为什么 .NET 内部 Hashtable 中有一个 Thread.Sleep(1)

发布于11月16日

最近我在读《实现》.NET Hashtable遇到了一段我不懂的代码.部分代码是:

int num3 = 0;
int num4;
do
{
   num4 = this.version;
   bucket = bucketArray[index];
   if (++num3 % 8 == 0)
     Thread.Sleep(1);
}
while (this.isWriterInProgress || num4 != this.version);

整个代码在System.Collections.Hashtable的public virtual object this[object key]以内(mscorlib版本=4.0.0.0).

问题是:

为什么会有Thread.Sleep(1)个呢？

推荐答案

Sleep(1)是Windows中一种有文档记录的方法，它可以让处理器运行，并允许其他线程运行.您可以在带有注释的参考源中找到此代码:

   // Our memory model guarantee if we pick up the change in bucket from another processor,
   // we will see the 'isWriterProgress' flag to be true or 'version' is changed in the reader.
   //
   int spinCount = 0;
   do {
       // this is violate read, following memory accesses can not be moved ahead of it.
       currentversion = version;
       b = lbuckets[bucketNumber];

       // The contention between reader and writer shouldn't happen frequently.
       // But just in case this will burn CPU, yield the control of CPU if we spinned a few times.
       // 8 is just a random number I pick.
       if( (++spinCount) % 8 == 0 ) {
           Thread.Sleep(1);   // 1 means we are yeilding control to all threads, including low-priority ones.
       }
   } while ( isWriterInProgress || (currentversion != version) );

isWriterInProgress变量是易失性布尔值.笔者在英语中遇到了一些麻烦，"违规阅读"是"易失性阅读".基本的 idea 是尽量避免让步，线程上下文切换是非常昂贵的，希望作者能尽快完成.如果这不起作用，那么显式让步以避免消耗CPU.这可能是今天用Spinlock写的，但是Hashtable已经很旧了.关于内存模型的假设也是如此.