实际应用场景是将用户上传的文件依次保存到阿里云 OSS 与腾讯云 COS ,实现方式是在启用 Request.EnableRewind() 的情况下通过 Request.Body 读取流,并依次通过 2 个 StreamContent 分别上传到阿里云 OSS 与 腾讯云 COS ,在集成测试中可以正常上传(用的是 TestServer 启动站点),而部署到服务器上通过浏览器上传却出现了奇怪的问题 —— 第一个 StreamContent 上传后的文件大小总是0,而第二个 StreamContent 上传正常。上传文件大小为 0 时,对应的 Request.Body.Length 也为 0 。(注:如果不使用 Request.EnableRewind ,Request.Body 只能被读取一次)
而如果在第一个 StreamContent 读取 Request.Body 之前先通过 MemoryStream 进行一次流的 Copy 操作,就能正常读取。
using (var ms = new MemoryStream()) { await Request.Body.CopyToAsync(ms); }
好奇怪的问题!要牺牲第一个流,才能让后面的 StreamContent 从 Request.Body 中读到数据。 为什么会这样?
先从 Request.EnableRewind() 下手,通过它的实现源码知道了 EnableRewind 之后 Request.Body 被替换为 FileBufferingReadStream ,所以 StreamContent 实际读取的是 FileBufferingReadStream ,问题可能与 FileBufferingReadStream 有关。
public static HttpRequest EnableRewind(this HttpRequest request, int bufferThreshold = DefaultBufferThreshold, long? bufferLimit = null) { //.. var body = request.Body; if (!body.CanSeek) { var fileStream = new FileBufferingReadStream(body, bufferThreshold, bufferLimit, _getTempDirectory); request.Body = fileStream; request.HttpContext.Response.RegisterForDispose(fileStream); } return request; }
向前进,查看 FileBufferingReadStream 的实现源码。
在构造函数中 _buffer 的长度被设置为 0 :
if (memoryThreshold < _maxRentedBufferSize) { _rentedBuffer = bytePool.Rent(memoryThreshold); _buffer = new MemoryStream(_rentedBuffer); _buffer.SetLength(0); } else { _buffer = new MemoryStream(); }
FileBufferingReadStream 的长度实际就是 _buffer 的长度:
public override long Length { get { return _buffer.Length; } }
ReadAsync 读取流的代码(已移除不相关的代码):
public override async Task<int> ReadAsync(byte[] buffer, int offset, int count, CancellationToken cancellationToken) { ThrowIfDisposed(); if (_buffer.Position < _buffer.Length || _completelyBuffered) { // Just read from the buffer return await _buffer.ReadAsync(buffer, offset, (int)Math.Min(count, _buffer.Length - _buffer.Position), cancellationToken); } int read = await _inner.ReadAsync(buffer, offset, count, cancellationToken); //... if (read > 0) { await _buffer.WriteAsync(buffer, offset, read, cancellationToken); } //... return read; }
从 FileBufferingReadStream 的实现代码中没有发现问题。
只有 StreamContent 才会出现这个问题吗?写了个简单的 ASP.NET Core 程序验证了一下:
public async Task<IActionResult> Index() { Request.EnableRewind(); Request.Body.Seek(0, 0); Console.WriteLine("First Read Request.Body"); await Request.Body.CopyToAsync(Console.OpenStandardOutput()); Console.WriteLine(); Request.Body.Seek(0, 0); Console.WriteLine("Second Read Request.Body"); await Request.Body.CopyToAsync(Console.OpenStandardOutput()); Console.WriteLine(); Request.Body.Seek(0, 0); using (var sr = new StreamReader(Request.Body)) { return Ok(await sr.ReadToEndAsync()); } }
控制台输出流(System.ConsolePal+WindowsConsoleStream)没这个问题。
是 StreamContent 的问题吗?用下面的代码验证一下
public async Task<IActionResult> Index() { Request.EnableRewind(); var streamContent = new StreamContent(Request.Body); return Ok(await streamContent.ReadAsStringAsync()); }
奇怪了,StreamContnent 也没问题,只是剩下唯一的嫌疑对象 —— HttpClient 。
写测试代码进行验证,站点A的代码(监听于5000端口)
public async Task<IActionResult> Index() { Request.EnableRewind(); var streamContent = new StreamContent(Request.Body); var httpClient = new HttpClient(); var response = await httpClient.PostAsync("http://localhost:5002", streamContent); return Ok(await response.Content.ReadAsStringAsync()); }
站点B的代码(监听于5002端口)
public async Task<IActionResult> Index() { using (var ms = new MemoryStream()) { await Request.Body.CopyToAsync(ms); return Ok(ms.Length); } }
站点 A 启用 EnableRewind 并直接将 Request.Body 流 POST 到站点 B ,模拟实际应用场景。
测试得到的返回值是 0 ,问题重现了。
为了进一步验证是否是 HttpClient 的问题,将 HttpClient 改为 WebRequest 。
