Re: [PATCH net] rtnl: don't send rtnl msg for unregistered iface

From: Andy Gospodarek <gospo@cumulusnetworks.com>
To: Jiri Pirko <jiri@resnulli.us>
Cc: Nicolas Dichtel <nicolas.dichtel@6wind.com>,
	davem@davemloft.net, netdev@vger.kernel.org,
	j.vosburgh@gmail.com, vfalico@gmail.com
Subject: Re: [PATCH net] rtnl: don't send rtnl msg for unregistered iface
Date: Mon, 11 May 2015 18:15:37 -0400	[thread overview]
Message-ID: <20150511221536.GB9559@gospo.home.greyhouse.net> (raw)
In-Reply-To: <20150511175355.GA3130@nanopsycho>

On Mon, May 11, 2015 at 07:53:55PM +0200, Jiri Pirko wrote:
> Mon, May 11, 2015 at 06:15:00PM CEST, nicolas.dichtel@6wind.com wrote:
> >Le 11/05/2015 18:04, Nicolas Dichtel a écrit :
> >>Le 11/05/2015 17:37, Jiri Pirko a écrit :
> >>[snip]
> >>>
> >>>I don't think this is the correct way to fix this.
> >>>
> >>>How ifindex can be 0 here? Ifindex is set in register_netdevice and
> >>>looking at bond_create, I don't see any call to __bond_opt_set before
> >>>that. But since it apparently is, the ordering should be changed so
> >>>register_netdevice is called first.
> >I also don't see why we would prevent to register a bonding interface directly
> >with the right mode.
> >
> >>bond_newlink => bond_changelink => __bond_opt_set
> >>and then back to bond_newlink => register_netdevice
> 
> I see it now. Why not to do register first, changelink later?
> Or, change __bond_opt_set to call call_netdevice_notifiers only in case
> dev->reg_state == NETREG_REGISTERED? Looking at the other places this

This is probably an excellent approach.

> check happens, looks like a little helper like "netdev_check_registered"
> might be convenient.