {
1106 struct task_struct *task;
1111 * Move (not current) task off this cpu, onto dest cpu. We're doing
1112 * this because either it can't run here any more (set_cpus_allowed()
1113 * away from this CPU, or CPU going down), or because we're
1114 * attempting to rebalance this task on exec (sched_exec).
1116 * So we race with normal scheduler movements, but that's OK, as long
1117 * as the task is no longer on this CPU.
1119 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1121 if (unlikely(!cpu_active(dest_cpu)))
1124 /* Affinity changed (again). */
1125 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1128 rq = move_queued_task(rq, p, dest_cpu);
1134 * migration_cpu_stop - this will be executed by a highprio stopper thread
1135 * and performs thread migration by bumping thread off CPU then
1136 * 'pushing' onto another runqueue.
1138 static int migration_cpu_stop(void *data)
1140 struct migration_arg *arg = data;
1141 struct task_struct *p = arg->task;
1142 struct rq *rq = this_rq();
1145 * The original target cpu might have gone down and we might
1146 * be on another cpu but it doesn't matter.
1148 local_irq_disable();
1150 * We need to explicitly wake pending tasks before running
1151 * __migrate_task() such that we will not miss enforcing cpus_allowed
1152 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1154 sched_ttwu_pending();
1156 raw_spin_lock(&p->pi_lock);
1157 raw_spin_lock(&rq->lock);
1159 * If task_rq(p) != rq, it cannot be migrated here, because we're
1160 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1161 * we're holding p->pi_lock.
1163 if (task_rq(p) == rq && task_on_rq_queued(p))
1164 rq = __migrate_task(rq, p, arg->dest_cpu);
1165 raw_spin_unlock(&rq->lock);
1166 raw_spin_unlock(&p->pi_lock);
1173 * sched_class::set_cpus_allowed must do the below, but is not required to
1174 * actually call this function.
1176 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1178 cpumask_copy(&p->cpus_allowed, new_mask);
1179 p->nr_cpus_allowed = cpumask_weight(new_mask);
1182 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1184 struct rq *rq = task_rq(p);
1185 bool queued, running;
1187 lockdep_assert_held(&p->pi_lock);
1189 queued = task_on_rq_queued(p);
1190 running = task_current(rq, p);
1194 * Because __kthread_bind() calls this on blocked tasks without
1197 lockdep_assert_held(&rq->lock);
1198 dequeue_task(rq, p, DEQUEUE_SAVE);
1201 put_prev_task(rq, p);
1203 p->sched_class->set_cpus_allowed(p, new_mask);
1206 p->sched_class->set_curr_task(rq);
1208 enqueue_task(rq, p, ENQUEUE_RESTORE);
1212 * Change a given task's CPU affinity. Migrate the thread to a
1213 * proper CPU and schedule it away if the CPU it's executing on
1214 * is removed from the allowed bitmask.
1216 * NOTE: the caller must have a valid reference to the task, the
1217 * task must not exit() & deallocate itself prematurely. The
1218 * call is not atomic; no spinlocks may be held.
1220 static int __set_cpus_allowed_ptr(struct task_struct *p,
1221 const struct cpumask *new_mask, bool check)
1223 unsigned long flags;
1225 unsigned int dest_cpu;
1228 rq = task_rq_lock(p, &flags);
1231 * Must re-check here, to close a race against __kthread_bind(),
1232 * sched_setaffinity() is not guaranteed to observe the flag.
1234 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1239 if (cpumask_equal(&p->cpus_allowed, new_mask))
1242 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1247 do_set_cpus_allowed(p, new_mask);
1249 /* Can the task run on the task's current CPU? If so, we're done */
1250 if (cpumask_test_cpu(task_cpu(p), new_mask))
1253 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1254 if (task_running(rq, p) || p->state == TASK_WAKING) {
1255 struct migration_arg arg = { p, dest_cpu };
1256 /* Need help from migration thread: drop lock and wait. */
1257 task_rq_unlock(rq, p, &flags);
1258 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1259 tlb_migrate_finish(p->mm);
1261 } else if (task_on_rq_queued(p)) {
1263 * OK, since we're going to drop the lock immediately
1264 * afterwards anyway.
1266 lockdep_unpin_lock(&rq->lock);
1267 rq = move_queued_task(rq, p, dest_cpu);
1268 lockdep_pin_lock(&rq->lock);
1271 task_rq_unlock(rq, p, &flags);
1276 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1278 return __set_cpus_allowed_ptr(p, new_mask, false);
1280 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1282 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1284 #ifdef CONFIG_SCHED_DEBUG
1286 * We should never call set_task_cpu() on a blocked task,
1287 * ttwu() will sort out the placement.
1289 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1292 #ifdef CONFIG_LOCKDEP
1294 * The caller should hold either p->pi_lock or rq->lock, when changing
1295 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1297 * sched_move_task() holds both and thus holding either pins the cgroup,
1300 * Furthermore, all task_rq users should acquire both locks, see
1303 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1304 lockdep_is_held(&task_rq(p)->lock)));
1308 trace_sched_migrate_task(p, new_cpu);
1310 if (task_cpu(p) != new_cpu) {
1311 if (p->sched_class->migrate_task_rq)
1312 p->sched_class->migrate_task_rq(p);
1313 p->se.nr_migrations++;
1314 perf_event_task_migrate(p);
1316 walt_fixup_busy_time(p, new_cpu);
1319 __set_task_cpu(p, new_cpu);
1322 static void __migrate_swap_task(struct task_struct *p, int cpu)
1324 if (task_on_rq_queued(p)) {
1325 struct rq *src_rq, *dst_rq;
1327 src_rq = task_rq(p);
1328 dst_rq = cpu_rq(cpu);
1330 deactivate_task(src_rq, p, 0);
1331 set_task_cpu(p, cpu);
1332 activate_task(dst_rq, p, 0);
1333 check_preempt_curr(dst_rq, p, 0);
1336 * Task isn't running anymore; make it appear like we migrated
1337 * it before it went to sleep. This means on wakeup we make the
1338 * previous cpu our targer instead of where it really is.
1344 struct migration_swap_arg {
1345 struct task_struct *src_task, *dst_task;
1346 int src_cpu, dst_cpu;
1349 static int migrate_swap_stop(void *data)
1351 struct migration_swap_arg *arg = data;
1352 struct rq *src_rq, *dst_rq;
1355 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1358 src_rq = cpu_rq(arg->src_cpu);
1359 dst_rq = cpu_rq(arg->dst_cpu);
1361 double_raw_lock(&arg->src_task->pi_lock,
1362 &arg->dst_task->pi_lock);
1363 double_rq_lock(src_rq, dst_rq);
1365 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1368 if (task_cpu(arg->src_task) != arg->src_cpu)
1371 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1374 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1377 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1378 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1383 double_rq_unlock(src_rq, dst_rq);
1384 raw_spin_unlock(&arg->dst_task->pi_lock);
1385 raw_spin_unlock(&arg->src_task->pi_lock);
1391 * Cross migrate two tasks
1393 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1395 struct migration_swap_arg arg;
1398 arg = (struct migration_swap_arg){
1400 .src_cpu = task_cpu(cur),
1402 .dst_cpu = task_cpu(p),
1405 if (arg.src_cpu == arg.dst_cpu)
1409 * These three tests are all lockless; this is OK since all of them
1410 * will be re-checked with proper locks held further down the line.
1412 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1415 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1418 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1421 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1422 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1429 * wait_task_inactive - wait for a thread to unschedule.
1431 * If @match_state is nonzero, it's the @p->state value just checked and
1432 * not expected to change. If it changes, i.e. @p might have woken up,
1433 * then return zero. When we succeed in waiting for @p to be off its CPU,
1434 * we return a positive number (its total switch count). If a second call
1435 * a short while later returns the same number, the caller can be sure that
1436 * @p has remained unscheduled the whole time.
1438 * The caller must ensure that the task *will* unschedule sometime soon,
1439 * else this function might spin for a *long* time. This function can't
1440 * be called with interrupts off, or it may introduce deadlock with
1441 * smp_call_function() if an IPI is sent by the same process we are
1442 * waiting to become inactive.
1444 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1446 unsigned long flags;
1447 int running, queued;
1453 * We do the initial early heuristics without holding
1454 * any task-queue locks at all. We'll only try to get
1455 * the runqueue lock when things look like they will
1461 * If the task is actively running on another CPU
1462 * still, just relax and busy-wait without holding
1465 * NOTE! Since we don't hold any locks, it's not
1466 * even sure that "rq" stays as the right runqueue!
1467 * But we don't care, since "task_running()" will
1468 * return false if the runqueue has changed and p
1469 * is actually now running somewhere else!
1471 while (task_running(rq, p)) {
1472 if (match_state && unlikely(p->state != match_state))
1478 * Ok, time to look more closely! We need the rq
1479 * lock now, to be *sure*. If we're wrong, we'll
1480 * just go back and repeat.
1482 rq = task_rq_lock(p, &flags);
1483 trace_sched_wait_task(p);
1484 running = task_running(rq, p);
1485 queued = task_on_rq_queued(p);
1487 if (!match_state || p->state == match_state)
1488 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1489 task_rq_unlock(rq, p, &flags);
1492 * If it changed from the expected state, bail out now.
1494 if (unlikely(!ncsw))
1498 * Was it really running after all now that we
1499 * checked with the proper locks actually held?
1501 * Oops. Go back and try again..
1503 if (unlikely(running)) {
1509 * It's not enough that it's not actively running,
1510 * it must be off the runqueue _entirely_, and not
1513 * So if it was still runnable (but just not actively
1514 * running right now), it's preempted, and we should
1515 * yield - it could be a while.
1517 if (unlikely(queued)) {
1518 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1520 set_current_state(TASK_UNINTERRUPTIBLE);
1521 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1526 * Ahh, all good. It wasn't running, and it wasn't
1527 * runnable, which means that it will never become
1528 * running in the future either. We're all done!
1537 * kick_process - kick a running thread to enter/exit the kernel
1538 * @p: the to-be-kicked thread
1540 * Cause a process which is running on another CPU to enter
1541 * kernel-mode, without any delay. (to get signals handled.)
1543 * NOTE: this function doesn't have to take the runqueue lock,
1544 * because all it wants to ensure is that the remote task enters
1545 * the kernel. If the IPI races and the task has been migrated
1546 * to another CPU then no harm is done and the purpose has been
1549 void kick_process(struct task_struct *p)
1555 if ((cpu != smp_processor_id()) && task_curr(p))
1556 smp_send_reschedule(cpu);
1559 EXPORT_SYMBOL_GPL(kick_process);
1562 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1564 static int select_fallback_rq(int cpu, struct task_struct *p)
1566 int nid = cpu_to_node(cpu);
1567 const struct cpumask *nodemask = NULL;
1568 enum { cpuset, possible, fail } state = cpuset;
1572 * If the node that the cpu is on has been offlined, cpu_to_node()
1573 * will return -1. There is no cpu on the node, and we should
1574 * select the cpu on the other node.
1577 nodemask = cpumask_of_node(nid);
1579 /* Look for allowed, online CPU in same node. */
1580 for_each_cpu(dest_cpu, nodemask) {
1581 if (!cpu_online(dest_cpu))
1583 if (!cpu_active(dest_cpu))
1585 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1591 /* Any allowed, online CPU? */
1592 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1593 if (!cpu_online(dest_cpu))
1595 if (!cpu_active(dest_cpu))
1600 /* No more Mr. Nice Guy. */
1603 if (IS_ENABLED(CONFIG_CPUSETS)) {
1604 cpuset_cpus_allowed_fallback(p);
1610 do_set_cpus_allowed(p, cpu_possible_mask);
1621 if (state != cpuset) {
1623 * Don't tell them about moving exiting tasks or
1624 * kernel threads (both mm NULL), since they never
1627 if (p->mm && printk_ratelimit()) {
1628 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1629 task_pid_nr(p), p->comm, cpu);
1637 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1640 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1642 lockdep_assert_held(&p->pi_lock);
1644 if (p->nr_cpus_allowed > 1)
1645 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1648 * In order not to call set_task_cpu() on a blocking task we need
1649 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1652 * Since this is common to all placement strategies, this lives here.
1654 * [ this allows ->select_task() to simply return task_cpu(p) and
1655 * not worry about this generic constraint ]
1657 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1659 cpu = select_fallback_rq(task_cpu(p), p);
1664 static void update_avg(u64 *avg, u64 sample)
1666 s64 diff = sample - *avg;
1672 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1673 const struct cpumask *new_mask, bool check)
1675 return set_cpus_allowed_ptr(p, new_mask);
1678 #endif /* CONFIG_SMP */
1681 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1683 #ifdef CONFIG_SCHEDSTATS
1684 struct rq *rq = this_rq();
1687 int this_cpu = smp_processor_id();
1689 if (cpu == this_cpu) {
1690 schedstat_inc(rq, ttwu_local);
1691 schedstat_inc(p, se.statistics.nr_wakeups_local);
1693 struct sched_domain *sd;
1695 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1697 for_each_domain(this_cpu, sd) {
1698 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1699 schedstat_inc(sd, ttwu_wake_remote);
1706 if (wake_flags & WF_MIGRATED)
1707 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1709 #endif /* CONFIG_SMP */
1711 schedstat_inc(rq, ttwu_count);
1712 schedstat_inc(p, se.statistics.nr_wakeups);
1714 if (wake_flags & WF_SYNC)
1715 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1717 #endif /* CONFIG_SCHEDSTATS */
1720 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1722 activate_task(rq, p, en_flags);
1723 p->on_rq = TASK_ON_RQ_QUEUED;
1725 /* if a worker is waking up, notify workqueue */
1726 if (p->flags & PF_WQ_WORKER)
1727 wq_worker_waking_up(p, cpu_of(rq));
1731 * Mark the task runnable and perform wakeup-preemption.
1734 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1736 check_preempt_curr(rq, p, wake_flags);
1737 p->state = TASK_RUNNING;
1738 trace_sched_wakeup(p);
1741 if (p->sched_class->task_woken) {
1743 * Our task @p is fully woken up and running; so its safe to
1744 * drop the rq->lock, hereafter rq is only used for statistics.
1746 lockdep_unpin_lock(&rq->lock);
1747 p->sched_class->task_woken(rq, p);
1748 lockdep_pin_lock(&rq->lock);
1751 if (rq->idle_stamp) {
1752 u64 delta = rq_clock(rq) - rq->idle_stamp;
1753 u64 max = 2*rq->max_idle_balance_cost;
1755 update_avg(&rq->avg_idle, delta);
1757 if (rq->avg_idle > max)
1766 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1768 lockdep_assert_held(&rq->lock);
1771 if (p->sched_contributes_to_load)
1772 rq->nr_uninterruptible--;
1775 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1776 ttwu_do_wakeup(rq, p, wake_flags);
1780 * Called in case the task @p isn't fully descheduled from its runqueue,
1781 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1782 * since all we need to do is flip p->state to TASK_RUNNING, since
1783 * the task is still ->on_rq.
1785 static int ttwu_remote(struct task_struct *p, int wake_flags)
1790 rq = __task_rq_lock(p);
1791 if (task_on_rq_queued(p)) {
1792 /* check_preempt_curr() may use rq clock */
1793 update_rq_clock(rq);
1794 ttwu_do_wakeup(rq, p, wake_flags);
1797 __task_rq_unlock(rq);
1803 void sched_ttwu_pending(void)
1805 struct rq *rq = this_rq();
1806 struct llist_node *llist = llist_del_all(&rq->wake_list);
1807 struct task_struct *p;
1808 unsigned long flags;
1813 raw_spin_lock_irqsave(&rq->lock, flags);
1814 lockdep_pin_lock(&rq->lock);
1817 p = llist_entry(llist, struct task_struct, wake_entry);
1818 llist = llist_next(llist);
1819 ttwu_do_activate(rq, p, 0);
1822 lockdep_unpin_lock(&rq->lock);
1823 raw_spin_unlock_irqrestore(&rq->lock, flags);
1826 void scheduler_ipi(void)
1829 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1830 * TIF_NEED_RESCHED remotely (for the first time) will also send
1833 preempt_fold_need_resched();
1835 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1839 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1840 * traditionally all their work was done from the interrupt return
1841 * path. Now that we actually do some work, we need to make sure
1844 * Some archs already do call them, luckily irq_enter/exit nest
1847 * Arguably we should visit all archs and update all handlers,
1848 * however a fair share of IPIs are still resched only so this would
1849 * somewhat pessimize the simple resched case.
1852 sched_ttwu_pending();
1855 * Check if someone kicked us for doing the nohz idle load balance.
1857 if (unlikely(got_nohz_idle_kick())) {
1858 this_rq()->idle_balance = 1;
1859 raise_softirq_irqoff(SCHED_SOFTIRQ);
1864 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1866 struct rq *rq = cpu_rq(cpu);
1868 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1869 if (!set_nr_if_polling(rq->idle))
1870 smp_send_reschedule(cpu);
1872 trace_sched_wake_idle_without_ipi(cpu);
1876 void wake_up_if_idle(int cpu)
1878 struct rq *rq = cpu_rq(cpu);
1879 unsigned long flags;
1883 if (!is_idle_task(rcu_dereference(rq->curr)))
1886 if (set_nr_if_polling(rq->idle)) {
1887 trace_sched_wake_idle_without_ipi(cpu);
1889 raw_spin_lock_irqsave(&rq->lock, flags);
1890 if (is_idle_task(rq->curr))
1891 smp_send_reschedule(cpu);
1892 /* Else cpu is not in idle, do nothing here */
1893 raw_spin_unlock_irqrestore(&rq->lock, flags);
1900 bool cpus_share_cache(int this_cpu, int that_cpu)
1902 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1904 #endif /* CONFIG_SMP */
1906 static void ttwu_queue(struct task_struct *p, int cpu)
1908 struct rq *rq = cpu_rq(cpu);
1910 #if defined(CONFIG_SMP)
1911 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1912 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1913 ttwu_queue_remote(p, cpu);
1918 raw_spin_lock(&rq->lock);
1919 lockdep_pin_lock(&rq->lock);
1920 ttwu_do_activate(rq, p, 0);
1921 lockdep_unpin_lock(&rq->lock);
1922 raw_spin_unlock(&rq->lock);
1926 * try_to_wake_up - wake up a thread
1927 * @p: the thread to be awakened
1928 * @state: the mask of task states that can be woken
1929 * @wake_flags: wake modifier flags (WF_*)
1931 * Put it on the run-queue if it's not already there. The "current"
1932 * thread is always on the run-queue (except when the actual
1933 * re-schedule is in progress), and as such you're allowed to do
1934 * the simpler "current->state = TASK_RUNNING" to mark yourself
1935 * runnable without the overhead of this.
1937 * Return: %true if @p was woken up, %false if it was already running.
1938 * or @state didn't match @p's state.
1941 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1943 unsigned long flags;
1944 int cpu, success = 0;
1951 * If we are going to wake up a thread waiting for CONDITION we
1952 * need to ensure that CONDITION=1 done by the caller can not be
1953 * reordered with p->state check below. This pairs with mb() in
1954 * set_current_state() the waiting thread does.
1956 smp_mb__before_spinlock();
1957 raw_spin_lock_irqsave(&p->pi_lock, flags);
1958 if (!(p->state & state))
1961 trace_sched_waking(p);
1963 success = 1; /* we're going to change ->state */
1966 if (p->on_rq && ttwu_remote(p, wake_flags))
1971 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1972 * possible to, falsely, observe p->on_cpu == 0.
1974 * One must be running (->on_cpu == 1) in order to remove oneself
1975 * from the runqueue.
1977 * [S] ->on_cpu = 1; [L] ->on_rq
1981 * [S] ->on_rq = 0; [L] ->on_cpu
1983 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1984 * from the consecutive calls to schedule(); the first switching to our
1985 * task, the second putting it to sleep.
1990 * If the owning (remote) cpu is still in the middle of schedule() with
1991 * this task as prev, wait until its done referencing the task.
1996 * Combined with the control dependency above, we have an effective
1997 * smp_load_acquire() without the need for full barriers.
1999 * Pairs with the smp_store_release() in finish_lock_switch().
2001 * This ensures that tasks getting woken will be fully ordered against
2002 * their previous state and preserve Program Order.
2006 rq = cpu_rq(task_cpu(p));
2008 raw_spin_lock(&rq->lock);
2009 wallclock = walt_ktime_clock();
2010 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2011 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2012 raw_spin_unlock(&rq->lock);
2014 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2015 p->state = TASK_WAKING;
2017 if (p->sched_class->task_waking)
2018 p->sched_class->task_waking(p);
2020 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2022 if (task_cpu(p) != cpu) {
2023 wake_flags |= WF_MIGRATED;
2024 set_task_cpu(p, cpu);
2027 #endif /* CONFIG_SMP */
2031 ttwu_stat(p, cpu, wake_flags);
2033 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2039 * try_to_wake_up_local - try to wake up a local task with rq lock held
2040 * @p: the thread to be awakened
2042 * Put @p on the run-queue if it's not already there. The caller must
2043 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2046 static void try_to_wake_up_local(struct task_struct *p)
2048 struct rq *rq = task_rq(p);
2050 if (WARN_ON_ONCE(rq != this_rq()) ||
2051 WARN_ON_ONCE(p == current))
2054 lockdep_assert_held(&rq->lock);
2056 if (!raw_spin_trylock(&p->pi_lock)) {
2058 * This is OK, because current is on_cpu, which avoids it being
2059 * picked for load-balance and preemption/IRQs are still
2060 * disabled avoiding further scheduler activity on it and we've
2061 * not yet picked a replacement task.
