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
2929 static unsigned long sum_capacity_reqs(unsigned long cfs_cap,
2930 struct sched_capacity_reqs *scr)
2932 unsigned long total = cfs_cap + scr->rt;
2934 total = total * capacity_margin;
2935 total /= SCHED_CAPACITY_SCALE;
2940 static void sched_freq_tick(int cpu)
2942 struct sched_capacity_reqs *scr;
2943 unsigned long capacity_orig, capacity_curr, capacity_sum;
2948 capacity_orig = capacity_orig_of(cpu);
2949 capacity_curr = capacity_curr_of(cpu);
2950 if (capacity_curr == capacity_orig)
2954 * To make free room for a task that is building up its "real"
2955 * utilization and to harm its performance the least, request
2956 * a jump to a higher OPP as soon as the margin of free capacity
2957 * is impacted (specified by capacity_margin).
2960 scr = &per_cpu(cpu_sched_capacity_reqs, cpu);
2961 capacity_sum = sum_capacity_reqs(cpu_util(cpu), scr);
2962 if (capacity_curr < capacity_sum) {
2963 set_cfs_cpu_capacity(cpu, true, capacity_sum);
2967 static inline void sched_freq_tick(int cpu) { }
2971 * This function gets called by the timer code, with HZ frequency.
2972 * We call it with interrupts disabled.
2974 void scheduler_tick(void)
2976 int cpu = smp_processor_id();
2977 struct rq *rq = cpu_rq(cpu);
2978 struct task_struct *curr = rq->curr;
2982 raw_spin_lock(&rq->lock);
2983 walt_set_window_start(rq);
2984 update_rq_clock(rq);
2985 curr->sched_class->task_tick(rq, curr, 0);
2986 update_cpu_load_active(rq);
2987 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
2988 walt_ktime_clock(), 0);
2989 calc_global_load_tick(rq);
2990 sched_freq_tick(cpu);
2991 raw_spin_unlock(&rq->lock);
2993 perf_event_task_tick();
2996 rq->idle_balance = idle_cpu(cpu);
2997 trigger_load_balance(rq);
2999 rq_last_tick_reset(rq);
3002 #ifdef CONFIG_NO_HZ_FULL
3004 * scheduler_tick_max_deferment
3006 * Keep at least one tick per second when a single
3007 * active task is running because the scheduler doesn't
3008 * yet completely support full dynticks environment.
3010 * This makes sure that uptime, CFS vruntime, load
3011 * balancing, etc... continue to move forward, even
3012 * with a very low granularity.
3014 * Return: Maximum deferment in nanoseconds.
3016 u64 scheduler_tick_max_deferment(void)
3018 struct rq *rq = this_rq();
3019 unsigned long next, now = READ_ONCE(jiffies);
3021 next = rq->last_sched_tick + HZ;
3023 if (time_before_eq(next, now))
3026 return jiffies_to_nsecs(next - now);
3030 notrace unsigned long get_parent_ip(unsigned long addr)
3032 if (in_lock_functions(addr)) {
3033 addr = CALLER_ADDR2;
3034 if (in_lock_functions(addr))
3035 addr = CALLER_ADDR3;
3040 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3041 defined(CONFIG_PREEMPT_TRACER))
3043 void preempt_count_add(int val)
3045 #ifdef CONFIG_DEBUG_PREEMPT
3049 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3052 __preempt_count_add(val);
3053 #ifdef CONFIG_DEBUG_PREEMPT
3055 * Spinlock count overflowing soon?
3057 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3060 if (preempt_count() == val) {
3061 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3062 #ifdef CONFIG_DEBUG_PREEMPT
3063 current->preempt_disable_ip = ip;
3065 trace_preempt_off(CALLER_ADDR0, ip);
3068 EXPORT_SYMBOL(preempt_count_add);
3069 NOKPROBE_SYMBOL(preempt_count_add);
3071 void preempt_count_sub(int val)
3073 #ifdef CONFIG_DEBUG_PREEMPT
3077 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3080 * Is the spinlock portion underflowing?
3082 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3083 !(preempt_count() & PREEMPT_MASK)))
3087 if (preempt_count() == val)
3088 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3089 __preempt_count_sub(val);
3091 EXPORT_SYMBOL(preempt_count_sub);
3092 NOKPROBE_SYMBOL(preempt_count_sub);
3097 * Print scheduling while atomic bug:
3099 static noinline void __schedule_bug(struct task_struct *prev)
3101 if (oops_in_progress)
3104 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3105 prev->comm, prev->pid, preempt_count());
3107 debug_show_held_locks(prev);
3109 if (irqs_disabled())
3110 print_irqtrace_events(prev);
3111 #ifdef CONFIG_DEBUG_PREEMPT
3112 if (in_atomic_preempt_off()) {
3113 pr_err("Preemption disabled at:");
3114 print_ip_sym(current->preempt_disable_ip);
3119 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3123 * Various schedule()-time debugging checks and statistics:
3125 static inline void schedule_debug(struct task_struct *prev)
3127 #ifdef CONFIG_SCHED_STACK_END_CHECK
3128 if (task_stack_end_corrupted(prev))
3129 panic("corrupted stack end detected inside scheduler\n");
3132 if (unlikely(in_atomic_preempt_off())) {
3133 __schedule_bug(prev);
3134 preempt_count_set(PREEMPT_DISABLED);
3138 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3140 schedstat_inc(this_rq(), sched_count);
3144 * Pick up the highest-prio task:
3146 static inline struct task_struct *
3147 pick_next_task(struct rq *rq, struct task_struct *prev)
3149 const struct sched_class *class = &fair_sched_class;
3150 struct task_struct *p;
3153 * Optimization: we know that if all tasks are in
3154 * the fair class we can call that function directly:
3156 if (likely(prev->sched_class == class &&
3157 rq->nr_running == rq->cfs.h_nr_running)) {
3158 p = fair_sched_class.pick_next_task(rq, prev);
3159 if (unlikely(p == RETRY_TASK))
3162 /* assumes fair_sched_class->next == idle_sched_class */
3164 p = idle_sched_class.pick_next_task(rq, prev);
3170 for_each_class(class) {
3171 p = class->pick_next_task(rq, prev);
3173 if (unlikely(p == RETRY_TASK))
3179 BUG(); /* the idle class will always have a runnable task */
3183 * __schedule() is the main scheduler function.
3185 * The main means of driving the scheduler and thus entering this function are:
3187 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3189 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3190 * paths. For example, see arch/x86/entry_64.S.
3192 * To drive preemption between tasks, the scheduler sets the flag in timer
3193 * interrupt handler scheduler_tick().
3195 * 3. Wakeups don't really cause entry into schedule(). They add a
3196 * task to the run-queue and that's it.
3198 * Now, if the new task added to the run-queue preempts the current
3199 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3200 * called on the nearest possible occasion:
3202 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3204 * - in syscall or exception context, at the next outmost
3205 * preempt_enable(). (this might be as soon as the wake_up()'s
3208 * - in IRQ context, return from interrupt-handler to
3209 * preemptible context
3211 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3214 * - cond_resched() call
3215 * - explicit schedule() call
3216 * - return from syscall or exception to user-space
3217 * - return from interrupt-handler to user-space
3219 * WARNING: must be called with preemption disabled!
3221 static void __sched notrace __schedule(bool preempt)
3223 struct task_struct *prev, *next;
3224 unsigned long *switch_count;
3229 cpu = smp_processor_id();
3231 rcu_note_context_switch();
3235 * do_exit() calls schedule() with preemption disabled as an exception;
3236 * however we must fix that up, otherwise the next task will see an
3237 * inconsistent (higher) preempt count.
3239 * It also avoids the below schedule_debug() test from complaining
3242 if (unlikely(prev->state == TASK_DEAD))
3243 preempt_enable_no_resched_notrace();
3245 schedule_debug(prev);
3247 if (sched_feat(HRTICK))
3251 * Make sure that signal_pending_state()->signal_pending() below
3252 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3253 * done by the caller to avoid the race with signal_wake_up().
3255 smp_mb__before_spinlock();
3256 raw_spin_lock_irq(&rq->lock);
3257 lockdep_pin_lock(&rq->lock);
3259 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3261 switch_count = &prev->nivcsw;
3262 if (!preempt && prev->state) {
3263 if (unlikely(signal_pending_state(prev->state, prev))) {
3264 prev->state = TASK_RUNNING;
3266 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3270 * If a worker went to sleep, notify and ask workqueue
3271 * whether it wants to wake up a task to maintain
3274 if (prev->flags & PF_WQ_WORKER) {
3275 struct task_struct *to_wakeup;
3277 to_wakeup = wq_worker_sleeping(prev, cpu);
3279 try_to_wake_up_local(to_wakeup);
3282 switch_count = &prev->nvcsw;
3285 if (task_on_rq_queued(prev))
3286 update_rq_clock(rq);
3288 next = pick_next_task(rq, prev);
3289 wallclock = walt_ktime_clock();
3290 walt_update_task_ravg(prev, rq, PUT_PREV_TASK, wallclock, 0);
3291 walt_update_task_ravg(next, rq, PICK_NEXT_TASK, wallclock, 0);
3292 clear_tsk_need_resched(prev);
3293 clear_preempt_need_resched();
3294 rq->clock_skip_update = 0;
3296 if (likely(prev != next)) {
3301 trace_sched_switch(preempt, prev, next);
3302 rq = context_switch(rq, prev, next); /* unlocks the rq */
3305 lockdep_unpin_lock(&rq->lock);
3306 raw_spin_unlock_irq(&rq->lock);
3309 balance_callback(rq);
3312 static inline void sched_submit_work(struct task_struct *tsk)
3314 if (!tsk->state || tsk_is_pi_blocked(tsk))
3317 * If we are going to sleep and we have plugged IO queued,
3318 * make sure to submit it to avoid deadlocks.
3320 if (blk_needs_flush_plug(tsk))
3321 blk_schedule_flush_plug(tsk);
3324 asmlinkage __visible void __sched schedule(void)
3326 struct task_struct *tsk = current;
3328 sched_submit_work(tsk);
3332 sched_preempt_enable_no_resched();
3333 } while (need_resched());
3335 EXPORT_SYMBOL(schedule);
3337 #ifdef CONFIG_CONTEXT_TRACKING
3338 asmlinkage __visible void __sched schedule_user(void)
3341 * If we come here after a random call to set_need_resched(),
3342 * or we have been woken up remotely but the IPI has not yet arrived,
3343 * we haven't yet exited the RCU idle mode. Do it here manually until
3344 * we find a better solution.
3346 * NB: There are buggy callers of this function. Ideally we
3347 * should warn if prev_state != CONTEXT_USER, but that will trigger
3348 * too frequently to make sense yet.
3350 enum ctx_state prev_state = exception_enter();
3352 exception_exit(prev_state);
3357 * schedule_preempt_disabled - called with preemption disabled
3359 * Returns with preemption disabled. Note: preempt_count must be 1
3361 void __sched schedule_preempt_disabled(void)
3363 sched_preempt_enable_no_resched();
3368 static void __sched notrace preempt_schedule_common(void)
3371 preempt_disable_notrace();
3373 preempt_enable_no_resched_notrace();
3376 * Check again in case we missed a preemption opportunity
3377 * between schedule and now.
3379 } while (need_resched());
3382 #ifdef CONFIG_PREEMPT
3384 * this is the entry point to schedule() from in-kernel preemption
3385 * off of preempt_enable. Kernel preemptions off return from interrupt
3386 * occur there and call schedule directly.
3388 asmlinkage __visible void __sched notrace preempt_schedule(void)
3391 * If there is a non-zero preempt_count or interrupts are disabled,
3392 * we do not want to preempt the current task. Just return..
3394 if (likely(!preemptible()))
3397 preempt_schedule_common();
3399 NOKPROBE_SYMBOL(preempt_schedule);
3400 EXPORT_SYMBOL(preempt_schedule);
3403 * preempt_schedule_notrace - preempt_schedule called by tracing
3405 * The tracing infrastructure uses preempt_enable_notrace to prevent
3406 * recursion and tracing preempt enabling caused by the tracing
3407 * infrastructure itself. But as tracing can happen in areas coming
3408 * from userspace or just about to enter userspace, a preempt enable
3409 * can occur before user_exit() is called. This will cause the scheduler
3410 * to be called when the system is still in usermode.
3412 * To prevent this, the preempt_enable_notrace will use this function
3413 * instead of preempt_schedule() to exit user context if needed before
3414 * calling the scheduler.
3416 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3418 enum ctx_state prev_ctx;
3420 if (likely(!preemptible()))
3424 preempt_disable_notrace();
3426 * Needs preempt disabled in case user_exit() is traced
3427 * and the tracer calls preempt_enable_notrace() causing
3428 * an infinite recursion.
3430 prev_ctx = exception_enter();
3432 exception_exit(prev_ctx);
3434 preempt_enable_no_resched_notrace();
3435 } while (need_resched());
3437 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3439 #endif /* CONFIG_PREEMPT */
3442 * this is the entry point to schedule() from kernel preemption
3443 * off of irq context.
3444 * Note, that this is called and return with irqs disabled. This will
3445 * protect us against recursive calling from irq.
3447 asmlinkage __visible void __sched preempt_schedule_irq(void)
3449 enum ctx_state prev_state;
3451 /* Catch callers which need to be fixed */
3452 BUG_ON(preempt_count() || !irqs_disabled());
3454 prev_state = exception_enter();
3460 local_irq_disable();
3461 sched_preempt_enable_no_resched();
3462 } while (need_resched());
3464 exception_exit(prev_state);
3467 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3470 return try_to_wake_up(curr->private, mode, wake_flags);
3472 EXPORT_SYMBOL(default_wake_function);
3474 #ifdef CONFIG_RT_MUTEXES
3477 * rt_mutex_setprio - set the current priority of a task
3479 * @prio: prio value (kernel-internal form)
3481 * This function changes the 'effective' priority of a task. It does
3482 * not touch ->normal_prio like __setscheduler().
3484 * Used by the rt_mutex code to implement priority inheritance
3485 * logic. Call site only calls if the priority of the task changed.
3487 void rt_mutex_setprio(struct task_struct *p, int prio)
3489 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3491 const struct sched_class *prev_class;
3493 BUG_ON(prio > MAX_PRIO);
3495 rq = __task_rq_lock(p);
3498 * Idle task boosting is a nono in general. There is one
3499 * exception, when PREEMPT_RT and NOHZ is active:
3501 * The idle task calls get_next_timer_interrupt() and holds
3502 * the timer wheel base->lock on the CPU and another CPU wants
3503 * to access the timer (probably to cancel it). We can safely
3504 * ignore the boosting request, as the idle CPU runs this code
3505 * with interrupts disabled and will complete the lock
3506 * protected section without being interrupted. So there is no
3507 * real need to boost.
3509 if (unlikely(p == rq->idle)) {
3510 WARN_ON(p != rq->curr);
3511 WARN_ON(p->pi_blocked_on);
3515 trace_sched_pi_setprio(p, prio);
3517 prev_class = p->sched_class;
3518 queued = task_on_rq_queued(p);
3519 running = task_current(rq, p);
3521 dequeue_task(rq, p, DEQUEUE_SAVE);
3523 put_prev_task(rq, p);
3526 * Boosting condition are:
3527 * 1. -rt task is running and holds mutex A
3528 * --> -dl task blocks on mutex A
3530 * 2. -dl task is running and holds mutex A
3531 * --> -dl task blocks on mutex A and could preempt the
3534 if (dl_prio(prio)) {
3535 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3536 if (!dl_prio(p->normal_prio) ||
3537 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3538 p->dl.dl_boosted = 1;
3539 enqueue_flag |= ENQUEUE_REPLENISH;
3541 p->dl.dl_boosted = 0;
3542 p->sched_class = &dl_sched_class;
3543 } else if (rt_prio(prio)) {
3544 if (dl_prio(oldprio))
3545 p->dl.dl_boosted = 0;
3547 enqueue_flag |= ENQUEUE_HEAD;
3548 p->sched_class = &rt_sched_class;
3550 if (dl_prio(oldprio))
3551 p->dl.dl_boosted = 0;
3552 if (rt_prio(oldprio))
3554 p->sched_class = &fair_sched_class;
3560 p->sched_class->set_curr_task(rq);
3562 enqueue_task(rq, p, enqueue_flag);
3564 check_class_changed(rq, p, prev_class, oldprio);
3566 preempt_disable(); /* avoid rq from going away on us */
3567 __task_rq_unlock(rq);
3569 balance_callback(rq);
3574 void set_user_nice(struct task_struct *p, long nice)
3576 int old_prio, delta, queued;
3577 unsigned long flags;
3580 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3583 * We have to be careful, if called from sys_setpriority(),
3584 * the task might be in the middle of scheduling on another CPU.
