Global ordering introduced in D22997 makes it possible. Since the current version of the routine discussed issues with it, I did the same with the new routine. I consider it to be enough of an improvement to go in as it is.
Tested by setting kern.maxvnodes=5000 and doing buildkernel. Results were similar to the stock kernel since vnode both vnode pressure is too much AND vlrureclaim only has one mode to go over everything (instead of only part of the list). This will be fixed later.