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>
95 static bool have_sched_energy_data(void);
98 DEFINE_MUTEX(sched_domains_mutex);
99 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
101 static void update_rq_clock_task(struct rq *rq, s64 delta);
103 void update_rq_clock(struct rq *rq)
107 lockdep_assert_held(&rq->lock);
109 if (rq->clock_skip_update & RQCF_ACT_SKIP)
112 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
116 update_rq_clock_task(rq, delta);
120 * Debugging: various feature bits
123 #define SCHED_FEAT(name, enabled) \
124 (1UL << __SCHED_FEAT_##name) * enabled |
126 const_debug unsigned int sysctl_sched_features =
127 #include "features.h"
132 #ifdef CONFIG_SCHED_DEBUG
133 #define SCHED_FEAT(name, enabled) \
136 static const char * const sched_feat_names[] = {
137 #include "features.h"
142 static int sched_feat_show(struct seq_file *m, void *v)
146 for (i = 0; i < __SCHED_FEAT_NR; i++) {
147 if (!(sysctl_sched_features & (1UL << i)))
149 seq_printf(m, "%s ", sched_feat_names[i]);
156 #ifdef HAVE_JUMP_LABEL
158 #define jump_label_key__true STATIC_KEY_INIT_TRUE
159 #define jump_label_key__false STATIC_KEY_INIT_FALSE
161 #define SCHED_FEAT(name, enabled) \
162 jump_label_key__##enabled ,
164 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
165 #include "features.h"
170 static void sched_feat_disable(int i)
172 static_key_disable(&sched_feat_keys[i]);
175 static void sched_feat_enable(int i)
177 static_key_enable(&sched_feat_keys[i]);
180 static void sched_feat_disable(int i) { };
181 static void sched_feat_enable(int i) { };
182 #endif /* HAVE_JUMP_LABEL */
184 static int sched_feat_set(char *cmp)
189 if (strncmp(cmp, "NO_", 3) == 0) {
194 for (i = 0; i < __SCHED_FEAT_NR; i++) {
195 if (strcmp(cmp, sched_feat_names[i]) == 0) {
197 sysctl_sched_features &= ~(1UL << i);
198 sched_feat_disable(i);
201 if (i == __SCHED_FEAT_ENERGY_AWARE)
202 WARN(!have_sched_energy_data(),
203 "Missing sched energy data\n");
205 sysctl_sched_features |= (1UL << i);
206 sched_feat_enable(i);
216 sched_feat_write(struct file *filp, const char __user *ubuf,
217 size_t cnt, loff_t *ppos)
227 if (copy_from_user(&buf, ubuf, cnt))
233 /* Ensure the static_key remains in a consistent state */
234 inode = file_inode(filp);
235 mutex_lock(&inode->i_mutex);
236 i = sched_feat_set(cmp);
237 mutex_unlock(&inode->i_mutex);
238 if (i == __SCHED_FEAT_NR)
246 static int sched_feat_open(struct inode *inode, struct file *filp)
248 return single_open(filp, sched_feat_show, NULL);
251 static const struct file_operations sched_feat_fops = {
252 .open = sched_feat_open,
253 .write = sched_feat_write,
256 .release = single_release,
259 static __init int sched_init_debug(void)
261 debugfs_create_file("sched_features", 0644, NULL, NULL,
266 late_initcall(sched_init_debug);
267 #endif /* CONFIG_SCHED_DEBUG */
270 * Number of tasks to iterate in a single balance run.
271 * Limited because this is done with IRQs disabled.
273 const_debug unsigned int sysctl_sched_nr_migrate = 32;
276 * period over which we average the RT time consumption, measured
281 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
284 * period over which we measure -rt task cpu usage in us.
287 unsigned int sysctl_sched_rt_period = 1000000;
289 __read_mostly int scheduler_running;
292 * part of the period that we allow rt tasks to run in us.
295 int sysctl_sched_rt_runtime = 950000;
297 /* cpus with isolated domains */
298 cpumask_var_t cpu_isolated_map;
301 lock_rq_of(struct task_struct *p, unsigned long *flags)
303 return task_rq_lock(p, flags);
307 unlock_rq_of(struct rq *rq, struct task_struct *p, unsigned long *flags)
309 task_rq_unlock(rq, p, flags);
313 * this_rq_lock - lock this runqueue and disable interrupts.
315 static struct rq *this_rq_lock(void)
322 raw_spin_lock(&rq->lock);
327 #ifdef CONFIG_SCHED_HRTICK
329 * Use HR-timers to deliver accurate preemption points.
332 static void hrtick_clear(struct rq *rq)
334 if (hrtimer_active(&rq->hrtick_timer))
335 hrtimer_cancel(&rq->hrtick_timer);
339 * High-resolution timer tick.
340 * Runs from hardirq context with interrupts disabled.
342 static enum hrtimer_restart hrtick(struct hrtimer *timer)
344 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
346 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
348 raw_spin_lock(&rq->lock);
350 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
351 raw_spin_unlock(&rq->lock);
353 return HRTIMER_NORESTART;
358 static void __hrtick_restart(struct rq *rq)
360 struct hrtimer *timer = &rq->hrtick_timer;
362 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
366 * called from hardirq (IPI) context
368 static void __hrtick_start(void *arg)
372 raw_spin_lock(&rq->lock);
373 __hrtick_restart(rq);
374 rq->hrtick_csd_pending = 0;
375 raw_spin_unlock(&rq->lock);
379 * Called to set the hrtick timer state.
381 * called with rq->lock held and irqs disabled
383 void hrtick_start(struct rq *rq, u64 delay)
385 struct hrtimer *timer = &rq->hrtick_timer;
390 * Don't schedule slices shorter than 10000ns, that just
391 * doesn't make sense and can cause timer DoS.
393 delta = max_t(s64, delay, 10000LL);
394 time = ktime_add_ns(timer->base->get_time(), delta);
396 hrtimer_set_expires(timer, time);
398 if (rq == this_rq()) {
399 __hrtick_restart(rq);
400 } else if (!rq->hrtick_csd_pending) {
401 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
402 rq->hrtick_csd_pending = 1;
407 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
409 int cpu = (int)(long)hcpu;
412 case CPU_UP_CANCELED:
413 case CPU_UP_CANCELED_FROZEN:
414 case CPU_DOWN_PREPARE:
415 case CPU_DOWN_PREPARE_FROZEN:
417 case CPU_DEAD_FROZEN:
418 hrtick_clear(cpu_rq(cpu));
425 static __init void init_hrtick(void)
427 hotcpu_notifier(hotplug_hrtick, 0);
431 * Called to set the hrtick timer state.
433 * called with rq->lock held and irqs disabled
435 void hrtick_start(struct rq *rq, u64 delay)
438 * Don't schedule slices shorter than 10000ns, that just
439 * doesn't make sense. Rely on vruntime for fairness.
441 delay = max_t(u64, delay, 10000LL);
442 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
443 HRTIMER_MODE_REL_PINNED);
446 static inline void init_hrtick(void)
449 #endif /* CONFIG_SMP */
451 static void init_rq_hrtick(struct rq *rq)
454 rq->hrtick_csd_pending = 0;
456 rq->hrtick_csd.flags = 0;
457 rq->hrtick_csd.func = __hrtick_start;
458 rq->hrtick_csd.info = rq;
461 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
462 rq->hrtick_timer.function = hrtick;
464 #else /* CONFIG_SCHED_HRTICK */
465 static inline void hrtick_clear(struct rq *rq)
469 static inline void init_rq_hrtick(struct rq *rq)
473 static inline void init_hrtick(void)
476 #endif /* CONFIG_SCHED_HRTICK */
479 * cmpxchg based fetch_or, macro so it works for different integer types
481 #define fetch_or(ptr, val) \
482 ({ typeof(*(ptr)) __old, __val = *(ptr); \
484 __old = cmpxchg((ptr), __val, __val | (val)); \
485 if (__old == __val) \
492 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
494 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
495 * this avoids any races wrt polling state changes and thereby avoids
498 static bool set_nr_and_not_polling(struct task_struct *p)
500 struct thread_info *ti = task_thread_info(p);
501 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
505 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
507 * If this returns true, then the idle task promises to call
508 * sched_ttwu_pending() and reschedule soon.
510 static bool set_nr_if_polling(struct task_struct *p)
512 struct thread_info *ti = task_thread_info(p);
513 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
516 if (!(val & _TIF_POLLING_NRFLAG))
518 if (val & _TIF_NEED_RESCHED)
520 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
529 static bool set_nr_and_not_polling(struct task_struct *p)
531 set_tsk_need_resched(p);
536 static bool set_nr_if_polling(struct task_struct *p)
543 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
545 struct wake_q_node *node = &task->wake_q;
548 * Atomically grab the task, if ->wake_q is !nil already it means
549 * its already queued (either by us or someone else) and will get the
550 * wakeup due to that.
552 * This cmpxchg() implies a full barrier, which pairs with the write
553 * barrier implied by the wakeup in wake_up_list().
555 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
560 get_task_struct(task);
563 * The head is context local, there can be no concurrency.
566 head->lastp = &node->next;
570 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags,
571 int sibling_count_hint);
573 void wake_up_q(struct wake_q_head *head)
575 struct wake_q_node *node = head->first;
577 while (node != WAKE_Q_TAIL) {
578 struct task_struct *task;
580 task = container_of(node, struct task_struct, wake_q);
582 /* task can safely be re-inserted now */
584 task->wake_q.next = NULL;
587 * try_to_wake_up() implies a wmb() to pair with the queueing
588 * in wake_q_add() so as not to miss wakeups.
590 try_to_wake_up(task, TASK_NORMAL, 0, head->count);
591 put_task_struct(task);
596 * resched_curr - mark rq's current task 'to be rescheduled now'.
598 * On UP this means the setting of the need_resched flag, on SMP it
599 * might also involve a cross-CPU call to trigger the scheduler on
602 void resched_curr(struct rq *rq)
604 struct task_struct *curr = rq->curr;
607 lockdep_assert_held(&rq->lock);
609 if (test_tsk_need_resched(curr))
614 if (cpu == smp_processor_id()) {
615 set_tsk_need_resched(curr);
616 set_preempt_need_resched();
620 if (set_nr_and_not_polling(curr))
621 smp_send_reschedule(cpu);
623 trace_sched_wake_idle_without_ipi(cpu);
626 void resched_cpu(int cpu)
628 struct rq *rq = cpu_rq(cpu);
631 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
634 raw_spin_unlock_irqrestore(&rq->lock, flags);
638 #ifdef CONFIG_NO_HZ_COMMON
640 * In the semi idle case, use the nearest busy cpu for migrating timers
641 * from an idle cpu. This is good for power-savings.
643 * We don't do similar optimization for completely idle system, as
644 * selecting an idle cpu will add more delays to the timers than intended
645 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
647 int get_nohz_timer_target(void)
649 int i, cpu = smp_processor_id();
650 struct sched_domain *sd;
652 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
656 for_each_domain(cpu, sd) {
657 for_each_cpu(i, sched_domain_span(sd)) {
661 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
668 if (!is_housekeeping_cpu(cpu))
669 cpu = housekeeping_any_cpu();
675 * When add_timer_on() enqueues a timer into the timer wheel of an
676 * idle CPU then this timer might expire before the next timer event
677 * which is scheduled to wake up that CPU. In case of a completely
678 * idle system the next event might even be infinite time into the
679 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
680 * leaves the inner idle loop so the newly added timer is taken into
681 * account when the CPU goes back to idle and evaluates the timer
682 * wheel for the next timer event.
684 static void wake_up_idle_cpu(int cpu)
686 struct rq *rq = cpu_rq(cpu);
688 if (cpu == smp_processor_id())
691 if (set_nr_and_not_polling(rq->idle))
692 smp_send_reschedule(cpu);
694 trace_sched_wake_idle_without_ipi(cpu);
697 static bool wake_up_full_nohz_cpu(int cpu)
700 * We just need the target to call irq_exit() and re-evaluate
701 * the next tick. The nohz full kick at least implies that.
702 * If needed we can still optimize that later with an
705 if (tick_nohz_full_cpu(cpu)) {
706 if (cpu != smp_processor_id() ||
707 tick_nohz_tick_stopped())
708 tick_nohz_full_kick_cpu(cpu);
715 void wake_up_nohz_cpu(int cpu)
717 if (!wake_up_full_nohz_cpu(cpu))
718 wake_up_idle_cpu(cpu);
721 static inline bool got_nohz_idle_kick(void)
723 int cpu = smp_processor_id();
725 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
728 if (idle_cpu(cpu) && !need_resched())
732 * We can't run Idle Load Balance on this CPU for this time so we
733 * cancel it and clear NOHZ_BALANCE_KICK
735 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
739 #else /* CONFIG_NO_HZ_COMMON */
741 static inline bool got_nohz_idle_kick(void)
746 #endif /* CONFIG_NO_HZ_COMMON */
748 #ifdef CONFIG_NO_HZ_FULL
749 bool sched_can_stop_tick(void)
752 * FIFO realtime policy runs the highest priority task. Other runnable
753 * tasks are of a lower priority. The scheduler tick does nothing.
755 if (current->policy == SCHED_FIFO)
759 * Round-robin realtime tasks time slice with other tasks at the same
760 * realtime priority. Is this task the only one at this priority?
762 if (current->policy == SCHED_RR) {
763 struct sched_rt_entity *rt_se = ¤t->rt;
765 return rt_se->run_list.prev == rt_se->run_list.next;
769 * More than one running task need preemption.
770 * nr_running update is assumed to be visible
771 * after IPI is sent from wakers.
773 if (this_rq()->nr_running > 1)
778 #endif /* CONFIG_NO_HZ_FULL */
780 void sched_avg_update(struct rq *rq)
782 s64 period = sched_avg_period();
784 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
786 * Inline assembly required to prevent the compiler
787 * optimising this loop into a divmod call.
788 * See __iter_div_u64_rem() for another example of this.
790 asm("" : "+rm" (rq->age_stamp));
791 rq->age_stamp += period;
796 #endif /* CONFIG_SMP */
798 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
799 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
801 * Iterate task_group tree rooted at *from, calling @down when first entering a
802 * node and @up when leaving it for the final time.
804 * Caller must hold rcu_lock or sufficient equivalent.
806 int walk_tg_tree_from(struct task_group *from,
807 tg_visitor down, tg_visitor up, void *data)
809 struct task_group *parent, *child;
815 ret = (*down)(parent, data);
818 list_for_each_entry_rcu(child, &parent->children, siblings) {
825 ret = (*up)(parent, data);
826 if (ret || parent == from)
830 parent = parent->parent;
837 int tg_nop(struct task_group *tg, void *data)
843 static void set_load_weight(struct task_struct *p)
845 int prio = p->static_prio - MAX_RT_PRIO;
846 struct load_weight *load = &p->se.load;
849 * SCHED_IDLE tasks get minimal weight:
851 if (idle_policy(p->policy)) {
852 load->weight = scale_load(WEIGHT_IDLEPRIO);
853 load->inv_weight = WMULT_IDLEPRIO;
857 load->weight = scale_load(prio_to_weight[prio]);
858 load->inv_weight = prio_to_wmult[prio];
861 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
864 if (!(flags & ENQUEUE_RESTORE))
865 sched_info_queued(rq, p);
866 p->sched_class->enqueue_task(rq, p, flags);
869 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
872 if (!(flags & DEQUEUE_SAVE))
873 sched_info_dequeued(rq, p);
874 p->sched_class->dequeue_task(rq, p, flags);
877 void activate_task(struct rq *rq, struct task_struct *p, int flags)
879 if (task_contributes_to_load(p))
880 rq->nr_uninterruptible--;
882 enqueue_task(rq, p, flags);
885 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
887 if (task_contributes_to_load(p))
888 rq->nr_uninterruptible++;
890 dequeue_task(rq, p, flags);
893 static void update_rq_clock_task(struct rq *rq, s64 delta)
896 * In theory, the compile should just see 0 here, and optimize out the call
897 * to sched_rt_avg_update. But I don't trust it...
899 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
900 s64 steal = 0, irq_delta = 0;
902 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
903 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
906 * Since irq_time is only updated on {soft,}irq_exit, we might run into
907 * this case when a previous update_rq_clock() happened inside a
910 * When this happens, we stop ->clock_task and only update the
911 * prev_irq_time stamp to account for the part that fit, so that a next
912 * update will consume the rest. This ensures ->clock_task is
915 * It does however cause some slight miss-attribution of {soft,}irq
916 * time, a more accurate solution would be to update the irq_time using
917 * the current rq->clock timestamp, except that would require using
920 if (irq_delta > delta)
923 rq->prev_irq_time += irq_delta;
926 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
927 if (static_key_false((¶virt_steal_rq_enabled))) {
928 steal = paravirt_steal_clock(cpu_of(rq));
929 steal -= rq->prev_steal_time_rq;
931 if (unlikely(steal > delta))
934 rq->prev_steal_time_rq += steal;
939 rq->clock_task += delta;
941 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
942 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
943 sched_rt_avg_update(rq, irq_delta + steal);
947 void sched_set_stop_task(int cpu, struct task_struct *stop)
949 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
950 struct task_struct *old_stop = cpu_rq(cpu)->stop;
954 * Make it appear like a SCHED_FIFO task, its something
955 * userspace knows about and won't get confused about.
957 * Also, it will make PI more or less work without too
958 * much confusion -- but then, stop work should not
959 * rely on PI working anyway.
961 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
963 stop->sched_class = &stop_sched_class;
966 cpu_rq(cpu)->stop = stop;
970 * Reset it back to a normal scheduling class so that
971 * it can die in pieces.
973 old_stop->sched_class = &rt_sched_class;
978 * __normal_prio - return the priority that is based on the static prio
980 static inline int __normal_prio(struct task_struct *p)
982 return p->static_prio;
986 * Calculate the expected normal priority: i.e. priority
987 * without taking RT-inheritance into account. Might be
988 * boosted by interactivity modifiers. Changes upon fork,
989 * setprio syscalls, and whenever the interactivity
990 * estimator recalculates.
992 static inline int normal_prio(struct task_struct *p)
996 if (task_has_dl_policy(p))
997 prio = MAX_DL_PRIO-1;
998 else if (task_has_rt_policy(p))
999 prio = MAX_RT_PRIO-1 - p->rt_priority;
1001 prio = __normal_prio(p);
1006 * Calculate the current priority, i.e. the priority
1007 * taken into account by the scheduler. This value might
1008 * be boosted by RT tasks, or might be boosted by
1009 * interactivity modifiers. Will be RT if the task got
1010 * RT-boosted. If not then it returns p->normal_prio.
1012 static int effective_prio(struct task_struct *p)
1014 p->normal_prio = normal_prio(p);
1016 * If we are RT tasks or we were boosted to RT priority,
1017 * keep the priority unchanged. Otherwise, update priority
1018 * to the normal priority:
1020 if (!rt_prio(p->prio))
1021 return p->normal_prio;
1026 * task_curr - is this task currently executing on a CPU?
1027 * @p: the task in question.
1029 * Return: 1 if the task is currently executing. 0 otherwise.
1031 inline int task_curr(const struct task_struct *p)
1033 return cpu_curr(task_cpu(p)) == p;
1037 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1038 * use the balance_callback list if you want balancing.
1040 * this means any call to check_class_changed() must be followed by a call to
1041 * balance_callback().
1043 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1044 const struct sched_class *prev_class,
1047 if (prev_class != p->sched_class) {
1048 if (prev_class->switched_from)
1049 prev_class->switched_from(rq, p);
1051 p->sched_class->switched_to(rq, p);
1052 } else if (oldprio != p->prio || dl_task(p))
1053 p->sched_class->prio_changed(rq, p, oldprio);
1056 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1058 const struct sched_class *class;
1060 if (p->sched_class == rq->curr->sched_class) {
1061 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1063 for_each_class(class) {
1064 if (class == rq->curr->sched_class)
1066 if (class == p->sched_class) {
1074 * A queue event has occurred, and we're going to schedule. In
1075 * this case, we can save a useless back to back clock update.
1077 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1078 rq_clock_skip_update(rq, true);
1083 * This is how migration works:
1085 * 1) we invoke migration_cpu_stop() on the target CPU using
1087 * 2) stopper starts to run (implicitly forcing the migrated thread
1089 * 3) it checks whether the migrated task is still in the wrong runqueue.
1090 * 4) if it's in the wrong runqueue then the migration thread removes
1091 * it and puts it into the right queue.
1092 * 5) stopper completes and stop_one_cpu() returns and the migration
1097 * move_queued_task - move a queued task to new rq.
1099 * Returns (locked) new rq. Old rq's lock is released.
1101 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1103 lockdep_assert_held(&rq->lock);
1105 dequeue_task(rq, p, 0);
1106 p->on_rq = TASK_ON_RQ_MIGRATING;
1107 double_lock_balance(rq, cpu_rq(new_cpu));
1108 set_task_cpu(p, new_cpu);
1109 double_unlock_balance(rq, cpu_rq(new_cpu));
1110 raw_spin_unlock(&rq->lock);
1112 rq = cpu_rq(new_cpu);
1114 raw_spin_lock(&rq->lock);
1115 BUG_ON(task_cpu(p) != new_cpu);
1116 p->on_rq = TASK_ON_RQ_QUEUED;
1117 enqueue_task(rq, p, 0);
1118 check_preempt_curr(rq, p, 0);
1123 struct migration_arg {
1124 struct task_struct *task;
1129 * Move (not current) task off this cpu, onto dest cpu. We're doing
1130 * this because either it can't run here any more (set_cpus_allowed()
1131 * away from this CPU, or CPU going down), or because we're
1132 * attempting to rebalance this task on exec (sched_exec).
1134 * So we race with normal scheduler movements, but that's OK, as long
1135 * as the task is no longer on this CPU.
1137 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1139 if (unlikely(!cpu_active(dest_cpu)))
1142 /* Affinity changed (again). */
1143 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1146 rq = move_queued_task(rq, p, dest_cpu);
1152 * migration_cpu_stop - this will be executed by a highprio stopper thread
1153 * and performs thread migration by bumping thread off CPU then
1154 * 'pushing' onto another runqueue.
1156 static int migration_cpu_stop(void *data)
1158 struct migration_arg *arg = data;
1159 struct task_struct *p = arg->task;
1160 struct rq *rq = this_rq();
1163 * The original target cpu might have gone down and we might
1164 * be on another cpu but it doesn't matter.
1166 local_irq_disable();
1168 * We need to explicitly wake pending tasks before running
1169 * __migrate_task() such that we will not miss enforcing cpus_allowed
1170 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1172 sched_ttwu_pending();
1174 raw_spin_lock(&p->pi_lock);
1175 raw_spin_lock(&rq->lock);
1177 * If task_rq(p) != rq, it cannot be migrated here, because we're
1178 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1179 * we're holding p->pi_lock.
1181 if (task_rq(p) == rq && task_on_rq_queued(p))
1182 rq = __migrate_task(rq, p, arg->dest_cpu);
1183 raw_spin_unlock(&rq->lock);
1184 raw_spin_unlock(&p->pi_lock);
1191 * sched_class::set_cpus_allowed must do the below, but is not required to
1192 * actually call this function.
1194 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1196 cpumask_copy(&p->cpus_allowed, new_mask);
1197 p->nr_cpus_allowed = cpumask_weight(new_mask);
1200 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1202 struct rq *rq = task_rq(p);
1203 bool queued, running;
1205 lockdep_assert_held(&p->pi_lock);
1207 queued = task_on_rq_queued(p);
1208 running = task_current(rq, p);
1212 * Because __kthread_bind() calls this on blocked tasks without
1215 lockdep_assert_held(&rq->lock);
1216 dequeue_task(rq, p, DEQUEUE_SAVE);
1219 put_prev_task(rq, p);
1221 p->sched_class->set_cpus_allowed(p, new_mask);
1224 p->sched_class->set_curr_task(rq);
1226 enqueue_task(rq, p, ENQUEUE_RESTORE);
1230 * Change a given task's CPU affinity. Migrate the thread to a
1231 * proper CPU and schedule it away if the CPU it's executing on
1232 * is removed from the allowed bitmask.
1234 * NOTE: the caller must have a valid reference to the task, the
1235 * task must not exit() & deallocate itself prematurely. The
1236 * call is not atomic; no spinlocks may be held.
1238 static int __set_cpus_allowed_ptr(struct task_struct *p,
1239 const struct cpumask *new_mask, bool check)
1241 unsigned long flags;
1243 unsigned int dest_cpu;
1246 rq = task_rq_lock(p, &flags);
1249 * Must re-check here, to close a race against __kthread_bind(),
1250 * sched_setaffinity() is not guaranteed to observe the flag.
