Re: [syzbot] [net?] general protection fault in rtnl_create_link

[Jakub Kicinski <kuba@kernel.org>]

On Mon, 14 Apr 2025 11:01:59 -0700 Kuniyuki Iwashima wrote:
> From: Jakub Kicinski <kuba@kernel.org>
> Date: Mon, 14 Apr 2025 10:37:27 -0700
> > On Sun, 13 Apr 2025 19:30:46 -0700 Kuniyuki Iwashima wrote:  
> > > diff --git a/include/net/netdev_lock.h b/include/net/netdev_lock.h
> > > index 5706835a660c..270e157a4a79 100644
> > > --- a/include/net/netdev_lock.h
> > > +++ b/include/net/netdev_lock.h
> > > @@ -30,7 +30,8 @@ static inline bool netdev_need_ops_lock(const struct net_device *dev)
> > >  	bool ret = dev->request_ops_lock || !!dev->queue_mgmt_ops;
> > >  
> > >  #if IS_ENABLED(CONFIG_NET_SHAPER)
> > > -	ret |= !!dev->netdev_ops->net_shaper_ops;
> > > +	if (dev->netdev_ops)
> > > +		ret |= !!dev->netdev_ops->net_shaper_ops;
> > >  #endif  
> > 
> > This is a bit surprising, we pretty much never validate if dev has ops.
> > 
> > I think we're guaranteed that IFF_UP will not be set if we just
> > allocated the device, so we can remove the locks in rtnl_create_link()
> > and to double confirm add a netdev_ops_assert_locked_or_invisible() 
> > in netif_state_change() ?  
> 
> Removing the lock from NEWLINK makes sense, but my concern
> was NETDEV_CHANGE, which will requires more caution ?
> 
> commit 04efcee6ef8d0f01eef495db047e7216d6e6e38f
> Author: Stanislav Fomichev <sdf@fomichev.me>
> Date:   Fri Apr 4 09:11:22 2025 -0700
> 
>     net: hold instance lock during NETDEV_CHANGE

How could we fire a notifier for a device that hasn't been initialized,
let alone registered?

I'm hoping that the _or_invisible assert in my suggestion will flag to
future developers trying to change netif_state_change() that the device
here may not be fully constructed.