Re: [PATCH 2/2] auth_gss: unregister gss_domain when unloading module

[Akinobu Mita]

2006/10/31, Trond Myklebust <Trond.Myklebust@xxxxxxxxxx>:
> On Mon, 2006-10-30 at 23:54 +0900, Akinobu Mita wrote:
> > On Sun, Oct 29, 2006 at 03:15:59PM -0500, Trond Myklebust wrote:
> >
> > > > +void svcauth_gss_unregister_pseudoflavor(char *name)
> > > > +{
> > > > + struct auth_domain *dom;
> > > > +
> > > > + dom = auth_domain_find(name);
> > > > + if (dom) {
> > > > +         auth_domain_put(dom);
> > > > +         auth_domain_put(dom);
> > > > + }
> > > > +}
> > >
> > > Strictly speaking, if you want to be smp-safe, you probably need
> > > something like the following:
> > >
> > >     dom = auth_domain_find(name);
> > >     if (dom) {
> > >             spin_lock(&auth_domain_lock);
> > >             if (!hlist_unhashed(dom->hash)) {
> > >                     hlist_del_init(dom->hash);
> > >                     spin_unlock(&auth_domain_lock);
> > >                     auth_domain_put(dom);
> > >             } else
> > >                     spin_unlock(&auth_domain_lock);
> > >             auth_domain_put(dom);
> > >     }
> > >
> > > and then add a test for hlist_unhashed into auth_domain_put(). If not,
> > > some other processor could race you inside
> > > svcauth_gss_unregister_pseudoflavor.
> >
> > But auth_domain_table is protected by auth_domain_lock while we are
> > using auth_domain_put()/auth_domain_lookup()/auth_domain_find().
> > So I think there is not big difference.
>
> No. The auth_domain_lock was released after the call to
> auth_domain_find(), and thus there is no guarantee that the entry is
> still referenced when you get round to that second call to
> auth_domain_put(). Testing for hlist_unhashed() and then removing the
> entry from the lookup table while under the spin lock fixes this
> problem: it ensures that you only call auth_domain_put() once if some
> other process has raced you.

Thanks, I understand it.

But I noticed that even if we have this kind of smp-safe code, there
is no guarantee that 2nd auth_domain_put() in
svcauth_gss_unregister_pseudoflavor() is the last reference of
this gss_domain.

So it is possible to happen invalid dereference by real last user of
this gss_domain after unloading module. If this is not wrong,
Is it neccesary to have try_get_module()/put_module() somewhere to
prevent this?
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/