1252 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1257 if (cpumask_equal(&p->cpus_allowed, new_mask))
1260 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1265 do_set_cpus_allowed(p, new_mask);
1267 /* Can the task run on the task's current CPU? If so, we're done */
1268 if (cpumask_test_cpu(task_cpu(p), new_mask))
1271 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1272 if (task_running(rq, p) || p->state == TASK_WAKING) {
1273 struct migration_arg arg = { p, dest_cpu };
1274 /* Need help from migration thread: drop lock and wait. */
1275 task_rq_unlock(rq, p, &flags);
1276 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1277 tlb_migrate_finish(p->mm);
1279 } else if (task_on_rq_queued(p)) {
1281 * OK, since we're going to drop the lock immediately
1282 * afterwards anyway.
1284 lockdep_unpin_lock(&rq->lock);
1285 rq = move_queued_task(rq, p, dest_cpu);
1286 lockdep_pin_lock(&rq->lock);
1289 task_rq_unlock(rq, p, &flags);
1294 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1296 return __set_cpus_allowed_ptr(p, new_mask, false);
1298 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1300 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1302 #ifdef CONFIG_SCHED_DEBUG
1304 * We should never call set_task_cpu() on a blocked task,
1305 * ttwu() will sort out the placement.
1307 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1310 #ifdef CONFIG_LOCKDEP
1312 * The caller should hold either p->pi_lock or rq->lock, when changing
1313 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1315 * sched_move_task() holds both and thus holding either pins the cgroup,
1318 * Furthermore, all task_rq users should acquire both locks, see
1321 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1322 lockdep_is_held(&task_rq(p)->lock)));
1326 trace_sched_migrate_task(p, new_cpu);
1328 if (task_cpu(p) != new_cpu) {
1329 if (p->sched_class->migrate_task_rq)
1330 p->sched_class->migrate_task_rq(p);
1331 p->se.nr_migrations++;
1332 perf_event_task_migrate(p);
1334 walt_fixup_busy_time(p, new_cpu);
1337 __set_task_cpu(p, new_cpu);
1340 static void __migrate_swap_task(struct task_struct *p, int cpu)
1342 if (task_on_rq_queued(p)) {
1343 struct rq *src_rq, *dst_rq;
1345 src_rq = task_rq(p);
1346 dst_rq = cpu_rq(cpu);
1348 deactivate_task(src_rq, p, 0);
1349 p->on_rq = TASK_ON_RQ_MIGRATING;
1350 set_task_cpu(p, cpu);
1351 p->on_rq = TASK_ON_RQ_QUEUED;
1352 activate_task(dst_rq, p, 0);
1353 check_preempt_curr(dst_rq, p, 0);
1356 * Task isn't running anymore; make it appear like we migrated
1357 * it before it went to sleep. This means on wakeup we make the
1358 * previous cpu our targer instead of where it really is.
1364 struct migration_swap_arg {
1365 struct task_struct *src_task, *dst_task;
1366 int src_cpu, dst_cpu;
1369 static int migrate_swap_stop(void *data)
1371 struct migration_swap_arg *arg = data;
1372 struct rq *src_rq, *dst_rq;
1375 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1378 src_rq = cpu_rq(arg->src_cpu);
1379 dst_rq = cpu_rq(arg->dst_cpu);
1381 double_raw_lock(&arg->src_task->pi_lock,
1382 &arg->dst_task->pi_lock);
1383 double_rq_lock(src_rq, dst_rq);
1385 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1388 if (task_cpu(arg->src_task) != arg->src_cpu)
1391 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1394 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1397 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1398 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1403 double_rq_unlock(src_rq, dst_rq);
1404 raw_spin_unlock(&arg->dst_task->pi_lock);
1405 raw_spin_unlock(&arg->src_task->pi_lock);
1411 * Cross migrate two tasks
1413 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1415 struct migration_swap_arg arg;
1418 arg = (struct migration_swap_arg){
1420 .src_cpu = task_cpu(cur),
1422 .dst_cpu = task_cpu(p),
1425 if (arg.src_cpu == arg.dst_cpu)
1429 * These three tests are all lockless; this is OK since all of them
1430 * will be re-checked with proper locks held further down the line.
1432 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1435 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1438 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1441 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1442 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1449 * wait_task_inactive - wait for a thread to unschedule.
1451 * If @match_state is nonzero, it's the @p->state value just checked and
1452 * not expected to change. If it changes, i.e. @p might have woken up,
1453 * then return zero. When we succeed in waiting for @p to be off its CPU,
1454 * we return a positive number (its total switch count). If a second call
1455 * a short while later returns the same number, the caller can be sure that
1456 * @p has remained unscheduled the whole time.
1458 * The caller must ensure that the task *will* unschedule sometime soon,
1459 * else this function might spin for a *long* time. This function can't
1460 * be called with interrupts off, or it may introduce deadlock with
1461 * smp_call_function() if an IPI is sent by the same process we are
1462 * waiting to become inactive.
1464 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1466 unsigned long flags;
1467 int running, queued;
1473 * We do the initial early heuristics without holding
1474 * any task-queue locks at all. We'll only try to get
1475 * the runqueue lock when things look like they will
1481 * If the task is actively running on another CPU
1482 * still, just relax and busy-wait without holding
1485 * NOTE! Since we don't hold any locks, it's not
1486 * even sure that "rq" stays as the right runqueue!
1487 * But we don't care, since "task_running()" will
1488 * return false if the runqueue has changed and p
1489 * is actually now running somewhere else!
1491 while (task_running(rq, p)) {
1492 if (match_state && unlikely(p->state != match_state))
1498 * Ok, time to look more closely! We need the rq
1499 * lock now, to be *sure*. If we're wrong, we'll
1500 * just go back and repeat.
1502 rq = task_rq_lock(p, &flags);
1503 trace_sched_wait_task(p);
1504 running = task_running(rq, p);
1505 queued = task_on_rq_queued(p);
1507 if (!match_state || p->state == match_state)
1508 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1509 task_rq_unlock(rq, p, &flags);
1512 * If it changed from the expected state, bail out now.
1514 if (unlikely(!ncsw))
1518 * Was it really running after all now that we
1519 * checked with the proper locks actually held?
1521 * Oops. Go back and try again..
1523 if (unlikely(running)) {
1529 * It's not enough that it's not actively running,
1530 * it must be off the runqueue _entirely_, and not
1533 * So if it was still runnable (but just not actively
1534 * running right now), it's preempted, and we should
1535 * yield - it could be a while.
1537 if (unlikely(queued)) {
1538 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1540 set_current_state(TASK_UNINTERRUPTIBLE);
1541 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1546 * Ahh, all good. It wasn't running, and it wasn't
1547 * runnable, which means that it will never become
1548 * running in the future either. We're all done!
1557 * kick_process - kick a running thread to enter/exit the kernel
1558 * @p: the to-be-kicked thread
1560 * Cause a process which is running on another CPU to enter
1561 * kernel-mode, without any delay. (to get signals handled.)
1563 * NOTE: this function doesn't have to take the runqueue lock,
1564 * because all it wants to ensure is that the remote task enters
1565 * the kernel. If the IPI races and the task has been migrated
1566 * to another CPU then no harm is done and the purpose has been
1569 void kick_process(struct task_struct *p)
1575 if ((cpu != smp_processor_id()) && task_curr(p))
1576 smp_send_reschedule(cpu);
1579 EXPORT_SYMBOL_GPL(kick_process);
1582 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1584 static int select_fallback_rq(int cpu, struct task_struct *p)
1586 int nid = cpu_to_node(cpu);
1587 const struct cpumask *nodemask = NULL;
1588 enum { cpuset, possible, fail } state = cpuset;
1592 * If the node that the cpu is on has been offlined, cpu_to_node()
1593 * will return -1. There is no cpu on the node, and we should
1594 * select the cpu on the other node.
1597 nodemask = cpumask_of_node(nid);
1599 /* Look for allowed, online CPU in same node. */
1600 for_each_cpu(dest_cpu, nodemask) {
1601 if (!cpu_online(dest_cpu))
1603 if (!cpu_active(dest_cpu))
1605 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1611 /* Any allowed, online CPU? */
1612 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1613 if (!cpu_online(dest_cpu))
1615 if (!cpu_active(dest_cpu))
1620 /* No more Mr. Nice Guy. */
1623 if (IS_ENABLED(CONFIG_CPUSETS)) {
1624 cpuset_cpus_allowed_fallback(p);
1630 do_set_cpus_allowed(p, cpu_possible_mask);
1641 if (state != cpuset) {
1643 * Don't tell them about moving exiting tasks or
1644 * kernel threads (both mm NULL), since they never
1647 if (p->mm && printk_ratelimit()) {
1648 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1649 task_pid_nr(p), p->comm, cpu);
1657 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1660 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags,
1661 int sibling_count_hint)
1663 lockdep_assert_held(&p->pi_lock);
1665 if (p->nr_cpus_allowed > 1)
1666 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags,
1667 sibling_count_hint);
1670 * In order not to call set_task_cpu() on a blocking task we need
1671 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1674 * Since this is common to all placement strategies, this lives here.
1676 * [ this allows ->select_task() to simply return task_cpu(p) and
1677 * not worry about this generic constraint ]
1679 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1681 cpu = select_fallback_rq(task_cpu(p), p);
1686 static void update_avg(u64 *avg, u64 sample)
1688 s64 diff = sample - *avg;
1694 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1695 const struct cpumask *new_mask, bool check)
1697 return set_cpus_allowed_ptr(p, new_mask);
1700 #endif /* CONFIG_SMP */
1703 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1705 #ifdef CONFIG_SCHEDSTATS
1706 struct rq *rq = this_rq();
1709 int this_cpu = smp_processor_id();
1711 if (cpu == this_cpu) {
1712 schedstat_inc(rq, ttwu_local);
1713 schedstat_inc(p, se.statistics.nr_wakeups_local);
1715 struct sched_domain *sd;
1717 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1719 for_each_domain(this_cpu, sd) {
1720 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1721 schedstat_inc(sd, ttwu_wake_remote);
1728 if (wake_flags & WF_MIGRATED)
1729 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1731 #endif /* CONFIG_SMP */
1733 schedstat_inc(rq, ttwu_count);
1734 schedstat_inc(p, se.statistics.nr_wakeups);
1736 if (wake_flags & WF_SYNC)
1737 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1739 #endif /* CONFIG_SCHEDSTATS */
1742 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1744 activate_task(rq, p, en_flags);
1745 p->on_rq = TASK_ON_RQ_QUEUED;
1747 /* if a worker is waking up, notify workqueue */
1748 if (p->flags & PF_WQ_WORKER)
1749 wq_worker_waking_up(p, cpu_of(rq));
1753 * Mark the task runnable and perform wakeup-preemption.
1756 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1758 check_preempt_curr(rq, p, wake_flags);
1759 p->state = TASK_RUNNING;
1760 trace_sched_wakeup(p);
1763 if (p->sched_class->task_woken) {
1765 * Our task @p is fully woken up and running; so its safe to
1766 * drop the rq->lock, hereafter rq is only used for statistics.
1768 lockdep_unpin_lock(&rq->lock);
1769 p->sched_class->task_woken(rq, p);
1770 lockdep_pin_lock(&rq->lock);
1773 if (rq->idle_stamp) {
1774 u64 delta = rq_clock(rq) - rq->idle_stamp;
1775 u64 max = 2*rq->max_idle_balance_cost;
1777 update_avg(&rq->avg_idle, delta);
1779 if (rq->avg_idle > max)
1788 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1790 lockdep_assert_held(&rq->lock);
1793 if (p->sched_contributes_to_load)
1794 rq->nr_uninterruptible--;
1797 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1798 ttwu_do_wakeup(rq, p, wake_flags);
1802 * Called in case the task @p isn't fully descheduled from its runqueue,
1803 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1804 * since all we need to do is flip p->state to TASK_RUNNING, since
1805 * the task is still ->on_rq.
1807 static int ttwu_remote(struct task_struct *p, int wake_flags)
1812 rq = __task_rq_lock(p);
1813 if (task_on_rq_queued(p)) {
1814 /* check_preempt_curr() may use rq clock */
1815 update_rq_clock(rq);
1816 ttwu_do_wakeup(rq, p, wake_flags);
1819 __task_rq_unlock(rq);
1825 void sched_ttwu_pending(void)
1827 struct rq *rq = this_rq();
1828 struct llist_node *llist = llist_del_all(&rq->wake_list);
1829 struct task_struct *p;
1830 unsigned long flags;
1835 raw_spin_lock_irqsave(&rq->lock, flags);
1836 lockdep_pin_lock(&rq->lock);
1839 p = llist_entry(llist, struct task_struct, wake_entry);
1840 llist = llist_next(llist);
1841 ttwu_do_activate(rq, p, 0);
1844 lockdep_unpin_lock(&rq->lock);
1845 raw_spin_unlock_irqrestore(&rq->lock, flags);
1848 void scheduler_ipi(void)
1851 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1852 * TIF_NEED_RESCHED remotely (for the first time) will also send
1855 preempt_fold_need_resched();
1857 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1861 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1862 * traditionally all their work was done from the interrupt return
1863 * path. Now that we actually do some work, we need to make sure
1866 * Some archs already do call them, luckily irq_enter/exit nest
1869 * Arguably we should visit all archs and update all handlers,
1870 * however a fair share of IPIs are still resched only so this would
1871 * somewhat pessimize the simple resched case.
1874 sched_ttwu_pending();
1877 * Check if someone kicked us for doing the nohz idle load balance.
1879 if (unlikely(got_nohz_idle_kick())) {
1880 this_rq()->idle_balance = 1;
1881 raise_softirq_irqoff(SCHED_SOFTIRQ);
1886 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1888 struct rq *rq = cpu_rq(cpu);
1890 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1891 if (!set_nr_if_polling(rq->idle))
1892 smp_send_reschedule(cpu);
1894 trace_sched_wake_idle_without_ipi(cpu);
1898 void wake_up_if_idle(int cpu)
1900 struct rq *rq = cpu_rq(cpu);
1901 unsigned long flags;
1905 if (!is_idle_task(rcu_dereference(rq->curr)))
1908 if (set_nr_if_polling(rq->idle)) {
1909 trace_sched_wake_idle_without_ipi(cpu);
1911 raw_spin_lock_irqsave(&rq->lock, flags);
1912 if (is_idle_task(rq->curr))
1913 smp_send_reschedule(cpu);
1914 /* Else cpu is not in idle, do nothing here */
1915 raw_spin_unlock_irqrestore(&rq->lock, flags);
1922 bool cpus_share_cache(int this_cpu, int that_cpu)
1924 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1926 #endif /* CONFIG_SMP */
1928 static void ttwu_queue(struct task_struct *p, int cpu)
1930 struct rq *rq = cpu_rq(cpu);
1932 #if defined(CONFIG_SMP)
1933 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1934 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1935 ttwu_queue_remote(p, cpu);
1940 raw_spin_lock(&rq->lock);
1941 lockdep_pin_lock(&rq->lock);
1942 ttwu_do_activate(rq, p, 0);
1943 lockdep_unpin_lock(&rq->lock);
1944 raw_spin_unlock(&rq->lock);
1948 * try_to_wake_up - wake up a thread
1949 * @p: the thread to be awakened
1950 * @state: the mask of task states that can be woken
1951 * @wake_flags: wake modifier flags (WF_*)
1952 * @sibling_count_hint: A hint at the number of threads that are being woken up
1955 * Put it on the run-queue if it's not already there. The "current"
1956 * thread is always on the run-queue (except when the actual
1957 * re-schedule is in progress), and as such you're allowed to do
1958 * the simpler "current->state = TASK_RUNNING" to mark yourself
1959 * runnable without the overhead of this.
1961 * Return: %true if @p was woken up, %false if it was already running.
1962 * or @state didn't match @p's state.
1965 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags,
1966 int sibling_count_hint)
1968 unsigned long flags;
1969 int cpu, success = 0;
1976 * If we are going to wake up a thread waiting for CONDITION we
1977 * need to ensure that CONDITION=1 done by the caller can not be
1978 * reordered with p->state check below. This pairs with mb() in
1979 * set_current_state() the waiting thread does.
1981 smp_mb__before_spinlock();
1982 raw_spin_lock_irqsave(&p->pi_lock, flags);
1983 if (!(p->state & state))
1986 trace_sched_waking(p);
1988 success = 1; /* we're going to change ->state */
1992 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1993 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1994 * in smp_cond_load_acquire() below.
1996 * sched_ttwu_pending() try_to_wake_up()
1997 * [S] p->on_rq = 1; [L] P->state
1998 * UNLOCK rq->lock -----.
2002 * LOCK rq->lock -----'
2006 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2008 * Pairs with the UNLOCK+LOCK on rq->lock from the
2009 * last wakeup of our task and the schedule that got our task
2013 if (p->on_rq && ttwu_remote(p, wake_flags))
2018 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2019 * possible to, falsely, observe p->on_cpu == 0.
2021 * One must be running (->on_cpu == 1) in order to remove oneself
2022 * from the runqueue.
2024 * [S] ->on_cpu = 1; [L] ->on_rq
2028 * [S] ->on_rq = 0; [L] ->on_cpu
2030 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2031 * from the consecutive calls to schedule(); the first switching to our
2032 * task, the second putting it to sleep.
2037 * If the owning (remote) cpu is still in the middle of schedule() with
2038 * this task as prev, wait until its done referencing the task.
2043 * Combined with the control dependency above, we have an effective
2044 * smp_load_acquire() without the need for full barriers.
2046 * Pairs with the smp_store_release() in finish_lock_switch().
2048 * This ensures that tasks getting woken will be fully ordered against
2049 * their previous state and preserve Program Order.
2053 rq = cpu_rq(task_cpu(p));
2055 raw_spin_lock(&rq->lock);
2056 wallclock = walt_ktime_clock();
2057 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2058 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2059 raw_spin_unlock(&rq->lock);
2061 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2062 p->state = TASK_WAKING;
2064 if (p->sched_class->task_waking)
2065 p->sched_class->task_waking(p);
2067 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags,
2068 sibling_count_hint);
2069 if (task_cpu(p) != cpu) {
2070 wake_flags |= WF_MIGRATED;
2071 set_task_cpu(p, cpu);
2074 #endif /* CONFIG_SMP */
2078 ttwu_stat(p, cpu, wake_flags);
2080 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2086 * try_to_wake_up_local - try to wake up a local task with rq lock held
2087 * @p: the thread to be awakened
2089 * Put @p on the run-queue if it's not already there. The caller must
2090 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2093 static void try_to_wake_up_local(struct task_struct *p)
2095 struct rq *rq = task_rq(p);
2097 if (WARN_ON_ONCE(rq != this_rq()) ||
2098 WARN_ON_ONCE(p == current))
2101 lockdep_assert_held(&rq->lock);
2103 if (!raw_spin_trylock(&p->pi_lock)) {
2105 * This is OK, because current is on_cpu, which avoids it being
2106 * picked for load-balance and preemption/IRQs are still
2107 * disabled avoiding further scheduler activity on it and we've
2108 * not yet picked a replacement task.
2110 lockdep_unpin_lock(&rq->lock);
2111 raw_spin_unlock(&rq->lock);
2112 raw_spin_lock(&p->pi_lock);
2113 raw_spin_lock(&rq->lock);
2114 lockdep_pin_lock(&rq->lock);
2117 if (!(p->state & TASK_NORMAL))
2120 trace_sched_waking(p);
2122 if (!task_on_rq_queued(p)) {
2123 u64 wallclock = walt_ktime_clock();
2125 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2126 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2127 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2130 ttwu_do_wakeup(rq, p, 0);
2131 ttwu_stat(p, smp_processor_id(), 0);
2133 raw_spin_unlock(&p->pi_lock);
2137 * wake_up_process - Wake up a specific process
2138 * @p: The process to be woken up.
2140 * Attempt to wake up the nominated process and move it to the set of runnable
2143 * Return: 1 if the process was woken up, 0 if it was already running.
2145 * It may be assumed that this function implies a write memory barrier before
2146 * changing the task state if and only if any tasks are woken up.
2148 int wake_up_process(struct task_struct *p)
2150 return try_to_wake_up(p, TASK_NORMAL, 0, 1);
2152 EXPORT_SYMBOL(wake_up_process);
2154 int wake_up_state(struct task_struct *p, unsigned int state)
2156 return try_to_wake_up(p, state, 0, 1);
2160 * This function clears the sched_dl_entity static params.
2162 void __dl_clear_params(struct task_struct *p)
2164 struct sched_dl_entity *dl_se = &p->dl;
2166 dl_se->dl_runtime = 0;
2167 dl_se->dl_deadline = 0;
2168 dl_se->dl_period = 0;
2172 dl_se->dl_throttled = 0;
2174 dl_se->dl_yielded = 0;
2178 * Perform scheduler related setup for a newly forked process p.
2179 * p is forked by current.
2181 * __sched_fork() is basic setup used by init_idle() too:
2183 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2188 p->se.exec_start = 0;
2189 p->se.sum_exec_runtime = 0;
2190 p->se.prev_sum_exec_runtime = 0;
2191 p->se.nr_migrations = 0;
2193 #ifdef CONFIG_SCHED_WALT
2194 p->last_sleep_ts = 0;
2197 INIT_LIST_HEAD(&p->se.group_node);
2198 walt_init_new_task_load(p);
2200 #ifdef CONFIG_FAIR_GROUP_SCHED
2201 p->se.cfs_rq = NULL;
2204 #ifdef CONFIG_SCHEDSTATS
2205 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2208 RB_CLEAR_NODE(&p->dl.rb_node);
2209 init_dl_task_timer(&p->dl);
2210 __dl_clear_params(p);
2212 INIT_LIST_HEAD(&p->rt.run_list);
2214 #ifdef CONFIG_PREEMPT_NOTIFIERS
2215 INIT_HLIST_HEAD(&p->preempt_notifiers);
2218 #ifdef CONFIG_NUMA_BALANCING
2219 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2220 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2221 p->mm->numa_scan_seq = 0;
2224 if (clone_flags & CLONE_VM)
2225 p->numa_preferred_nid = current->numa_preferred_nid;
2227 p->numa_preferred_nid = -1;
2229 p->node_stamp = 0ULL;
2230 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2231 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2232 p->numa_work.next = &p->numa_work;
2233 p->numa_faults = NULL;
2234 p->last_task_numa_placement = 0;
2235 p->last_sum_exec_runtime = 0;
2237 p->numa_group = NULL;
2238 #endif /* CONFIG_NUMA_BALANCING */
2241 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2243 #ifdef CONFIG_NUMA_BALANCING
2245 void set_numabalancing_state(bool enabled)
2248 static_branch_enable(&sched_numa_balancing);
2250 static_branch_disable(&sched_numa_balancing);
2253 #ifdef CONFIG_PROC_SYSCTL
2254 int sysctl_numa_balancing(struct ctl_table *table, int write,
2255 void __user *buffer, size_t *lenp, loff_t *ppos)
2259 int state = static_branch_likely(&sched_numa_balancing);
2261 if (write && !capable(CAP_SYS_ADMIN))
2266 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2270 set_numabalancing_state(state);
2277 * fork()/clone()-time setup:
2279 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2281 unsigned long flags;
2282 int cpu = get_cpu();
2284 __sched_fork(clone_flags, p);
2286 * We mark the process as NEW here. This guarantees that
2287 * nobody will actually run it, and a signal or other external
2288 * event cannot wake it up and insert it on the runqueue either.
2290 p->state = TASK_NEW;
2293 * Make sure we do not leak PI boosting priority to the child.
2295 p->prio = current->normal_prio;
2298 * Revert to default priority/policy on fork if requested.
2300 if (unlikely(p->sched_reset_on_fork)) {
2301 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2302 p->policy = SCHED_NORMAL;
2303 p->static_prio = NICE_TO_PRIO(0);
2305 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2306 p->static_prio = NICE_TO_PRIO(0);
2308 p->prio = p->normal_prio = __normal_prio(p);
2312 * We don't need the reset flag anymore after the fork. It has
2313 * fulfilled its duty:
2315 p->sched_reset_on_fork = 0;
2318 if (dl_prio(p->prio)) {
2321 } else if (rt_prio(p->prio)) {
2322 p->sched_class = &rt_sched_class;
2324 p->sched_class = &fair_sched_class;
2327 init_entity_runnable_average(&p->se);
2330 * The child is not yet in the pid-hash so no cgroup attach races,
2331 * and the cgroup is pinned to this child due to cgroup_fork()
2332 * is ran before sched_fork().
2334 * Silence PROVE_RCU.
2336 raw_spin_lock_irqsave(&p->pi_lock, flags);
2338 * We're setting the cpu for the first time, we don't migrate,
2339 * so use __set_task_cpu().
