Re: [PATCH v2] fs: block cross-uid sticky symlinks

[Eric Paris <eparis@redhat.com>]

On Tue, 2010-06-01 at 07:52 -0700, Kees Cook wrote:
> On Tue, Jun 01, 2010 at 07:55:02AM -0400, Eric Paris wrote:
> > On Tue, 2010-06-01 at 03:55 -0400, Christoph Hellwig wrote:
> > > On Mon, May 31, 2010 at 08:24:23PM -0700, Kees Cook wrote:
> > 
> > > > My rationale is that if it's in commoncaps, it's effective for everyone, so
> > > > it might as well be in core VFS.  If the VFS objections really do boil down
> > > > to "not in fs/" then I'm curious if doing this in commoncaps is acceptable.
> > > 
> > > If you think the objection is about having things in fs/ you're smoking
> > > some really bad stuff.
> > 
> > Sounds to me like we should probably follow the same path as
> > mmap_min_addr.  We should add these hooks right in the VFS where they
> > belong (much like mmap_min_addr hooks into the vm) and control them 2
> > ways.
> > 
> > 1) a Kconfig so distros can choose to turn it on or off by default
> > 2) a /proc interface so root can turn it off
> > 
> > Nothing about that precludes additional similar checks inside an LSM
> > (like CONFIG_LSM_MMAP_MIN_ADDR) which can be more finely controlled.  So
> > maybe we want to follow up with the core VFS check with new checks in
> > SELinux (and maybe apparmour).  This allows the user to disable the
> > general check and still be provided with some modicum of protection.
> > You might ask why not ONLY do the check in SELinux and drop the generic
> > check, but we have seen with mmap_min_addr that the SELinux unconfined
> > user can do damn well anything it wants to, so having a non-LSM version
> > of appropriate security checks is highly regarded.
> 
> Would a CONFIG for this be overkill?  mmap_min_addr is a little different
> in that there was desire to control a bottom limit on it, etc.  Given this
> is either "on" or "off", I think just a sysctl is needed?

Seems like one of Alan's main arguments is that you should not turn it
on 'by default.'  I assume most distros will want it on by default.
Alan made the same argument against mmap_min_addr (known to break dos
emu) but I think most major distros have it on by default these days
even if it does break those weird obscure use cases.  I guess distros
can do it through sysctl but Fedora, at least, likes to keep those
default if possible, which is why I suggested the CONFIG.  In any case,
putting this right square in the VFS where it happens makes the most
sense to me.

I'd also like to point out that I don't buy the argument that per
user /tmp/ is a 'better' solution for the general case.  Any application
that would be broken by this change will also be broken by per
user /tmp.  Now, if we used filesystem namespaces regularly for years
and users, administrators, and developers dealt with them often I agree
that would probably be the preferred solution.  It would solve this
issue, but in introduces a whole host of other problems that are even
more obvious and even likely to bite people.

I probably would move the security hook down into __do_follow_link and
put this check down there as well, but I think you still have a problem
with d_parent.  I don't see what keeps d_parent from being freed while
you are using it....

-Eric