[Bug 50955] .netCore app can't bind to port shortly after another .netCore program binding to the same port was terminated

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=50955

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[e-mail]

--- Comment #4 from [e-mail] ---
Created attachment 71900
  --> https://bugs.winehq.org/attachment.cgi?id=71900
Wine test suite patch to (hopefully) reproduce this issue

ZusiDisplay is affected by this when a train reverses at a station and
embedded/integrated displays are re-started by Zusi.

The attached wine testsuite patch reproduces this in different ways (but it
fails on the Wine Testbot on Debian, probably due to a race condition where the
child process takes longer than 100ms to open the listening socket). After a
rather complex first attempt that actually spawns child processes to stay close
to what ZusiDisplay and the BrokenClient/BrokenServer reproducer does, I found
an easier way:

The same / a very similar issue occurs with just one single-threaded test case
(called "test_port_reuse()") that sets up a listening socket, connects to that
socket, then closes the connection and the listening socket, then does it again
using the same listening port. On Windows, the test passes. Using wine on
Linux, the second attempt fails because the port is still "in use" (actually in
TIME_WAIT state).

It seems like TIME_WAIT on Windows doesn't affect listening at all, so
applications are free to bind to and listen on a port just moments after the
previous listening socket was closed. It looks like the Windows kernel just
takes extra care to make TCP work despite TIME_WAIT, compared to Linux. As I
understand it, TIME_WAIT should prevent new connections using the same
<local_ip:local_port, remote_ip:remote_port> combination, so just listening on
a port is not an issue as no connection has been made at that point. The remote
end will probably use a different port next time (or the connection can be
refused otherwise, or even accepted using a larger-than-before initial TCP
sequence number).

I'm not sure if there is a good way to emulate this behavior in wine. One way
that comes to mind is to set the SO_REUSEADDR socket option for AF_INET sockets
before the call to bind() in wineserver . However, this will break other test
cases because they expect bind()ing to the same address twice to fail. So the
current wine behavior is too restrictive compared to Windows, but always using
SO_REUSEADDR is too permissive, and I'm not sure if there is any sensible
middle ground. Maybe being too permissive is acceptable in this case? I still
need to test if that actually fixes the ZusiDisplay issue (via Proton) though.