Server Name Indication support broken when reusing a (dead) session
RFC3546 recommend that the client includes the server_name in each client hello message when possible.
The ruby openssl client implementation doesn't send a server_name when it has a session to resume. Normally the server can resume the session and doesn't need the server_name. However if the server for what ever reason does not recognize the session ID, it is unable to determine what certificate to serve. This can cause intermittent failures.
This issue surfaced due to broken session_id based persistence in a VMWare load balancer, causing every second connect to fail due to a invalid certificate. (The second connection was load balanced to another server that didn't know the session). Since this might also occur if the server forgets the session more quickly than the client, I think this can be considered a bug.
I have tried to figure out how to patch this, but my C knowledge is not sufficient to figure this out.
Steps to reproduce¶
When monitoring the following code with wireshark, it shows that when reopening a https connection with a session, there is no server_name part included in the message.
I used the following few lines of code to test:
uri = URI('https://www.example.com/') req = Net::HTTP::Get.new uri.request_uri con = Net::HTTP.new uri.host, uri.port con.use_ssl = true con.start con.finish con.start # Produces a certificate error if the session is lost by the server