2341 __set_task_cpu(p, cpu);
2342 if (p->sched_class->task_fork)
2343 p->sched_class->task_fork(p);
2344 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2346 #ifdef CONFIG_SCHED_INFO
2347 if (likely(sched_info_on()))
2348 memset(&p->sched_info, 0, sizeof(p->sched_info));
2350 #if defined(CONFIG_SMP)
2353 init_task_preempt_count(p);
2355 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2356 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2363 unsigned long to_ratio(u64 period, u64 runtime)
2365 if (runtime == RUNTIME_INF)
2369 * Doing this here saves a lot of checks in all
2370 * the calling paths, and returning zero seems
2371 * safe for them anyway.
2376 return div64_u64(runtime << 20, period);
2380 inline struct dl_bw *dl_bw_of(int i)
2382 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2383 "sched RCU must be held");
2384 return &cpu_rq(i)->rd->dl_bw;
2387 static inline int dl_bw_cpus(int i)
2389 struct root_domain *rd = cpu_rq(i)->rd;
2392 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2393 "sched RCU must be held");
2394 for_each_cpu_and(i, rd->span, cpu_active_mask)
2400 inline struct dl_bw *dl_bw_of(int i)
2402 return &cpu_rq(i)->dl.dl_bw;
2405 static inline int dl_bw_cpus(int i)
2412 * We must be sure that accepting a new task (or allowing changing the
2413 * parameters of an existing one) is consistent with the bandwidth
2414 * constraints. If yes, this function also accordingly updates the currently
2415 * allocated bandwidth to reflect the new situation.
2417 * This function is called while holding p's rq->lock.
2419 * XXX we should delay bw change until the task's 0-lag point, see
2422 static int dl_overflow(struct task_struct *p, int policy,
2423 const struct sched_attr *attr)
2426 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2427 u64 period = attr->sched_period ?: attr->sched_deadline;
2428 u64 runtime = attr->sched_runtime;
2429 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2432 if (new_bw == p->dl.dl_bw)
2436 * Either if a task, enters, leave, or stays -deadline but changes
2437 * its parameters, we may need to update accordingly the total
2438 * allocated bandwidth of the container.
2440 raw_spin_lock(&dl_b->lock);
2441 cpus = dl_bw_cpus(task_cpu(p));
2442 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2443 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2444 __dl_add(dl_b, new_bw);
2446 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2447 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2448 __dl_clear(dl_b, p->dl.dl_bw);
2449 __dl_add(dl_b, new_bw);
2451 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2452 __dl_clear(dl_b, p->dl.dl_bw);
2455 raw_spin_unlock(&dl_b->lock);
2460 extern void init_dl_bw(struct dl_bw *dl_b);
2463 * wake_up_new_task - wake up a newly created task for the first time.
2465 * This function will do some initial scheduler statistics housekeeping
2466 * that must be done for every newly created context, then puts the task
2467 * on the runqueue and wakes it.
2469 void wake_up_new_task(struct task_struct *p)
2471 unsigned long flags;
2474 raw_spin_lock_irqsave(&p->pi_lock, flags);
2475 p->state = TASK_RUNNING;
2477 walt_init_new_task_load(p);
2479 /* Initialize new task's runnable average */
2480 init_entity_runnable_average(&p->se);
2483 * Fork balancing, do it here and not earlier because:
2484 * - cpus_allowed can change in the fork path
2485 * - any previously selected cpu might disappear through hotplug
2487 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2488 * as we're not fully set-up yet.
2490 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0, 1));
2492 rq = __task_rq_lock(p);
2493 update_rq_clock(rq);
2494 post_init_entity_util_avg(&p->se);
2496 walt_mark_task_starting(p);
2497 activate_task(rq, p, ENQUEUE_WAKEUP_NEW);
2498 p->on_rq = TASK_ON_RQ_QUEUED;
2499 trace_sched_wakeup_new(p);
2500 check_preempt_curr(rq, p, WF_FORK);
2502 if (p->sched_class->task_woken) {
2504 * Nothing relies on rq->lock after this, so its fine to
2507 lockdep_unpin_lock(&rq->lock);
2508 p->sched_class->task_woken(rq, p);
2509 lockdep_pin_lock(&rq->lock);
2512 task_rq_unlock(rq, p, &flags);
2515 #ifdef CONFIG_PREEMPT_NOTIFIERS
2517 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2519 void preempt_notifier_inc(void)
2521 static_key_slow_inc(&preempt_notifier_key);
2523 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2525 void preempt_notifier_dec(void)
2527 static_key_slow_dec(&preempt_notifier_key);
2529 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2532 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2533 * @notifier: notifier struct to register
2535 void preempt_notifier_register(struct preempt_notifier *notifier)
2537 if (!static_key_false(&preempt_notifier_key))
2538 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2540 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2542 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2545 * preempt_notifier_unregister - no longer interested in preemption notifications
2546 * @notifier: notifier struct to unregister
2548 * This is *not* safe to call from within a preemption notifier.
2550 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2552 hlist_del(¬ifier->link);
2554 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2556 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2558 struct preempt_notifier *notifier;
2560 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2561 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2564 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2566 if (static_key_false(&preempt_notifier_key))
2567 __fire_sched_in_preempt_notifiers(curr);
2571 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2572 struct task_struct *next)
2574 struct preempt_notifier *notifier;
2576 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2577 notifier->ops->sched_out(notifier, next);
2580 static __always_inline void
2581 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2582 struct task_struct *next)
2584 if (static_key_false(&preempt_notifier_key))
2585 __fire_sched_out_preempt_notifiers(curr, next);
2588 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2590 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2595 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2596 struct task_struct *next)
2600 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2603 * prepare_task_switch - prepare to switch tasks
2604 * @rq: the runqueue preparing to switch
2605 * @prev: the current task that is being switched out
2606 * @next: the task we are going to switch to.
2608 * This is called with the rq lock held and interrupts off. It must
2609 * be paired with a subsequent finish_task_switch after the context
2612 * prepare_task_switch sets up locking and calls architecture specific
2616 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2617 struct task_struct *next)
2619 sched_info_switch(rq, prev, next);
2620 perf_event_task_sched_out(prev, next);
2621 fire_sched_out_preempt_notifiers(prev, next);
2622 prepare_lock_switch(rq, next);
2623 prepare_arch_switch(next);
2627 * finish_task_switch - clean up after a task-switch
2628 * @prev: the thread we just switched away from.
2630 * finish_task_switch must be called after the context switch, paired
2631 * with a prepare_task_switch call before the context switch.
2632 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2633 * and do any other architecture-specific cleanup actions.
2635 * Note that we may have delayed dropping an mm in context_switch(). If
2636 * so, we finish that here outside of the runqueue lock. (Doing it
2637 * with the lock held can cause deadlocks; see schedule() for
2640 * The context switch have flipped the stack from under us and restored the
2641 * local variables which were saved when this task called schedule() in the
2642 * past. prev == current is still correct but we need to recalculate this_rq
2643 * because prev may have moved to another CPU.
2645 static struct rq *finish_task_switch(struct task_struct *prev)
2646 __releases(rq->lock)
2648 struct rq *rq = this_rq();
2649 struct mm_struct *mm = rq->prev_mm;
2653 * The previous task will have left us with a preempt_count of 2
2654 * because it left us after:
2657 * preempt_disable(); // 1
2659 * raw_spin_lock_irq(&rq->lock) // 2
2661 * Also, see FORK_PREEMPT_COUNT.
2663 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2664 "corrupted preempt_count: %s/%d/0x%x\n",
2665 current->comm, current->pid, preempt_count()))
2666 preempt_count_set(FORK_PREEMPT_COUNT);
2671 * A task struct has one reference for the use as "current".
2672 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2673 * schedule one last time. The schedule call will never return, and
2674 * the scheduled task must drop that reference.
2676 * We must observe prev->state before clearing prev->on_cpu (in
2677 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2678 * running on another CPU and we could rave with its RUNNING -> DEAD
2679 * transition, resulting in a double drop.
2681 prev_state = prev->state;
2682 vtime_task_switch(prev);
2683 perf_event_task_sched_in(prev, current);
2684 finish_lock_switch(rq, prev);
2685 finish_arch_post_lock_switch();
2687 fire_sched_in_preempt_notifiers(current);
2690 if (unlikely(prev_state == TASK_DEAD)) {
2691 if (prev->sched_class->task_dead)
2692 prev->sched_class->task_dead(prev);
2695 * Remove function-return probe instances associated with this
2696 * task and put them back on the free list.
2698 kprobe_flush_task(prev);
2699 put_task_struct(prev);
2702 tick_nohz_task_switch();
2708 /* rq->lock is NOT held, but preemption is disabled */
2709 static void __balance_callback(struct rq *rq)
2711 struct callback_head *head, *next;
2712 void (*func)(struct rq *rq);
2713 unsigned long flags;
2715 raw_spin_lock_irqsave(&rq->lock, flags);
2716 head = rq->balance_callback;
2717 rq->balance_callback = NULL;
2719 func = (void (*)(struct rq *))head->func;
2726 raw_spin_unlock_irqrestore(&rq->lock, flags);
2729 static inline void balance_callback(struct rq *rq)
2731 if (unlikely(rq->balance_callback))
2732 __balance_callback(rq);
2737 static inline void balance_callback(struct rq *rq)
2744 * schedule_tail - first thing a freshly forked thread must call.
2745 * @prev: the thread we just switched away from.
2747 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2748 __releases(rq->lock)
2753 * New tasks start with FORK_PREEMPT_COUNT, see there and
2754 * finish_task_switch() for details.
2756 * finish_task_switch() will drop rq->lock() and lower preempt_count
2757 * and the preempt_enable() will end up enabling preemption (on
2758 * PREEMPT_COUNT kernels).
2761 rq = finish_task_switch(prev);
2762 balance_callback(rq);
2765 if (current->set_child_tid)
2766 put_user(task_pid_vnr(current), current->set_child_tid);
2770 * context_switch - switch to the new MM and the new thread's register state.
2772 static inline struct rq *
2773 context_switch(struct rq *rq, struct task_struct *prev,
2774 struct task_struct *next)
2776 struct mm_struct *mm, *oldmm;
2778 prepare_task_switch(rq, prev, next);
2781 oldmm = prev->active_mm;
2783 * For paravirt, this is coupled with an exit in switch_to to
2784 * combine the page table reload and the switch backend into
2787 arch_start_context_switch(prev);
2790 next->active_mm = oldmm;
2791 atomic_inc(&oldmm->mm_count);
2792 enter_lazy_tlb(oldmm, next);
2794 switch_mm(oldmm, mm, next);
2797 prev->active_mm = NULL;
2798 rq->prev_mm = oldmm;
2801 * Since the runqueue lock will be released by the next
2802 * task (which is an invalid locking op but in the case
2803 * of the scheduler it's an obvious special-case), so we
2804 * do an early lockdep release here:
2806 lockdep_unpin_lock(&rq->lock);
2807 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2809 /* Here we just switch the register state and the stack. */
2810 switch_to(prev, next, prev);
2813 return finish_task_switch(prev);
2817 * nr_running and nr_context_switches:
2819 * externally visible scheduler statistics: current number of runnable
2820 * threads, total number of context switches performed since bootup.
2822 unsigned long nr_running(void)
2824 unsigned long i, sum = 0;
2826 for_each_online_cpu(i)
2827 sum += cpu_rq(i)->nr_running;
2833 * Check if only the current task is running on the cpu.
2835 * Caution: this function does not check that the caller has disabled
2836 * preemption, thus the result might have a time-of-check-to-time-of-use
2837 * race. The caller is responsible to use it correctly, for example:
2839 * - from a non-preemptable section (of course)
2841 * - from a thread that is bound to a single CPU
2843 * - in a loop with very short iterations (e.g. a polling loop)
2845 bool single_task_running(void)
2847 return raw_rq()->nr_running == 1;
2849 EXPORT_SYMBOL(single_task_running);
2851 unsigned long long nr_context_switches(void)
2854 unsigned long long sum = 0;
2856 for_each_possible_cpu(i)
2857 sum += cpu_rq(i)->nr_switches;
2862 unsigned long nr_iowait(void)
2864 unsigned long i, sum = 0;
2866 for_each_possible_cpu(i)
2867 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2872 unsigned long nr_iowait_cpu(int cpu)
2874 struct rq *this = cpu_rq(cpu);
2875 return atomic_read(&this->nr_iowait);
2878 #ifdef CONFIG_CPU_QUIET
2879 u64 nr_running_integral(unsigned int cpu)
2881 unsigned int seqcnt;
2885 if (cpu >= nr_cpu_ids)
2891 * Update average to avoid reading stalled value if there were
2892 * no run-queue changes for a long time. On the other hand if
2893 * the changes are happening right now, just read current value
2897 seqcnt = read_seqcount_begin(&q->ave_seqcnt);
2898 integral = do_nr_running_integral(q);
2899 if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
2900 read_seqcount_begin(&q->ave_seqcnt);
2901 integral = q->nr_running_integral;
2908 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2910 struct rq *rq = this_rq();
2911 *nr_waiters = atomic_read(&rq->nr_iowait);
2912 *load = rq->load.weight;
2918 * sched_exec - execve() is a valuable balancing opportunity, because at
2919 * this point the task has the smallest effective memory and cache footprint.
2921 void sched_exec(void)
2923 struct task_struct *p = current;
2924 unsigned long flags;
2927 raw_spin_lock_irqsave(&p->pi_lock, flags);
2928 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0, 1);
2929 if (dest_cpu == smp_processor_id())
2932 if (likely(cpu_active(dest_cpu))) {
2933 struct migration_arg arg = { p, dest_cpu };
2935 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2936 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2940 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2945 DEFINE_PER_CPU(struct kernel_stat, kstat);
2946 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2948 EXPORT_PER_CPU_SYMBOL(kstat);
2949 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2952 * Return accounted runtime for the task.
2953 * In case the task is currently running, return the runtime plus current's
2954 * pending runtime that have not been accounted yet.
2956 unsigned long long task_sched_runtime(struct task_struct *p)
2958 unsigned long flags;
2962 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2964 * 64-bit doesn't need locks to atomically read a 64bit value.
2965 * So we have a optimization chance when the task's delta_exec is 0.
2966 * Reading ->on_cpu is racy, but this is ok.
2968 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2969 * If we race with it entering cpu, unaccounted time is 0. This is
2970 * indistinguishable from the read occurring a few cycles earlier.
2971 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2972 * been accounted, so we're correct here as well.
2974 if (!p->on_cpu || !task_on_rq_queued(p))
2975 return p->se.sum_exec_runtime;
2978 rq = task_rq_lock(p, &flags);
2980 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2981 * project cycles that may never be accounted to this
2982 * thread, breaking clock_gettime().
2984 if (task_current(rq, p) && task_on_rq_queued(p)) {
2985 update_rq_clock(rq);
2986 p->sched_class->update_curr(rq);
2988 ns = p->se.sum_exec_runtime;
2989 task_rq_unlock(rq, p, &flags);
2994 #ifdef CONFIG_CPU_FREQ_GOV_SCHED
2997 unsigned long add_capacity_margin(unsigned long cpu_capacity)
2999 cpu_capacity = cpu_capacity * capacity_margin;
3000 cpu_capacity /= SCHED_CAPACITY_SCALE;
3001 return cpu_capacity;
3005 unsigned long sum_capacity_reqs(unsigned long cfs_cap,
3006 struct sched_capacity_reqs *scr)
3008 unsigned long total = add_capacity_margin(cfs_cap + scr->rt);
3009 return total += scr->dl;
3012 unsigned long boosted_cpu_util(int cpu);
3013 static void sched_freq_tick_pelt(int cpu)
3015 unsigned long cpu_utilization = boosted_cpu_util(cpu);
3016 unsigned long capacity_curr = capacity_curr_of(cpu);
3017 struct sched_capacity_reqs *scr;
3019 scr = &per_cpu(cpu_sched_capacity_reqs, cpu);
3020 if (sum_capacity_reqs(cpu_utilization, scr) < capacity_curr)
3024 * To make free room for a task that is building up its "real"
3025 * utilization and to harm its performance the least, request
3026 * a jump to a higher OPP as soon as the margin of free capacity
3027 * is impacted (specified by capacity_margin).
3028 * Remember CPU utilization in sched_capacity_reqs should be normalised.
3030 cpu_utilization = cpu_utilization * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
3031 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
3034 #ifdef CONFIG_SCHED_WALT
3035 static void sched_freq_tick_walt(int cpu)
3037 unsigned long cpu_utilization = cpu_util_freq(cpu);
3038 unsigned long capacity_curr = capacity_curr_of(cpu);
3040 if (walt_disabled || !sysctl_sched_use_walt_cpu_util)
3041 return sched_freq_tick_pelt(cpu);
3044 * Add a margin to the WALT utilization to check if we will need to
3045 * increase frequency.
3046 * NOTE: WALT tracks a single CPU signal for all the scheduling
3047 * classes, thus this margin is going to be added to the DL class as
3048 * well, which is something we do not do in sched_freq_tick_pelt case.
3050 if (add_capacity_margin(cpu_utilization) <= capacity_curr)
3054 * It is likely that the load is growing so we
3055 * keep the added margin in our request as an
3057 * Remember CPU utilization in sched_capacity_reqs should be normalised.
3059 cpu_utilization = cpu_utilization * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
3060 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
3063 #define _sched_freq_tick(cpu) sched_freq_tick_walt(cpu)
3065 #define _sched_freq_tick(cpu) sched_freq_tick_pelt(cpu)
3066 #endif /* CONFIG_SCHED_WALT */
3068 static void sched_freq_tick(int cpu)
3073 _sched_freq_tick(cpu);
3076 static inline void sched_freq_tick(int cpu) { }
3077 #endif /* CONFIG_CPU_FREQ_GOV_SCHED */
3080 * This function gets called by the timer code, with HZ frequency.
3081 * We call it with interrupts disabled.
3083 void scheduler_tick(void)
3085 int cpu = smp_processor_id();
3086 struct rq *rq = cpu_rq(cpu);
3087 struct task_struct *curr = rq->curr;
3091 raw_spin_lock(&rq->lock);
3092 walt_set_window_start(rq);
3093 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
3094 walt_ktime_clock(), 0);
3095 update_rq_clock(rq);
3096 curr->sched_class->task_tick(rq, curr, 0);
3097 update_cpu_load_active(rq);
3098 calc_global_load_tick(rq);
3099 sched_freq_tick(cpu);
3100 raw_spin_unlock(&rq->lock);
3102 perf_event_task_tick();
3105 rq->idle_balance = idle_cpu(cpu);
3106 trigger_load_balance(rq);
3108 rq_last_tick_reset(rq);
3111 #ifdef CONFIG_NO_HZ_FULL
3113 * scheduler_tick_max_deferment
3115 * Keep at least one tick per second when a single
3116 * active task is running because the scheduler doesn't
3117 * yet completely support full dynticks environment.
3119 * This makes sure that uptime, CFS vruntime, load
3120 * balancing, etc... continue to move forward, even
3121 * with a very low granularity.
3123 * Return: Maximum deferment in nanoseconds.
3125 u64 scheduler_tick_max_deferment(void)
3127 struct rq *rq = this_rq();
3128 unsigned long next, now = READ_ONCE(jiffies);
3130 next = rq->last_sched_tick + HZ;
3132 if (time_before_eq(next, now))
3135 return jiffies_to_nsecs(next - now);
3139 notrace unsigned long get_parent_ip(unsigned long addr)
3141 if (in_lock_functions(addr)) {
3142 addr = CALLER_ADDR2;
3143 if (in_lock_functions(addr))
3144 addr = CALLER_ADDR3;
3149 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3150 defined(CONFIG_PREEMPT_TRACER))
3152 void preempt_count_add(int val)
3154 #ifdef CONFIG_DEBUG_PREEMPT
3158 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3161 __preempt_count_add(val);
3162 #ifdef CONFIG_DEBUG_PREEMPT
3164 * Spinlock count overflowing soon?
3166 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3169 if (preempt_count() == val) {
3170 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3171 #ifdef CONFIG_DEBUG_PREEMPT
3172 current->preempt_disable_ip = ip;
3174 trace_preempt_off(CALLER_ADDR0, ip);
3177 EXPORT_SYMBOL(preempt_count_add);
3178 NOKPROBE_SYMBOL(preempt_count_add);
3180 void preempt_count_sub(int val)
3182 #ifdef CONFIG_DEBUG_PREEMPT
3186 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3189 * Is the spinlock portion underflowing?
3191 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3192 !(preempt_count() & PREEMPT_MASK)))
3196 if (preempt_count() == val)
3197 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3198 __preempt_count_sub(val);
3200 EXPORT_SYMBOL(preempt_count_sub);
3201 NOKPROBE_SYMBOL(preempt_count_sub);
3206 * Print scheduling while atomic bug:
3208 static noinline void __schedule_bug(struct task_struct *prev)
3210 if (oops_in_progress)
3213 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3214 prev->comm, prev->pid, preempt_count());
3216 debug_show_held_locks(prev);
3218 if (irqs_disabled())
3219 print_irqtrace_events(prev);
3220 #ifdef CONFIG_DEBUG_PREEMPT
3221 if (in_atomic_preempt_off()) {
3222 pr_err("Preemption disabled at:");
3223 print_ip_sym(current->preempt_disable_ip);
3228 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3232 * Various schedule()-time debugging checks and statistics:
3234 static inline void schedule_debug(struct task_struct *prev)
3236 #ifdef CONFIG_SCHED_STACK_END_CHECK
3237 if (task_stack_end_corrupted(prev))
3238 panic("corrupted stack end detected inside scheduler\n");
3241 if (unlikely(in_atomic_preempt_off())) {
3242 __schedule_bug(prev);
3243 preempt_count_set(PREEMPT_DISABLED);
3247 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3249 schedstat_inc(this_rq(), sched_count);
3253 * Pick up the highest-prio task:
3255 static inline struct task_struct *
3256 pick_next_task(struct rq *rq, struct task_struct *prev)
3258 const struct sched_class *class = &fair_sched_class;
3259 struct task_struct *p;
3262 * Optimization: we know that if all tasks are in
3263 * the fair class we can call that function directly:
3265 if (likely(prev->sched_class == class &&
3266 rq->nr_running == rq->cfs.h_nr_running)) {
3267 p = fair_sched_class.pick_next_task(rq, prev);
3268 if (unlikely(p == RETRY_TASK))
3271 /* assumes fair_sched_class->next == idle_sched_class */
3273 p = idle_sched_class.pick_next_task(rq, prev);
3279 for_each_class(class) {
3280 p = class->pick_next_task(rq, prev);
3282 if (unlikely(p == RETRY_TASK))
3288 BUG(); /* the idle class will always have a runnable task */
3292 * __schedule() is the main scheduler function.
3294 * The main means of driving the scheduler and thus entering this function are:
3296 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3298 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3299 * paths. For example, see arch/x86/entry_64.S.
3301 * To drive preemption between tasks, the scheduler sets the flag in timer
3302 * interrupt handler scheduler_tick().
3304 * 3. Wakeups don't really cause entry into schedule(). They add a
3305 * task to the run-queue and that's it.
3307 * Now, if the new task added to the run-queue preempts the current
3308 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3309 * called on the nearest possible occasion:
3311 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3313 * - in syscall or exception context, at the next outmost
3314 * preempt_enable(). (this might be as soon as the wake_up()'s
3317 * - in IRQ context, return from interrupt-handler to
3318 * preemptible context
3320 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3323 * - cond_resched() call
3324 * - explicit schedule() call
3325 * - return from syscall or exception to user-space
3326 * - return from interrupt-handler to user-space
3328 * WARNING: must be called with preemption disabled!
3330 static void __sched notrace __schedule(bool preempt)
3332 struct task_struct *prev, *next;
3333 unsigned long *switch_count;
3338 cpu = smp_processor_id();
3340 rcu_note_context_switch();
3344 * do_exit() calls schedule() with preemption disabled as an exception;
3345 * however we must fix that up, otherwise the next task will see an
3346 * inconsistent (higher) preempt count.
3348 * It also avoids the below schedule_debug() test from complaining
3351 if (unlikely(prev->state == TASK_DEAD))
3352 preempt_enable_no_resched_notrace();
3354 schedule_debug(prev);
3356 if (sched_feat(HRTICK))
3360 * Make sure that signal_pending_state()->signal_pending() below
3361 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3362 * done by the caller to avoid the race with signal_wake_up().