2063 lockdep_unpin_lock(&rq->lock);
2064 raw_spin_unlock(&rq->lock);
2065 raw_spin_lock(&p->pi_lock);
2066 raw_spin_lock(&rq->lock);
2067 lockdep_pin_lock(&rq->lock);
2070 if (!(p->state & TASK_NORMAL))
2073 trace_sched_waking(p);
2075 if (!task_on_rq_queued(p)) {
2076 u64 wallclock = walt_ktime_clock();
2078 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2079 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2080 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2083 ttwu_do_wakeup(rq, p, 0);
2084 ttwu_stat(p, smp_processor_id(), 0);
2086 raw_spin_unlock(&p->pi_lock);
2090 * wake_up_process - Wake up a specific process
2091 * @p: The process to be woken up.
2093 * Attempt to wake up the nominated process and move it to the set of runnable
2096 * Return: 1 if the process was woken up, 0 if it was already running.
2098 * It may be assumed that this function implies a write memory barrier before
2099 * changing the task state if and only if any tasks are woken up.
2101 int wake_up_process(struct task_struct *p)
2103 return try_to_wake_up(p, TASK_NORMAL, 0);
2105 EXPORT_SYMBOL(wake_up_process);
2107 int wake_up_state(struct task_struct *p, unsigned int state)
2109 return try_to_wake_up(p, state, 0);
2113 * This function clears the sched_dl_entity static params.
2115 void __dl_clear_params(struct task_struct *p)
2117 struct sched_dl_entity *dl_se = &p->dl;
2119 dl_se->dl_runtime = 0;
2120 dl_se->dl_deadline = 0;
2121 dl_se->dl_period = 0;
2125 dl_se->dl_throttled = 0;
2127 dl_se->dl_yielded = 0;
2131 * Perform scheduler related setup for a newly forked process p.
2132 * p is forked by current.
2134 * __sched_fork() is basic setup used by init_idle() too:
2136 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2141 p->se.exec_start = 0;
2142 p->se.sum_exec_runtime = 0;
2143 p->se.prev_sum_exec_runtime = 0;
2144 p->se.nr_migrations = 0;
2146 INIT_LIST_HEAD(&p->se.group_node);
2147 walt_init_new_task_load(p);
2149 #ifdef CONFIG_SCHEDSTATS
2150 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2153 RB_CLEAR_NODE(&p->dl.rb_node);
2154 init_dl_task_timer(&p->dl);
2155 __dl_clear_params(p);
2157 INIT_LIST_HEAD(&p->rt.run_list);
2159 #ifdef CONFIG_PREEMPT_NOTIFIERS
2160 INIT_HLIST_HEAD(&p->preempt_notifiers);
2163 #ifdef CONFIG_NUMA_BALANCING
2164 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2165 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2166 p->mm->numa_scan_seq = 0;
2169 if (clone_flags & CLONE_VM)
2170 p->numa_preferred_nid = current->numa_preferred_nid;
2172 p->numa_preferred_nid = -1;
2174 p->node_stamp = 0ULL;
2175 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2176 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2177 p->numa_work.next = &p->numa_work;
2178 p->numa_faults = NULL;
2179 p->last_task_numa_placement = 0;
2180 p->last_sum_exec_runtime = 0;
2182 p->numa_group = NULL;
2183 #endif /* CONFIG_NUMA_BALANCING */
2186 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2188 #ifdef CONFIG_NUMA_BALANCING
2190 void set_numabalancing_state(bool enabled)
2193 static_branch_enable(&sched_numa_balancing);
2195 static_branch_disable(&sched_numa_balancing);
2198 #ifdef CONFIG_PROC_SYSCTL
2199 int sysctl_numa_balancing(struct ctl_table *table, int write,
2200 void __user *buffer, size_t *lenp, loff_t *ppos)
2204 int state = static_branch_likely(&sched_numa_balancing);
2206 if (write && !capable(CAP_SYS_ADMIN))
2211 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2215 set_numabalancing_state(state);
2222 * fork()/clone()-time setup:
2224 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2226 unsigned long flags;
2227 int cpu = get_cpu();
2229 __sched_fork(clone_flags, p);
2231 * We mark the process as running here. This guarantees that
2232 * nobody will actually run it, and a signal or other external
2233 * event cannot wake it up and insert it on the runqueue either.
2235 p->state = TASK_RUNNING;
2238 * Make sure we do not leak PI boosting priority to the child.
2240 p->prio = current->normal_prio;
2243 * Revert to default priority/policy on fork if requested.
2245 if (unlikely(p->sched_reset_on_fork)) {
2246 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2247 p->policy = SCHED_NORMAL;
2248 p->static_prio = NICE_TO_PRIO(0);
2250 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2251 p->static_prio = NICE_TO_PRIO(0);
2253 p->prio = p->normal_prio = __normal_prio(p);
2257 * We don't need the reset flag anymore after the fork. It has
2258 * fulfilled its duty:
2260 p->sched_reset_on_fork = 0;
2263 if (dl_prio(p->prio)) {
2266 } else if (rt_prio(p->prio)) {
2267 p->sched_class = &rt_sched_class;
2269 p->sched_class = &fair_sched_class;
2272 if (p->sched_class->task_fork)
2273 p->sched_class->task_fork(p);
2276 * The child is not yet in the pid-hash so no cgroup attach races,
2277 * and the cgroup is pinned to this child due to cgroup_fork()
2278 * is ran before sched_fork().
2280 * Silence PROVE_RCU.
2282 raw_spin_lock_irqsave(&p->pi_lock, flags);
2283 set_task_cpu(p, cpu);
2284 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2286 #ifdef CONFIG_SCHED_INFO
2287 if (likely(sched_info_on()))
2288 memset(&p->sched_info, 0, sizeof(p->sched_info));
2290 #if defined(CONFIG_SMP)
2293 init_task_preempt_count(p);
2295 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2296 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2303 unsigned long to_ratio(u64 period, u64 runtime)
2305 if (runtime == RUNTIME_INF)
2309 * Doing this here saves a lot of checks in all
2310 * the calling paths, and returning zero seems
2311 * safe for them anyway.
2316 return div64_u64(runtime << 20, period);
2320 inline struct dl_bw *dl_bw_of(int i)
2322 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2323 "sched RCU must be held");
2324 return &cpu_rq(i)->rd->dl_bw;
2327 static inline int dl_bw_cpus(int i)
2329 struct root_domain *rd = cpu_rq(i)->rd;
2332 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2333 "sched RCU must be held");
2334 for_each_cpu_and(i, rd->span, cpu_active_mask)
2340 inline struct dl_bw *dl_bw_of(int i)
2342 return &cpu_rq(i)->dl.dl_bw;
2345 static inline int dl_bw_cpus(int i)
2352 * We must be sure that accepting a new task (or allowing changing the
2353 * parameters of an existing one) is consistent with the bandwidth
2354 * constraints. If yes, this function also accordingly updates the currently
2355 * allocated bandwidth to reflect the new situation.
2357 * This function is called while holding p's rq->lock.
2359 * XXX we should delay bw change until the task's 0-lag point, see
2362 static int dl_overflow(struct task_struct *p, int policy,
2363 const struct sched_attr *attr)
2366 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2367 u64 period = attr->sched_period ?: attr->sched_deadline;
2368 u64 runtime = attr->sched_runtime;
2369 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2372 if (new_bw == p->dl.dl_bw)
2376 * Either if a task, enters, leave, or stays -deadline but changes
2377 * its parameters, we may need to update accordingly the total
2378 * allocated bandwidth of the container.
2380 raw_spin_lock(&dl_b->lock);
2381 cpus = dl_bw_cpus(task_cpu(p));
2382 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2383 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2384 __dl_add(dl_b, new_bw);
2386 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2387 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2388 __dl_clear(dl_b, p->dl.dl_bw);
2389 __dl_add(dl_b, new_bw);
2391 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2392 __dl_clear(dl_b, p->dl.dl_bw);
2395 raw_spin_unlock(&dl_b->lock);
2400 extern void init_dl_bw(struct dl_bw *dl_b);
2403 * wake_up_new_task - wake up a newly created task for the first time.
2405 * This function will do some initial scheduler statistics housekeeping
2406 * that must be done for every newly created context, then puts the task
2407 * on the runqueue and wakes it.
2409 void wake_up_new_task(struct task_struct *p)
2411 unsigned long flags;
2414 raw_spin_lock_irqsave(&p->pi_lock, flags);
2416 walt_init_new_task_load(p);
2418 /* Initialize new task's runnable average */
2419 init_entity_runnable_average(&p->se);
2422 * Fork balancing, do it here and not earlier because:
2423 * - cpus_allowed can change in the fork path
2424 * - any previously selected cpu might disappear through hotplug
2426 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2429 rq = __task_rq_lock(p);
2430 walt_mark_task_starting(p);
2431 activate_task(rq, p, ENQUEUE_WAKEUP_NEW);
2432 p->on_rq = TASK_ON_RQ_QUEUED;
2433 trace_sched_wakeup_new(p);
2434 check_preempt_curr(rq, p, WF_FORK);
2436 if (p->sched_class->task_woken) {
2438 * Nothing relies on rq->lock after this, so its fine to
2441 lockdep_unpin_lock(&rq->lock);
2442 p->sched_class->task_woken(rq, p);
2443 lockdep_pin_lock(&rq->lock);
2446 task_rq_unlock(rq, p, &flags);
2449 #ifdef CONFIG_PREEMPT_NOTIFIERS
2451 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2453 void preempt_notifier_inc(void)
2455 static_key_slow_inc(&preempt_notifier_key);
2457 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2459 void preempt_notifier_dec(void)
2461 static_key_slow_dec(&preempt_notifier_key);
2463 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2466 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2467 * @notifier: notifier struct to register
2469 void preempt_notifier_register(struct preempt_notifier *notifier)
2471 if (!static_key_false(&preempt_notifier_key))
2472 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2474 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2476 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2479 * preempt_notifier_unregister - no longer interested in preemption notifications
2480 * @notifier: notifier struct to unregister
2482 * This is *not* safe to call from within a preemption notifier.
2484 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2486 hlist_del(¬ifier->link);
2488 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2490 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2492 struct preempt_notifier *notifier;
2494 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2495 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2498 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2500 if (static_key_false(&preempt_notifier_key))
2501 __fire_sched_in_preempt_notifiers(curr);
2505 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2506 struct task_struct *next)
2508 struct preempt_notifier *notifier;
2510 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2511 notifier->ops->sched_out(notifier, next);
2514 static __always_inline void
2515 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2516 struct task_struct *next)
2518 if (static_key_false(&preempt_notifier_key))
2519 __fire_sched_out_preempt_notifiers(curr, next);
2522 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2524 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2529 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2530 struct task_struct *next)
2534 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2537 * prepare_task_switch - prepare to switch tasks
2538 * @rq: the runqueue preparing to switch
2539 * @prev: the current task that is being switched out
2540 * @next: the task we are going to switch to.
2542 * This is called with the rq lock held and interrupts off. It must
2543 * be paired with a subsequent finish_task_switch after the context
2546 * prepare_task_switch sets up locking and calls architecture specific
2550 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2551 struct task_struct *next)
2553 sched_info_switch(rq, prev, next);
2554 perf_event_task_sched_out(prev, next);
2555 fire_sched_out_preempt_notifiers(prev, next);
2556 prepare_lock_switch(rq, next);
2557 prepare_arch_switch(next);
2561 * finish_task_switch - clean up after a task-switch
2562 * @prev: the thread we just switched away from.
2564 * finish_task_switch must be called after the context switch, paired
2565 * with a prepare_task_switch call before the context switch.
2566 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2567 * and do any other architecture-specific cleanup actions.
2569 * Note that we may have delayed dropping an mm in context_switch(). If
2570 * so, we finish that here outside of the runqueue lock. (Doing it
2571 * with the lock held can cause deadlocks; see schedule() for
2574 * The context switch have flipped the stack from under us and restored the
2575 * local variables which were saved when this task called schedule() in the
2576 * past. prev == current is still correct but we need to recalculate this_rq
2577 * because prev may have moved to another CPU.
2579 static struct rq *finish_task_switch(struct task_struct *prev)
2580 __releases(rq->lock)
2582 struct rq *rq = this_rq();
2583 struct mm_struct *mm = rq->prev_mm;
2587 * The previous task will have left us with a preempt_count of 2
2588 * because it left us after:
2591 * preempt_disable(); // 1
2593 * raw_spin_lock_irq(&rq->lock) // 2
2595 * Also, see FORK_PREEMPT_COUNT.
2597 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2598 "corrupted preempt_count: %s/%d/0x%x\n",
2599 current->comm, current->pid, preempt_count()))
2600 preempt_count_set(FORK_PREEMPT_COUNT);
2605 * A task struct has one reference for the use as "current".
2606 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2607 * schedule one last time. The schedule call will never return, and
2608 * the scheduled task must drop that reference.
2610 * We must observe prev->state before clearing prev->on_cpu (in
2611 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2612 * running on another CPU and we could rave with its RUNNING -> DEAD
2613 * transition, resulting in a double drop.
2615 prev_state = prev->state;
2616 vtime_task_switch(prev);
2617 perf_event_task_sched_in(prev, current);
2618 finish_lock_switch(rq, prev);
2619 finish_arch_post_lock_switch();
2621 fire_sched_in_preempt_notifiers(current);
2624 if (unlikely(prev_state == TASK_DEAD)) {
2625 if (prev->sched_class->task_dead)
2626 prev->sched_class->task_dead(prev);
2629 * Remove function-return probe instances associated with this
2630 * task and put them back on the free list.
2632 kprobe_flush_task(prev);
2633 put_task_struct(prev);
2636 tick_nohz_task_switch();
2642 /* rq->lock is NOT held, but preemption is disabled */
2643 static void __balance_callback(struct rq *rq)
2645 struct callback_head *head, *next;
2646 void (*func)(struct rq *rq);
2647 unsigned long flags;
2649 raw_spin_lock_irqsave(&rq->lock, flags);
2650 head = rq->balance_callback;
2651 rq->balance_callback = NULL;
2653 func = (void (*)(struct rq *))head->func;
2660 raw_spin_unlock_irqrestore(&rq->lock, flags);
2663 static inline void balance_callback(struct rq *rq)
2665 if (unlikely(rq->balance_callback))
2666 __balance_callback(rq);
2671 static inline void balance_callback(struct rq *rq)
2678 * schedule_tail - first thing a freshly forked thread must call.
2679 * @prev: the thread we just switched away from.
2681 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2682 __releases(rq->lock)
2687 * New tasks start with FORK_PREEMPT_COUNT, see there and
2688 * finish_task_switch() for details.
2690 * finish_task_switch() will drop rq->lock() and lower preempt_count
2691 * and the preempt_enable() will end up enabling preemption (on
2692 * PREEMPT_COUNT kernels).
2695 rq = finish_task_switch(prev);
2696 balance_callback(rq);
2699 if (current->set_child_tid)
2700 put_user(task_pid_vnr(current), current->set_child_tid);
2704 * context_switch - switch to the new MM and the new thread's register state.
2706 static inline struct rq *
2707 context_switch(struct rq *rq, struct task_struct *prev,
2708 struct task_struct *next)
2710 struct mm_struct *mm, *oldmm;
2712 prepare_task_switch(rq, prev, next);
2715 oldmm = prev->active_mm;
2717 * For paravirt, this is coupled with an exit in switch_to to
2718 * combine the page table reload and the switch backend into
2721 arch_start_context_switch(prev);
2724 next->active_mm = oldmm;
2725 atomic_inc(&oldmm->mm_count);
2726 enter_lazy_tlb(oldmm, next);
2728 switch_mm(oldmm, mm, next);
2731 prev->active_mm = NULL;
2732 rq->prev_mm = oldmm;
2735 * Since the runqueue lock will be released by the next
2736 * task (which is an invalid locking op but in the case
2737 * of the scheduler it's an obvious special-case), so we
2738 * do an early lockdep release here:
2740 lockdep_unpin_lock(&rq->lock);
2741 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2743 /* Here we just switch the register state and the stack. */
2744 switch_to(prev, next, prev);
2747 return finish_task_switch(prev);
2751 * nr_running and nr_context_switches:
2753 * externally visible scheduler statistics: current number of runnable
2754 * threads, total number of context switches performed since bootup.
2756 unsigned long nr_running(void)
2758 unsigned long i, sum = 0;
2760 for_each_online_cpu(i)
2761 sum += cpu_rq(i)->nr_running;
2767 * Check if only the current task is running on the cpu.
2769 * Caution: this function does not check that the caller has disabled
2770 * preemption, thus the result might have a time-of-check-to-time-of-use
2771 * race. The caller is responsible to use it correctly, for example:
2773 * - from a non-preemptable section (of course)
2775 * - from a thread that is bound to a single CPU
2777 * - in a loop with very short iterations (e.g. a polling loop)
2779 bool single_task_running(void)
2781 return raw_rq()->nr_running == 1;
2783 EXPORT_SYMBOL(single_task_running);
2785 unsigned long long nr_context_switches(void)
2788 unsigned long long sum = 0;
2790 for_each_possible_cpu(i)
2791 sum += cpu_rq(i)->nr_switches;
2796 unsigned long nr_iowait(void)
2798 unsigned long i, sum = 0;
2800 for_each_possible_cpu(i)
2801 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2806 unsigned long nr_iowait_cpu(int cpu)
2808 struct rq *this = cpu_rq(cpu);
2809 return atomic_read(&this->nr_iowait);
2812 #ifdef CONFIG_CPU_QUIET
2813 u64 nr_running_integral(unsigned int cpu)
2815 unsigned int seqcnt;
2819 if (cpu >= nr_cpu_ids)
2825 * Update average to avoid reading stalled value if there were
2826 * no run-queue changes for a long time. On the other hand if
2827 * the changes are happening right now, just read current value
2831 seqcnt = read_seqcount_begin(&q->ave_seqcnt);
2832 integral = do_nr_running_integral(q);
2833 if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
2834 read_seqcount_begin(&q->ave_seqcnt);
2835 integral = q->nr_running_integral;
2842 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2844 struct rq *rq = this_rq();
2845 *nr_waiters = atomic_read(&rq->nr_iowait);
2846 *load = rq->load.weight;
2852 * sched_exec - execve() is a valuable balancing opportunity, because at
2853 * this point the task has the smallest effective memory and cache footprint.
2855 void sched_exec(void)
2857 struct task_struct *p = current;
2858 unsigned long flags;
2861 raw_spin_lock_irqsave(&p->pi_lock, flags);
2862 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2863 if (dest_cpu == smp_processor_id())
2866 if (likely(cpu_active(dest_cpu))) {
2867 struct migration_arg arg = { p, dest_cpu };
2869 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2870 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2874 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2879 DEFINE_PER_CPU(struct kernel_stat, kstat);
2880 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2882 EXPORT_PER_CPU_SYMBOL(kstat);
2883 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2886 * Return accounted runtime for the task.
2887 * In case the task is currently running, return the runtime plus current's
2888 * pending runtime that have not been accounted yet.
2890 unsigned long long task_sched_runtime(struct task_struct *p)
2892 unsigned long flags;
2896 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2898 * 64-bit doesn't need locks to atomically read a 64bit value.
2899 * So we have a optimization chance when the task's delta_exec is 0.
2900 * Reading ->on_cpu is racy, but this is ok.
2902 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2903 * If we race with it entering cpu, unaccounted time is 0. This is
2904 * indistinguishable from the read occurring a few cycles earlier.
2905 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2906 * been accounted, so we're correct here as well.
2908 if (!p->on_cpu || !task_on_rq_queued(p))
2909 return p->se.sum_exec_runtime;
2912 rq = task_rq_lock(p, &flags);
2914 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2915 * project cycles that may never be accounted to this
2916 * thread, breaking clock_gettime().
2918 if (task_current(rq, p) && task_on_rq_queued(p)) {
2919 update_rq_clock(rq);
2920 p->sched_class->update_curr(rq);
2922 ns = p->se.sum_exec_runtime;
2923 task_rq_unlock(rq, p, &flags);
2928 #ifdef CONFIG_CPU_FREQ_GOV_SCHED
2931 unsigned long add_capacity_margin(unsigned long cpu_capacity)
2933 cpu_capacity = cpu_capacity * capacity_margin;
2934 cpu_capacity /= SCHED_CAPACITY_SCALE;
2935 return cpu_capacity;
2939 unsigned long sum_capacity_reqs(unsigned long cfs_cap,
2940 struct sched_capacity_reqs *scr)
2942 unsigned long total = add_capacity_margin(cfs_cap + scr->rt);
2943 return total += scr->dl;
2946 static void sched_freq_tick_pelt(int cpu)
2948 unsigned long cpu_utilization = capacity_max;
2949 unsigned long capacity_curr = capacity_curr_of(cpu);
2950 struct sched_capacity_reqs *scr;
2952 scr = &per_cpu(cpu_sched_capacity_reqs, cpu);
2953 if (sum_capacity_reqs(cpu_utilization, scr) < capacity_curr)
2957 * To make free room for a task that is building up its "real"
2958 * utilization and to harm its performance the least, request
2959 * a jump to a higher OPP as soon as the margin of free capacity
2960 * is impacted (specified by capacity_margin).
2962 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
2965 #ifdef CONFIG_SCHED_WALT
2966 static void sched_freq_tick_walt(int cpu)
2968 unsigned long cpu_utilization = cpu_util(cpu);
2969 unsigned long capacity_curr = capacity_curr_of(cpu);
2971 if (walt_disabled || !sysctl_sched_use_walt_cpu_util)
2972 return sched_freq_tick_pelt(cpu);
2975 * Add a margin to the WALT utilization.
