A few months ago a bug involving NAT and locally bound sockeys was reported. NAT could assign a new source port to a connection that would conflict with a port already locally bound; when this happened certain traffic was prevented from being delivered to the local process. In particular, this really kills NFS over UDP - if a connection to the NFS server gets NAT'd and takes a port used by one of the NFS daemons, massive NFS timeouts can happen. I've seen this to be the case very often; 100 clients hidden behind a single NAT box, all trying to mount the same NFS server, demonstrates this problem quite spectacularly.
Has this bug been fixed? -- Mike Shuey