3586 rq = task_rq_lock(p, &flags);
3588 * The RT priorities are set via sched_setscheduler(), but we still
3589 * allow the 'normal' nice value to be set - but as expected
3590 * it wont have any effect on scheduling until the task is
3591 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3593 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3594 p->static_prio = NICE_TO_PRIO(nice);
3597 queued = task_on_rq_queued(p);
3599 dequeue_task(rq, p, DEQUEUE_SAVE);
3601 p->static_prio = NICE_TO_PRIO(nice);
3604 p->prio = effective_prio(p);
3605 delta = p->prio - old_prio;
3608 enqueue_task(rq, p, ENQUEUE_RESTORE);
3610 * If the task increased its priority or is running and
3611 * lowered its priority, then reschedule its CPU:
3613 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3617 task_rq_unlock(rq, p, &flags);
3619 EXPORT_SYMBOL(set_user_nice);
3622 * can_nice - check if a task can reduce its nice value
3626 int can_nice(const struct task_struct *p, const int nice)
3628 /* convert nice value [19,-20] to rlimit style value [1,40] */
3629 int nice_rlim = nice_to_rlimit(nice);
3631 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3632 capable(CAP_SYS_NICE));
3635 #ifdef __ARCH_WANT_SYS_NICE
3638 * sys_nice - change the priority of the current process.
3639 * @increment: priority increment
3641 * sys_setpriority is a more generic, but much slower function that
3642 * does similar things.
3644 SYSCALL_DEFINE1(nice, int, increment)
3649 * Setpriority might change our priority at the same moment.
3650 * We don't have to worry. Conceptually one call occurs first
3651 * and we have a single winner.
3653 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3654 nice = task_nice(current) + increment;
3656 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3657 if (increment < 0 && !can_nice(current, nice))
3660 retval = security_task_setnice(current, nice);
3664 set_user_nice(current, nice);
3671 * task_prio - return the priority value of a given task.
3672 * @p: the task in question.
3674 * Return: The priority value as seen by users in /proc.
3675 * RT tasks are offset by -200. Normal tasks are centered
3676 * around 0, value goes from -16 to +15.
3678 int task_prio(const struct task_struct *p)
3680 return p->prio - MAX_RT_PRIO;
3684 * idle_cpu - is a given cpu idle currently?
3685 * @cpu: the processor in question.
3687 * Return: 1 if the CPU is currently idle. 0 otherwise.
3689 int idle_cpu(int cpu)
3691 struct rq *rq = cpu_rq(cpu);
3693 if (rq->curr != rq->idle)
3700 if (!llist_empty(&rq->wake_list))
3708 * idle_task - return the idle task for a given cpu.
3709 * @cpu: the processor in question.
3711 * Return: The idle task for the cpu @cpu.
3713 struct task_struct *idle_task(int cpu)
3715 return cpu_rq(cpu)->idle;
3719 * find_process_by_pid - find a process with a matching PID value.
3720 * @pid: the pid in question.
3722 * The task of @pid, if found. %NULL otherwise.
3724 static struct task_struct *find_process_by_pid(pid_t pid)
3726 return pid ? find_task_by_vpid(pid) : current;
3730 * This function initializes the sched_dl_entity of a newly becoming
3731 * SCHED_DEADLINE task.
3733 * Only the static values are considered here, the actual runtime and the
3734 * absolute deadline will be properly calculated when the task is enqueued
3735 * for the first time with its new policy.
3738 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3740 struct sched_dl_entity *dl_se = &p->dl;
3742 dl_se->dl_runtime = attr->sched_runtime;
3743 dl_se->dl_deadline = attr->sched_deadline;
3744 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3745 dl_se->flags = attr->sched_flags;
3746 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3749 * Changing the parameters of a task is 'tricky' and we're not doing
3750 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3752 * What we SHOULD do is delay the bandwidth release until the 0-lag
3753 * point. This would include retaining the task_struct until that time
3754 * and change dl_overflow() to not immediately decrement the current
3757 * Instead we retain the current runtime/deadline and let the new
3758 * parameters take effect after the current reservation period lapses.
3759 * This is safe (albeit pessimistic) because the 0-lag point is always
3760 * before the current scheduling deadline.
3762 * We can still have temporary overloads because we do not delay the
3763 * change in bandwidth until that time; so admission control is
3764 * not on the safe side. It does however guarantee tasks will never
3765 * consume more than promised.
3770 * sched_setparam() passes in -1 for its policy, to let the functions
3771 * it calls know not to change it.
3773 #define SETPARAM_POLICY -1
3775 static void __setscheduler_params(struct task_struct *p,
3776 const struct sched_attr *attr)
3778 int policy = attr->sched_policy;
3780 if (policy == SETPARAM_POLICY)
3785 if (dl_policy(policy))
3786 __setparam_dl(p, attr);
3787 else if (fair_policy(policy))
3788 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3791 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3792 * !rt_policy. Always setting this ensures that things like
3793 * getparam()/getattr() don't report silly values for !rt tasks.
3795 p->rt_priority = attr->sched_priority;
3796 p->normal_prio = normal_prio(p);
3800 /* Actually do priority change: must hold pi & rq lock. */
3801 static void __setscheduler(struct rq *rq, struct task_struct *p,
3802 const struct sched_attr *attr, bool keep_boost)
3804 __setscheduler_params(p, attr);
3807 * Keep a potential priority boosting if called from
3808 * sched_setscheduler().
3811 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3813 p->prio = normal_prio(p);
3815 if (dl_prio(p->prio))
3816 p->sched_class = &dl_sched_class;
3817 else if (rt_prio(p->prio))
3818 p->sched_class = &rt_sched_class;
3820 p->sched_class = &fair_sched_class;
3824 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3826 struct sched_dl_entity *dl_se = &p->dl;
3828 attr->sched_priority = p->rt_priority;
3829 attr->sched_runtime = dl_se->dl_runtime;
3830 attr->sched_deadline = dl_se->dl_deadline;
3831 attr->sched_period = dl_se->dl_period;
3832 attr->sched_flags = dl_se->flags;
3836 * This function validates the new parameters of a -deadline task.
3837 * We ask for the deadline not being zero, and greater or equal
3838 * than the runtime, as well as the period of being zero or
3839 * greater than deadline. Furthermore, we have to be sure that
3840 * user parameters are above the internal resolution of 1us (we
3841 * check sched_runtime only since it is always the smaller one) and
3842 * below 2^63 ns (we have to check both sched_deadline and
3843 * sched_period, as the latter can be zero).
3846 __checkparam_dl(const struct sched_attr *attr)
3849 if (attr->sched_deadline == 0)
3853 * Since we truncate DL_SCALE bits, make sure we're at least
3856 if (attr->sched_runtime < (1ULL << DL_SCALE))
3860 * Since we use the MSB for wrap-around and sign issues, make
3861 * sure it's not set (mind that period can be equal to zero).
3863 if (attr->sched_deadline & (1ULL << 63) ||
3864 attr->sched_period & (1ULL << 63))
3867 /* runtime <= deadline <= period (if period != 0) */
3868 if ((attr->sched_period != 0 &&
3869 attr->sched_period < attr->sched_deadline) ||
3870 attr->sched_deadline < attr->sched_runtime)
3877 * check the target process has a UID that matches the current process's
3879 static bool check_same_owner(struct task_struct *p)
3881 const struct cred *cred = current_cred(), *pcred;
3885 pcred = __task_cred(p);
3886 match = (uid_eq(cred->euid, pcred->euid) ||
3887 uid_eq(cred->euid, pcred->uid));
3892 static bool dl_param_changed(struct task_struct *p,
3893 const struct sched_attr *attr)
3895 struct sched_dl_entity *dl_se = &p->dl;
3897 if (dl_se->dl_runtime != attr->sched_runtime ||
3898 dl_se->dl_deadline != attr->sched_deadline ||
3899 dl_se->dl_period != attr->sched_period ||
3900 dl_se->flags != attr->sched_flags)
3906 static int __sched_setscheduler(struct task_struct *p,
3907 const struct sched_attr *attr,
3910 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3911 MAX_RT_PRIO - 1 - attr->sched_priority;
3912 int retval, oldprio, oldpolicy = -1, queued, running;
3913 int new_effective_prio, policy = attr->sched_policy;
3914 unsigned long flags;
3915 const struct sched_class *prev_class;
3919 /* may grab non-irq protected spin_locks */
3920 BUG_ON(in_interrupt());
3922 /* double check policy once rq lock held */
3924 reset_on_fork = p->sched_reset_on_fork;
3925 policy = oldpolicy = p->policy;
3927 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3929 if (!valid_policy(policy))
3933 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3937 * Valid priorities for SCHED_FIFO and SCHED_RR are
3938 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3939 * SCHED_BATCH and SCHED_IDLE is 0.
3941 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3942 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3944 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3945 (rt_policy(policy) != (attr->sched_priority != 0)))
3949 * Allow unprivileged RT tasks to decrease priority:
3951 if (user && !capable(CAP_SYS_NICE)) {
3952 if (fair_policy(policy)) {
3953 if (attr->sched_nice < task_nice(p) &&
3954 !can_nice(p, attr->sched_nice))
3958 if (rt_policy(policy)) {
3959 unsigned long rlim_rtprio =
3960 task_rlimit(p, RLIMIT_RTPRIO);
3962 /* can't set/change the rt policy */
3963 if (policy != p->policy && !rlim_rtprio)
3966 /* can't increase priority */
3967 if (attr->sched_priority > p->rt_priority &&
3968 attr->sched_priority > rlim_rtprio)
3973 * Can't set/change SCHED_DEADLINE policy at all for now
3974 * (safest behavior); in the future we would like to allow
3975 * unprivileged DL tasks to increase their relative deadline
3976 * or reduce their runtime (both ways reducing utilization)
3978 if (dl_policy(policy))
3982 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3983 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3985 if (idle_policy(p->policy) && !idle_policy(policy)) {
3986 if (!can_nice(p, task_nice(p)))
3990 /* can't change other user's priorities */
3991 if (!check_same_owner(p))
3994 /* Normal users shall not reset the sched_reset_on_fork flag */
3995 if (p->sched_reset_on_fork && !reset_on_fork)
4000 retval = security_task_setscheduler(p);
4006 * make sure no PI-waiters arrive (or leave) while we are
4007 * changing the priority of the task:
4009 * To be able to change p->policy safely, the appropriate
4010 * runqueue lock must be held.
4012 rq = task_rq_lock(p, &flags);
4015 * Changing the policy of the stop threads its a very bad idea
4017 if (p == rq->stop) {
4018 task_rq_unlock(rq, p, &flags);
4023 * If not changing anything there's no need to proceed further,
4024 * but store a possible modification of reset_on_fork.
4026 if (unlikely(policy == p->policy)) {
4027 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4029 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4031 if (dl_policy(policy) && dl_param_changed(p, attr))
4034 p->sched_reset_on_fork = reset_on_fork;
4035 task_rq_unlock(rq, p, &flags);
4041 #ifdef CONFIG_RT_GROUP_SCHED
4043 * Do not allow realtime tasks into groups that have no runtime
4046 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4047 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4048 !task_group_is_autogroup(task_group(p))) {
4049 task_rq_unlock(rq, p, &flags);
4054 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4055 cpumask_t *span = rq->rd->span;
4058 * Don't allow tasks with an affinity mask smaller than
4059 * the entire root_domain to become SCHED_DEADLINE. We
4060 * will also fail if there's no bandwidth available.
4062 if (!cpumask_subset(span, &p->cpus_allowed) ||
4063 rq->rd->dl_bw.bw == 0) {
4064 task_rq_unlock(rq, p, &flags);
4071 /* recheck policy now with rq lock held */
4072 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4073 policy = oldpolicy = -1;
4074 task_rq_unlock(rq, p, &flags);
4079 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4080 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4083 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4084 task_rq_unlock(rq, p, &flags);
4088 p->sched_reset_on_fork = reset_on_fork;
4093 * Take priority boosted tasks into account. If the new
4094 * effective priority is unchanged, we just store the new
4095 * normal parameters and do not touch the scheduler class and
4096 * the runqueue. This will be done when the task deboost
4099 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4100 if (new_effective_prio == oldprio) {
4101 __setscheduler_params(p, attr);
4102 task_rq_unlock(rq, p, &flags);
4107 queued = task_on_rq_queued(p);
4108 running = task_current(rq, p);
4110 dequeue_task(rq, p, DEQUEUE_SAVE);
4112 put_prev_task(rq, p);
4114 prev_class = p->sched_class;
4115 __setscheduler(rq, p, attr, pi);
4118 p->sched_class->set_curr_task(rq);
4120 int enqueue_flags = ENQUEUE_RESTORE;
4122 * We enqueue to tail when the priority of a task is
4123 * increased (user space view).
4125 if (oldprio <= p->prio)
4126 enqueue_flags |= ENQUEUE_HEAD;
4128 enqueue_task(rq, p, enqueue_flags);
4131 check_class_changed(rq, p, prev_class, oldprio);
4132 preempt_disable(); /* avoid rq from going away on us */
4133 task_rq_unlock(rq, p, &flags);
4136 rt_mutex_adjust_pi(p);
4139 * Run balance callbacks after we've adjusted the PI chain.
4141 balance_callback(rq);
4147 static int _sched_setscheduler(struct task_struct *p, int policy,
4148 const struct sched_param *param, bool check)
4150 struct sched_attr attr = {
4151 .sched_policy = policy,
4152 .sched_priority = param->sched_priority,
4153 .sched_nice = PRIO_TO_NICE(p->static_prio),
4156 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4157 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4158 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4159 policy &= ~SCHED_RESET_ON_FORK;
4160 attr.sched_policy = policy;
4163 return __sched_setscheduler(p, &attr, check, true);
4166 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4167 * @p: the task in question.
4168 * @policy: new policy.
4169 * @param: structure containing the new RT priority.
4171 * Return: 0 on success. An error code otherwise.
4173 * NOTE that the task may be already dead.
4175 int sched_setscheduler(struct task_struct *p, int policy,
4176 const struct sched_param *param)
4178 return _sched_setscheduler(p, policy, param, true);
4180 EXPORT_SYMBOL_GPL(sched_setscheduler);
4182 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4184 return __sched_setscheduler(p, attr, true, true);
4186 EXPORT_SYMBOL_GPL(sched_setattr);
4189 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4190 * @p: the task in question.
4191 * @policy: new policy.
4192 * @param: structure containing the new RT priority.
4194 * Just like sched_setscheduler, only don't bother checking if the
4195 * current context has permission. For example, this is needed in
4196 * stop_machine(): we create temporary high priority worker threads,
4197 * but our caller might not have that capability.
4199 * Return: 0 on success. An error code otherwise.
4201 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4202 const struct sched_param *param)
4204 return _sched_setscheduler(p, policy, param, false);
4206 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4209 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4211 struct sched_param lparam;
4212 struct task_struct *p;
4215 if (!param || pid < 0)
4217 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4222 p = find_process_by_pid(pid);
4224 retval = sched_setscheduler(p, policy, &lparam);
4231 * Mimics kernel/events/core.c perf_copy_attr().
4233 static int sched_copy_attr(struct sched_attr __user *uattr,
4234 struct sched_attr *attr)
4239 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4243 * zero the full structure, so that a short copy will be nice.
4245 memset(attr, 0, sizeof(*attr));
4247 ret = get_user(size, &uattr->size);
4251 if (size > PAGE_SIZE) /* silly large */
4254 if (!size) /* abi compat */
4255 size = SCHED_ATTR_SIZE_VER0;
4257 if (size < SCHED_ATTR_SIZE_VER0)
4261 * If we're handed a bigger struct than we know of,
4262 * ensure all the unknown bits are 0 - i.e. new
4263 * user-space does not rely on any kernel feature
4264 * extensions we dont know about yet.