3364 smp_mb__before_spinlock();
3365 raw_spin_lock_irq(&rq->lock);
3366 lockdep_pin_lock(&rq->lock);
3368 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3370 switch_count = &prev->nivcsw;
3371 if (!preempt && prev->state) {
3372 if (unlikely(signal_pending_state(prev->state, prev))) {
3373 prev->state = TASK_RUNNING;
3375 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3379 * If a worker went to sleep, notify and ask workqueue
3380 * whether it wants to wake up a task to maintain
3383 if (prev->flags & PF_WQ_WORKER) {
3384 struct task_struct *to_wakeup;
3386 to_wakeup = wq_worker_sleeping(prev, cpu);
3388 try_to_wake_up_local(to_wakeup);
3391 switch_count = &prev->nvcsw;
3394 if (task_on_rq_queued(prev))
3395 update_rq_clock(rq);
3397 next = pick_next_task(rq, prev);
3398 wallclock = walt_ktime_clock();
3399 walt_update_task_ravg(prev, rq, PUT_PREV_TASK, wallclock, 0);
3400 walt_update_task_ravg(next, rq, PICK_NEXT_TASK, wallclock, 0);
3401 clear_tsk_need_resched(prev);
3402 clear_preempt_need_resched();
3403 rq->clock_skip_update = 0;
3405 if (likely(prev != next)) {
3406 #ifdef CONFIG_SCHED_WALT
3408 prev->last_sleep_ts = wallclock;
3414 trace_sched_switch(preempt, prev, next);
3415 rq = context_switch(rq, prev, next); /* unlocks the rq */
3418 lockdep_unpin_lock(&rq->lock);
3419 raw_spin_unlock_irq(&rq->lock);
3422 balance_callback(rq);
3425 static inline void sched_submit_work(struct task_struct *tsk)
3427 if (!tsk->state || tsk_is_pi_blocked(tsk))
3430 * If we are going to sleep and we have plugged IO queued,
3431 * make sure to submit it to avoid deadlocks.
3433 if (blk_needs_flush_plug(tsk))
3434 blk_schedule_flush_plug(tsk);
3437 asmlinkage __visible void __sched schedule(void)
3439 struct task_struct *tsk = current;
3441 sched_submit_work(tsk);
3445 sched_preempt_enable_no_resched();
3446 } while (need_resched());
3448 EXPORT_SYMBOL(schedule);
3450 #ifdef CONFIG_CONTEXT_TRACKING
3451 asmlinkage __visible void __sched schedule_user(void)
3454 * If we come here after a random call to set_need_resched(),
3455 * or we have been woken up remotely but the IPI has not yet arrived,
3456 * we haven't yet exited the RCU idle mode. Do it here manually until
3457 * we find a better solution.
3459 * NB: There are buggy callers of this function. Ideally we
3460 * should warn if prev_state != CONTEXT_USER, but that will trigger
3461 * too frequently to make sense yet.
3463 enum ctx_state prev_state = exception_enter();
3465 exception_exit(prev_state);
3470 * schedule_preempt_disabled - called with preemption disabled
3472 * Returns with preemption disabled. Note: preempt_count must be 1
3474 void __sched schedule_preempt_disabled(void)
3476 sched_preempt_enable_no_resched();
3481 static void __sched notrace preempt_schedule_common(void)
3484 preempt_disable_notrace();
3486 preempt_enable_no_resched_notrace();
3489 * Check again in case we missed a preemption opportunity
3490 * between schedule and now.
3492 } while (need_resched());
3495 #ifdef CONFIG_PREEMPT
3497 * this is the entry point to schedule() from in-kernel preemption
3498 * off of preempt_enable. Kernel preemptions off return from interrupt
3499 * occur there and call schedule directly.
3501 asmlinkage __visible void __sched notrace preempt_schedule(void)
3504 * If there is a non-zero preempt_count or interrupts are disabled,
3505 * we do not want to preempt the current task. Just return..
3507 if (likely(!preemptible()))
3510 preempt_schedule_common();
3512 NOKPROBE_SYMBOL(preempt_schedule);
3513 EXPORT_SYMBOL(preempt_schedule);
3516 * preempt_schedule_notrace - preempt_schedule called by tracing
3518 * The tracing infrastructure uses preempt_enable_notrace to prevent
3519 * recursion and tracing preempt enabling caused by the tracing
3520 * infrastructure itself. But as tracing can happen in areas coming
3521 * from userspace or just about to enter userspace, a preempt enable
3522 * can occur before user_exit() is called. This will cause the scheduler
3523 * to be called when the system is still in usermode.
3525 * To prevent this, the preempt_enable_notrace will use this function
3526 * instead of preempt_schedule() to exit user context if needed before
3527 * calling the scheduler.
3529 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3531 enum ctx_state prev_ctx;
3533 if (likely(!preemptible()))
3537 preempt_disable_notrace();
3539 * Needs preempt disabled in case user_exit() is traced
3540 * and the tracer calls preempt_enable_notrace() causing
3541 * an infinite recursion.
3543 prev_ctx = exception_enter();
3545 exception_exit(prev_ctx);
3547 preempt_enable_no_resched_notrace();
3548 } while (need_resched());
3550 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3552 #endif /* CONFIG_PREEMPT */
3555 * this is the entry point to schedule() from kernel preemption
3556 * off of irq context.
3557 * Note, that this is called and return with irqs disabled. This will
3558 * protect us against recursive calling from irq.
3560 asmlinkage __visible void __sched preempt_schedule_irq(void)
3562 enum ctx_state prev_state;
3564 /* Catch callers which need to be fixed */
3565 BUG_ON(preempt_count() || !irqs_disabled());
3567 prev_state = exception_enter();
3573 local_irq_disable();
3574 sched_preempt_enable_no_resched();
3575 } while (need_resched());
3577 exception_exit(prev_state);
3580 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3583 return try_to_wake_up(curr->private, mode, wake_flags, 1);
3585 EXPORT_SYMBOL(default_wake_function);
3587 #ifdef CONFIG_RT_MUTEXES
3590 * rt_mutex_setprio - set the current priority of a task
3592 * @prio: prio value (kernel-internal form)
3594 * This function changes the 'effective' priority of a task. It does
3595 * not touch ->normal_prio like __setscheduler().
3597 * Used by the rt_mutex code to implement priority inheritance
3598 * logic. Call site only calls if the priority of the task changed.
3600 void rt_mutex_setprio(struct task_struct *p, int prio)
3602 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3604 const struct sched_class *prev_class;
3606 BUG_ON(prio > MAX_PRIO);
3608 rq = __task_rq_lock(p);
3609 update_rq_clock(rq);
3612 * Idle task boosting is a nono in general. There is one
3613 * exception, when PREEMPT_RT and NOHZ is active:
3615 * The idle task calls get_next_timer_interrupt() and holds
3616 * the timer wheel base->lock on the CPU and another CPU wants
3617 * to access the timer (probably to cancel it). We can safely
3618 * ignore the boosting request, as the idle CPU runs this code
3619 * with interrupts disabled and will complete the lock
3620 * protected section without being interrupted. So there is no
3621 * real need to boost.
3623 if (unlikely(p == rq->idle)) {
3624 WARN_ON(p != rq->curr);
3625 WARN_ON(p->pi_blocked_on);
3629 trace_sched_pi_setprio(p, prio);
3631 prev_class = p->sched_class;
3632 queued = task_on_rq_queued(p);
3633 running = task_current(rq, p);
3635 dequeue_task(rq, p, DEQUEUE_SAVE);
3637 put_prev_task(rq, p);
3640 * Boosting condition are:
3641 * 1. -rt task is running and holds mutex A
3642 * --> -dl task blocks on mutex A
3644 * 2. -dl task is running and holds mutex A
3645 * --> -dl task blocks on mutex A and could preempt the
3648 if (dl_prio(prio)) {
3649 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3650 if (!dl_prio(p->normal_prio) ||
3651 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3652 p->dl.dl_boosted = 1;
3653 enqueue_flag |= ENQUEUE_REPLENISH;
3655 p->dl.dl_boosted = 0;
3656 p->sched_class = &dl_sched_class;
3657 } else if (rt_prio(prio)) {
3658 if (dl_prio(oldprio))
3659 p->dl.dl_boosted = 0;
3661 enqueue_flag |= ENQUEUE_HEAD;
3662 p->sched_class = &rt_sched_class;
3664 if (dl_prio(oldprio))
3665 p->dl.dl_boosted = 0;
3666 if (rt_prio(oldprio))
3668 p->sched_class = &fair_sched_class;
3674 p->sched_class->set_curr_task(rq);
3676 enqueue_task(rq, p, enqueue_flag);
3678 check_class_changed(rq, p, prev_class, oldprio);
3680 preempt_disable(); /* avoid rq from going away on us */
3681 __task_rq_unlock(rq);
3683 balance_callback(rq);
3688 void set_user_nice(struct task_struct *p, long nice)
3690 int old_prio, delta, queued;
3691 unsigned long flags;
3694 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3697 * We have to be careful, if called from sys_setpriority(),
3698 * the task might be in the middle of scheduling on another CPU.
3700 rq = task_rq_lock(p, &flags);
3701 update_rq_clock(rq);
3704 * The RT priorities are set via sched_setscheduler(), but we still
3705 * allow the 'normal' nice value to be set - but as expected
3706 * it wont have any effect on scheduling until the task is
3707 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3709 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3710 p->static_prio = NICE_TO_PRIO(nice);
3713 queued = task_on_rq_queued(p);
3715 dequeue_task(rq, p, DEQUEUE_SAVE);
3717 p->static_prio = NICE_TO_PRIO(nice);
3720 p->prio = effective_prio(p);
3721 delta = p->prio - old_prio;
3724 enqueue_task(rq, p, ENQUEUE_RESTORE);
3726 * If the task increased its priority or is running and
3727 * lowered its priority, then reschedule its CPU:
3729 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3733 task_rq_unlock(rq, p, &flags);
3735 EXPORT_SYMBOL(set_user_nice);
3738 * can_nice - check if a task can reduce its nice value
3742 int can_nice(const struct task_struct *p, const int nice)
3744 /* convert nice value [19,-20] to rlimit style value [1,40] */
3745 int nice_rlim = nice_to_rlimit(nice);
3747 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3748 capable(CAP_SYS_NICE));
3751 #ifdef __ARCH_WANT_SYS_NICE
3754 * sys_nice - change the priority of the current process.
3755 * @increment: priority increment
3757 * sys_setpriority is a more generic, but much slower function that
3758 * does similar things.
3760 SYSCALL_DEFINE1(nice, int, increment)
3765 * Setpriority might change our priority at the same moment.
3766 * We don't have to worry. Conceptually one call occurs first
3767 * and we have a single winner.
3769 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3770 nice = task_nice(current) + increment;
3772 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3773 if (increment < 0 && !can_nice(current, nice))
3776 retval = security_task_setnice(current, nice);
3780 set_user_nice(current, nice);
3787 * task_prio - return the priority value of a given task.
3788 * @p: the task in question.
3790 * Return: The priority value as seen by users in /proc.
3791 * RT tasks are offset by -200. Normal tasks are centered
3792 * around 0, value goes from -16 to +15.
3794 int task_prio(const struct task_struct *p)
3796 return p->prio - MAX_RT_PRIO;
3800 * idle_cpu - is a given cpu idle currently?
3801 * @cpu: the processor in question.
3803 * Return: 1 if the CPU is currently idle. 0 otherwise.
3805 int idle_cpu(int cpu)
3807 struct rq *rq = cpu_rq(cpu);
3809 if (rq->curr != rq->idle)
3816 if (!llist_empty(&rq->wake_list))
3824 * idle_task - return the idle task for a given cpu.
3825 * @cpu: the processor in question.
3827 * Return: The idle task for the cpu @cpu.
3829 struct task_struct *idle_task(int cpu)
3831 return cpu_rq(cpu)->idle;
3835 * find_process_by_pid - find a process with a matching PID value.
3836 * @pid: the pid in question.
3838 * The task of @pid, if found. %NULL otherwise.
3840 static struct task_struct *find_process_by_pid(pid_t pid)
3842 return pid ? find_task_by_vpid(pid) : current;
3846 * This function initializes the sched_dl_entity of a newly becoming
3847 * SCHED_DEADLINE task.
3849 * Only the static values are considered here, the actual runtime and the
3850 * absolute deadline will be properly calculated when the task is enqueued
3851 * for the first time with its new policy.
3854 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3856 struct sched_dl_entity *dl_se = &p->dl;
3858 dl_se->dl_runtime = attr->sched_runtime;
3859 dl_se->dl_deadline = attr->sched_deadline;
3860 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3861 dl_se->flags = attr->sched_flags;
3862 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3865 * Changing the parameters of a task is 'tricky' and we're not doing
3866 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3868 * What we SHOULD do is delay the bandwidth release until the 0-lag
3869 * point. This would include retaining the task_struct until that time
3870 * and change dl_overflow() to not immediately decrement the current
3873 * Instead we retain the current runtime/deadline and let the new
3874 * parameters take effect after the current reservation period lapses.
3875 * This is safe (albeit pessimistic) because the 0-lag point is always
3876 * before the current scheduling deadline.
3878 * We can still have temporary overloads because we do not delay the
3879 * change in bandwidth until that time; so admission control is
3880 * not on the safe side. It does however guarantee tasks will never
3881 * consume more than promised.
3886 * sched_setparam() passes in -1 for its policy, to let the functions
3887 * it calls know not to change it.
3889 #define SETPARAM_POLICY -1
3891 static void __setscheduler_params(struct task_struct *p,
3892 const struct sched_attr *attr)
3894 int policy = attr->sched_policy;
3896 if (policy == SETPARAM_POLICY)
3901 if (dl_policy(policy))
3902 __setparam_dl(p, attr);
3903 else if (fair_policy(policy))
3904 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3907 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3908 * !rt_policy. Always setting this ensures that things like
3909 * getparam()/getattr() don't report silly values for !rt tasks.
3911 p->rt_priority = attr->sched_priority;
3912 p->normal_prio = normal_prio(p);
3916 /* Actually do priority change: must hold pi & rq lock. */
3917 static void __setscheduler(struct rq *rq, struct task_struct *p,
3918 const struct sched_attr *attr, bool keep_boost)
3920 __setscheduler_params(p, attr);
3923 * Keep a potential priority boosting if called from
3924 * sched_setscheduler().
3927 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3929 p->prio = normal_prio(p);
3931 if (dl_prio(p->prio))
3932 p->sched_class = &dl_sched_class;
3933 else if (rt_prio(p->prio))
3934 p->sched_class = &rt_sched_class;
3936 p->sched_class = &fair_sched_class;
3940 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3942 struct sched_dl_entity *dl_se = &p->dl;
3944 attr->sched_priority = p->rt_priority;
3945 attr->sched_runtime = dl_se->dl_runtime;
3946 attr->sched_deadline = dl_se->dl_deadline;
3947 attr->sched_period = dl_se->dl_period;
3948 attr->sched_flags = dl_se->flags;
3952 * This function validates the new parameters of a -deadline task.
3953 * We ask for the deadline not being zero, and greater or equal
3954 * than the runtime, as well as the period of being zero or
3955 * greater than deadline. Furthermore, we have to be sure that
3956 * user parameters are above the internal resolution of 1us (we
3957 * check sched_runtime only since it is always the smaller one) and
3958 * below 2^63 ns (we have to check both sched_deadline and
3959 * sched_period, as the latter can be zero).
3962 __checkparam_dl(const struct sched_attr *attr)
3965 if (attr->sched_deadline == 0)
3969 * Since we truncate DL_SCALE bits, make sure we're at least
3972 if (attr->sched_runtime < (1ULL << DL_SCALE))
3976 * Since we use the MSB for wrap-around and sign issues, make
3977 * sure it's not set (mind that period can be equal to zero).
3979 if (attr->sched_deadline & (1ULL << 63) ||
3980 attr->sched_period & (1ULL << 63))
3983 /* runtime <= deadline <= period (if period != 0) */
3984 if ((attr->sched_period != 0 &&
3985 attr->sched_period < attr->sched_deadline) ||
3986 attr->sched_deadline < attr->sched_runtime)
3993 * check the target process has a UID that matches the current process's
3995 static bool check_same_owner(struct task_struct *p)
3997 const struct cred *cred = current_cred(), *pcred;
4001 pcred = __task_cred(p);
4002 match = (uid_eq(cred->euid, pcred->euid) ||
4003 uid_eq(cred->euid, pcred->uid));
4008 static bool dl_param_changed(struct task_struct *p,
4009 const struct sched_attr *attr)
4011 struct sched_dl_entity *dl_se = &p->dl;
4013 if (dl_se->dl_runtime != attr->sched_runtime ||
4014 dl_se->dl_deadline != attr->sched_deadline ||
4015 dl_se->dl_period != attr->sched_period ||
4016 dl_se->flags != attr->sched_flags)
4022 static int __sched_setscheduler(struct task_struct *p,
4023 const struct sched_attr *attr,
4026 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4027 MAX_RT_PRIO - 1 - attr->sched_priority;
4028 int retval, oldprio, oldpolicy = -1, queued, running;
4029 int new_effective_prio, policy = attr->sched_policy;
4030 unsigned long flags;
4031 const struct sched_class *prev_class;
4035 /* may grab non-irq protected spin_locks */
4036 BUG_ON(in_interrupt());
4038 /* double check policy once rq lock held */
4040 reset_on_fork = p->sched_reset_on_fork;
4041 policy = oldpolicy = p->policy;
4043 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4045 if (!valid_policy(policy))
4049 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4053 * Valid priorities for SCHED_FIFO and SCHED_RR are
4054 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4055 * SCHED_BATCH and SCHED_IDLE is 0.
4057 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4058 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4060 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4061 (rt_policy(policy) != (attr->sched_priority != 0)))
4065 * Allow unprivileged RT tasks to decrease priority:
4067 if (user && !capable(CAP_SYS_NICE)) {
4068 if (fair_policy(policy)) {
4069 if (attr->sched_nice < task_nice(p) &&
4070 !can_nice(p, attr->sched_nice))
4074 if (rt_policy(policy)) {
4075 unsigned long rlim_rtprio =
4076 task_rlimit(p, RLIMIT_RTPRIO);
4078 /* can't set/change the rt policy */
4079 if (policy != p->policy && !rlim_rtprio)
4082 /* can't increase priority */
4083 if (attr->sched_priority > p->rt_priority &&
4084 attr->sched_priority > rlim_rtprio)
4089 * Can't set/change SCHED_DEADLINE policy at all for now
4090 * (safest behavior); in the future we would like to allow
4091 * unprivileged DL tasks to increase their relative deadline
4092 * or reduce their runtime (both ways reducing utilization)
4094 if (dl_policy(policy))
4098 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4099 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4101 if (idle_policy(p->policy) && !idle_policy(policy)) {
4102 if (!can_nice(p, task_nice(p)))
4106 /* can't change other user's priorities */
4107 if (!check_same_owner(p))
4110 /* Normal users shall not reset the sched_reset_on_fork flag */
4111 if (p->sched_reset_on_fork && !reset_on_fork)
4116 retval = security_task_setscheduler(p);
4122 * make sure no PI-waiters arrive (or leave) while we are
4123 * changing the priority of the task:
4125 * To be able to change p->policy safely, the appropriate
4126 * runqueue lock must be held.
4128 rq = task_rq_lock(p, &flags);
4129 update_rq_clock(rq);
4132 * Changing the policy of the stop threads its a very bad idea
4134 if (p == rq->stop) {
4135 task_rq_unlock(rq, p, &flags);
4140 * If not changing anything there's no need to proceed further,
4141 * but store a possible modification of reset_on_fork.
4143 if (unlikely(policy == p->policy)) {
4144 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4146 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4148 if (dl_policy(policy) && dl_param_changed(p, attr))
4151 p->sched_reset_on_fork = reset_on_fork;
4152 task_rq_unlock(rq, p, &flags);
4158 #ifdef CONFIG_RT_GROUP_SCHED
4160 * Do not allow realtime tasks into groups that have no runtime
4163 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4164 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4165 !task_group_is_autogroup(task_group(p))) {
4166 task_rq_unlock(rq, p, &flags);
4171 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4172 cpumask_t *span = rq->rd->span;
4175 * Don't allow tasks with an affinity mask smaller than
4176 * the entire root_domain to become SCHED_DEADLINE. We
4177 * will also fail if there's no bandwidth available.
4179 if (!cpumask_subset(span, &p->cpus_allowed) ||
4180 rq->rd->dl_bw.bw == 0) {
4181 task_rq_unlock(rq, p, &flags);
4188 /* recheck policy now with rq lock held */
4189 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4190 policy = oldpolicy = -1;
4191 task_rq_unlock(rq, p, &flags);
4196 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4197 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4200 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4201 task_rq_unlock(rq, p, &flags);
4205 p->sched_reset_on_fork = reset_on_fork;
4210 * Take priority boosted tasks into account. If the new
4211 * effective priority is unchanged, we just store the new
4212 * normal parameters and do not touch the scheduler class and
4213 * the runqueue. This will be done when the task deboost
4216 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4217 if (new_effective_prio == oldprio) {
4218 __setscheduler_params(p, attr);
4219 task_rq_unlock(rq, p, &flags);
4224 queued = task_on_rq_queued(p);
4225 running = task_current(rq, p);
4227 dequeue_task(rq, p, DEQUEUE_SAVE);
4229 put_prev_task(rq, p);
4231 prev_class = p->sched_class;
4232 __setscheduler(rq, p, attr, pi);
4235 p->sched_class->set_curr_task(rq);
4237 int enqueue_flags = ENQUEUE_RESTORE;
4239 * We enqueue to tail when the priority of a task is
4240 * increased (user space view).
4242 if (oldprio <= p->prio)
4243 enqueue_flags |= ENQUEUE_HEAD;
4245 enqueue_task(rq, p, enqueue_flags);
4248 check_class_changed(rq, p, prev_class, oldprio);
4249 preempt_disable(); /* avoid rq from going away on us */
4250 task_rq_unlock(rq, p, &flags);
4253 rt_mutex_adjust_pi(p);
4256 * Run balance callbacks after we've adjusted the PI chain.
4258 balance_callback(rq);
4264 static int _sched_setscheduler(struct task_struct *p, int policy,
4265 const struct sched_param *param, bool check)
4267 struct sched_attr attr = {
4268 .sched_policy = policy,
4269 .sched_priority = param->sched_priority,
4270 .sched_nice = PRIO_TO_NICE(p->static_prio),
4273 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4274 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4275 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4276 policy &= ~SCHED_RESET_ON_FORK;
4277 attr.sched_policy = policy;
4280 return __sched_setscheduler(p, &attr, check, true);
4283 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4284 * @p: the task in question.
4285 * @policy: new policy.
4286 * @param: structure containing the new RT priority.
4288 * Return: 0 on success. An error code otherwise.
4290 * NOTE that the task may be already dead.
4292 int sched_setscheduler(struct task_struct *p, int policy,
4293 const struct sched_param *param)
4295 return _sched_setscheduler(p, policy, param, true);
4297 EXPORT_SYMBOL_GPL(sched_setscheduler);
4299 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4301 return __sched_setscheduler(p, attr, true, true);
4303 EXPORT_SYMBOL_GPL(sched_setattr);
4306 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4307 * @p: the task in question.
4308 * @policy: new policy.
4309 * @param: structure containing the new RT priority.
4311 * Just like sched_setscheduler, only don't bother checking if the
4312 * current context has permission. For example, this is needed in
4313 * stop_machine(): we create temporary high priority worker threads,
4314 * but our caller might not have that capability.
4316 * Return: 0 on success. An error code otherwise.
4318 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4319 const struct sched_param *param)
4321 return _sched_setscheduler(p, policy, param, false);
4323 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4326 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4328 struct sched_param lparam;
4329 struct task_struct *p;
4332 if (!param || pid < 0)
4334 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4339 p = find_process_by_pid(pid);
4341 retval = sched_setscheduler(p, policy, &lparam);
4348 * Mimics kernel/events/core.c perf_copy_attr().
4350 static int sched_copy_attr(struct sched_attr __user *uattr,
4351 struct sched_attr *attr)
4356 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4360 * zero the full structure, so that a short copy will be nice.
4362 memset(attr, 0, sizeof(*attr));
4364 ret = get_user(size, &uattr->size);
4368 if (size > PAGE_SIZE) /* silly large */
4371 if (!size) /* abi compat */
4372 size = SCHED_ATTR_SIZE_VER0;
4374 if (size < SCHED_ATTR_SIZE_VER0)
4378 * If we're handed a bigger struct than we know of,
4379 * ensure all the unknown bits are 0 - i.e. new
4380 * user-space does not rely on any kernel feature
4381 * extensions we dont know about yet.
4383 if (size > sizeof(*attr)) {
4384 unsigned char __user *addr;
4385 unsigned char __user *end;
4388 addr = (void __user *)uattr + sizeof(*attr);
4389 end = (void __user *)uattr + size;
4391 for (; addr < end; addr++) {
4392 ret = get_user(val, addr);
4398 size = sizeof(*attr);
4401 ret = copy_from_user(attr, uattr, size);
4406 * XXX: do we want to be lenient like existing syscalls; or do we want
4407 * to be strict and return an error on out-of-bounds values?
4409 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4414 put_user(sizeof(*attr), &uattr->size);
4419 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4420 * @pid: the pid in question.
4421 * @policy: new policy.
4422 * @param: structure containing the new RT priority.
4424 * Return: 0 on success. An error code otherwise.