public async Task<IActionResult> Index() { Request.EnableRewind(); var request = WebRequest.CreateHttp("http://localhost:5002"); request.Method = "POST"; using (var requestStream = await request.GetRequestStreamAsync()) { await Request.Body.CopyToAsync(requestStream); } using (var response = await request.GetResponseAsync()) { using (var sr = new StreamReader(response.GetResponseStream())) { return Ok(await sr.ReadToEndAsync()); } } }
测试结果显示 WebRequest 没这个问题,果然与 HttpClient 有关。
向 HttpClient 的源代码进军。。。
从 HttpClient.SendAsync 到 HttpMessageInvoker.SendAsync 再到 HttpMessageHandler.SendAsync ,默认用的是 SocketsHttpHandler ,从 SocketsHttpHandler.SendAsync 到 HttpConnectionHandler.SendAsync 到 HttpConnectionPoolManager.SendAsync 。。。翻山越岭,长途跋涉,来到了 HttpConnection 的 SendAsyncCore 方法。
public async Task<HttpResponseMessage> SendAsyncCore(HttpRequestMessage request, CancellationToken cancellationToken) { //.. await SendRequestContentAsync(request, CreateRequestContentStream(request), cancellationToken).ConfigureAwait(false); //... }
原来是调用 SendRequestContentAsync 方法发送请求内容的
private async Task SendRequestContentAsync(HttpRequestMessage request, HttpContentWriteStream stream, CancellationToken cancellationToken) { // Now that we're sending content, prohibit retries on this connection. _canRetry = false; // Copy all of the data to the server. await request.Content.CopyToAsync(stream, _transportContext, cancellationToken).ConfigureAwait(false); // Finish the content; with a chunked upload, this includes writing the terminating chunk. await stream.FinishAsync().ConfigureAwait(false); // Flush any content that might still be buffered. await FlushAsync().ConfigureAwait(false); }
从 request.Content.CopyToAsync 追踪到 HttpConnection 的 WriteAsync 方法
private async Task WriteAsync(ReadOnlyMemory<byte> source) { int remaining = _writeBuffer.Length - _writeOffset; if (source.Length <= remaining) { // Fits in current write buffer. Just copy and return. WriteToBuffer(source); return; } if (_writeOffset != 0) { // Fit what we can in the current write buffer and flush it. WriteToBuffer(source.Slice(0, remaining)); source = source.Slice(remaining); await FlushAsync().ConfigureAwait(false); } if (source.Length >= _writeBuffer.Length) { // Large write. No sense buffering this. Write directly to stream. await WriteToStreamAsync(source).ConfigureAwait(false); } else { // Copy remainder into buffer WriteToBuffer(source); } }
看这部分代码实在毫无头绪,于是采用笨方法——手工打点在控制台显示信息,在 WriteToStreamAsync 进行打点
private ValueTask WriteToStreamAsync(ReadOnlyMemory<byte> source) { if (NetEventSource.IsEnabled) Trace($"Writing {source.Length} bytes."); Console.WriteLine($"{_stream} Writing {source.Length} bytes."); Console.WriteLine("source text: " + System.Text.Encoding.Default.GetString(source.ToArray())); return _stream.WriteAsync(source); }
编译 System.Net.Http 解决方案,将编译输出的 corefxinWindows_NT.AnyCPU.DebugSystem.Net.Http etcoreappSystem.Net.Http.dll 复制到 C:Program FilesdotnetsharedMicrosoft.NETCore.App2.1.2 文件夹中,然后就可以使用自己编译的 System.Net.Http.dll 运行 ASP.NET Core 程序。
运行测试站点(见之前的站点A与站点B的代码,站点 A 将 Request.Body 流中的内容通过 HttpClient POST 到站点 B ),站点 A 的控制台显示了下面的打点信息:
System.Net.Sockets.NetworkStream Writing 10 bytes. source text: POST / HTT System.Net.Sockets.NetworkStream Writing 10 bytes. source text: P/1.1 Con System.Net.Sockets.NetworkStream Writing 10 bytes. source text: tent-Lengt System.Net.Sockets.NetworkStream Writing 10 bytes. source text: h: 0 Host System.Net.Sockets.NetworkStream Writing 10 bytes. source text: : localhos _writeBuffer.Length: 10 _writeOffset: 10 remaining: 0 source.Length: 4 System.Net.Sockets.NetworkStream Writing 10 bytes. source text: t:5002 System.Net.Sockets.NetworkStream Writing 4 bytes. source text: test
将上面的 source text 内容连接起来,到下面的 http 请求内容:
POST / HTTP/1.1 Content-Length: 0 Host: localhost:5002 test
立马就发现了问题:Content-Length: 0 ,原来是 Content-Length 惹的祸,怎么会是 0 ?
