This is tracking the upstream effort needed to deliver the feature request described below.

In several places we use glibc's getifaddrs.

In it's standard implementation (glibc's), this function makes 2 really expensive netlink calls:

• GETLINK. This collects all interfaces (with full statistics, including inet ones because it does not use RTEXT_FILTER_SKIP_STATS)
• GETADDR: This collects all the Eth/IPv4/IPv6 addresses in the system

The obvious first thought is to use RTEXT_FILTER_SKIP_STATS in the first call but on second thought, why use GETLINK at all? The answer is the API exposed by libc.

The struct returned by getifaddrs contains interface name and its addresses. However, GETADDR returns interface id alongside the addresses. So, the only purpose of the other, very expensive netlink command is to map interface ids to interface names.

OVS already has the interface id <-> interface name mapping in the netdev_linux struct.

Therefore, the best approach IMHO would be to replace the getifaddrs implementation altogether with one that only sends GETADDR.