Well, fullnat is the fourth mode for LVS, it's technically identical to LVS-NAT But it tries to preserve actual client IP addresses to real servers. In order to do that, FULLNAT server alters the incoming packet to inject the actual IP address into TCP/SYN packet, and real servers must load a kernel module to extract the information and present it to the user mode application.
The real server module (client) is implemented long ago by some Baidu folks for Linux 2.6.9, and they failed to find a perfect solution that minimize the impact on altering IP addresses the user mode see, aka NAT. So they came to an interesting but hacking way: hijack the getpeername() kernel API. The catch is you have to patch the real server kernel, thus is dangerous and unfriendly to kernel changes.
This is an alternative. I provide this kernel module for real servers to see actual client IP through LVS-FULLNAT without patching, via NAT implemented by netfilter. It's actually the reverse way SNAT and DNAT works: it does SNAT on incoming packets and does DNAT on outgoing packets. There's no need to patch anything, and as long as netfilter/conntrack is present in the kernel, you can simply load the module and see what happens.
Since 3.14, need_ipv4_conntrack() is removed, so you'll have to modprobe nf_conntrack_ipv4 manually.
Also UDP and IPv6 are not yet implemented.
Well, it's my first time doing kernel programming, so be gentle please.