Here we are again - VSV00003 in perspective¶
So it probably helps if you first re-read what I wrote two years ago about our first first major security hole.
Statistically, it is incredibly hard to derive any information from a single binary datapoint.
If some event happens after X years, and that is all you know, there is no way to meaningfully tell if that was a once per millenium event arriving embarrassingly early or a biannual event arriving fashionably late.
That allows us to cabin the expectations for the discovery of major security problems in Varnish Cache to “probably about 3 per decade” 2.
Given that one of my goals with Varnish Cache was to see how well systems-programming in C can be done in the FOSS world 3 and even though we are doing a lot better than most of the FOSS world, that is a bit of a disappointment 4.
The nature of the beast¶
VSV00003 is a buffer overflow, a kind of bug which could have manifested itself in many different ways, but here it runs directly into the maw of an assert statement before it can do any harm, so it is “merely” a Denial-Of-Service vulnerability.
A DoS is of course bad enough, but not nearly as bad as a remote code execution or information disclosure vulnerability would have been.
That, again, validates our strategy of littering our source code with asserts, about one in ten source lines contain an assert, and even more so that we leaving the asserts in the production code.
I really wish more FOSS projects would pick up this practice.
How did we find it¶
This is a bit embarrasing for me.
For ages I have been muttering about wanting to “fuzz”5 Varnish, to see what would happen, but between all the many other items on the TODO list, it never really bubbled to the top.
A new employee at Varnish-Software needed a way to get to know the source code, so he did, and struck this nugget of gold far too fast.
Hat-tip to Alf-André Walla.
Dealing with it¶
Martin Grydeland from Varnish Software has been the Senior Wrangler of this security issue, while I deliberatly have taken a hands-off stance, a decision I have no reason to regret.
Thanks a lot Martin!
As I explained at length in context of VSV00001, we really like to be able to offer a VCL-based mitigation, so that people who for one reason or another cannot update right away, still can protect themselves.
Initially we did not think that would even be possible, but tell that to a German Engineer…
Nils Goroll from UPLEX didn’t quite say “Halten Sie Mein Bier…”, but he did produce a VCL workaround right away, once again using the inline-C capability, to frob things which are normally “No User Serviceable Parts Behind This Door”.
Are we barking up the wrong tree ?¶
An event like this is a good chance to “recalculate the route” so to speak, and the first question we need to answer is if we are barking up the wrong tree?
Does it matter in the real world, that Varnish does not spit out a handful of CVE’s per year ?
Would the significant amount of time we spend on trying to prevent that be better used to extend Varnish ?
There is no doubt that part of Varnish Cache’s success is that it is largely “fire & forget”.
Every so often I get an email from “the new guy” who just found a Varnish instance which has been running for years, unbeknownst to everybody still in the company.
There are still Varnish 2.x and 3.x out there, running serious workloads without making a fuzz about it.
But is that actually a good thing ?
Dan Geer thinks not, he has argued that all software should have a firm expiry date, to prevent cyberspace ending as a “Cybersecurity SuperFund Site”.
So far our two big security issues have both been DoS vulnerabilities, and Varnish recovers as soon as the attack ends, but what if the next one is a data-disclosure issue ?
When Varnish users are not used to patch their Varnish instance, would they even notice the security advisory, or would they obliviously keep running the vulnerable code for years on end ?
Of course, updating a software package has never been easier, in a well-run installation it should be a non-event which happens automatically.
And in a world where August 2019 saw a grand total of 2004 CVEs, how much should we (still) cater to people who “fire & forget” ?
And finally we must ask ourselves if all the effort we spend on code quality is worth it, if we still face a major security issue as often as every other year ?
We will be discussing these and many other issues at our next VDD.
User input would be very welcome.
I’m not counting VSV00002, it only affected a very small fraction of our users.
Sandia has a really fascinating introduction to this obscure corner of statistics: Sensitivity in Risk Analyses with Uncertain Numbers
As distinct from for instance Aerospace or Automotive organizations who must set aside serious resources for Quality Assurance, meet legal requirements.
A disappointment not in any way reduced by the fact that this is a bug of my own creation.
“Fuzzing” is a testing method to where you send random garbage into your program, to see what happens. In practice it is a lot more involved like that if you want it to be an efficient process. John Regehr writes a lot about it on his blog “Embedded in Academia” and I fully agree with most of what he writes.