I got naive microbenchmark. It allocates ncpu threads, and each threads creates n UDP sockets that are connect(2)-ed to a random valid address:port. In a test virtual machine with 8 CPUs and 10000 sockets per thread I got 6x-7x speedup. If the test is adjusted to close(2) sockets when 50% allocation is reached, so that both insertions and removals are exercised at the same time, the speedup is roughly the same. I didn't yet do any real tests on hardware, but benefits are clear.

The patch doesn't yet address O(n) cycle when we are allocating ports to connect to a single remote IP address.

I got another naive benchmark. It exercises situation when we create as much as possible connections from us to a single peer:port eventually running out of ports. This exercises the worst case of the cycle in in_pcb_lport_dest(), which at the end of the test is executed for 10^4 times to find a port. While the cycle itself is not optimized at all, on a virtual machine with 8 CPUs the test passes 6-7 times faster than before the change.

The patch doesn't apply to main, is there some prerequisite patch?