possible causes for EADDRINUSE in client code

[Steven Seagull <steven7779@gmx.net>]

Hi!

I'm running a 2.6.26 kernel where I have to create short-lived TCP
connections with a high frequency. In order to do this, I am using
non-blocking sockets which are bound to different IP addresses in a
round-robin manner, and to port 0, so the kernel can choose a free
local port.

Now I observed that I am getting EADDRINUSE errors on the bind system
call, even though there should be ports left according to the current
net.ipv4.ip_local_port_range sysctl setting.

I tried to enumerate the used ports at the time of the error using
netstat, which shows only rather few ports being in use, but I
realize that this is an inaccurate measure because it takes time for
netstat to be forked and run.

What I wonder is: Is there any limit beyond
net.ipv4.ip_local_port_range which may lead to EADDRINUSE errors?
Maybe any limits in the kernel data structures I should be aware of
when working with huge amounts of TCP connections?

Thanks!

--
Steven