Re: [PATCH net-next v9 06/11] net: ipv6: Use link netns in newlink() of rtnl_link_ops

[Kuniyuki Iwashima <kuniyu@amazon.com>]

From: Xiao Liang <shaw.leon@gmail.com>
Date: Fri, 14 Feb 2025 17:22:28 +0800
> On Thu, Feb 13, 2025 at 7:00 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> >
> > From: Xiao Liang <shaw.leon@gmail.com>
> > Date: Thu, 13 Feb 2025 17:55:32 +0800
> > > On Thu, Feb 13, 2025 at 4:37 PM Xiao Liang <shaw.leon@gmail.com> wrote:
> > > >
> > > > On Thu, Feb 13, 2025 at 3:05 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> > > > >
> > > > [...]
> > > > > > diff --git a/net/ipv6/ip6_gre.c b/net/ipv6/ip6_gre.c
> > > > > > index 863852abe8ea..108600dc716f 100644
> > > > > > --- a/net/ipv6/ip6_gre.c
> > > > > > +++ b/net/ipv6/ip6_gre.c
> > > > > > @@ -1498,7 +1498,8 @@ static int ip6gre_tunnel_init_common(struct net_device *dev)
> > > > > >       tunnel = netdev_priv(dev);
> > > > > >
> > > > > >       tunnel->dev = dev;
> > > > > > -     tunnel->net = dev_net(dev);
> > > > > > +     if (!tunnel->net)
> > > > > > +             tunnel->net = dev_net(dev);
> > > > >
> > > > > Same question as patch 5 for here and other parts.
> > > > > Do we need this check and assignment ?
> > > > >
> > > > > ip6gre_newlink_common
> > > > > -> nt->net = dev_net(dev)
> > > > > -> register_netdevice
> > > > >   -> ndo_init / ip6gre_tunnel_init()
> > > > >     -> ip6gre_tunnel_init_common
> > > > >       -> tunnel->net = dev_net(dev)
> > > >
> > > > Will remove this line.
> > >
> > > However, fb tunnel of ip6_tunnel, ip6_vti and sit can have
> > > tunnel->net == NULL here. Take ip6_tunnel for example:
> > >
> > > ip6_tnl_init_net()
> > >     -> ip6_fb_tnl_dev_init()
> > >     -> register_netdev()
> > >         -> register_netdevice()
> > >             -> ip6_tnl_dev_init()
> > >
> > > This code path (including ip6_fb_tnl_dev_init()) doesn't set
> > > tunnel->net. But for ip6_gre, ip6gre_fb_tunnel_init() does.
> >
> > Ah, okay.  Then, let's set net in a single place, which would
> > be better than spreading net assignment and adding null check
> > in ->ndo_init(), and maybe apply the same to IPv4 tunnels ?
> 
> Tunnels are created in three ways: a) rtnetlink newlink,
> b) ioctl SIOCADDTUNNEL and c) during per netns init (fb).
> The code paths don't have much in common, and refactoring
> to set net in a single place is somewhat beyond the scope
> of this series. But for now I think we could put a general rule:
> net should be set prior to register_netdevice().
> 
> For IPv4 tunnels, tunnel->net of a) is set in ip_tunnel_newlink().
> b) and c) are set in __ip_tunnel_create():
> ip_tunnel_init_net() -> __ip_tunnel_create()
> ip_tunnel_ctl() -> ip_tunnel_create() -> __ip_tunnel_create()
> So net has already been initialized when register_netdevice()
> is called.
> 
> But it varies for IPv6 tunnels. Some set net for a) or c) while
> some don't. This patch has "fixed" for a). As for c) we can
> adopt the way of ip6_gre - setting net in *_fb_tunnel_init(),
> then remove the check in ndo_init().
> 
> Is it reasonable?

Yes, fair enough.