Re: [RFC PATCH v2] libsepol/cil: raise default attrs_expand_size to 2

[Stephen Smalley <stephen.smalley.work@gmail.com>]

On Tue, May 12, 2020 at 1:16 PM Petr Lautrbach <plautrba@redhat.com> wrote:
>
> On Mon, May 11, 2020 at 03:01:06PM -0400, James Carter wrote:
> > On Mon, May 11, 2020 at 2:09 PM Stephen Smalley
> > <stephen.smalley.work@gmail.com> wrote:
> > >
> > > On Mon, May 11, 2020 at 2:03 PM James Carter <jwcart2@gmail.com> wrote:
> > > >
> > > > On Mon, May 11, 2020 at 9:25 AM James Carter <jwcart2@gmail.com> wrote:
> > > > >
> > > > > On Mon, May 11, 2020 at 8:27 AM Petr Lautrbach <plautrba@redhat.com> wrote:
> > > > > >
> > > > > > On Thu, Mar 05, 2020 at 02:53:37PM +0100, Ondrej Mosnacek wrote:
> > > > > > > The value attrs_expand_size == 1 removes all empty attributes, but it
> > > > > > > also makes sense to expand all attributes that have only one type. This
> > > > > > > removes some redundant rules (there is sometimes the same rule for the
> > > > > > > type and the attribute) and reduces the number of attributes that the
> > > > > > > kernel has to go through when looking up rules.
> > > > > > >
> > > > > > > Signed-off-by: Ondrej Mosnacek <omosnace@redhat.com>
> > > > > > > ---
> > > > > > >
> > > > > > > v2: fix typos (Tne -> The; cointains -> contains)
> > > > > > >
> > > > > > >  libsepol/cil/src/cil.c | 3 ++-
> > > > > > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > > > > >
> > > > > > > diff --git a/libsepol/cil/src/cil.c b/libsepol/cil/src/cil.c
> > > > > > > index d222ad3a..c010ca2a 100644
> > > > > > > --- a/libsepol/cil/src/cil.c
> > > > > > > +++ b/libsepol/cil/src/cil.c
> > > > > > > @@ -452,7 +452,8 @@ void cil_db_init(struct cil_db **db)
> > > > > > >       (*db)->disable_dontaudit = CIL_FALSE;
> > > > > > >       (*db)->disable_neverallow = CIL_FALSE;
> > > > > > >       (*db)->attrs_expand_generated = CIL_FALSE;
> > > > > > > -     (*db)->attrs_expand_size = 1;
> > > > > > > +     /* 2 == remove attributes that contain none or just 1 type */
> > > > > > > +     (*db)->attrs_expand_size = 2;
> > > > > > >       (*db)->preserve_tunables = CIL_FALSE;
> > > > > > >       (*db)->handle_unknown = -1;
> > > > > > >       (*db)->mls = -1;
> > > > > > > --
> > > > > > > 2.24.1
> > > > > > >
> > > > > >
> > > > > >
> > > > > > This patch broke `semanage node -l` on Fedora [1]
> > > > > >
> > > > > > :: [ 21:25:25 ] :: [  BEGIN   ] :: Running 'make LIBDIR=/usr/lib64 SHLIBDIR=/lib64 test'
> > > > > > ...
> > > > > > test_list (__main__.SemanageTests) ... Traceback (most recent call last):
> > > > > >   File "/usr/sbin/semanage", line 967, in <module>
> > > > > >     do_parser()
> > > > > >   File "/usr/sbin/semanage", line 946, in do_parser
> > > > > >     args.func(args)
> > > > > >   File "/usr/sbin/semanage", line 649, in handleNode
> > > > > >     OBJECT = object_dict['node'](args)
> > > > > >   File "/usr/lib/python3.8/site-packages/seobject.py", line 1849, in __init__
> > > > > >     self.valid_types = list(list(sepolicy.info(sepolicy.ATTRIBUTE, "node_type"))[0]["types"])
> > > > > > IndexError: list index out of range
> > > > > >
> > > > > > While the `IndexError: list index out of range` error can be simply fixed, it
> > > > > > uncovered the problem that semanage uses attibutes to list certain records -
> > > > > > node_type, port_type, file_type, device_node, ... and these attributes can disappear when
> > > > > > there's only 1 type assigned.
> > > > > >
> > > > > > I guess it should be reverted as there's no other way how to find out that a
> > > > > > type node_t is node_type.
> > > > > >
> > > > > > [1] https://jenkins-continuous-infra.apps.ci.centos.org/job/fedora-rawhide-pr-pipeline/3462/artifact/package-tests/logs/FAIL-upstream-err.log
> > > > > >
> > > >
> > > > I see now.  python/semanage/seobject.py and
> > > > python/semanage/semanage-bash-completion.sh both assume that node_type
> > > > is always defined as an attribute in a policy. There seems to be quite
> > > > a lot that is assumed about policy in the python directory.
> > > >
> > > > This is not a bug in CIL or libsepol. Ideally the tests should have
> > > > their own policy and the python code should gracefully handle the
> > > > situation when its assumptions are wrong. If we need to revert this
> > > > patch in the short-term than I am ok with that.
> > >
> > > We can mark these attributes explicitly in policy to prevent their
> > > expansion, right?  So while we cannot make this change right now
> > > without breaking compatibility with selinux userspace tools (not just
> > > tests), we could start marking these attributes on which the tools
> > > depend in policy and then later we can re-apply this?
> >
> > Yes, we can add the rule "expandattribute node_type false;" in policy
> > for node_type and any other attributes that are required to exist.
> >
>
>
> It doesn't seem to be propagated to cil:
>
> # cat mypolicy.te
> policy_module(mypolicy,1.0)
>
> type myapp_t;
> type myapp_log_t;
>
> attribute myattribute;
> expandattribute myattribute false;
> typeattribute myapp_t myattribute;
>
> allow myattribute myapp_log_t:file read;
>
> # make -f /usr/share/selinux/devel/Makefile mypolicy.pp
> make: 'mypolicy.pp' is up to date.
>
> # /usr/libexec/selinux/hll/pp mypolicy.pp
> (type myapp_t)
> (roletype object_r myapp_t)
> (type myapp_log_t)
> (roletype object_r myapp_log_t)
> (typeattribute myattribute)
> (typeattributeset myattribute (myapp_t ))
> (roleattributeset cil_gen_require system_r)
> (allow myattribute myapp_log_t (file (read)))
>
>
> But it works with mypolicy.cil which contains:
> (expandtypeattribute myattribute false)
>
>
> I'm not really experienced in writing policy but I still find the behaviour
> confusing.
> It's not only about `semanage`, you can hit this using `sesearch` as well, e.g.
> before policy is rebuilt with new libsepol, `sesearch -A -t node_type` finds
> about 535 rules with node_type, then you rebuild policy and it's 0. But if
> assign node_type to another type, it's again more than 536.
>
> For this particular attribute we can/should use `expandtypeattribute` but should
> be this expression used every time you have an attribute assigned to only 1 type?
>
>
> If it stays as it is, it definitely needs to be part of release notes.

I think we should revert the change prior to release since it breaks
selinux userspace.
We'll need to fix the separate bug regarding pp to cil conversion that
is losing the expandattribute statement that you mentioned above.
And we should start annotating all attributes in policy on which
selinux userspace depends with
explicit expandattribute false statements to ensure that they are not
optimized away in the future.