起因是发现一个同事编写的程序运行两个月左右,占用了服务器20G左右的内存。用WinDbg查看发现存在大量的Async Pinned
Handles,而它们的gcroot都来自于SocketAsyncEventArgs。下面是场景的简易模拟代码(为了说明问题添加了手动GC):
for (var i = 0; i < 1000; ++i)
{
var endPoint = new IPEndPoint(IPAddress.Parse(host), port);
var socket = new Socket(endPoint.AddressFamily, SocketType.Stream, ProtocolType.Tcp);
socket.ReceiveBufferSize = bufferSize;
socket.SendBufferSize = bufferSize;var iocp = new SocketAsyncEventArgs();
iocp.Completed += new EventHandler<SocketAsyncEventArgs>(OnIoSocketCompleted);
iocp.SetBuffer(new Byte[bufferSize], 0, bufferSize);
iocp.AcceptSocket = socket;try
{
socket.Connect(endPoint);
Console.WriteLine(i);
}
catch (SocketException ex)
{
Console.WriteLine(ex.Message);
}socket.Close();
//iocp.Dispose();
}GC.Collect();
GC.WaitForPendingFinalizers();
GC.Collect();
SocketAsyncEventArgs的SetBuffer函数内部会pin住buffer数据(查阅SetBufferInternal实现),它的析构函数会调用FreeOverlapped函数释放资源。而问题就是在于FreeOverlapped函数的实现和传递参数。来看一下Dispose、~SocketAsyncEventArgs的实现:
public void Dispose()
{
this.m_DisposeCalled = true;
if (Interlocked.CompareExchange(ref this.m_Operating, 2, 0) != 0)
{
return;
}
this.FreeOverlapped(false);
GC.SuppressFinalize(this);
}
~SocketAsyncEventArgs()
{
this.FreeOverlapped(true);
}
两者都会调用FreeOverlapped函数释放,但是一个传递了false、一个传递了true参数。再来看FreeOverlapped实现:
private void FreeOverlapped(bool checkForShutdown)
{
if (!checkForShutdown || !NclUtilities.HasShutdownStarted)
{
if (this.m_PtrNativeOverlapped != null && !this.m_PtrNativeOverlapped.IsInvalid)
{
this.m_PtrNativeOverlapped.Dispose();
this.m_PtrNativeOverlapped = null;
this.m_Overlapped = null;
this.m_PinState = SocketAsyncEventArgs.PinState.None;
this.m_PinnedAcceptBuffer = null;
this.m_PinnedSingleBuffer = null;
this.m_PinnedSingleBufferOffset = 0;
this.m_PinnedSingleBufferCount = 0;
}
if (this.m_SocketAddressGCHandle.IsAllocated)
{
this.m_SocketAddressGCHandle.Free();
}
if (this.m_WSAMessageBufferGCHandle.IsAllocated)
{
this.m_WSAMessageBufferGCHandle.Free();
}
if (this.m_WSARecvMsgWSABufferArrayGCHandle.IsAllocated)
{
this.m_WSARecvMsgWSABufferArrayGCHandle.Free();
}
if (this.m_ControlBufferGCHandle.IsAllocated)
{
this.m_ControlBufferGCHandle.Free();
}
}
}
用WinDbg查看gchandles统计:
Statistics:
MT Count TotalSize Class Name
000007fb1bdb6ae8 1 24 System.Object
000007fb1bdb6b80 1 48 System.SharedStatics
000007fb1bdb7f58 1 64 System.Security.PermissionSet
000007fb1bdb6a10 1 160 System.ExecutionEngineException
000007fb1bdb6998 1 160 System.StackOverflowException
000007fb1bdb6920 1 160 System.OutOfMemoryException
000007fb1bdb6738 1 160 System.Exception
000007fb1bdb7b90 2 192 System.Threading.Thread
000007fb1bdb6c40 1 216 System.AppDomain
000007fb1bdb6a88 2 320 System.Threading.ThreadAbortException
000007fb1ae19770 6 336 System.Net.Logging+NclTraceSource
000007fb1bdbf958 3 480 System.RuntimeType+RuntimeTypeCache
000007fb1ae19900 6 480 System.Diagnostics.SourceSwitch
000007fb1bd64458 6 34520 System.Object[]
000007fb1b6f5e40 1000 112000 System.Threading.OverlappedData
Total 1033 objectsHandles:
Strong Handles: 12
Pinned Handles: 5
Async Pinned Handles: 1000
Weak Long Handles: 3
Weak Short Handles: 13
确实存在1000个Async Pinned
Handles,也就是说无法通过SocketAsyncEventArgs的析构函数释放SafeNativeOverlapped相关的资源。将示例代码的"iocp.Dispose();"注释去除,并重新执行再次查看gchandles:
Statistics:
MT Count TotalSize Class Name
000007fb1bdb6ae8 1 24 System.Object
000007fb1bdb6b80 1 48 System.SharedStatics
000007fb1bdb7f58 1 64 System.Security.PermissionSet
000007fb1bdb6a10 1 160 System.ExecutionEngineException
000007fb1bdb6998 1 160 System.StackOverflowException
000007fb1bdb6920 1 160 System.OutOfMemoryException
000007fb1bdb6738 1 160 System.Exception
000007fb1bdb7b90 2 192 System.Threading.Thread
000007fb1bdb6c40 1 216 System.AppDomain
000007fb1bdb6a88 2 320 System.Threading.ThreadAbortException
000007fb1ae19770 6 336 System.Net.Logging+NclTraceSource
000007fb1bdbf958 3 480 System.RuntimeType+RuntimeTypeCache
000007fb1ae19900 6 480 System.Diagnostics.SourceSwitch
000007fb1bd64458 6 34520 System.Object[]
Total 33 objectsHandles:
Strong Handles: 12
Pinned Handles: 5
Weak Long Handles: 3
Weak Short Handles: 13
1000个Async Pinned
Handles已不存在,但SocketAsyncEventArgs的析构函数从实现来看应该也可以完成释放,为什么失败了?
用!bpmd命令添加NclUtilities.HasShutdownStarted、 m_PtrNativeOverlapped.Dispose()的断点。
!bpmd System.dll System.Net.NclUtilities.get_HasShutdownStarted
!bpmd mscorlib.dll System.Runtime.InteropServices.SafeHandle.Dispose
SocketAsyncEventArgs的释放问题