2976 * NOTE: WALT tracks a single CPU signal for all the scheduling
2977 * classes, thus this margin is going to be added to the DL class as
2978 * well, which is something we do not do in sched_freq_tick_pelt case.
2980 cpu_utilization = add_capacity_margin(cpu_utilization);
2981 if (cpu_utilization <= capacity_curr)
2985 * It is likely that the load is growing so we
2986 * keep the added margin in our request as an
2989 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
2992 #define _sched_freq_tick(cpu) sched_freq_tick_walt(cpu)
2994 #define _sched_freq_tick(cpu) sched_freq_tick_pelt(cpu)
2995 #endif /* CONFIG_SCHED_WALT */
2997 static void sched_freq_tick(int cpu)
2999 unsigned long capacity_orig, capacity_curr;
3004 capacity_orig = capacity_orig_of(cpu);
3005 capacity_curr = capacity_curr_of(cpu);
3006 if (capacity_curr == capacity_orig)
3009 _sched_freq_tick(cpu);
3012 static inline void sched_freq_tick(int cpu) { }
3013 #endif /* CONFIG_CPU_FREQ_GOV_SCHED */
3016 * This function gets called by the timer code, with HZ frequency.
3017 * We call it with interrupts disabled.
3019 void scheduler_tick(void)
3021 int cpu = smp_processor_id();
3022 struct rq *rq = cpu_rq(cpu);
3023 struct task_struct *curr = rq->curr;
3027 raw_spin_lock(&rq->lock);
3028 walt_set_window_start(rq);
3029 update_rq_clock(rq);
3030 curr->sched_class->task_tick(rq, curr, 0);
3031 update_cpu_load_active(rq);
3032 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
3033 walt_ktime_clock(), 0);
3034 calc_global_load_tick(rq);
3035 sched_freq_tick(cpu);
3036 raw_spin_unlock(&rq->lock);
3038 perf_event_task_tick();
3041 rq->idle_balance = idle_cpu(cpu);
3042 trigger_load_balance(rq);
3044 rq_last_tick_reset(rq);
3047 #ifdef CONFIG_NO_HZ_FULL
3049 * scheduler_tick_max_deferment
3051 * Keep at least one tick per second when a single
3052 * active task is running because the scheduler doesn't
3053 * yet completely support full dynticks environment.
3055 * This makes sure that uptime, CFS vruntime, load
3056 * balancing, etc... continue to move forward, even
3057 * with a very low granularity.
3059 * Return: Maximum deferment in nanoseconds.
3061 u64 scheduler_tick_max_deferment(void)
3063 struct rq *rq = this_rq();
3064 unsigned long next, now = READ_ONCE(jiffies);
3066 next = rq->last_sched_tick + HZ;
3068 if (time_before_eq(next, now))
3071 return jiffies_to_nsecs(next - now);
3075 notrace unsigned long get_parent_ip(unsigned long addr)
3077 if (in_lock_functions(addr)) {
3078 addr = CALLER_ADDR2;
3079 if (in_lock_functions(addr))
3080 addr = CALLER_ADDR3;
3085 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3086 defined(CONFIG_PREEMPT_TRACER))
3088 void preempt_count_add(int val)
3090 #ifdef CONFIG_DEBUG_PREEMPT
3094 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3097 __preempt_count_add(val);
3098 #ifdef CONFIG_DEBUG_PREEMPT
3100 * Spinlock count overflowing soon?
3102 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3105 if (preempt_count() == val) {
3106 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3107 #ifdef CONFIG_DEBUG_PREEMPT
3108 current->preempt_disable_ip = ip;
3110 trace_preempt_off(CALLER_ADDR0, ip);
3113 EXPORT_SYMBOL(preempt_count_add);
3114 NOKPROBE_SYMBOL(preempt_count_add);
3116 void preempt_count_sub(int val)
3118 #ifdef CONFIG_DEBUG_PREEMPT
3122 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3125 * Is the spinlock portion underflowing?
3127 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3128 !(preempt_count() & PREEMPT_MASK)))
3132 if (preempt_count() == val)
3133 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3134 __preempt_count_sub(val);
3136 EXPORT_SYMBOL(preempt_count_sub);
3137 NOKPROBE_SYMBOL(preempt_count_sub);
3142 * Print scheduling while atomic bug:
3144 static noinline void __schedule_bug(struct task_struct *prev)
3146 if (oops_in_progress)
3149 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3150 prev->comm, prev->pid, preempt_count());
3152 debug_show_held_locks(prev);
3154 if (irqs_disabled())
3155 print_irqtrace_events(prev);
3156 #ifdef CONFIG_DEBUG_PREEMPT
3157 if (in_atomic_preempt_off()) {
3158 pr_err("Preemption disabled at:");
3159 print_ip_sym(current->preempt_disable_ip);
3164 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3168 * Various schedule()-time debugging checks and statistics:
3170 static inline void schedule_debug(struct task_struct *prev)
3172 #ifdef CONFIG_SCHED_STACK_END_CHECK
3173 if (task_stack_end_corrupted(prev))
3174 panic("corrupted stack end detected inside scheduler\n");
3177 if (unlikely(in_atomic_preempt_off())) {
3178 __schedule_bug(prev);
3179 preempt_count_set(PREEMPT_DISABLED);
3183 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3185 schedstat_inc(this_rq(), sched_count);
3189 * Pick up the highest-prio task:
3191 static inline struct task_struct *
3192 pick_next_task(struct rq *rq, struct task_struct *prev)
3194 const struct sched_class *class = &fair_sched_class;
3195 struct task_struct *p;
3198 * Optimization: we know that if all tasks are in
3199 * the fair class we can call that function directly:
3201 if (likely(prev->sched_class == class &&
3202 rq->nr_running == rq->cfs.h_nr_running)) {
3203 p = fair_sched_class.pick_next_task(rq, prev);
3204 if (unlikely(p == RETRY_TASK))
3207 /* assumes fair_sched_class->next == idle_sched_class */
3209 p = idle_sched_class.pick_next_task(rq, prev);
3215 for_each_class(class) {
3216 p = class->pick_next_task(rq, prev);
3218 if (unlikely(p == RETRY_TASK))
3224 BUG(); /* the idle class will always have a runnable task */
3228 * __schedule() is the main scheduler function.
3230 * The main means of driving the scheduler and thus entering this function are:
3232 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3234 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3235 * paths. For example, see arch/x86/entry_64.S.
3237 * To drive preemption between tasks, the scheduler sets the flag in timer
3238 * interrupt handler scheduler_tick().
3240 * 3. Wakeups don't really cause entry into schedule(). They add a
3241 * task to the run-queue and that's it.
3243 * Now, if the new task added to the run-queue preempts the current
3244 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3245 * called on the nearest possible occasion:
3247 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3249 * - in syscall or exception context, at the next outmost
3250 * preempt_enable(). (this might be as soon as the wake_up()'s
3253 * - in IRQ context, return from interrupt-handler to
3254 * preemptible context
3256 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3259 * - cond_resched() call
3260 * - explicit schedule() call
3261 * - return from syscall or exception to user-space
3262 * - return from interrupt-handler to user-space
3264 * WARNING: must be called with preemption disabled!
3266 static void __sched notrace __schedule(bool preempt)
3268 struct task_struct *prev, *next;
3269 unsigned long *switch_count;
3274 cpu = smp_processor_id();
3276 rcu_note_context_switch();
3280 * do_exit() calls schedule() with preemption disabled as an exception;
3281 * however we must fix that up, otherwise the next task will see an
3282 * inconsistent (higher) preempt count.
3284 * It also avoids the below schedule_debug() test from complaining
3287 if (unlikely(prev->state == TASK_DEAD))
3288 preempt_enable_no_resched_notrace();
3290 schedule_debug(prev);
3292 if (sched_feat(HRTICK))
3296 * Make sure that signal_pending_state()->signal_pending() below
3297 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3298 * done by the caller to avoid the race with signal_wake_up().
3300 smp_mb__before_spinlock();
3301 raw_spin_lock_irq(&rq->lock);
3302 lockdep_pin_lock(&rq->lock);
3304 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3306 switch_count = &prev->nivcsw;
3307 if (!preempt && prev->state) {
3308 if (unlikely(signal_pending_state(prev->state, prev))) {
3309 prev->state = TASK_RUNNING;
3311 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3315 * If a worker went to sleep, notify and ask workqueue
3316 * whether it wants to wake up a task to maintain
3319 if (prev->flags & PF_WQ_WORKER) {
3320 struct task_struct *to_wakeup;
3322 to_wakeup = wq_worker_sleeping(prev, cpu);
3324 try_to_wake_up_local(to_wakeup);
3327 switch_count = &prev->nvcsw;
3330 if (task_on_rq_queued(prev))
3331 update_rq_clock(rq);
3333 next = pick_next_task(rq, prev);
3334 wallclock = walt_ktime_clock();
3335 walt_update_task_ravg(prev, rq, PUT_PREV_TASK, wallclock, 0);
3336 walt_update_task_ravg(next, rq, PICK_NEXT_TASK, wallclock, 0);
3337 clear_tsk_need_resched(prev);
3338 clear_preempt_need_resched();
3339 rq->clock_skip_update = 0;
3341 if (likely(prev != next)) {
3346 trace_sched_switch(preempt, prev, next);
3347 rq = context_switch(rq, prev, next); /* unlocks the rq */
3350 lockdep_unpin_lock(&rq->lock);
3351 raw_spin_unlock_irq(&rq->lock);
3354 balance_callback(rq);
3357 static inline void sched_submit_work(struct task_struct *tsk)
3359 if (!tsk->state || tsk_is_pi_blocked(tsk))
3362 * If we are going to sleep and we have plugged IO queued,
3363 * make sure to submit it to avoid deadlocks.
3365 if (blk_needs_flush_plug(tsk))
3366 blk_schedule_flush_plug(tsk);
3369 asmlinkage __visible void __sched schedule(void)
3371 struct task_struct *tsk = current;
3373 sched_submit_work(tsk);
3377 sched_preempt_enable_no_resched();
3378 } while (need_resched());
3380 EXPORT_SYMBOL(schedule);
3382 #ifdef CONFIG_CONTEXT_TRACKING
3383 asmlinkage __visible void __sched schedule_user(void)
3386 * If we come here after a random call to set_need_resched(),
3387 * or we have been woken up remotely but the IPI has not yet arrived,
3388 * we haven't yet exited the RCU idle mode. Do it here manually until
3389 * we find a better solution.
3391 * NB: There are buggy callers of this function. Ideally we
3392 * should warn if prev_state != CONTEXT_USER, but that will trigger
3393 * too frequently to make sense yet.
3395 enum ctx_state prev_state = exception_enter();
3397 exception_exit(prev_state);
3402 * schedule_preempt_disabled - called with preemption disabled
3404 * Returns with preemption disabled. Note: preempt_count must be 1
3406 void __sched schedule_preempt_disabled(void)
3408 sched_preempt_enable_no_resched();
3413 static void __sched notrace preempt_schedule_common(void)
3416 preempt_disable_notrace();
3418 preempt_enable_no_resched_notrace();
3421 * Check again in case we missed a preemption opportunity
3422 * between schedule and now.
3424 } while (need_resched());
3427 #ifdef CONFIG_PREEMPT
3429 * this is the entry point to schedule() from in-kernel preemption
3430 * off of preempt_enable. Kernel preemptions off return from interrupt
3431 * occur there and call schedule directly.
3433 asmlinkage __visible void __sched notrace preempt_schedule(void)
3436 * If there is a non-zero preempt_count or interrupts are disabled,
3437 * we do not want to preempt the current task. Just return..
3439 if (likely(!preemptible()))
3442 preempt_schedule_common();
3444 NOKPROBE_SYMBOL(preempt_schedule);
3445 EXPORT_SYMBOL(preempt_schedule);
3448 * preempt_schedule_notrace - preempt_schedule called by tracing
3450 * The tracing infrastructure uses preempt_enable_notrace to prevent
3451 * recursion and tracing preempt enabling caused by the tracing
3452 * infrastructure itself. But as tracing can happen in areas coming
3453 * from userspace or just about to enter userspace, a preempt enable
3454 * can occur before user_exit() is called. This will cause the scheduler
3455 * to be called when the system is still in usermode.
3457 * To prevent this, the preempt_enable_notrace will use this function
3458 * instead of preempt_schedule() to exit user context if needed before
3459 * calling the scheduler.
3461 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3463 enum ctx_state prev_ctx;
3465 if (likely(!preemptible()))
3469 preempt_disable_notrace();
3471 * Needs preempt disabled in case user_exit() is traced
3472 * and the tracer calls preempt_enable_notrace() causing
3473 * an infinite recursion.
3475 prev_ctx = exception_enter();
3477 exception_exit(prev_ctx);
3479 preempt_enable_no_resched_notrace();
3480 } while (need_resched());
3482 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3484 #endif /* CONFIG_PREEMPT */
3487 * this is the entry point to schedule() from kernel preemption
3488 * off of irq context.
3489 * Note, that this is called and return with irqs disabled. This will
3490 * protect us against recursive calling from irq.
3492 asmlinkage __visible void __sched preempt_schedule_irq(void)
3494 enum ctx_state prev_state;
3496 /* Catch callers which need to be fixed */
3497 BUG_ON(preempt_count() || !irqs_disabled());
3499 prev_state = exception_enter();
3505 local_irq_disable();
3506 sched_preempt_enable_no_resched();
3507 } while (need_resched());
3509 exception_exit(prev_state);
3512 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3515 return try_to_wake_up(curr->private, mode, wake_flags);
3517 EXPORT_SYMBOL(default_wake_function);
3519 #ifdef CONFIG_RT_MUTEXES
3522 * rt_mutex_setprio - set the current priority of a task
3524 * @prio: prio value (kernel-internal form)
3526 * This function changes the 'effective' priority of a task. It does
3527 * not touch ->normal_prio like __setscheduler().
3529 * Used by the rt_mutex code to implement priority inheritance
3530 * logic. Call site only calls if the priority of the task changed.
3532 void rt_mutex_setprio(struct task_struct *p, int prio)
3534 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3536 const struct sched_class *prev_class;
3538 BUG_ON(prio > MAX_PRIO);
3540 rq = __task_rq_lock(p);
3543 * Idle task boosting is a nono in general. There is one
3544 * exception, when PREEMPT_RT and NOHZ is active:
3546 * The idle task calls get_next_timer_interrupt() and holds
3547 * the timer wheel base->lock on the CPU and another CPU wants
3548 * to access the timer (probably to cancel it). We can safely
3549 * ignore the boosting request, as the idle CPU runs this code
3550 * with interrupts disabled and will complete the lock
3551 * protected section without being interrupted. So there is no
3552 * real need to boost.
3554 if (unlikely(p == rq->idle)) {
3555 WARN_ON(p != rq->curr);
3556 WARN_ON(p->pi_blocked_on);
3560 trace_sched_pi_setprio(p, prio);
3562 prev_class = p->sched_class;
3563 queued = task_on_rq_queued(p);
3564 running = task_current(rq, p);
3566 dequeue_task(rq, p, DEQUEUE_SAVE);
3568 put_prev_task(rq, p);
3571 * Boosting condition are:
3572 * 1. -rt task is running and holds mutex A
3573 * --> -dl task blocks on mutex A
3575 * 2. -dl task is running and holds mutex A
3576 * --> -dl task blocks on mutex A and could preempt the
3579 if (dl_prio(prio)) {
3580 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3581 if (!dl_prio(p->normal_prio) ||
3582 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3583 p->dl.dl_boosted = 1;
3584 enqueue_flag |= ENQUEUE_REPLENISH;
3586 p->dl.dl_boosted = 0;
3587 p->sched_class = &dl_sched_class;
3588 } else if (rt_prio(prio)) {
3589 if (dl_prio(oldprio))
3590 p->dl.dl_boosted = 0;
3592 enqueue_flag |= ENQUEUE_HEAD;
3593 p->sched_class = &rt_sched_class;
3595 if (dl_prio(oldprio))
3596 p->dl.dl_boosted = 0;
3597 if (rt_prio(oldprio))
3599 p->sched_class = &fair_sched_class;
3605 p->sched_class->set_curr_task(rq);
3607 enqueue_task(rq, p, enqueue_flag);
3609 check_class_changed(rq, p, prev_class, oldprio);
3611 preempt_disable(); /* avoid rq from going away on us */
3612 __task_rq_unlock(rq);
3614 balance_callback(rq);
3619 void set_user_nice(struct task_struct *p, long nice)
3621 int old_prio, delta, queued;
3622 unsigned long flags;
3625 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3628 * We have to be careful, if called from sys_setpriority(),
3629 * the task might be in the middle of scheduling on another CPU.
3631 rq = task_rq_lock(p, &flags);
3633 * The RT priorities are set via sched_setscheduler(), but we still
3634 * allow the 'normal' nice value to be set - but as expected
3635 * it wont have any effect on scheduling until the task is
3636 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3638 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3639 p->static_prio = NICE_TO_PRIO(nice);
3642 queued = task_on_rq_queued(p);
3644 dequeue_task(rq, p, DEQUEUE_SAVE);
3646 p->static_prio = NICE_TO_PRIO(nice);
3649 p->prio = effective_prio(p);
3650 delta = p->prio - old_prio;
3653 enqueue_task(rq, p, ENQUEUE_RESTORE);
3655 * If the task increased its priority or is running and
3656 * lowered its priority, then reschedule its CPU:
3658 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3662 task_rq_unlock(rq, p, &flags);
3664 EXPORT_SYMBOL(set_user_nice);
3667 * can_nice - check if a task can reduce its nice value
3671 int can_nice(const struct task_struct *p, const int nice)
3673 /* convert nice value [19,-20] to rlimit style value [1,40] */
3674 int nice_rlim = nice_to_rlimit(nice);
3676 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3677 capable(CAP_SYS_NICE));
3680 #ifdef __ARCH_WANT_SYS_NICE
3683 * sys_nice - change the priority of the current process.
3684 * @increment: priority increment
3686 * sys_setpriority is a more generic, but much slower function that
3687 * does similar things.
3689 SYSCALL_DEFINE1(nice, int, increment)
3694 * Setpriority might change our priority at the same moment.
3695 * We don't have to worry. Conceptually one call occurs first
3696 * and we have a single winner.
3698 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3699 nice = task_nice(current) + increment;
3701 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3702 if (increment < 0 && !can_nice(current, nice))
3705 retval = security_task_setnice(current, nice);
3709 set_user_nice(current, nice);
3716 * task_prio - return the priority value of a given task.
3717 * @p: the task in question.
3719 * Return: The priority value as seen by users in /proc.
3720 * RT tasks are offset by -200. Normal tasks are centered
3721 * around 0, value goes from -16 to +15.
3723 int task_prio(const struct task_struct *p)
3725 return p->prio - MAX_RT_PRIO;
3729 * idle_cpu - is a given cpu idle currently?
3730 * @cpu: the processor in question.
3732 * Return: 1 if the CPU is currently idle. 0 otherwise.
3734 int idle_cpu(int cpu)
3736 struct rq *rq = cpu_rq(cpu);
3738 if (rq->curr != rq->idle)
3745 if (!llist_empty(&rq->wake_list))
3753 * idle_task - return the idle task for a given cpu.
3754 * @cpu: the processor in question.
3756 * Return: The idle task for the cpu @cpu.
3758 struct task_struct *idle_task(int cpu)
3760 return cpu_rq(cpu)->idle;
3764 * find_process_by_pid - find a process with a matching PID value.
3765 * @pid: the pid in question.
3767 * The task of @pid, if found. %NULL otherwise.
3769 static struct task_struct *find_process_by_pid(pid_t pid)
3771 return pid ? find_task_by_vpid(pid) : current;
3775 * This function initializes the sched_dl_entity of a newly becoming
3776 * SCHED_DEADLINE task.
3778 * Only the static values are considered here, the actual runtime and the
3779 * absolute deadline will be properly calculated when the task is enqueued
3780 * for the first time with its new policy.
3783 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3785 struct sched_dl_entity *dl_se = &p->dl;
3787 dl_se->dl_runtime = attr->sched_runtime;
3788 dl_se->dl_deadline = attr->sched_deadline;
3789 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3790 dl_se->flags = attr->sched_flags;
3791 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3794 * Changing the parameters of a task is 'tricky' and we're not doing
3795 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3797 * What we SHOULD do is delay the bandwidth release until the 0-lag
3798 * point. This would include retaining the task_struct until that time
3799 * and change dl_overflow() to not immediately decrement the current
3802 * Instead we retain the current runtime/deadline and let the new
3803 * parameters take effect after the current reservation period lapses.
3804 * This is safe (albeit pessimistic) because the 0-lag point is always
3805 * before the current scheduling deadline.
3807 * We can still have temporary overloads because we do not delay the
3808 * change in bandwidth until that time; so admission control is
3809 * not on the safe side. It does however guarantee tasks will never
3810 * consume more than promised.
3815 * sched_setparam() passes in -1 for its policy, to let the functions
3816 * it calls know not to change it.
3818 #define SETPARAM_POLICY -1
3820 static void __setscheduler_params(struct task_struct *p,
3821 const struct sched_attr *attr)
3823 int policy = attr->sched_policy;
3825 if (policy == SETPARAM_POLICY)
3830 if (dl_policy(policy))
3831 __setparam_dl(p, attr);
3832 else if (fair_policy(policy))
3833 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3836 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3837 * !rt_policy. Always setting this ensures that things like
3838 * getparam()/getattr() don't report silly values for !rt tasks.
