Hi Shannon,
Bummer :-( Please don't give up!
Sorry if this is a really stupid question (my knowledge of proxy-servers is
minimal) but what issues are involved here that have not been solved with a
product such as Orbited? http://www.orbited.org/
Look, I've never used it but, if I understand it correctly, it sits behind
the proxy-server and traps the Socket traffic and forwards it onto the
designated run-o'-the-mill TCP/IP server. Could a similar mechanism(s) not
work here? Is it that we're dealing with runtime discovery of public
proxy-servers outside any firewalls? Something else?
Cheers Richard Maher
PS. Much prefer the websocket.read(bytes) and websocket.write(bytes) well
done! Please don't forget websocket.peek(bytes).
- Original Message -
From: Shannon [EMAIL PROTECTED]
To: WHAT working group whatwg@lists.whatwg.org
Sent: Tuesday, October 14, 2008 7:22 AM
Subject: [whatwg] WebSocket and proxies
In the process of testing my WebSocket proposal I discovered the CONNECT
method has a major restriction. Most proxies disable CONNECT to anything
but port 443.
The following is from Squid and the Blowfish:
--
It is very important that you stop CONNECT type requests to non-SSL
ports. The CONNECT method allows data transfer in any direction at any
time, regardless of the transport protocol used. As a consequence, a
malicious user could telnet(1) to a (very) badly configured proxy, enter
something like:
... snip example ...
and end up connected to the remote server, as if the connection was
originated by the proxy.
---
I verified that Squid and all public proxies I tried disable CONNECT by
default to non-SSL ports. It's unlikely many internet hosts will have
443 available for WebSockets if they also run a webserver. It could be
done with virtual IPs or dedicated hosts but this imposes complex
requirements and costs over alternatives like CGI.
The availability and capabilities of the OPTIONS and GET protocols also
varied from proxy to proxy. The IETF draft related to TLS
(http://tools.ietf.org/html/draft-ietf-tls-http-upgrade-05) has this to
say:
---
3.2 Mandatory Upgrade
If an unsecured response would be unacceptable, a client MUST send
an OPTIONS request first to complete the switch to TLS/1.0 (if
possible).
OPTIONS * HTTP/1.1
Host: example.bank.com
Upgrade: TLS/1.0
Connection: Upgrade
---
So according to this draft spec OPTIONS is the only way to do a
*mandatory* upgrade of our connection. Once again this failed in testing
---
= OPTIONS * HTTP/1.1
= Proxy-Connection: keep-alive
= Connection: Upgrade
= Upgrade: WebSocket/1.0
= Host: warriorhut.org:8000
=
= HTTP/1.0 400 Bad Request
= Server: squid/3.0.STABLE8
Other proxies gave different errors or simply returned nothing. The
problem may be related to the Upgrade and Connection headers rather than
OPTIONS, since I had similar issues using Connection: Upgrade with GET.
I had the most success using GET without a Connection: Upgrade header.
It seems that the proxy thinks the header is directed at it so it does
not pass it on to the remote host. In many cases it will abort the
connection. Using the Upgrade: header without Connection allows the
Upgrade header through to the actual websocket service.
It seems to me that whatever we try in many cases the connection will be
silently dropped by the proxy and the reasons will be unclear due to the
lack of error handling. There seems to be a wide variation in proxy
behaviour for uncommon operations. I suppose proxy developers could fix
these issues but whether a significant rollout could be achieved before
HTML5 is released is questionable.
Given that an asynchronous connection cannot be cached the only reasons
remaining for going through a proxy are anonymity and firewall
traversal. Automatically bypassing the users proxy configuration to
solve the issues above has the potential to break both of these. It
would be a significant breach of trust for a UA to bypass the users
proxy and some networks only allow connections via a proxy (for security
and monitoring).
It seems that we're stuck between a rock and hard place here. In light
of this I reiterate my earlier suggestion that the time could be better
spent providing guidelines for communication via an asynchronous CGI
interface. This would allow reuse of existing port 80 and 443 web
services which would resolve the cross-domain issues (the CGI can relay
the actual service via a backend connection) and most of the proxy
issues above (since proxy GET and CONNECT are more reliable on these
ports).
Shannon