4426 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4427 struct sched_param __user *, param)
4429 /* negative values for policy are not valid */
4433 return do_sched_setscheduler(pid, policy, param);
4437 * sys_sched_setparam - set/change the RT priority of a thread
4438 * @pid: the pid in question.
4439 * @param: structure containing the new RT priority.
4441 * Return: 0 on success. An error code otherwise.
4443 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4445 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4449 * sys_sched_setattr - same as above, but with extended sched_attr
4450 * @pid: the pid in question.
4451 * @uattr: structure containing the extended parameters.
4452 * @flags: for future extension.
4454 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4455 unsigned int, flags)
4457 struct sched_attr attr;
4458 struct task_struct *p;
4461 if (!uattr || pid < 0 || flags)
4464 retval = sched_copy_attr(uattr, &attr);
4468 if ((int)attr.sched_policy < 0)
4473 p = find_process_by_pid(pid);
4475 retval = sched_setattr(p, &attr);
4482 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4483 * @pid: the pid in question.
4485 * Return: On success, the policy of the thread. Otherwise, a negative error
4488 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4490 struct task_struct *p;
4498 p = find_process_by_pid(pid);
4500 retval = security_task_getscheduler(p);
4503 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4510 * sys_sched_getparam - get the RT priority of a thread
4511 * @pid: the pid in question.
4512 * @param: structure containing the RT priority.
4514 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4517 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4519 struct sched_param lp = { .sched_priority = 0 };
4520 struct task_struct *p;
4523 if (!param || pid < 0)
4527 p = find_process_by_pid(pid);
4532 retval = security_task_getscheduler(p);
4536 if (task_has_rt_policy(p))
4537 lp.sched_priority = p->rt_priority;
4541 * This one might sleep, we cannot do it with a spinlock held ...
4543 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4552 static int sched_read_attr(struct sched_attr __user *uattr,
4553 struct sched_attr *attr,
4558 if (!access_ok(VERIFY_WRITE, uattr, usize))
4562 * If we're handed a smaller struct than we know of,
4563 * ensure all the unknown bits are 0 - i.e. old
4564 * user-space does not get uncomplete information.
4566 if (usize < sizeof(*attr)) {
4567 unsigned char *addr;
4570 addr = (void *)attr + usize;
4571 end = (void *)attr + sizeof(*attr);
4573 for (; addr < end; addr++) {
4581 ret = copy_to_user(uattr, attr, attr->size);
4589 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4590 * @pid: the pid in question.
4591 * @uattr: structure containing the extended parameters.
4592 * @size: sizeof(attr) for fwd/bwd comp.
4593 * @flags: for future extension.
4595 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4596 unsigned int, size, unsigned int, flags)
4598 struct sched_attr attr = {
4599 .size = sizeof(struct sched_attr),
4601 struct task_struct *p;
4604 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4605 size < SCHED_ATTR_SIZE_VER0 || flags)
4609 p = find_process_by_pid(pid);
4614 retval = security_task_getscheduler(p);
4618 attr.sched_policy = p->policy;
4619 if (p->sched_reset_on_fork)
4620 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4621 if (task_has_dl_policy(p))
4622 __getparam_dl(p, &attr);
4623 else if (task_has_rt_policy(p))
4624 attr.sched_priority = p->rt_priority;
4626 attr.sched_nice = task_nice(p);
4630 retval = sched_read_attr(uattr, &attr, size);
4638 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4640 cpumask_var_t cpus_allowed, new_mask;
4641 struct task_struct *p;
4646 p = find_process_by_pid(pid);
4652 /* Prevent p going away */
4656 if (p->flags & PF_NO_SETAFFINITY) {
4660 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4664 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4666 goto out_free_cpus_allowed;
4669 if (!check_same_owner(p)) {
4671 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4673 goto out_free_new_mask;
4678 retval = security_task_setscheduler(p);
4680 goto out_free_new_mask;
4683 cpuset_cpus_allowed(p, cpus_allowed);
4684 cpumask_and(new_mask, in_mask, cpus_allowed);
4687 * Since bandwidth control happens on root_domain basis,
4688 * if admission test is enabled, we only admit -deadline
4689 * tasks allowed to run on all the CPUs in the task's
4693 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4695 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4698 goto out_free_new_mask;
4704 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4707 cpuset_cpus_allowed(p, cpus_allowed);
4708 if (!cpumask_subset(new_mask, cpus_allowed)) {
4710 * We must have raced with a concurrent cpuset
4711 * update. Just reset the cpus_allowed to the
4712 * cpuset's cpus_allowed
4714 cpumask_copy(new_mask, cpus_allowed);
4719 free_cpumask_var(new_mask);
4720 out_free_cpus_allowed:
4721 free_cpumask_var(cpus_allowed);
4727 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4728 struct cpumask *new_mask)
4730 if (len < cpumask_size())
4731 cpumask_clear(new_mask);
4732 else if (len > cpumask_size())
4733 len = cpumask_size();
4735 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4739 * sys_sched_setaffinity - set the cpu affinity of a process
4740 * @pid: pid of the process
4741 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4742 * @user_mask_ptr: user-space pointer to the new cpu mask
4744 * Return: 0 on success. An error code otherwise.
4746 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4747 unsigned long __user *, user_mask_ptr)
4749 cpumask_var_t new_mask;
4752 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4755 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4757 retval = sched_setaffinity(pid, new_mask);
4758 free_cpumask_var(new_mask);
4762 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4764 struct task_struct *p;
4765 unsigned long flags;
4771 p = find_process_by_pid(pid);
4775 retval = security_task_getscheduler(p);
4779 raw_spin_lock_irqsave(&p->pi_lock, flags);
4780 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4781 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4790 * sys_sched_getaffinity - get the cpu affinity of a process
4791 * @pid: pid of the process
4792 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4793 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4795 * Return: 0 on success. An error code otherwise.
4797 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4798 unsigned long __user *, user_mask_ptr)
4803 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4805 if (len & (sizeof(unsigned long)-1))
4808 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4811 ret = sched_getaffinity(pid, mask);
4813 size_t retlen = min_t(size_t, len, cpumask_size());
4815 if (copy_to_user(user_mask_ptr, mask, retlen))
4820 free_cpumask_var(mask);
4826 * sys_sched_yield - yield the current processor to other threads.
4828 * This function yields the current CPU to other tasks. If there are no
4829 * other threads running on this CPU then this function will return.
4833 SYSCALL_DEFINE0(sched_yield)
4835 struct rq *rq = this_rq_lock();
4837 schedstat_inc(rq, yld_count);
4838 current->sched_class->yield_task(rq);
4841 * Since we are going to call schedule() anyway, there's
4842 * no need to preempt or enable interrupts:
4844 __release(rq->lock);
4845 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4846 do_raw_spin_unlock(&rq->lock);
4847 sched_preempt_enable_no_resched();
4854 int __sched _cond_resched(void)
4856 if (should_resched(0)) {
4857 preempt_schedule_common();
4862 EXPORT_SYMBOL(_cond_resched);
4865 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4866 * call schedule, and on return reacquire the lock.
4868 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4869 * operations here to prevent schedule() from being called twice (once via
4870 * spin_unlock(), once by hand).
4872 int __cond_resched_lock(spinlock_t *lock)
4874 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4877 lockdep_assert_held(lock);
4879 if (spin_needbreak(lock) || resched) {
4882 preempt_schedule_common();
4890 EXPORT_SYMBOL(__cond_resched_lock);
4892 int __sched __cond_resched_softirq(void)
4894 BUG_ON(!in_softirq());
4896 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4898 preempt_schedule_common();
4904 EXPORT_SYMBOL(__cond_resched_softirq);
4907 * yield - yield the current processor to other threads.
4909 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4911 * The scheduler is at all times free to pick the calling task as the most
4912 * eligible task to run, if removing the yield() call from your code breaks
4913 * it, its already broken.
4915 * Typical broken usage is:
4920 * where one assumes that yield() will let 'the other' process run that will
4921 * make event true. If the current task is a SCHED_FIFO task that will never
4922 * happen. Never use yield() as a progress guarantee!!
4924 * If you want to use yield() to wait for something, use wait_event().
4925 * If you want to use yield() to be 'nice' for others, use cond_resched().
4926 * If you still want to use yield(), do not!
4928 void __sched yield(void)
4930 set_current_state(TASK_RUNNING);
4933 EXPORT_SYMBOL(yield);
4936 * yield_to - yield the current processor to another thread in
4937 * your thread group, or accelerate that thread toward the
4938 * processor it's on.
4940 * @preempt: whether task preemption is allowed or not
4942 * It's the caller's job to ensure that the target task struct
4943 * can't go away on us before we can do any checks.
4946 * true (>0) if we indeed boosted the target task.
4947 * false (0) if we failed to boost the target.
4948 * -ESRCH if there's no task to yield to.
4950 int __sched yield_to(struct task_struct *p, bool preempt)
4952 struct task_struct *curr = current;
4953 struct rq *rq, *p_rq;
4954 unsigned long flags;
4957 local_irq_save(flags);
4963 * If we're the only runnable task on the rq and target rq also
4964 * has only one task, there's absolutely no point in yielding.
4966 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4971 double_rq_lock(rq, p_rq);
4972 if (task_rq(p) != p_rq) {
4973 double_rq_unlock(rq, p_rq);
4977 if (!curr->sched_class->yield_to_task)
4980 if (curr->sched_class != p->sched_class)
4983 if (task_running(p_rq, p) || p->state)
4986 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4988 schedstat_inc(rq, yld_count);
4990 * Make p's CPU reschedule; pick_next_entity takes care of
4993 if (preempt && rq != p_rq)
4998 double_rq_unlock(rq, p_rq);
5000 local_irq_restore(flags);
5007 EXPORT_SYMBOL_GPL(yield_to);
5010 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5011 * that process accounting knows that this is a task in IO wait state.
5013 long __sched io_schedule_timeout(long timeout)
5015 int old_iowait = current->in_iowait;
5019 current->in_iowait = 1;
5020 blk_schedule_flush_plug(current);
5022 delayacct_blkio_start();
5024 atomic_inc(&rq->nr_iowait);
5025 ret = schedule_timeout(timeout);
5026 current->in_iowait = old_iowait;
5027 atomic_dec(&rq->nr_iowait);
5028 delayacct_blkio_end();
5032 EXPORT_SYMBOL(io_schedule_timeout);
5035 * sys_sched_get_priority_max - return maximum RT priority.
5036 * @policy: scheduling class.
5038 * Return: On success, this syscall returns the maximum
5039 * rt_priority that can be used by a given scheduling class.
5040 * On failure, a negative error code is returned.
5042 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5049 ret = MAX_USER_RT_PRIO-1;
5051 case SCHED_DEADLINE:
5062 * sys_sched_get_priority_min - return minimum RT priority.
5063 * @policy: scheduling class.
5065 * Return: On success, this syscall returns the minimum
5066 * rt_priority that can be used by a given scheduling class.
5067 * On failure, a negative error code is returned.
5069 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5078 case SCHED_DEADLINE:
5088 * sys_sched_rr_get_interval - return the default timeslice of a process.
5089 * @pid: pid of the process.
5090 * @interval: userspace pointer to the timeslice value.
5092 * this syscall writes the default timeslice value of a given process
5093 * into the user-space timespec buffer. A value of '0' means infinity.
5095 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5098 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5099 struct timespec __user *, interval)
5101 struct task_struct *p;
5102 unsigned int time_slice;
5103 unsigned long flags;
5113 p = find_process_by_pid(pid);
5117 retval = security_task_getscheduler(p);
5121 rq = task_rq_lock(p, &flags);
5123 if (p->sched_class->get_rr_interval)
5124 time_slice = p->sched_class->get_rr_interval(rq, p);
5125 task_rq_unlock(rq, p, &flags);
5128 jiffies_to_timespec(time_slice, &t);
5129 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5137 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5139 void sched_show_task(struct task_struct *p)
5141 unsigned long free = 0;
5143 unsigned long state = p->state;
5146 state = __ffs(state) + 1;
5147 printk(KERN_INFO "%-15.15s %c", p->comm,
5148 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5149 #if BITS_PER_LONG == 32
5150 if (state == TASK_RUNNING)
5151 printk(KERN_CONT " running ");
5153 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5155 if (state == TASK_RUNNING)
5156 printk(KERN_CONT " running task ");
5158 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5160 #ifdef CONFIG_DEBUG_STACK_USAGE
5161 free = stack_not_used(p);
5166 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5168 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5169 task_pid_nr(p), ppid,
5170 (unsigned long)task_thread_info(p)->flags);
5172 print_worker_info(KERN_INFO, p);
5173 show_stack(p, NULL);
5176 void show_state_filter(unsigned long state_filter)
5178 struct task_struct *g, *p;
5180 #if BITS_PER_LONG == 32
5182 " task PC stack pid father\n");
5185 " task PC stack pid father\n");
5188 for_each_process_thread(g, p) {
5190 * reset the NMI-timeout, listing all files on a slow
5191 * console might take a lot of time:
5192 * Also, reset softlockup watchdogs on all CPUs, because
5193 * another CPU might be blocked waiting for us to process
5196 touch_nmi_watchdog();
5197 touch_all_softlockup_watchdogs();
5198 if (!state_filter || (p->state & state_filter))
5202 #ifdef CONFIG_SCHED_DEBUG
5203 sysrq_sched_debug_show();
5207 * Only show locks if all tasks are dumped:
5210 debug_show_all_locks();
5213 void init_idle_bootup_task(struct task_struct *idle)
5215 idle->sched_class = &idle_sched_class;
5219 * init_idle - set up an idle thread for a given CPU
5220 * @idle: task in question
5221 * @cpu: cpu the idle task belongs to
5223 * NOTE: this function does not set the idle thread's NEED_RESCHED
5224 * flag, to make booting more robust.
5226 void init_idle(struct task_struct *idle, int cpu)
5228 struct rq *rq = cpu_rq(cpu);
5229 unsigned long flags;
5231 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5232 raw_spin_lock(&rq->lock);
5234 __sched_fork(0, idle);
5236 idle->state = TASK_RUNNING;
5237 idle->se.exec_start = sched_clock();
5241 * Its possible that init_idle() gets called multiple times on a task,
5242 * in that case do_set_cpus_allowed() will not do the right thing.
5244 * And since this is boot we can forgo the serialization.
5246 set_cpus_allowed_common(idle, cpumask_of(cpu));
5249 * We're having a chicken and egg problem, even though we are
5250 * holding rq->lock, the cpu isn't yet set to this cpu so the
5251 * lockdep check in task_group() will fail.
5253 * Similar case to sched_fork(). / Alternatively we could
5254 * use task_rq_lock() here and obtain the other rq->lock.
5259 __set_task_cpu(idle, cpu);
5262 rq->curr = rq->idle = idle;
5263 idle->on_rq = TASK_ON_RQ_QUEUED;
5267 raw_spin_unlock(&rq->lock);
5268 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5270 /* Set the preempt count _outside_ the spinlocks! */
5271 init_idle_preempt_count(idle, cpu);
5274 * The idle tasks have their own, simple scheduling class:
5276 idle->sched_class = &idle_sched_class;
5277 ftrace_graph_init_idle_task(idle, cpu);
5278 vtime_init_idle(idle, cpu);
5280 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5284 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5285 const struct cpumask *trial)
5287 int ret = 1, trial_cpus;
5288 struct dl_bw *cur_dl_b;
5289 unsigned long flags;
5291 if (!cpumask_weight(cur))
5294 rcu_read_lock_sched();
5295 cur_dl_b = dl_bw_of(cpumask_any(cur));
5296 trial_cpus = cpumask_weight(trial);
5298 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5299 if (cur_dl_b->bw != -1 &&
5300 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5302 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5303 rcu_read_unlock_sched();
5308 int task_can_attach(struct task_struct *p,
5309 const struct cpumask *cs_cpus_allowed)
5314 * Kthreads which disallow setaffinity shouldn't be moved
5315 * to a new cpuset; we don't want to change their cpu
5316 * affinity and isolating such threads by their set of
5317 * allowed nodes is unnecessary. Thus, cpusets are not
5318 * applicable for such threads. This prevents checking for
5319 * success of set_cpus_allowed_ptr() on all attached tasks
5320 * before cpus_allowed may be changed.
5322 if (p->flags & PF_NO_SETAFFINITY) {
5328 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5330 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5335 unsigned long flags;
5337 rcu_read_lock_sched();
5338 dl_b = dl_bw_of(dest_cpu);
5339 raw_spin_lock_irqsave(&dl_b->lock, flags);
5340 cpus = dl_bw_cpus(dest_cpu);
5341 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5346 * We reserve space for this task in the destination
5347 * root_domain, as we can't fail after this point.
5348 * We will free resources in the source root_domain
5349 * later on (see set_cpus_allowed_dl()).
5351 __dl_add(dl_b, p->dl.dl_bw);
5353 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5354 rcu_read_unlock_sched();
5364 #ifdef CONFIG_NUMA_BALANCING
5365 /* Migrate current task p to target_cpu */
5366 int migrate_task_to(struct task_struct *p, int target_cpu)
5368 struct migration_arg arg = { p, target_cpu };
5369 int curr_cpu = task_cpu(p);
5371 if (curr_cpu == target_cpu)
5374 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5377 /* TODO: This is not properly updating schedstats */
5379 trace_sched_move_numa(p, curr_cpu, target_cpu);
5380 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5384 * Requeue a task on a given node and accurately track the number of NUMA
5385 * tasks on the runqueues
5387 void sched_setnuma(struct task_struct *p, int nid)
5390 unsigned long flags;
5391 bool queued, running;
5393 rq = task_rq_lock(p, &flags);
5394 queued = task_on_rq_queued(p);
5395 running = task_current(rq, p);
5398 dequeue_task(rq, p, DEQUEUE_SAVE);
5400 put_prev_task(rq, p);
5402 p->numa_preferred_nid = nid;
5405 p->sched_class->set_curr_task(rq);
5407 enqueue_task(rq, p, ENQUEUE_RESTORE);
5408 task_rq_unlock(rq, p, &flags);
5410 #endif /* CONFIG_NUMA_BALANCING */
5412 #ifdef CONFIG_HOTPLUG_CPU
5414 * Ensures that the idle task is using init_mm right before its cpu goes
5417 void idle_task_exit(void)
5419 struct mm_struct *mm = current->active_mm;
5421 BUG_ON(cpu_online(smp_processor_id()));
5423 if (mm != &init_mm) {
5424 switch_mm(mm, &init_mm, current);
5425 finish_arch_post_lock_switch();
5431 * Since this CPU is going 'away' for a while, fold any nr_active delta
5432 * we might have. Assumes we're called after migrate_tasks() so that the
5433 * nr_active count is stable.
5435 * Also see the comment "Global load-average calculations".
5437 static void calc_load_migrate(struct rq *rq)
5439 long delta = calc_load_fold_active(rq);
5441 atomic_long_add(delta, &calc_load_tasks);
5444 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5448 static const struct sched_class fake_sched_class = {
5449 .put_prev_task = put_prev_task_fake,
5452 static struct task_struct fake_task = {
5454 * Avoid pull_{rt,dl}_task()
5456 .prio = MAX_PRIO + 1,
5457 .sched_class = &fake_sched_class,
5461 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5462 * try_to_wake_up()->select_task_rq().
5464 * Called with rq->lock held even though we'er in stop_machine() and
5465 * there's no concurrency possible, we hold the required locks anyway
5466 * because of lock validation efforts.
5468 static void migrate_tasks(struct rq *dead_rq)
5470 struct rq *rq = dead_rq;
5471 struct task_struct *next, *stop = rq->stop;
5475 * Fudge the rq selection such that the below task selection loop
5476 * doesn't get stuck on the currently eligible stop task.
5478 * We're currently inside stop_machine() and the rq is either stuck
5479 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5480 * either way we should never end up calling schedule() until we're
5486 * put_prev_task() and pick_next_task() sched
5487 * class method both need to have an up-to-date
5488 * value of rq->clock[_task]
5490 update_rq_clock(rq);
5494 * There's this thread running, bail when that's the only
5497 if (rq->nr_running == 1)
5501 * pick_next_task assumes pinned rq->lock.
5503 lockdep_pin_lock(&rq->lock);
5504 next = pick_next_task(rq, &fake_task);
5506 next->sched_class->put_prev_task(rq, next);
5509 * Rules for changing task_struct::cpus_allowed are holding
5510 * both pi_lock and rq->lock, such that holding either
5511 * stabilizes the mask.
5513 * Drop rq->lock is not quite as disastrous as it usually is
5514 * because !cpu_active at this point, which means load-balance
5515 * will not interfere. Also, stop-machine.
5517 lockdep_unpin_lock(&rq->lock);
5518 raw_spin_unlock(&rq->lock);
5519 raw_spin_lock(&next->pi_lock);
5520 raw_spin_lock(&rq->lock);
5523 * Since we're inside stop-machine, _nothing_ should have
5524 * changed the task, WARN if weird stuff happened, because in
5525 * that case the above rq->lock drop is a fail too.
5527 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5528 raw_spin_unlock(&next->pi_lock);
5532 /* Find suitable destination for @next, with force if needed. */
5533 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5535 rq = __migrate_task(rq, next, dest_cpu);
5536 if (rq != dead_rq) {
5537 raw_spin_unlock(&rq->lock);
5539 raw_spin_lock(&rq->lock);
5541 raw_spin_unlock(&next->pi_lock);
5546 #endif /* CONFIG_HOTPLUG_CPU */
5548 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5550 static struct ctl_table sd_ctl_dir[] = {
5552 .procname = "sched_domain",
5558 static struct ctl_table sd_ctl_root[] = {
5560 .procname = "kernel",
5562 .child = sd_ctl_dir,
5567 static struct ctl_table *sd_alloc_ctl_entry(int n)
5569 struct ctl_table *entry =
5570 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5575 static void sd_free_ctl_entry(struct ctl_table **tablep)
5577 struct ctl_table *entry;
5580 * In the intermediate directories, both the child directory and
5581 * procname are dynamically allocated and could fail but the mode
5582 * will always be set. In the lowest directory the names are
5583 * static strings and all have proc handlers.
5585 for (entry = *tablep; entry->mode; entry++) {
5587 sd_free_ctl_entry(&entry->child);
5588 if (entry->proc_handler == NULL)
5589 kfree(entry->procname);
5596 static int min_load_idx = 0;
5597 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5600 set_table_entry(struct ctl_table *entry,
5601 const char *procname, void *data, int maxlen,
5602 umode_t mode, proc_handler *proc_handler,
5605 entry->procname = procname;
5607 entry->maxlen = maxlen;
5609 entry->proc_handler = proc_handler;
5612 entry->extra1 = &min_load_idx;
5613 entry->extra2 = &max_load_idx;
5617 static struct ctl_table *
5618 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5620 struct ctl_table *table = sd_alloc_ctl_entry(5);
5625 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5626 sizeof(int), 0644, proc_dointvec_minmax, false);
5627 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5628 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5629 proc_doulongvec_minmax, false);
5630 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5631 sizeof(int), 0644, proc_dointvec_minmax, false);
5632 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5633 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5634 proc_doulongvec_minmax, false);
5639 static struct ctl_table *
5640 sd_alloc_ctl_group_table(struct sched_group *sg)
5642 struct ctl_table *table = sd_alloc_ctl_entry(2);
5647 table->procname = kstrdup("energy", GFP_KERNEL);
5649 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5654 static struct ctl_table *
5655 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5657 struct ctl_table *table;
5658 unsigned int nr_entries = 14;
5661 struct sched_group *sg = sd->groups;
5666 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5668 nr_entries += nr_sgs;
5671 table = sd_alloc_ctl_entry(nr_entries);
5676 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5677 sizeof(long), 0644, proc_doulongvec_minmax, false);
5678 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5679 sizeof(long), 0644, proc_doulongvec_minmax, false);
5680 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5681 sizeof(int), 0644, proc_dointvec_minmax, true);
5682 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5683 sizeof(int), 0644, proc_dointvec_minmax, true);
5684 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5685 sizeof(int), 0644, proc_dointvec_minmax, true);
5686 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5687 sizeof(int), 0644, proc_dointvec_minmax, true);
5688 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5689 sizeof(int), 0644, proc_dointvec_minmax, true);
5690 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5691 sizeof(int), 0644, proc_dointvec_minmax, false);
5692 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5693 sizeof(int), 0644, proc_dointvec_minmax, false);
5694 set_table_entry(&table[9], "cache_nice_tries",
5695 &sd->cache_nice_tries,
5696 sizeof(int), 0644, proc_dointvec_minmax, false);
5697 set_table_entry(&table[10], "flags", &sd->flags,
5698 sizeof(int), 0644, proc_dointvec_minmax, false);
5699 set_table_entry(&table[11], "max_newidle_lb_cost",
5700 &sd->max_newidle_lb_cost,
5701 sizeof(long), 0644, proc_doulongvec_minmax, false);
5702 set_table_entry(&table[12], "name", sd->name,
5703 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5707 struct ctl_table *entry = &table[13];
5710 snprintf(buf, 32, "group%d", i);
5711 entry->procname = kstrdup(buf, GFP_KERNEL);
5713 entry->child = sd_alloc_ctl_group_table(sg);
5714 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5716 /* &table[nr_entries-1] is terminator */
5721 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5723 struct ctl_table *entry, *table;
5724 struct sched_domain *sd;
5725 int domain_num = 0, i;
5728 for_each_domain(cpu, sd)
5730 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5735 for_each_domain(cpu, sd) {
5736 snprintf(buf, 32, "domain%d", i);
5737 entry->procname = kstrdup(buf, GFP_KERNEL);
5739 entry->child = sd_alloc_ctl_domain_table(sd);
5746 static struct ctl_table_header *sd_sysctl_header;
5747 static void register_sched_domain_sysctl(void)
5749 int i, cpu_num = num_possible_cpus();
5750 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5753 WARN_ON(sd_ctl_dir[0].child);
5754 sd_ctl_dir[0].child = entry;
5759 for_each_possible_cpu(i) {
5760 snprintf(buf, 32, "cpu%d", i);
5761 entry->procname = kstrdup(buf, GFP_KERNEL);
5763 entry->child = sd_alloc_ctl_cpu_table(i);
5767 WARN_ON(sd_sysctl_header);
5768 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5771 /* may be called multiple times per register */
5772 static void unregister_sched_domain_sysctl(void)
5774 unregister_sysctl_table(sd_sysctl_header);
5775 sd_sysctl_header = NULL;
5776 if (sd_ctl_dir[0].child)
5777 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5780 static void register_sched_domain_sysctl(void)
5783 static void unregister_sched_domain_sysctl(void)
5786 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5788 static void set_rq_online(struct rq *rq)
5791 const struct sched_class *class;
5793 cpumask_set_cpu(rq->cpu, rq->rd->online);
5796 for_each_class(class) {
5797 if (class->rq_online)
5798 class->rq_online(rq);
5803 static void set_rq_offline(struct rq *rq)
5806 const struct sched_class *class;
5808 for_each_class(class) {
5809 if (class->rq_offline)
5810 class->rq_offline(rq);
5813 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5819 * migration_call - callback that gets triggered when a CPU is added.