3840 p->rt_priority = attr->sched_priority;
3841 p->normal_prio = normal_prio(p);
3845 /* Actually do priority change: must hold pi & rq lock. */
3846 static void __setscheduler(struct rq *rq, struct task_struct *p,
3847 const struct sched_attr *attr, bool keep_boost)
3849 __setscheduler_params(p, attr);
3852 * Keep a potential priority boosting if called from
3853 * sched_setscheduler().
3856 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3858 p->prio = normal_prio(p);
3860 if (dl_prio(p->prio))
3861 p->sched_class = &dl_sched_class;
3862 else if (rt_prio(p->prio))
3863 p->sched_class = &rt_sched_class;
3865 p->sched_class = &fair_sched_class;
3869 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3871 struct sched_dl_entity *dl_se = &p->dl;
3873 attr->sched_priority = p->rt_priority;
3874 attr->sched_runtime = dl_se->dl_runtime;
3875 attr->sched_deadline = dl_se->dl_deadline;
3876 attr->sched_period = dl_se->dl_period;
3877 attr->sched_flags = dl_se->flags;
3881 * This function validates the new parameters of a -deadline task.
3882 * We ask for the deadline not being zero, and greater or equal
3883 * than the runtime, as well as the period of being zero or
3884 * greater than deadline. Furthermore, we have to be sure that
3885 * user parameters are above the internal resolution of 1us (we
3886 * check sched_runtime only since it is always the smaller one) and
3887 * below 2^63 ns (we have to check both sched_deadline and
3888 * sched_period, as the latter can be zero).
3891 __checkparam_dl(const struct sched_attr *attr)
3894 if (attr->sched_deadline == 0)
3898 * Since we truncate DL_SCALE bits, make sure we're at least
3901 if (attr->sched_runtime < (1ULL << DL_SCALE))
3905 * Since we use the MSB for wrap-around and sign issues, make
3906 * sure it's not set (mind that period can be equal to zero).
3908 if (attr->sched_deadline & (1ULL << 63) ||
3909 attr->sched_period & (1ULL << 63))
3912 /* runtime <= deadline <= period (if period != 0) */
3913 if ((attr->sched_period != 0 &&
3914 attr->sched_period < attr->sched_deadline) ||
3915 attr->sched_deadline < attr->sched_runtime)
3922 * check the target process has a UID that matches the current process's
3924 static bool check_same_owner(struct task_struct *p)
3926 const struct cred *cred = current_cred(), *pcred;
3930 pcred = __task_cred(p);
3931 match = (uid_eq(cred->euid, pcred->euid) ||
3932 uid_eq(cred->euid, pcred->uid));
3937 static bool dl_param_changed(struct task_struct *p,
3938 const struct sched_attr *attr)
3940 struct sched_dl_entity *dl_se = &p->dl;
3942 if (dl_se->dl_runtime != attr->sched_runtime ||
3943 dl_se->dl_deadline != attr->sched_deadline ||
3944 dl_se->dl_period != attr->sched_period ||
3945 dl_se->flags != attr->sched_flags)
3951 static int __sched_setscheduler(struct task_struct *p,
3952 const struct sched_attr *attr,
3955 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3956 MAX_RT_PRIO - 1 - attr->sched_priority;
3957 int retval, oldprio, oldpolicy = -1, queued, running;
3958 int new_effective_prio, policy = attr->sched_policy;
3959 unsigned long flags;
3960 const struct sched_class *prev_class;
3964 /* may grab non-irq protected spin_locks */
3965 BUG_ON(in_interrupt());
3967 /* double check policy once rq lock held */
3969 reset_on_fork = p->sched_reset_on_fork;
3970 policy = oldpolicy = p->policy;
3972 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3974 if (!valid_policy(policy))
3978 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3982 * Valid priorities for SCHED_FIFO and SCHED_RR are
3983 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3984 * SCHED_BATCH and SCHED_IDLE is 0.
3986 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3987 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3989 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3990 (rt_policy(policy) != (attr->sched_priority != 0)))
3994 * Allow unprivileged RT tasks to decrease priority:
3996 if (user && !capable(CAP_SYS_NICE)) {
3997 if (fair_policy(policy)) {
3998 if (attr->sched_nice < task_nice(p) &&
3999 !can_nice(p, attr->sched_nice))
4003 if (rt_policy(policy)) {
4004 unsigned long rlim_rtprio =
4005 task_rlimit(p, RLIMIT_RTPRIO);
4007 /* can't set/change the rt policy */
4008 if (policy != p->policy && !rlim_rtprio)
4011 /* can't increase priority */
4012 if (attr->sched_priority > p->rt_priority &&
4013 attr->sched_priority > rlim_rtprio)
4018 * Can't set/change SCHED_DEADLINE policy at all for now
4019 * (safest behavior); in the future we would like to allow
4020 * unprivileged DL tasks to increase their relative deadline
4021 * or reduce their runtime (both ways reducing utilization)
4023 if (dl_policy(policy))
4027 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4028 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4030 if (idle_policy(p->policy) && !idle_policy(policy)) {
4031 if (!can_nice(p, task_nice(p)))
4035 /* can't change other user's priorities */
4036 if (!check_same_owner(p))
4039 /* Normal users shall not reset the sched_reset_on_fork flag */
4040 if (p->sched_reset_on_fork && !reset_on_fork)
4045 retval = security_task_setscheduler(p);
4051 * make sure no PI-waiters arrive (or leave) while we are
4052 * changing the priority of the task:
4054 * To be able to change p->policy safely, the appropriate
4055 * runqueue lock must be held.
4057 rq = task_rq_lock(p, &flags);
4060 * Changing the policy of the stop threads its a very bad idea
4062 if (p == rq->stop) {
4063 task_rq_unlock(rq, p, &flags);
4068 * If not changing anything there's no need to proceed further,
4069 * but store a possible modification of reset_on_fork.
4071 if (unlikely(policy == p->policy)) {
4072 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4074 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4076 if (dl_policy(policy) && dl_param_changed(p, attr))
4079 p->sched_reset_on_fork = reset_on_fork;
4080 task_rq_unlock(rq, p, &flags);
4086 #ifdef CONFIG_RT_GROUP_SCHED
4088 * Do not allow realtime tasks into groups that have no runtime
4091 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4092 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4093 !task_group_is_autogroup(task_group(p))) {
4094 task_rq_unlock(rq, p, &flags);
4099 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4100 cpumask_t *span = rq->rd->span;
4103 * Don't allow tasks with an affinity mask smaller than
4104 * the entire root_domain to become SCHED_DEADLINE. We
4105 * will also fail if there's no bandwidth available.
4107 if (!cpumask_subset(span, &p->cpus_allowed) ||
4108 rq->rd->dl_bw.bw == 0) {
4109 task_rq_unlock(rq, p, &flags);
4116 /* recheck policy now with rq lock held */
4117 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4118 policy = oldpolicy = -1;
4119 task_rq_unlock(rq, p, &flags);
4124 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4125 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4128 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4129 task_rq_unlock(rq, p, &flags);
4133 p->sched_reset_on_fork = reset_on_fork;
4138 * Take priority boosted tasks into account. If the new
4139 * effective priority is unchanged, we just store the new
4140 * normal parameters and do not touch the scheduler class and
4141 * the runqueue. This will be done when the task deboost
4144 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4145 if (new_effective_prio == oldprio) {
4146 __setscheduler_params(p, attr);
4147 task_rq_unlock(rq, p, &flags);
4152 queued = task_on_rq_queued(p);
4153 running = task_current(rq, p);
4155 dequeue_task(rq, p, DEQUEUE_SAVE);
4157 put_prev_task(rq, p);
4159 prev_class = p->sched_class;
4160 __setscheduler(rq, p, attr, pi);
4163 p->sched_class->set_curr_task(rq);
4165 int enqueue_flags = ENQUEUE_RESTORE;
4167 * We enqueue to tail when the priority of a task is
4168 * increased (user space view).
4170 if (oldprio <= p->prio)
4171 enqueue_flags |= ENQUEUE_HEAD;
4173 enqueue_task(rq, p, enqueue_flags);
4176 check_class_changed(rq, p, prev_class, oldprio);
4177 preempt_disable(); /* avoid rq from going away on us */
4178 task_rq_unlock(rq, p, &flags);
4181 rt_mutex_adjust_pi(p);
4184 * Run balance callbacks after we've adjusted the PI chain.
4186 balance_callback(rq);
4192 static int _sched_setscheduler(struct task_struct *p, int policy,
4193 const struct sched_param *param, bool check)
4195 struct sched_attr attr = {
4196 .sched_policy = policy,
4197 .sched_priority = param->sched_priority,
4198 .sched_nice = PRIO_TO_NICE(p->static_prio),
4201 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4202 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4203 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4204 policy &= ~SCHED_RESET_ON_FORK;
4205 attr.sched_policy = policy;
4208 return __sched_setscheduler(p, &attr, check, true);
4211 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4212 * @p: the task in question.
4213 * @policy: new policy.
4214 * @param: structure containing the new RT priority.
4216 * Return: 0 on success. An error code otherwise.
4218 * NOTE that the task may be already dead.
4220 int sched_setscheduler(struct task_struct *p, int policy,
4221 const struct sched_param *param)
4223 return _sched_setscheduler(p, policy, param, true);
4225 EXPORT_SYMBOL_GPL(sched_setscheduler);
4227 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4229 return __sched_setscheduler(p, attr, true, true);
4231 EXPORT_SYMBOL_GPL(sched_setattr);
4234 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4235 * @p: the task in question.
4236 * @policy: new policy.
4237 * @param: structure containing the new RT priority.
4239 * Just like sched_setscheduler, only don't bother checking if the
4240 * current context has permission. For example, this is needed in
4241 * stop_machine(): we create temporary high priority worker threads,
4242 * but our caller might not have that capability.
4244 * Return: 0 on success. An error code otherwise.
4246 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4247 const struct sched_param *param)
4249 return _sched_setscheduler(p, policy, param, false);
4251 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4254 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4256 struct sched_param lparam;
4257 struct task_struct *p;
4260 if (!param || pid < 0)
4262 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4267 p = find_process_by_pid(pid);
4269 retval = sched_setscheduler(p, policy, &lparam);
4276 * Mimics kernel/events/core.c perf_copy_attr().
4278 static int sched_copy_attr(struct sched_attr __user *uattr,
4279 struct sched_attr *attr)
4284 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4288 * zero the full structure, so that a short copy will be nice.
4290 memset(attr, 0, sizeof(*attr));
4292 ret = get_user(size, &uattr->size);
4296 if (size > PAGE_SIZE) /* silly large */
4299 if (!size) /* abi compat */
4300 size = SCHED_ATTR_SIZE_VER0;
4302 if (size < SCHED_ATTR_SIZE_VER0)
4306 * If we're handed a bigger struct than we know of,
4307 * ensure all the unknown bits are 0 - i.e. new
4308 * user-space does not rely on any kernel feature
4309 * extensions we dont know about yet.
4311 if (size > sizeof(*attr)) {
4312 unsigned char __user *addr;
4313 unsigned char __user *end;
4316 addr = (void __user *)uattr + sizeof(*attr);
4317 end = (void __user *)uattr + size;
4319 for (; addr < end; addr++) {
4320 ret = get_user(val, addr);
4326 size = sizeof(*attr);
4329 ret = copy_from_user(attr, uattr, size);
4334 * XXX: do we want to be lenient like existing syscalls; or do we want
4335 * to be strict and return an error on out-of-bounds values?
4337 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4342 put_user(sizeof(*attr), &uattr->size);
4347 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4348 * @pid: the pid in question.
4349 * @policy: new policy.
4350 * @param: structure containing the new RT priority.
4352 * Return: 0 on success. An error code otherwise.
4354 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4355 struct sched_param __user *, param)
4357 /* negative values for policy are not valid */
4361 return do_sched_setscheduler(pid, policy, param);
4365 * sys_sched_setparam - set/change the RT priority of a thread
4366 * @pid: the pid in question.
4367 * @param: structure containing the new RT priority.
4369 * Return: 0 on success. An error code otherwise.
4371 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4373 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4377 * sys_sched_setattr - same as above, but with extended sched_attr
4378 * @pid: the pid in question.
4379 * @uattr: structure containing the extended parameters.
4380 * @flags: for future extension.
4382 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4383 unsigned int, flags)
4385 struct sched_attr attr;
4386 struct task_struct *p;
4389 if (!uattr || pid < 0 || flags)
4392 retval = sched_copy_attr(uattr, &attr);
4396 if ((int)attr.sched_policy < 0)
4401 p = find_process_by_pid(pid);
4403 retval = sched_setattr(p, &attr);
4410 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4411 * @pid: the pid in question.
4413 * Return: On success, the policy of the thread. Otherwise, a negative error
4416 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4418 struct task_struct *p;
4426 p = find_process_by_pid(pid);
4428 retval = security_task_getscheduler(p);
4431 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4438 * sys_sched_getparam - get the RT priority of a thread
4439 * @pid: the pid in question.
4440 * @param: structure containing the RT priority.
4442 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4445 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4447 struct sched_param lp = { .sched_priority = 0 };
4448 struct task_struct *p;
4451 if (!param || pid < 0)
4455 p = find_process_by_pid(pid);
4460 retval = security_task_getscheduler(p);
4464 if (task_has_rt_policy(p))
4465 lp.sched_priority = p->rt_priority;
4469 * This one might sleep, we cannot do it with a spinlock held ...
4471 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4480 static int sched_read_attr(struct sched_attr __user *uattr,
4481 struct sched_attr *attr,
4486 if (!access_ok(VERIFY_WRITE, uattr, usize))
4490 * If we're handed a smaller struct than we know of,
4491 * ensure all the unknown bits are 0 - i.e. old
4492 * user-space does not get uncomplete information.
4494 if (usize < sizeof(*attr)) {
4495 unsigned char *addr;
4498 addr = (void *)attr + usize;
4499 end = (void *)attr + sizeof(*attr);
4501 for (; addr < end; addr++) {
4509 ret = copy_to_user(uattr, attr, attr->size);
4517 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4518 * @pid: the pid in question.
4519 * @uattr: structure containing the extended parameters.
4520 * @size: sizeof(attr) for fwd/bwd comp.
4521 * @flags: for future extension.
4523 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4524 unsigned int, size, unsigned int, flags)
4526 struct sched_attr attr = {
4527 .size = sizeof(struct sched_attr),
4529 struct task_struct *p;
4532 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4533 size < SCHED_ATTR_SIZE_VER0 || flags)
4537 p = find_process_by_pid(pid);
4542 retval = security_task_getscheduler(p);
4546 attr.sched_policy = p->policy;
4547 if (p->sched_reset_on_fork)
4548 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4549 if (task_has_dl_policy(p))
4550 __getparam_dl(p, &attr);
4551 else if (task_has_rt_policy(p))
4552 attr.sched_priority = p->rt_priority;
4554 attr.sched_nice = task_nice(p);
4558 retval = sched_read_attr(uattr, &attr, size);
4566 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4568 cpumask_var_t cpus_allowed, new_mask;
4569 struct task_struct *p;
4574 p = find_process_by_pid(pid);
4580 /* Prevent p going away */
4584 if (p->flags & PF_NO_SETAFFINITY) {
4588 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4592 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4594 goto out_free_cpus_allowed;
4597 if (!check_same_owner(p)) {
4599 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4601 goto out_free_new_mask;
4606 retval = security_task_setscheduler(p);
4608 goto out_free_new_mask;
4611 cpuset_cpus_allowed(p, cpus_allowed);
4612 cpumask_and(new_mask, in_mask, cpus_allowed);
4615 * Since bandwidth control happens on root_domain basis,
4616 * if admission test is enabled, we only admit -deadline
4617 * tasks allowed to run on all the CPUs in the task's
4621 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4623 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4626 goto out_free_new_mask;
4632 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4635 cpuset_cpus_allowed(p, cpus_allowed);
4636 if (!cpumask_subset(new_mask, cpus_allowed)) {
4638 * We must have raced with a concurrent cpuset
4639 * update. Just reset the cpus_allowed to the
4640 * cpuset's cpus_allowed
4642 cpumask_copy(new_mask, cpus_allowed);
4647 free_cpumask_var(new_mask);
4648 out_free_cpus_allowed:
4649 free_cpumask_var(cpus_allowed);
4655 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4656 struct cpumask *new_mask)
4658 if (len < cpumask_size())
4659 cpumask_clear(new_mask);
4660 else if (len > cpumask_size())
4661 len = cpumask_size();
4663 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4667 * sys_sched_setaffinity - set the cpu affinity of a process
4668 * @pid: pid of the process
4669 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4670 * @user_mask_ptr: user-space pointer to the new cpu mask
4672 * Return: 0 on success. An error code otherwise.
4674 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4675 unsigned long __user *, user_mask_ptr)
4677 cpumask_var_t new_mask;
4680 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4683 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4685 retval = sched_setaffinity(pid, new_mask);
4686 free_cpumask_var(new_mask);
4690 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4692 struct task_struct *p;
4693 unsigned long flags;
4699 p = find_process_by_pid(pid);
4703 retval = security_task_getscheduler(p);
4707 raw_spin_lock_irqsave(&p->pi_lock, flags);
4708 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4709 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4718 * sys_sched_getaffinity - get the cpu affinity of a process
4719 * @pid: pid of the process
4720 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4721 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4723 * Return: 0 on success. An error code otherwise.
4725 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4726 unsigned long __user *, user_mask_ptr)
4731 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4733 if (len & (sizeof(unsigned long)-1))
4736 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4739 ret = sched_getaffinity(pid, mask);
4741 size_t retlen = min_t(size_t, len, cpumask_size());
4743 if (copy_to_user(user_mask_ptr, mask, retlen))
4748 free_cpumask_var(mask);
4754 * sys_sched_yield - yield the current processor to other threads.
4756 * This function yields the current CPU to other tasks. If there are no
4757 * other threads running on this CPU then this function will return.
4761 SYSCALL_DEFINE0(sched_yield)
4763 struct rq *rq = this_rq_lock();
4765 schedstat_inc(rq, yld_count);
4766 current->sched_class->yield_task(rq);
4769 * Since we are going to call schedule() anyway, there's
4770 * no need to preempt or enable interrupts:
4772 __release(rq->lock);
4773 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4774 do_raw_spin_unlock(&rq->lock);
4775 sched_preempt_enable_no_resched();
4782 int __sched _cond_resched(void)
4784 if (should_resched(0)) {
4785 preempt_schedule_common();
4790 EXPORT_SYMBOL(_cond_resched);
4793 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4794 * call schedule, and on return reacquire the lock.
4796 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4797 * operations here to prevent schedule() from being called twice (once via
4798 * spin_unlock(), once by hand).
4800 int __cond_resched_lock(spinlock_t *lock)
4802 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4805 lockdep_assert_held(lock);
4807 if (spin_needbreak(lock) || resched) {
4810 preempt_schedule_common();
4818 EXPORT_SYMBOL(__cond_resched_lock);
4820 int __sched __cond_resched_softirq(void)
4822 BUG_ON(!in_softirq());
4824 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4826 preempt_schedule_common();
4832 EXPORT_SYMBOL(__cond_resched_softirq);
4835 * yield - yield the current processor to other threads.
4837 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4839 * The scheduler is at all times free to pick the calling task as the most
4840 * eligible task to run, if removing the yield() call from your code breaks
4841 * it, its already broken.
4843 * Typical broken usage is:
4848 * where one assumes that yield() will let 'the other' process run that will
4849 * make event true. If the current task is a SCHED_FIFO task that will never
4850 * happen. Never use yield() as a progress guarantee!!
4852 * If you want to use yield() to wait for something, use wait_event().
4853 * If you want to use yield() to be 'nice' for others, use cond_resched().
4854 * If you still want to use yield(), do not!
4856 void __sched yield(void)
4858 set_current_state(TASK_RUNNING);
4861 EXPORT_SYMBOL(yield);
4864 * yield_to - yield the current processor to another thread in
4865 * your thread group, or accelerate that thread toward the
4866 * processor it's on.
4868 * @preempt: whether task preemption is allowed or not
4870 * It's the caller's job to ensure that the target task struct
4871 * can't go away on us before we can do any checks.
4874 * true (>0) if we indeed boosted the target task.
4875 * false (0) if we failed to boost the target.
4876 * -ESRCH if there's no task to yield to.
4878 int __sched yield_to(struct task_struct *p, bool preempt)
4880 struct task_struct *curr = current;
4881 struct rq *rq, *p_rq;
4882 unsigned long flags;
4885 local_irq_save(flags);
4891 * If we're the only runnable task on the rq and target rq also
4892 * has only one task, there's absolutely no point in yielding.
4894 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4899 double_rq_lock(rq, p_rq);
4900 if (task_rq(p) != p_rq) {
4901 double_rq_unlock(rq, p_rq);
4905 if (!curr->sched_class->yield_to_task)
4908 if (curr->sched_class != p->sched_class)
4911 if (task_running(p_rq, p) || p->state)
4914 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4916 schedstat_inc(rq, yld_count);
4918 * Make p's CPU reschedule; pick_next_entity takes care of
4921 if (preempt && rq != p_rq)
4926 double_rq_unlock(rq, p_rq);
4928 local_irq_restore(flags);
4935 EXPORT_SYMBOL_GPL(yield_to);
4938 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4939 * that process accounting knows that this is a task in IO wait state.
