Squid-3.2 mythbusting NAT

One of the more frequently mentioned “problems” with Squid-3.2 since its release is a change in how it handles NAT failures.

The Myths

“Squid used to work when I NAT traffic to it from my router.”

“Squid used to work with one port when I configure the browser and NAT traffic.”

The Reality

No. Squid up until 3.1 would silence the NAT errors and treat the router as if it were the client browser. Any differences between the Host header and the requested URL were also completely ignored. All of this would be invisible to you the administrator, hidden at debug levels not normally shown. With several of the problems being completely unrecorded as well.

This last part seems to be behind a fair bit of skepticism about whether the problem we solved really exists. Nothing was showing up in the logs before and even a full packet trace from Squid or a gateway server would not reveal how JavaScript hijacks a client browser to pass internal documents to an external attacker.

What Changed?

Squid-3.2 finally added Host header validation to protect against the nasty behaviour and a few 0-day attacks using CVE-2009-0801 security vulnerability. Which has been a known flaw with NAT interception by HTTP proxies for at least 12 years now. This meant two things had to change:

* ignoring NAT errors had to stop. We can’t rely on validation results if the TCP details are already known to be corrupted.

* traffic directly from a browser to a NAT intercept port on Squid had to stop. There is no way to separate the NAT lookup result of error and unknown entry. A pity really, but that is what we are forced to work with.

During testing we uncovered a lot of really quite nasty behaviour by various clients, and indeed some public services setup to take advantage of the NAT bug as if it were a feature. On the whole a lot of client software sends a Host header with values strangely unrelated to what is being requested of the proxy. All of this had to go, or did it? for most of the year or so of testing it was banned completely and a HTTP error message returned for any such garbage. But in the end it was clear that we had to let it through somehow or we would be pitting Squid against the biggest players on the Internet … yeah.

(Almost-) Final Results

Squid-3.2 and later will reject traffic where NAT lookups fail on an intercept port. This includes NAT done on external devices and browsers directly sending proxy requests to the Squid intercept port. At the very least accurate reporting about the traffic and what it contains is the critical factor. If we let this traffic into Squid it would seriously compromise the validation results and also allow malicious clients to attack internal network resources with a large measure of anonymity. Neither of which is acceptible for a proxy trusted with controlling user traffic.

The best practice guideline for some years has been that NAT MUST be done on the Squid device. Squid-3.2 are now enforcing it as a basic requirement. If you are one of the network administrators running into this requirement change, please investigate the Policy Routing functionality of the device you originally had doing the NAT.

Squid-3.2 and later will accept invalid Host headers and produce a response to the client. But they will not cache the untrustworthy transaction results. They will contact the same server which the client TCP connection would have originally reached had Squid not been there (visible as ORIGINAL_DST in the logs). Notice that this is now deserving of the name “transparent proxy” a lot more than previous intercepting Squid. Even so NAT interception is still only half-transparent with the server being able to easily identify the proxies existence.

The loss of caching is intended to be a temporary solution until we can properly implement per-client caching of objects. As the saying goes there is nothing so permanent as the temporary – Squid still contains this workaround several series later. Support with this work is very welcome.

The situation when upstream peers are involved is also quite dangerous. For now we have had to permit Squid to pass the traffic to peers, which opens all multi-hop systems of proxies to the same vulnerabilities that were previously possible wth a single intercepting proxy. The solution is also going to require a substantial amount more work and be some years away. For the meanwhile it is a good idea to avoid passing intercepted traffic to cache_peer when possible.

More information on the problems, log entries, when they occur and what can be done about each is detailed in the Squid wiki Host header forgery page.

Happy Eyeballs

Geoff Huston wrote up a very interesting analysis of the RFC 6555 “Happy Eyeballs” features being added to web browsers recently.

As these features reach the mainstream stable browser releases and more people being using them Squid in the role of intercepting proxy are starting to face the same issues mentioned for CGN gateways. For all the same reasons. Whether or not you are operating an existing interception proxy or installing a new one this is one major new feature of the modern web which needs to be taken into account when provisioning the network and Squid socket/FD resources.

Squid operating as forward proxy do not face this issue as each browser only opens a limited number of connections to the proxy. Although Firefox implementation of the  “Happy Eyeballs” algorithm appears to have been instrumental in uncovering a certain major bug in Squids new connection handling recently.

A Squid Implementation

For those interested, Squid-3.2 does implement by default a variation of the “Happy Eyeballs” algorithm.

DNS lookups are performed in parallel now, as opposed to serial as they were in 3.1. As a result the maximum DNS lookup time is reduced from the sum of A and AAAA response times, to the maximum of both.

TCP connection attempts are still run in serial, but where older versions of Squid interspersed a DNS lookup with each set of TCP attempts the new 3.2 code identifies all the possible destinations first and tries each individual address until a working connection is found. Retries under the new version are also now limited per-address where in the older versions each retry meant a full DNS result set of addresses was re-tried.

As a result dns_timout is separated from connect_timeout which is now fully controlling only one individual TCP connection handshake.