4266 if (size > sizeof(*attr)) {
4267 unsigned char __user *addr;
4268 unsigned char __user *end;
4271 addr = (void __user *)uattr + sizeof(*attr);
4272 end = (void __user *)uattr + size;
4274 for (; addr < end; addr++) {
4275 ret = get_user(val, addr);
4281 size = sizeof(*attr);
4284 ret = copy_from_user(attr, uattr, size);
4289 * XXX: do we want to be lenient like existing syscalls; or do we want
4290 * to be strict and return an error on out-of-bounds values?
4292 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4297 put_user(sizeof(*attr), &uattr->size);
4302 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4303 * @pid: the pid in question.
4304 * @policy: new policy.
4305 * @param: structure containing the new RT priority.
4307 * Return: 0 on success. An error code otherwise.
4309 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4310 struct sched_param __user *, param)
4312 /* negative values for policy are not valid */
4316 return do_sched_setscheduler(pid, policy, param);
4320 * sys_sched_setparam - set/change the RT priority of a thread
4321 * @pid: the pid in question.
4322 * @param: structure containing the new RT priority.
4324 * Return: 0 on success. An error code otherwise.
4326 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4328 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4332 * sys_sched_setattr - same as above, but with extended sched_attr
4333 * @pid: the pid in question.
4334 * @uattr: structure containing the extended parameters.
4335 * @flags: for future extension.
4337 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4338 unsigned int, flags)
4340 struct sched_attr attr;
4341 struct task_struct *p;
4344 if (!uattr || pid < 0 || flags)
4347 retval = sched_copy_attr(uattr, &attr);
4351 if ((int)attr.sched_policy < 0)
4356 p = find_process_by_pid(pid);
4358 retval = sched_setattr(p, &attr);
4365 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4366 * @pid: the pid in question.
4368 * Return: On success, the policy of the thread. Otherwise, a negative error
4371 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4373 struct task_struct *p;
4381 p = find_process_by_pid(pid);
4383 retval = security_task_getscheduler(p);
4386 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4393 * sys_sched_getparam - get the RT priority of a thread
4394 * @pid: the pid in question.
4395 * @param: structure containing the RT priority.
4397 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4400 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4402 struct sched_param lp = { .sched_priority = 0 };
4403 struct task_struct *p;
4406 if (!param || pid < 0)
4410 p = find_process_by_pid(pid);
4415 retval = security_task_getscheduler(p);
4419 if (task_has_rt_policy(p))
4420 lp.sched_priority = p->rt_priority;
4424 * This one might sleep, we cannot do it with a spinlock held ...
4426 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4435 static int sched_read_attr(struct sched_attr __user *uattr,
4436 struct sched_attr *attr,
4441 if (!access_ok(VERIFY_WRITE, uattr, usize))
4445 * If we're handed a smaller struct than we know of,
4446 * ensure all the unknown bits are 0 - i.e. old
4447 * user-space does not get uncomplete information.
4449 if (usize < sizeof(*attr)) {
4450 unsigned char *addr;
4453 addr = (void *)attr + usize;
4454 end = (void *)attr + sizeof(*attr);
4456 for (; addr < end; addr++) {
4464 ret = copy_to_user(uattr, attr, attr->size);
4472 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4473 * @pid: the pid in question.
4474 * @uattr: structure containing the extended parameters.
4475 * @size: sizeof(attr) for fwd/bwd comp.
4476 * @flags: for future extension.
4478 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4479 unsigned int, size, unsigned int, flags)
4481 struct sched_attr attr = {
4482 .size = sizeof(struct sched_attr),
4484 struct task_struct *p;
4487 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4488 size < SCHED_ATTR_SIZE_VER0 || flags)
4492 p = find_process_by_pid(pid);
4497 retval = security_task_getscheduler(p);
4501 attr.sched_policy = p->policy;
4502 if (p->sched_reset_on_fork)
4503 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4504 if (task_has_dl_policy(p))
4505 __getparam_dl(p, &attr);
4506 else if (task_has_rt_policy(p))
4507 attr.sched_priority = p->rt_priority;
4509 attr.sched_nice = task_nice(p);
4513 retval = sched_read_attr(uattr, &attr, size);
4521 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4523 cpumask_var_t cpus_allowed, new_mask;
4524 struct task_struct *p;
4529 p = find_process_by_pid(pid);
4535 /* Prevent p going away */
4539 if (p->flags & PF_NO_SETAFFINITY) {
4543 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4547 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4549 goto out_free_cpus_allowed;
4552 if (!check_same_owner(p)) {
4554 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4556 goto out_free_new_mask;
4561 retval = security_task_setscheduler(p);
4563 goto out_free_new_mask;
4566 cpuset_cpus_allowed(p, cpus_allowed);
4567 cpumask_and(new_mask, in_mask, cpus_allowed);
4570 * Since bandwidth control happens on root_domain basis,
4571 * if admission test is enabled, we only admit -deadline
4572 * tasks allowed to run on all the CPUs in the task's
4576 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4578 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4581 goto out_free_new_mask;
4587 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4590 cpuset_cpus_allowed(p, cpus_allowed);
4591 if (!cpumask_subset(new_mask, cpus_allowed)) {
4593 * We must have raced with a concurrent cpuset
4594 * update. Just reset the cpus_allowed to the
4595 * cpuset's cpus_allowed
4597 cpumask_copy(new_mask, cpus_allowed);
4602 free_cpumask_var(new_mask);
4603 out_free_cpus_allowed:
4604 free_cpumask_var(cpus_allowed);
4610 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4611 struct cpumask *new_mask)
4613 if (len < cpumask_size())
4614 cpumask_clear(new_mask);
4615 else if (len > cpumask_size())
4616 len = cpumask_size();
4618 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4622 * sys_sched_setaffinity - set the cpu affinity of a process
4623 * @pid: pid of the process
4624 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4625 * @user_mask_ptr: user-space pointer to the new cpu mask
4627 * Return: 0 on success. An error code otherwise.
4629 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4630 unsigned long __user *, user_mask_ptr)
4632 cpumask_var_t new_mask;
4635 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4638 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4640 retval = sched_setaffinity(pid, new_mask);
4641 free_cpumask_var(new_mask);
4645 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4647 struct task_struct *p;
4648 unsigned long flags;
4654 p = find_process_by_pid(pid);
4658 retval = security_task_getscheduler(p);
4662 raw_spin_lock_irqsave(&p->pi_lock, flags);
4663 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4664 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4673 * sys_sched_getaffinity - get the cpu affinity of a process
4674 * @pid: pid of the process
4675 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4676 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4678 * Return: 0 on success. An error code otherwise.
4680 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4681 unsigned long __user *, user_mask_ptr)
4686 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4688 if (len & (sizeof(unsigned long)-1))
4691 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4694 ret = sched_getaffinity(pid, mask);
4696 size_t retlen = min_t(size_t, len, cpumask_size());
4698 if (copy_to_user(user_mask_ptr, mask, retlen))
4703 free_cpumask_var(mask);
4709 * sys_sched_yield - yield the current processor to other threads.
4711 * This function yields the current CPU to other tasks. If there are no
4712 * other threads running on this CPU then this function will return.
4716 SYSCALL_DEFINE0(sched_yield)
4718 struct rq *rq = this_rq_lock();
4720 schedstat_inc(rq, yld_count);
4721 current->sched_class->yield_task(rq);
4724 * Since we are going to call schedule() anyway, there's
4725 * no need to preempt or enable interrupts:
4727 __release(rq->lock);
4728 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4729 do_raw_spin_unlock(&rq->lock);
4730 sched_preempt_enable_no_resched();
4737 int __sched _cond_resched(void)
4739 if (should_resched(0)) {
4740 preempt_schedule_common();
4745 EXPORT_SYMBOL(_cond_resched);
4748 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4749 * call schedule, and on return reacquire the lock.
4751 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4752 * operations here to prevent schedule() from being called twice (once via
4753 * spin_unlock(), once by hand).
4755 int __cond_resched_lock(spinlock_t *lock)
4757 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4760 lockdep_assert_held(lock);
4762 if (spin_needbreak(lock) || resched) {
4765 preempt_schedule_common();
4773 EXPORT_SYMBOL(__cond_resched_lock);
4775 int __sched __cond_resched_softirq(void)
4777 BUG_ON(!in_softirq());
4779 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4781 preempt_schedule_common();
4787 EXPORT_SYMBOL(__cond_resched_softirq);
4790 * yield - yield the current processor to other threads.
4792 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4794 * The scheduler is at all times free to pick the calling task as the most
4795 * eligible task to run, if removing the yield() call from your code breaks
4796 * it, its already broken.
4798 * Typical broken usage is:
4803 * where one assumes that yield() will let 'the other' process run that will
4804 * make event true. If the current task is a SCHED_FIFO task that will never
4805 * happen. Never use yield() as a progress guarantee!!
4807 * If you want to use yield() to wait for something, use wait_event().
4808 * If you want to use yield() to be 'nice' for others, use cond_resched().
4809 * If you still want to use yield(), do not!
4811 void __sched yield(void)
4813 set_current_state(TASK_RUNNING);
4816 EXPORT_SYMBOL(yield);
4819 * yield_to - yield the current processor to another thread in
4820 * your thread group, or accelerate that thread toward the
4821 * processor it's on.
4823 * @preempt: whether task preemption is allowed or not
4825 * It's the caller's job to ensure that the target task struct
4826 * can't go away on us before we can do any checks.
4829 * true (>0) if we indeed boosted the target task.
4830 * false (0) if we failed to boost the target.
4831 * -ESRCH if there's no task to yield to.
4833 int __sched yield_to(struct task_struct *p, bool preempt)
4835 struct task_struct *curr = current;
4836 struct rq *rq, *p_rq;
4837 unsigned long flags;
4840 local_irq_save(flags);
4846 * If we're the only runnable task on the rq and target rq also
4847 * has only one task, there's absolutely no point in yielding.
4849 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4854 double_rq_lock(rq, p_rq);
4855 if (task_rq(p) != p_rq) {
4856 double_rq_unlock(rq, p_rq);
4860 if (!curr->sched_class->yield_to_task)
4863 if (curr->sched_class != p->sched_class)
4866 if (task_running(p_rq, p) || p->state)
4869 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4871 schedstat_inc(rq, yld_count);
4873 * Make p's CPU reschedule; pick_next_entity takes care of
4876 if (preempt && rq != p_rq)
4881 double_rq_unlock(rq, p_rq);
4883 local_irq_restore(flags);
4890 EXPORT_SYMBOL_GPL(yield_to);
4893 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4894 * that process accounting knows that this is a task in IO wait state.
4896 long __sched io_schedule_timeout(long timeout)
4898 int old_iowait = current->in_iowait;
4902 current->in_iowait = 1;
4903 blk_schedule_flush_plug(current);
4905 delayacct_blkio_start();
4907 atomic_inc(&rq->nr_iowait);
4908 ret = schedule_timeout(timeout);
4909 current->in_iowait = old_iowait;
4910 atomic_dec(&rq->nr_iowait);
4911 delayacct_blkio_end();
4915 EXPORT_SYMBOL(io_schedule_timeout);
4918 * sys_sched_get_priority_max - return maximum RT priority.
4919 * @policy: scheduling class.
4921 * Return: On success, this syscall returns the maximum
4922 * rt_priority that can be used by a given scheduling class.
4923 * On failure, a negative error code is returned.
4925 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4932 ret = MAX_USER_RT_PRIO-1;
4934 case SCHED_DEADLINE:
4945 * sys_sched_get_priority_min - return minimum RT priority.
4946 * @policy: scheduling class.
4948 * Return: On success, this syscall returns the minimum
4949 * rt_priority that can be used by a given scheduling class.
4950 * On failure, a negative error code is returned.
4952 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4961 case SCHED_DEADLINE:
4971 * sys_sched_rr_get_interval - return the default timeslice of a process.
4972 * @pid: pid of the process.
4973 * @interval: userspace pointer to the timeslice value.
4975 * this syscall writes the default timeslice value of a given process
4976 * into the user-space timespec buffer. A value of '0' means infinity.
4978 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4981 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4982 struct timespec __user *, interval)
4984 struct task_struct *p;
4985 unsigned int time_slice;
4986 unsigned long flags;
4996 p = find_process_by_pid(pid);
5000 retval = security_task_getscheduler(p);
5004 rq = task_rq_lock(p, &flags);
5006 if (p->sched_class->get_rr_interval)
5007 time_slice = p->sched_class->get_rr_interval(rq, p);
5008 task_rq_unlock(rq, p, &flags);
5011 jiffies_to_timespec(time_slice, &t);
5012 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5020 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5022 void sched_show_task(struct task_struct *p)
5024 unsigned long free = 0;
5026 unsigned long state = p->state;
5029 state = __ffs(state) + 1;
5030 printk(KERN_INFO "%-15.15s %c", p->comm,
5031 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5032 #if BITS_PER_LONG == 32
5033 if (state == TASK_RUNNING)
5034 printk(KERN_CONT " running ");
5036 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5038 if (state == TASK_RUNNING)
5039 printk(KERN_CONT " running task ");
5041 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5043 #ifdef CONFIG_DEBUG_STACK_USAGE
5044 free = stack_not_used(p);
5049 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5051 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5052 task_pid_nr(p), ppid,
5053 (unsigned long)task_thread_info(p)->flags);
5055 print_worker_info(KERN_INFO, p);
5056 show_stack(p, NULL);
5059 void show_state_filter(unsigned long state_filter)
5061 struct task_struct *g, *p;
5063 #if BITS_PER_LONG == 32
5065 " task PC stack pid father\n");
5068 " task PC stack pid father\n");
5071 for_each_process_thread(g, p) {
5073 * reset the NMI-timeout, listing all files on a slow
5074 * console might take a lot of time:
5075 * Also, reset softlockup watchdogs on all CPUs, because
5076 * another CPU might be blocked waiting for us to process
5079 touch_nmi_watchdog();
5080 touch_all_softlockup_watchdogs();
5081 if (!state_filter || (p->state & state_filter))
5085 #ifdef CONFIG_SCHED_DEBUG
5086 sysrq_sched_debug_show();
5090 * Only show locks if all tasks are dumped:
5093 debug_show_all_locks();
5096 void init_idle_bootup_task(struct task_struct *idle)
5098 idle->sched_class = &idle_sched_class;
5102 * init_idle - set up an idle thread for a given CPU
5103 * @idle: task in question
5104 * @cpu: cpu the idle task belongs to
5106 * NOTE: this function does not set the idle thread's NEED_RESCHED
5107 * flag, to make booting more robust.
5109 void init_idle(struct task_struct *idle, int cpu)
5111 struct rq *rq = cpu_rq(cpu);
5112 unsigned long flags;
5114 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5115 raw_spin_lock(&rq->lock);
5117 __sched_fork(0, idle);
5118 idle->state = TASK_RUNNING;
5119 idle->se.exec_start = sched_clock();
5123 * Its possible that init_idle() gets called multiple times on a task,
5124 * in that case do_set_cpus_allowed() will not do the right thing.
5126 * And since this is boot we can forgo the serialization.
5128 set_cpus_allowed_common(idle, cpumask_of(cpu));
5131 * We're having a chicken and egg problem, even though we are
5132 * holding rq->lock, the cpu isn't yet set to this cpu so the
5133 * lockdep check in task_group() will fail.
5135 * Similar case to sched_fork(). / Alternatively we could
5136 * use task_rq_lock() here and obtain the other rq->lock.