4941 long __sched io_schedule_timeout(long timeout)
4943 int old_iowait = current->in_iowait;
4947 current->in_iowait = 1;
4948 blk_schedule_flush_plug(current);
4950 delayacct_blkio_start();
4952 atomic_inc(&rq->nr_iowait);
4953 ret = schedule_timeout(timeout);
4954 current->in_iowait = old_iowait;
4955 atomic_dec(&rq->nr_iowait);
4956 delayacct_blkio_end();
4960 EXPORT_SYMBOL(io_schedule_timeout);
4963 * sys_sched_get_priority_max - return maximum RT priority.
4964 * @policy: scheduling class.
4966 * Return: On success, this syscall returns the maximum
4967 * rt_priority that can be used by a given scheduling class.
4968 * On failure, a negative error code is returned.
4970 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4977 ret = MAX_USER_RT_PRIO-1;
4979 case SCHED_DEADLINE:
4990 * sys_sched_get_priority_min - return minimum RT priority.
4991 * @policy: scheduling class.
4993 * Return: On success, this syscall returns the minimum
4994 * rt_priority that can be used by a given scheduling class.
4995 * On failure, a negative error code is returned.
4997 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5006 case SCHED_DEADLINE:
5016 * sys_sched_rr_get_interval - return the default timeslice of a process.
5017 * @pid: pid of the process.
5018 * @interval: userspace pointer to the timeslice value.
5020 * this syscall writes the default timeslice value of a given process
5021 * into the user-space timespec buffer. A value of '0' means infinity.
5023 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5026 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5027 struct timespec __user *, interval)
5029 struct task_struct *p;
5030 unsigned int time_slice;
5031 unsigned long flags;
5041 p = find_process_by_pid(pid);
5045 retval = security_task_getscheduler(p);
5049 rq = task_rq_lock(p, &flags);
5051 if (p->sched_class->get_rr_interval)
5052 time_slice = p->sched_class->get_rr_interval(rq, p);
5053 task_rq_unlock(rq, p, &flags);
5056 jiffies_to_timespec(time_slice, &t);
5057 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5065 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5067 void sched_show_task(struct task_struct *p)
5069 unsigned long free = 0;
5071 unsigned long state = p->state;
5074 state = __ffs(state) + 1;
5075 printk(KERN_INFO "%-15.15s %c", p->comm,
5076 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5077 #if BITS_PER_LONG == 32
5078 if (state == TASK_RUNNING)
5079 printk(KERN_CONT " running ");
5081 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5083 if (state == TASK_RUNNING)
5084 printk(KERN_CONT " running task ");
5086 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5088 #ifdef CONFIG_DEBUG_STACK_USAGE
5089 free = stack_not_used(p);
5094 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5096 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5097 task_pid_nr(p), ppid,
5098 (unsigned long)task_thread_info(p)->flags);
5100 print_worker_info(KERN_INFO, p);
5101 show_stack(p, NULL);
5104 void show_state_filter(unsigned long state_filter)
5106 struct task_struct *g, *p;
5108 #if BITS_PER_LONG == 32
5110 " task PC stack pid father\n");
5113 " task PC stack pid father\n");
5116 for_each_process_thread(g, p) {
5118 * reset the NMI-timeout, listing all files on a slow
5119 * console might take a lot of time:
5120 * Also, reset softlockup watchdogs on all CPUs, because
5121 * another CPU might be blocked waiting for us to process
5124 touch_nmi_watchdog();
5125 touch_all_softlockup_watchdogs();
5126 if (!state_filter || (p->state & state_filter))
5130 #ifdef CONFIG_SCHED_DEBUG
5131 sysrq_sched_debug_show();
5135 * Only show locks if all tasks are dumped:
5138 debug_show_all_locks();
5141 void init_idle_bootup_task(struct task_struct *idle)
5143 idle->sched_class = &idle_sched_class;
5147 * init_idle - set up an idle thread for a given CPU
5148 * @idle: task in question
5149 * @cpu: cpu the idle task belongs to
5151 * NOTE: this function does not set the idle thread's NEED_RESCHED
5152 * flag, to make booting more robust.
5154 void init_idle(struct task_struct *idle, int cpu)
5156 struct rq *rq = cpu_rq(cpu);
5157 unsigned long flags;
5159 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5160 raw_spin_lock(&rq->lock);
5162 __sched_fork(0, idle);
5164 idle->state = TASK_RUNNING;
5165 idle->se.exec_start = sched_clock();
5169 * Its possible that init_idle() gets called multiple times on a task,
5170 * in that case do_set_cpus_allowed() will not do the right thing.
5172 * And since this is boot we can forgo the serialization.
5174 set_cpus_allowed_common(idle, cpumask_of(cpu));
5177 * We're having a chicken and egg problem, even though we are
5178 * holding rq->lock, the cpu isn't yet set to this cpu so the
5179 * lockdep check in task_group() will fail.
5181 * Similar case to sched_fork(). / Alternatively we could
5182 * use task_rq_lock() here and obtain the other rq->lock.
5187 __set_task_cpu(idle, cpu);
5190 rq->curr = rq->idle = idle;
5191 idle->on_rq = TASK_ON_RQ_QUEUED;
5195 raw_spin_unlock(&rq->lock);
5196 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5198 /* Set the preempt count _outside_ the spinlocks! */
5199 init_idle_preempt_count(idle, cpu);
5202 * The idle tasks have their own, simple scheduling class:
5204 idle->sched_class = &idle_sched_class;
5205 ftrace_graph_init_idle_task(idle, cpu);
5206 vtime_init_idle(idle, cpu);
5208 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5212 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5213 const struct cpumask *trial)
5215 int ret = 1, trial_cpus;
5216 struct dl_bw *cur_dl_b;
5217 unsigned long flags;
5219 if (!cpumask_weight(cur))
5222 rcu_read_lock_sched();
5223 cur_dl_b = dl_bw_of(cpumask_any(cur));
5224 trial_cpus = cpumask_weight(trial);
5226 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5227 if (cur_dl_b->bw != -1 &&
5228 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5230 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5231 rcu_read_unlock_sched();
5236 int task_can_attach(struct task_struct *p,
5237 const struct cpumask *cs_cpus_allowed)
5242 * Kthreads which disallow setaffinity shouldn't be moved
5243 * to a new cpuset; we don't want to change their cpu
5244 * affinity and isolating such threads by their set of
5245 * allowed nodes is unnecessary. Thus, cpusets are not
5246 * applicable for such threads. This prevents checking for
5247 * success of set_cpus_allowed_ptr() on all attached tasks
5248 * before cpus_allowed may be changed.
5250 if (p->flags & PF_NO_SETAFFINITY) {
5256 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5258 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5263 unsigned long flags;
5265 rcu_read_lock_sched();
5266 dl_b = dl_bw_of(dest_cpu);
5267 raw_spin_lock_irqsave(&dl_b->lock, flags);
5268 cpus = dl_bw_cpus(dest_cpu);
5269 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5274 * We reserve space for this task in the destination
5275 * root_domain, as we can't fail after this point.
5276 * We will free resources in the source root_domain
5277 * later on (see set_cpus_allowed_dl()).
5279 __dl_add(dl_b, p->dl.dl_bw);
5281 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5282 rcu_read_unlock_sched();
5292 #ifdef CONFIG_NUMA_BALANCING
5293 /* Migrate current task p to target_cpu */
5294 int migrate_task_to(struct task_struct *p, int target_cpu)
5296 struct migration_arg arg = { p, target_cpu };
5297 int curr_cpu = task_cpu(p);
5299 if (curr_cpu == target_cpu)
5302 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5305 /* TODO: This is not properly updating schedstats */
5307 trace_sched_move_numa(p, curr_cpu, target_cpu);
5308 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5312 * Requeue a task on a given node and accurately track the number of NUMA
5313 * tasks on the runqueues
5315 void sched_setnuma(struct task_struct *p, int nid)
5318 unsigned long flags;
5319 bool queued, running;
5321 rq = task_rq_lock(p, &flags);
5322 queued = task_on_rq_queued(p);
5323 running = task_current(rq, p);
5326 dequeue_task(rq, p, DEQUEUE_SAVE);
5328 put_prev_task(rq, p);
5330 p->numa_preferred_nid = nid;
5333 p->sched_class->set_curr_task(rq);
5335 enqueue_task(rq, p, ENQUEUE_RESTORE);
5336 task_rq_unlock(rq, p, &flags);
5338 #endif /* CONFIG_NUMA_BALANCING */
5340 #ifdef CONFIG_HOTPLUG_CPU
5342 * Ensures that the idle task is using init_mm right before its cpu goes
5345 void idle_task_exit(void)
5347 struct mm_struct *mm = current->active_mm;
5349 BUG_ON(cpu_online(smp_processor_id()));
5351 if (mm != &init_mm) {
5352 switch_mm(mm, &init_mm, current);
5353 finish_arch_post_lock_switch();
5359 * Since this CPU is going 'away' for a while, fold any nr_active delta
5360 * we might have. Assumes we're called after migrate_tasks() so that the
5361 * nr_active count is stable.
5363 * Also see the comment "Global load-average calculations".
5365 static void calc_load_migrate(struct rq *rq)
5367 long delta = calc_load_fold_active(rq);
5369 atomic_long_add(delta, &calc_load_tasks);
5372 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5376 static const struct sched_class fake_sched_class = {
5377 .put_prev_task = put_prev_task_fake,
5380 static struct task_struct fake_task = {
5382 * Avoid pull_{rt,dl}_task()
5384 .prio = MAX_PRIO + 1,
5385 .sched_class = &fake_sched_class,
5389 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5390 * try_to_wake_up()->select_task_rq().
5392 * Called with rq->lock held even though we'er in stop_machine() and
5393 * there's no concurrency possible, we hold the required locks anyway
5394 * because of lock validation efforts.
5396 static void migrate_tasks(struct rq *dead_rq)
5398 struct rq *rq = dead_rq;
5399 struct task_struct *next, *stop = rq->stop;
5403 * Fudge the rq selection such that the below task selection loop
5404 * doesn't get stuck on the currently eligible stop task.
5406 * We're currently inside stop_machine() and the rq is either stuck
5407 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5408 * either way we should never end up calling schedule() until we're
5414 * put_prev_task() and pick_next_task() sched
5415 * class method both need to have an up-to-date
5416 * value of rq->clock[_task]
5418 update_rq_clock(rq);
5422 * There's this thread running, bail when that's the only
5425 if (rq->nr_running == 1)
5429 * pick_next_task assumes pinned rq->lock.
5431 lockdep_pin_lock(&rq->lock);
5432 next = pick_next_task(rq, &fake_task);
5434 next->sched_class->put_prev_task(rq, next);
5437 * Rules for changing task_struct::cpus_allowed are holding
5438 * both pi_lock and rq->lock, such that holding either
5439 * stabilizes the mask.
5441 * Drop rq->lock is not quite as disastrous as it usually is
5442 * because !cpu_active at this point, which means load-balance
5443 * will not interfere. Also, stop-machine.
5445 lockdep_unpin_lock(&rq->lock);
5446 raw_spin_unlock(&rq->lock);
5447 raw_spin_lock(&next->pi_lock);
5448 raw_spin_lock(&rq->lock);
5451 * Since we're inside stop-machine, _nothing_ should have
5452 * changed the task, WARN if weird stuff happened, because in
5453 * that case the above rq->lock drop is a fail too.
5455 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5456 raw_spin_unlock(&next->pi_lock);
5460 /* Find suitable destination for @next, with force if needed. */
5461 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5463 rq = __migrate_task(rq, next, dest_cpu);
5464 if (rq != dead_rq) {
5465 raw_spin_unlock(&rq->lock);
5467 raw_spin_lock(&rq->lock);
5469 raw_spin_unlock(&next->pi_lock);
5474 #endif /* CONFIG_HOTPLUG_CPU */
5476 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5478 static struct ctl_table sd_ctl_dir[] = {
5480 .procname = "sched_domain",
5486 static struct ctl_table sd_ctl_root[] = {
5488 .procname = "kernel",
5490 .child = sd_ctl_dir,
5495 static struct ctl_table *sd_alloc_ctl_entry(int n)
5497 struct ctl_table *entry =
5498 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5503 static void sd_free_ctl_entry(struct ctl_table **tablep)
5505 struct ctl_table *entry;
5508 * In the intermediate directories, both the child directory and
5509 * procname are dynamically allocated and could fail but the mode
5510 * will always be set. In the lowest directory the names are
5511 * static strings and all have proc handlers.
5513 for (entry = *tablep; entry->mode; entry++) {
5515 sd_free_ctl_entry(&entry->child);
5516 if (entry->proc_handler == NULL)
5517 kfree(entry->procname);
5524 static int min_load_idx = 0;
5525 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5528 set_table_entry(struct ctl_table *entry,
5529 const char *procname, void *data, int maxlen,
5530 umode_t mode, proc_handler *proc_handler,
5533 entry->procname = procname;
5535 entry->maxlen = maxlen;
5537 entry->proc_handler = proc_handler;
5540 entry->extra1 = &min_load_idx;
5541 entry->extra2 = &max_load_idx;
5545 static struct ctl_table *
5546 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5548 struct ctl_table *table = sd_alloc_ctl_entry(5);
5553 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5554 sizeof(int), 0644, proc_dointvec_minmax, false);
5555 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5556 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5557 proc_doulongvec_minmax, false);
5558 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5559 sizeof(int), 0644, proc_dointvec_minmax, false);
5560 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5561 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5562 proc_doulongvec_minmax, false);
5567 static struct ctl_table *
5568 sd_alloc_ctl_group_table(struct sched_group *sg)
5570 struct ctl_table *table = sd_alloc_ctl_entry(2);
5575 table->procname = kstrdup("energy", GFP_KERNEL);
5577 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5582 static struct ctl_table *
5583 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5585 struct ctl_table *table;
5586 unsigned int nr_entries = 14;
5589 struct sched_group *sg = sd->groups;
5594 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5596 nr_entries += nr_sgs;
5599 table = sd_alloc_ctl_entry(nr_entries);
5604 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5605 sizeof(long), 0644, proc_doulongvec_minmax, false);
5606 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5607 sizeof(long), 0644, proc_doulongvec_minmax, false);
5608 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5609 sizeof(int), 0644, proc_dointvec_minmax, true);
5610 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5611 sizeof(int), 0644, proc_dointvec_minmax, true);
5612 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5613 sizeof(int), 0644, proc_dointvec_minmax, true);
5614 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5615 sizeof(int), 0644, proc_dointvec_minmax, true);
5616 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5617 sizeof(int), 0644, proc_dointvec_minmax, true);
5618 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5619 sizeof(int), 0644, proc_dointvec_minmax, false);
5620 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5621 sizeof(int), 0644, proc_dointvec_minmax, false);
5622 set_table_entry(&table[9], "cache_nice_tries",
5623 &sd->cache_nice_tries,
5624 sizeof(int), 0644, proc_dointvec_minmax, false);
5625 set_table_entry(&table[10], "flags", &sd->flags,
5626 sizeof(int), 0644, proc_dointvec_minmax, false);
5627 set_table_entry(&table[11], "max_newidle_lb_cost",
5628 &sd->max_newidle_lb_cost,
5629 sizeof(long), 0644, proc_doulongvec_minmax, false);
5630 set_table_entry(&table[12], "name", sd->name,
5631 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5635 struct ctl_table *entry = &table[13];
5638 snprintf(buf, 32, "group%d", i);
5639 entry->procname = kstrdup(buf, GFP_KERNEL);
5641 entry->child = sd_alloc_ctl_group_table(sg);
5642 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5644 /* &table[nr_entries-1] is terminator */
5649 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5651 struct ctl_table *entry, *table;
5652 struct sched_domain *sd;
5653 int domain_num = 0, i;
5656 for_each_domain(cpu, sd)
5658 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5663 for_each_domain(cpu, sd) {
5664 snprintf(buf, 32, "domain%d", i);
5665 entry->procname = kstrdup(buf, GFP_KERNEL);
5667 entry->child = sd_alloc_ctl_domain_table(sd);
5674 static struct ctl_table_header *sd_sysctl_header;
5675 static void register_sched_domain_sysctl(void)
5677 int i, cpu_num = num_possible_cpus();
5678 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5681 WARN_ON(sd_ctl_dir[0].child);
5682 sd_ctl_dir[0].child = entry;
5687 for_each_possible_cpu(i) {
5688 snprintf(buf, 32, "cpu%d", i);
5689 entry->procname = kstrdup(buf, GFP_KERNEL);
5691 entry->child = sd_alloc_ctl_cpu_table(i);
5695 WARN_ON(sd_sysctl_header);
5696 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5699 /* may be called multiple times per register */
5700 static void unregister_sched_domain_sysctl(void)
5702 unregister_sysctl_table(sd_sysctl_header);
5703 sd_sysctl_header = NULL;
5704 if (sd_ctl_dir[0].child)
5705 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5708 static void register_sched_domain_sysctl(void)
5711 static void unregister_sched_domain_sysctl(void)
5714 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5716 static void set_rq_online(struct rq *rq)
5719 const struct sched_class *class;
5721 cpumask_set_cpu(rq->cpu, rq->rd->online);
5724 for_each_class(class) {
5725 if (class->rq_online)
5726 class->rq_online(rq);
5731 static void set_rq_offline(struct rq *rq)
5734 const struct sched_class *class;
5736 for_each_class(class) {
5737 if (class->rq_offline)
5738 class->rq_offline(rq);
5741 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5747 * migration_call - callback that gets triggered when a CPU is added.
5748 * Here we can start up the necessary migration thread for the new CPU.
5751 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5753 int cpu = (long)hcpu;
5754 unsigned long flags;
5755 struct rq *rq = cpu_rq(cpu);
5757 switch (action & ~CPU_TASKS_FROZEN) {
5759 case CPU_UP_PREPARE:
5760 raw_spin_lock_irqsave(&rq->lock, flags);
5761 walt_set_window_start(rq);
5762 raw_spin_unlock_irqrestore(&rq->lock, flags);
5763 rq->calc_load_update = calc_load_update;
5764 account_reset_rq(rq);
5768 /* Update our root-domain */
5769 raw_spin_lock_irqsave(&rq->lock, flags);
5771 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5775 raw_spin_unlock_irqrestore(&rq->lock, flags);
5778 #ifdef CONFIG_HOTPLUG_CPU
5780 sched_ttwu_pending();
5781 /* Update our root-domain */
5782 raw_spin_lock_irqsave(&rq->lock, flags);
5783 walt_migrate_sync_cpu(cpu);
5785 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5789 BUG_ON(rq->nr_running != 1); /* the migration thread */
5790 raw_spin_unlock_irqrestore(&rq->lock, flags);
5794 calc_load_migrate(rq);
5799 update_max_interval();
5805 * Register at high priority so that task migration (migrate_all_tasks)
5806 * happens before everything else. This has to be lower priority than
5807 * the notifier in the perf_event subsystem, though.
5809 static struct notifier_block migration_notifier = {
5810 .notifier_call = migration_call,
5811 .priority = CPU_PRI_MIGRATION,
5814 static void set_cpu_rq_start_time(void)
5816 int cpu = smp_processor_id();
5817 struct rq *rq = cpu_rq(cpu);
5818 rq->age_stamp = sched_clock_cpu(cpu);
5821 static int sched_cpu_active(struct notifier_block *nfb,
5822 unsigned long action, void *hcpu)
5824 int cpu = (long)hcpu;
5826 switch (action & ~CPU_TASKS_FROZEN) {
5828 set_cpu_rq_start_time();
5833 * At this point a starting CPU has marked itself as online via
5834 * set_cpu_online(). But it might not yet have marked itself
5835 * as active, which is essential from here on.