5820 * Here we can start up the necessary migration thread for the new CPU.
5823 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5825 int cpu = (long)hcpu;
5826 unsigned long flags;
5827 struct rq *rq = cpu_rq(cpu);
5829 switch (action & ~CPU_TASKS_FROZEN) {
5831 case CPU_UP_PREPARE:
5832 raw_spin_lock_irqsave(&rq->lock, flags);
5833 walt_set_window_start(rq);
5834 raw_spin_unlock_irqrestore(&rq->lock, flags);
5835 rq->calc_load_update = calc_load_update;
5839 /* Update our root-domain */
5840 raw_spin_lock_irqsave(&rq->lock, flags);
5842 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5846 raw_spin_unlock_irqrestore(&rq->lock, flags);
5849 #ifdef CONFIG_HOTPLUG_CPU
5851 sched_ttwu_pending();
5852 /* Update our root-domain */
5853 raw_spin_lock_irqsave(&rq->lock, flags);
5854 walt_migrate_sync_cpu(cpu);
5856 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5860 BUG_ON(rq->nr_running != 1); /* the migration thread */
5861 raw_spin_unlock_irqrestore(&rq->lock, flags);
5865 calc_load_migrate(rq);
5870 update_max_interval();
5876 * Register at high priority so that task migration (migrate_all_tasks)
5877 * happens before everything else. This has to be lower priority than
5878 * the notifier in the perf_event subsystem, though.
5880 static struct notifier_block migration_notifier = {
5881 .notifier_call = migration_call,
5882 .priority = CPU_PRI_MIGRATION,
5885 static void set_cpu_rq_start_time(void)
5887 int cpu = smp_processor_id();
5888 struct rq *rq = cpu_rq(cpu);
5889 rq->age_stamp = sched_clock_cpu(cpu);
5892 static int sched_cpu_active(struct notifier_block *nfb,
5893 unsigned long action, void *hcpu)
5895 int cpu = (long)hcpu;
5897 switch (action & ~CPU_TASKS_FROZEN) {
5899 set_cpu_rq_start_time();
5904 * At this point a starting CPU has marked itself as online via
5905 * set_cpu_online(). But it might not yet have marked itself
5906 * as active, which is essential from here on.
5908 set_cpu_active(cpu, true);
5909 stop_machine_unpark(cpu);
5912 case CPU_DOWN_FAILED:
5913 set_cpu_active(cpu, true);
5921 static int sched_cpu_inactive(struct notifier_block *nfb,
5922 unsigned long action, void *hcpu)
5924 switch (action & ~CPU_TASKS_FROZEN) {
5925 case CPU_DOWN_PREPARE:
5926 set_cpu_active((long)hcpu, false);
5933 static int __init migration_init(void)
5935 void *cpu = (void *)(long)smp_processor_id();
5938 /* Initialize migration for the boot CPU */
5939 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5940 BUG_ON(err == NOTIFY_BAD);
5941 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5942 register_cpu_notifier(&migration_notifier);
5944 /* Register cpu active notifiers */
5945 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5946 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5950 early_initcall(migration_init);
5952 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5954 #ifdef CONFIG_SCHED_DEBUG
5956 static __read_mostly int sched_debug_enabled;
5958 static int __init sched_debug_setup(char *str)
5960 sched_debug_enabled = 1;
5964 early_param("sched_debug", sched_debug_setup);
5966 static inline bool sched_debug(void)
5968 return sched_debug_enabled;
5971 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5972 struct cpumask *groupmask)
5974 struct sched_group *group = sd->groups;
5976 cpumask_clear(groupmask);
5978 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5980 if (!(sd->flags & SD_LOAD_BALANCE)) {
5981 printk("does not load-balance\n");
5985 printk(KERN_CONT "span %*pbl level %s\n",
5986 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5988 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5989 printk(KERN_ERR "ERROR: domain->span does not contain "
5992 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5993 printk(KERN_ERR "ERROR: domain->groups does not contain"
5997 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6001 printk(KERN_ERR "ERROR: group is NULL\n");
6005 if (!cpumask_weight(sched_group_cpus(group))) {
6006 printk(KERN_CONT "\n");
6007 printk(KERN_ERR "ERROR: empty group\n");
6011 if (!(sd->flags & SD_OVERLAP) &&
6012 cpumask_intersects(groupmask, sched_group_cpus(group))) {
6013 printk(KERN_CONT "\n");
6014 printk(KERN_ERR "ERROR: repeated CPUs\n");
6018 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6020 printk(KERN_CONT " %*pbl",
6021 cpumask_pr_args(sched_group_cpus(group)));
6022 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
6023 printk(KERN_CONT " (cpu_capacity = %lu)",
6024 group->sgc->capacity);
6027 group = group->next;
6028 } while (group != sd->groups);
6029 printk(KERN_CONT "\n");
6031 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6032 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6035 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6036 printk(KERN_ERR "ERROR: parent span is not a superset "
6037 "of domain->span\n");
6041 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6045 if (!sched_debug_enabled)
6049 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6053 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6056 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6064 #else /* !CONFIG_SCHED_DEBUG */
6065 # define sched_domain_debug(sd, cpu) do { } while (0)
6066 static inline bool sched_debug(void)
6070 #endif /* CONFIG_SCHED_DEBUG */
6072 static int sd_degenerate(struct sched_domain *sd)
6074 if (cpumask_weight(sched_domain_span(sd)) == 1) {
6075 if (sd->groups->sge)
6076 sd->flags &= ~SD_LOAD_BALANCE;
6081 /* Following flags need at least 2 groups */
6082 if (sd->flags & (SD_LOAD_BALANCE |
6083 SD_BALANCE_NEWIDLE |
6086 SD_SHARE_CPUCAPACITY |
6087 SD_ASYM_CPUCAPACITY |
6088 SD_SHARE_PKG_RESOURCES |
6089 SD_SHARE_POWERDOMAIN |
6090 SD_SHARE_CAP_STATES)) {
6091 if (sd->groups != sd->groups->next)
6095 /* Following flags don't use groups */
6096 if (sd->flags & (SD_WAKE_AFFINE))
6103 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6105 unsigned long cflags = sd->flags, pflags = parent->flags;
6107 if (sd_degenerate(parent))
6110 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6113 /* Flags needing groups don't count if only 1 group in parent */
6114 if (parent->groups == parent->groups->next) {
6115 pflags &= ~(SD_LOAD_BALANCE |
6116 SD_BALANCE_NEWIDLE |
6119 SD_ASYM_CPUCAPACITY |
6120 SD_SHARE_CPUCAPACITY |
6121 SD_SHARE_PKG_RESOURCES |
6123 SD_SHARE_POWERDOMAIN |
6124 SD_SHARE_CAP_STATES);
6125 if (parent->groups->sge) {
6126 parent->flags &= ~SD_LOAD_BALANCE;
6129 if (nr_node_ids == 1)
6130 pflags &= ~SD_SERIALIZE;
6132 if (~cflags & pflags)
6138 static void free_rootdomain(struct rcu_head *rcu)
6140 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6142 cpupri_cleanup(&rd->cpupri);
6143 cpudl_cleanup(&rd->cpudl);
6144 free_cpumask_var(rd->dlo_mask);
6145 free_cpumask_var(rd->rto_mask);
6146 free_cpumask_var(rd->online);
6147 free_cpumask_var(rd->span);
6151 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6153 struct root_domain *old_rd = NULL;
6154 unsigned long flags;
6156 raw_spin_lock_irqsave(&rq->lock, flags);
6161 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6164 cpumask_clear_cpu(rq->cpu, old_rd->span);
6167 * If we dont want to free the old_rd yet then
6168 * set old_rd to NULL to skip the freeing later
6171 if (!atomic_dec_and_test(&old_rd->refcount))
6175 atomic_inc(&rd->refcount);
6178 cpumask_set_cpu(rq->cpu, rd->span);
6179 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6182 raw_spin_unlock_irqrestore(&rq->lock, flags);
6185 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6188 static int init_rootdomain(struct root_domain *rd)
6190 memset(rd, 0, sizeof(*rd));
6192 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6194 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6196 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6198 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6201 init_dl_bw(&rd->dl_bw);
6202 if (cpudl_init(&rd->cpudl) != 0)
6205 if (cpupri_init(&rd->cpupri) != 0)
6208 init_max_cpu_capacity(&rd->max_cpu_capacity);
6210 rd->max_cap_orig_cpu = rd->min_cap_orig_cpu = -1;
6215 free_cpumask_var(rd->rto_mask);
6217 free_cpumask_var(rd->dlo_mask);
6219 free_cpumask_var(rd->online);
6221 free_cpumask_var(rd->span);
6227 * By default the system creates a single root-domain with all cpus as
6228 * members (mimicking the global state we have today).
6230 struct root_domain def_root_domain;
6232 static void init_defrootdomain(void)
6234 init_rootdomain(&def_root_domain);
6236 atomic_set(&def_root_domain.refcount, 1);
6239 static struct root_domain *alloc_rootdomain(void)
6241 struct root_domain *rd;
6243 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6247 if (init_rootdomain(rd) != 0) {
6255 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6257 struct sched_group *tmp, *first;
6266 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6271 } while (sg != first);
6274 static void free_sched_domain(struct rcu_head *rcu)
6276 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6279 * If its an overlapping domain it has private groups, iterate and
6282 if (sd->flags & SD_OVERLAP) {
6283 free_sched_groups(sd->groups, 1);
6284 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6285 kfree(sd->groups->sgc);
6291 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6293 call_rcu(&sd->rcu, free_sched_domain);
6296 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6298 for (; sd; sd = sd->parent)
6299 destroy_sched_domain(sd, cpu);
6303 * Keep a special pointer to the highest sched_domain that has
6304 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6305 * allows us to avoid some pointer chasing select_idle_sibling().
6307 * Also keep a unique ID per domain (we use the first cpu number in
6308 * the cpumask of the domain), this allows us to quickly tell if
6309 * two cpus are in the same cache domain, see cpus_share_cache().
6311 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6312 DEFINE_PER_CPU(int, sd_llc_size);
6313 DEFINE_PER_CPU(int, sd_llc_id);
6314 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6315 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6316 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6317 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6318 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6320 static void update_top_cache_domain(int cpu)
6322 struct sched_domain *sd;
6323 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6327 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6329 id = cpumask_first(sched_domain_span(sd));
6330 size = cpumask_weight(sched_domain_span(sd));
6331 busy_sd = sd->parent; /* sd_busy */
6333 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6335 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6336 per_cpu(sd_llc_size, cpu) = size;
6337 per_cpu(sd_llc_id, cpu) = id;
6339 sd = lowest_flag_domain(cpu, SD_NUMA);
6340 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6342 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6343 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6345 for_each_domain(cpu, sd) {
6346 if (sd->groups->sge)
6351 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6353 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6354 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6358 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6359 * hold the hotplug lock.
6362 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6364 struct rq *rq = cpu_rq(cpu);
6365 struct sched_domain *tmp;
6367 /* Remove the sched domains which do not contribute to scheduling. */
6368 for (tmp = sd; tmp; ) {
6369 struct sched_domain *parent = tmp->parent;
6373 if (sd_parent_degenerate(tmp, parent)) {
6374 tmp->parent = parent->parent;
6376 parent->parent->child = tmp;
6378 * Transfer SD_PREFER_SIBLING down in case of a
6379 * degenerate parent; the spans match for this
6380 * so the property transfers.
6382 if (parent->flags & SD_PREFER_SIBLING)
6383 tmp->flags |= SD_PREFER_SIBLING;
6384 destroy_sched_domain(parent, cpu);
6389 if (sd && sd_degenerate(sd)) {
6392 destroy_sched_domain(tmp, cpu);
6397 sched_domain_debug(sd, cpu);
6399 rq_attach_root(rq, rd);
6401 rcu_assign_pointer(rq->sd, sd);
6402 destroy_sched_domains(tmp, cpu);
6404 update_top_cache_domain(cpu);
6407 /* Setup the mask of cpus configured for isolated domains */
6408 static int __init isolated_cpu_setup(char *str)
6410 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6411 cpulist_parse(str, cpu_isolated_map);
6415 __setup("isolcpus=", isolated_cpu_setup);
6418 struct sched_domain ** __percpu sd;
6419 struct root_domain *rd;
6430 * Build an iteration mask that can exclude certain CPUs from the upwards
6433 * Only CPUs that can arrive at this group should be considered to continue
6436 * Asymmetric node setups can result in situations where the domain tree is of
6437 * unequal depth, make sure to skip domains that already cover the entire
6440 * In that case build_sched_domains() will have terminated the iteration early
6441 * and our sibling sd spans will be empty. Domains should always include the
6442 * cpu they're built on, so check that.
6445 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6447 const struct cpumask *sg_span = sched_group_cpus(sg);
6448 struct sd_data *sdd = sd->private;
6449 struct sched_domain *sibling;
6452 for_each_cpu(i, sg_span) {
6453 sibling = *per_cpu_ptr(sdd->sd, i);
6456 * Can happen in the asymmetric case, where these siblings are
6457 * unused. The mask will not be empty because those CPUs that
6458 * do have the top domain _should_ span the domain.
6460 if (!sibling->child)
6463 /* If we would not end up here, we can't continue from here */
6464 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
6467 cpumask_set_cpu(i, sched_group_mask(sg));
6470 /* We must not have empty masks here */
6471 WARN_ON_ONCE(cpumask_empty(sched_group_mask(sg)));
6475 * Return the canonical balance cpu for this group, this is the first cpu
6476 * of this group that's also in the iteration mask.
6478 int group_balance_cpu(struct sched_group *sg)
6480 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6484 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6486 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6487 const struct cpumask *span = sched_domain_span(sd);
6488 struct cpumask *covered = sched_domains_tmpmask;
6489 struct sd_data *sdd = sd->private;
6490 struct sched_domain *sibling;
6493 cpumask_clear(covered);
6495 for_each_cpu(i, span) {
6496 struct cpumask *sg_span;
6498 if (cpumask_test_cpu(i, covered))
6501 sibling = *per_cpu_ptr(sdd->sd, i);
6503 /* See the comment near build_group_mask(). */
6504 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6507 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6508 GFP_KERNEL, cpu_to_node(cpu));
6513 sg_span = sched_group_cpus(sg);
6515 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6517 cpumask_set_cpu(i, sg_span);
6519 cpumask_or(covered, covered, sg_span);
6521 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6522 if (atomic_inc_return(&sg->sgc->ref) == 1)
6523 build_group_mask(sd, sg);
6526 * Initialize sgc->capacity such that even if we mess up the
6527 * domains and no possible iteration will get us here, we won't
6530 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6531 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6532 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
6535 * Make sure the first group of this domain contains the
6536 * canonical balance cpu. Otherwise the sched_domain iteration
6537 * breaks. See update_sg_lb_stats().
6539 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6540 group_balance_cpu(sg) == cpu)
6550 sd->groups = groups;
6555 free_sched_groups(first, 0);
6560 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6562 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6563 struct sched_domain *child = sd->child;
6566 cpu = cpumask_first(sched_domain_span(child));
6569 *sg = *per_cpu_ptr(sdd->sg, cpu);
6570 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6571 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6578 * build_sched_groups will build a circular linked list of the groups
6579 * covered by the given span, and will set each group's ->cpumask correctly,
6580 * and ->cpu_capacity to 0.
6582 * Assumes the sched_domain tree is fully constructed
6585 build_sched_groups(struct sched_domain *sd, int cpu)
6587 struct sched_group *first = NULL, *last = NULL;
6588 struct sd_data *sdd = sd->private;
6589 const struct cpumask *span = sched_domain_span(sd);
6590 struct cpumask *covered;
6593 get_group(cpu, sdd, &sd->groups);
6594 atomic_inc(&sd->groups->ref);
6596 if (cpu != cpumask_first(span))
6599 lockdep_assert_held(&sched_domains_mutex);
6600 covered = sched_domains_tmpmask;
6602 cpumask_clear(covered);
6604 for_each_cpu(i, span) {
6605 struct sched_group *sg;
6608 if (cpumask_test_cpu(i, covered))
6611 group = get_group(i, sdd, &sg);
6612 cpumask_setall(sched_group_mask(sg));
6614 for_each_cpu(j, span) {
6615 if (get_group(j, sdd, NULL) != group)
6618 cpumask_set_cpu(j, covered);
6619 cpumask_set_cpu(j, sched_group_cpus(sg));
6634 * Initialize sched groups cpu_capacity.
6636 * cpu_capacity indicates the capacity of sched group, which is used while
6637 * distributing the load between different sched groups in a sched domain.
6638 * Typically cpu_capacity for all the groups in a sched domain will be same
6639 * unless there are asymmetries in the topology. If there are asymmetries,
6640 * group having more cpu_capacity will pickup more load compared to the
6641 * group having less cpu_capacity.
6643 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6645 struct sched_group *sg = sd->groups;
6650 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6652 } while (sg != sd->groups);
6654 if (cpu != group_balance_cpu(sg))
6657 update_group_capacity(sd, cpu);
6658 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6661 static bool have_sched_energy_data(void)
6665 for_each_possible_cpu(cpu) {
6666 if (!rcu_dereference(per_cpu(sd_scs, cpu)) ||
6667 !rcu_dereference(per_cpu(sd_ea, cpu)))
6675 * Check that the per-cpu provided sd energy data is consistent for all cpus
6678 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6679 const struct cpumask *cpumask)
6681 const struct sched_group_energy * const sge = fn(cpu);
6682 struct cpumask mask;
6685 if (cpumask_weight(cpumask) <= 1)
6688 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6690 for_each_cpu(i, &mask) {
6691 const struct sched_group_energy * const e = fn(i);
6694 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6696 for (y = 0; y < (e->nr_idle_states); y++) {
6697 BUG_ON(e->idle_states[y].power !=
6698 sge->idle_states[y].power);
6701 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6703 for (y = 0; y < (e->nr_cap_states); y++) {
6704 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6705 BUG_ON(e->cap_states[y].power !=
6706 sge->cap_states[y].power);
6711 static void init_sched_energy(int cpu, struct sched_domain *sd,
6712 sched_domain_energy_f fn)
6714 if (!(fn && fn(cpu)))
6717 if (cpu != group_balance_cpu(sd->groups))
6720 if (sd->child && !sd->child->groups->sge) {
6721 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6722 #ifdef CONFIG_SCHED_DEBUG
6723 pr_err(" energy data on %s but not on %s domain\n",
6724 sd->name, sd->child->name);
6729 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6731 sd->groups->sge = fn(cpu);
6735 * Initializers for schedule domains
6736 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6739 static int default_relax_domain_level = -1;
6740 int sched_domain_level_max;
6742 static int __init setup_relax_domain_level(char *str)
6744 if (kstrtoint(str, 0, &default_relax_domain_level))
6745 pr_warn("Unable to set relax_domain_level\n");
6749 __setup("relax_domain_level=", setup_relax_domain_level);
6751 static void set_domain_attribute(struct sched_domain *sd,
6752 struct sched_domain_attr *attr)
6756 if (!attr || attr->relax_domain_level < 0) {
6757 if (default_relax_domain_level < 0)
6760 request = default_relax_domain_level;
6762 request = attr->relax_domain_level;
6763 if (request < sd->level) {
6764 /* turn off idle balance on this domain */
6765 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6767 /* turn on idle balance on this domain */
6768 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6772 static void __sdt_free(const struct cpumask *cpu_map);
6773 static int __sdt_alloc(const struct cpumask *cpu_map);
6775 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6776 const struct cpumask *cpu_map)
6780 if (!atomic_read(&d->rd->refcount))
6781 free_rootdomain(&d->rd->rcu); /* fall through */
6783 free_percpu(d->sd); /* fall through */
6785 __sdt_free(cpu_map); /* fall through */
6791 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6792 const struct cpumask *cpu_map)
6794 memset(d, 0, sizeof(*d));
6796 if (__sdt_alloc(cpu_map))
6797 return sa_sd_storage;
6798 d->sd = alloc_percpu(struct sched_domain *);
6800 return sa_sd_storage;
6801 d->rd = alloc_rootdomain();
6804 return sa_rootdomain;
6808 * NULL the sd_data elements we've used to build the sched_domain and
6809 * sched_group structure so that the subsequent __free_domain_allocs()
6810 * will not free the data we're using.
6812 static void claim_allocations(int cpu, struct sched_domain *sd)
6814 struct sd_data *sdd = sd->private;
6816 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6817 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6819 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6820 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6822 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6823 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6827 static int sched_domains_numa_levels;
6828 enum numa_topology_type sched_numa_topology_type;
6829 static int *sched_domains_numa_distance;
6830 int sched_max_numa_distance;
6831 static struct cpumask ***sched_domains_numa_masks;
6832 static int sched_domains_curr_level;
6836 * SD_flags allowed in topology descriptions.
6838 * These flags are purely descriptive of the topology and do not prescribe
6839 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6842 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6843 * SD_SHARE_PKG_RESOURCES - describes shared caches
6844 * SD_NUMA - describes NUMA topologies
6845 * SD_SHARE_POWERDOMAIN - describes shared power domain
6846 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6847 * SD_SHARE_CAP_STATES - describes shared capacity states
6849 * Odd one out, which beside describing the topology has a quirk also
6850 * prescribes the desired behaviour that goes along with it:
6853 * SD_ASYM_PACKING - describes SMT quirks
6855 #define TOPOLOGY_SD_FLAGS \
6856 (SD_SHARE_CPUCAPACITY | \
6857 SD_SHARE_PKG_RESOURCES | \
6860 SD_ASYM_CPUCAPACITY | \
6861 SD_SHARE_POWERDOMAIN | \
6862 SD_SHARE_CAP_STATES)
6864 static struct sched_domain *
6865 sd_init(struct sched_domain_topology_level *tl,
6866 struct sched_domain *child, int cpu)
6868 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6869 int sd_weight, sd_flags = 0;
6873 * Ugly hack to pass state to sd_numa_mask()...
6875 sched_domains_curr_level = tl->numa_level;
6878 sd_weight = cpumask_weight(tl->mask(cpu));
6881 sd_flags = (*tl->sd_flags)();
6882 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6883 "wrong sd_flags in topology description\n"))
6884 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6886 *sd = (struct sched_domain){
6887 .min_interval = sd_weight,
6888 .max_interval = 2*sd_weight,
6890 .imbalance_pct = 125,
6892 .cache_nice_tries = 0,
6899 .flags = 1*SD_LOAD_BALANCE
6900 | 1*SD_BALANCE_NEWIDLE
6905 | 0*SD_SHARE_CPUCAPACITY
6906 | 0*SD_SHARE_PKG_RESOURCES
6908 | 0*SD_PREFER_SIBLING
6913 .last_balance = jiffies,
6914 .balance_interval = sd_weight,
6916 .max_newidle_lb_cost = 0,
6917 .next_decay_max_lb_cost = jiffies,
6919 #ifdef CONFIG_SCHED_DEBUG
6925 * Convert topological properties into behaviour.
