[omniORB] Server hangs when client crashes

[bjornw@fairplay.no]

Sai-Lai Lo <S.Lo@orl.co.uk> writes:

You are corrent about the dce-threads.

I said I had a very simple server (which was not totally correct). I have set up
several threads at my server and one queue where I place queries (requests
from clients) to be performed. This queue was wrapper in a omni_mutex that used
a omni_condition. Each thread waited for the condition to give a signal back to
the actual thread. Well, this did not work for HPUX-10.20 but for linux RedHat-5.1
it worked just great. So it seems like our concern about upgrading to HPUX-11.0
just got upgraded.... (well, since sybase has release sybase for linux, we might
just throw our hp out of the window :-/ ..... )


> I want to understand why this is happening.
> 
> Firstly, I assume you are using HP's C++ compiler and not egcs/gcc. If not
> I suspect it is most likely that the compiler is generating wrong code. So
> far I have not received positive confirmation that egcs/gcc on HPUX is usable
> with omniORB2 (which can be seen as a non-trival multithreaded C++ testsuite).
> 
> Secondly, one factor you have to consider is the fact that HPUX-10.x is
> using dce-threads. There is no kernel thread support. Therefore all the
> blocking I/Os are done via a wrapper within the pthread library to give the
> illusion that each thread can block on a socket independently. Correct me
> if I'm wrong- I think the wrapper would only come into play if you have
> included pthread.h in all your {.c/.cc} files that contain blocking I/O
> calls. This is how dce-threads on Ultrix used to work. Now if your main
> thread calls into a library independent of omniORB2 and if that library
> does a blocking I/O call and if that library was not compiled with
> pthread.h, your whole process would be blocked until that call returns! I
> wonder if this situation applies to your application.
> 
> Finally, your system may not have all the OS patches related to dce-threads
> installed. Someone from HP once admitted to me that the pthread
> implementation in HPUX 10-x is poor. He promises that HPUX 11 is much
> better. Let wait and see. 
> 
> Sai-Lai
> 
> >>>>> bjornw  writes:
> 
> > (I'm running the latest omniorb-snapshot on HPUX-10.20, with dce-threads)
> 
> > When running two or more clients against one server, and one of the client
> > crashes, the server blocks.
> 
> > The server implements one simple method like this:
> > query_result *
> > ServerFactory::Query(int id)
> > {
> >  // create a result
> >  return result;
> > }
> 
> > When one of the client stops, it seems like the server blocks in the
> > return-statement while trying to send to the client that recently
> > crashed.
> 
> > Does this mean that the server does not create a new thread per
> > client-connection?  Or, better how can I prevent this from happening.
> 
> -- 
> Dr. Sai-Lai Lo                          |       Research Scientist
>                                         |
> E-mail:         S.Lo@orl.co.uk          |       Olivetti & Oracle Research Lab
>                                         |       24a Trumpington Street
> Tel:            +44 223 343000          |       Cambridge CB2 1QA
> Fax:            +44 223 313542          |       ENGLAND