On Wed, 2008-03-26 at 09:46 +0100, Daniel J Walsh wrote: > -----BEGIN PGP SIGNED MESSAGE----- > Hash: SHA1 > > Stephen Smalley wrote: > > On Tue, 2008-03-25 at 07:04 -0400, Joshua Brindle wrote: > >> Stephen Smalley wrote: > >>> On Mon, 2008-03-24 at 16:27 -0400, Joshua Brindle wrote: > >>> > >>>> Stephen Smalley wrote: > >>>> > >>>>> On Mon, 2008-03-24 at 13:40 -0400, Joshua Brindle wrote: > >>>>> > >>>>> > >>>>>> This implements user_transition in the toolchain. It should help on > >>>>>> distro's like Ubuntu that can't use run_init due to the user not knowing > >>>>>> the root password. It also seems like a more eloquent way to handle > >>>>>> service restarts than assigning system_r to user accounts and having the > >>>>>> daemons run as someuser:system_r:foo_t. > >>>>>> > >>>>>> > >>>>> Yes, that's something that has been wanted in Fedora for quite some > >>>>> time. > >>>>> > >>>>> The real issue with run_init isn't the re-authentication stage, as that > >>>>> can always be disabled via pam config (and was just a weak form of > >>>>> confirming user intent, not an authorization mechanism), but rather the > >>>>> difficulty in transparently interposing it into all situations where > >>>>> services get started/re-started. Only Gentoo seemed to have a good > >>>>> story there. > >>>>> > >>>>> > >>>>> > >>>>>> This has some issues in policy due to users not always being known in > >>>>>> the policy (eg., semanage users). I hope Chris or Dan will be able to > >>>>>> give some suggestions there. > >>>>>> > >>>>>> > >>>>> I'm not sure why anyone needs to add users to policy via semanage users > >>>>> given the base set of generic users and the ability to map Linux users > >>>>> to them via seusers aka semanage login. > >>>>> > >>>>> > >>>>> > >>>>>> The kernel patch (forthcoming after this is accepted) so far only > >>>>>> implements the transition on process transitions. Later on I plan on > >>>>>> doing a patch to expand role_transition to object classes (this is a > >>>>>> change needed for policy rbac support to work). I suspect I'll do the > >>>>>> same for user at that time. The question here is, do we think its worth > >>>>>> it to do fine grained transitions like we did for range_trans? (I don't). > >>>>>> > >>>>>> > >>>>> Offhand, I can't see a use for per-class user transitions, if that is > >>>>> what you mean. > >>>>> > >>>>> I don't think per-class role transitions is really the fundamental > >>>>> obstacle to enabling use of roles on objects - more thought is required > >>>>> there. What will be fun there is role/type and user/range validation, > >>>>> which presently gets to ignore everything that has object_r. > >>>>> > >>>>> > >>>> Ah, another thing. While going through the policyrep implementation the > >>>> question of object_r came up. My thought is to start adding object_r > >>>> magic into the toolchain (adding all types, etc) and eventually purge > >>>> object_r from the kernel. at least one magic instance of object_r will > >>>> be removed by object role_transitions, the others are really short > >>>> circuits in the security server that can be removed after sufficient > >>>> support is in the toolchain. What are your thoughts on that (for future > >>>> reference)? > >>>> > >>> Well, the interesting question is what should the default role be in the > >>> new context in security_compute_sid, if not object_r. Even aside from > >>> the support for per-class role transitions. User defaults to the source > >>> context, type defaults to the related object context, and MLS range > >>> defaults to the low level of the source context. Role could be the > >>> subject's role or the related object's role. > >>> > >>> > >> Good question. My original assumption was that we'd use the related > >> object role. That would require that home directories be correctly > >> labeled with the role of the user. If we start using the source role > >> then things will quickly change from object_r to system_r, so maybe the > >> policy should do that anyway. Chris, any opinions on this? > > > > Yes, related object role would likely cause the least breakage. It > > would preserve the existing default for existing filesystems (as they > > already have object_r in the directory contexts), while allowing us to > > switch over to the user's role for home directories upon a relabel or > > new filesystem. Source role might create more conflicts, as we enforce > > the role/type relationship for contexts and there might be a mismatch > > between the creating process role and the parent directory type. > > > I am not sure where this is going, but I believe that separation based > on role in the home directory is a mistake. It assumes that the home > directory will always be used by the same user with the same role. And > will not work when you have a network file system that supports labels. > > In Red Hat I can login to people.redhat.com people.fedoraproject.com > which I should use the guest_r. While logging into my laptop I would be > unconfined_t and on test machines I might get staff_r or user_r. All of > them would use the same homedirectory. So how would this work in this > environment? That's an interesting question, and I know that in some operating systems, roles are actually separate user accounts altogether. However, the mechanism being described here doesn't prevent you from continuing to use a generic role (object_r or otherwise) on all files; it just allows for people who want to apply distinct roles on files to do so. -- Stephen Smalley National Security Agency -- This message was distributed to subscribers of the selinux mailing list. If you no longer wish to subscribe, send mail to majordomo@xxxxxxxxxxxxx with the words "unsubscribe selinux" without quotes as the message.