5141 __set_task_cpu(idle, cpu);
5144 rq->curr = rq->idle = idle;
5145 idle->on_rq = TASK_ON_RQ_QUEUED;
5149 raw_spin_unlock(&rq->lock);
5150 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5152 /* Set the preempt count _outside_ the spinlocks! */
5153 init_idle_preempt_count(idle, cpu);
5156 * The idle tasks have their own, simple scheduling class:
5158 idle->sched_class = &idle_sched_class;
5159 ftrace_graph_init_idle_task(idle, cpu);
5160 vtime_init_idle(idle, cpu);
5162 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5166 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5167 const struct cpumask *trial)
5169 int ret = 1, trial_cpus;
5170 struct dl_bw *cur_dl_b;
5171 unsigned long flags;
5173 if (!cpumask_weight(cur))
5176 rcu_read_lock_sched();
5177 cur_dl_b = dl_bw_of(cpumask_any(cur));
5178 trial_cpus = cpumask_weight(trial);
5180 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5181 if (cur_dl_b->bw != -1 &&
5182 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5184 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5185 rcu_read_unlock_sched();
5190 int task_can_attach(struct task_struct *p,
5191 const struct cpumask *cs_cpus_allowed)
5196 * Kthreads which disallow setaffinity shouldn't be moved
5197 * to a new cpuset; we don't want to change their cpu
5198 * affinity and isolating such threads by their set of
5199 * allowed nodes is unnecessary. Thus, cpusets are not
5200 * applicable for such threads. This prevents checking for
5201 * success of set_cpus_allowed_ptr() on all attached tasks
5202 * before cpus_allowed may be changed.
5204 if (p->flags & PF_NO_SETAFFINITY) {
5210 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5212 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5217 unsigned long flags;
5219 rcu_read_lock_sched();
5220 dl_b = dl_bw_of(dest_cpu);
5221 raw_spin_lock_irqsave(&dl_b->lock, flags);
5222 cpus = dl_bw_cpus(dest_cpu);
5223 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5228 * We reserve space for this task in the destination
5229 * root_domain, as we can't fail after this point.
5230 * We will free resources in the source root_domain
5231 * later on (see set_cpus_allowed_dl()).
5233 __dl_add(dl_b, p->dl.dl_bw);
5235 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5236 rcu_read_unlock_sched();
5246 #ifdef CONFIG_NUMA_BALANCING
5247 /* Migrate current task p to target_cpu */
5248 int migrate_task_to(struct task_struct *p, int target_cpu)
5250 struct migration_arg arg = { p, target_cpu };
5251 int curr_cpu = task_cpu(p);
5253 if (curr_cpu == target_cpu)
5256 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5259 /* TODO: This is not properly updating schedstats */
5261 trace_sched_move_numa(p, curr_cpu, target_cpu);
5262 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5266 * Requeue a task on a given node and accurately track the number of NUMA
5267 * tasks on the runqueues
5269 void sched_setnuma(struct task_struct *p, int nid)
5272 unsigned long flags;
5273 bool queued, running;
5275 rq = task_rq_lock(p, &flags);
5276 queued = task_on_rq_queued(p);
5277 running = task_current(rq, p);
5280 dequeue_task(rq, p, DEQUEUE_SAVE);
5282 put_prev_task(rq, p);
5284 p->numa_preferred_nid = nid;
5287 p->sched_class->set_curr_task(rq);
5289 enqueue_task(rq, p, ENQUEUE_RESTORE);
5290 task_rq_unlock(rq, p, &flags);
5292 #endif /* CONFIG_NUMA_BALANCING */
5294 #ifdef CONFIG_HOTPLUG_CPU
5296 * Ensures that the idle task is using init_mm right before its cpu goes
5299 void idle_task_exit(void)
5301 struct mm_struct *mm = current->active_mm;
5303 BUG_ON(cpu_online(smp_processor_id()));
5305 if (mm != &init_mm) {
5306 switch_mm(mm, &init_mm, current);
5307 finish_arch_post_lock_switch();
5313 * Since this CPU is going 'away' for a while, fold any nr_active delta
5314 * we might have. Assumes we're called after migrate_tasks() so that the
5315 * nr_active count is stable.
5317 * Also see the comment "Global load-average calculations".
5319 static void calc_load_migrate(struct rq *rq)
5321 long delta = calc_load_fold_active(rq);
5323 atomic_long_add(delta, &calc_load_tasks);
5326 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5330 static const struct sched_class fake_sched_class = {
5331 .put_prev_task = put_prev_task_fake,
5334 static struct task_struct fake_task = {
5336 * Avoid pull_{rt,dl}_task()
5338 .prio = MAX_PRIO + 1,
5339 .sched_class = &fake_sched_class,
5343 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5344 * try_to_wake_up()->select_task_rq().
5346 * Called with rq->lock held even though we'er in stop_machine() and
5347 * there's no concurrency possible, we hold the required locks anyway
5348 * because of lock validation efforts.
5350 static void migrate_tasks(struct rq *dead_rq)
5352 struct rq *rq = dead_rq;
5353 struct task_struct *next, *stop = rq->stop;
5357 * Fudge the rq selection such that the below task selection loop
5358 * doesn't get stuck on the currently eligible stop task.
5360 * We're currently inside stop_machine() and the rq is either stuck
5361 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5362 * either way we should never end up calling schedule() until we're
5368 * put_prev_task() and pick_next_task() sched
5369 * class method both need to have an up-to-date
5370 * value of rq->clock[_task]
5372 update_rq_clock(rq);
5376 * There's this thread running, bail when that's the only
5379 if (rq->nr_running == 1)
5383 * pick_next_task assumes pinned rq->lock.
5385 lockdep_pin_lock(&rq->lock);
5386 next = pick_next_task(rq, &fake_task);
5388 next->sched_class->put_prev_task(rq, next);
5391 * Rules for changing task_struct::cpus_allowed are holding
5392 * both pi_lock and rq->lock, such that holding either
5393 * stabilizes the mask.
5395 * Drop rq->lock is not quite as disastrous as it usually is
5396 * because !cpu_active at this point, which means load-balance
5397 * will not interfere. Also, stop-machine.
5399 lockdep_unpin_lock(&rq->lock);
5400 raw_spin_unlock(&rq->lock);
5401 raw_spin_lock(&next->pi_lock);
5402 raw_spin_lock(&rq->lock);
5405 * Since we're inside stop-machine, _nothing_ should have
5406 * changed the task, WARN if weird stuff happened, because in
5407 * that case the above rq->lock drop is a fail too.
5409 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5410 raw_spin_unlock(&next->pi_lock);
5414 /* Find suitable destination for @next, with force if needed. */
5415 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5417 rq = __migrate_task(rq, next, dest_cpu);
5418 if (rq != dead_rq) {
5419 raw_spin_unlock(&rq->lock);
5421 raw_spin_lock(&rq->lock);
5423 raw_spin_unlock(&next->pi_lock);
5428 #endif /* CONFIG_HOTPLUG_CPU */
5430 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5432 static struct ctl_table sd_ctl_dir[] = {
5434 .procname = "sched_domain",
5440 static struct ctl_table sd_ctl_root[] = {
5442 .procname = "kernel",
5444 .child = sd_ctl_dir,
5449 static struct ctl_table *sd_alloc_ctl_entry(int n)
5451 struct ctl_table *entry =
5452 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5457 static void sd_free_ctl_entry(struct ctl_table **tablep)
5459 struct ctl_table *entry;
5462 * In the intermediate directories, both the child directory and
5463 * procname are dynamically allocated and could fail but the mode
5464 * will always be set. In the lowest directory the names are
5465 * static strings and all have proc handlers.
5467 for (entry = *tablep; entry->mode; entry++) {
5469 sd_free_ctl_entry(&entry->child);
5470 if (entry->proc_handler == NULL)
5471 kfree(entry->procname);
5478 static int min_load_idx = 0;
5479 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5482 set_table_entry(struct ctl_table *entry,
5483 const char *procname, void *data, int maxlen,
5484 umode_t mode, proc_handler *proc_handler,
5487 entry->procname = procname;
5489 entry->maxlen = maxlen;
5491 entry->proc_handler = proc_handler;
5494 entry->extra1 = &min_load_idx;
5495 entry->extra2 = &max_load_idx;
5499 static struct ctl_table *
5500 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5502 struct ctl_table *table = sd_alloc_ctl_entry(5);
5507 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5508 sizeof(int), 0644, proc_dointvec_minmax, false);
5509 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5510 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5511 proc_doulongvec_minmax, false);
5512 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5513 sizeof(int), 0644, proc_dointvec_minmax, false);
5514 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5515 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5516 proc_doulongvec_minmax, false);
5521 static struct ctl_table *
5522 sd_alloc_ctl_group_table(struct sched_group *sg)
5524 struct ctl_table *table = sd_alloc_ctl_entry(2);
5529 table->procname = kstrdup("energy", GFP_KERNEL);
5531 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5536 static struct ctl_table *
5537 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5539 struct ctl_table *table;
5540 unsigned int nr_entries = 14;
5543 struct sched_group *sg = sd->groups;
5548 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5550 nr_entries += nr_sgs;
5553 table = sd_alloc_ctl_entry(nr_entries);
5558 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5559 sizeof(long), 0644, proc_doulongvec_minmax, false);
5560 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5561 sizeof(long), 0644, proc_doulongvec_minmax, false);
5562 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5563 sizeof(int), 0644, proc_dointvec_minmax, true);
5564 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5565 sizeof(int), 0644, proc_dointvec_minmax, true);
5566 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5567 sizeof(int), 0644, proc_dointvec_minmax, true);
5568 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5569 sizeof(int), 0644, proc_dointvec_minmax, true);
5570 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5571 sizeof(int), 0644, proc_dointvec_minmax, true);
5572 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5573 sizeof(int), 0644, proc_dointvec_minmax, false);
5574 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5575 sizeof(int), 0644, proc_dointvec_minmax, false);
5576 set_table_entry(&table[9], "cache_nice_tries",
5577 &sd->cache_nice_tries,
5578 sizeof(int), 0644, proc_dointvec_minmax, false);
5579 set_table_entry(&table[10], "flags", &sd->flags,
5580 sizeof(int), 0644, proc_dointvec_minmax, false);
5581 set_table_entry(&table[11], "max_newidle_lb_cost",
5582 &sd->max_newidle_lb_cost,
5583 sizeof(long), 0644, proc_doulongvec_minmax, false);
5584 set_table_entry(&table[12], "name", sd->name,
5585 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5589 struct ctl_table *entry = &table[13];
5592 snprintf(buf, 32, "group%d", i);
5593 entry->procname = kstrdup(buf, GFP_KERNEL);
5595 entry->child = sd_alloc_ctl_group_table(sg);
5596 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5598 /* &table[nr_entries-1] is terminator */
5603 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5605 struct ctl_table *entry, *table;
5606 struct sched_domain *sd;
5607 int domain_num = 0, i;
5610 for_each_domain(cpu, sd)
5612 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5617 for_each_domain(cpu, sd) {
5618 snprintf(buf, 32, "domain%d", i);
5619 entry->procname = kstrdup(buf, GFP_KERNEL);
5621 entry->child = sd_alloc_ctl_domain_table(sd);
5628 static struct ctl_table_header *sd_sysctl_header;
5629 static void register_sched_domain_sysctl(void)
5631 int i, cpu_num = num_possible_cpus();
5632 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5635 WARN_ON(sd_ctl_dir[0].child);
5636 sd_ctl_dir[0].child = entry;
5641 for_each_possible_cpu(i) {
5642 snprintf(buf, 32, "cpu%d", i);
5643 entry->procname = kstrdup(buf, GFP_KERNEL);
5645 entry->child = sd_alloc_ctl_cpu_table(i);
5649 WARN_ON(sd_sysctl_header);
5650 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5653 /* may be called multiple times per register */
5654 static void unregister_sched_domain_sysctl(void)
5656 unregister_sysctl_table(sd_sysctl_header);
5657 sd_sysctl_header = NULL;
5658 if (sd_ctl_dir[0].child)
5659 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5662 static void register_sched_domain_sysctl(void)
5665 static void unregister_sched_domain_sysctl(void)
5668 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5670 static void set_rq_online(struct rq *rq)
5673 const struct sched_class *class;
5675 cpumask_set_cpu(rq->cpu, rq->rd->online);
5678 for_each_class(class) {
5679 if (class->rq_online)
5680 class->rq_online(rq);
5685 static void set_rq_offline(struct rq *rq)
5688 const struct sched_class *class;
5690 for_each_class(class) {
5691 if (class->rq_offline)
5692 class->rq_offline(rq);
5695 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5701 * migration_call - callback that gets triggered when a CPU is added.
5702 * Here we can start up the necessary migration thread for the new CPU.
5705 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5707 int cpu = (long)hcpu;
5708 unsigned long flags;
5709 struct rq *rq = cpu_rq(cpu);
5711 switch (action & ~CPU_TASKS_FROZEN) {
5713 case CPU_UP_PREPARE:
5714 raw_spin_lock_irqsave(&rq->lock, flags);
5715 walt_set_window_start(rq);
5716 raw_spin_unlock_irqrestore(&rq->lock, flags);
5717 rq->calc_load_update = calc_load_update;
5718 account_reset_rq(rq);
5722 /* Update our root-domain */
5723 raw_spin_lock_irqsave(&rq->lock, flags);
5725 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5729 raw_spin_unlock_irqrestore(&rq->lock, flags);
5732 #ifdef CONFIG_HOTPLUG_CPU
5734 sched_ttwu_pending();
5735 /* Update our root-domain */
5736 raw_spin_lock_irqsave(&rq->lock, flags);
5737 walt_migrate_sync_cpu(cpu);
5739 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5743 BUG_ON(rq->nr_running != 1); /* the migration thread */
5744 raw_spin_unlock_irqrestore(&rq->lock, flags);
5748 calc_load_migrate(rq);
5753 update_max_interval();
5759 * Register at high priority so that task migration (migrate_all_tasks)
5760 * happens before everything else. This has to be lower priority than
5761 * the notifier in the perf_event subsystem, though.
5763 static struct notifier_block migration_notifier = {
5764 .notifier_call = migration_call,
5765 .priority = CPU_PRI_MIGRATION,
5768 static void set_cpu_rq_start_time(void)
5770 int cpu = smp_processor_id();
5771 struct rq *rq = cpu_rq(cpu);
5772 rq->age_stamp = sched_clock_cpu(cpu);
5775 static int sched_cpu_active(struct notifier_block *nfb,
5776 unsigned long action, void *hcpu)
5778 int cpu = (long)hcpu;
5780 switch (action & ~CPU_TASKS_FROZEN) {
5782 set_cpu_rq_start_time();
5787 * At this point a starting CPU has marked itself as online via
5788 * set_cpu_online(). But it might not yet have marked itself
5789 * as active, which is essential from here on.
