Re: [PATCH v2] fuse: Add module param for non-descendant userns access to allow_other

[Christian Brauner <brauner@kernel.org>]

On Mon, Jun 13, 2022 at 12:34:05PM +0200, Miklos Szeredi wrote:
> On Mon, 13 Jun 2022 at 11:37, Christian Brauner <brauner@kernel.org> wrote:
> >
> > On Mon, Jun 13, 2022 at 10:23:47AM +0200, Miklos Szeredi wrote:
> > > On Fri, 10 Jun 2022 at 23:39, Andrii Nakryiko <andriin@fb.com> wrote:
> > > >
> > > >
> > > >
> > > > On 6/7/22 1:47 AM, Christian Brauner wrote:
> > > > > On Wed, Jun 01, 2022 at 11:44:07AM -0700, Dave Marchevsky wrote:
> > >
> > > [...]
> > >
> > > > >> +static bool __read_mostly allow_other_parent_userns;
> > > > >> +module_param(allow_other_parent_userns, bool, 0644);
> > > > >> +MODULE_PARM_DESC(allow_other_parent_userns,
> > > > >> + "Allow users not in mounting or descendant userns "
> > > > >> + "to access FUSE with allow_other set");
> > > > >
> > > > > The name of the parameter also suggests that access is granted to parent
> > > > > userns tasks whereas the change seems to me to allows every task access
> > > > > to that fuse filesystem independent of what userns they are in.
> > > > >
> > > > > So even a task in a sibling userns could - probably with rather
> > > > > elaborate mount propagation trickery - access that fuse filesystem.
> > > > >
> > > > > AFaict, either the module parameter is misnamed or the patch doesn't
> > > > > implement the behavior expressed in the name.
> > > > >
> > > > > The original patch restricted access to a CAP_SYS_ADMIN capable task.
> > > > > Did we agree that it was a good idea to weaken it to all tasks?
> > > > > Shouldn't we still just restrict this to CAP_SYS_ADMIN capable tasks in
> > > > > the initial userns?
> > > >
> > > > I think it's fine to allow for CAP_SYS_ADMIN only, but can we then
> > > > ignore the allow_other mount option in such case? The idea is that
> > > > CAP_SYS_ADMIN allows you to read FUSE-backed contents no matter what, so
> > > > user not mounting with allow_other preventing root from reading contents
> > > > defeats the purpose at least partially.
> > >
> > > If we want to be compatible with "user_allow_other", then it should be
> > > checking if the uid/gid of the current task is mapped in the
> > > filesystems user_ns (fsuidgid_has_mapping()).  Right?
> >
> > I think that's doable. So assuming we're still talking about requiring
> > cap_sys_admin then we'd roughly have sm like:
> >
> >         if (fc->allow_other)
> >                 return current_in_userns(fc->user_ns) ||
> >                         (capable(CAP_SYS_ADMIN) &&
> >                         fsuidgid_has_mapping(..., &init_user_ns));
> 
> No, I meant this:
> 
>         if (fc->allow_other)
>                 return current_in_userns(fc->user_ns) ||
>                         (userns_allow_other &&
>                         fsuidgid_has_mapping(..., &init_user_ns));
> 
> But I think the OP wanted to allow real root to access the fs, which
> this doesn't allow (since 0 will have no mapping in the user ns), so
> I'm not sure what's the right solution...

I aimed to show that. You can setfs*id() and retain capabilities and
still access the filesystem.

> 
> Maybe the original patch is fine: this check isn't meant to protect
> the filesystem from access, it's meant to protect the accessor.

I don't have specific worries here. I'm just a bit hesitant to just let
anyone access the fs. But if we go for allow other semantics then that's
probably fine. Though I wonder why then we don't just do:

if (fc->allow_other)
        return current_in_userns(fc->user_ns) ||
                (userns_allow_other &&
                ns_capable(fc->user_ns, CAP_SYS_ADMIN));

? That'll let any ancestor userns access the fs not just descendants of
fc->user_ns.