Alan Bateman wrote:
....It still seems very fragile and doesn't take into account that the number of file descriptors might legitimately increase. How about changing the test to use ulimit on Solaris/Linux to set the hard limit to a smallish limit (say 1024) and then run and accept >1024 connections. If the original leak were to come back then it should fail reliably.
I've updated the webrev as per your suggestion. http://cr.openjdk.java.net/~chegar/6935233/webrev.00/webrev/ -Chris.
-Alan.
