Handling of SIOC{ADD,DEL}RT for 32bit is somewhat odd. AFAICS, the rules for native ioctl look so:
AF_APPLETALK, AF_INET, AF_IPX, AF_PACKET: take struct rtentry. The last one doesn't have ->compat_ioctl() and 32bit automatically hits routing_ioctl() in net/socket.c, the rest have ->compat_ioctl() but it doesn't recognize SIOC{ADD,DEL}RT, so it ends up handled by the same code. AF_INET6: takes struct in6_rtmsg. Hits routing_ioctl(), which recognizes ipv6 and does the right thing. AF_X25: takes x25_route_struct. Layout is apparently identical for 32bit and 64bit. Has ->compat_ioctl(), which does the same thing as ->ioctl() on those two. AF_AX25: takes struct ax25_routes_struct. Again, identical layout on 32bit and 64bit. Unfortunately, there's no ->compat_ioctl() in this one, so we end up hitting routing_ioctl() and get screwed. AF_NETROM: same as previous, except that it takes struct nr_route_struct. Apparently broken. AF_ROSE: ditto, with struct rose_route_struct. AF_QIPCRTR: explicitly recognizes and fails with -EINVAL. Odd (other protocol families without SIOCADDRT support fail with -ENOTTY), but clearly not an issue for compat code. Everything else: fails with -ENOTTY. Are AF_{AX25,NETROM,ROSE} really broken for 32bit processes on biarch hosts, or am I missing something subtle in there?