Bug 1445054

Description of problem:
When booting with ipv6.disable=1, vxlan will fail to initialize with the error "vxlan: Cannot bind port 4789, err=-97" which is EAFNOSUPPORT.  This is normally fine, except that should this IPv6 check fail due to EAFNOSUPPORT on the first call of __vxlan_sock_add(), the second non-IPv6 call of __vxlan_sock_add() will not occur.  This is due to the check done in vxlan_add_sock():

When __vxlan_add_sock() will return failure due to EAFNOSUPPORT via vxlan_sock_add->__vxlan_sock_add->vxlan_create_sock->udp_sock_create() refusing the AF_INET6 address, the first check fails and the second will never retry, so no vxlan socket is made.

static int vxlan_sock_add(struct vxlan_dev *vxlan)
{
        bool ipv6 = vxlan->flags & VXLAN_F_IPV6;
        bool metadata = vxlan->flags & VXLAN_F_COLLECT_METADATA;
        int ret = 0; 

        RCU_INIT_POINTER(vxlan->vn4_sock, NULL);
#if IS_ENABLED(CONFIG_IPV6)
        RCU_INIT_POINTER(vxlan->vn6_sock, NULL);
        if (ipv6 || metadata)
                ret = __vxlan_sock_add(vxlan, true);
#endif
        if ((!ret || ret == -EAFNOSUPPORT) && (!ipv6 || metadata))
                ret = __vxlan_sock_add(vxlan, false);
        if (ret < 0) 
                vxlan_sock_release(vxlan);
        return ret; 
}

Fix, as authored by Marcelo Leitner, is as below:

diff --git a/drivers/net/vxlan.c b/drivers/net/vxlan.c
index ebc98bb..2dfeda8 100644
--- a/drivers/net/vxlan.c
+++ b/drivers/net/vxlan.c
@@ -2832,7 +2832,7 @@ static int vxlan_sock_add(struct vxlan_dev *vxlan)
        if (ipv6 || metadata)
                ret = __vxlan_sock_add(vxlan, true);
 #endif
-       if (!ret && (!ipv6 || metadata))
+       if ((!ret || ret == -EAFNOSUPPORT) && (!ipv6 || metadata))
                ret = __vxlan_sock_add(vxlan, false);
        if (ret < 0)
                vxlan_sock_release(vxlan);

We now track to see if EAFNOSUPPORT is returned, and if it is we will fall through to an IPv4 socket using __vxlan_sock_add().

Version-Release number of selected component (if applicable):
Occurs in latest net-next and affects RHEL7 by proxy.

How reproducible:
Every time.

Steps to Reproduce:
1. Configure vxlan tunnel (I used an OVS bridge) on a host and attach a tap or veth to it:
	# ovs-vsctl add-br vxlan-br vx1
	# ovs-vsctl add-port vxlan-br vx1 -- set interface vx1 type=vxlan options:remote_ip=192.168.222.21
	# ovs-vsctl add-port vxlan-br $TAP/VETH
2. Attach a seperate network to the IP of the tap or veth device.
3. Reboot with ipv6.disable=1 as a command line option, vxlan tunnel will no longer function.

Actual results:
vxlan socket is not created for IPv4 despite only IPv6 being disabled

Expected results:
Only IPv6 socket should not be created, IPv4 socket should also be made.