Cause: When creating a child process (e.g. to run another binary), libvirt closes filedescriptors right after the child process is created and keep only handful of them open. This is to enhance security, because we can't blindly trust that binary will not misuse any of the passed FDs. So far, this so called mass close of FDs was done by iterating throug every possible FD in range of 3 to 'limit of open files' and either called close() or made FD survive exec(). Problem is, only a handful of FDs need to be preserved and also only very few FDs need to be closed. The majority of FDs that we are trying to close() are not open really. While there is no technical problem, every close() is a syscall and as such adds some overhead. This is more visible if the limit for open files is way bigger than the default (1024). It that's the case, spawning a runnig a binary in a child process becomes very expensive (time-wise). Consequence: Fetching stats for an OVS interface is delayed and very annoying. Fix: Closing of FDs was changed. Now libvirt utilizes /proc filesystem to learn all FDs that are open and iterate only through them instead of the full range of possible FDs. Because the proportion between the two, this approach takes way less time and achieves the same result. I mean, like 99% of FDs within the range were never open really and thus are closed already, but we were trying to close() them anyway. Result: Fetching statistics is now as smooth as baby skin.