【scheduler】调度器如何判断cpu_overutilized？

计算过程如下：

static bool cpu_overutilized(int cpu)  
{  
    return __cpu_overutilized(cpu, 0, NULL);  
}  
static bool __cpu_overutilized(int cpu, int delta, struct task_struct *p)  
{  
    struct rq *rq = cpu_rq(cpu);  
    //下面两个数值在update_cpu_capacity函数中update的。下面会解释
    unsigned long capacity_orig = capacity_orig_of(cpu);  
    unsigned long max_capacity = rq->rd->max_cpu_capacity.val;  

    if (capacity_orig == max_capacity) {  
        /* 
         * Waking task on idle big cpu or 
         * less than one task running on big cpu 
         */  
        if (p && ((p->state == TASK_WAKING &&  
                 idle_cpu(cpu)) ||  
                (p->state != TASK_WAKING &&  
                 rq->nr_running <= 1)))  
            return false;  

        if (!p && rq->nr_running <= 1)  
            return false;  
    }  

    return (capacity_of(cpu) * 1024) < ((cpu_util(cpu) + delta) * capacity_margin);  
}

具体获取的函数如下：

//初始化root_domain 成员变量 struct max_cpu_capacity
void init_max_cpu_capacity(struct max_cpu_capacity *mcc)  
{  
    raw_spin_lock_init(&mcc->lock);  
    mcc->val = 0;  
    mcc->cpu = -1;  
}  
/*update rq 成员变量cpu_capacity_orig和root_domain 里面max_cpu_capacity成
员变量数值*/
static void update_cpu_capacity(struct sched_domain *sd, int cpu)  
{  
    /*由于CONFIG_64BIT_ONLY_CPU=n，所以按照我们之前讲过的scale cpu capacity处理
    即相当于这个函数scale_cpu_capacity*/
    unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);  
    struct sched_group *sdg = sd->groups;  
    struct max_cpu_capacity *mcc;  
    unsigned long max_capacity;  
    int max_cap_cpu;  
    unsigned long flags;  
    /*即为此cpu的capacity，也是实际的capacity，dts里面规定的capacity数值*/
    cpu_rq(cpu)->cpu_capacity_orig = capacity;  

    mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;  

    raw_spin_lock_irqsave(&mcc->lock, flags);  
    max_capacity = mcc->val;  
    max_cap_cpu = mcc->cpu;  

    if ((max_capacity > capacity && max_cap_cpu == cpu) ||  
        (max_capacity < capacity)) {  
        /*update max_cpu_capacity结构体成员，获取整个topology cpu的max capacity
        存储在 rd结构体变量max_cpu_capacity中*/
        mcc->val = capacity;  
        mcc->cpu = cpu;  
#ifdef CONFIG_SCHED_DEBUG  
        raw_spin_unlock_irqrestore(&mcc->lock, flags);  
        printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",  
                cpu, capacity);  
        goto skip_unlock;  
#endif  
    }  
    raw_spin_unlock_irqrestore(&mcc->lock, flags);  

skip_unlock: __attribute__ ((unused));  
    capacity *= scale_rt_capacity(cpu);  
    capacity >>= SCHED_CAPACITY_SHIFT;  

    if (!capacity)  
        capacity = 1;  
    /*update sched domain group中sched group capacity成员变量和rq cpu_capacity
     成员变量*/
    cpu_rq(cpu)->cpu_capacity = capacity;  
    sdg->sgc->capacity = capacity;  
    sdg->sgc->max_capacity = capacity;  
    sdg->sgc->min_capacity = capacity;  
}

从上面可以知道，如果capacity_orig == max_capacity成立。则接下来判断如干个条件是否满足cpu已经过载？

task 真实存在 & task状态是TASK_WAKING，正在被waking过程 & cpu是处在idle状态
task状态不是TASK_WAKING & rq的nr_running数量 <= 1
没有task存在 & rq rn_running数量 <= 1

只要满足上面三个条件之一，cpu就直接返回false，cpu不存过载情况。
判断上面的三个条件之后，就开始根据cpu负载情况来真实的计算cpu是否over utilization了。

(capacity_of(cpu) * 1024) < ((cpu_util(cpu) + delta) * capacity_margin);

对于cpu的util的计算在下面文章中讲解了：https://blog.csdn.net/wukongmingjing/article/details/81739394
cpu_util(cpu)数值在[0,1024]这范围之内
capacity_margin是一个常量：1138,（约等于1024*1.1）
delta=0
capacity_of(cpu)数值是一个变量，<= capacity_orig，区分big/little架构不同，比如在我现在的手机上little core的capacity_orig=785，big core的capacity_orig=1024
最后计算他们的数值就可以判断是否过载了。打印的信息如下：

cpu=0,capacity_orig=782,max_capacity=1024,capacity_of=735,util=782,overutil=1  
cpu=1,capacity_orig=782,max_capacity=1024,capacity_of=745,util=720,overutil=1  
cpu=2,capacity_orig=782,max_capacity=1024,capacity_of=730,util=537,overutil=0  
cpu=3,capacity_orig=782,max_capacity=1024,capacity_of=737,util=10,overutil=0  
cpu=4,capacity_orig=1024,max_capacity=1024,capacity_of=1022,util=0,overutil=0  
cpu=5,capacity_orig=1024,max_capacity=1024,capacity_of=1023,util=0,overutil=0  
cpu=6,capacity_orig=1024,max_capacity=1024,capacity_of=1022,util=0,overutil=0  
cpu=7,capacity_orig=1024,max_capacity=1024,capacity_of=1024,util=0,overutil=0

计算原理比较简单，但是存在一个遗留的action：
capacity_of(cpu)是会变化的，目前还没有找是在哪里修改的，但是知道这个涉及到domain/group的balance。

【scheduler】调度器如何判断cpu_overutilized？

猜你喜欢