RE: [PATCH net-next v02 1/2] af_packet: allow fanout_add when socket is not RUNNING

[Gur Stavi]

> Gur Stavi wrote:
> > >
> > > If we don't care about opening up fanout groups to ETH_P_NONE, then
> > > patch v2 seems sufficient. If explicitly blocking this, the ENXIO
> > > return can be added, but ideally without touching the other lines.
> >
> > I don't think that allowing ETH_P_NONE is relevant.
> > In my opinion the 2 options that should be considered to fail
> > fanout_add are:
> > 1. Testing proto == 0
> > 2. Testing proto == 0 || ifindex == -1
> >
> > The only corner case that is caught by [2] and missed by [1] is
> > the "unlisted" case during do_bind. It is such a rare case that
> > probably no one will ever encounter bind "unlisted" followed by
> > FANOUT_ADD. And this is not a dangerous corner case that leads to
> > system crash.
> >
> > However, being a purist, I see the major goal of code review to promote
> > correctness by identifying corner cases while improving style is a
> > secondary priority. Since we did identify this corner case in our
> > discussion I think we should still use [2].
> > I don't consider the code complex. In fact, to me, the ifindex clause
> > is a more understandable direct reason for failure than the proto which
> > is indirect. Having the ifindex clause helps figuring out the proto
> > clause.
> 
> It's interesting that the unlisted fix does not return ENODEV, but
> returns success and leaves the socket in an unbound state, equivalent
> to binding to ETH_P_NONE and ifindex 0. This seems surprising behavior
> to the caller.
> 
> On rereading that, I still do not see a purpose of special ifindex -1.
> 
>

Can this code be relevant?

		case NETDEV_UP:
			if (dev->ifindex == po->ifindex) {
				spin_lock(&po->bind_lock);
				if (po->num)
					register_prot_hook(sk);
				spin_unlock(&po->bind_lock);
			}
			break;

Perhaps, although the socket failed to (re) find the device, the device
is still aware of the socket and we need the ifindex condition to fail.