Question

_{This question is related to: Is it possible to have several ORB objects in the same process?}

So, thanks to @BrianKelly I found information about the ORB identifier (even though there was no such information in all ORBACUS docs, that I have) and I successfully created a simple application, that connects to different CORBA servers and successfully executed several CORBA requests.

So far, so good.

Now, what I want to do, is to make this application multithreaded and to start a separate thread for the connection to the different servers. But ORB_init crashes.

Here’s a very short code, that I use for testing:

#include <OB/CORBA.h>

pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
void* run( void * );

struct config { const char* nameservice; const char* id; const char* exe; };

const bool mt = true;

int main()
{
    config cfg1 = { "NameService=corbaloc::10.102.8.15:13069/NameService", "1", "test" };
    config cfg2 = { "NameService=corbaloc::192.168.1.99:13069/NameService", "2", "test" };

    if( mt )
    {   
        pthread_t t1, t2;

        pthread_create( &t1, NULL, run, (void*)&cfg1 ); 
        pthread_create( &t2, NULL, run, (void*)&cfg2 ); 

        pthread_join( t1, NULL ); pthread_join( t2, NULL );
    }
    else
    {
        run( (void*)&cfg1 );
        run( (void*)&cfg2 );
    }

    printf( "SUCCESS!\n" );
    return 0;
}

void* run( void* arg )
{
    pthread_mutex_lock( &mutex );

    int argc = 2; char* argv[3];

    config* cfg = (config*)arg;
    argv[0] = (char*)cfg->exe;
    argv[1] = (char*)cfg->nameservice;
    argv[2] = NULL;

    CORBA::ORB_var m_varOrb = CORBA::ORB_init( argc, argv, cfg->id );

    pthread_mutex_unlock( &mutex );
    return NULL;
}

So, when mt is false, everything’s fine, I can extend the code to create some server specific objects, to execute different requests, etc. But then mt is true, the second thread fails calling ORB_init. See the stack trace below.

I’m pretty sure that I’m missing something very simple and stupid, but what?

$ g++ -g3 -ggdb -Wall -Wshadow -march=i486 
      -DUNIX -DLINUX -DPTHREADS -DMULTITHREAD -D_REENTRANT
      -I. -I/usr/local/include/OB/ -I/usr/local/include/JTC/ 
      -I/usr/include/OB/ -I/usr/include/JTC/ -L/usr/local/lib 
      -lpthread -lm -lz -lrt -ldl -lOB -lJTC -lCosNaming 
      test.cpp

Stacktrace:

#0  0x00566402 in __kernel_vsyscall ()
#1  0x0080dfd0 in raise () from /lib/i686/nosegneg/libc.so.6
#2  0x0080f9b1 in abort () from /lib/i686/nosegneg/libc.so.6
#3  0x03dc490b in ~RefCount 
    (this=Could not find the frame base for "~RefCount".) 
    at ../../include/OB/RefCount_Ts_Linux-x86-32.h:43
#4  0x03ef8965 in ORBInstance 
    (this=Could not find the frame base for "ORBInstance".) 
    at ORBInstance.cpp:276
#5  0x03f134fe in ORB_impl 
    (this=Could not find the frame base for "ORB_impl".) 
    at ORB_impl.cpp:281
#6  0x03f24740 in OBCORBA::ORB_init 
    (ac=Could not find the frame base for 
        "OBCORBA::ORB_init(int&, char**, OB::Properties*, 
                           OB::Logger*, OB::Reactor*, 
                           char const*, char const*)". ) 
    at ORB_init.cpp:994
#7  0x03f249d9 in CORBA::ORB_init 
    (ac=Could not find the frame base for 
         "CORBA::ORB_init(int&, char**, char const*, char const*)".) 
    at ORB_init.cpp:1014
#8  0x0804895d in run (arg=0xbfe8b544) at test_server.cpp:45
#9  0x007334d2 in start_thread () from /lib/i686/nosegneg/libpthread.so.0
#10 0x008b848e in clone () from /lib/i686/nosegneg/libc.so.6

Answer 1

I found something like a workaround. Makes my code really ugly and not easy for support, but it’s still something.

Here’s what I did:

• add a mechanism (in my application), that will count the necessary threads, before starting them
• read, in advance, the configurations – I need to know the necessary parameters for the naming services (used in ORB_init)
• before starting any threads, a “manager” will execute just once ORB_init, but it will pass several times the -ORBInitRef parameter, with different values – one for each thread/connection
• after this is done, the threads are started, but instead of executing ORB_init, they directly execute resolve_initial_references and continue with the server specific things

Note: My example does not containt resolve_initial_references, because the crash is in ORB_init.

---

So, applying this “algorithm” for this “workaround” would look like:

#include <OB/CORBA.h>

void* run( void * );
CORBA::ORB_var varORB;

int main()
{
    /** The necessary configurations */
    //-------------------------------------v
    const char* nameservice1 = "NameService1=corbaloc::10.102.8.15:13069/NameService";
    const char* nameservice2 = "NameService2=corbaloc::192.168.1.99:13069/NameService";
    //-------------------------------------^

    /** INIT the ORB **/
    int argc = 5; char* argv[ 6 ];
    const char* initref = "-ORBInitRef";
    const char* exe = "test";

    argv[0] = (char*)exe;
    argv[1] = (char*)initref; argv[2] = (char*)nameservice1;
    argv[3] = (char*)initref; argv[4] = (char*)nameservice2;
    argv[5] = NULL;

    varORB = CORBA::ORB_init( argc, argv );

    pthread_t t1, t2; 

    char ns_id1 = '1', ns_id2 = '2';
    pthread_create( &t1, NULL, run, (void*)&ns_id1 );  
    pthread_create( &t2, NULL, run, (void*)&ns_id2 );  

    pthread_join( t1, NULL ); pthread_join( t2, NULL );

    varORB->destroy();

    return 0;
}
void* run( void* arg )
{
    char nameservice[] = "NameServiceN";

    // set the right number of the nameservice
    nameservice[ 11 ] = *((char*)arg);  

    varORB->resolve_initial_references( nameservice );

    // do some CORBA-specific stuff

    printf( "SUCCESS %c\n", *(char*)arg );
    return NULL;
}

---

NOTE

I still can’t believe this is the only option. If you look at my code (in the question) carefully, you’ll see, that:

• it IS possible to have multiple ORBs (see the case with mt == false)
• the call to ORB_init IS synchronized
• the ORB identifier IS implemented and it works fine (again with mt == false)

So, this is not actual answer to my question, it’s kind of a workaround.

It doesn’t make sense (at least to me) to be possible to create several ORBs in a single thread, but not in multiple threads.