5837 set_cpu_active(cpu, true);
5838 stop_machine_unpark(cpu);
5841 case CPU_DOWN_FAILED:
5842 set_cpu_active(cpu, true);
5850 static int sched_cpu_inactive(struct notifier_block *nfb,
5851 unsigned long action, void *hcpu)
5853 switch (action & ~CPU_TASKS_FROZEN) {
5854 case CPU_DOWN_PREPARE:
5855 set_cpu_active((long)hcpu, false);
5862 static int __init migration_init(void)
5864 void *cpu = (void *)(long)smp_processor_id();
5867 /* Initialize migration for the boot CPU */
5868 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5869 BUG_ON(err == NOTIFY_BAD);
5870 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5871 register_cpu_notifier(&migration_notifier);
5873 /* Register cpu active notifiers */
5874 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5875 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5879 early_initcall(migration_init);
5881 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5883 #ifdef CONFIG_SCHED_DEBUG
5885 static __read_mostly int sched_debug_enabled;
5887 static int __init sched_debug_setup(char *str)
5889 sched_debug_enabled = 1;
5893 early_param("sched_debug", sched_debug_setup);
5895 static inline bool sched_debug(void)
5897 return sched_debug_enabled;
5900 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5901 struct cpumask *groupmask)
5903 struct sched_group *group = sd->groups;
5905 cpumask_clear(groupmask);
5907 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5909 if (!(sd->flags & SD_LOAD_BALANCE)) {
5910 printk("does not load-balance\n");
5912 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5917 printk(KERN_CONT "span %*pbl level %s\n",
5918 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5920 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5921 printk(KERN_ERR "ERROR: domain->span does not contain "
5924 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5925 printk(KERN_ERR "ERROR: domain->groups does not contain"
5929 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5933 printk(KERN_ERR "ERROR: group is NULL\n");
5937 if (!cpumask_weight(sched_group_cpus(group))) {
5938 printk(KERN_CONT "\n");
5939 printk(KERN_ERR "ERROR: empty group\n");
5943 if (!(sd->flags & SD_OVERLAP) &&
5944 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5945 printk(KERN_CONT "\n");
5946 printk(KERN_ERR "ERROR: repeated CPUs\n");
5950 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5952 printk(KERN_CONT " %*pbl",
5953 cpumask_pr_args(sched_group_cpus(group)));
5954 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5955 printk(KERN_CONT " (cpu_capacity = %lu)",
5956 group->sgc->capacity);
5959 group = group->next;
5960 } while (group != sd->groups);
5961 printk(KERN_CONT "\n");
5963 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5964 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5967 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5968 printk(KERN_ERR "ERROR: parent span is not a superset "
5969 "of domain->span\n");
5973 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5977 if (!sched_debug_enabled)
5981 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5985 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5988 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5996 #else /* !CONFIG_SCHED_DEBUG */
5997 # define sched_domain_debug(sd, cpu) do { } while (0)
5998 static inline bool sched_debug(void)
6002 #endif /* CONFIG_SCHED_DEBUG */
6004 static int sd_degenerate(struct sched_domain *sd)
6006 if (cpumask_weight(sched_domain_span(sd)) == 1)
6009 /* Following flags need at least 2 groups */
6010 if (sd->flags & (SD_LOAD_BALANCE |
6011 SD_BALANCE_NEWIDLE |
6014 SD_SHARE_CPUCAPACITY |
6015 SD_SHARE_PKG_RESOURCES |
6016 SD_SHARE_POWERDOMAIN |
6017 SD_SHARE_CAP_STATES)) {
6018 if (sd->groups != sd->groups->next)
6022 /* Following flags don't use groups */
6023 if (sd->flags & (SD_WAKE_AFFINE))
6030 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6032 unsigned long cflags = sd->flags, pflags = parent->flags;
6034 if (sd_degenerate(parent))
6037 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6040 /* Flags needing groups don't count if only 1 group in parent */
6041 if (parent->groups == parent->groups->next) {
6042 pflags &= ~(SD_LOAD_BALANCE |
6043 SD_BALANCE_NEWIDLE |
6046 SD_SHARE_CPUCAPACITY |
6047 SD_SHARE_PKG_RESOURCES |
6049 SD_SHARE_POWERDOMAIN |
6050 SD_SHARE_CAP_STATES);
6051 if (nr_node_ids == 1)
6052 pflags &= ~SD_SERIALIZE;
6054 if (~cflags & pflags)
6060 static void free_rootdomain(struct rcu_head *rcu)
6062 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6064 cpupri_cleanup(&rd->cpupri);
6065 cpudl_cleanup(&rd->cpudl);
6066 free_cpumask_var(rd->dlo_mask);
6067 free_cpumask_var(rd->rto_mask);
6068 free_cpumask_var(rd->online);
6069 free_cpumask_var(rd->span);
6073 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6075 struct root_domain *old_rd = NULL;
6076 unsigned long flags;
6078 raw_spin_lock_irqsave(&rq->lock, flags);
6083 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6086 cpumask_clear_cpu(rq->cpu, old_rd->span);
6089 * If we dont want to free the old_rd yet then
6090 * set old_rd to NULL to skip the freeing later
6093 if (!atomic_dec_and_test(&old_rd->refcount))
6097 atomic_inc(&rd->refcount);
6100 cpumask_set_cpu(rq->cpu, rd->span);
6101 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6104 raw_spin_unlock_irqrestore(&rq->lock, flags);
6107 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6110 static int init_rootdomain(struct root_domain *rd)
6112 memset(rd, 0, sizeof(*rd));
6114 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6116 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6118 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6120 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6123 init_dl_bw(&rd->dl_bw);
6124 if (cpudl_init(&rd->cpudl) != 0)
6127 if (cpupri_init(&rd->cpupri) != 0)
6130 init_max_cpu_capacity(&rd->max_cpu_capacity);
6134 free_cpumask_var(rd->rto_mask);
6136 free_cpumask_var(rd->dlo_mask);
6138 free_cpumask_var(rd->online);
6140 free_cpumask_var(rd->span);
6146 * By default the system creates a single root-domain with all cpus as
6147 * members (mimicking the global state we have today).
6149 struct root_domain def_root_domain;
6151 static void init_defrootdomain(void)
6153 init_rootdomain(&def_root_domain);
6155 atomic_set(&def_root_domain.refcount, 1);
6158 static struct root_domain *alloc_rootdomain(void)
6160 struct root_domain *rd;
6162 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6166 if (init_rootdomain(rd) != 0) {
6174 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6176 struct sched_group *tmp, *first;
6185 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6190 } while (sg != first);
6193 static void free_sched_domain(struct rcu_head *rcu)
6195 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6198 * If its an overlapping domain it has private groups, iterate and
6201 if (sd->flags & SD_OVERLAP) {
6202 free_sched_groups(sd->groups, 1);
6203 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6204 kfree(sd->groups->sgc);
6210 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6212 call_rcu(&sd->rcu, free_sched_domain);
6215 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6217 for (; sd; sd = sd->parent)
6218 destroy_sched_domain(sd, cpu);
6222 * Keep a special pointer to the highest sched_domain that has
6223 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6224 * allows us to avoid some pointer chasing select_idle_sibling().
6226 * Also keep a unique ID per domain (we use the first cpu number in
6227 * the cpumask of the domain), this allows us to quickly tell if
6228 * two cpus are in the same cache domain, see cpus_share_cache().
6230 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6231 DEFINE_PER_CPU(int, sd_llc_size);
6232 DEFINE_PER_CPU(int, sd_llc_id);
6233 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6234 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6235 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6236 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6237 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6239 static void update_top_cache_domain(int cpu)
6241 struct sched_domain *sd;
6242 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6246 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6248 id = cpumask_first(sched_domain_span(sd));
6249 size = cpumask_weight(sched_domain_span(sd));
6250 busy_sd = sd->parent; /* sd_busy */
6252 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6254 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6255 per_cpu(sd_llc_size, cpu) = size;
6256 per_cpu(sd_llc_id, cpu) = id;
6258 sd = lowest_flag_domain(cpu, SD_NUMA);
6259 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6261 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6262 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6264 for_each_domain(cpu, sd) {
6265 if (sd->groups->sge)
6270 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6272 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6273 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6277 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6278 * hold the hotplug lock.
6281 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6283 struct rq *rq = cpu_rq(cpu);
6284 struct sched_domain *tmp;
6286 /* Remove the sched domains which do not contribute to scheduling. */
6287 for (tmp = sd; tmp; ) {
6288 struct sched_domain *parent = tmp->parent;
6292 if (sd_parent_degenerate(tmp, parent)) {
6293 tmp->parent = parent->parent;
6295 parent->parent->child = tmp;
6297 * Transfer SD_PREFER_SIBLING down in case of a
6298 * degenerate parent; the spans match for this
6299 * so the property transfers.
6301 if (parent->flags & SD_PREFER_SIBLING)
6302 tmp->flags |= SD_PREFER_SIBLING;
6303 destroy_sched_domain(parent, cpu);
6308 if (sd && sd_degenerate(sd)) {
6311 destroy_sched_domain(tmp, cpu);
6316 sched_domain_debug(sd, cpu);
6318 rq_attach_root(rq, rd);
6320 rcu_assign_pointer(rq->sd, sd);
6321 destroy_sched_domains(tmp, cpu);
6323 update_top_cache_domain(cpu);
6326 /* Setup the mask of cpus configured for isolated domains */
6327 static int __init isolated_cpu_setup(char *str)
6329 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6330 cpulist_parse(str, cpu_isolated_map);
6334 __setup("isolcpus=", isolated_cpu_setup);
6337 struct sched_domain ** __percpu sd;
6338 struct root_domain *rd;
6349 * Build an iteration mask that can exclude certain CPUs from the upwards
6352 * Asymmetric node setups can result in situations where the domain tree is of
6353 * unequal depth, make sure to skip domains that already cover the entire
6356 * In that case build_sched_domains() will have terminated the iteration early
6357 * and our sibling sd spans will be empty. Domains should always include the
6358 * cpu they're built on, so check that.
6361 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6363 const struct cpumask *span = sched_domain_span(sd);
6364 struct sd_data *sdd = sd->private;
6365 struct sched_domain *sibling;
6368 for_each_cpu(i, span) {
6369 sibling = *per_cpu_ptr(sdd->sd, i);
6370 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6373 cpumask_set_cpu(i, sched_group_mask(sg));
6378 * Return the canonical balance cpu for this group, this is the first cpu
6379 * of this group that's also in the iteration mask.
6381 int group_balance_cpu(struct sched_group *sg)
6383 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6387 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6389 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6390 const struct cpumask *span = sched_domain_span(sd);
6391 struct cpumask *covered = sched_domains_tmpmask;
6392 struct sd_data *sdd = sd->private;
6393 struct sched_domain *sibling;
6396 cpumask_clear(covered);
6398 for_each_cpu(i, span) {
6399 struct cpumask *sg_span;
6401 if (cpumask_test_cpu(i, covered))
6404 sibling = *per_cpu_ptr(sdd->sd, i);
6406 /* See the comment near build_group_mask(). */
6407 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6410 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6411 GFP_KERNEL, cpu_to_node(cpu));
6416 sg_span = sched_group_cpus(sg);
6418 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6420 cpumask_set_cpu(i, sg_span);
6422 cpumask_or(covered, covered, sg_span);
6424 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6425 if (atomic_inc_return(&sg->sgc->ref) == 1)
6426 build_group_mask(sd, sg);
6429 * Initialize sgc->capacity such that even if we mess up the
6430 * domains and no possible iteration will get us here, we won't
6433 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6434 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6437 * Make sure the first group of this domain contains the
6438 * canonical balance cpu. Otherwise the sched_domain iteration
6439 * breaks. See update_sg_lb_stats().
6441 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6442 group_balance_cpu(sg) == cpu)
6452 sd->groups = groups;
6457 free_sched_groups(first, 0);
6462 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6464 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6465 struct sched_domain *child = sd->child;
6468 cpu = cpumask_first(sched_domain_span(child));
6471 *sg = *per_cpu_ptr(sdd->sg, cpu);
6472 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6473 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6480 * build_sched_groups will build a circular linked list of the groups
6481 * covered by the given span, and will set each group's ->cpumask correctly,
6482 * and ->cpu_capacity to 0.
6484 * Assumes the sched_domain tree is fully constructed
6487 build_sched_groups(struct sched_domain *sd, int cpu)
6489 struct sched_group *first = NULL, *last = NULL;
6490 struct sd_data *sdd = sd->private;
6491 const struct cpumask *span = sched_domain_span(sd);
6492 struct cpumask *covered;
6495 get_group(cpu, sdd, &sd->groups);
6496 atomic_inc(&sd->groups->ref);
6498 if (cpu != cpumask_first(span))
6501 lockdep_assert_held(&sched_domains_mutex);
6502 covered = sched_domains_tmpmask;
6504 cpumask_clear(covered);
6506 for_each_cpu(i, span) {
6507 struct sched_group *sg;
6510 if (cpumask_test_cpu(i, covered))
6513 group = get_group(i, sdd, &sg);
6514 cpumask_setall(sched_group_mask(sg));
6516 for_each_cpu(j, span) {
6517 if (get_group(j, sdd, NULL) != group)
6520 cpumask_set_cpu(j, covered);
6521 cpumask_set_cpu(j, sched_group_cpus(sg));
6536 * Initialize sched groups cpu_capacity.
6538 * cpu_capacity indicates the capacity of sched group, which is used while
6539 * distributing the load between different sched groups in a sched domain.
6540 * Typically cpu_capacity for all the groups in a sched domain will be same
6541 * unless there are asymmetries in the topology. If there are asymmetries,
6542 * group having more cpu_capacity will pickup more load compared to the
6543 * group having less cpu_capacity.
6545 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6547 struct sched_group *sg = sd->groups;
6552 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6554 } while (sg != sd->groups);
6556 if (cpu != group_balance_cpu(sg))
6559 update_group_capacity(sd, cpu);
6560 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6564 * Check that the per-cpu provided sd energy data is consistent for all cpus
6567 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6568 const struct cpumask *cpumask)
6570 const struct sched_group_energy * const sge = fn(cpu);
6571 struct cpumask mask;
6574 if (cpumask_weight(cpumask) <= 1)
6577 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6579 for_each_cpu(i, &mask) {
6580 const struct sched_group_energy * const e = fn(i);
6583 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6585 for (y = 0; y < (e->nr_idle_states); y++) {
6586 BUG_ON(e->idle_states[y].power !=
6587 sge->idle_states[y].power);
6590 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6592 for (y = 0; y < (e->nr_cap_states); y++) {
6593 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6594 BUG_ON(e->cap_states[y].power !=
6595 sge->cap_states[y].power);
6600 static void init_sched_energy(int cpu, struct sched_domain *sd,
6601 sched_domain_energy_f fn)
6603 if (!(fn && fn(cpu)))
6606 if (cpu != group_balance_cpu(sd->groups))
6609 if (sd->child && !sd->child->groups->sge) {
6610 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6611 #ifdef CONFIG_SCHED_DEBUG
6612 pr_err(" energy data on %s but not on %s domain\n",
6613 sd->name, sd->child->name);
6618 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6620 sd->groups->sge = fn(cpu);
6624 * Initializers for schedule domains
6625 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6628 static int default_relax_domain_level = -1;
6629 int sched_domain_level_max;
6631 static int __init setup_relax_domain_level(char *str)
6633 if (kstrtoint(str, 0, &default_relax_domain_level))
6634 pr_warn("Unable to set relax_domain_level\n");
6638 __setup("relax_domain_level=", setup_relax_domain_level);
6640 static void set_domain_attribute(struct sched_domain *sd,
6641 struct sched_domain_attr *attr)
6645 if (!attr || attr->relax_domain_level < 0) {
6646 if (default_relax_domain_level < 0)
6649 request = default_relax_domain_level;
6651 request = attr->relax_domain_level;
6652 if (request < sd->level) {
6653 /* turn off idle balance on this domain */
6654 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6656 /* turn on idle balance on this domain */
6657 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6661 static void __sdt_free(const struct cpumask *cpu_map);
6662 static int __sdt_alloc(const struct cpumask *cpu_map);
6664 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6665 const struct cpumask *cpu_map)
6669 if (!atomic_read(&d->rd->refcount))
6670 free_rootdomain(&d->rd->rcu); /* fall through */
6672 free_percpu(d->sd); /* fall through */
6674 __sdt_free(cpu_map); /* fall through */
6680 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6681 const struct cpumask *cpu_map)
6683 memset(d, 0, sizeof(*d));
6685 if (__sdt_alloc(cpu_map))
6686 return sa_sd_storage;
6687 d->sd = alloc_percpu(struct sched_domain *);
6689 return sa_sd_storage;
6690 d->rd = alloc_rootdomain();
6693 return sa_rootdomain;
6697 * NULL the sd_data elements we've used to build the sched_domain and
6698 * sched_group structure so that the subsequent __free_domain_allocs()
6699 * will not free the data we're using.
6701 static void claim_allocations(int cpu, struct sched_domain *sd)
6703 struct sd_data *sdd = sd->private;
6705 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6706 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6708 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6709 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6711 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6712 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6716 static int sched_domains_numa_levels;
6717 enum numa_topology_type sched_numa_topology_type;
6718 static int *sched_domains_numa_distance;
6719 int sched_max_numa_distance;
6720 static struct cpumask ***sched_domains_numa_masks;
6721 static int sched_domains_curr_level;
6725 * SD_flags allowed in topology descriptions.
6727 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6728 * SD_SHARE_PKG_RESOURCES - describes shared caches
6729 * SD_NUMA - describes NUMA topologies
6730 * SD_SHARE_POWERDOMAIN - describes shared power domain
6731 * SD_SHARE_CAP_STATES - describes shared capacity states
6734 * SD_ASYM_PACKING - describes SMT quirks
6736 #define TOPOLOGY_SD_FLAGS \
6737 (SD_SHARE_CPUCAPACITY | \
6738 SD_SHARE_PKG_RESOURCES | \
6741 SD_SHARE_POWERDOMAIN | \
6742 SD_SHARE_CAP_STATES)
6744 static struct sched_domain *
6745 sd_init(struct sched_domain_topology_level *tl, int cpu)
6747 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6748 int sd_weight, sd_flags = 0;
6752 * Ugly hack to pass state to sd_numa_mask()...
6754 sched_domains_curr_level = tl->numa_level;
6757 sd_weight = cpumask_weight(tl->mask(cpu));
6760 sd_flags = (*tl->sd_flags)();
6761 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6762 "wrong sd_flags in topology description\n"))
6763 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6765 *sd = (struct sched_domain){
6766 .min_interval = sd_weight,
6767 .max_interval = 2*sd_weight,
6769 .imbalance_pct = 125,
6771 .cache_nice_tries = 0,
6778 .flags = 1*SD_LOAD_BALANCE
6779 | 1*SD_BALANCE_NEWIDLE
6784 | 0*SD_SHARE_CPUCAPACITY
6785 | 0*SD_SHARE_PKG_RESOURCES
6787 | 0*SD_PREFER_SIBLING
6792 .last_balance = jiffies,
6793 .balance_interval = sd_weight,
6795 .max_newidle_lb_cost = 0,
6796 .next_decay_max_lb_cost = jiffies,
6797 #ifdef CONFIG_SCHED_DEBUG
6803 * Convert topological properties into behaviour.
6806 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6807 sd->flags |= SD_PREFER_SIBLING;
6808 sd->imbalance_pct = 110;
6809 sd->smt_gain = 1178; /* ~15% */
6811 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6812 sd->imbalance_pct = 117;
6813 sd->cache_nice_tries = 1;
6817 } else if (sd->flags & SD_NUMA) {
6818 sd->cache_nice_tries = 2;
6822 sd->flags |= SD_SERIALIZE;
6823 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6824 sd->flags &= ~(SD_BALANCE_EXEC |
6831 sd->flags |= SD_PREFER_SIBLING;
6832 sd->cache_nice_tries = 1;
6837 sd->private = &tl->data;
6843 * Topology list, bottom-up.
6845 static struct sched_domain_topology_level default_topology[] = {
6846 #ifdef CONFIG_SCHED_SMT
6847 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6849 #ifdef CONFIG_SCHED_MC
6850 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6852 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6856 static struct sched_domain_topology_level *sched_domain_topology =
6859 #define for_each_sd_topology(tl) \
6860 for (tl = sched_domain_topology; tl->mask; tl++)
6862 void set_sched_topology(struct sched_domain_topology_level *tl)
6864 sched_domain_topology = tl;
6869 static const struct cpumask *sd_numa_mask(int cpu)
6871 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6874 static void sched_numa_warn(const char *str)
6876 static int done = false;
6884 printk(KERN_WARNING "ERROR: %s\n\n", str);
6886 for (i = 0; i < nr_node_ids; i++) {
6887 printk(KERN_WARNING " ");
6888 for (j = 0; j < nr_node_ids; j++)
6889 printk(KERN_CONT "%02d ", node_distance(i,j));
6890 printk(KERN_CONT "\n");
6892 printk(KERN_WARNING "\n");
6895 bool find_numa_distance(int distance)
6899 if (distance == node_distance(0, 0))
6902 for (i = 0; i < sched_domains_numa_levels; i++) {
6903 if (sched_domains_numa_distance[i] == distance)
6911 * A system can have three types of NUMA topology:
6912 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6913 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6914 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6916 * The difference between a glueless mesh topology and a backplane
6917 * topology lies in whether communication between not directly
6918 * connected nodes goes through intermediary nodes (where programs
6919 * could run), or through backplane controllers. This affects
6920 * placement of programs.
6922 * The type of topology can be discerned with the following tests:
6923 * - If the maximum distance between any nodes is 1 hop, the system
6924 * is directly connected.
6925 * - If for two nodes A and B, located N > 1 hops away from each other,
6926 * there is an intermediary node C, which is < N hops away from both
6927 * nodes A and B, the system is a glueless mesh.
6929 static void init_numa_topology_type(void)
6933 n = sched_max_numa_distance;
6935 if (sched_domains_numa_levels <= 1) {
6936 sched_numa_topology_type = NUMA_DIRECT;
6940 for_each_online_node(a) {
6941 for_each_online_node(b) {
6942 /* Find two nodes furthest removed from each other. */
6943 if (node_distance(a, b) < n)
6946 /* Is there an intermediary node between a and b? */
6947 for_each_online_node(c) {
6948 if (node_distance(a, c) < n &&
6949 node_distance(b, c) < n) {
6950 sched_numa_topology_type =
6956 sched_numa_topology_type = NUMA_BACKPLANE;
6962 static void sched_init_numa(void)
6964 int next_distance, curr_distance = node_distance(0, 0);
6965 struct sched_domain_topology_level *tl;
6969 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6970 if (!sched_domains_numa_distance)
6974 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6975 * unique distances in the node_distance() table.
6977 * Assumes node_distance(0,j) includes all distances in
6978 * node_distance(i,j) in order to avoid cubic time.
6980 next_distance = curr_distance;
6981 for (i = 0; i < nr_node_ids; i++) {
6982 for (j = 0; j < nr_node_ids; j++) {
6983 for (k = 0; k < nr_node_ids; k++) {
6984 int distance = node_distance(i, k);
6986 if (distance > curr_distance &&
6987 (distance < next_distance ||
6988 next_distance == curr_distance))
6989 next_distance = distance;
6992 * While not a strong assumption it would be nice to know
6993 * about cases where if node A is connected to B, B is not
6994 * equally connected to A.