6928 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6929 struct sched_domain *t = sd;
6931 for_each_lower_domain(t)
6932 t->flags |= SD_BALANCE_WAKE;
6935 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6936 sd->flags |= SD_PREFER_SIBLING;
6937 sd->imbalance_pct = 110;
6938 sd->smt_gain = 1178; /* ~15% */
6940 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6941 sd->imbalance_pct = 117;
6942 sd->cache_nice_tries = 1;
6946 } else if (sd->flags & SD_NUMA) {
6947 sd->cache_nice_tries = 2;
6951 sd->flags |= SD_SERIALIZE;
6952 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6953 sd->flags &= ~(SD_BALANCE_EXEC |
6960 sd->flags |= SD_PREFER_SIBLING;
6961 sd->cache_nice_tries = 1;
6966 sd->private = &tl->data;
6972 * Topology list, bottom-up.
6974 static struct sched_domain_topology_level default_topology[] = {
6975 #ifdef CONFIG_SCHED_SMT
6976 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6978 #ifdef CONFIG_SCHED_MC
6979 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6981 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6985 static struct sched_domain_topology_level *sched_domain_topology =
6988 #define for_each_sd_topology(tl) \
6989 for (tl = sched_domain_topology; tl->mask; tl++)
6991 void set_sched_topology(struct sched_domain_topology_level *tl)
6993 sched_domain_topology = tl;
6998 static const struct cpumask *sd_numa_mask(int cpu)
7000 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
7003 static void sched_numa_warn(const char *str)
7005 static int done = false;
7013 printk(KERN_WARNING "ERROR: %s\n\n", str);
7015 for (i = 0; i < nr_node_ids; i++) {
7016 printk(KERN_WARNING " ");
7017 for (j = 0; j < nr_node_ids; j++)
7018 printk(KERN_CONT "%02d ", node_distance(i,j));
7019 printk(KERN_CONT "\n");
7021 printk(KERN_WARNING "\n");
7024 bool find_numa_distance(int distance)
7028 if (distance == node_distance(0, 0))
7031 for (i = 0; i < sched_domains_numa_levels; i++) {
7032 if (sched_domains_numa_distance[i] == distance)
7040 * A system can have three types of NUMA topology:
7041 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
7042 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
7043 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
7045 * The difference between a glueless mesh topology and a backplane
7046 * topology lies in whether communication between not directly
7047 * connected nodes goes through intermediary nodes (where programs
7048 * could run), or through backplane controllers. This affects
7049 * placement of programs.
7051 * The type of topology can be discerned with the following tests:
7052 * - If the maximum distance between any nodes is 1 hop, the system
7053 * is directly connected.
7054 * - If for two nodes A and B, located N > 1 hops away from each other,
7055 * there is an intermediary node C, which is < N hops away from both
7056 * nodes A and B, the system is a glueless mesh.
7058 static void init_numa_topology_type(void)
7062 n = sched_max_numa_distance;
7064 if (sched_domains_numa_levels <= 1) {
7065 sched_numa_topology_type = NUMA_DIRECT;
7069 for_each_online_node(a) {
7070 for_each_online_node(b) {
7071 /* Find two nodes furthest removed from each other. */
7072 if (node_distance(a, b) < n)
7075 /* Is there an intermediary node between a and b? */
7076 for_each_online_node(c) {
7077 if (node_distance(a, c) < n &&
7078 node_distance(b, c) < n) {
7079 sched_numa_topology_type =
7085 sched_numa_topology_type = NUMA_BACKPLANE;
7091 static void sched_init_numa(void)
7093 int next_distance, curr_distance = node_distance(0, 0);
7094 struct sched_domain_topology_level *tl;
7098 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
7099 if (!sched_domains_numa_distance)
7103 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
7104 * unique distances in the node_distance() table.
7106 * Assumes node_distance(0,j) includes all distances in
7107 * node_distance(i,j) in order to avoid cubic time.
7109 next_distance = curr_distance;
7110 for (i = 0; i < nr_node_ids; i++) {
7111 for (j = 0; j < nr_node_ids; j++) {
7112 for (k = 0; k < nr_node_ids; k++) {
7113 int distance = node_distance(i, k);
7115 if (distance > curr_distance &&
7116 (distance < next_distance ||
7117 next_distance == curr_distance))
7118 next_distance = distance;
7121 * While not a strong assumption it would be nice to know
7122 * about cases where if node A is connected to B, B is not
7123 * equally connected to A.
7125 if (sched_debug() && node_distance(k, i) != distance)
7126 sched_numa_warn("Node-distance not symmetric");
7128 if (sched_debug() && i && !find_numa_distance(distance))
7129 sched_numa_warn("Node-0 not representative");
7131 if (next_distance != curr_distance) {
7132 sched_domains_numa_distance[level++] = next_distance;
7133 sched_domains_numa_levels = level;
7134 curr_distance = next_distance;
7139 * In case of sched_debug() we verify the above assumption.
7149 * 'level' contains the number of unique distances, excluding the
7150 * identity distance node_distance(i,i).
7152 * The sched_domains_numa_distance[] array includes the actual distance
7157 * Here, we should temporarily reset sched_domains_numa_levels to 0.
7158 * If it fails to allocate memory for array sched_domains_numa_masks[][],
7159 * the array will contain less then 'level' members. This could be
7160 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
7161 * in other functions.
7163 * We reset it to 'level' at the end of this function.
7165 sched_domains_numa_levels = 0;
7167 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
7168 if (!sched_domains_numa_masks)
7172 * Now for each level, construct a mask per node which contains all
7173 * cpus of nodes that are that many hops away from us.
7175 for (i = 0; i < level; i++) {
7176 sched_domains_numa_masks[i] =
7177 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
7178 if (!sched_domains_numa_masks[i])
7181 for (j = 0; j < nr_node_ids; j++) {
7182 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
7186 sched_domains_numa_masks[i][j] = mask;
7189 if (node_distance(j, k) > sched_domains_numa_distance[i])
7192 cpumask_or(mask, mask, cpumask_of_node(k));
7197 /* Compute default topology size */
7198 for (i = 0; sched_domain_topology[i].mask; i++);
7200 tl = kzalloc((i + level + 1) *
7201 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7206 * Copy the default topology bits..
7208 for (i = 0; sched_domain_topology[i].mask; i++)
7209 tl[i] = sched_domain_topology[i];
7212 * .. and append 'j' levels of NUMA goodness.
7214 for (j = 0; j < level; i++, j++) {
7215 tl[i] = (struct sched_domain_topology_level){
7216 .mask = sd_numa_mask,
7217 .sd_flags = cpu_numa_flags,
7218 .flags = SDTL_OVERLAP,
7224 sched_domain_topology = tl;
7226 sched_domains_numa_levels = level;
7227 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7229 init_numa_topology_type();
7232 static void sched_domains_numa_masks_set(int cpu)
7235 int node = cpu_to_node(cpu);
7237 for (i = 0; i < sched_domains_numa_levels; i++) {
7238 for (j = 0; j < nr_node_ids; j++) {
7239 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7240 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7245 static void sched_domains_numa_masks_clear(int cpu)
7248 for (i = 0; i < sched_domains_numa_levels; i++) {
7249 for (j = 0; j < nr_node_ids; j++)
7250 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7255 * Update sched_domains_numa_masks[level][node] array when new cpus
7258 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7259 unsigned long action,
7262 int cpu = (long)hcpu;
7264 switch (action & ~CPU_TASKS_FROZEN) {
7266 sched_domains_numa_masks_set(cpu);
7270 sched_domains_numa_masks_clear(cpu);
7280 static inline void sched_init_numa(void)
7284 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7285 unsigned long action,
7290 #endif /* CONFIG_NUMA */
7292 static int __sdt_alloc(const struct cpumask *cpu_map)
7294 struct sched_domain_topology_level *tl;
7297 for_each_sd_topology(tl) {
7298 struct sd_data *sdd = &tl->data;
7300 sdd->sd = alloc_percpu(struct sched_domain *);
7304 sdd->sg = alloc_percpu(struct sched_group *);
7308 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7312 for_each_cpu(j, cpu_map) {
7313 struct sched_domain *sd;
7314 struct sched_group *sg;
7315 struct sched_group_capacity *sgc;
7317 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7318 GFP_KERNEL, cpu_to_node(j));
7322 *per_cpu_ptr(sdd->sd, j) = sd;
7324 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7325 GFP_KERNEL, cpu_to_node(j));
7331 *per_cpu_ptr(sdd->sg, j) = sg;
7333 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7334 GFP_KERNEL, cpu_to_node(j));
7338 *per_cpu_ptr(sdd->sgc, j) = sgc;
7345 static void __sdt_free(const struct cpumask *cpu_map)
7347 struct sched_domain_topology_level *tl;
7350 for_each_sd_topology(tl) {
7351 struct sd_data *sdd = &tl->data;
7353 for_each_cpu(j, cpu_map) {
7354 struct sched_domain *sd;
7357 sd = *per_cpu_ptr(sdd->sd, j);
7358 if (sd && (sd->flags & SD_OVERLAP))
7359 free_sched_groups(sd->groups, 0);
7360 kfree(*per_cpu_ptr(sdd->sd, j));
7364 kfree(*per_cpu_ptr(sdd->sg, j));
7366 kfree(*per_cpu_ptr(sdd->sgc, j));
7368 free_percpu(sdd->sd);
7370 free_percpu(sdd->sg);
7372 free_percpu(sdd->sgc);
7377 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7378 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7379 struct sched_domain *child, int cpu)
7381 struct sched_domain *sd = sd_init(tl, child, cpu);
7383 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7385 sd->level = child->level + 1;
7386 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7389 if (!cpumask_subset(sched_domain_span(child),
7390 sched_domain_span(sd))) {
7391 pr_err("BUG: arch topology borken\n");
7392 #ifdef CONFIG_SCHED_DEBUG
7393 pr_err(" the %s domain not a subset of the %s domain\n",
7394 child->name, sd->name);
7396 /* Fixup, ensure @sd has at least @child cpus. */
7397 cpumask_or(sched_domain_span(sd),
7398 sched_domain_span(sd),
7399 sched_domain_span(child));
7403 set_domain_attribute(sd, attr);
7409 * Build sched domains for a given set of cpus and attach the sched domains
7410 * to the individual cpus
7412 static int build_sched_domains(const struct cpumask *cpu_map,
7413 struct sched_domain_attr *attr)
7415 enum s_alloc alloc_state;
7416 struct sched_domain *sd;
7418 int i, ret = -ENOMEM;
7420 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7421 if (alloc_state != sa_rootdomain)
7424 /* Set up domains for cpus specified by the cpu_map. */
7425 for_each_cpu(i, cpu_map) {
7426 struct sched_domain_topology_level *tl;
7429 for_each_sd_topology(tl) {
7430 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7431 if (tl == sched_domain_topology)
7432 *per_cpu_ptr(d.sd, i) = sd;
7433 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7434 sd->flags |= SD_OVERLAP;
7438 /* Build the groups for the domains */
7439 for_each_cpu(i, cpu_map) {
7440 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7441 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7442 if (sd->flags & SD_OVERLAP) {
7443 if (build_overlap_sched_groups(sd, i))
7446 if (build_sched_groups(sd, i))
7452 /* Calculate CPU capacity for physical packages and nodes */
7453 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7454 struct sched_domain_topology_level *tl = sched_domain_topology;
7456 if (!cpumask_test_cpu(i, cpu_map))
7459 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7460 init_sched_energy(i, sd, tl->energy);
7461 claim_allocations(i, sd);
7462 init_sched_groups_capacity(i, sd);
7466 /* Attach the domains */
7468 for_each_cpu(i, cpu_map) {
7469 int max_cpu = READ_ONCE(d.rd->max_cap_orig_cpu);
7470 int min_cpu = READ_ONCE(d.rd->min_cap_orig_cpu);
7472 if ((max_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig >
7473 cpu_rq(max_cpu)->cpu_capacity_orig))
7474 WRITE_ONCE(d.rd->max_cap_orig_cpu, i);
7476 if ((min_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig <
7477 cpu_rq(min_cpu)->cpu_capacity_orig))
7478 WRITE_ONCE(d.rd->min_cap_orig_cpu, i);
7480 sd = *per_cpu_ptr(d.sd, i);
7482 cpu_attach_domain(sd, d.rd, i);
7486 WARN(sched_feat(ENERGY_AWARE) && !have_sched_energy_data(),
7487 "Missing data for energy aware scheduling\n");
7491 __free_domain_allocs(&d, alloc_state, cpu_map);
7495 static cpumask_var_t *doms_cur; /* current sched domains */
7496 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7497 static struct sched_domain_attr *dattr_cur;
7498 /* attribues of custom domains in 'doms_cur' */
7501 * Special case: If a kmalloc of a doms_cur partition (array of
7502 * cpumask) fails, then fallback to a single sched domain,
7503 * as determined by the single cpumask fallback_doms.
7505 static cpumask_var_t fallback_doms;
7508 * arch_update_cpu_topology lets virtualized architectures update the
7509 * cpu core maps. It is supposed to return 1 if the topology changed
7510 * or 0 if it stayed the same.
7512 int __weak arch_update_cpu_topology(void)
7517 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7520 cpumask_var_t *doms;
7522 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7525 for (i = 0; i < ndoms; i++) {
7526 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7527 free_sched_domains(doms, i);
7534 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7537 for (i = 0; i < ndoms; i++)
7538 free_cpumask_var(doms[i]);
7543 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7544 * For now this just excludes isolated cpus, but could be used to
7545 * exclude other special cases in the future.
7547 static int init_sched_domains(const struct cpumask *cpu_map)
7551 arch_update_cpu_topology();
7553 doms_cur = alloc_sched_domains(ndoms_cur);
7555 doms_cur = &fallback_doms;
7556 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7557 err = build_sched_domains(doms_cur[0], NULL);
7558 register_sched_domain_sysctl();
7564 * Detach sched domains from a group of cpus specified in cpu_map
7565 * These cpus will now be attached to the NULL domain
7567 static void detach_destroy_domains(const struct cpumask *cpu_map)
7572 for_each_cpu(i, cpu_map)
7573 cpu_attach_domain(NULL, &def_root_domain, i);
7577 /* handle null as "default" */
7578 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7579 struct sched_domain_attr *new, int idx_new)
7581 struct sched_domain_attr tmp;
7588 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7589 new ? (new + idx_new) : &tmp,
7590 sizeof(struct sched_domain_attr));
7594 * Partition sched domains as specified by the 'ndoms_new'
7595 * cpumasks in the array doms_new[] of cpumasks. This compares
7596 * doms_new[] to the current sched domain partitioning, doms_cur[].
7597 * It destroys each deleted domain and builds each new domain.
7599 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7600 * The masks don't intersect (don't overlap.) We should setup one
7601 * sched domain for each mask. CPUs not in any of the cpumasks will
7602 * not be load balanced. If the same cpumask appears both in the
7603 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7606 * The passed in 'doms_new' should be allocated using
7607 * alloc_sched_domains. This routine takes ownership of it and will
7608 * free_sched_domains it when done with it. If the caller failed the
7609 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7610 * and partition_sched_domains() will fallback to the single partition
7611 * 'fallback_doms', it also forces the domains to be rebuilt.
7613 * If doms_new == NULL it will be replaced with cpu_online_mask.
7614 * ndoms_new == 0 is a special case for destroying existing domains,
7615 * and it will not create the default domain.
7617 * Call with hotplug lock held
7619 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7620 struct sched_domain_attr *dattr_new)
7625 mutex_lock(&sched_domains_mutex);
7627 /* always unregister in case we don't destroy any domains */
7628 unregister_sched_domain_sysctl();
7630 /* Let architecture update cpu core mappings. */
7631 new_topology = arch_update_cpu_topology();
7633 n = doms_new ? ndoms_new : 0;
7635 /* Destroy deleted domains */
7636 for (i = 0; i < ndoms_cur; i++) {
7637 for (j = 0; j < n && !new_topology; j++) {
7638 if (cpumask_equal(doms_cur[i], doms_new[j])
7639 && dattrs_equal(dattr_cur, i, dattr_new, j))
7642 /* no match - a current sched domain not in new doms_new[] */
7643 detach_destroy_domains(doms_cur[i]);
7649 if (doms_new == NULL) {
7651 doms_new = &fallback_doms;
7652 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7653 WARN_ON_ONCE(dattr_new);
7656 /* Build new domains */
7657 for (i = 0; i < ndoms_new; i++) {
7658 for (j = 0; j < n && !new_topology; j++) {
7659 if (cpumask_equal(doms_new[i], doms_cur[j])
7660 && dattrs_equal(dattr_new, i, dattr_cur, j))
7663 /* no match - add a new doms_new */
7664 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7669 /* Remember the new sched domains */
7670 if (doms_cur != &fallback_doms)
7671 free_sched_domains(doms_cur, ndoms_cur);
7672 kfree(dattr_cur); /* kfree(NULL) is safe */
7673 doms_cur = doms_new;
7674 dattr_cur = dattr_new;
7675 ndoms_cur = ndoms_new;
7677 register_sched_domain_sysctl();
7679 mutex_unlock(&sched_domains_mutex);
7682 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7685 * Update cpusets according to cpu_active mask. If cpusets are
7686 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7687 * around partition_sched_domains().
7689 * If we come here as part of a suspend/resume, don't touch cpusets because we
7690 * want to restore it back to its original state upon resume anyway.
7692 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7696 case CPU_ONLINE_FROZEN:
7697 case CPU_DOWN_FAILED_FROZEN:
7700 * num_cpus_frozen tracks how many CPUs are involved in suspend
7701 * resume sequence. As long as this is not the last online
7702 * operation in the resume sequence, just build a single sched
7703 * domain, ignoring cpusets.
7705 partition_sched_domains(1, NULL, NULL);
7706 if (--num_cpus_frozen)
7710 * This is the last CPU online operation. So fall through and
7711 * restore the original sched domains by considering the
7712 * cpuset configurations.
7714 cpuset_force_rebuild();
7717 cpuset_update_active_cpus(true);
7725 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7728 unsigned long flags;
7729 long cpu = (long)hcpu;
7735 case CPU_DOWN_PREPARE:
7736 rcu_read_lock_sched();
7737 dl_b = dl_bw_of(cpu);
7739 raw_spin_lock_irqsave(&dl_b->lock, flags);
7740 cpus = dl_bw_cpus(cpu);
7741 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7742 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7744 rcu_read_unlock_sched();
7747 return notifier_from_errno(-EBUSY);
7748 cpuset_update_active_cpus(false);
7750 case CPU_DOWN_PREPARE_FROZEN:
7752 partition_sched_domains(1, NULL, NULL);
7760 void __init sched_init_smp(void)
7762 cpumask_var_t non_isolated_cpus;
7764 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7765 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7770 * There's no userspace yet to cause hotplug operations; hence all the
7771 * cpu masks are stable and all blatant races in the below code cannot
7774 mutex_lock(&sched_domains_mutex);
7775 init_sched_domains(cpu_active_mask);
7776 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7777 if (cpumask_empty(non_isolated_cpus))
7778 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7779 mutex_unlock(&sched_domains_mutex);
7781 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7782 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7783 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7787 /* Move init over to a non-isolated CPU */
7788 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7790 sched_init_granularity();
7791 free_cpumask_var(non_isolated_cpus);
7793 init_sched_rt_class();
7794 init_sched_dl_class();
7797 void __init sched_init_smp(void)
7799 sched_init_granularity();
7801 #endif /* CONFIG_SMP */
7803 int in_sched_functions(unsigned long addr)
7805 return in_lock_functions(addr) ||
7806 (addr >= (unsigned long)__sched_text_start
7807 && addr < (unsigned long)__sched_text_end);
7810 #ifdef CONFIG_CGROUP_SCHED
7812 * Default task group.
7813 * Every task in system belongs to this group at bootup.
7815 struct task_group root_task_group;
7816 LIST_HEAD(task_groups);
7819 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7821 void __init sched_init(void)
7824 unsigned long alloc_size = 0, ptr;
7826 #ifdef CONFIG_FAIR_GROUP_SCHED
7827 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7829 #ifdef CONFIG_RT_GROUP_SCHED
7830 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7833 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7835 #ifdef CONFIG_FAIR_GROUP_SCHED
7836 root_task_group.se = (struct sched_entity **)ptr;
7837 ptr += nr_cpu_ids * sizeof(void **);
7839 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7840 ptr += nr_cpu_ids * sizeof(void **);
7842 #endif /* CONFIG_FAIR_GROUP_SCHED */
7843 #ifdef CONFIG_RT_GROUP_SCHED
7844 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7845 ptr += nr_cpu_ids * sizeof(void **);
7847 root_task_group.rt_rq = (struct rt_rq **)ptr;
7848 ptr += nr_cpu_ids * sizeof(void **);
7850 #endif /* CONFIG_RT_GROUP_SCHED */
7852 #ifdef CONFIG_CPUMASK_OFFSTACK
7853 for_each_possible_cpu(i) {
7854 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7855 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7857 #endif /* CONFIG_CPUMASK_OFFSTACK */
7859 init_rt_bandwidth(&def_rt_bandwidth,
7860 global_rt_period(), global_rt_runtime());
7861 init_dl_bandwidth(&def_dl_bandwidth,
7862 global_rt_period(), global_rt_runtime());
7865 init_defrootdomain();
7868 #ifdef CONFIG_RT_GROUP_SCHED
7869 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7870 global_rt_period(), global_rt_runtime());
7871 #endif /* CONFIG_RT_GROUP_SCHED */
7873 #ifdef CONFIG_CGROUP_SCHED
7874 list_add(&root_task_group.list, &task_groups);
7875 INIT_LIST_HEAD(&root_task_group.children);
7876 INIT_LIST_HEAD(&root_task_group.siblings);
7877 autogroup_init(&init_task);
7879 #endif /* CONFIG_CGROUP_SCHED */
7881 for_each_possible_cpu(i) {
7885 raw_spin_lock_init(&rq->lock);
7887 rq->calc_load_active = 0;
7888 rq->calc_load_update = jiffies + LOAD_FREQ;
7889 init_cfs_rq(&rq->cfs);
7890 init_rt_rq(&rq->rt);
7891 init_dl_rq(&rq->dl);
7892 #ifdef CONFIG_FAIR_GROUP_SCHED
7893 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7894 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7895 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7897 * How much cpu bandwidth does root_task_group get?
7899 * In case of task-groups formed thr' the cgroup filesystem, it
7900 * gets 100% of the cpu resources in the system. This overall
7901 * system cpu resource is divided among the tasks of
7902 * root_task_group and its child task-groups in a fair manner,
7903 * based on each entity's (task or task-group's) weight
7904 * (se->load.weight).
7906 * In other words, if root_task_group has 10 tasks of weight
7907 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7908 * then A0's share of the cpu resource is:
7910 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7912 * We achieve this by letting root_task_group's tasks sit
7913 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7915 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7916 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7917 #endif /* CONFIG_FAIR_GROUP_SCHED */
7919 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7920 #ifdef CONFIG_RT_GROUP_SCHED
7921 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7924 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7925 rq->cpu_load[j] = 0;
7927 rq->last_load_update_tick = jiffies;
7932 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7933 rq->balance_callback = NULL;
7934 rq->active_balance = 0;
7935 rq->next_balance = jiffies;
7940 rq->avg_idle = 2*sysctl_sched_migration_cost;
7941 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7942 #ifdef CONFIG_SCHED_WALT
7943 rq->cur_irqload = 0;
7944 rq->avg_irqload = 0;
7948 INIT_LIST_HEAD(&rq->cfs_tasks);
7950 rq_attach_root(rq, &def_root_domain);
7951 #ifdef CONFIG_NO_HZ_COMMON
7954 #ifdef CONFIG_NO_HZ_FULL
7955 rq->last_sched_tick = 0;
7959 atomic_set(&rq->nr_iowait, 0);
7962 set_load_weight(&init_task);
7964 #ifdef CONFIG_PREEMPT_NOTIFIERS
7965 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7969 * The boot idle thread does lazy MMU switching as well:
7971 atomic_inc(&init_mm.mm_count);
7972 enter_lazy_tlb(&init_mm, current);
7975 * During early bootup we pretend to be a normal task:
7977 current->sched_class = &fair_sched_class;
7980 * Make us the idle thread. Technically, schedule() should not be
7981 * called from this thread, however somewhere below it might be,
7982 * but because we are the idle thread, we just pick up running again
7983 * when this runqueue becomes "idle".
