This is the mail archive of the pthreads-win32@sourceware.org mailing list for the pthreas-win32 project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: pthread_join problem


Stefan Eilemann wrote:
Hello,

I am in the situation that a pthread_join does not return, even
though the thread has called pthread_exit.

I read the cleanup notes, but I think it does not apply here.

I am using the C cleanup code. One thread calls pthread_exit,
the other phtread_join. I've verified that the thread calling
pthread_exit does the longjmp to the thread start code, which
calls _endthreadex.

The main thread calling pthread_join does hang in
WaitForMultipleObjects.

The problem only occurs when I am using some unrelated(?)
external code (the Mellanox SDP Infiniband implementation),
so it could be caused by that, or just be a race appearing
with this code.

There are other pthreads in my application, which terminate
correctly with pthread_exit/pthread_join. Only one thread -
the network receive thread ;)- does exhibit the problem.

Do you have an idea what could be the cause of this problem?
Anything else I could try to find the problem?
One thing that comes to mind that seems to fit the evidence assumes that the external code that you mentioned is a DLL and it executes it's own dllMain routine which somehow interferes with pthread-win32's thread exit cleanup. This would be occurring after _endthreadex() is called, which you've verified is called.

I don't know how Win32 determines which and in what order these dllMain's are called (is it the order the DLLs are loaded?), but pthreads-win32 does rely on this mechanism to do some final cleanup and status setting for each 'POSIX' thread, and if this doesn't get done I imagine it's possible you would see symptoms like this.

Pthreads-win32's dllMain() calls pthread_win32_thread_detach_np() in pthread_win32_attach_detach_np.c. To verify that this is happening you could set up a thread-specific data key and give it a destructor routine, have your problem thread set it to a non-null value, and then see if the destructor routine is called.

Regards.
Ross

Best Regards,


Stefan.

PS: I've tested the Win64 version, and it works like a charm.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]