In my example case, there are 2 threads - one is a worker thread created by OpenThreads::Thread. The other thread is the main thread i.e. the thread that is intrinsically created when you execute the application. The crucial problem is that for the main thread, OpenThreads::Thread::CurrentThread() will return null.
I'll demonstrate this by breaking ReentrantMutex::lock() into sub-statements:
1.) if (_threadHoldingMutex==OpenThreads::Thread::CurrentThread())
2.) if (_lockCount>0){
3.)
OpenThreads::ScopedLock<OpenThreads::Mutex> lock(_lockCountMutex);
++_lockCount;
return 0;
4.)
int result = Mutex::lock();
if (result==0)
{
OpenThreads::ScopedLock<OpenThreads::Mutex> lock(_lockCountMutex);
5.)
_threadHoldingMutex = OpenThreads::Thread::CurrentThread();
_lockCount = 1;
return result;
An error will occur in the following case:
1) The worker thread calls lock(), it gets to the start of statement 5.
2) The main thread calls lock(). Statement 1 is evaluated as true as _threadHoldingMutex is null, and OpenThreads::Thread::CurrentThread() returns null.
3) The worker thread executes statement 5.
4) The main thread executes statement 2 and evaluates it as true, because the worker thread has set _lockCount to 1. The main thread executes statement 3, and now can access the mutexed-data at the same time as the worker thread!
The simple solution to this is to always protect access to _lockCount and _threadHoldingMutex using _lockCountMutex. I have done this in the file I am submitting."
"
and a later post the same osg-submissions thread:
"it's been a while since I have made the changes but I think it was due to problems with static builds of OpenThreads on windows. I was using
OpenThreads in a communication/synchronisation library (without
OpenSceneGraph). It seems I forgot to post a small change in the CMakeLists file of OpenThreads. If a user turns DYNAMIC_OPENTHREADS to OFF (static build) OT_LIBRARY_STATIC will be defined in the Config.
Without these changes a windows user will always end up with a "__declspec(dllexport)" or "__declspec(dllimport)" which is a problem for static builds."
And another post from Blasius on this topic:
"I tested with VS2005 and VS2008. For 32 bit everything works as expected. For x64 and VS2008 I could successfully do the cmake-configure and then the compilation but I had occasional crashes of cmTryCompileExec.exe (during the cmake-configure phase) which seems to be a cmake bug. With VS2005 and 64bit cmake does not set _OPENTHREADS_ATOMIC_USE_WIN32_INTERLOCKED although the interlocked functionality should be there. If I place the source snippet from the CHECK_CXX_SOURCE_RUNS macro to a separate sourcefile I can compile and run the resulting executable successfully. Forcing OPENTHREADS_ATOMIC_USE_WIN32_INTERLOCKED (on VS2005/x64) reveals a bug in "intrin.h" which seems to be fixed in VS2008 but not in VS2005.
In case anyone is interested the lines:
__MACHINEI(unsigned char _interlockedbittestandset(long *a, long b))
__MACHINEI(unsigned char _interlockedbittestandreset(long *a, long b))
__MACHINEX64(unsigned char _interlockedbittestandset64(__int64 *a, __int64 b))
__MACHINEX64(unsigned char _interlockedbittestandreset64(__int64 *a, __int64 b))
should be changed to:
__MACHINEI(unsigned char _interlockedbittestandset(long volatile *a, long b))
__MACHINEI(unsigned char _interlockedbittestandreset(long volatile *a, long b))
__MACHINEX64(unsigned char _interlockedbittestandset64(__int64 volatile *a, __int64 b))
__MACHINEX64(unsigned char _interlockedbittestandreset64(__int64 volatile *a, __int64 b))
The worst thing that can happen is that interlocked funtionality is not detected during cmake-configure and the mutex fallback is used.
Which reminds me another small glitch in the Atomic header so I attached a corrected version.
Why is the OT_LIBRARY_STATIC added to the config file? It is not needed anywhere.
OT_LIBRARY_STATIC is needed if you are doing static-builds on Windows. See my previous post on that.
"
implementation of the atomic increment and decrement into a implementation
file.
This way inlining and compiler optimization can no longer happen for these
implementations, but it fixes compilation on win32 msvc targets. I expect
that this is still faster than with with mutexes.
Also the i386 gcc target gets atomic operations with this patch. By using an
implementation file we can guarantee that we have the right compiler flags
available."
atomic related compile failures with msvs2005. Attached changes to make win32
really use the atomic stuff. There are pointer typecast problems and some
historic alignment restrictions that I just took from a previous similar
implementation of mine without looking deep enough. "
This change has been done to make it easier for OpenSceneGraph users to check out the svn via https
without any conflicts introduced with a http externals.
