Masquerade of outbound IPVS traffic uses wrong source ip when loopback has an ip alias

[lucasmundim]

If I add an ip alias on loopback interface IPVS outbound traffic masquerade uses that ip instead of nodeIP.

Example:

Given a container with host-network=true and dnsPolicy=ClusterFirstWithHostNet

From container (works)

$ getent hosts kubernetes.default.svc.cluster.local.
10.255.2.1        kubernetes.default.svc.cluster.local.  kubernetes.default.svc.cluster.local.

tcpdump

[Interface:tun-1022413915] 06:23:59.104003 IP 10.224.64.6.31403 > 10.244.2.125.53: 43707+ AAAA? kubernetes.default.svc.cluster.local. (54)
[Interface:tun-1022413915] 06:23:59.104841 IP 10.244.2.125.53 > 10.224.64.6.31403: 43707* 0/1/0 (147)
[Interface:tun-1022413915] 06:23:59.104983 IP 10.224.64.6.44205 > 10.244.2.124.53: 8538+ A? kubernetes.default.svc.cluster.local. (54)
[Interface:tun-1022413915] 06:23:59.105618 IP 10.244.2.124.53 > 10.224.64.6.44205: 8538* 1/0/0 A 10.255.2.1 (106)
[Interface:tun-1022413915] 06:23:59.105732 IP 10.224.64.6.52080 > 10.244.2.125.53: 53493+ A? kubernetes.default.svc.cluster.local. (54)
[Interface:tun-1022413915] 06:23:59.106206 IP 10.244.2.125.53 > 10.224.64.6.52080: 53493* 1/0/0 A 10.255.2.1 (106)

From host

$ ip addr add 192.168.124.81/32 dev lo label lo:2

From container (stops working)

$ getent hosts kubernetes.default.svc.cluster.local.
# no output = timeout

tcpdump

[Interface:tun-1022413915] 06:23:02.411767 IP 192.168.124.81.53081 > 10.244.2.124.53: 51862+ AAAA? kubernetes.default.svc.cluster.local. (54)