[PATCH] selinux: log anon inode class name

[Christian Göttsche]

On Fri, 25 Feb 2022 at 01:25, Paul Moore <paul at paul-moore.com> wrote:
>
> On Thu, Feb 17, 2022 at 9:35 AM Christian Göttsche
> <cgzones at googlemail.com> wrote:
> >
> > Log the anonymous inode class name in the security hook
> > inode_init_security_anon.  This name is the key for name based type
> > transitions on the anon_inode security class on creation.  Example:
> >
> >     type=AVC msg=audit(02/16/22 22:02:50.585:216) : avc:  granted  { create } for  pid=2136 comm=mariadbd anonclass="[io_uring]" dev="anon_inodefs" ino=6871 scontext=system_u:system_r:mysqld_t:s0 tcontext=system_u:system_r:mysqld_iouring_t:s0 tclass=anon_inode
> >
> > Add a new LSM audit data type holding the inode and the class name.
> >
> > Also warn if the security hook gets called with no name set; currently
> > the only caller fs/anon_inodes.c:anon_inode_make_secure_inode() passes
> > one.
> >
> > Signed-off-by: Christian Göttsche <cgzones at googlemail.com>
> > ---
> >  include/linux/lsm_audit.h |  5 +++++
> >  security/lsm_audit.c      | 21 +++++++++++++++++++++
> >  security/selinux/hooks.c  |  7 +++++--
> >  3 files changed, 31 insertions(+), 2 deletions(-)
>
> ...
>
> > diff --git a/include/linux/lsm_audit.h b/include/linux/lsm_audit.h
> > index 17d02eda9538..8135a88d6d82 100644
> > --- a/include/linux/lsm_audit.h
> > +++ b/include/linux/lsm_audit.h
> > @@ -76,6 +76,7 @@ struct common_audit_data {
> >  #define LSM_AUDIT_DATA_IBENDPORT 14
> >  #define LSM_AUDIT_DATA_LOCKDOWN 15
> >  #define LSM_AUDIT_DATA_NOTIFICATION 16
> > +#define LSM_AUDIT_DATA_ANONINODE       17
> >         union   {
> >                 struct path path;
> >                 struct dentry *dentry;
> > @@ -96,6 +97,10 @@ struct common_audit_data {
> >                 struct lsm_ibpkey_audit *ibpkey;
> >                 struct lsm_ibendport_audit *ibendport;
> >                 int reason;
> > +               struct {
> > +                       struct inode *inode;
> > +                       const char *anonclass;
> > +               } anoninode_struct;
> >         } u;
> >         /* this union contains LSM specific data */
> >         union {
> > diff --git a/security/lsm_audit.c b/security/lsm_audit.c
> > index 1897cbf6fc69..5545fed35539 100644
> > --- a/security/lsm_audit.c
> > +++ b/security/lsm_audit.c
> > @@ -433,6 +433,27 @@ static void dump_common_audit_data(struct audit_buffer *ab,
> >                 audit_log_format(ab, " lockdown_reason=\"%s\"",
> >                                  lockdown_reasons[a->u.reason]);
> >                 break;
> > +       case LSM_AUDIT_DATA_ANONINODE: {
> > +               struct dentry *dentry;
> > +               struct inode *inode;
> > +
> > +               rcu_read_lock();
> > +               inode = a->u.anoninode_struct.inode;
> > +               dentry = d_find_alias_rcu(inode);
> > +               if (dentry) {
> > +                       audit_log_format(ab, " name=");
> > +                       spin_lock(&dentry->d_lock);
> > +                       audit_log_untrustedstring(ab, dentry->d_name.name);
> > +                       spin_unlock(&dentry->d_lock);
> > +               }
>
> I'm not sure we are ever going to get a useful dentry name for
> anonymous inodes, I think we can probably drop this.  The "anonclass="
> field will likely be much more interesting and helpful.
>
> > +               audit_log_format(ab, " anonclass=");
> > +               audit_log_untrustedstring(ab, a->u.anoninode_struct.anonclass);
> > +               audit_log_format(ab, " dev=");
> > +               audit_log_untrustedstring(ab, inode->i_sb->s_id);
>
> I'm pretty sure this is always going to end up being "anon_inodefs"
> and thus not very useful.
>
> > +               audit_log_format(ab, " ino=%lu", inode->i_ino);
>
> Similarly, I'm not sure how useful the inode number is in practice.
> I've never tried, but can a user lookup an anonymous inode via the
> inode number?
>
> > diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c
> > index dafabb4dcc64..19c831d94d9b 100644
> > --- a/security/selinux/hooks.c
> > +++ b/security/selinux/hooks.c
> > @@ -2932,6 +2932,8 @@ static int selinux_inode_init_security_anon(struct inode *inode,
> >         if (unlikely(!selinux_initialized(&selinux_state)))
> >                 return 0;
> >
> > +       WARN_ON(!name);
> > +
> >         isec = selinux_inode(inode);
> >
> >         /*
> > @@ -2965,8 +2967,9 @@ static int selinux_inode_init_security_anon(struct inode *inode,
> >          * allowed to actually create this type of anonymous inode.
> >          */
> >
> > -       ad.type = LSM_AUDIT_DATA_INODE;
> > -       ad.u.inode = inode;
> > +       ad.type = LSM_AUDIT_DATA_ANONINODE;
> > +       ad.u.anoninode_struct.inode = inode;
> > +       ad.u.anoninode_struct.anonclass = name ? (const char *)name->name : "unknown(null)";
>
> This doesn't seem to match well with the newly added WARN_ON()
> assertion above.  I would suggest dropping the WARN_ON() assertion as
> security_transition_sid() can already handle that safely, and leaving
> the tertiary statement above; however I think we should probably
> change the anonclass string to "?" as that is the common unset field
> value used by audit.

Is the hook inode_init_security_anon expected to be called with an
empty name though?
The condition was just a fallback to not crash the kernel.
(Dropped in v2.)

>
> --
> paul-moore.com