A Box API without the Sand

[Allan Jude]

On 2021-01-07 14:42, Mina Galić wrote:
> Hello FreeBSD developers
> 
> I’m trying to find an (existing) API on FreeBSD which allows me to box in a process and all of its descendants. That box should be easy to identify, and allow process management, accounting and resource management.
> 
> Unfortunately all boxes Iʼm finding come pre-filled with Sand that’s impossible to get rid of, or lack some of the other requirements.
> 
> An analogue for such an API can be found in Solaris Contracts and Linux Cgroups.
> 
> Cgroups are extensively used in many init implementations on Linux to contain, identify and control services.
> Personally, i used contracts on Solaris for that very purpose, in an in-house deployment software, long before systemd was hip.
> 
> I would like to do something like that in init/rc on FreeBSD. But i can’t seem to find an API that fits the bill.
> 
> The normal POSIX process management facilities, like process groups and process sessions are too easy to escape. But given their primary use is job control, that’s a feature, not a bug. Either way, they can be ruled out.
> 
> The closest we have are jail(2) and Capsicum(4), as well as the functions supplementing the latter, such as pdfork(2) & co.
> 
> Jails are really good at containing and identifying process groups, and for managing their resources.
> However, jailing every single daemon on a system, while desirable by some, would make many daemons useless. sshd and monitoring software are my prime examples, but there are probably many more.
> 
> Capsicum, again, seems more concerned with the Sand than the box: i cannot seem to find anything that would help identify a process group. Except for pdfork(2)'s procdesc(4).
> 
> Fortuitously, these can be used independently of Capsicum, so that means we have an identity and control, but i can’t find any information on whether all descentants will remain within the same procdesc.
> 
> That’s as far as my research has gotten me, so before i dig myself any deeper, i’d like to ask for your help.
> 
> And, most importantly, if no such API exists yet, where to start in adding it, or modifying an existing API to suit these requirements?
> 
> Best regards,
> 
> Mina
> https://alpha.pkgbase.live/
> _______________________________________________
> freebsd-hackers at freebsd.org mailing list
> https://lists.freebsd.org/mailman/listinfo/freebsd-hackers
> To unsubscribe, send any mail to "freebsd-hackers-unsubscribe at freebsd.org"
> 

A few different ideas in this area have been discussed before. I
remember a conversation with John Baldwin at one of the FreeBSD Storage
Summits, about the idea of a 'null' jail.

Basically a jail with no restrictions, one that uses / instead of a
chroot, inherits the network and shared memory namespaces, etc.

The main advantage to this abstraction is that it builds on a lot of
existing functionality. It gives you a top level identifier (jail id),
and can already have a lot of policies applied to it (cpuset, RCTL
resource limits, etc)

You can mostly already control what devices from /dev are visible.

I think what we would need to do is define what else needs to be
different from a jail. Like do we need the ability to allow processes
inside the container to see the processes from the host?

There may be a few places that will be more problematic. ZFS
automatically enforces its restrictions if your jail ID is not 0.


-- 
Allan Jude