[zeromq-dev] zmq-signaler-port-sync

[Pieter Hintjens]

Hi Pau,

So there are two different problems here, one is that we're hitting a
socket limit on WXP, and the other is that the asserts are happening
inside a critical section.

I don't think we can fix the first one easily but we can presumably
assert in a smarter way. Do you want to try making a patch for this?

-Pieter

On Mon, Jan 21, 2013 at 6:23 PM, Pau <pau at teleopsia.com> wrote:
>
> Hi,
>
>
> I am using (not yet in production) ZMQ on Windows and I have found what
> I think is a big problem for Windows users.
> We use WXP and W7 and Visual C++ different versions. ZMQ version 3.2.0
> (as far as I see the same problem happens in 3.2.2)
>
> I do not fully understand ZMQ internals but I've seen that every time a
> socket is created the function make_fdpair(..) is called and in
> signaler.cpp(line244) a system event "zmq-signaler-port-sync" is created.
> This event is used as a system-wide critical section and, so all
> applications that try to create an event will WaitForSingleObject (sync,
> INFINITE) until  SetEvent (...) is called.
> The problem is that the code between:
>   HANDLE sync = CreateEvent (NULL, FALSE, TRUE, TEXT
> ("zmq-signaler-port-sync"));
> and
> SetEvent (sync);
> is full of wsa_asserts(..) that will terminate the application if
> something goes wrong.
>
> It is clear that terminating the application not leaving the system-wide
> critical section is a bad idea because all applications in the system
> will hang and you will have to stop all them to start again.
> I understand that no errors should happen but anyway to escape from the
> error is not a good idea in this case.
>
> I do not know all possible reasons to generate a fatal wsa_assert(..)
> but there is at least one:
>
> I have seen that in XP it is possible that line 301  rc = connect (*w_,
> (sockaddr *) &addr, sizeof (addr)); returns an error when a socket tries
> to connect to 5905 and this has happened many times.
> Windows uses port numbers starting near 1400 and XP has a limit at 5000.
> In W7 this does not look as a problem because maximum is 65000
> It sounds as if the number was big enough but apart from the fact that
> ZMQ uses a big number of connections (at least in my tests) I have
> experienced that Windows jumps port numbers by 7, so 5000 happens
> sometimes with catastrophic consequences.
>
> best,
>
> Pau Ceano
> _______________________________________________
> zeromq-dev mailing list
> zeromq-dev at lists.zeromq.org
> http://lists.zeromq.org/mailman/listinfo/zeromq-dev