hc-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Christian Klauser (JIRA)" <j...@apache.org>
Subject [jira] [Created] (HTTPCORE-442) SSLIOSession + ChunkDecoder: MalformedChunkCodingException with last chunk left in inEncryptedBuffer
Date Fri, 20 Jan 2017 12:57:26 GMT
Christian Klauser created HTTPCORE-442:

             Summary: SSLIOSession + ChunkDecoder: MalformedChunkCodingException with last
chunk left in inEncryptedBuffer
                 Key: HTTPCORE-442
                 URL: https://issues.apache.org/jira/browse/HTTPCORE-442
             Project: HttpComponents HttpCore
          Issue Type: Bug
            Reporter: Christian Klauser
         Attachments: httpcore-nio-SSLIOSession-eof.path

I'm running into a rather esoteric problem when reading the response to an HTTP request using
{{httpcore-nio}} (used via CXF). On a high level we get a MalfromedChunkCodingException because
the ChunkDecoder doesn't receive the last chunk ( {{"0\r\n\r\n"}} is missing  ) We see the
following situation:

 * The plain text content fed into the SOAP/JAXB deserialization machinery cuts off mid-way
 * *because* the {{ChunkDecoder}} threw a {{MalformedChunkCodingException}}'
 * *because* {{this.buffer.readLine}} returned false
 * *because* (not 100% sure here) {{fillBufferFromChannel}} returned -1 and we are looking
for the next chunk size
 * *because* the last 106 bytes of the encrypted stream (see below) have never been decrypted
 * *because* the loop inside {{SSLIOSession#decryptData}} gets exited early if the {{SSLIOSession#endOfStream}}
flag is set _even if the input stream still contains encrypted data_

state of the SSLIOSession in `this.channel` when the exception occurs:<->[ACTIVE][r:r][ACTIVE][r][NOT_HANDSHAKING][EOF][][106][0][0][0]

Looking at past issues/commit messages, I understand why we have an early exit here (avoid
infinite loops on connections that broke in the middle of a TLS packet). I'm not sure if just
exiting is correct here. The documentation for the java [{{SSLEngine#unwrap}}|https://docs.oracle.com/javase/8/docs/api/javax/net/ssl/SSLEngine.html]
method says that the engine tries to decrypt "one complete TLS network packet". There is no
guarantee that it decrypts all remaining packets in the buffer.

In our particular case (106 still encrypted bytes), we have to run the {{unwrap}} method two
more times until we have the complete response from the server (including the final chunk
"{{0\r\n\r\n}}"). After the first additional call, there are still 53 encrypted bytes left.

I have attached a patch that solves this issue for us. The patch monitors whether the decryption
loop is making progress and keeps looping even if {{endOfStream}} is set. It aborts after
a number of iterations without clear progress. With this patch, we can run our entire integration
test suite without error. Without this patch, about 1 in SOAP 200 calls fails.

The "progress measurement" logic could perhaps be simplified (only measure plain bytes, not
also consumed encrypted bytes), but I'm not sure if that would still be correct.

I could also imagine that the actual bug is in the ChunkDecoder, that doesn't give the Channel
another chance at filling the buffer, but the SSLIOSession code looks more suspicious to me
(signalling EOF with 106 bytes left encrypted).

What do you think?

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: dev-unsubscribe@hc.apache.org
For additional commands, e-mail: dev-help@hc.apache.org

View raw message