这种方法是否比在 Task.Run 中触发 stream.Read() 更好？

Question

编辑：我最终没有按照下面的 with Stephen Cleary. If you are interested in how I did it differently, take a look at my 使用这种方法。

我正在寻找一种从 NetworkStream 异步读取超时的方法。当然，问题是无法取消 NetworkStream 上的 ReadAsync()，因为它只是忽略了 CancellationToken。我读到一个建议在令牌取消时关闭流的答案，但在我的情况下，这不是一个选项，因为 Tcp 连接必须保持打开状态。所以我想出了以下代码，但我想知道这是否比

更好

Task.Run(() => stream.Read(buffer, offset, count)

并且只是阻塞了一个线程。

public static class TcpStreamExtension
{
    public static async Task<int> ReadAsyncWithTimeout(this NetworkStream stream, byte[] buffer, int offset, int count)
    {
        CancellationTokenSource cts = new CancellationTokenSource();
        bool keepTrying = true;
        Timer timer = new Timer(stream.ReadTimeout);
        timer.Elapsed += new ElapsedEventHandler((sender, args) => stopTrying(sender, args, cts, out keepTrying));
        timer.Start();

        try
        {
            if (stream.CanRead)
            {
                while (true)
                {
                    if (stream.DataAvailable)
                    {
                        return await stream.ReadAsync(buffer, offset, count, cts.Token).ConfigureAwait(false);
                    }

                    if (keepTrying)
                    {
                        await Task.Delay(300, cts.Token).ConfigureAwait(false);
                    }
                    else
                    {
                        cts.Dispose();
                        timer.Dispose();
                        throw new IOException();
                    }
                }
            }
        }
        catch (TaskCanceledException tce)
        {
            // do nothing
        }
        finally
        {
            cts.Dispose();
            timer.Dispose();
        }
        if (stream.DataAvailable)
        {
            return await stream.ReadAsync(buffer, offset, count).ConfigureAwait(false);
        }

        throw new IOException();
    }

    private static void stopTrying(object sender, ElapsedEventArgs args, CancellationTokenSource cts, out bool keepTrying)
    {
        keepTrying = false;
        cts.Cancel();
    }

}

该应用程序可能必须能够与数千个端点进行通信，我希望以一种不会阻塞一堆线程的方式创建它，因为它所做的大部分工作都是 IO。另外，超时的情况应该是

Answer 1

对于类似的用例，我使用了 Task.Delay() 任务来超时。看起来像这样：

public static async Task<int> ReadAsync(
        NetworkStream stream, byte[] buffer, int offset, int count, int timeoutMillis)
{
        if (timeoutMillis < 0) throw new ArgumentException(nameof(timeoutMillis));
        else if (timeoutMillis == 0)
        {
            // No timeout
            return await stream.ReadAsync(buffer, offset, count);
        }

        var cts = new CancellationTokenSource();
        var readTask = stream.ReadAsync(buffer, offset, count, cts.Token);
        var timerTask = Task.Delay(timeoutMillis, cts.Token);

        var finishedTask = await Task.WhenAny(readTask, timerTask);
        var hasTimeout = ReferenceEquals(timerTask, finishedTask);
        // Cancel the timer which might be still running
        cts.Cancel();
        cts.Dispose();

        if (hasTimeout) throw new TimeoutException();
        // No timeout occured
        return readTask.Result;
}

Answer 2

首先，您尝试做的事情存在根本性缺陷。您应该一直从开放的 TCP/IP 流读取 - 一旦一次读取获得一些数据，就将其传递并开始下一次读取。

因此，我的第一个建议是 不需要 首先需要可取消读取。相反，请始终阅读。同样，使用 DataAvailable 是一种代码味道。

更多解释...

对于不可取消的代码，"enforcing" 没有很好的取消方法。关闭 TCP/IP 套接字是最简单和最干净的方法。您现有的解决方案将不起作用，因为 ReadAsync 忽略了 CancellationToken。所以它并不比只使用 CancellationToken 没有计时器更好。如果 ReadAsync 忽略 CancellationToken，您唯一真正的选择是关闭套接字。任何其他解决方案都可能导致 "lost data" - 从套接字读取但被丢弃的数据。

Answer 3

根据 with Stephen Cleary 和他的建议，我重新审视了如何实现它，我采用了一种每次读取都不会超时但只要 TcpClient 就保持打开状态的方法打开然后控制来自不同代码的超时。我使用 Task.Run(() => beginReading());，这样当然会使用池中的线程，但我认为这没关系，因为大多数时候该线程会触及 await，因此是空闲的

这是我的实现：

private readonly Queue<byte> bigBuffer = new Queue<byte>();
private readonly SemaphoreSlim _signal = new SemaphoreSlim(0, 1);

// This is called in a Task.Run() 
private async Task beginReading()
{
    byte[] buffer = new byte[1024];

    using (_shutdownToken.Register(() => m_TcpStream.Close()))
    {
        while (!_shutdownToken.IsCancellationRequested)
        {
            try
            {
                int bytesReceived = 0;
                if (m_TcpStream.CanRead)
                {
                    bytesReceived = await m_TcpStream.ReadAsync(buffer, 0, buffer.Length).ConfigureAwait(false);
                }
                else
                {
                    // in case the stream is not working, wait a little bit
                    await Task.Delay(3000, _shutdownToken);
                }

                if (bytesReceived > 0)
                {
                    for (int i = 0; i < bytesReceived; i++)
                    {
                        bigBuffer.Enqueue(buffer[i]);
                    }

                    _signal.Release();
                    Array.Clear(buffer, 0, buffer.Length);

                }
            }
            catch (Exception e)
            {
                LoggingService.Log(e);
            }
        }
    }
}

private async Task<int> ReadAsyncWithTimeout(byte[] buffer, int offset, int count)
{
    int bytesToBeRead = 0;

    if (!m_TcpClient.Connected)
    {
        throw new ObjectDisposedException("Socket is not connected");
    }

    if (bigBuffer.Count > 0)
    {
        bytesToBeRead = bigBuffer.Count < count ? bigBuffer.Count : count;

        for (int i = offset; i < bytesToBeRead; i++)
        {
            buffer[i] = bigBuffer.Dequeue();
        }

        // Clear up the semaphore in case of a race condition where the writer just wrote and then this came in and read it without waiting
        if (_signal.CurrentCount > 0)
            await _signal.WaitAsync(BabelfishConst.TCPIP_READ_TIME_OUT_IN_MS, _shutdownToken).ConfigureAwait(false);

        return bytesToBeRead;
    }

    // In case there is nothing in the Q, wait up to timeout to get data from the writer
    await _signal.WaitAsync(15000, _shutdownToken).ConfigureAwait(false);

    // read again in case the semaphore was signaled by an Enqueue
    if (bigBuffer.Count > 0)
    {
        bytesToBeRead = bigBuffer.Count < count ? bigBuffer.Count : count;

        for (int i = offset; i < bytesToBeRead; i++)
        {
            buffer[i] = bigBuffer.Dequeue();
        }


        return bytesToBeRead;
    }

    // This is because the synchronous NetworkStream Read() method throws this exception when it times out
    throw new IOException();
}

这种方法是否比在 Task.Run 中触发 stream.Read() 更好？

Is this approach better than just firing stream.Read() in a Task.Run?

c#

networkstream

async-await