Re: drop SECURITY_FILE_CAPABILITIES?

["Serge E. Hallyn" <serue@us.ibm.com>]

Quoting Steve Grubb (sgrubb@redhat.com):
> On Tuesday 10 November 2009 10:53:49 am Serge E. Hallyn wrote:
> > > > Does anyone know of cases where CONFIG_SECURITY_FILE_CAPABILITIES=n
> > > > is still perceived as useful?
> > >
> > > 
> > > As a library writer, I wished that the kernel behavior was either
> > > consistent,  or there is an API that I can use to find out what model we
> > > are operating under. The biggest issue is that for a distribution we know
> > > the assumptions the distribution should be running under. But end users
> > > are free to build their own kernel that has it disabled. This has already
> > > lead to dbus not working at all.
> > > 
> > > I also take issue with probing the capability version number returning
> > > EINVAL  when its the only way to find out what the preferred version is.
> > 
> > In 2007/2008, KaiGai had floated patches to export capability info
> > over securityfs.  If it was something library writers and distros
> > wanted, we could resurrect those patches - and tack on some info
> > about cap-related kernel config.
> 
> Unfortunately, I would have to support the kernels from 2.6.26->2.6.32 which 
> presumably don't have this facility. So, I'm kind of stuck. I think in a 
> previous discussion you mentioned that I could call getcap or 
> prctl(PR_CAPBSET_READ) and check for CAP_SETPCAP. I think I have to go that 
> direction for backwards compatibility.

Yes, I'm afraid so - unless /proc/config.gz happened to be available.
I suppose looking through /proc/1/status might be more reliable
actually, in case you were running in an already-partially-restricted
process tree.

> But back to detecting the capability version number...if I pass 0 as the 
> version in the header, why can't the kernel just say oh you want the preferred 
> version number, stuff it in the header, and return the syscall with success and 
> not EINVAL?

This is something I believe Andrew has advocated in the past, but I
forget why.  Andrew?

> Another irritation...if I want to clear the bounding set, I have to make a for 
> loop and call prctl 34 times (once for each bit). I'd rather see a v4 
> capability that takes the bounding set as part of the same syscall. Maybe all 
> 3 of these could be fixed in the same OS release so that changing to v4 also 
> signifies the other behavior changes.

I worry a bit about people confusing the bounding set as something
more  flexible than it is, and/or getting lazy and using the bounding
set instead of fI|pI .vs. fP, but am not solidly against this.

Anyway, maybe we should get on thsi sooner rather than later...
Are there any other deficiencies people see  in the current
API?

-serge