We found an issue in the kernel bug tracking that seems to explain it, but we never confirmed for sure. Fortunately updating either of those things fixed the issue. This was of course also hard because this was back in the golden days of vmware managed by client's IT team, so getting them to update things on a VM was hell as well.