Skip to content

3.3.0

Latest

Choose a tag to compare

@benoitc benoitc released this 20 Nov 22:20
03cbe6c

Changed

  • quic:set_owner_sync/2 on a connection that is closing now returns {error, closed}, the same as once the connection is gone, instead of {error, {invalid_state, draining}}. The owner still receives {quic, Conn, {closed, Reason}} with the reason.

Fixed

  • quic_h3:connect/3 could crash the calling process when the server closed the connection right after the handshake, for example a server that requires a client certificate refusing a client without one. Depending on timing the caller got {error, _}, crashed, or waited until connect_timeout. It now always returns {error, Reason} with the reason the server closed with.
  • With a host name that resolves to several addresses, quic:connect/4 could return {error, connect_timeout} after the full timeout when the server completed the handshake and closed the connection at once. It now returns the reason the server closed with: the attempt whose handshake completed has won the race (RFC 8305 Section 5), and the other attempts are no longer waited for.
  • An HTTP/3 server no longer logs a crash for each client it refuses right after the handshake.
  • An HTTP/3 server declines a connection that the client closed before the server had set it up, instead of logging a connection handler error.
  • A distribution server no longer crashes when the peer closes the connection as it is being set up.
  • A server that refused a client during the handshake, for example one that requires a client certificate and got none, could give the client the wrong reason: after refusing, it went on reading what the client had sent and closed a second time with an "unexpected message" error, which the client sometimes saw instead of the real one. The server also logged a warning for each refused client. A refusal now ends the handshake (RFC 9846 Section 6.2), and a connection only ever sends its first close (RFC 9000 Section 10.2.1).
  • With qpack_never_indexed, the headers of push_promise and push_stream events lost the never-indexed mark, so code forwarding a pushed request could not keep it on the next hop (RFC 9204 Section 7.1.3). They now carry it like the other events. A promise repeated with the same fields but different marks is still accepted as the same promise (RFC 9114 Section 7.2.5).