Re: [RFC] lsm: namespace hooks

[ebiederm-aS9lmoZGLiVWk0Htik3J/w@public.gmane.org (Eric W. Biederman)]

Lukasz Pawelczyk <l.pawelczyk-Sze3O3UU22JBDgjK7y7TUQ@public.gmane.org> writes:

> On czw, 2014-11-27 at 09:42 -0600, Eric W. Biederman wrote:
>> Lukasz Pawelczyk <l.pawelczyk-Sze3O3UU22JBDgjK7y7TUQ@public.gmane.org> writes:
>> 
>> > On czw, 2014-11-27 at 16:01 +0100, Richard Weinberger wrote:
>> >> Am 27.11.2014 um 15:44 schrieb Lukasz Pawelczyk:
>> >> > True, the last one is 0x80000000. I did not notice that. Thanks for
>> >> > pointing out.
>> >> 
>> >> Isn't this CLONE_IO?
>> >
>> > Yes, I was merely noticing out loud that it's the last bit of 32bit.
>> >
>> > After close look though the 0x00001000 appears to be unused
>> >
>> >> > Any suggestion on what can be done here? New syscal with flags2?
>> >> 
>> >> I'm not sure. But a new syscall would be a candidate.
>> 
>> We are probably going to need to go a couple rounds with this but at
>> first approximation I think this functionality needs to be tied to the
>> user namespace.  This functionality already looks half tied to it.
>> 
>> When mounting filesystems with user namespaces priveleges matures a
>> little more you should be able to use unmapped labels.  In the near term
>> we are looking at filesystems such as tmpfs, fuse and posibly extN.
>
> I presume you are referring to the Smack namespace readme where I
> mentioned mounts with specifying smack labels in the mount options, not
> to the quote above?
>
> I was referring the to the check here that has been changed to
> smack_ns_privileged() using ns_capable():
> http://lxr.free-electrons.com/source/security/smack/smack_lsm.c#L462
>
> And you can't use an unmapped Smack label inside the namespace, this
> would be completely against its idea.
>
> Anyway, at this point I'm more interested in the LSM namespace. I'll be
> doing an RFC for Smack namespace later.
>
> Unless I misunderstood your mail.

I had two points.
a) Tie the label mapping to the user namespace, then we don't need any
   new namespaces.

   Is there a reason not to tie the label mapping to the user namespace?

   Needing to modify every userspace that create containers to know
   about every different lsm looks like a maintenance difficulty I would
   prefer to avoid.

b) For filesystems that don't need uid mapping (say ext2 mounted with
   user namespace permissions) we shouldn't need LSM mapping either.

Eric