5791 set_cpu_active(cpu, true);
5792 stop_machine_unpark(cpu);
5795 case CPU_DOWN_FAILED:
5796 set_cpu_active(cpu, true);
5804 static int sched_cpu_inactive(struct notifier_block *nfb,
5805 unsigned long action, void *hcpu)
5807 switch (action & ~CPU_TASKS_FROZEN) {
5808 case CPU_DOWN_PREPARE:
5809 set_cpu_active((long)hcpu, false);
5816 static int __init migration_init(void)
5818 void *cpu = (void *)(long)smp_processor_id();
5821 /* Initialize migration for the boot CPU */
5822 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5823 BUG_ON(err == NOTIFY_BAD);
5824 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5825 register_cpu_notifier(&migration_notifier);
5827 /* Register cpu active notifiers */
5828 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5829 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5833 early_initcall(migration_init);
5835 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5837 #ifdef CONFIG_SCHED_DEBUG
5839 static __read_mostly int sched_debug_enabled;
5841 static int __init sched_debug_setup(char *str)
5843 sched_debug_enabled = 1;
5847 early_param("sched_debug", sched_debug_setup);
5849 static inline bool sched_debug(void)
5851 return sched_debug_enabled;
5854 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5855 struct cpumask *groupmask)
5857 struct sched_group *group = sd->groups;
5859 cpumask_clear(groupmask);
5861 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5863 if (!(sd->flags & SD_LOAD_BALANCE)) {
5864 printk("does not load-balance\n");
5866 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5871 printk(KERN_CONT "span %*pbl level %s\n",
5872 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5874 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5875 printk(KERN_ERR "ERROR: domain->span does not contain "
5878 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5879 printk(KERN_ERR "ERROR: domain->groups does not contain"
5883 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5887 printk(KERN_ERR "ERROR: group is NULL\n");
5891 if (!cpumask_weight(sched_group_cpus(group))) {
5892 printk(KERN_CONT "\n");
5893 printk(KERN_ERR "ERROR: empty group\n");
5897 if (!(sd->flags & SD_OVERLAP) &&
5898 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5899 printk(KERN_CONT "\n");
5900 printk(KERN_ERR "ERROR: repeated CPUs\n");
5904 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5906 printk(KERN_CONT " %*pbl",
5907 cpumask_pr_args(sched_group_cpus(group)));
5908 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5909 printk(KERN_CONT " (cpu_capacity = %lu)",
5910 group->sgc->capacity);
5913 group = group->next;
5914 } while (group != sd->groups);
5915 printk(KERN_CONT "\n");
5917 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5918 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5921 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5922 printk(KERN_ERR "ERROR: parent span is not a superset "
5923 "of domain->span\n");
5927 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5931 if (!sched_debug_enabled)
5935 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5939 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5942 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5950 #else /* !CONFIG_SCHED_DEBUG */
5951 # define sched_domain_debug(sd, cpu) do { } while (0)
5952 static inline bool sched_debug(void)
5956 #endif /* CONFIG_SCHED_DEBUG */
5958 static int sd_degenerate(struct sched_domain *sd)
5960 if (cpumask_weight(sched_domain_span(sd)) == 1)
5963 /* Following flags need at least 2 groups */
5964 if (sd->flags & (SD_LOAD_BALANCE |
5965 SD_BALANCE_NEWIDLE |
5968 SD_SHARE_CPUCAPACITY |
5969 SD_SHARE_PKG_RESOURCES |
5970 SD_SHARE_POWERDOMAIN |
5971 SD_SHARE_CAP_STATES)) {
5972 if (sd->groups != sd->groups->next)
5976 /* Following flags don't use groups */
5977 if (sd->flags & (SD_WAKE_AFFINE))
5984 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5986 unsigned long cflags = sd->flags, pflags = parent->flags;
5988 if (sd_degenerate(parent))
5991 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5994 /* Flags needing groups don't count if only 1 group in parent */
5995 if (parent->groups == parent->groups->next) {
5996 pflags &= ~(SD_LOAD_BALANCE |
5997 SD_BALANCE_NEWIDLE |
6000 SD_SHARE_CPUCAPACITY |
6001 SD_SHARE_PKG_RESOURCES |
6003 SD_SHARE_POWERDOMAIN |
6004 SD_SHARE_CAP_STATES);
6005 if (nr_node_ids == 1)
6006 pflags &= ~SD_SERIALIZE;
6008 if (~cflags & pflags)
6014 static void free_rootdomain(struct rcu_head *rcu)
6016 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6018 cpupri_cleanup(&rd->cpupri);
6019 cpudl_cleanup(&rd->cpudl);
6020 free_cpumask_var(rd->dlo_mask);
6021 free_cpumask_var(rd->rto_mask);
6022 free_cpumask_var(rd->online);
6023 free_cpumask_var(rd->span);
6027 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6029 struct root_domain *old_rd = NULL;
6030 unsigned long flags;
6032 raw_spin_lock_irqsave(&rq->lock, flags);
6037 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6040 cpumask_clear_cpu(rq->cpu, old_rd->span);
6043 * If we dont want to free the old_rd yet then
6044 * set old_rd to NULL to skip the freeing later
6047 if (!atomic_dec_and_test(&old_rd->refcount))
6051 atomic_inc(&rd->refcount);
6054 cpumask_set_cpu(rq->cpu, rd->span);
6055 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6058 raw_spin_unlock_irqrestore(&rq->lock, flags);
6061 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6064 static int init_rootdomain(struct root_domain *rd)
6066 memset(rd, 0, sizeof(*rd));
6068 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6070 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6072 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6074 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6077 init_dl_bw(&rd->dl_bw);
6078 if (cpudl_init(&rd->cpudl) != 0)
6081 if (cpupri_init(&rd->cpupri) != 0)
6084 init_max_cpu_capacity(&rd->max_cpu_capacity);
6088 free_cpumask_var(rd->rto_mask);
6090 free_cpumask_var(rd->dlo_mask);
6092 free_cpumask_var(rd->online);
6094 free_cpumask_var(rd->span);
6100 * By default the system creates a single root-domain with all cpus as
6101 * members (mimicking the global state we have today).
6103 struct root_domain def_root_domain;
6105 static void init_defrootdomain(void)
6107 init_rootdomain(&def_root_domain);
6109 atomic_set(&def_root_domain.refcount, 1);
6112 static struct root_domain *alloc_rootdomain(void)
6114 struct root_domain *rd;
6116 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6120 if (init_rootdomain(rd) != 0) {
6128 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6130 struct sched_group *tmp, *first;
6139 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6144 } while (sg != first);
6147 static void free_sched_domain(struct rcu_head *rcu)
6149 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6152 * If its an overlapping domain it has private groups, iterate and
6155 if (sd->flags & SD_OVERLAP) {
6156 free_sched_groups(sd->groups, 1);
6157 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6158 kfree(sd->groups->sgc);
6164 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6166 call_rcu(&sd->rcu, free_sched_domain);
6169 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6171 for (; sd; sd = sd->parent)
6172 destroy_sched_domain(sd, cpu);
6176 * Keep a special pointer to the highest sched_domain that has
6177 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6178 * allows us to avoid some pointer chasing select_idle_sibling().
6180 * Also keep a unique ID per domain (we use the first cpu number in
6181 * the cpumask of the domain), this allows us to quickly tell if
6182 * two cpus are in the same cache domain, see cpus_share_cache().
6184 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6185 DEFINE_PER_CPU(int, sd_llc_size);
6186 DEFINE_PER_CPU(int, sd_llc_id);
6187 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6188 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6189 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6190 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6191 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6193 static void update_top_cache_domain(int cpu)
6195 struct sched_domain *sd;
6196 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6200 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6202 id = cpumask_first(sched_domain_span(sd));
6203 size = cpumask_weight(sched_domain_span(sd));
6204 busy_sd = sd->parent; /* sd_busy */
6206 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6208 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6209 per_cpu(sd_llc_size, cpu) = size;
6210 per_cpu(sd_llc_id, cpu) = id;
6212 sd = lowest_flag_domain(cpu, SD_NUMA);
6213 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6215 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6216 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6218 for_each_domain(cpu, sd) {
6219 if (sd->groups->sge)
6224 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6226 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6227 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6231 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6232 * hold the hotplug lock.
6235 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6237 struct rq *rq = cpu_rq(cpu);
6238 struct sched_domain *tmp;
6240 /* Remove the sched domains which do not contribute to scheduling. */
6241 for (tmp = sd; tmp; ) {
6242 struct sched_domain *parent = tmp->parent;
6246 if (sd_parent_degenerate(tmp, parent)) {
6247 tmp->parent = parent->parent;
6249 parent->parent->child = tmp;
6251 * Transfer SD_PREFER_SIBLING down in case of a
6252 * degenerate parent; the spans match for this
6253 * so the property transfers.
6255 if (parent->flags & SD_PREFER_SIBLING)
6256 tmp->flags |= SD_PREFER_SIBLING;
6257 destroy_sched_domain(parent, cpu);
6262 if (sd && sd_degenerate(sd)) {
6265 destroy_sched_domain(tmp, cpu);
6270 sched_domain_debug(sd, cpu);
6272 rq_attach_root(rq, rd);
6274 rcu_assign_pointer(rq->sd, sd);
6275 destroy_sched_domains(tmp, cpu);
6277 update_top_cache_domain(cpu);
6280 /* Setup the mask of cpus configured for isolated domains */
6281 static int __init isolated_cpu_setup(char *str)
6283 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6284 cpulist_parse(str, cpu_isolated_map);
6288 __setup("isolcpus=", isolated_cpu_setup);
6291 struct sched_domain ** __percpu sd;
6292 struct root_domain *rd;
6303 * Build an iteration mask that can exclude certain CPUs from the upwards
6306 * Asymmetric node setups can result in situations where the domain tree is of
6307 * unequal depth, make sure to skip domains that already cover the entire
6310 * In that case build_sched_domains() will have terminated the iteration early
6311 * and our sibling sd spans will be empty. Domains should always include the
6312 * cpu they're built on, so check that.
6315 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6317 const struct cpumask *span = sched_domain_span(sd);
6318 struct sd_data *sdd = sd->private;
6319 struct sched_domain *sibling;
6322 for_each_cpu(i, span) {
6323 sibling = *per_cpu_ptr(sdd->sd, i);
6324 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6327 cpumask_set_cpu(i, sched_group_mask(sg));
6332 * Return the canonical balance cpu for this group, this is the first cpu
6333 * of this group that's also in the iteration mask.
6335 int group_balance_cpu(struct sched_group *sg)
6337 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6341 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6343 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6344 const struct cpumask *span = sched_domain_span(sd);
6345 struct cpumask *covered = sched_domains_tmpmask;
6346 struct sd_data *sdd = sd->private;
6347 struct sched_domain *sibling;
6350 cpumask_clear(covered);
6352 for_each_cpu(i, span) {
6353 struct cpumask *sg_span;
6355 if (cpumask_test_cpu(i, covered))
6358 sibling = *per_cpu_ptr(sdd->sd, i);
6360 /* See the comment near build_group_mask(). */
6361 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6364 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6365 GFP_KERNEL, cpu_to_node(cpu));
6370 sg_span = sched_group_cpus(sg);
6372 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6374 cpumask_set_cpu(i, sg_span);
6376 cpumask_or(covered, covered, sg_span);
6378 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6379 if (atomic_inc_return(&sg->sgc->ref) == 1)
6380 build_group_mask(sd, sg);
6383 * Initialize sgc->capacity such that even if we mess up the
6384 * domains and no possible iteration will get us here, we won't
6387 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6388 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6391 * Make sure the first group of this domain contains the
6392 * canonical balance cpu. Otherwise the sched_domain iteration
6393 * breaks. See update_sg_lb_stats().
6395 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6396 group_balance_cpu(sg) == cpu)
6406 sd->groups = groups;
6411 free_sched_groups(first, 0);
6416 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6418 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6419 struct sched_domain *child = sd->child;
6422 cpu = cpumask_first(sched_domain_span(child));
6425 *sg = *per_cpu_ptr(sdd->sg, cpu);
6426 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6427 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6434 * build_sched_groups will build a circular linked list of the groups
6435 * covered by the given span, and will set each group's ->cpumask correctly,
6436 * and ->cpu_capacity to 0.
6438 * Assumes the sched_domain tree is fully constructed
6441 build_sched_groups(struct sched_domain *sd, int cpu)
6443 struct sched_group *first = NULL, *last = NULL;
6444 struct sd_data *sdd = sd->private;
6445 const struct cpumask *span = sched_domain_span(sd);
6446 struct cpumask *covered;
6449 get_group(cpu, sdd, &sd->groups);
6450 atomic_inc(&sd->groups->ref);
6452 if (cpu != cpumask_first(span))
6455 lockdep_assert_held(&sched_domains_mutex);
6456 covered = sched_domains_tmpmask;
6458 cpumask_clear(covered);
6460 for_each_cpu(i, span) {
6461 struct sched_group *sg;
6464 if (cpumask_test_cpu(i, covered))
6467 group = get_group(i, sdd, &sg);
6468 cpumask_setall(sched_group_mask(sg));
6470 for_each_cpu(j, span) {
6471 if (get_group(j, sdd, NULL) != group)
6474 cpumask_set_cpu(j, covered);
6475 cpumask_set_cpu(j, sched_group_cpus(sg));
6490 * Initialize sched groups cpu_capacity.
6492 * cpu_capacity indicates the capacity of sched group, which is used while
6493 * distributing the load between different sched groups in a sched domain.
6494 * Typically cpu_capacity for all the groups in a sched domain will be same
6495 * unless there are asymmetries in the topology. If there are asymmetries,
6496 * group having more cpu_capacity will pickup more load compared to the
6497 * group having less cpu_capacity.
6499 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6501 struct sched_group *sg = sd->groups;
6506 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6508 } while (sg != sd->groups);
6510 if (cpu != group_balance_cpu(sg))
6513 update_group_capacity(sd, cpu);
6514 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6518 * Check that the per-cpu provided sd energy data is consistent for all cpus
6521 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6522 const struct cpumask *cpumask)
6524 const struct sched_group_energy * const sge = fn(cpu);
6525 struct cpumask mask;
6528 if (cpumask_weight(cpumask) <= 1)
6531 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6533 for_each_cpu(i, &mask) {
6534 const struct sched_group_energy * const e = fn(i);
6537 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6539 for (y = 0; y < (e->nr_idle_states); y++) {
6540 BUG_ON(e->idle_states[y].power !=
6541 sge->idle_states[y].power);
6544 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6546 for (y = 0; y < (e->nr_cap_states); y++) {
6547 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6548 BUG_ON(e->cap_states[y].power !=
6549 sge->cap_states[y].power);
6554 static void init_sched_energy(int cpu, struct sched_domain *sd,
6555 sched_domain_energy_f fn)
6557 if (!(fn && fn(cpu)))
6560 if (cpu != group_balance_cpu(sd->groups))
6563 if (sd->child && !sd->child->groups->sge) {
6564 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6565 #ifdef CONFIG_SCHED_DEBUG
6566 pr_err(" energy data on %s but not on %s domain\n",
6567 sd->name, sd->child->name);
6572 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6574 sd->groups->sge = fn(cpu);
6578 * Initializers for schedule domains
6579 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6582 static int default_relax_domain_level = -1;
6583 int sched_domain_level_max;
6585 static int __init setup_relax_domain_level(char *str)
6587 if (kstrtoint(str, 0, &default_relax_domain_level))
6588 pr_warn("Unable to set relax_domain_level\n");
6592 __setup("relax_domain_level=", setup_relax_domain_level);
6594 static void set_domain_attribute(struct sched_domain *sd,
6595 struct sched_domain_attr *attr)
6599 if (!attr || attr->relax_domain_level < 0) {
6600 if (default_relax_domain_level < 0)
6603 request = default_relax_domain_level;
6605 request = attr->relax_domain_level;
6606 if (request < sd->level) {
6607 /* turn off idle balance on this domain */
6608 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6610 /* turn on idle balance on this domain */
6611 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6615 static void __sdt_free(const struct cpumask *cpu_map);
6616 static int __sdt_alloc(const struct cpumask *cpu_map);
6618 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6619 const struct cpumask *cpu_map)
6623 if (!atomic_read(&d->rd->refcount))
6624 free_rootdomain(&d->rd->rcu); /* fall through */
6626 free_percpu(d->sd); /* fall through */
6628 __sdt_free(cpu_map); /* fall through */
6634 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6635 const struct cpumask *cpu_map)
6637 memset(d, 0, sizeof(*d));
6639 if (__sdt_alloc(cpu_map))
6640 return sa_sd_storage;
6641 d->sd = alloc_percpu(struct sched_domain *);
6643 return sa_sd_storage;
6644 d->rd = alloc_rootdomain();
6647 return sa_rootdomain;
6651 * NULL the sd_data elements we've used to build the sched_domain and
6652 * sched_group structure so that the subsequent __free_domain_allocs()
6653 * will not free the data we're using.
6655 static void claim_allocations(int cpu, struct sched_domain *sd)
6657 struct sd_data *sdd = sd->private;
6659 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6660 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6662 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6663 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6665 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6666 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6670 static int sched_domains_numa_levels;
6671 enum numa_topology_type sched_numa_topology_type;
6672 static int *sched_domains_numa_distance;
6673 int sched_max_numa_distance;
6674 static struct cpumask ***sched_domains_numa_masks;
6675 static int sched_domains_curr_level;
6679 * SD_flags allowed in topology descriptions.
6681 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6682 * SD_SHARE_PKG_RESOURCES - describes shared caches
6683 * SD_NUMA - describes NUMA topologies
6684 * SD_SHARE_POWERDOMAIN - describes shared power domain
6685 * SD_SHARE_CAP_STATES - describes shared capacity states
6688 * SD_ASYM_PACKING - describes SMT quirks
6690 #define TOPOLOGY_SD_FLAGS \
6691 (SD_SHARE_CPUCAPACITY | \
6692 SD_SHARE_PKG_RESOURCES | \
6695 SD_SHARE_POWERDOMAIN | \
6696 SD_SHARE_CAP_STATES)
6698 static struct sched_domain *
6699 sd_init(struct sched_domain_topology_level *tl, int cpu)
6701 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6702 int sd_weight, sd_flags = 0;
6706 * Ugly hack to pass state to sd_numa_mask()...
