Issue with BinaryReader.ReadChars()

I've run into what I believe is an issue with the BinaryReader.ReadChars() method. When I wrap a BinaryReader around a raw socket NetworkStream occasionally I get a stream corruption where the stream being read gets out of sync. The stream in question contains messages in a binary serialisation protocol.

I've tracked this down to the following

It only happens when reading a unicode string (encoded using the Encoding.BigEndian)
It only happens when the string in question is split across two tcp packets (confirmed using wireshark)

I think what is happening is the following (in the context of the example below)

BinaryReader.ReadChars() is called asking it to read 3 characters (string lengths are encoded before the string itself)
First loop internally requests a read of 6 bytes (3 remaining characters * 2 bytes/char) off the network stream
Network stream only has 3 bytes available
3 bytes read into local buffer
Buffer handed to Decoder
Decoder decodes 1 char, and keeps the other byte in it's own internal buffer
Second loop internally requests a read of 4 bytes! (2 remaining characters * 2 bytes/char)
Network stream has all 4 bytes available
4 bytes read into local buffer
Buffer handed to Decoder
Decoder decodes 2 char, and keeps the remaining 4th bytes internally
String decode is complete

Serialisation code attempts to unmarshal the next item and croaks because of stream corruption.

char[] buffer = new char[3];
int charIndex = 0;

Decoder decoder = Encoding.BigEndianUnicode.GetDecoder();

// pretend 3 of the 6 bytes arrives in one packet
byte[] b1 = new byte[] { 0, 83, 0 };
int charsRead = decoder.GetChars(b1, 0, 3, buffer, charIndex);
charIndex += charsRead;

// pretend the remaining 3 bytes plus a final byte, for something unrelated,
// arrive next
byte[] b2 = new byte[] { 71, 0, 114, 3 };
charsRead = decoder.GetChars(b2, 0, 4, buffer, charIndex);
charIndex += charsRead;

I think the root is a bug in the .NET code which uses charsRemaining * bytes/char each loop to calculate the remaining bytes required. Because of the extra byte hidden in the Decoder this calculation can be off by one causing an extra byte to be consumed off the input stream.

Here's the .NET framework code in question

    while (charsRemaining>0) { 
        // We really want to know what the minimum number of bytes per char 
        // is for our encoding.  Otherwise for UnicodeEncoding we'd have to
        // do ~1+log(n) reads to read n characters. 
        numBytes = charsRemaining;
        if (m_2BytesPerChar)
            numBytes <<= 1;

        numBytes = m_stream.Read(m_charBytes, 0, numBytes);
        if (numBytes==0) { 
            return (count - charsRemaining); 
        } 
        charsRead = m_decoder.GetChars(m_charBytes, 0, numBytes, buffer, index);

        charsRemaining -= charsRead;
        index+=charsRead;
    }

I'm not entirely sure if this is a bug or just a misuse of the API. To work round this issue I'm just calculating the bytes required myself, reading them, and then running the byte[] through the relevant Encoding.GetString(). However this wouldn't work for something like UTF-8.

Be interested to hear people's thoughts on this and whether I'm doing something wrong or not. And maybe it will save the next person a few hours/days of tedious debugging.

EDIT: posted to connect Connect tracking item

Solution

I have reproduced the problem you mentioned with BinaryReader.ReadChars.

Although the developer always needs to account for lookahead when composing things like streams and decoders, this seems like a fairly significant bug in BinaryReader because that class is intended for reading data structures composed of various types of data. In this case, I agree that ReadChars should have been more conservative in what it read to avoid losing that byte.

There is nothing wrong with your workaround of using the Decoder directly, after all that is what ReadChars does behind the scenes.

Unicode is a simple case. If you think about an arbitrary encoding, there really is no general purpose way to ensure that the correct number of bytes are consumed when you pass in a character count instead of a byte count (think about varying length characters and cases involving malformed input). For this reason, avoiding BinaryReader.ReadChars in favor of reading the specific number of bytes provides a more robust, general solution.

I would suggest that you bring this to Microsoft's attention via http://connect.microsoft.com/visualstudio.