Linux proc schedstat调度统计信息与se.statistics/proc//schedstat输出由schedstat_open()在procfs中注册通过show_schedstat()遍历系统中所有任务并聚合其se.statistics字段。输出格式为三个无符号长整型数列sched_yield()触发次数实际递增的是rq-yld_count但每个任务独立计数在schedstat中称为yld_count、调度延迟总和sum_sleep_runtime、先占次数sched_info上通过sched_info_on的计数器。c/* kernel/sched/stats.c */static int show_schedstat(struct seq_file *seq, void *v){struct rq *rq;int cpu;seq_printf(seq, version %d\n, SCHEDSTAT_VERSION);seq_printf(seq, timestamp %llu\n, sched_clock());for_each_online_cpu(cpu) {rq cpu_rq(cpu);seq_printf(seq, cpu%d %u %u %u %u %u %u %u %u\n,cpu, rq-yld_count, rq-sched_count,rq-sched_goidle, rq-avg_idle,rq-ttwu_count, rq-ttwu_local,rq-sched_switch, rq-sched_cfs_count);}return 0;}struct sched_statistics是每个调度实体se的内嵌统计结构包含sum_sleep_runtime、sum_block_runtime、wait_start、wait_sum、wait_count、iowait_sum、iowait_count、exec_start、sum_exec_runtime、nr_migrations等计数器。这些计数器在内核编译CONFIG_SCHEDSTATSy时启用否则退化为空结构以消除运行时开销。c#ifdef CONFIG_SCHEDSTATSstruct sched_statistics {u64 wait_start;u64 wait_max;u64 wait_count;u64 wait_sum;u64 iowait_count;u64 iowait_sum;u64 sleep_start;u64 sleep_max;u64 sum_sleep_runtime;u64 block_start;u64 block_max;u64 sum_block_runtime;u64 exec_max;u64 slice_max;u64 nr_migrations_cold;u64 nr_failed_migrations_affine;u64 nr_failed_migrations_running;u64 nr_failed_migrations_hot;u64 nr_forced_migrations;u64 nr_wakeups;u64 nr_wakeups_sync;u64 nr_wakeups_migrate;u64 nr_wakeups_local;u64 nr_wakeups_remote;u64 nr_wakeups_affine;u64 nr_wakeups_affine_attempts;u64 nr_wakeups_passive;u64 nr_wakeups_idle;};sum_sleep_runtime在进程从TASK_RUNNING退出到TASK_INTERRUPTIBLE时累积。__schedule()中context_switch前调用deactivate_task()并触发__schedstat_set_sleep_start()记录sleep_start当进程再次被wakeup时wake_up_process()的ttwu_do_activate()调用__schedstat_add_sleep_runtime()计算sleep_start到当前时间的差值并累加到sum_sleep_runtime。这里的关键是sleep_start必须在持有rq-lock时更新否则wakeup和schedule的race会导致sleep差值为负——__schedstat_add_sleep_runtime内部对溢出做了saturationcstatic inline void__schedstat_add_sleep_runtime(struct task_struct *tsk, u64 delta){tsk-se.statistics.sum_sleep_runtime delta;}/* enqueue_entity的wakeup路径 */static voidenqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags){bool sleep_update (flags ENQUEUE_WAKEUP) !(flags ENQUEUE_MIGRATED);if (sleep_update) {__schedstat_set_sleep_start(se, 0); /* 清零触发重新记录 */u64 sl se-statistics.sleep_start;if (sl) {u64 delta rq_clock(rq_of(cfs_rq)) - sl;if (delta 0) {__schedstat_add_sleep_runtime(task_of(se), delta);}}}}nr_migrations与nr_forced_migrations的区别代表了迁移的不同原因。nr_migrations是se在每次set_task_cpu()包括wakeup affinity迁移和load balance pull时递增的计数器。而nr_forced_migrations只在load_balance中通过can_migrate_task()决定强制迁移时递增——即任务由于cpu过载或misfit被强行从源rq拉走而非自愿或cache-hot迁移。can_migrate_task()检查三要素task_cache_hot、busiest-nr_running和target-nr_running的负载差额。cstatic inline intcan_migrate_task(struct task_struct *p, struct lb_env *env){int tsk_cache_hot;if (throttled_lb_pair(task_group(p), env-src_cpu, env-dst_cpu))return 0;if (task_running(env-src_rq, p)) {schedstat_inc(p-se.statistics.nr_failed_migrations_running);return 0;}tsk_cache_hot task_hot(p, env);if (tsk_cache_hot 0)goto migrate;if (env-sd-nr_balance_failed env-sd-cache_nice_tries)goto migrate; /* 多次balance失败缓存热点忽略 */schedstat_inc(p-se.statistics.nr_failed_migrations_hot);return 0;migrate:schedstat_inc(p-se.statistics.nr_forced_migrations);return 1;}nr_wakeups及其子计数在try_to_wake_up()的wakeup路径中更新。ttwu_do_wakeup()检查是否跨CPU迁移若task-cpu ! smp_processor_id()则递增nr_wakeups_remote否则递增nr_wakeups_local。wake_affine()成功时递增nr_wakeups_affine。这些区分用于诊断wakeup性能问题的交叉引用——例如nr_wakeups_local比例过低时意味着waker和wakee频繁在不同CPU上暴露出cgroup拓扑配置不当。/proc/schedstat中cpu维度的行是每个rq的聚合数据。rq-avg_idle是一个指数加权移动平均值记录cpu在idle loop中平均等待下一个wakeup的时长。该值用于idle balance决策当avg_idle小于sysctl_sched_migration_cost时idle cpu不参与pull迁移——因为idle时间太短不值得迁移开销。avg_idle更新在cpu_idle_poll()或mwait_idle()退出时c/* kernel/sched/idle.c */static void cpuidle_idle_call(void){int next_state;if (!idle_should_enter_rt()) {rq-avg_idle calc_avg_idle(rq-avg_idle, delta);return;}next_state cpuidle_select(drv, dev);if (next_state 0) {entered_state cpuidle_enter(drv, dev, next_state);rq-avg_idle calc_avg_idle(rq-avg_idle,ktime_to_ns(dev-last_residency_ns));}}一个性能陷阱schedstat在NUMA系统中的wait_sum累加可能引入严重的cache ping-pong。wait_sum由__schedule()中dequeue_task()时记录wait_start而累加发生在wakeup路径即另一个CPU上。每次task在NUMA节点间迁移更新其se.statistics时两个CPU都必须获取se所在cache line的ownership。高频率wakeup如nginx worker的accept在CONFIG_SCHEDSTATSy下可造成数%的性能损失——这也是发行版内核通常关闭CONFIG_SCHEDSTATS的原因。