6996 if (sched_debug() && node_distance(k, i) != distance)
6997 sched_numa_warn("Node-distance not symmetric");
6999 if (sched_debug() && i && !find_numa_distance(distance))
7000 sched_numa_warn("Node-0 not representative");
7002 if (next_distance != curr_distance) {
7003 sched_domains_numa_distance[level++] = next_distance;
7004 sched_domains_numa_levels = level;
7005 curr_distance = next_distance;
7010 * In case of sched_debug() we verify the above assumption.
7020 * 'level' contains the number of unique distances, excluding the
7021 * identity distance node_distance(i,i).
7023 * The sched_domains_numa_distance[] array includes the actual distance
7028 * Here, we should temporarily reset sched_domains_numa_levels to 0.
7029 * If it fails to allocate memory for array sched_domains_numa_masks[][],
7030 * the array will contain less then 'level' members. This could be
7031 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
7032 * in other functions.
7034 * We reset it to 'level' at the end of this function.
7036 sched_domains_numa_levels = 0;
7038 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
7039 if (!sched_domains_numa_masks)
7043 * Now for each level, construct a mask per node which contains all
7044 * cpus of nodes that are that many hops away from us.
7046 for (i = 0; i < level; i++) {
7047 sched_domains_numa_masks[i] =
7048 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
7049 if (!sched_domains_numa_masks[i])
7052 for (j = 0; j < nr_node_ids; j++) {
7053 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
7057 sched_domains_numa_masks[i][j] = mask;
7060 if (node_distance(j, k) > sched_domains_numa_distance[i])
7063 cpumask_or(mask, mask, cpumask_of_node(k));
7068 /* Compute default topology size */
7069 for (i = 0; sched_domain_topology[i].mask; i++);
7071 tl = kzalloc((i + level + 1) *
7072 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7077 * Copy the default topology bits..
7079 for (i = 0; sched_domain_topology[i].mask; i++)
7080 tl[i] = sched_domain_topology[i];
7083 * .. and append 'j' levels of NUMA goodness.
7085 for (j = 0; j < level; i++, j++) {
7086 tl[i] = (struct sched_domain_topology_level){
7087 .mask = sd_numa_mask,
7088 .sd_flags = cpu_numa_flags,
7089 .flags = SDTL_OVERLAP,
7095 sched_domain_topology = tl;
7097 sched_domains_numa_levels = level;
7098 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7100 init_numa_topology_type();
7103 static void sched_domains_numa_masks_set(int cpu)
7106 int node = cpu_to_node(cpu);
7108 for (i = 0; i < sched_domains_numa_levels; i++) {
7109 for (j = 0; j < nr_node_ids; j++) {
7110 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7111 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7116 static void sched_domains_numa_masks_clear(int cpu)
7119 for (i = 0; i < sched_domains_numa_levels; i++) {
7120 for (j = 0; j < nr_node_ids; j++)
7121 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7126 * Update sched_domains_numa_masks[level][node] array when new cpus
7129 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7130 unsigned long action,
7133 int cpu = (long)hcpu;
7135 switch (action & ~CPU_TASKS_FROZEN) {
7137 sched_domains_numa_masks_set(cpu);
7141 sched_domains_numa_masks_clear(cpu);
7151 static inline void sched_init_numa(void)
7155 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7156 unsigned long action,
7161 #endif /* CONFIG_NUMA */
7163 static int __sdt_alloc(const struct cpumask *cpu_map)
7165 struct sched_domain_topology_level *tl;
7168 for_each_sd_topology(tl) {
7169 struct sd_data *sdd = &tl->data;
7171 sdd->sd = alloc_percpu(struct sched_domain *);
7175 sdd->sg = alloc_percpu(struct sched_group *);
7179 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7183 for_each_cpu(j, cpu_map) {
7184 struct sched_domain *sd;
7185 struct sched_group *sg;
7186 struct sched_group_capacity *sgc;
7188 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7189 GFP_KERNEL, cpu_to_node(j));
7193 *per_cpu_ptr(sdd->sd, j) = sd;
7195 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7196 GFP_KERNEL, cpu_to_node(j));
7202 *per_cpu_ptr(sdd->sg, j) = sg;
7204 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7205 GFP_KERNEL, cpu_to_node(j));
7209 *per_cpu_ptr(sdd->sgc, j) = sgc;
7216 static void __sdt_free(const struct cpumask *cpu_map)
7218 struct sched_domain_topology_level *tl;
7221 for_each_sd_topology(tl) {
7222 struct sd_data *sdd = &tl->data;
7224 for_each_cpu(j, cpu_map) {
7225 struct sched_domain *sd;
7228 sd = *per_cpu_ptr(sdd->sd, j);
7229 if (sd && (sd->flags & SD_OVERLAP))
7230 free_sched_groups(sd->groups, 0);
7231 kfree(*per_cpu_ptr(sdd->sd, j));
7235 kfree(*per_cpu_ptr(sdd->sg, j));
7237 kfree(*per_cpu_ptr(sdd->sgc, j));
7239 free_percpu(sdd->sd);
7241 free_percpu(sdd->sg);
7243 free_percpu(sdd->sgc);
7248 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7249 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7250 struct sched_domain *child, int cpu)
7252 struct sched_domain *sd = sd_init(tl, cpu);
7256 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7258 sd->level = child->level + 1;
7259 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7263 if (!cpumask_subset(sched_domain_span(child),
7264 sched_domain_span(sd))) {
7265 pr_err("BUG: arch topology borken\n");
7266 #ifdef CONFIG_SCHED_DEBUG
7267 pr_err(" the %s domain not a subset of the %s domain\n",
7268 child->name, sd->name);
7270 /* Fixup, ensure @sd has at least @child cpus. */
7271 cpumask_or(sched_domain_span(sd),
7272 sched_domain_span(sd),
7273 sched_domain_span(child));
7277 set_domain_attribute(sd, attr);
7283 * Build sched domains for a given set of cpus and attach the sched domains
7284 * to the individual cpus
7286 static int build_sched_domains(const struct cpumask *cpu_map,
7287 struct sched_domain_attr *attr)
7289 enum s_alloc alloc_state;
7290 struct sched_domain *sd;
7292 struct rq *rq = NULL;
7293 int i, ret = -ENOMEM;
7295 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7296 if (alloc_state != sa_rootdomain)
7299 /* Set up domains for cpus specified by the cpu_map. */
7300 for_each_cpu(i, cpu_map) {
7301 struct sched_domain_topology_level *tl;
7304 for_each_sd_topology(tl) {
7305 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7306 if (tl == sched_domain_topology)
7307 *per_cpu_ptr(d.sd, i) = sd;
7308 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7309 sd->flags |= SD_OVERLAP;
7310 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7315 /* Build the groups for the domains */
7316 for_each_cpu(i, cpu_map) {
7317 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7318 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7319 if (sd->flags & SD_OVERLAP) {
7320 if (build_overlap_sched_groups(sd, i))
7323 if (build_sched_groups(sd, i))
7329 /* Calculate CPU capacity for physical packages and nodes */
7330 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7331 struct sched_domain_topology_level *tl = sched_domain_topology;
7333 if (!cpumask_test_cpu(i, cpu_map))
7336 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7337 init_sched_energy(i, sd, tl->energy);
7338 claim_allocations(i, sd);
7339 init_sched_groups_capacity(i, sd);
7343 /* Attach the domains */
7345 for_each_cpu(i, cpu_map) {
7347 sd = *per_cpu_ptr(d.sd, i);
7348 cpu_attach_domain(sd, d.rd, i);
7354 __free_domain_allocs(&d, alloc_state, cpu_map);
7358 static cpumask_var_t *doms_cur; /* current sched domains */
7359 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7360 static struct sched_domain_attr *dattr_cur;
7361 /* attribues of custom domains in 'doms_cur' */
7364 * Special case: If a kmalloc of a doms_cur partition (array of
7365 * cpumask) fails, then fallback to a single sched domain,
7366 * as determined by the single cpumask fallback_doms.
7368 static cpumask_var_t fallback_doms;
7371 * arch_update_cpu_topology lets virtualized architectures update the
7372 * cpu core maps. It is supposed to return 1 if the topology changed
7373 * or 0 if it stayed the same.
7375 int __weak arch_update_cpu_topology(void)
7380 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7383 cpumask_var_t *doms;
7385 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7388 for (i = 0; i < ndoms; i++) {
7389 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7390 free_sched_domains(doms, i);
7397 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7400 for (i = 0; i < ndoms; i++)
7401 free_cpumask_var(doms[i]);
7406 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7407 * For now this just excludes isolated cpus, but could be used to
7408 * exclude other special cases in the future.
7410 static int init_sched_domains(const struct cpumask *cpu_map)
7414 arch_update_cpu_topology();
7416 doms_cur = alloc_sched_domains(ndoms_cur);
7418 doms_cur = &fallback_doms;
7419 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7420 err = build_sched_domains(doms_cur[0], NULL);
7421 register_sched_domain_sysctl();
7427 * Detach sched domains from a group of cpus specified in cpu_map
7428 * These cpus will now be attached to the NULL domain
7430 static void detach_destroy_domains(const struct cpumask *cpu_map)
7435 for_each_cpu(i, cpu_map)
7436 cpu_attach_domain(NULL, &def_root_domain, i);
7440 /* handle null as "default" */
7441 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7442 struct sched_domain_attr *new, int idx_new)
7444 struct sched_domain_attr tmp;
7451 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7452 new ? (new + idx_new) : &tmp,
7453 sizeof(struct sched_domain_attr));
7457 * Partition sched domains as specified by the 'ndoms_new'
7458 * cpumasks in the array doms_new[] of cpumasks. This compares
7459 * doms_new[] to the current sched domain partitioning, doms_cur[].
7460 * It destroys each deleted domain and builds each new domain.
7462 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7463 * The masks don't intersect (don't overlap.) We should setup one
7464 * sched domain for each mask. CPUs not in any of the cpumasks will
7465 * not be load balanced. If the same cpumask appears both in the
7466 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7469 * The passed in 'doms_new' should be allocated using
7470 * alloc_sched_domains. This routine takes ownership of it and will
7471 * free_sched_domains it when done with it. If the caller failed the
7472 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7473 * and partition_sched_domains() will fallback to the single partition
7474 * 'fallback_doms', it also forces the domains to be rebuilt.
7476 * If doms_new == NULL it will be replaced with cpu_online_mask.
7477 * ndoms_new == 0 is a special case for destroying existing domains,
7478 * and it will not create the default domain.
7480 * Call with hotplug lock held
7482 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7483 struct sched_domain_attr *dattr_new)
7488 mutex_lock(&sched_domains_mutex);
7490 /* always unregister in case we don't destroy any domains */
7491 unregister_sched_domain_sysctl();
7493 /* Let architecture update cpu core mappings. */
7494 new_topology = arch_update_cpu_topology();
7496 n = doms_new ? ndoms_new : 0;
7498 /* Destroy deleted domains */
7499 for (i = 0; i < ndoms_cur; i++) {
7500 for (j = 0; j < n && !new_topology; j++) {
7501 if (cpumask_equal(doms_cur[i], doms_new[j])
7502 && dattrs_equal(dattr_cur, i, dattr_new, j))
7505 /* no match - a current sched domain not in new doms_new[] */
7506 detach_destroy_domains(doms_cur[i]);
7512 if (doms_new == NULL) {
7514 doms_new = &fallback_doms;
7515 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7516 WARN_ON_ONCE(dattr_new);
7519 /* Build new domains */
7520 for (i = 0; i < ndoms_new; i++) {
7521 for (j = 0; j < n && !new_topology; j++) {
7522 if (cpumask_equal(doms_new[i], doms_cur[j])
7523 && dattrs_equal(dattr_new, i, dattr_cur, j))
7526 /* no match - add a new doms_new */
7527 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7532 /* Remember the new sched domains */
7533 if (doms_cur != &fallback_doms)
7534 free_sched_domains(doms_cur, ndoms_cur);
7535 kfree(dattr_cur); /* kfree(NULL) is safe */
7536 doms_cur = doms_new;
7537 dattr_cur = dattr_new;
7538 ndoms_cur = ndoms_new;
7540 register_sched_domain_sysctl();
7542 mutex_unlock(&sched_domains_mutex);
7545 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7548 * Update cpusets according to cpu_active mask. If cpusets are
7549 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7550 * around partition_sched_domains().
7552 * If we come here as part of a suspend/resume, don't touch cpusets because we
7553 * want to restore it back to its original state upon resume anyway.
7555 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7559 case CPU_ONLINE_FROZEN:
7560 case CPU_DOWN_FAILED_FROZEN:
7563 * num_cpus_frozen tracks how many CPUs are involved in suspend
7564 * resume sequence. As long as this is not the last online
7565 * operation in the resume sequence, just build a single sched
7566 * domain, ignoring cpusets.
7569 if (likely(num_cpus_frozen)) {
7570 partition_sched_domains(1, NULL, NULL);
7575 * This is the last CPU online operation. So fall through and
7576 * restore the original sched domains by considering the
7577 * cpuset configurations.
7581 cpuset_update_active_cpus(true);
7589 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7592 unsigned long flags;
7593 long cpu = (long)hcpu;
7599 case CPU_DOWN_PREPARE:
7600 rcu_read_lock_sched();
7601 dl_b = dl_bw_of(cpu);
7603 raw_spin_lock_irqsave(&dl_b->lock, flags);
7604 cpus = dl_bw_cpus(cpu);
7605 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7606 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7608 rcu_read_unlock_sched();
7611 return notifier_from_errno(-EBUSY);
7612 cpuset_update_active_cpus(false);
7614 case CPU_DOWN_PREPARE_FROZEN:
7616 partition_sched_domains(1, NULL, NULL);
7624 void __init sched_init_smp(void)
7626 cpumask_var_t non_isolated_cpus;
7628 walt_init_cpu_efficiency();
7629 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7630 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7635 * There's no userspace yet to cause hotplug operations; hence all the
7636 * cpu masks are stable and all blatant races in the below code cannot
7639 mutex_lock(&sched_domains_mutex);
7640 init_sched_domains(cpu_active_mask);
7641 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7642 if (cpumask_empty(non_isolated_cpus))
7643 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7644 mutex_unlock(&sched_domains_mutex);
7646 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7647 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7648 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7652 /* Move init over to a non-isolated CPU */
7653 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7655 sched_init_granularity();
7656 free_cpumask_var(non_isolated_cpus);
7658 init_sched_rt_class();
7659 init_sched_dl_class();
7662 void __init sched_init_smp(void)
7664 sched_init_granularity();
7666 #endif /* CONFIG_SMP */
7668 int in_sched_functions(unsigned long addr)
7670 return in_lock_functions(addr) ||
7671 (addr >= (unsigned long)__sched_text_start
7672 && addr < (unsigned long)__sched_text_end);
7675 #ifdef CONFIG_CGROUP_SCHED
7677 * Default task group.
7678 * Every task in system belongs to this group at bootup.
7680 struct task_group root_task_group;
7681 LIST_HEAD(task_groups);
7684 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7686 void __init sched_init(void)
7689 unsigned long alloc_size = 0, ptr;
7691 #ifdef CONFIG_FAIR_GROUP_SCHED
7692 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7694 #ifdef CONFIG_RT_GROUP_SCHED
7695 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7698 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7700 #ifdef CONFIG_FAIR_GROUP_SCHED
7701 root_task_group.se = (struct sched_entity **)ptr;
7702 ptr += nr_cpu_ids * sizeof(void **);
7704 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7705 ptr += nr_cpu_ids * sizeof(void **);
7707 #endif /* CONFIG_FAIR_GROUP_SCHED */
7708 #ifdef CONFIG_RT_GROUP_SCHED
7709 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7710 ptr += nr_cpu_ids * sizeof(void **);
7712 root_task_group.rt_rq = (struct rt_rq **)ptr;
7713 ptr += nr_cpu_ids * sizeof(void **);
7715 #endif /* CONFIG_RT_GROUP_SCHED */
7717 #ifdef CONFIG_CPUMASK_OFFSTACK
7718 for_each_possible_cpu(i) {
7719 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7720 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7722 #endif /* CONFIG_CPUMASK_OFFSTACK */
7724 init_rt_bandwidth(&def_rt_bandwidth,
7725 global_rt_period(), global_rt_runtime());
7726 init_dl_bandwidth(&def_dl_bandwidth,
7727 global_rt_period(), global_rt_runtime());
7730 init_defrootdomain();
7733 #ifdef CONFIG_RT_GROUP_SCHED
7734 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7735 global_rt_period(), global_rt_runtime());
7736 #endif /* CONFIG_RT_GROUP_SCHED */
7738 #ifdef CONFIG_CGROUP_SCHED
7739 list_add(&root_task_group.list, &task_groups);
7740 INIT_LIST_HEAD(&root_task_group.children);
7741 INIT_LIST_HEAD(&root_task_group.siblings);
7742 autogroup_init(&init_task);
7744 #endif /* CONFIG_CGROUP_SCHED */
7746 for_each_possible_cpu(i) {
7750 raw_spin_lock_init(&rq->lock);
7752 rq->calc_load_active = 0;
7753 rq->calc_load_update = jiffies + LOAD_FREQ;
7754 init_cfs_rq(&rq->cfs);
7755 init_rt_rq(&rq->rt);
7756 init_dl_rq(&rq->dl);
7757 #ifdef CONFIG_FAIR_GROUP_SCHED
7758 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7759 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7761 * How much cpu bandwidth does root_task_group get?
7763 * In case of task-groups formed thr' the cgroup filesystem, it
7764 * gets 100% of the cpu resources in the system. This overall
7765 * system cpu resource is divided among the tasks of
7766 * root_task_group and its child task-groups in a fair manner,
7767 * based on each entity's (task or task-group's) weight
7768 * (se->load.weight).
7770 * In other words, if root_task_group has 10 tasks of weight
7771 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7772 * then A0's share of the cpu resource is:
7774 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7776 * We achieve this by letting root_task_group's tasks sit
7777 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7779 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7780 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7781 #endif /* CONFIG_FAIR_GROUP_SCHED */
7783 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7784 #ifdef CONFIG_RT_GROUP_SCHED
7785 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7788 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7789 rq->cpu_load[j] = 0;
7791 rq->last_load_update_tick = jiffies;
7796 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7797 rq->balance_callback = NULL;
7798 rq->active_balance = 0;
7799 rq->next_balance = jiffies;
7804 rq->avg_idle = 2*sysctl_sched_migration_cost;
7805 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7806 #ifdef CONFIG_SCHED_WALT
7807 rq->cur_irqload = 0;
7808 rq->avg_irqload = 0;
7812 INIT_LIST_HEAD(&rq->cfs_tasks);
7814 rq_attach_root(rq, &def_root_domain);
7815 #ifdef CONFIG_NO_HZ_COMMON
7818 #ifdef CONFIG_NO_HZ_FULL
7819 rq->last_sched_tick = 0;
7823 atomic_set(&rq->nr_iowait, 0);
7826 set_load_weight(&init_task);
7828 #ifdef CONFIG_PREEMPT_NOTIFIERS
7829 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7833 * The boot idle thread does lazy MMU switching as well:
7835 atomic_inc(&init_mm.mm_count);
7836 enter_lazy_tlb(&init_mm, current);
7839 * During early bootup we pretend to be a normal task:
7841 current->sched_class = &fair_sched_class;
7844 * Make us the idle thread. Technically, schedule() should not be
7845 * called from this thread, however somewhere below it might be,
7846 * but because we are the idle thread, we just pick up running again
7847 * when this runqueue becomes "idle".
7849 init_idle(current, smp_processor_id());
7851 calc_load_update = jiffies + LOAD_FREQ;
7854 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7855 /* May be allocated at isolcpus cmdline parse time */
7856 if (cpu_isolated_map == NULL)
7857 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7858 idle_thread_set_boot_cpu();
7859 set_cpu_rq_start_time();
7861 init_sched_fair_class();
7863 scheduler_running = 1;
7866 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7867 static inline int preempt_count_equals(int preempt_offset)
7869 int nested = preempt_count() + rcu_preempt_depth();
7871 return (nested == preempt_offset);
7874 static int __might_sleep_init_called;
7875 int __init __might_sleep_init(void)
7877 __might_sleep_init_called = 1;
7880 early_initcall(__might_sleep_init);
7882 void __might_sleep(const char *file, int line, int preempt_offset)
7885 * Blocking primitives will set (and therefore destroy) current->state,
7886 * since we will exit with TASK_RUNNING make sure we enter with it,
7887 * otherwise we will destroy state.
