Re: Varnish - no CVE == bug regression

[Stefan Bühler <stbuehler () lighttpd net>]

On Thu, 03 Jul 2014 13:41:58 -0600
Kurt Seifried <kseifried () redhat com> wrote:

> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
>
>
> On 03/07/14 11:12 AM, Stefan Bühler wrote:
> > On Thu, 3 Jul 2014 08:15:06 +0000 Sven Kieske
> > <S.Kieske () mittwald de> wrote:
> >
> > > -----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
> > >
> > > I'd agree with this. And I don't get the argument from
> > > poul-henning kamp, what I understand is: "hey, we trust our
> > > backend server" well, but your backend server can make you crash,
> > > so you probably shouldn't trust it in the first place?
> > >
> > > you _never_ can trust input, so you have to validate it, either
> > > way, at least enough to not crash or perform malicious actions.
> > >
> > > Am 03.07.2014 09:48, schrieb Kurt Seifried:
> > > > So as I understand this: Varnish front end for web servers, the
> > > > web servers can trigger varnish to restart. Are the back end
> > > > servers supposed to be able to cause varnish to restart?
> > > >
> > > > I'm guessing not. Scenario: hosting env, or a website with a
> > > > vuln, whatever, you can now cause the varnish front ends to
> > > > restart constantly, effectively causing a permanent denial of
> > > > service.
> > > >
> > > > That sounds CVE worthy. Or am I missing something?
> >
> > you should never trust *untrusted* input. your root shell usually 
> > trusts the input it gets...
> >
> > so the valgrind developers decided that they consider the backend 
> > webservers trusted, at least regarding the capability to cause a
> > DoS.
> >
> > for the record - so does lighttpd (a backend can trigger OOM as
> > lighty reads (nearly) as fast as possible from a backend, as
> > backends often only handle one request at a time); we usually tell
> > people to use X-sendfile instead of sending ISOs through php.
>
> That also sounds like it needs a CVE then. You should not be able to
> trivially DoS stuff, especially OOM, things should protect themselves
> from OOM'ing especially if they accept user controlled input from the
> network.

And again "user controlled input"... a root shell also uses "user
controlled input".

> > just because you disagree with such decisions doesn't make it CVE 
> > worthy (missing or wrong documentation could).
>
> So to be clear your argument is that the http backends serviced by
> Varnish are supposed to be able to shut down Varnish, not by using an
> administrative channel/command but by executing a denial of service
> against Varnish? And that this is intended behaviour and thus not a
> security vulnerability?

If you can get it for free to prevent it, it is of course desirable to
prevent it; and this is what valgrind did in this case: they could fix
it, so they did. But I guess if it would have meant changing a lot of
core code they might have refused to fix it.

So the question is whether it is a priority for you; the original
author of lighty decided performance was more important, and even if we
now may have changed our mind about it, we don't want to break
everything just trying to fix it now. (Our new development version
includes a design to protect against such problems.)

> > in case you actually want to assign a CVE here, maybe we can get
> > one for the bad openssl default cipherstring too? because for that
> > it is really obvious that it is f*** wrong, but i think that none
> > was assigned because upstream didn't agree with it.

So you really want to tell me that it is intended to use openssl with
the crappy default cipher? - just to keep the analogy here, as you
somehow seem to have missed it.

regards,
Stefan