Interoperability between Windows and Linux is bugged

Alexandros Panagiotidis — Thu, 03 Sep 2009 11:42:33 -0000

Hello Equalizer-Team,

I tried setting up the following scenario using the 0.9 release and HEAD-revision of Equalizer (checked out 02.09.2009):

- server on cluster headnode running "Linux XXX 2.6.18-92.1.10.el5_lustre.1.6.6custom #1 SMP Mon May 25 17:23:20 CEST 2009 x86_64 x86_64 x86_64 GNU/Linux"
- client and applicatiuon on Windows XP x64 SP2

Equalizer was built as 64-bit.
On the Linux machine GCC 4.3.2 is used. On Windows I used Visual Studio 2005 8.0.50727.42 to build and debug Equalizer.
The following programm outputs "1" on both machines, which is why I assume they are both little-endian.

#include <stdio.h>

int main() {
unsigned char EndianTest[2] = { 1, 0 };
short x;

x = *(short *) EndianTest;

printf("%d\n", x);

return 0;
}

Using the attached configuration for eqServer I received the error "Node not found on server".
Trying to fix this problem, I did some debugging in Visual Studio. Essentially, the offender for this error seems to be in Session::_cmdGetIDMasterReply():

NodeID nodeID = packet->masterID;
nodeID.convertToHost();

if( nodeID != NodeID::ZERO )
{
ScopedMutex mutex( _idMasterMutex );
_idMasters[ packet->id ] = packet->masterID;
}
// else not found

The nodeID that the master returns is converted via "nodeID.convertToHost()" but the unconverted ID from the original answer is used inside the if-case.
Applying the following patch gets rid of the above error:

713c713
> _idMasters[ packet->id ] = packet->masterID;
---
< _idMasters[ packet->id ] = nodeID;

After patching and recompiling on both machines I get the error:

3344 3976656 ..\lib\net\session.cpp:359 4380 Can't find master node for object id 0
3344 3976656 ..\lib\client\pipe.cpp:361 4380 Assert: eqResult [session->mapObject( view, viewVersion.id )]

At this point I gave up debugging and tried running the client and application under Linux on the Windows machine:

- eqServer on cluster headnode (see above)
- eqPly as client on "Linux XXX 2.6.27.25-0.1-default #1 SMP 2009-07-01 15:37:09 +0200 x86_64 x86_64 x86_64 GNU/Linux"
- eqPly as application on the same machine as the client

In this setup no errors occured with and without patching Session::_cmdGetIDMasterReply().
A colleague of mine noticed that under Linux "UUID::convertToHost()" and "UUID::convertToNetwork()" are not implemented but should be.
I also assume that this is related to the error in the first setup but we did not investigate any further as we settled for running on Linux only for now.

Attached is my Equalizer configuration and several logs of server, client and application, both unpatched and patched and running client/application under Windows and Linux.
Due to privacy reasons I had to remove hostnames and IP-addresses, the rest of the log is unchanged.
The binaries were always called like this:

Server: eqServer test.eqc
Client: eqPly -- --eq-client --eq-listen CLIENT:4242 --eq-server LINUX-CLUSTER-HEAD:4242
Application: eqPly -m screwdriver.ply -- --eq-server LINUX-CLUSTER-HEAD:4242

I also included the call stack from Visual Studio of the second error, where I stopped debugging.

If it helps, I got on to the error in Session::_cmdGetIDMasterReply() because the bytes of the node IDs were swapped in the log.
If you need any more information or testing regarding this issue feel free to contact me anytime.

Kind regards, Alex

Recent changes to support-requests

Interoperability between Windows and Linux is bugged