use shared memory

[Lex-2008]

Issue was that different worker processes had different value of "nonce", thus producing different ipscrub value for same IP.

Consider this nginx.conf:

load_module modules/ngx_ipscrub_module.so;
events {
  accept_mutex on;
}
worker_processes 2;
daemon off;
http {
  server {
listen 8000 default_server reuseport;
location / { return 200 "hello from $pid $remote_addr_ipscrub\n"; }
  }
}

interesting parts here are accept_mutex on and worker_processes 2 - these two options mean that nginx will start two worker processes, and they will accept connections by turn. For some reason, I also had to add reuseport to listen for this to work, too.

If you run nginx with above config and run curl in a loop, you would get output like this:

$ while true; do curl localhost:8000/; done
hello from 533073 1xGVi2
hello from 533072 +G8Qbi
hello from 533073 1xGVi2
hello from 533072 +G8Qbi
hello from 533072 +G8Qbi
hello from 533072 +G8Qbi
hello from 533073 1xGVi2
hello from 533072 +G8Qbi
^C

Probably that's not what you wanted :)

This commit fixes it by using shared memory.
As another benefit, nonce now also "survives" nginx reconfiguration events.