There are a couple of features missing from the MSRP relay implementation on Kamailio. These are required for any large-scale/real-world deployment.
RFC 4976 section 6.4.1 paragraph 3 states:
If the Failure-Report header is "yes" or "partial", and if there is a problem processing the SEND request or if an error response is received for that SEND request, then the relay MUST respond with an appropriate error response in a REPORT back to the original source of the message.
This means that when a SEND request that contains a Failure-Report: header with value of "yes" or "partial" is relayed by Kamailio we need to maintain some state so that when the TCP send fails (immediately or timeout) or we receive a failure response from the next hop, Kamailio can generate and send a REPORT back to the sender.
An MSRP relay has the option of splitting large MSRP messages (including those that are already chunked) into smaller chunks. The reason is that there is no maximum size for an MSRP chunk, so it is entirely possible for a client to attempt to send a large file in a single, very-large, MSRP message. In this case it becomes important to pull the message from the receive buffer in small pieces (instead of waiting for a complete message to arrive and over-flowing the receive buffer) and relaying it in smaller chunks.
Also, the flow control for MSRP is based around TCP windowing, so if you want to limit the rate that someone is sending you MSRP messages (for example during the transfer of a large file) you do so by pulling from the TCP buffer at a lower rate. One way of avoiding overloading an MSRP relay is to have fairly small TCP receive buffers thereby limiting the rate at which clients and relays can send data. However, if the TCP buffers are small the chances of a client sending a message or chunk that is too large (and needs to be chunked by the relay) increases further.
Of course, to do this the MSRP module will need to be able to remember the (From|To)-Path, transaction ID, and so on from the start of the MSRP message so that it can use these (along with calculated Byte-Range) in each additional chunk.
I was thinking a new parameter indicating the maximum MSRP send chunk size is required. If an MSRP message in the receive buffer reaches that number then it has to be pulled from the buffer in parts and relayed as multiple chunks. This maximum send chunk size should be set to be much smaller than the TCP receive buffer size.