Andy Lutomirski wrote:
> > So I bet the google chrome people are not at all interested in
> > "running random binaries", and might want execve() very much for
> > "running some specific binaries that we ship with or install from the
> > browser".
> 
> I want to be able to run "magic_seccomp_sandbox gv somefile.ps".  And
> I want that to still work even if my distro gets fancy and uses
> selinux policy to prevent gv from accessing the network (which would
> not be an unreasonable thing to do).  The chromium project already has
> a little seccomp wrapper that can (sort of) do this.

Chrome also comes with NaCL.  They *are* "running random binaries",
but they use code analysis to check the code first.  Maybe because
seccomp wasn't good enough ;-)

> > So I really think that the *only* valid model is the "fail the execve
> > on any changes", not the "mnt-nosuid" approach of trying to run things
> > with the wrong permissions and get perhaps odd results. And I think it
> > works even - and perhaps *especially* - in models like selinux or
> > apparmor that do have a lot of "implicit MAC knowledge" on specific
> > binaries.
> 
> Is the current exec_no_trans check enough for you?  With my patch,
> selinux can already block the execve if it wants.  (The policy is the
> same as it would be if a program explicitly asked to run the new
> executable with an unchanged security context.)  I'd be happy to fail
> the exec in AppArmor, and then maybe AppArmor will change its mind
> if/when users get annoyed :)

Does SELinux block if userspace does exec entirely in userspace using
mmap() and not execve()?  If not, why is execve() allowed to be different?

-- Jamie