accept

This patch should solve the problem, servers do not shutdown itselfs but are
teared down from within run_test.pl. Applying the patch run_test.pl succeeds.

Index: Servant_Activator.cpp
===================================================================
RCS file:
/project/cvs-repository/ACE_wrappers-repository/TAO/examples/POA/Forwarding/Servant_Activator.cpp,v
retrieving revision 1.7
diff -u -u -b -r1.7 Servant_Activator.cpp
--- Servant_Activator.cpp	22 Aug 2002 18:08:04 -0000	1.7
+++ Servant_Activator.cpp	13 Apr 2006 06:52:41 -0000
@@ -19,10 +19,6 @@
   ACE_THROW_SPEC ((CORBA::SystemException,
                    PortableServer::ForwardRequest))
 {
-  this->orb_->shutdown (0
-                        ACE_ENV_ARG_PARAMETER);
-  ACE_CHECK_RETURN (0);
-
   // Throw forward exception
   ACE_THROW_RETURN (
                     PortableServer::ForwardRequest (
Index: run_test.pl
===================================================================
RCS file:
/project/cvs-repository/ACE_wrappers-repository/TAO/examples/POA/Forwarding/run_test.pl,v
retrieving revision 1.15
diff -u -u -b -r1.15 run_test.pl
--- run_test.pl	5 Mar 2001 03:23:46 -0000	1.15
+++ run_test.pl	13 Apr 2006 06:53:37 -0000
@@ -53,9 +53,9 @@
 }
     
 $client  = $CL->SpawnWaitKill (60);
-$server1 = $SV1->WaitKill (10);
-$server2 = $SV2->WaitKill (10);
-$server3 = $SV3->WaitKill (10);
+$server1 = $SV1->Kill ();

+$server2 = $SV2->Kill ();

+$server3 = $SV3->Kill ();

 
 unlink $iorfile1, $iorfile2, $iorfile3;
     

Frank, killing the server from the script is something I consider as bad
practice.  First tools like valgrind/purify then can't correctly give results,
also on for example VxWorks this is not possible. Is it maybe possible to just
run the orb.run for some time and then let the server exit itself?

AFAICS, calling orb.run with timeout might succeed in most cases, but in few
cases thread might return from ORB context and terminate (closing connections)
before client was able to read data from socket. This might happen even if
server-thread checks a "terminate"-flag within server frequently, which might be
set in the moment the forwarding has happened. The problem is that we dont know
how long to wait, to make sure client read GIOP reply from socket successfully.
AFAICS, at best server-termination should be synchronized with "close
connection" initiated by client (eg. its termination).

But, sleeping a few seconds would do also the job, just to make sure, client is
able to read reply from socket within specific time before closing the
connection, eg:

  orb->run(ACE_ENV_SINGLE_ARG_PARAMETER);
  ACE_TRY_CHECK;

+ // sleep few seconds to make sure client has sufficient time to read
+ // reply from connection stream
+ ACE_OS::sleep (1);

  orb->destroy (...


Sniffing the traffic, another thing that makes me wonder is, that the
terminating server (executed on WinXP) is doing a TCP-RST instead of TCP-FIN.
This might happen if orb.destroy does not terminate the existing connection and
the operating system is forced to tear down the resource. Or (maybe) it might
happen if orb.destroy just calls ::close instead of combination
::shutdown(SHUT_RDWR)/::close per connection.

Created attachment 517 [details]
2 network sniffs, client interacting with server3, the first server doing a forward

Add Frank to CC list, 
BTW the network sniff is gzip-ed.

I've committed a fix to the doc head. The problem was that the ServantActivator
was inheriting from PortableServer::ServantActivator, but not also from
TAO_Local_RefCounted_Object. It was then expecting refcounted behavior, ie
having the activator cleaned up automatically. 

Adding the refcounting makes this test work.

moved this example to the tests directory and enabled it on our daily test suite so that we can see any failure on the scoreboard

Updated summary to reflect new location of this example

test stats shows this works on head now, so closing