Re: [PATCH] veth: delay peer link configuration after interfaces are tied

[Nicolas Dichtel <nicolas.dichtel@6wind.com>]

Le 30/05/2016 17:26, Vincent Bernat a écrit :
>  ❦ 30 mai 2016 17:19 CEST, Nicolas Dichtel <nicolas.dichtel@6wind.com> :
> 
>>>>>  	priv = netdev_priv(peer);
>>>>>  	rcu_assign_pointer(priv->peer, dev);
>>>>> +
>>>>> +	err = rtnl_configure_link(peer, ifmp);
>>>>> +	if (err < 0)
>>>>> +		goto err_configure_peer;
>>>
>>>> You should fix the error path. 'unregister_netdevice(dev)' is missing.
>>>
>>> I am sending another patch to fix that. I am quite unsure if I do the
>>> right thing here.
>>>
>> A less intrusive fix is to call 'rtmsg_ifinfo(RTM_NEWLINK, peer, ~0U,
>> GFP_KERNEL);' a the end of veth_newlink().
> 
> I did that at first. Maybe this would make more sense to do
> that. Otherwise, the first message contains an iflink value that we
> cannot resolve with just the received netlink messages (since the
> information is in the next netlink message). "ip monitor" seems to be
> able to get the info, but I suppose it does an additional
> lookup.
> 
Yes, it's a chicken and egg problem ;-)
I think that the first message with an iflink set to '0' is not a problem if a
second one update this value. Daemons that listen to those rtnl messages must
always update their caches. When the peer is put in another netns, its ifindex
may change again.