[RFC PATCH 0/2] Containerise nproc count

[Eric W. Biederman]

Nikolay Borisov <kernel at kyup.com> writes:

> From: Nikolay Borisov <n.borisov at siteground.com>
>
> Hello, 
>
> This is an initial try to have nproc count apply per-userns, 
> rather than per the global user struct. The implementation is 
> really simple - a hashtable holding uid->nproc mapping for each
> id inside the respective namespace. In its current form I have also
> left the debugging code so that people who want to have a play with 
> it can easily see what's happening. 
>
> Now, this is only an RFC and I'd like to gather your thoughts about
> the semantics. Currently as it stands I have tested the patchset by 
> invoking multiple LXC containers, with identical uid mappings and 
> users with the same uid inside the containers and it was working
> correctly. 
>
> There is an issue however, when using the unshare syscall and then doing
> the mappings e.g. using "unshare -r" util from util-linux the initial process 
> (the one which have done the unsharing) is accounted to the overflowuid but 
> then again when exiting from the resulting shell the UID for user 0 is being 
> freed which causes the BUG_ON in nsuser_nproc_dec to trigger. My initial idea 
> for fixing this was to add code which upon writing to /proc/[pid]/uid_map  
> would map all current processes from overflowuid to the 'ns->uid_map.extent[0].first'. 
> This was working correctly but it was breaking the use case of lxc, since lxc is 
> changing the uids after creating the uid_mapping (maybe this is a deficiency in the 
> unshare util implementation?)
>
> Another thing that needs improving is the locking occuring on the nsuser_nproc_hash, 
> since in its current coarse-grained form it is serialisign process/thread creation on 
> a per-usernamespace basis. 
>
> I'm happy to discuss any concerns and improvements that people might have 
> regarding this patchset. 

So. no.

Changing rlimit nproc this way breaks per user process accounting.
Effectively this allows any user to escape their NPROC limit by creating
a new user namespace.  Which means that to even consider anything like
this we need hierarchical limits.

I am not particularly convinced that reusing uids between containers is
all that smart.  Certainly it is bad to assume that there is no leakage
and containers can never interact.  So something like this also needs a
description of why this new set of semantics is a good direction to go
in.

Eric