6708 sched_domains_curr_level = tl->numa_level;
6711 sd_weight = cpumask_weight(tl->mask(cpu));
6714 sd_flags = (*tl->sd_flags)();
6715 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6716 "wrong sd_flags in topology description\n"))
6717 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6719 *sd = (struct sched_domain){
6720 .min_interval = sd_weight,
6721 .max_interval = 2*sd_weight,
6723 .imbalance_pct = 125,
6725 .cache_nice_tries = 0,
6732 .flags = 1*SD_LOAD_BALANCE
6733 | 1*SD_BALANCE_NEWIDLE
6738 | 0*SD_SHARE_CPUCAPACITY
6739 | 0*SD_SHARE_PKG_RESOURCES
6741 | 0*SD_PREFER_SIBLING
6746 .last_balance = jiffies,
6747 .balance_interval = sd_weight,
6749 .max_newidle_lb_cost = 0,
6750 .next_decay_max_lb_cost = jiffies,
6751 #ifdef CONFIG_SCHED_DEBUG
6757 * Convert topological properties into behaviour.
6760 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6761 sd->flags |= SD_PREFER_SIBLING;
6762 sd->imbalance_pct = 110;
6763 sd->smt_gain = 1178; /* ~15% */
6765 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6766 sd->imbalance_pct = 117;
6767 sd->cache_nice_tries = 1;
6771 } else if (sd->flags & SD_NUMA) {
6772 sd->cache_nice_tries = 2;
6776 sd->flags |= SD_SERIALIZE;
6777 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6778 sd->flags &= ~(SD_BALANCE_EXEC |
6785 sd->flags |= SD_PREFER_SIBLING;
6786 sd->cache_nice_tries = 1;
6791 sd->private = &tl->data;
6797 * Topology list, bottom-up.
6799 static struct sched_domain_topology_level default_topology[] = {
6800 #ifdef CONFIG_SCHED_SMT
6801 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6803 #ifdef CONFIG_SCHED_MC
6804 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6806 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6810 static struct sched_domain_topology_level *sched_domain_topology =
6813 #define for_each_sd_topology(tl) \
6814 for (tl = sched_domain_topology; tl->mask; tl++)
6816 void set_sched_topology(struct sched_domain_topology_level *tl)
6818 sched_domain_topology = tl;
6823 static const struct cpumask *sd_numa_mask(int cpu)
6825 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6828 static void sched_numa_warn(const char *str)
6830 static int done = false;
6838 printk(KERN_WARNING "ERROR: %s\n\n", str);
6840 for (i = 0; i < nr_node_ids; i++) {
6841 printk(KERN_WARNING " ");
6842 for (j = 0; j < nr_node_ids; j++)
6843 printk(KERN_CONT "%02d ", node_distance(i,j));
6844 printk(KERN_CONT "\n");
6846 printk(KERN_WARNING "\n");
6849 bool find_numa_distance(int distance)
6853 if (distance == node_distance(0, 0))
6856 for (i = 0; i < sched_domains_numa_levels; i++) {
6857 if (sched_domains_numa_distance[i] == distance)
6865 * A system can have three types of NUMA topology:
6866 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6867 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6868 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6870 * The difference between a glueless mesh topology and a backplane
6871 * topology lies in whether communication between not directly
6872 * connected nodes goes through intermediary nodes (where programs
6873 * could run), or through backplane controllers. This affects
6874 * placement of programs.
6876 * The type of topology can be discerned with the following tests:
6877 * - If the maximum distance between any nodes is 1 hop, the system
6878 * is directly connected.
6879 * - If for two nodes A and B, located N > 1 hops away from each other,
6880 * there is an intermediary node C, which is < N hops away from both
6881 * nodes A and B, the system is a glueless mesh.
6883 static void init_numa_topology_type(void)
6887 n = sched_max_numa_distance;
6889 if (sched_domains_numa_levels <= 1) {
6890 sched_numa_topology_type = NUMA_DIRECT;
6894 for_each_online_node(a) {
6895 for_each_online_node(b) {
6896 /* Find two nodes furthest removed from each other. */
6897 if (node_distance(a, b) < n)
6900 /* Is there an intermediary node between a and b? */
6901 for_each_online_node(c) {
6902 if (node_distance(a, c) < n &&
6903 node_distance(b, c) < n) {
6904 sched_numa_topology_type =
6910 sched_numa_topology_type = NUMA_BACKPLANE;
6916 static void sched_init_numa(void)
6918 int next_distance, curr_distance = node_distance(0, 0);
6919 struct sched_domain_topology_level *tl;
6923 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6924 if (!sched_domains_numa_distance)
6928 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6929 * unique distances in the node_distance() table.
6931 * Assumes node_distance(0,j) includes all distances in
6932 * node_distance(i,j) in order to avoid cubic time.
6934 next_distance = curr_distance;
6935 for (i = 0; i < nr_node_ids; i++) {
6936 for (j = 0; j < nr_node_ids; j++) {
6937 for (k = 0; k < nr_node_ids; k++) {
6938 int distance = node_distance(i, k);
6940 if (distance > curr_distance &&
6941 (distance < next_distance ||
6942 next_distance == curr_distance))
6943 next_distance = distance;
6946 * While not a strong assumption it would be nice to know
6947 * about cases where if node A is connected to B, B is not
6948 * equally connected to A.
6950 if (sched_debug() && node_distance(k, i) != distance)
6951 sched_numa_warn("Node-distance not symmetric");
6953 if (sched_debug() && i && !find_numa_distance(distance))
6954 sched_numa_warn("Node-0 not representative");
6956 if (next_distance != curr_distance) {
6957 sched_domains_numa_distance[level++] = next_distance;
6958 sched_domains_numa_levels = level;
6959 curr_distance = next_distance;
6964 * In case of sched_debug() we verify the above assumption.
6974 * 'level' contains the number of unique distances, excluding the
6975 * identity distance node_distance(i,i).
6977 * The sched_domains_numa_distance[] array includes the actual distance
6982 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6983 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6984 * the array will contain less then 'level' members. This could be
6985 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6986 * in other functions.
6988 * We reset it to 'level' at the end of this function.
6990 sched_domains_numa_levels = 0;
6992 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6993 if (!sched_domains_numa_masks)
6997 * Now for each level, construct a mask per node which contains all
6998 * cpus of nodes that are that many hops away from us.
7000 for (i = 0; i < level; i++) {
7001 sched_domains_numa_masks[i] =
7002 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
7003 if (!sched_domains_numa_masks[i])
7006 for (j = 0; j < nr_node_ids; j++) {
7007 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
7011 sched_domains_numa_masks[i][j] = mask;
7014 if (node_distance(j, k) > sched_domains_numa_distance[i])
7017 cpumask_or(mask, mask, cpumask_of_node(k));
7022 /* Compute default topology size */
7023 for (i = 0; sched_domain_topology[i].mask; i++);
7025 tl = kzalloc((i + level + 1) *
7026 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7031 * Copy the default topology bits..
7033 for (i = 0; sched_domain_topology[i].mask; i++)
7034 tl[i] = sched_domain_topology[i];
7037 * .. and append 'j' levels of NUMA goodness.
7039 for (j = 0; j < level; i++, j++) {
7040 tl[i] = (struct sched_domain_topology_level){
7041 .mask = sd_numa_mask,
7042 .sd_flags = cpu_numa_flags,
7043 .flags = SDTL_OVERLAP,
7049 sched_domain_topology = tl;
7051 sched_domains_numa_levels = level;
7052 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7054 init_numa_topology_type();
7057 static void sched_domains_numa_masks_set(int cpu)
7060 int node = cpu_to_node(cpu);
7062 for (i = 0; i < sched_domains_numa_levels; i++) {
7063 for (j = 0; j < nr_node_ids; j++) {
7064 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7065 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7070 static void sched_domains_numa_masks_clear(int cpu)
7073 for (i = 0; i < sched_domains_numa_levels; i++) {
7074 for (j = 0; j < nr_node_ids; j++)
7075 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7080 * Update sched_domains_numa_masks[level][node] array when new cpus
7083 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7084 unsigned long action,
7087 int cpu = (long)hcpu;
7089 switch (action & ~CPU_TASKS_FROZEN) {
7091 sched_domains_numa_masks_set(cpu);
7095 sched_domains_numa_masks_clear(cpu);
7105 static inline void sched_init_numa(void)
7109 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7110 unsigned long action,
7115 #endif /* CONFIG_NUMA */
7117 static int __sdt_alloc(const struct cpumask *cpu_map)
7119 struct sched_domain_topology_level *tl;
7122 for_each_sd_topology(tl) {
7123 struct sd_data *sdd = &tl->data;
7125 sdd->sd = alloc_percpu(struct sched_domain *);
7129 sdd->sg = alloc_percpu(struct sched_group *);
7133 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7137 for_each_cpu(j, cpu_map) {
7138 struct sched_domain *sd;
7139 struct sched_group *sg;
7140 struct sched_group_capacity *sgc;
7142 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7143 GFP_KERNEL, cpu_to_node(j));
7147 *per_cpu_ptr(sdd->sd, j) = sd;
7149 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7150 GFP_KERNEL, cpu_to_node(j));
7156 *per_cpu_ptr(sdd->sg, j) = sg;
7158 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7159 GFP_KERNEL, cpu_to_node(j));
7163 *per_cpu_ptr(sdd->sgc, j) = sgc;
7170 static void __sdt_free(const struct cpumask *cpu_map)
7172 struct sched_domain_topology_level *tl;
7175 for_each_sd_topology(tl) {
7176 struct sd_data *sdd = &tl->data;
7178 for_each_cpu(j, cpu_map) {
7179 struct sched_domain *sd;
7182 sd = *per_cpu_ptr(sdd->sd, j);
7183 if (sd && (sd->flags & SD_OVERLAP))
7184 free_sched_groups(sd->groups, 0);
7185 kfree(*per_cpu_ptr(sdd->sd, j));
7189 kfree(*per_cpu_ptr(sdd->sg, j));
7191 kfree(*per_cpu_ptr(sdd->sgc, j));
7193 free_percpu(sdd->sd);
7195 free_percpu(sdd->sg);
7197 free_percpu(sdd->sgc);
7202 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7203 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7204 struct sched_domain *child, int cpu)
7206 struct sched_domain *sd = sd_init(tl, cpu);
7210 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7212 sd->level = child->level + 1;
7213 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7217 if (!cpumask_subset(sched_domain_span(child),
7218 sched_domain_span(sd))) {
7219 pr_err("BUG: arch topology borken\n");
7220 #ifdef CONFIG_SCHED_DEBUG
7221 pr_err(" the %s domain not a subset of the %s domain\n",
7222 child->name, sd->name);
7224 /* Fixup, ensure @sd has at least @child cpus. */
7225 cpumask_or(sched_domain_span(sd),
7226 sched_domain_span(sd),
7227 sched_domain_span(child));
7231 set_domain_attribute(sd, attr);
7237 * Build sched domains for a given set of cpus and attach the sched domains
7238 * to the individual cpus
7240 static int build_sched_domains(const struct cpumask *cpu_map,
7241 struct sched_domain_attr *attr)
7243 enum s_alloc alloc_state;
7244 struct sched_domain *sd;
7246 struct rq *rq = NULL;
7247 int i, ret = -ENOMEM;
7249 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7250 if (alloc_state != sa_rootdomain)
7253 /* Set up domains for cpus specified by the cpu_map. */
7254 for_each_cpu(i, cpu_map) {
7255 struct sched_domain_topology_level *tl;
7258 for_each_sd_topology(tl) {
7259 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7260 if (tl == sched_domain_topology)
7261 *per_cpu_ptr(d.sd, i) = sd;
7262 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7263 sd->flags |= SD_OVERLAP;
7264 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7269 /* Build the groups for the domains */
7270 for_each_cpu(i, cpu_map) {
7271 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7272 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7273 if (sd->flags & SD_OVERLAP) {
7274 if (build_overlap_sched_groups(sd, i))
7277 if (build_sched_groups(sd, i))
7283 /* Calculate CPU capacity for physical packages and nodes */
7284 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7285 struct sched_domain_topology_level *tl = sched_domain_topology;
7287 if (!cpumask_test_cpu(i, cpu_map))
7290 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7291 init_sched_energy(i, sd, tl->energy);
7292 claim_allocations(i, sd);
7293 init_sched_groups_capacity(i, sd);
7297 /* Attach the domains */
7299 for_each_cpu(i, cpu_map) {
7301 sd = *per_cpu_ptr(d.sd, i);
7302 cpu_attach_domain(sd, d.rd, i);
7308 __free_domain_allocs(&d, alloc_state, cpu_map);
7312 static cpumask_var_t *doms_cur; /* current sched domains */
7313 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7314 static struct sched_domain_attr *dattr_cur;
7315 /* attribues of custom domains in 'doms_cur' */
7318 * Special case: If a kmalloc of a doms_cur partition (array of
7319 * cpumask) fails, then fallback to a single sched domain,
7320 * as determined by the single cpumask fallback_doms.
7322 static cpumask_var_t fallback_doms;
7325 * arch_update_cpu_topology lets virtualized architectures update the
7326 * cpu core maps. It is supposed to return 1 if the topology changed
7327 * or 0 if it stayed the same.
7329 int __weak arch_update_cpu_topology(void)
7334 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7337 cpumask_var_t *doms;
7339 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7342 for (i = 0; i < ndoms; i++) {
7343 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7344 free_sched_domains(doms, i);
7351 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7354 for (i = 0; i < ndoms; i++)
7355 free_cpumask_var(doms[i]);
7360 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7361 * For now this just excludes isolated cpus, but could be used to
7362 * exclude other special cases in the future.
7364 static int init_sched_domains(const struct cpumask *cpu_map)
7368 arch_update_cpu_topology();
7370 doms_cur = alloc_sched_domains(ndoms_cur);
7372 doms_cur = &fallback_doms;
7373 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7374 err = build_sched_domains(doms_cur[0], NULL);
7375 register_sched_domain_sysctl();
7381 * Detach sched domains from a group of cpus specified in cpu_map
7382 * These cpus will now be attached to the NULL domain
7384 static void detach_destroy_domains(const struct cpumask *cpu_map)
7389 for_each_cpu(i, cpu_map)
7390 cpu_attach_domain(NULL, &def_root_domain, i);
7394 /* handle null as "default" */
7395 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7396 struct sched_domain_attr *new, int idx_new)
7398 struct sched_domain_attr tmp;
7405 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7406 new ? (new + idx_new) : &tmp,
7407 sizeof(struct sched_domain_attr));
7411 * Partition sched domains as specified by the 'ndoms_new'
7412 * cpumasks in the array doms_new[] of cpumasks. This compares
7413 * doms_new[] to the current sched domain partitioning, doms_cur[].
7414 * It destroys each deleted domain and builds each new domain.
7416 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7417 * The masks don't intersect (don't overlap.) We should setup one
7418 * sched domain for each mask. CPUs not in any of the cpumasks will
7419 * not be load balanced. If the same cpumask appears both in the
7420 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7423 * The passed in 'doms_new' should be allocated using
7424 * alloc_sched_domains. This routine takes ownership of it and will
7425 * free_sched_domains it when done with it. If the caller failed the
7426 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7427 * and partition_sched_domains() will fallback to the single partition
7428 * 'fallback_doms', it also forces the domains to be rebuilt.
7430 * If doms_new == NULL it will be replaced with cpu_online_mask.
7431 * ndoms_new == 0 is a special case for destroying existing domains,
7432 * and it will not create the default domain.
