[systemd-devel] SELinux support takes up ~15MB of memory?

[Lennart Poettering]

On Fri, 07.01.11 10:18, Daniel J Walsh (dwalsh at redhat.com) wrote:

> On 01/07/2011 09:44 AM, Lennart Poettering wrote:
> > On Fri, 07.01.11 09:40, Daniel J Walsh (dwalsh at redhat.com) wrote:
> > 
> >>> Hmm, can we start with an empty loaded policy and then load additional
> >>> parts of it as we go? i.e. if we encounter a socket /foo/bar/waldo we
> >>> ask libselinux to load /foo/bar, and so on? Most likely 90% of the
> >>> sockets will be in the same dir anyway (/var/run), so after the first
> >>> socket everything we need should be loaded most of the time. However,
> >>> since sockets can be configured dynamically to any place we might need
> >>> to load policy for other areas, too. Hence if we could load hte policy
> >>> bit by bit we should get relatively nice behaviour and only load a
> >>> minimal subset of the policy into memory.
> >>>
> >>> Lennart
> >>>
> >> I think the library functions are there to do this, but you would have
> >> to do the management of the paths.  libselinux I believe does not have
> >> the capability to add a path after the initial load but you could have a
> >> link list of paths connected to blobs of regexes.
> > 
> > So, instead of loading one single policy blob we would basically load a
> > number of independent policy blobs, but always only parts of the real
> > thing? I guess that is quite doable, though I do wonder how the prefix
> > finding algorithm should best look like...
> > 
> > Lennart
> > 
> Yes and you have a speed versus size thing.   Do you want to compile
> /var once or /var/lock and /var/run or much worse /var/run/app1 and
> /var/run/app2 ...
> 
> I would not go above two directories.
> 
> If you really want size versus speed you could free the blobs once you
> got to a steady state, if you can figure out you are in a steady state.
>  And then reload them if you need too.

Hmm, so, I wonder how far we could get by just hardcoding that we preload
/var/run, /var/lock and /dev (the latter for /dev/initctl). That should
cover 100% of the sockets/fifos of a usual install. And if that doesn't
suffice we load the real deal on-demand.

So, loading parts of the policy apparently takes up less memory. How
exactly is the speed of loading the policy influenced by this? I.e. is
loading a part of the policy as CPU/IO intensive as loading the full thing?
Or even more CPU intensive due to a lot of compares? Or cheaper?

Lennart

-- 
Lennart Poettering - Red Hat, Inc.