Re: Solaris 10 Containers / Zones Security Flaw

[Jonathan Katz <jonathan.katz () gmail com>]

Jim, et al,

I hate to sound like a shill, especially because I'm not working at
Sun, but starting with Solaris 9, Solaris ships with Sun Resource
Manager bundled with the OS. Resources can be allocated in "pools"
which can be a set number of processors, memory, etc. Rules for pools
are far more abstract than "bind this zone or process to this
processor" (see pbind(1m) for the old way of doing things.) The rules
can be more along the lines of "pool A cannot take up more than 50% of
the CPU resources on the system, but pool A has a lower priority than
pool B and can pool A give up 50% of its available CPU if pool B
requires the resources." There are a lot of knobs that can be turned.
See pooladm(1m) and poolcfg(1m) for more info. zonecfg(1m) shows how a
zone can be linked to a pool, and the zonecfg manpage references the
pooladm and poolcfg pages.

This goes back to a fundamental question of computer security; if
there is a new option (in this case zones) is to be enabled the
ramifications of enabling that option must be weighed and
investigated. Documentation needs to be read and researched and any
production deployment requires some kind of functional testing... by
running that bash "bomb" in a zone, Jim has done some great functional
testing.

Depending on the day of the week it could be considered a troll or a
good philosophical question of whether or not Sun needs to ship zones
with some kind of resource limits enabled, but at the end of the day,
one thing holds true; zones aren't created "out of the box." A user or
admin needs to manually create a zone, and just as they can create a
zone without any resource limits they can also create a .rhosts
containing "+ +"

Just my $.02.

-- 
-Jon
Jonathan Katz -- J. Random BOFH

On 1 Apr 2005 07:38:04 -0000, jim allan <intehnet () gmail com> wrote:
> all,
>
> thought i'd share something from a bit of home research. It's a bit trivial, and the "hole" (so to speak) is easily 
> patched up, but it defies the claims of Sun in regards to Solaris 10 security.
>
> Solaris 10 contains a feature called containers, or zones, which are kind of like a "VMware" "session" embedded 
> inside the kernel. These seperate zones have their own ip address (virtual interface off a physical interface, eg; 
> bge0:1), their own /proc /dev /etc and file system, entirely their own operating system, and unable to affect the 
> master, or other zones.
> Sun suggest zones are good for running separate internet facing applications, for example, a sol10 box runs a 
> webserver in one zone, and an internal DNS on another zone. If the internet facing web server gets compromised, and 
> an attacker drops them selves to root on that zone, whilst they are physically connected to the box, they cannot go 
> outside that zone, often, they'll have to be wise to solaris 10 to even know they are in a zone, and it's not it's 
> own box.
> They can compromise and wreck havoc in that zone, without any other zones, or the master zone, from which all zones 
> are controlled, being affected. There is NO way to drop out of a slave zone into a master zone (yet...) unless you 
> logged into the master zone first. I hope that makes sense.. read suns webpage if you wanna know more. 
> http://www.sun.com/software/solaris/
>
> Here's where it gets interesting. By default, there is no limit on virtual memory or cpu time for each zone. By doing 
> a standard bash fork bomb, I was able to take down an entire Solaris 10 box, from within a non-master zone. All zones 
> were locked up, including the master zone.
>
> It's nothing ground breaking, but I just found it interesting/poor that Sun didn't place, by default, CPU or memory 
> limits on zones, which are meant to be, essentially, master of their own domain, and unable to affect other zones. 
> One would have to go out of their way to configure CPU limits.
>
> See bash fork bomb below.
>
> #!/usr/local/bin/bash
> :(){ :|:& };:
>
> ps; if you wish to patch this, either set a ulimit to the amount of virtual memory a user can have, or explore the 
> set up of zones, i've been told there is a way to configure a limit to cpu time, although i haven't been able to find 
> any relevant documentation after a brief search.
> I'm considering writing a patch using solaris 10's dtrace D language to capture a process that is forking X amount in 
> Y time, given some miracle that I have some free time once in a while :)
>
> look forward to your replies
>
> jim allan
>
> intehnet at g mail dot com