Re: ?????: VFS, NFS security bug? Should CAP_MKNOD and CAP_LINUX_IMMUTABLE be added to CAP_FS_MASK?

["Serge E. Hallyn" <serue@us.ibm.com>]

Quoting Casey Schaufler (casey@schaufler-ca.com):
> Serge E. Hallyn wrote:
> > Quoting Stephen Smalley (sds@tycho.nsa.gov):
> >   
> >> On Tue, 2009-03-17 at 12:39 -0500, Serge E. Hallyn wrote:
> >>     
> >>> Quoting Stephen Smalley (sds@tycho.nsa.gov):
> >>>       
> >>>>> So do you think it makes sense to have CAP_MAC_ADMIN and CAP_FOWNER
> >>>>> in CAP_FS_MASK?  In other words are you objecting to CAP_SYS_ADMIN
> >>>>> because of all of its other implications, or because you disagree
> >>>>> that labels for security modules should be treated as mere fs data
> >>>>> here?
> >>>>>           
> >>>> For CAP_FOWNER, yes (and it is already there).  CAP_MAC_ADMIN is less
> >>>>         
> >>> Sorry, I meant CAP_SETFCAP.  Should it be added?
> >>>       
> >> Sure - it is only used for filesystem operations.
> >>     
> >
> > Ok, so then:
> >
> >   
> >>>> ideal as it isn't clearly tied to filesystem accesses, and likewise for
> >>>> CAP_MAC_OVERRIDE (but that one is included in CAP_FS_MASK already).
> >>>>         
> >>> So it is.  I didn't realize that.
> >>>
> >>>       
> >>>> Ideally the capability space would be partitioned into capabilities that
> >>>> affect filesystem accesses and the rest so that setfsuid() would yield
> >>>> the expected behavior of only affecting filesystem access.
> >>>> CAP_SYS_ADMIN is even less suitable due to its pervasive use outside of
> >>>> the filesystem.  So that's the first concern.
> >>>>
> >>>> The second one is that we don't want CAP_SYS_ADMIN (or CAP_MAC_ADMIN) to
> >>>> be required when setting SELinux labels.  Only the SELinux permission
> >>>> checks should govern setting those labels (aside from the usual DAC
> >>>> ownership || CAP_FOWNER check).
> >>>>         
> >>> So if a non-selinux kernel is booted, then you think only the usual
> >>> DAC checks should be required to set selinux labels?
> >>>       
> >> I'm talking about the dumb NFS server case (non-SELinux NFS server
> >> providing label and data storage to SELinux clients, MAC enforcement
> >> handled client-side).  But we aren't there yet, so I don't think we have
> >> to worry about it right now.
> >>     
> >
> > But in cap_inode_setxattr, any security.* xattrs are controlled by
> > CAP_SYS_ADMIN.  So do you think that this should be changed to a
> > CAP_XATTR_SECURITY capability which can be added to CAP_FS_MASK?
> >   
> 
> Hum. The intention of CAP_MAC_ADMIN was that it control the explicit
> setting of the access control attributes used by the Smack LSM. I
> personally prefer a single capability for the action over multiple
> capabilities based on the objects involved. If you introduce
> CAP_XATTR_SECURITY I would think that CAP_PROC_XATTR,
> CAP_SVIPC_XATTR, CAP_NETWORK_XATTR, ... would follow in short order
> and I hate the idea of having hundreds of capabilities. If you
> must decouple the capability from MAC, how about a new name?

Oh I didn't say that we must, I'm just trying to figure out what we want
to do in the case that a security.foo xattr is being set, and the foo
LSM is not compiled in.

What is being done right now is that CAP_SYS_ADMIN is required to do
the setting, and so doing

	setresuid(500,500,0);
	setfsuid(0);
	setxattr(somefilename, "security.SMACK64", LABEL, sizeof(LABEL), 0);

will fail the setxattr.

-serge