7985 init_idle(current, smp_processor_id());
7987 calc_load_update = jiffies + LOAD_FREQ;
7990 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7991 /* May be allocated at isolcpus cmdline parse time */
7992 if (cpu_isolated_map == NULL)
7993 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7994 idle_thread_set_boot_cpu();
7995 set_cpu_rq_start_time();
7997 init_sched_fair_class();
7999 scheduler_running = 1;
8002 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8003 static inline int preempt_count_equals(int preempt_offset)
8005 int nested = preempt_count() + rcu_preempt_depth();
8007 return (nested == preempt_offset);
8010 static int __might_sleep_init_called;
8011 int __init __might_sleep_init(void)
8013 __might_sleep_init_called = 1;
8016 early_initcall(__might_sleep_init);
8018 void __might_sleep(const char *file, int line, int preempt_offset)
8021 * Blocking primitives will set (and therefore destroy) current->state,
8022 * since we will exit with TASK_RUNNING make sure we enter with it,
8023 * otherwise we will destroy state.
8025 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
8026 "do not call blocking ops when !TASK_RUNNING; "
8027 "state=%lx set at [<%p>] %pS\n",
8029 (void *)current->task_state_change,
8030 (void *)current->task_state_change);
8032 ___might_sleep(file, line, preempt_offset);
8034 EXPORT_SYMBOL(__might_sleep);
8036 void ___might_sleep(const char *file, int line, int preempt_offset)
8038 static unsigned long prev_jiffy; /* ratelimiting */
8040 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8041 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
8042 !is_idle_task(current)) || oops_in_progress)
8044 if (system_state != SYSTEM_RUNNING &&
8045 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
8047 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8049 prev_jiffy = jiffies;
8052 "BUG: sleeping function called from invalid context at %s:%d\n",
8055 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8056 in_atomic(), irqs_disabled(),
8057 current->pid, current->comm);
8059 if (task_stack_end_corrupted(current))
8060 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
8062 debug_show_held_locks(current);
8063 if (irqs_disabled())
8064 print_irqtrace_events(current);
8065 #ifdef CONFIG_DEBUG_PREEMPT
8066 if (!preempt_count_equals(preempt_offset)) {
8067 pr_err("Preemption disabled at:");
8068 print_ip_sym(current->preempt_disable_ip);
8074 EXPORT_SYMBOL(___might_sleep);
8077 #ifdef CONFIG_MAGIC_SYSRQ
8078 void normalize_rt_tasks(void)
8080 struct task_struct *g, *p;
8081 struct sched_attr attr = {
8082 .sched_policy = SCHED_NORMAL,
8085 read_lock(&tasklist_lock);
8086 for_each_process_thread(g, p) {
8088 * Only normalize user tasks:
8090 if (p->flags & PF_KTHREAD)
8093 p->se.exec_start = 0;
8094 #ifdef CONFIG_SCHEDSTATS
8095 p->se.statistics.wait_start = 0;
8096 p->se.statistics.sleep_start = 0;
8097 p->se.statistics.block_start = 0;
8100 if (!dl_task(p) && !rt_task(p)) {
8102 * Renice negative nice level userspace
8105 if (task_nice(p) < 0)
8106 set_user_nice(p, 0);
8110 __sched_setscheduler(p, &attr, false, false);
8112 read_unlock(&tasklist_lock);
8115 #endif /* CONFIG_MAGIC_SYSRQ */
8117 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8119 * These functions are only useful for the IA64 MCA handling, or kdb.
8121 * They can only be called when the whole system has been
8122 * stopped - every CPU needs to be quiescent, and no scheduling
8123 * activity can take place. Using them for anything else would
8124 * be a serious bug, and as a result, they aren't even visible
8125 * under any other configuration.
8129 * curr_task - return the current task for a given cpu.
8130 * @cpu: the processor in question.
8132 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8134 * Return: The current task for @cpu.
8136 struct task_struct *curr_task(int cpu)
8138 return cpu_curr(cpu);
8141 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8145 * set_curr_task - set the current task for a given cpu.
8146 * @cpu: the processor in question.
8147 * @p: the task pointer to set.
8149 * Description: This function must only be used when non-maskable interrupts
8150 * are serviced on a separate stack. It allows the architecture to switch the
8151 * notion of the current task on a cpu in a non-blocking manner. This function
8152 * must be called with all CPU's synchronized, and interrupts disabled, the
8153 * and caller must save the original value of the current task (see
8154 * curr_task() above) and restore that value before reenabling interrupts and
8155 * re-starting the system.
8157 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8159 void set_curr_task(int cpu, struct task_struct *p)
8166 #ifdef CONFIG_CGROUP_SCHED
8167 /* task_group_lock serializes the addition/removal of task groups */
8168 static DEFINE_SPINLOCK(task_group_lock);
8170 static void sched_free_group(struct task_group *tg)
8172 free_fair_sched_group(tg);
8173 free_rt_sched_group(tg);
8178 /* allocate runqueue etc for a new task group */
8179 struct task_group *sched_create_group(struct task_group *parent)
8181 struct task_group *tg;
8183 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8185 return ERR_PTR(-ENOMEM);
8187 if (!alloc_fair_sched_group(tg, parent))
8190 if (!alloc_rt_sched_group(tg, parent))
8196 sched_free_group(tg);
8197 return ERR_PTR(-ENOMEM);
8200 void sched_online_group(struct task_group *tg, struct task_group *parent)
8202 unsigned long flags;
8204 spin_lock_irqsave(&task_group_lock, flags);
8205 list_add_rcu(&tg->list, &task_groups);
8207 WARN_ON(!parent); /* root should already exist */
8209 tg->parent = parent;
8210 INIT_LIST_HEAD(&tg->children);
8211 list_add_rcu(&tg->siblings, &parent->children);
8212 spin_unlock_irqrestore(&task_group_lock, flags);
8215 /* rcu callback to free various structures associated with a task group */
8216 static void sched_free_group_rcu(struct rcu_head *rhp)
8218 /* now it should be safe to free those cfs_rqs */
8219 sched_free_group(container_of(rhp, struct task_group, rcu));
8222 void sched_destroy_group(struct task_group *tg)
8224 /* wait for possible concurrent references to cfs_rqs complete */
8225 call_rcu(&tg->rcu, sched_free_group_rcu);
8228 void sched_offline_group(struct task_group *tg)
8230 unsigned long flags;
8233 /* end participation in shares distribution */
8234 for_each_possible_cpu(i)
8235 unregister_fair_sched_group(tg, i);
8237 spin_lock_irqsave(&task_group_lock, flags);
8238 list_del_rcu(&tg->list);
8239 list_del_rcu(&tg->siblings);
8240 spin_unlock_irqrestore(&task_group_lock, flags);
8243 static void sched_change_group(struct task_struct *tsk, int type)
8245 struct task_group *tg;
8248 * All callers are synchronized by task_rq_lock(); we do not use RCU
8249 * which is pointless here. Thus, we pass "true" to task_css_check()
8250 * to prevent lockdep warnings.
8252 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8253 struct task_group, css);
8254 tg = autogroup_task_group(tsk, tg);
8255 tsk->sched_task_group = tg;
8257 #ifdef CONFIG_FAIR_GROUP_SCHED
8258 if (tsk->sched_class->task_change_group)
8259 tsk->sched_class->task_change_group(tsk, type);
8262 set_task_rq(tsk, task_cpu(tsk));
8266 * Change task's runqueue when it moves between groups.
8268 * The caller of this function should have put the task in its new group by
8269 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8272 void sched_move_task(struct task_struct *tsk)
8274 int queued, running;
8275 unsigned long flags;
8278 rq = task_rq_lock(tsk, &flags);
8280 running = task_current(rq, tsk);
8281 queued = task_on_rq_queued(tsk);
8284 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8285 if (unlikely(running))
8286 put_prev_task(rq, tsk);
8288 sched_change_group(tsk, TASK_MOVE_GROUP);
8290 if (unlikely(running))
8291 tsk->sched_class->set_curr_task(rq);
8293 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8295 task_rq_unlock(rq, tsk, &flags);
8297 #endif /* CONFIG_CGROUP_SCHED */
8299 #ifdef CONFIG_RT_GROUP_SCHED
8301 * Ensure that the real time constraints are schedulable.
8303 static DEFINE_MUTEX(rt_constraints_mutex);
8305 /* Must be called with tasklist_lock held */
8306 static inline int tg_has_rt_tasks(struct task_group *tg)
8308 struct task_struct *g, *p;
8311 * Autogroups do not have RT tasks; see autogroup_create().
8313 if (task_group_is_autogroup(tg))
8316 for_each_process_thread(g, p) {
8317 if (rt_task(p) && task_group(p) == tg)
8324 struct rt_schedulable_data {
8325 struct task_group *tg;
8330 static int tg_rt_schedulable(struct task_group *tg, void *data)
8332 struct rt_schedulable_data *d = data;
8333 struct task_group *child;
8334 unsigned long total, sum = 0;
8335 u64 period, runtime;
8337 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8338 runtime = tg->rt_bandwidth.rt_runtime;
8341 period = d->rt_period;
8342 runtime = d->rt_runtime;
8346 * Cannot have more runtime than the period.
8348 if (runtime > period && runtime != RUNTIME_INF)
8352 * Ensure we don't starve existing RT tasks.
8354 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8357 total = to_ratio(period, runtime);
8360 * Nobody can have more than the global setting allows.
8362 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8366 * The sum of our children's runtime should not exceed our own.
8368 list_for_each_entry_rcu(child, &tg->children, siblings) {
8369 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8370 runtime = child->rt_bandwidth.rt_runtime;
8372 if (child == d->tg) {
8373 period = d->rt_period;
8374 runtime = d->rt_runtime;
8377 sum += to_ratio(period, runtime);
8386 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8390 struct rt_schedulable_data data = {
8392 .rt_period = period,
8393 .rt_runtime = runtime,
8397 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8403 static int tg_set_rt_bandwidth(struct task_group *tg,
8404 u64 rt_period, u64 rt_runtime)
8409 * Disallowing the root group RT runtime is BAD, it would disallow the
8410 * kernel creating (and or operating) RT threads.
8412 if (tg == &root_task_group && rt_runtime == 0)
8415 /* No period doesn't make any sense. */
8419 mutex_lock(&rt_constraints_mutex);
8420 read_lock(&tasklist_lock);
8421 err = __rt_schedulable(tg, rt_period, rt_runtime);
8425 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8426 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8427 tg->rt_bandwidth.rt_runtime = rt_runtime;
8429 for_each_possible_cpu(i) {
8430 struct rt_rq *rt_rq = tg->rt_rq[i];
8432 raw_spin_lock(&rt_rq->rt_runtime_lock);
8433 rt_rq->rt_runtime = rt_runtime;
8434 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8436 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8438 read_unlock(&tasklist_lock);
8439 mutex_unlock(&rt_constraints_mutex);
8444 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8446 u64 rt_runtime, rt_period;
8448 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8449 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8450 if (rt_runtime_us < 0)
8451 rt_runtime = RUNTIME_INF;
8453 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8456 static long sched_group_rt_runtime(struct task_group *tg)
8460 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8463 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8464 do_div(rt_runtime_us, NSEC_PER_USEC);
8465 return rt_runtime_us;
8468 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8470 u64 rt_runtime, rt_period;
8472 rt_period = rt_period_us * NSEC_PER_USEC;
8473 rt_runtime = tg->rt_bandwidth.rt_runtime;
8475 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8478 static long sched_group_rt_period(struct task_group *tg)
8482 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8483 do_div(rt_period_us, NSEC_PER_USEC);
8484 return rt_period_us;
8486 #endif /* CONFIG_RT_GROUP_SCHED */
8488 #ifdef CONFIG_RT_GROUP_SCHED
8489 static int sched_rt_global_constraints(void)
8493 mutex_lock(&rt_constraints_mutex);
8494 read_lock(&tasklist_lock);
8495 ret = __rt_schedulable(NULL, 0, 0);
8496 read_unlock(&tasklist_lock);
8497 mutex_unlock(&rt_constraints_mutex);
8502 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8504 /* Don't accept realtime tasks when there is no way for them to run */
8505 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8511 #else /* !CONFIG_RT_GROUP_SCHED */
8512 static int sched_rt_global_constraints(void)
8514 unsigned long flags;
8517 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8518 for_each_possible_cpu(i) {
8519 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8521 raw_spin_lock(&rt_rq->rt_runtime_lock);
8522 rt_rq->rt_runtime = global_rt_runtime();
8523 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8525 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8529 #endif /* CONFIG_RT_GROUP_SCHED */
8531 static int sched_dl_global_validate(void)
8533 u64 runtime = global_rt_runtime();
8534 u64 period = global_rt_period();
8535 u64 new_bw = to_ratio(period, runtime);
8538 unsigned long flags;
8541 * Here we want to check the bandwidth not being set to some
8542 * value smaller than the currently allocated bandwidth in
8543 * any of the root_domains.
8545 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8546 * cycling on root_domains... Discussion on different/better
8547 * solutions is welcome!
8549 for_each_possible_cpu(cpu) {
8550 rcu_read_lock_sched();
8551 dl_b = dl_bw_of(cpu);
8553 raw_spin_lock_irqsave(&dl_b->lock, flags);
8554 if (new_bw < dl_b->total_bw)
8556 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8558 rcu_read_unlock_sched();
8567 static void sched_dl_do_global(void)
8572 unsigned long flags;
8574 def_dl_bandwidth.dl_period = global_rt_period();
8575 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8577 if (global_rt_runtime() != RUNTIME_INF)
8578 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8581 * FIXME: As above...
8583 for_each_possible_cpu(cpu) {
8584 rcu_read_lock_sched();
8585 dl_b = dl_bw_of(cpu);
8587 raw_spin_lock_irqsave(&dl_b->lock, flags);
8589 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8591 rcu_read_unlock_sched();
8595 static int sched_rt_global_validate(void)
8597 if (sysctl_sched_rt_period <= 0)
8600 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8601 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8607 static void sched_rt_do_global(void)
8609 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8610 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8613 int sched_rt_handler(struct ctl_table *table, int write,
8614 void __user *buffer, size_t *lenp,
8617 int old_period, old_runtime;
8618 static DEFINE_MUTEX(mutex);
8622 old_period = sysctl_sched_rt_period;
8623 old_runtime = sysctl_sched_rt_runtime;
8625 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8627 if (!ret && write) {
8628 ret = sched_rt_global_validate();
8632 ret = sched_dl_global_validate();
8636 ret = sched_rt_global_constraints();
8640 sched_rt_do_global();
8641 sched_dl_do_global();
8645 sysctl_sched_rt_period = old_period;
8646 sysctl_sched_rt_runtime = old_runtime;
8648 mutex_unlock(&mutex);
8653 int sched_rr_handler(struct ctl_table *table, int write,
8654 void __user *buffer, size_t *lenp,
8658 static DEFINE_MUTEX(mutex);
8661 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8662 /* make sure that internally we keep jiffies */
8663 /* also, writing zero resets timeslice to default */
8664 if (!ret && write) {
8665 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8666 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8668 mutex_unlock(&mutex);
8672 #ifdef CONFIG_CGROUP_SCHED
8674 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8676 return css ? container_of(css, struct task_group, css) : NULL;
8679 static struct cgroup_subsys_state *
8680 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8682 struct task_group *parent = css_tg(parent_css);
8683 struct task_group *tg;
8686 /* This is early initialization for the top cgroup */
8687 return &root_task_group.css;
8690 tg = sched_create_group(parent);
8692 return ERR_PTR(-ENOMEM);
8697 /* Expose task group only after completing cgroup initialization */
8698 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8700 struct task_group *tg = css_tg(css);
8701 struct task_group *parent = css_tg(css->parent);
8704 sched_online_group(tg, parent);
8708 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8710 struct task_group *tg = css_tg(css);
8712 sched_offline_group(tg);
8715 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8717 struct task_group *tg = css_tg(css);
8720 * Relies on the RCU grace period between css_released() and this.
8722 sched_free_group(tg);
8726 * This is called before wake_up_new_task(), therefore we really only
8727 * have to set its group bits, all the other stuff does not apply.
8729 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8731 unsigned long flags;
8734 rq = task_rq_lock(task, &flags);
8736 update_rq_clock(rq);
8737 sched_change_group(task, TASK_SET_GROUP);
8739 task_rq_unlock(rq, task, &flags);
8742 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8744 struct task_struct *task;
8745 struct cgroup_subsys_state *css;
8748 cgroup_taskset_for_each(task, css, tset) {
8749 #ifdef CONFIG_RT_GROUP_SCHED
8750 if (!sched_rt_can_attach(css_tg(css), task))
8753 /* We don't support RT-tasks being in separate groups */
8754 if (task->sched_class != &fair_sched_class)
8758 * Serialize against wake_up_new_task() such that if its
8759 * running, we're sure to observe its full state.
8761 raw_spin_lock_irq(&task->pi_lock);
8763 * Avoid calling sched_move_task() before wake_up_new_task()
8764 * has happened. This would lead to problems with PELT, due to
8765 * move wanting to detach+attach while we're not attached yet.
8767 if (task->state == TASK_NEW)
8769 raw_spin_unlock_irq(&task->pi_lock);
8777 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8779 struct task_struct *task;
8780 struct cgroup_subsys_state *css;
8782 cgroup_taskset_for_each(task, css, tset)
8783 sched_move_task(task);
8786 #ifdef CONFIG_FAIR_GROUP_SCHED
8787 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8788 struct cftype *cftype, u64 shareval)
8790 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8793 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8796 struct task_group *tg = css_tg(css);
8798 return (u64) scale_load_down(tg->shares);
8801 #ifdef CONFIG_CFS_BANDWIDTH
8802 static DEFINE_MUTEX(cfs_constraints_mutex);
8804 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8805 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8807 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8809 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8811 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8812 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8814 if (tg == &root_task_group)
8818 * Ensure we have at some amount of bandwidth every period. This is
8819 * to prevent reaching a state of large arrears when throttled via
8820 * entity_tick() resulting in prolonged exit starvation.
8822 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8826 * Likewise, bound things on the otherside by preventing insane quota
8827 * periods. This also allows us to normalize in computing quota
8830 if (period > max_cfs_quota_period)
8834 * Prevent race between setting of cfs_rq->runtime_enabled and
8835 * unthrottle_offline_cfs_rqs().
8838 mutex_lock(&cfs_constraints_mutex);
8839 ret = __cfs_schedulable(tg, period, quota);
8843 runtime_enabled = quota != RUNTIME_INF;
8844 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8846 * If we need to toggle cfs_bandwidth_used, off->on must occur
8847 * before making related changes, and on->off must occur afterwards
8849 if (runtime_enabled && !runtime_was_enabled)
8850 cfs_bandwidth_usage_inc();
8851 raw_spin_lock_irq(&cfs_b->lock);
8852 cfs_b->period = ns_to_ktime(period);
8853 cfs_b->quota = quota;
8855 __refill_cfs_bandwidth_runtime(cfs_b);
8856 /* restart the period timer (if active) to handle new period expiry */
8857 if (runtime_enabled)
8858 start_cfs_bandwidth(cfs_b);
8859 raw_spin_unlock_irq(&cfs_b->lock);
8861 for_each_online_cpu(i) {
8862 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8863 struct rq *rq = cfs_rq->rq;
8865 raw_spin_lock_irq(&rq->lock);
8866 cfs_rq->runtime_enabled = runtime_enabled;
8867 cfs_rq->runtime_remaining = 0;
8869 if (cfs_rq->throttled)
8870 unthrottle_cfs_rq(cfs_rq);
8871 raw_spin_unlock_irq(&rq->lock);
8873 if (runtime_was_enabled && !runtime_enabled)
8874 cfs_bandwidth_usage_dec();
8876 mutex_unlock(&cfs_constraints_mutex);
8882 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8886 period = ktime_to_ns(tg->cfs_bandwidth.period);
8887 if (cfs_quota_us < 0)
8888 quota = RUNTIME_INF;
8890 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8892 return tg_set_cfs_bandwidth(tg, period, quota);
8895 long tg_get_cfs_quota(struct task_group *tg)
8899 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8902 quota_us = tg->cfs_bandwidth.quota;
8903 do_div(quota_us, NSEC_PER_USEC);
8908 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8912 period = (u64)cfs_period_us * NSEC_PER_USEC;
8913 quota = tg->cfs_bandwidth.quota;
8915 return tg_set_cfs_bandwidth(tg, period, quota);
8918 long tg_get_cfs_period(struct task_group *tg)
8922 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8923 do_div(cfs_period_us, NSEC_PER_USEC);
8925 return cfs_period_us;
8928 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8931 return tg_get_cfs_quota(css_tg(css));
8934 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8935 struct cftype *cftype, s64 cfs_quota_us)
8937 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8940 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8943 return tg_get_cfs_period(css_tg(css));
8946 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8947 struct cftype *cftype, u64 cfs_period_us)
8949 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8952 struct cfs_schedulable_data {
8953 struct task_group *tg;
8958 * normalize group quota/period to be quota/max_period
8959 * note: units are usecs
8961 static u64 normalize_cfs_quota(struct task_group *tg,
8962 struct cfs_schedulable_data *d)
8970 period = tg_get_cfs_period(tg);
8971 quota = tg_get_cfs_quota(tg);
8974 /* note: these should typically be equivalent */
8975 if (quota == RUNTIME_INF || quota == -1)
8978 return to_ratio(period, quota);
8981 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8983 struct cfs_schedulable_data *d = data;
8984 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8985 s64 quota = 0, parent_quota = -1;
8988 quota = RUNTIME_INF;
8990 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8992 quota = normalize_cfs_quota(tg, d);
8993 parent_quota = parent_b->hierarchical_quota;
8996 * ensure max(child_quota) <= parent_quota, inherit when no
8999 if (quota == RUNTIME_INF)
9000 quota = parent_quota;
9001 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9004 cfs_b->hierarchical_quota = quota;
9009 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9012 struct cfs_schedulable_data data = {
9018 if (quota != RUNTIME_INF) {
9019 do_div(data.period, NSEC_PER_USEC);
9020 do_div(data.quota, NSEC_PER_USEC);
9024 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9030 static int cpu_stats_show(struct seq_file *sf, void *v)
9032 struct task_group *tg = css_tg(seq_css(sf));
9033 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
9035 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
9036 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
9037 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
9041 #endif /* CONFIG_CFS_BANDWIDTH */
9042 #endif /* CONFIG_FAIR_GROUP_SCHED */
9044 #ifdef CONFIG_RT_GROUP_SCHED
9045 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
9046 struct cftype *cft, s64 val)
9048 return sched_group_set_rt_runtime(css_tg(css), val);
9051 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
9054 return sched_group_rt_runtime(css_tg(css));
9057 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
9058 struct cftype *cftype, u64 rt_period_us)
9060 return sched_group_set_rt_period(css_tg(css), rt_period_us);
9063 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
9066 return sched_group_rt_period(css_tg(css));
9068 #endif /* CONFIG_RT_GROUP_SCHED */
9070 static struct cftype cpu_files[] = {
9071 #ifdef CONFIG_FAIR_GROUP_SCHED
9074 .read_u64 = cpu_shares_read_u64,
9075 .write_u64 = cpu_shares_write_u64,
9078 #ifdef CONFIG_CFS_BANDWIDTH
9080 .name = "cfs_quota_us",
9081 .read_s64 = cpu_cfs_quota_read_s64,
9082 .write_s64 = cpu_cfs_quota_write_s64,
9085 .name = "cfs_period_us",
9086 .read_u64 = cpu_cfs_period_read_u64,
9087 .write_u64 = cpu_cfs_period_write_u64,
9091 .seq_show = cpu_stats_show,
9094 #ifdef CONFIG_RT_GROUP_SCHED
9096 .name = "rt_runtime_us",
9097 .read_s64 = cpu_rt_runtime_read,
9098 .write_s64 = cpu_rt_runtime_write,
9101 .name = "rt_period_us",
9102 .read_u64 = cpu_rt_period_read_uint,
9103 .write_u64 = cpu_rt_period_write_uint,
9109 struct cgroup_subsys cpu_cgrp_subsys = {
9110 .css_alloc = cpu_cgroup_css_alloc,
9111 .css_online = cpu_cgroup_css_online,
9112 .css_released = cpu_cgroup_css_released,
9113 .css_free = cpu_cgroup_css_free,
9114 .fork = cpu_cgroup_fork,
9115 .can_attach = cpu_cgroup_can_attach,
9116 .attach = cpu_cgroup_attach,
9117 .legacy_cftypes = cpu_files,
9121 #endif /* CONFIG_CGROUP_SCHED */
9123 void dump_cpu_task(int cpu)
9125 pr_info("Task dump for CPU %d:\n", cpu);
9126 sched_show_task(cpu_curr(cpu));