Re: [PATCH net v2] net: vlan: fix a UAF in vlan_dev_real_dev()

[Petr Machata <petrm@nvidia.com>]

Ziyang Xuan <william.xuanziyang@huawei.com> writes:

> diff --git a/net/8021q/vlan.c b/net/8021q/vlan.c
> index 55275ef9a31a..a3a0a5e994f5 100644
> --- a/net/8021q/vlan.c
> +++ b/net/8021q/vlan.c
> @@ -123,9 +123,6 @@ void unregister_vlan_dev(struct net_device *dev, struct list_head *head)
>  	}
>  
>  	vlan_vid_del(real_dev, vlan->vlan_proto, vlan_id);
> -
> -	/* Get rid of the vlan's reference to real_dev */
> -	dev_put(real_dev);
>  }
>  
>  int vlan_check_real_dev(struct net_device *real_dev,
> diff --git a/net/8021q/vlan_dev.c b/net/8021q/vlan_dev.c
> index 0c21d1fec852..aeeb5f90417b 100644
> --- a/net/8021q/vlan_dev.c
> +++ b/net/8021q/vlan_dev.c
> @@ -843,6 +843,9 @@ static void vlan_dev_free(struct net_device *dev)
>  
>  	free_percpu(vlan->vlan_pcpu_stats);
>  	vlan->vlan_pcpu_stats = NULL;
> +
> +	/* Get rid of the vlan's reference to real_dev */
> +	dev_put(vlan->real_dev);
>  }
>  
>  void vlan_setup(struct net_device *dev)

This is causing reference counting issues when vetoing is involved.
Consider the following snippet:

    ip link add name bond1 type bond mode 802.3ad
    ip link set dev swp1 master bond1
    ip link add name bond1.100 link bond1 type vlan protocol 802.1ad id 100
    # ^ vetoed, no netdevice created
    ip link del dev bond1

The setup process goes like this: vlan_newlink() calls
register_vlan_dev() calls netdev_upper_dev_link() calls
__netdev_upper_dev_link(), which issues a notifier
NETDEV_PRECHANGEUPPER, which yields a non-zero error,
because a listener vetoed it.

So it unwinds, skipping dev_hold(real_dev), but eventually the VLAN ends
up decreasing reference count of the real_dev. Then when when the bond
netdevice is removed, we get an endless loop of:

    kernel:unregister_netdevice: waiting for bond1 to become free. Usage count = 0 

Moving the dev_hold(real_dev) to always happen even if the
netdev_upper_dev_link() call makes the issue go away.

I'm not sure why this wasn't happening before. After the veto,
register_vlan_dev() follows with a goto out_unregister_netdev, which
calls unregister_netdevice() calls unregister_netdevice_queue(), which
issues a notifier NETDEV_UNREGISTER, which invokes vlan_device_event(),
which calls unregister_vlan_dev(), which used to dev_put(real_dev),
which seems like it should have caused the same issue. Dunno.