继续打点。。。
找到了 "Content-Length: 0" 是 StreamContent 中的 TryComputeLength 方法引起的
protected internal override bool TryComputeLength(out long length) { if (_content.CanSeek) { length = _content.Length - _start;
return true; } else { length = 0; return false; } }
上面的代码中 _content.Length 的值为 0 (在博文的开头我们提到过 FileBufferingReadStream 在未被读取时 Length 的值为 0 ),于是 length 为 0 并返回 true ,所以生成了 "Content-Length: 0" 请求头。
如果当 length 为 0 时,让 TryComputeLength 返回 false ,这样就不会生成 "Content-Length: 0" 请求头,是不是可以解决问题呢?
protected internal override bool TryComputeLength(out long length) { if (_content.CanSeek) { length = _content.Length - _start; return length > 0; } else { length = 0; return false; } }
这样会产生下面的请求内容:
POST / HTTP/1.1 Transfer-Encoding: chunked Host: localhost:5002 4 test 0
这样的请求内容在示例程序服务端就可以正常读取到 Request.Body ,但是无法将文件上传到阿里云 OSS 与腾讯云 COS ,应该是 "Transfer-Encoding: chunked" 请求头的原因。
后来改为从 HttpAbstractions 下手,修改了 BufferingHelper.cs 与 FileBufferingReadStream.cs 的代码,终于解决了这个问题。
给 FileBufferingReadStream.cs 添加一个私有字段 _innerLength ,在 Request.EnableRewind 时通过构造函数将 Request.ContentLength 的值传给 _innerLength 。
var fileStream = new FileBufferingReadStream(body, request.ContentLength, bufferThreshold, bufferLimit, _getTempDirectory);
在 FileBufferingReadStream 的 Length 属性中,如果流还没被读取过,就返回 _innerLength 的值。
public override long Length { get { var useInnerLength = _innerLength.HasValue && _innerLength > 0 && !_completelyBuffered && _buffer.Position == 0; return useInnerLength ?_innerLength.Value : _buffer.Length; } }
修改 HttpAbstractions 的源代码后,需要将编译生成的下面5个文件都复制到 C:Program FilesdotnetsharedMicrosoft.NETCore.App2.1.2 文件夹中。
Microsoft.AspNetCore.Http.Abstractions.dll Microsoft.AspNetCore.Http.dll Microsoft.AspNetCore.Http.Features.dll Microsoft.Net.Http.Headers.dll Microsoft.AspNetCore.WebUtilities.dll
如果用的是 Linux ,需要复制到 /usr/share/dotnet/shared/Microsoft.AspNetCore.App/2.1.2/ 目录中。
后来发现基于 request.ContentLength 的解决方法不适用于 chunked requests ,将 CanSeek 属性改为下面的实现(原先是直接返回true)
public override bool CanSeek { get { return Length > 0; } }
这样第一读 Request.Body 正常,但之后继续读会出现下面的错误:
System.ObjectDisposedException: Cannot access a disposed object. Object name: 'FileBufferingReadStream'. at WebsiteA.FileBufferingReadStream.ThrowIfDisposed()