Inconsistencies with trusted.SGI_ACL_{FILE,DEFAULT}

[Dave Chinner]

On Tue, Oct 27, 2015 at 10:39:51PM +0100, Andreas Gruenbacher wrote:
> On Tue, Oct 27, 2015 at 9:18 PM, Dave Chinner <david at fromorbit.com> wrote:
> > Further, user namespaces are irrelevant here - you can't run
> > xfsdump/restore outside the init_ns.  xfsdump requires access to the
> > handle interface, which is unsafe to use inside a user ns because it
> > allows complete access to any inode in the filesystem without
> > limitations. xfs_restore requires unfettered access to directly
> > manipulate the uid/gid/security attrs of inodes, which once again is
> > something that isn't allowed inside user namespaces.
> >
> > Setting Posix acls by directly poking the on-disk attr format rather
> > than going through the proper kernel ACL namespace is not a *general
> > purpose user interface*.  Thi exists for backup/restore utilities to
> > do things like restore ACLs and security labels simply by treating
> > them as opaque xattrs.  If a user sets ACLs using this low level
> > "opaque xattr" method, then they get to keep all the broken bits to
> > themselves.
> 
> Any process capable of CAP_SYS_ADMIN can getxattr and setxattr those

CAP_SYS_ADMIN = enough rope to hang yourself.

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com