Re: [Regression?] 1ea0ce4069 ("selinux: allow changing labels for cgroupfs") stops Android from booting

[Stephen Smalley <sds@tycho.nsa.gov>]

On Mon, 2017-02-27 at 19:18 -0500, Paul Moore wrote:
> On Mon, Feb 27, 2017 at 4:23 PM, Stephen Smalley <sds@tycho.nsa.gov>
> wrote:
> > 
> > On Mon, 2017-02-27 at 12:48 -0800, Nick Kralevich wrote:
> > > 
> > > On Mon, Feb 27, 2017 at 11:53 AM, Stephen Smalley <sds@tycho.nsa.
> > > gov>
> > > wrote:
> > > > 
> > > > 
> > > > > 
> > > > > 
> > > > > I can reproduce it on angler (with a back-port of just that
> > > > > patch),
> > > > > although I am unclear on the cause.  The patch is only
> > > > > supposed
> > > > > to
> > > > > enable explicit setting of security labels by userspace on
> > > > > cgroup
> > > > > files, so it isn't supposed to cause any breakage under
> > > > > existing
> > > > > policy.  Prior to the patch, the kernel would always just
> > > > > return
> > > > > -1
> > > > > with errno EOPNOTSUPP upon attempts to set security labels on
> > > > > cgroup
> > > > > files; with the patch, the kernel may instead return -1 with
> > > > > errno
> > > > > EACCES if not allowed.  So I suppose if userspace was
> > > > > explicitly
> > > > > testing for EOPNOTSUPP and not failing hard in that case, it
> > > > > might
> > > > > cause breakage.  Not sure why existing userspace would be
> > > > > trying
> > > > > to
> > > > > relabel cgroup files, unless it is just a recursive
> > > > > restorecon
> > > > > that
> > > > > happens to traverse into a cgroup mount (and in that case,
> > > > > not
> > > > > sure
> > > > > why
> > > > > it would be fatal).  Other possible interaction would be use
> > > > > of
> > > > > setfscreatecon() prior to creating a file in cgroup.
> > > > 
> > > > Oh, I see - it is the latter.
> > > > 
> > > > For example, init.rc does mkdir /dev/cpuctl/bg_non_interactive,
> > > > which
> > > > internally looks up the context for that directory from
> > > > file_contexts
> > > > and does a setfscreatecon() followed by a mkdir().  Previously,
> > > > that
> > > > was ignored because cgroup did not support anything other than
> > > > the
> > > > policy-defined label.  But now it will try to use that label,
> > > > which
> > > > in
> > > > turn will trigger a denial in enforcing mode and the create
> > > > will
> > > > fail.
> > > > 
> > > > So this is an incompatible change and needs to be reverted.
> > > > We'll need to wrap it up with a policy capability or something
> > > > to
> > > > allow
> > > > it to be enabled only if the policy correctly supports
> > > > it.  Even
> > > > better, we should instead just allow the policy to specify
> > > > which
> > > > filesystems should support this behavior (already on the issues
> > > > list).
> > > > 
> > > 
> > > If Android is the only system affected by this bug, I would
> > > prefer to
> > > just fix Android to allow for this patch, rather than having
> > > additional kernel complexity.
> > 
> > Well, it does break userspace (even if it happens to only affect
> > Android, which isn't clear, e.g. possibly a distribution would
> > likewise
> > suffer breakage under a tighter policy), and we already have a
> > long-
> > standing open issue to replace the current set of whitelisted
> > filesystem types with something configuration-driven.  So I'm ok
> > with
> > reverting it and requiring it to be done in a more general
> > way.  The
> > latter is something we want regardless.
> 
> This went up to Linus during the current merge window via the
> stable-4.11 branch and I know the container guys really want this so
> I'd prefer to fix this up in 4.11 with a policy capability if
> possible
> (and I believe it should be).  I agree with Stephen that we need a
> better long term solution, but I think a policy capability should
> work
> in the short term.
> 
> Who wants to send me a patch? ;)

So, there are a couple of caveats with doing that:

1) It still requires the container folks to update their kernel,
libsepol, and policy in order to make use of the new policy capability.

2) The determination of whether a given mount should be assigned this
flag is made at mount time, so you can't simply reload policy with a
policy that defines this capability and have it automatically applied
to existing cgroup mounts.  You'd have to unmount and re-mount them
(more likely reboot).

Not saying you can't do that, just understand what is required.