Re: [capabilities] Allow normal inheritance for a configurable set of capabilities

["Serge E. Hallyn" <serge@hallyn.com>]

Quoting Andrew G. Morgan (morgan@kernel.org):
> I'm not generally in favor of this. Mostly because this seems to be a
> mini-root kind of inheritance that propagates privilege to binaries
> that aren't prepared for privilege.

Earlier in this thread, Casey said:

| One of the holes in the 1003.1e spec is what to do with a program file
| that does not have a capability set attached to it. The two options are
| drop all capabilities and leave the capabilities alone. The latter gives
| you what you're asking for. The former is arguably safer.

and

| It's what we did in Trusted Irix. It made life much easier.

I'm going to need to clear my head a bit before I try to compare that to
the root cause of the sendmail capabilities bug.

>  I don't really buy the mmap code
> concern because the model as it stands says that you trust the binary
> (and all of the various ways it was programmed to execute code) with
> specific privileges. If the binary mmap's some code (PAM modules come
> to mind) then that is part of what it was programmed to/allowed to do.

That's not really the point...  The point is that yes, a mini-root is
exactly what is being asked for :)  I'm not saying I expect an adversary
to do the mmap+jump, but that currently it is a, and the only, way to
do unprivileged userid with retaining some privileges to run legacy
programs.

> That being said, if you really really want this kind of thing, then
> make it a single secure bit (with another lock on/off bit) which, when
> set, makes: fI default to X.
> 
>    pP' = (X & fP) | (pI & fI)
> 
> That way the per-process bounding set still works as advertised and
> you don't need to worry about the existing semantics being violated.

Maybe that is the way to go...

-serge