Re: bind() should not return -EADDRINUSE

From: Phillip Susi <redacted>
Date: Thu, 11 Dec 2014 14:50:48 -0500

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 12/11/2014 2:23 PM, David Miller wrote:

quoted

From: Phillip Susi <redacted> Date: Thu, 11 Dec 2014
14:19:11 -0500

quoted

Attempting to establish a connection to a remote host using the
same local port, but a different remote port as the previous
connection ( that is still in TIME_WAIT ) results in bind()
returning -EADDRINUSE. By changing the remote port, you avoid the
conflict with the other connection that is in TIME_WAIT, but
since the remote port is not known when bind() is called, it
incorrectly returns -EADDRINUSE.  This check should not be done
in bind(), but deferred until the remote port is known in the
call to connect(), or listen().

Bind has to allocate and hold from everyone else on the system the 
local port at bind() time, so we cannot defer this decision.

What on earth for?  If two processes are going to connect to different
remotes using the same source, that is perfectly fine.  The only
contention is if two processes want to listen() with the same local
addr and wildcard remote.

But you don't know ahead of time what the processes are going to
do, that's the problem.

You cannot leave the port available and pretend to another process
that he will be able to use it.

Port allocation failures must be signalled at bind() time.