On Tue, 2021-12-07 at 10:16 -0500, James Bottomley wrote:
> On Tue, 2021-12-07 at 15:59 +0100, Christian Brauner wrote:
> > On Mon, Dec 06, 2021 at 04:14:15PM -0500, James Bottomley wrote:
[...]
> > > static int securityfs_fill_super(struct super_block *sb, struct
> > > fs_context *fc)
> > > {
> > > static const struct tree_descr files[] = {{""}};
> > > int error;
> > > + struct user_namespace *ns = fc->user_ns;
> > >
> > > error = simple_fill_super(sb, SECURITYFS_MAGIC, files);
> > > if (error)
> > > return error;
> > >
> > > + ns->securityfs_root = dget(sb->s_root);
> > > +
> > > sb->s_op = &securityfs_super_operations;
> > >
> > > + if (ns != &init_user_ns)
> > > + blocking_notifier_call_chain(&securityfs_ns_notifier,
> > > + SECURITYFS_NS_ADD, ns);
> >
> > I would propose not to use the notifier logic. While it might be
> > nifty it's over-engineered in my opinion.
>
> The reason for a notifier is that this current patch set only
> namespaces ima, but we also have integrity and evm to do. Plus, as
> Casey said, we might get apparmour and selinux. Since each of those
> will also want to add entries in fill_super, the notifier mechanism
> seemed fairly tailor made for this. The alternative is to have a
> load of
>
> #if CONFIG_securityfeature
> callback()
> #endif
>
> Inside securityfs_fill_super which is a bit inelegant.
>
> > The dentry stashing in struct user_namespace currently serves the
> > purpose to make it retrievable in ima_fs_ns_init(). That doesn't
> > justify its existence imho.
>
> I can thread the root as part of the callback. I think I can still
> use the standard securityfs calls because the only reason for the
> dentry in the namespace is so the callee can pass NULL and have the
> dentry created at the top level. We can insist in the namespaced use
> case that the callee always pass in the dentry, even for the top
> level.
>
> > There is one central place were all users of namespaced securityfs
> > can create the files that they need to and that is in
> > securityfs_fill_super(). (If you want to make that more obvious
> > then give it a subdirectory securityfs and move inode.c in there.)
>
> Right, that's what the patch does.
>
> > We simply will expect users to add:
> >
> > ima_init_securityfs()
> > mylsm_init_securityfs()
>
> Yes, plus all the #ifdefs because securityfs can exist independently
> of each of the features. We can hide the ifdefs in the header files
> and make the functions static do nothing if not defined, but the
> ifdeffery has to live somewhere.
Actually, I've got a much better reason: securityfs is a bool; all the
other LSMs and IMA are tristates. We can't call module init functions
from core code, it has to be done by something like a notifier.
James