7889 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7890 "do not call blocking ops when !TASK_RUNNING; "
7891 "state=%lx set at [<%p>] %pS\n",
7893 (void *)current->task_state_change,
7894 (void *)current->task_state_change);
7896 ___might_sleep(file, line, preempt_offset);
7898 EXPORT_SYMBOL(__might_sleep);
7900 void ___might_sleep(const char *file, int line, int preempt_offset)
7902 static unsigned long prev_jiffy; /* ratelimiting */
7904 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7905 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7906 !is_idle_task(current)) || oops_in_progress)
7908 if (system_state != SYSTEM_RUNNING &&
7909 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
7911 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7913 prev_jiffy = jiffies;
7916 "BUG: sleeping function called from invalid context at %s:%d\n",
7919 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7920 in_atomic(), irqs_disabled(),
7921 current->pid, current->comm);
7923 if (task_stack_end_corrupted(current))
7924 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7926 debug_show_held_locks(current);
7927 if (irqs_disabled())
7928 print_irqtrace_events(current);
7929 #ifdef CONFIG_DEBUG_PREEMPT
7930 if (!preempt_count_equals(preempt_offset)) {
7931 pr_err("Preemption disabled at:");
7932 print_ip_sym(current->preempt_disable_ip);
7938 EXPORT_SYMBOL(___might_sleep);
7941 #ifdef CONFIG_MAGIC_SYSRQ
7942 void normalize_rt_tasks(void)
7944 struct task_struct *g, *p;
7945 struct sched_attr attr = {
7946 .sched_policy = SCHED_NORMAL,
7949 read_lock(&tasklist_lock);
7950 for_each_process_thread(g, p) {
7952 * Only normalize user tasks:
7954 if (p->flags & PF_KTHREAD)
7957 p->se.exec_start = 0;
7958 #ifdef CONFIG_SCHEDSTATS
7959 p->se.statistics.wait_start = 0;
7960 p->se.statistics.sleep_start = 0;
7961 p->se.statistics.block_start = 0;
7964 if (!dl_task(p) && !rt_task(p)) {
7966 * Renice negative nice level userspace
7969 if (task_nice(p) < 0)
7970 set_user_nice(p, 0);
7974 __sched_setscheduler(p, &attr, false, false);
7976 read_unlock(&tasklist_lock);
7979 #endif /* CONFIG_MAGIC_SYSRQ */
7981 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7983 * These functions are only useful for the IA64 MCA handling, or kdb.
7985 * They can only be called when the whole system has been
7986 * stopped - every CPU needs to be quiescent, and no scheduling
7987 * activity can take place. Using them for anything else would
7988 * be a serious bug, and as a result, they aren't even visible
7989 * under any other configuration.
7993 * curr_task - return the current task for a given cpu.
7994 * @cpu: the processor in question.
7996 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7998 * Return: The current task for @cpu.
8000 struct task_struct *curr_task(int cpu)
8002 return cpu_curr(cpu);
8005 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8009 * set_curr_task - set the current task for a given cpu.
8010 * @cpu: the processor in question.
8011 * @p: the task pointer to set.
8013 * Description: This function must only be used when non-maskable interrupts
8014 * are serviced on a separate stack. It allows the architecture to switch the
8015 * notion of the current task on a cpu in a non-blocking manner. This function
8016 * must be called with all CPU's synchronized, and interrupts disabled, the
8017 * and caller must save the original value of the current task (see
8018 * curr_task() above) and restore that value before reenabling interrupts and
8019 * re-starting the system.
8021 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8023 void set_curr_task(int cpu, struct task_struct *p)
8030 #ifdef CONFIG_CGROUP_SCHED
8031 /* task_group_lock serializes the addition/removal of task groups */
8032 static DEFINE_SPINLOCK(task_group_lock);
8034 static void sched_free_group(struct task_group *tg)
8036 free_fair_sched_group(tg);
8037 free_rt_sched_group(tg);
8042 /* allocate runqueue etc for a new task group */
8043 struct task_group *sched_create_group(struct task_group *parent)
8045 struct task_group *tg;
8047 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8049 return ERR_PTR(-ENOMEM);
8051 if (!alloc_fair_sched_group(tg, parent))
8054 if (!alloc_rt_sched_group(tg, parent))
8060 sched_free_group(tg);
8061 return ERR_PTR(-ENOMEM);
8064 void sched_online_group(struct task_group *tg, struct task_group *parent)
8066 unsigned long flags;
8068 spin_lock_irqsave(&task_group_lock, flags);
8069 list_add_rcu(&tg->list, &task_groups);
8071 WARN_ON(!parent); /* root should already exist */
8073 tg->parent = parent;
8074 INIT_LIST_HEAD(&tg->children);
8075 list_add_rcu(&tg->siblings, &parent->children);
8076 spin_unlock_irqrestore(&task_group_lock, flags);
8079 /* rcu callback to free various structures associated with a task group */
8080 static void sched_free_group_rcu(struct rcu_head *rhp)
8082 /* now it should be safe to free those cfs_rqs */
8083 sched_free_group(container_of(rhp, struct task_group, rcu));
8086 void sched_destroy_group(struct task_group *tg)
8088 /* wait for possible concurrent references to cfs_rqs complete */
8089 call_rcu(&tg->rcu, sched_free_group_rcu);
8092 void sched_offline_group(struct task_group *tg)
8094 unsigned long flags;
8097 /* end participation in shares distribution */
8098 for_each_possible_cpu(i)
8099 unregister_fair_sched_group(tg, i);
8101 spin_lock_irqsave(&task_group_lock, flags);
8102 list_del_rcu(&tg->list);
8103 list_del_rcu(&tg->siblings);
8104 spin_unlock_irqrestore(&task_group_lock, flags);
8107 /* change task's runqueue when it moves between groups.
8108 * The caller of this function should have put the task in its new group
8109 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8110 * reflect its new group.
8112 void sched_move_task(struct task_struct *tsk)
8114 struct task_group *tg;
8115 int queued, running;
8116 unsigned long flags;
8119 rq = task_rq_lock(tsk, &flags);
8121 running = task_current(rq, tsk);
8122 queued = task_on_rq_queued(tsk);
8125 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8126 if (unlikely(running))
8127 put_prev_task(rq, tsk);
8130 * All callers are synchronized by task_rq_lock(); we do not use RCU
8131 * which is pointless here. Thus, we pass "true" to task_css_check()
8132 * to prevent lockdep warnings.
8134 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8135 struct task_group, css);
8136 tg = autogroup_task_group(tsk, tg);
8137 tsk->sched_task_group = tg;
8139 #ifdef CONFIG_FAIR_GROUP_SCHED
8140 if (tsk->sched_class->task_move_group)
8141 tsk->sched_class->task_move_group(tsk);
8144 set_task_rq(tsk, task_cpu(tsk));
8146 if (unlikely(running))
8147 tsk->sched_class->set_curr_task(rq);
8149 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8151 task_rq_unlock(rq, tsk, &flags);
8153 #endif /* CONFIG_CGROUP_SCHED */
8155 #ifdef CONFIG_RT_GROUP_SCHED
8157 * Ensure that the real time constraints are schedulable.
8159 static DEFINE_MUTEX(rt_constraints_mutex);
8161 /* Must be called with tasklist_lock held */
8162 static inline int tg_has_rt_tasks(struct task_group *tg)
8164 struct task_struct *g, *p;
8167 * Autogroups do not have RT tasks; see autogroup_create().
8169 if (task_group_is_autogroup(tg))
8172 for_each_process_thread(g, p) {
8173 if (rt_task(p) && task_group(p) == tg)
8180 struct rt_schedulable_data {
8181 struct task_group *tg;
8186 static int tg_rt_schedulable(struct task_group *tg, void *data)
8188 struct rt_schedulable_data *d = data;
8189 struct task_group *child;
8190 unsigned long total, sum = 0;
8191 u64 period, runtime;
8193 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8194 runtime = tg->rt_bandwidth.rt_runtime;
8197 period = d->rt_period;
8198 runtime = d->rt_runtime;
8202 * Cannot have more runtime than the period.
8204 if (runtime > period && runtime != RUNTIME_INF)
8208 * Ensure we don't starve existing RT tasks.
8210 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8213 total = to_ratio(period, runtime);
8216 * Nobody can have more than the global setting allows.
8218 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8222 * The sum of our children's runtime should not exceed our own.
8224 list_for_each_entry_rcu(child, &tg->children, siblings) {
8225 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8226 runtime = child->rt_bandwidth.rt_runtime;
8228 if (child == d->tg) {
8229 period = d->rt_period;
8230 runtime = d->rt_runtime;
8233 sum += to_ratio(period, runtime);
8242 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8246 struct rt_schedulable_data data = {
8248 .rt_period = period,
8249 .rt_runtime = runtime,
8253 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8259 static int tg_set_rt_bandwidth(struct task_group *tg,
8260 u64 rt_period, u64 rt_runtime)
8265 * Disallowing the root group RT runtime is BAD, it would disallow the
8266 * kernel creating (and or operating) RT threads.
8268 if (tg == &root_task_group && rt_runtime == 0)
8271 /* No period doesn't make any sense. */
8275 mutex_lock(&rt_constraints_mutex);
8276 read_lock(&tasklist_lock);
8277 err = __rt_schedulable(tg, rt_period, rt_runtime);
8281 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8282 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8283 tg->rt_bandwidth.rt_runtime = rt_runtime;
8285 for_each_possible_cpu(i) {
8286 struct rt_rq *rt_rq = tg->rt_rq[i];
8288 raw_spin_lock(&rt_rq->rt_runtime_lock);
8289 rt_rq->rt_runtime = rt_runtime;
8290 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8292 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8294 read_unlock(&tasklist_lock);
8295 mutex_unlock(&rt_constraints_mutex);
8300 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8302 u64 rt_runtime, rt_period;
8304 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8305 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8306 if (rt_runtime_us < 0)
8307 rt_runtime = RUNTIME_INF;
8309 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8312 static long sched_group_rt_runtime(struct task_group *tg)
8316 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8319 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8320 do_div(rt_runtime_us, NSEC_PER_USEC);
8321 return rt_runtime_us;
8324 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8326 u64 rt_runtime, rt_period;
8328 rt_period = rt_period_us * NSEC_PER_USEC;
8329 rt_runtime = tg->rt_bandwidth.rt_runtime;
8331 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8334 static long sched_group_rt_period(struct task_group *tg)
8338 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8339 do_div(rt_period_us, NSEC_PER_USEC);
8340 return rt_period_us;
8342 #endif /* CONFIG_RT_GROUP_SCHED */
8344 #ifdef CONFIG_RT_GROUP_SCHED
8345 static int sched_rt_global_constraints(void)
8349 mutex_lock(&rt_constraints_mutex);
8350 read_lock(&tasklist_lock);
8351 ret = __rt_schedulable(NULL, 0, 0);
8352 read_unlock(&tasklist_lock);
8353 mutex_unlock(&rt_constraints_mutex);
8358 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8360 /* Don't accept realtime tasks when there is no way for them to run */
8361 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8367 #else /* !CONFIG_RT_GROUP_SCHED */
8368 static int sched_rt_global_constraints(void)
8370 unsigned long flags;
8373 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8374 for_each_possible_cpu(i) {
8375 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8377 raw_spin_lock(&rt_rq->rt_runtime_lock);
8378 rt_rq->rt_runtime = global_rt_runtime();
8379 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8381 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8385 #endif /* CONFIG_RT_GROUP_SCHED */
8387 static int sched_dl_global_validate(void)
8389 u64 runtime = global_rt_runtime();
8390 u64 period = global_rt_period();
8391 u64 new_bw = to_ratio(period, runtime);
8394 unsigned long flags;
8397 * Here we want to check the bandwidth not being set to some
8398 * value smaller than the currently allocated bandwidth in
8399 * any of the root_domains.
8401 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8402 * cycling on root_domains... Discussion on different/better
8403 * solutions is welcome!
8405 for_each_possible_cpu(cpu) {
8406 rcu_read_lock_sched();
8407 dl_b = dl_bw_of(cpu);
8409 raw_spin_lock_irqsave(&dl_b->lock, flags);
8410 if (new_bw < dl_b->total_bw)
8412 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8414 rcu_read_unlock_sched();
8423 static void sched_dl_do_global(void)
8428 unsigned long flags;
8430 def_dl_bandwidth.dl_period = global_rt_period();
8431 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8433 if (global_rt_runtime() != RUNTIME_INF)
8434 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8437 * FIXME: As above...
8439 for_each_possible_cpu(cpu) {
8440 rcu_read_lock_sched();
8441 dl_b = dl_bw_of(cpu);
8443 raw_spin_lock_irqsave(&dl_b->lock, flags);
8445 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8447 rcu_read_unlock_sched();
8451 static int sched_rt_global_validate(void)
8453 if (sysctl_sched_rt_period <= 0)
8456 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8457 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8463 static void sched_rt_do_global(void)
8465 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8466 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8469 int sched_rt_handler(struct ctl_table *table, int write,
8470 void __user *buffer, size_t *lenp,
8473 int old_period, old_runtime;
8474 static DEFINE_MUTEX(mutex);
8478 old_period = sysctl_sched_rt_period;
8479 old_runtime = sysctl_sched_rt_runtime;
8481 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8483 if (!ret && write) {
8484 ret = sched_rt_global_validate();
8488 ret = sched_dl_global_validate();
8492 ret = sched_rt_global_constraints();
8496 sched_rt_do_global();
8497 sched_dl_do_global();
8501 sysctl_sched_rt_period = old_period;
8502 sysctl_sched_rt_runtime = old_runtime;
8504 mutex_unlock(&mutex);
8509 int sched_rr_handler(struct ctl_table *table, int write,
8510 void __user *buffer, size_t *lenp,
8514 static DEFINE_MUTEX(mutex);
8517 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8518 /* make sure that internally we keep jiffies */
8519 /* also, writing zero resets timeslice to default */
8520 if (!ret && write) {
8521 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8522 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8524 mutex_unlock(&mutex);
8528 #ifdef CONFIG_CGROUP_SCHED
8530 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8532 return css ? container_of(css, struct task_group, css) : NULL;
8535 static struct cgroup_subsys_state *
8536 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8538 struct task_group *parent = css_tg(parent_css);
8539 struct task_group *tg;
8542 /* This is early initialization for the top cgroup */
8543 return &root_task_group.css;
8546 tg = sched_create_group(parent);
8548 return ERR_PTR(-ENOMEM);
8550 sched_online_group(tg, parent);
8555 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8557 struct task_group *tg = css_tg(css);
8559 sched_offline_group(tg);
8562 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8564 struct task_group *tg = css_tg(css);
8567 * Relies on the RCU grace period between css_released() and this.
8569 sched_free_group(tg);
8572 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8574 sched_move_task(task);
8577 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8579 struct task_struct *task;
8580 struct cgroup_subsys_state *css;
8582 cgroup_taskset_for_each(task, css, tset) {
8583 #ifdef CONFIG_RT_GROUP_SCHED
8584 if (!sched_rt_can_attach(css_tg(css), task))
8587 /* We don't support RT-tasks being in separate groups */
8588 if (task->sched_class != &fair_sched_class)
8595 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8597 struct task_struct *task;
8598 struct cgroup_subsys_state *css;
8600 cgroup_taskset_for_each(task, css, tset)
8601 sched_move_task(task);
8604 #ifdef CONFIG_FAIR_GROUP_SCHED
8605 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8606 struct cftype *cftype, u64 shareval)
8608 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8611 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8614 struct task_group *tg = css_tg(css);
8616 return (u64) scale_load_down(tg->shares);
8619 #ifdef CONFIG_CFS_BANDWIDTH
8620 static DEFINE_MUTEX(cfs_constraints_mutex);
8622 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8623 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8625 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8627 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8629 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8630 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8632 if (tg == &root_task_group)
8636 * Ensure we have at some amount of bandwidth every period. This is
8637 * to prevent reaching a state of large arrears when throttled via
8638 * entity_tick() resulting in prolonged exit starvation.
8640 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8644 * Likewise, bound things on the otherside by preventing insane quota
8645 * periods. This also allows us to normalize in computing quota
8648 if (period > max_cfs_quota_period)
8652 * Prevent race between setting of cfs_rq->runtime_enabled and
8653 * unthrottle_offline_cfs_rqs().
8656 mutex_lock(&cfs_constraints_mutex);
8657 ret = __cfs_schedulable(tg, period, quota);
8661 runtime_enabled = quota != RUNTIME_INF;
8662 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8664 * If we need to toggle cfs_bandwidth_used, off->on must occur
8665 * before making related changes, and on->off must occur afterwards
8667 if (runtime_enabled && !runtime_was_enabled)
8668 cfs_bandwidth_usage_inc();
8669 raw_spin_lock_irq(&cfs_b->lock);
8670 cfs_b->period = ns_to_ktime(period);
8671 cfs_b->quota = quota;
8673 __refill_cfs_bandwidth_runtime(cfs_b);
8674 /* restart the period timer (if active) to handle new period expiry */
8675 if (runtime_enabled)
8676 start_cfs_bandwidth(cfs_b);
8677 raw_spin_unlock_irq(&cfs_b->lock);
8679 for_each_online_cpu(i) {
8680 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8681 struct rq *rq = cfs_rq->rq;
8683 raw_spin_lock_irq(&rq->lock);
8684 cfs_rq->runtime_enabled = runtime_enabled;
8685 cfs_rq->runtime_remaining = 0;
8687 if (cfs_rq->throttled)
8688 unthrottle_cfs_rq(cfs_rq);
8689 raw_spin_unlock_irq(&rq->lock);
8691 if (runtime_was_enabled && !runtime_enabled)
8692 cfs_bandwidth_usage_dec();
8694 mutex_unlock(&cfs_constraints_mutex);
8700 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8704 period = ktime_to_ns(tg->cfs_bandwidth.period);
8705 if (cfs_quota_us < 0)
8706 quota = RUNTIME_INF;
8708 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8710 return tg_set_cfs_bandwidth(tg, period, quota);
8713 long tg_get_cfs_quota(struct task_group *tg)
8717 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8720 quota_us = tg->cfs_bandwidth.quota;
8721 do_div(quota_us, NSEC_PER_USEC);
8726 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8730 period = (u64)cfs_period_us * NSEC_PER_USEC;
8731 quota = tg->cfs_bandwidth.quota;
8733 return tg_set_cfs_bandwidth(tg, period, quota);
8736 long tg_get_cfs_period(struct task_group *tg)
8740 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8741 do_div(cfs_period_us, NSEC_PER_USEC);
8743 return cfs_period_us;
8746 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8749 return tg_get_cfs_quota(css_tg(css));
8752 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8753 struct cftype *cftype, s64 cfs_quota_us)
8755 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8758 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8761 return tg_get_cfs_period(css_tg(css));
8764 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8765 struct cftype *cftype, u64 cfs_period_us)
8767 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8770 struct cfs_schedulable_data {
8771 struct task_group *tg;
8776 * normalize group quota/period to be quota/max_period
8777 * note: units are usecs
8779 static u64 normalize_cfs_quota(struct task_group *tg,
8780 struct cfs_schedulable_data *d)
8788 period = tg_get_cfs_period(tg);
8789 quota = tg_get_cfs_quota(tg);
8792 /* note: these should typically be equivalent */
8793 if (quota == RUNTIME_INF || quota == -1)
8796 return to_ratio(period, quota);
8799 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8801 struct cfs_schedulable_data *d = data;
8802 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8803 s64 quota = 0, parent_quota = -1;
8806 quota = RUNTIME_INF;
8808 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8810 quota = normalize_cfs_quota(tg, d);
8811 parent_quota = parent_b->hierarchical_quota;
8814 * ensure max(child_quota) <= parent_quota, inherit when no
8817 if (quota == RUNTIME_INF)
8818 quota = parent_quota;
8819 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8822 cfs_b->hierarchical_quota = quota;
8827 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8830 struct cfs_schedulable_data data = {
8836 if (quota != RUNTIME_INF) {
8837 do_div(data.period, NSEC_PER_USEC);
8838 do_div(data.quota, NSEC_PER_USEC);
8842 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8848 static int cpu_stats_show(struct seq_file *sf, void *v)
8850 struct task_group *tg = css_tg(seq_css(sf));
8851 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8853 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8854 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8855 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8859 #endif /* CONFIG_CFS_BANDWIDTH */
8860 #endif /* CONFIG_FAIR_GROUP_SCHED */
8862 #ifdef CONFIG_RT_GROUP_SCHED
8863 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8864 struct cftype *cft, s64 val)
8866 return sched_group_set_rt_runtime(css_tg(css), val);
8869 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8872 return sched_group_rt_runtime(css_tg(css));
8875 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8876 struct cftype *cftype, u64 rt_period_us)
8878 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8881 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8884 return sched_group_rt_period(css_tg(css));
8886 #endif /* CONFIG_RT_GROUP_SCHED */
8888 static struct cftype cpu_files[] = {
8889 #ifdef CONFIG_FAIR_GROUP_SCHED
8892 .read_u64 = cpu_shares_read_u64,
8893 .write_u64 = cpu_shares_write_u64,
8896 #ifdef CONFIG_CFS_BANDWIDTH
8898 .name = "cfs_quota_us",
8899 .read_s64 = cpu_cfs_quota_read_s64,
8900 .write_s64 = cpu_cfs_quota_write_s64,
8903 .name = "cfs_period_us",
8904 .read_u64 = cpu_cfs_period_read_u64,
8905 .write_u64 = cpu_cfs_period_write_u64,
8909 .seq_show = cpu_stats_show,
8912 #ifdef CONFIG_RT_GROUP_SCHED
8914 .name = "rt_runtime_us",
8915 .read_s64 = cpu_rt_runtime_read,
8916 .write_s64 = cpu_rt_runtime_write,
8919 .name = "rt_period_us",
8920 .read_u64 = cpu_rt_period_read_uint,
8921 .write_u64 = cpu_rt_period_write_uint,
8927 struct cgroup_subsys cpu_cgrp_subsys = {
8928 .css_alloc = cpu_cgroup_css_alloc,
8929 .css_released = cpu_cgroup_css_released,
8930 .css_free = cpu_cgroup_css_free,
8931 .fork = cpu_cgroup_fork,
8932 .can_attach = cpu_cgroup_can_attach,
8933 .attach = cpu_cgroup_attach,
8934 .allow_attach = subsys_cgroup_allow_attach,
8935 .legacy_cftypes = cpu_files,
8939 #endif /* CONFIG_CGROUP_SCHED */
8941 void dump_cpu_task(int cpu)
8943 pr_info("Task dump for CPU %d:\n", cpu);
8944 sched_show_task(cpu_curr(cpu));