CONNECT method (Proof-of-Concept)

[zuiderkwast]

I've been experimenting with CONNECT to write a proxy server: https://github.com/zuiderkwast/deputy. (Is the name Deputy already taken or reserved for something else in the Cowboy world?) I've tested it only manually by configuring it as a proxy running on localhost in Firefox.

Feel free to pick selected stuff from this PR, now or later. I don't really need it for any serious project.

Changes to Cowboy:

• Request parsing and validation for CONNECT update in cowboy_http and cowboy_http2 and restrictions forbidding CONNECT removed. Path is set to "/" rather than omitted, to make the request pass through cowboy_router and cowboy_handler unmodified. (It's a hack.)
• A cowboy_tunnel modelled after cowboy_websocket is added.
• Adaptations to cowboy_http for taking over the connection process without sending 101 Switching Protocols.

The flow, in short:

1. Handler:init is called. For CONNECT requests, it may return {cowboy_tunnel, ...}.
2. cowboy_tunnel:upgrade is called. It calls cowboy_http with a new command {takeover, ...} which has the same effect as {switch_protocol, ...}, except 101 is not sent. Takover is called in the connection process.
3. cowboy_tunnel calls Handler:tunnel_init in the connection process. Here, a TCP connection (tunnel) is supposed to be created.
4. If successful, cowboy_tunnel sends 200 OK to the client and the tunnel is open.
5. On info (such as incoming TCP tunnel data from the peer) Handler:tunnel_info is called. On data from the client, Handler:tunnel_handle is called.

TODO:

• Decide how the interface should actually look like
• Test cases
• Remove commented-out code

[essen]

Just answering the comments you made, didn't look at the code yet:

• Yes deputy is fair game.
• The path hack is definitely not good. I was thinking a CONNECT handler should be configured in the protocol options, question is where exactly (in env probably, in which case cowboy_router can just take it from there or abort if none is configured).
• Perhaps switch_protocol can be replaced by takeover entirely.
• I'm not convinced a special handler is needed otherwise. cowboy_loop should be enough once initialization is done. And it can use "read_body auto" to receive packets from the socket as they come (or better cowboy_stream_h problems in grpc streaming #1430 (comment) but the code doesn't currently exist).