7434 * Call with hotplug lock held
7436 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7437 struct sched_domain_attr *dattr_new)
7442 mutex_lock(&sched_domains_mutex);
7444 /* always unregister in case we don't destroy any domains */
7445 unregister_sched_domain_sysctl();
7447 /* Let architecture update cpu core mappings. */
7448 new_topology = arch_update_cpu_topology();
7450 n = doms_new ? ndoms_new : 0;
7452 /* Destroy deleted domains */
7453 for (i = 0; i < ndoms_cur; i++) {
7454 for (j = 0; j < n && !new_topology; j++) {
7455 if (cpumask_equal(doms_cur[i], doms_new[j])
7456 && dattrs_equal(dattr_cur, i, dattr_new, j))
7459 /* no match - a current sched domain not in new doms_new[] */
7460 detach_destroy_domains(doms_cur[i]);
7466 if (doms_new == NULL) {
7468 doms_new = &fallback_doms;
7469 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7470 WARN_ON_ONCE(dattr_new);
7473 /* Build new domains */
7474 for (i = 0; i < ndoms_new; i++) {
7475 for (j = 0; j < n && !new_topology; j++) {
7476 if (cpumask_equal(doms_new[i], doms_cur[j])
7477 && dattrs_equal(dattr_new, i, dattr_cur, j))
7480 /* no match - add a new doms_new */
7481 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7486 /* Remember the new sched domains */
7487 if (doms_cur != &fallback_doms)
7488 free_sched_domains(doms_cur, ndoms_cur);
7489 kfree(dattr_cur); /* kfree(NULL) is safe */
7490 doms_cur = doms_new;
7491 dattr_cur = dattr_new;
7492 ndoms_cur = ndoms_new;
7494 register_sched_domain_sysctl();
7496 mutex_unlock(&sched_domains_mutex);
7499 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7502 * Update cpusets according to cpu_active mask. If cpusets are
7503 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7504 * around partition_sched_domains().
7506 * If we come here as part of a suspend/resume, don't touch cpusets because we
7507 * want to restore it back to its original state upon resume anyway.
7509 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7513 case CPU_ONLINE_FROZEN:
7514 case CPU_DOWN_FAILED_FROZEN:
7517 * num_cpus_frozen tracks how many CPUs are involved in suspend
7518 * resume sequence. As long as this is not the last online
7519 * operation in the resume sequence, just build a single sched
7520 * domain, ignoring cpusets.
7523 if (likely(num_cpus_frozen)) {
7524 partition_sched_domains(1, NULL, NULL);
7529 * This is the last CPU online operation. So fall through and
7530 * restore the original sched domains by considering the
7531 * cpuset configurations.
7535 cpuset_update_active_cpus(true);
7543 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7546 unsigned long flags;
7547 long cpu = (long)hcpu;
7553 case CPU_DOWN_PREPARE:
7554 rcu_read_lock_sched();
7555 dl_b = dl_bw_of(cpu);
7557 raw_spin_lock_irqsave(&dl_b->lock, flags);
7558 cpus = dl_bw_cpus(cpu);
7559 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7560 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7562 rcu_read_unlock_sched();
7565 return notifier_from_errno(-EBUSY);
7566 cpuset_update_active_cpus(false);
7568 case CPU_DOWN_PREPARE_FROZEN:
7570 partition_sched_domains(1, NULL, NULL);
7578 void __init sched_init_smp(void)
7580 cpumask_var_t non_isolated_cpus;
7582 walt_init_cpu_efficiency();
7583 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7584 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7589 * There's no userspace yet to cause hotplug operations; hence all the
7590 * cpu masks are stable and all blatant races in the below code cannot
7593 mutex_lock(&sched_domains_mutex);
7594 init_sched_domains(cpu_active_mask);
7595 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7596 if (cpumask_empty(non_isolated_cpus))
7597 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7598 mutex_unlock(&sched_domains_mutex);
7600 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7601 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7602 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7606 /* Move init over to a non-isolated CPU */
7607 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7609 sched_init_granularity();
7610 free_cpumask_var(non_isolated_cpus);
7612 init_sched_rt_class();
7613 init_sched_dl_class();
7616 void __init sched_init_smp(void)
7618 sched_init_granularity();
7620 #endif /* CONFIG_SMP */
7622 int in_sched_functions(unsigned long addr)
7624 return in_lock_functions(addr) ||
7625 (addr >= (unsigned long)__sched_text_start
7626 && addr < (unsigned long)__sched_text_end);
7629 #ifdef CONFIG_CGROUP_SCHED
7631 * Default task group.
7632 * Every task in system belongs to this group at bootup.
7634 struct task_group root_task_group;
7635 LIST_HEAD(task_groups);
7638 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7640 void __init sched_init(void)
7643 unsigned long alloc_size = 0, ptr;
7645 #ifdef CONFIG_FAIR_GROUP_SCHED
7646 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7648 #ifdef CONFIG_RT_GROUP_SCHED
7649 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7652 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7654 #ifdef CONFIG_FAIR_GROUP_SCHED
7655 root_task_group.se = (struct sched_entity **)ptr;
7656 ptr += nr_cpu_ids * sizeof(void **);
7658 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7659 ptr += nr_cpu_ids * sizeof(void **);
7661 #endif /* CONFIG_FAIR_GROUP_SCHED */
7662 #ifdef CONFIG_RT_GROUP_SCHED
7663 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7664 ptr += nr_cpu_ids * sizeof(void **);
7666 root_task_group.rt_rq = (struct rt_rq **)ptr;
7667 ptr += nr_cpu_ids * sizeof(void **);
7669 #endif /* CONFIG_RT_GROUP_SCHED */
7671 #ifdef CONFIG_CPUMASK_OFFSTACK
7672 for_each_possible_cpu(i) {
7673 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7674 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7676 #endif /* CONFIG_CPUMASK_OFFSTACK */
7678 init_rt_bandwidth(&def_rt_bandwidth,
7679 global_rt_period(), global_rt_runtime());
7680 init_dl_bandwidth(&def_dl_bandwidth,
7681 global_rt_period(), global_rt_runtime());
7684 init_defrootdomain();
7687 #ifdef CONFIG_RT_GROUP_SCHED
7688 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7689 global_rt_period(), global_rt_runtime());
7690 #endif /* CONFIG_RT_GROUP_SCHED */
7692 #ifdef CONFIG_CGROUP_SCHED
7693 list_add(&root_task_group.list, &task_groups);
7694 INIT_LIST_HEAD(&root_task_group.children);
7695 INIT_LIST_HEAD(&root_task_group.siblings);
7696 autogroup_init(&init_task);
7698 #endif /* CONFIG_CGROUP_SCHED */
7700 for_each_possible_cpu(i) {
7704 raw_spin_lock_init(&rq->lock);
7706 rq->calc_load_active = 0;
7707 rq->calc_load_update = jiffies + LOAD_FREQ;
7708 init_cfs_rq(&rq->cfs);
7709 init_rt_rq(&rq->rt);
7710 init_dl_rq(&rq->dl);
7711 #ifdef CONFIG_FAIR_GROUP_SCHED
7712 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7713 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7715 * How much cpu bandwidth does root_task_group get?
7717 * In case of task-groups formed thr' the cgroup filesystem, it
7718 * gets 100% of the cpu resources in the system. This overall
7719 * system cpu resource is divided among the tasks of
7720 * root_task_group and its child task-groups in a fair manner,
7721 * based on each entity's (task or task-group's) weight
7722 * (se->load.weight).
7724 * In other words, if root_task_group has 10 tasks of weight
7725 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7726 * then A0's share of the cpu resource is:
7728 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7730 * We achieve this by letting root_task_group's tasks sit
7731 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7733 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7734 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7735 #endif /* CONFIG_FAIR_GROUP_SCHED */
7737 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7738 #ifdef CONFIG_RT_GROUP_SCHED
7739 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7742 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7743 rq->cpu_load[j] = 0;
7745 rq->last_load_update_tick = jiffies;
7750 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7751 rq->balance_callback = NULL;
7752 rq->active_balance = 0;
7753 rq->next_balance = jiffies;
7758 rq->avg_idle = 2*sysctl_sched_migration_cost;
7759 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7760 #ifdef CONFIG_SCHED_WALT
7761 rq->cur_irqload = 0;
7762 rq->avg_irqload = 0;
7766 INIT_LIST_HEAD(&rq->cfs_tasks);
7768 rq_attach_root(rq, &def_root_domain);
7769 #ifdef CONFIG_NO_HZ_COMMON
7772 #ifdef CONFIG_NO_HZ_FULL
7773 rq->last_sched_tick = 0;
7777 atomic_set(&rq->nr_iowait, 0);
7780 set_load_weight(&init_task);
7782 #ifdef CONFIG_PREEMPT_NOTIFIERS
7783 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7787 * The boot idle thread does lazy MMU switching as well:
7789 atomic_inc(&init_mm.mm_count);
7790 enter_lazy_tlb(&init_mm, current);
7793 * During early bootup we pretend to be a normal task:
7795 current->sched_class = &fair_sched_class;
7798 * Make us the idle thread. Technically, schedule() should not be
7799 * called from this thread, however somewhere below it might be,
7800 * but because we are the idle thread, we just pick up running again
7801 * when this runqueue becomes "idle".
7803 init_idle(current, smp_processor_id());
7805 calc_load_update = jiffies + LOAD_FREQ;
7808 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7809 /* May be allocated at isolcpus cmdline parse time */
7810 if (cpu_isolated_map == NULL)
7811 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7812 idle_thread_set_boot_cpu();
7813 set_cpu_rq_start_time();
7815 init_sched_fair_class();
7817 scheduler_running = 1;
7820 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7821 static inline int preempt_count_equals(int preempt_offset)
7823 int nested = preempt_count() + rcu_preempt_depth();
7825 return (nested == preempt_offset);
7828 static int __might_sleep_init_called;
7829 int __init __might_sleep_init(void)
7831 __might_sleep_init_called = 1;
7834 early_initcall(__might_sleep_init);
7836 void __might_sleep(const char *file, int line, int preempt_offset)
7839 * Blocking primitives will set (and therefore destroy) current->state,
7840 * since we will exit with TASK_RUNNING make sure we enter with it,
7841 * otherwise we will destroy state.
7843 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7844 "do not call blocking ops when !TASK_RUNNING; "
7845 "state=%lx set at [<%p>] %pS\n",
7847 (void *)current->task_state_change,
7848 (void *)current->task_state_change);
7850 ___might_sleep(file, line, preempt_offset);
7852 EXPORT_SYMBOL(__might_sleep);
7854 void ___might_sleep(const char *file, int line, int preempt_offset)
7856 static unsigned long prev_jiffy; /* ratelimiting */
7858 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7859 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7860 !is_idle_task(current)) || oops_in_progress)
7862 if (system_state != SYSTEM_RUNNING &&
7863 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
7865 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7867 prev_jiffy = jiffies;
7870 "BUG: sleeping function called from invalid context at %s:%d\n",
7873 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7874 in_atomic(), irqs_disabled(),
7875 current->pid, current->comm);
7877 if (task_stack_end_corrupted(current))
7878 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7880 debug_show_held_locks(current);
7881 if (irqs_disabled())
7882 print_irqtrace_events(current);
7883 #ifdef CONFIG_DEBUG_PREEMPT
7884 if (!preempt_count_equals(preempt_offset)) {
7885 pr_err("Preemption disabled at:");
7886 print_ip_sym(current->preempt_disable_ip);
7892 EXPORT_SYMBOL(___might_sleep);
7895 #ifdef CONFIG_MAGIC_SYSRQ
7896 void normalize_rt_tasks(void)
7898 struct task_struct *g, *p;
7899 struct sched_attr attr = {
7900 .sched_policy = SCHED_NORMAL,
7903 read_lock(&tasklist_lock);
7904 for_each_process_thread(g, p) {
7906 * Only normalize user tasks:
7908 if (p->flags & PF_KTHREAD)
7911 p->se.exec_start = 0;
7912 #ifdef CONFIG_SCHEDSTATS
7913 p->se.statistics.wait_start = 0;
7914 p->se.statistics.sleep_start = 0;
7915 p->se.statistics.block_start = 0;
7918 if (!dl_task(p) && !rt_task(p)) {
7920 * Renice negative nice level userspace
7923 if (task_nice(p) < 0)
7924 set_user_nice(p, 0);
7928 __sched_setscheduler(p, &attr, false, false);
7930 read_unlock(&tasklist_lock);
7933 #endif /* CONFIG_MAGIC_SYSRQ */
7935 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7937 * These functions are only useful for the IA64 MCA handling, or kdb.
7939 * They can only be called when the whole system has been
7940 * stopped - every CPU needs to be quiescent, and no scheduling
7941 * activity can take place. Using them for anything else would
7942 * be a serious bug, and as a result, they aren't even visible
7943 * under any other configuration.
7947 * curr_task - return the current task for a given cpu.
7948 * @cpu: the processor in question.
7950 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7952 * Return: The current task for @cpu.
7954 struct task_struct *curr_task(int cpu)
7956 return cpu_curr(cpu);
7959 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7963 * set_curr_task - set the current task for a given cpu.
7964 * @cpu: the processor in question.
7965 * @p: the task pointer to set.
7967 * Description: This function must only be used when non-maskable interrupts
7968 * are serviced on a separate stack. It allows the architecture to switch the
7969 * notion of the current task on a cpu in a non-blocking manner. This function
7970 * must be called with all CPU's synchronized, and interrupts disabled, the
7971 * and caller must save the original value of the current task (see
7972 * curr_task() above) and restore that value before reenabling interrupts and
7973 * re-starting the system.
7975 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7977 void set_curr_task(int cpu, struct task_struct *p)
7984 #ifdef CONFIG_CGROUP_SCHED
7985 /* task_group_lock serializes the addition/removal of task groups */
7986 static DEFINE_SPINLOCK(task_group_lock);
7988 static void sched_free_group(struct task_group *tg)
7990 free_fair_sched_group(tg);
7991 free_rt_sched_group(tg);
7996 /* allocate runqueue etc for a new task group */
7997 struct task_group *sched_create_group(struct task_group *parent)
7999 struct task_group *tg;
8001 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8003 return ERR_PTR(-ENOMEM);
8005 if (!alloc_fair_sched_group(tg, parent))
8008 if (!alloc_rt_sched_group(tg, parent))
8014 sched_free_group(tg);
8015 return ERR_PTR(-ENOMEM);
8018 void sched_online_group(struct task_group *tg, struct task_group *parent)
8020 unsigned long flags;
8022 spin_lock_irqsave(&task_group_lock, flags);
8023 list_add_rcu(&tg->list, &task_groups);
8025 WARN_ON(!parent); /* root should already exist */
8027 tg->parent = parent;
8028 INIT_LIST_HEAD(&tg->children);
8029 list_add_rcu(&tg->siblings, &parent->children);
8030 spin_unlock_irqrestore(&task_group_lock, flags);
8033 /* rcu callback to free various structures associated with a task group */
8034 static void sched_free_group_rcu(struct rcu_head *rhp)
8036 /* now it should be safe to free those cfs_rqs */
8037 sched_free_group(container_of(rhp, struct task_group, rcu));
8040 void sched_destroy_group(struct task_group *tg)
8042 /* wait for possible concurrent references to cfs_rqs complete */
8043 call_rcu(&tg->rcu, sched_free_group_rcu);
8046 void sched_offline_group(struct task_group *tg)
8048 unsigned long flags;
8051 /* end participation in shares distribution */
8052 for_each_possible_cpu(i)
8053 unregister_fair_sched_group(tg, i);
8055 spin_lock_irqsave(&task_group_lock, flags);
8056 list_del_rcu(&tg->list);
8057 list_del_rcu(&tg->siblings);
8058 spin_unlock_irqrestore(&task_group_lock, flags);
8061 /* change task's runqueue when it moves between groups.
8062 * The caller of this function should have put the task in its new group
8063 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8064 * reflect its new group.
8066 void sched_move_task(struct task_struct *tsk)
8068 struct task_group *tg;
8069 int queued, running;
8070 unsigned long flags;
8073 rq = task_rq_lock(tsk, &flags);
8075 running = task_current(rq, tsk);
8076 queued = task_on_rq_queued(tsk);
8079 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8080 if (unlikely(running))
8081 put_prev_task(rq, tsk);
8084 * All callers are synchronized by task_rq_lock(); we do not use RCU
8085 * which is pointless here. Thus, we pass "true" to task_css_check()
8086 * to prevent lockdep warnings.
