Re: [PATCH net v2 02/11] net: hold instance lock during NETDEV_REGISTER/UP/UNREGISTER

[Jakub Kicinski <kuba@kernel.org>]

On Thu, 27 Mar 2025 06:56:50 -0700 Stanislav Fomichev wrote:
>  static void call_netdevice_unregister_notifiers(struct notifier_block *nb,
> -						struct net_device *dev)
> +						struct net_device *dev,
> +						struct net_device *locked)
>  {
>  	if (dev->flags & IFF_UP) {
>  		call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
>  					dev);
>  		call_netdevice_notifier(nb, NETDEV_DOWN, dev);
>  	}
> +	if (dev != locked)
> +		netdev_lock_ops(dev);
>  	call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
> +	if (dev != locked)
> +		netdev_unlock_ops(dev);
>  }
>  
>  static int call_netdevice_register_net_notifiers(struct notifier_block *nb,
> -						 struct net *net)
> +						 struct net *net,
> +						 struct net_device *locked)
>  {
>  	struct net_device *dev;
>  	int err;
>  
>  	for_each_netdev(net, dev) {
> +		if (locked != dev)
> +			netdev_lock_ops(dev);
>  		err = call_netdevice_register_notifiers(nb, dev);
> +		if (locked != dev)
> +			netdev_unlock_ops(dev);
>  		if (err)
>  			goto rollback;
>  	}

Any strong reason we wouldn't split the netns change into two chunks, 
and release the ops lock while we move the notifiers?
I'm thinking rename netif_change_net_namespace() back to
dev_change_net_namespace(), move it up in setlink, and let it take 
the lock.