8088 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8089 struct task_group, css);
8090 tg = autogroup_task_group(tsk, tg);
8091 tsk->sched_task_group = tg;
8093 #ifdef CONFIG_FAIR_GROUP_SCHED
8094 if (tsk->sched_class->task_move_group)
8095 tsk->sched_class->task_move_group(tsk);
8098 set_task_rq(tsk, task_cpu(tsk));
8100 if (unlikely(running))
8101 tsk->sched_class->set_curr_task(rq);
8103 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8105 task_rq_unlock(rq, tsk, &flags);
8107 #endif /* CONFIG_CGROUP_SCHED */
8109 #ifdef CONFIG_RT_GROUP_SCHED
8111 * Ensure that the real time constraints are schedulable.
8113 static DEFINE_MUTEX(rt_constraints_mutex);
8115 /* Must be called with tasklist_lock held */
8116 static inline int tg_has_rt_tasks(struct task_group *tg)
8118 struct task_struct *g, *p;
8121 * Autogroups do not have RT tasks; see autogroup_create().
8123 if (task_group_is_autogroup(tg))
8126 for_each_process_thread(g, p) {
8127 if (rt_task(p) && task_group(p) == tg)
8134 struct rt_schedulable_data {
8135 struct task_group *tg;
8140 static int tg_rt_schedulable(struct task_group *tg, void *data)
8142 struct rt_schedulable_data *d = data;
8143 struct task_group *child;
8144 unsigned long total, sum = 0;
8145 u64 period, runtime;
8147 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8148 runtime = tg->rt_bandwidth.rt_runtime;
8151 period = d->rt_period;
8152 runtime = d->rt_runtime;
8156 * Cannot have more runtime than the period.
8158 if (runtime > period && runtime != RUNTIME_INF)
8162 * Ensure we don't starve existing RT tasks.
8164 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8167 total = to_ratio(period, runtime);
8170 * Nobody can have more than the global setting allows.
8172 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8176 * The sum of our children's runtime should not exceed our own.
8178 list_for_each_entry_rcu(child, &tg->children, siblings) {
8179 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8180 runtime = child->rt_bandwidth.rt_runtime;
8182 if (child == d->tg) {
8183 period = d->rt_period;
8184 runtime = d->rt_runtime;
8187 sum += to_ratio(period, runtime);
8196 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8200 struct rt_schedulable_data data = {
8202 .rt_period = period,
8203 .rt_runtime = runtime,
8207 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8213 static int tg_set_rt_bandwidth(struct task_group *tg,
8214 u64 rt_period, u64 rt_runtime)
8219 * Disallowing the root group RT runtime is BAD, it would disallow the
8220 * kernel creating (and or operating) RT threads.
8222 if (tg == &root_task_group && rt_runtime == 0)
8225 /* No period doesn't make any sense. */
8229 mutex_lock(&rt_constraints_mutex);
8230 read_lock(&tasklist_lock);
8231 err = __rt_schedulable(tg, rt_period, rt_runtime);
8235 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8236 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8237 tg->rt_bandwidth.rt_runtime = rt_runtime;
8239 for_each_possible_cpu(i) {
8240 struct rt_rq *rt_rq = tg->rt_rq[i];
8242 raw_spin_lock(&rt_rq->rt_runtime_lock);
8243 rt_rq->rt_runtime = rt_runtime;
8244 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8246 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8248 read_unlock(&tasklist_lock);
8249 mutex_unlock(&rt_constraints_mutex);
8254 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8256 u64 rt_runtime, rt_period;
8258 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8259 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8260 if (rt_runtime_us < 0)
8261 rt_runtime = RUNTIME_INF;
8263 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8266 static long sched_group_rt_runtime(struct task_group *tg)
8270 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8273 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8274 do_div(rt_runtime_us, NSEC_PER_USEC);
8275 return rt_runtime_us;
8278 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8280 u64 rt_runtime, rt_period;
8282 rt_period = rt_period_us * NSEC_PER_USEC;
8283 rt_runtime = tg->rt_bandwidth.rt_runtime;
8285 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8288 static long sched_group_rt_period(struct task_group *tg)
8292 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8293 do_div(rt_period_us, NSEC_PER_USEC);
8294 return rt_period_us;
8296 #endif /* CONFIG_RT_GROUP_SCHED */
8298 #ifdef CONFIG_RT_GROUP_SCHED
8299 static int sched_rt_global_constraints(void)
8303 mutex_lock(&rt_constraints_mutex);
8304 read_lock(&tasklist_lock);
8305 ret = __rt_schedulable(NULL, 0, 0);
8306 read_unlock(&tasklist_lock);
8307 mutex_unlock(&rt_constraints_mutex);
8312 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8314 /* Don't accept realtime tasks when there is no way for them to run */
8315 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8321 #else /* !CONFIG_RT_GROUP_SCHED */
8322 static int sched_rt_global_constraints(void)
8324 unsigned long flags;
8327 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8328 for_each_possible_cpu(i) {
8329 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8331 raw_spin_lock(&rt_rq->rt_runtime_lock);
8332 rt_rq->rt_runtime = global_rt_runtime();
8333 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8335 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8339 #endif /* CONFIG_RT_GROUP_SCHED */
8341 static int sched_dl_global_validate(void)
8343 u64 runtime = global_rt_runtime();
8344 u64 period = global_rt_period();
8345 u64 new_bw = to_ratio(period, runtime);
8348 unsigned long flags;
8351 * Here we want to check the bandwidth not being set to some
8352 * value smaller than the currently allocated bandwidth in
8353 * any of the root_domains.
8355 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8356 * cycling on root_domains... Discussion on different/better
8357 * solutions is welcome!
8359 for_each_possible_cpu(cpu) {
8360 rcu_read_lock_sched();
8361 dl_b = dl_bw_of(cpu);
8363 raw_spin_lock_irqsave(&dl_b->lock, flags);
8364 if (new_bw < dl_b->total_bw)
8366 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8368 rcu_read_unlock_sched();
8377 static void sched_dl_do_global(void)
8382 unsigned long flags;
8384 def_dl_bandwidth.dl_period = global_rt_period();
8385 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8387 if (global_rt_runtime() != RUNTIME_INF)
8388 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8391 * FIXME: As above...
8393 for_each_possible_cpu(cpu) {
8394 rcu_read_lock_sched();
8395 dl_b = dl_bw_of(cpu);
8397 raw_spin_lock_irqsave(&dl_b->lock, flags);
8399 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8401 rcu_read_unlock_sched();
8405 static int sched_rt_global_validate(void)
8407 if (sysctl_sched_rt_period <= 0)
8410 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8411 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8417 static void sched_rt_do_global(void)
8419 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8420 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8423 int sched_rt_handler(struct ctl_table *table, int write,
8424 void __user *buffer, size_t *lenp,
8427 int old_period, old_runtime;
8428 static DEFINE_MUTEX(mutex);
8432 old_period = sysctl_sched_rt_period;
8433 old_runtime = sysctl_sched_rt_runtime;
8435 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8437 if (!ret && write) {
8438 ret = sched_rt_global_validate();
8442 ret = sched_dl_global_validate();
8446 ret = sched_rt_global_constraints();
8450 sched_rt_do_global();
8451 sched_dl_do_global();
8455 sysctl_sched_rt_period = old_period;
8456 sysctl_sched_rt_runtime = old_runtime;
8458 mutex_unlock(&mutex);
8463 int sched_rr_handler(struct ctl_table *table, int write,
8464 void __user *buffer, size_t *lenp,
8468 static DEFINE_MUTEX(mutex);
8471 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8472 /* make sure that internally we keep jiffies */
8473 /* also, writing zero resets timeslice to default */
8474 if (!ret && write) {
8475 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8476 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8478 mutex_unlock(&mutex);
8482 #ifdef CONFIG_CGROUP_SCHED
8484 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8486 return css ? container_of(css, struct task_group, css) : NULL;
8489 static struct cgroup_subsys_state *
8490 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8492 struct task_group *parent = css_tg(parent_css);
8493 struct task_group *tg;
8496 /* This is early initialization for the top cgroup */
8497 return &root_task_group.css;
8500 tg = sched_create_group(parent);
8502 return ERR_PTR(-ENOMEM);
8504 sched_online_group(tg, parent);
8509 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8511 struct task_group *tg = css_tg(css);
8513 sched_offline_group(tg);
8516 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8518 struct task_group *tg = css_tg(css);
8521 * Relies on the RCU grace period between css_released() and this.
8523 sched_free_group(tg);
8526 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8528 sched_move_task(task);
8531 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8533 struct task_struct *task;
8534 struct cgroup_subsys_state *css;
8536 cgroup_taskset_for_each(task, css, tset) {
8537 #ifdef CONFIG_RT_GROUP_SCHED
8538 if (!sched_rt_can_attach(css_tg(css), task))
8541 /* We don't support RT-tasks being in separate groups */
8542 if (task->sched_class != &fair_sched_class)
8549 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8551 struct task_struct *task;
8552 struct cgroup_subsys_state *css;
8554 cgroup_taskset_for_each(task, css, tset)
8555 sched_move_task(task);
8558 #ifdef CONFIG_FAIR_GROUP_SCHED
8559 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8560 struct cftype *cftype, u64 shareval)
8562 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8565 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8568 struct task_group *tg = css_tg(css);
8570 return (u64) scale_load_down(tg->shares);
8573 #ifdef CONFIG_CFS_BANDWIDTH
8574 static DEFINE_MUTEX(cfs_constraints_mutex);
8576 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8577 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8579 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8581 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8583 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8584 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8586 if (tg == &root_task_group)
8590 * Ensure we have at some amount of bandwidth every period. This is
8591 * to prevent reaching a state of large arrears when throttled via
8592 * entity_tick() resulting in prolonged exit starvation.
8594 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8598 * Likewise, bound things on the otherside by preventing insane quota
8599 * periods. This also allows us to normalize in computing quota
8602 if (period > max_cfs_quota_period)
8606 * Prevent race between setting of cfs_rq->runtime_enabled and
8607 * unthrottle_offline_cfs_rqs().
8610 mutex_lock(&cfs_constraints_mutex);
8611 ret = __cfs_schedulable(tg, period, quota);
8615 runtime_enabled = quota != RUNTIME_INF;
8616 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8618 * If we need to toggle cfs_bandwidth_used, off->on must occur
8619 * before making related changes, and on->off must occur afterwards
8621 if (runtime_enabled && !runtime_was_enabled)
8622 cfs_bandwidth_usage_inc();
8623 raw_spin_lock_irq(&cfs_b->lock);
8624 cfs_b->period = ns_to_ktime(period);
8625 cfs_b->quota = quota;
8627 __refill_cfs_bandwidth_runtime(cfs_b);
8628 /* restart the period timer (if active) to handle new period expiry */
8629 if (runtime_enabled)
8630 start_cfs_bandwidth(cfs_b);
8631 raw_spin_unlock_irq(&cfs_b->lock);
8633 for_each_online_cpu(i) {
8634 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8635 struct rq *rq = cfs_rq->rq;
8637 raw_spin_lock_irq(&rq->lock);
8638 cfs_rq->runtime_enabled = runtime_enabled;
8639 cfs_rq->runtime_remaining = 0;
8641 if (cfs_rq->throttled)
8642 unthrottle_cfs_rq(cfs_rq);
8643 raw_spin_unlock_irq(&rq->lock);
8645 if (runtime_was_enabled && !runtime_enabled)
8646 cfs_bandwidth_usage_dec();
8648 mutex_unlock(&cfs_constraints_mutex);
8654 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8658 period = ktime_to_ns(tg->cfs_bandwidth.period);
8659 if (cfs_quota_us < 0)
8660 quota = RUNTIME_INF;
8662 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8664 return tg_set_cfs_bandwidth(tg, period, quota);
8667 long tg_get_cfs_quota(struct task_group *tg)
8671 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8674 quota_us = tg->cfs_bandwidth.quota;
8675 do_div(quota_us, NSEC_PER_USEC);
8680 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8684 period = (u64)cfs_period_us * NSEC_PER_USEC;
8685 quota = tg->cfs_bandwidth.quota;
8687 return tg_set_cfs_bandwidth(tg, period, quota);
8690 long tg_get_cfs_period(struct task_group *tg)
8694 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8695 do_div(cfs_period_us, NSEC_PER_USEC);
8697 return cfs_period_us;
8700 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8703 return tg_get_cfs_quota(css_tg(css));
8706 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8707 struct cftype *cftype, s64 cfs_quota_us)
8709 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8712 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8715 return tg_get_cfs_period(css_tg(css));
8718 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8719 struct cftype *cftype, u64 cfs_period_us)
8721 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8724 struct cfs_schedulable_data {
8725 struct task_group *tg;
8730 * normalize group quota/period to be quota/max_period
8731 * note: units are usecs
8733 static u64 normalize_cfs_quota(struct task_group *tg,
8734 struct cfs_schedulable_data *d)
8742 period = tg_get_cfs_period(tg);
8743 quota = tg_get_cfs_quota(tg);
8746 /* note: these should typically be equivalent */
8747 if (quota == RUNTIME_INF || quota == -1)
8750 return to_ratio(period, quota);
8753 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8755 struct cfs_schedulable_data *d = data;
8756 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8757 s64 quota = 0, parent_quota = -1;
8760 quota = RUNTIME_INF;
8762 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8764 quota = normalize_cfs_quota(tg, d);
8765 parent_quota = parent_b->hierarchical_quota;
8768 * ensure max(child_quota) <= parent_quota, inherit when no
8771 if (quota == RUNTIME_INF)
8772 quota = parent_quota;
8773 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8776 cfs_b->hierarchical_quota = quota;
8781 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8784 struct cfs_schedulable_data data = {
8790 if (quota != RUNTIME_INF) {
8791 do_div(data.period, NSEC_PER_USEC);
8792 do_div(data.quota, NSEC_PER_USEC);
8796 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8802 static int cpu_stats_show(struct seq_file *sf, void *v)
8804 struct task_group *tg = css_tg(seq_css(sf));
8805 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8807 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8808 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8809 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8813 #endif /* CONFIG_CFS_BANDWIDTH */
8814 #endif /* CONFIG_FAIR_GROUP_SCHED */
8816 #ifdef CONFIG_RT_GROUP_SCHED
8817 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8818 struct cftype *cft, s64 val)
8820 return sched_group_set_rt_runtime(css_tg(css), val);
8823 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8826 return sched_group_rt_runtime(css_tg(css));
8829 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8830 struct cftype *cftype, u64 rt_period_us)
8832 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8835 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8838 return sched_group_rt_period(css_tg(css));
8840 #endif /* CONFIG_RT_GROUP_SCHED */
8842 static struct cftype cpu_files[] = {
8843 #ifdef CONFIG_FAIR_GROUP_SCHED
8846 .read_u64 = cpu_shares_read_u64,
8847 .write_u64 = cpu_shares_write_u64,
8850 #ifdef CONFIG_CFS_BANDWIDTH
8852 .name = "cfs_quota_us",
8853 .read_s64 = cpu_cfs_quota_read_s64,
8854 .write_s64 = cpu_cfs_quota_write_s64,
8857 .name = "cfs_period_us",
8858 .read_u64 = cpu_cfs_period_read_u64,
8859 .write_u64 = cpu_cfs_period_write_u64,
8863 .seq_show = cpu_stats_show,
8866 #ifdef CONFIG_RT_GROUP_SCHED
8868 .name = "rt_runtime_us",
8869 .read_s64 = cpu_rt_runtime_read,
8870 .write_s64 = cpu_rt_runtime_write,
8873 .name = "rt_period_us",
8874 .read_u64 = cpu_rt_period_read_uint,
8875 .write_u64 = cpu_rt_period_write_uint,
8881 struct cgroup_subsys cpu_cgrp_subsys = {
8882 .css_alloc = cpu_cgroup_css_alloc,
8883 .css_released = cpu_cgroup_css_released,
8884 .css_free = cpu_cgroup_css_free,
8885 .fork = cpu_cgroup_fork,
8886 .can_attach = cpu_cgroup_can_attach,
8887 .attach = cpu_cgroup_attach,
8888 .allow_attach = subsys_cgroup_allow_attach,
8889 .legacy_cftypes = cpu_files,
8893 #endif /* CONFIG_CGROUP_SCHED */
8895 void dump_cpu_task(int cpu)
8897 pr_info("Task dump for CPU %d:\n", cpu);
8898 sched_show_task(cpu_curr(cpu));