Contact data race in MT about physx HOT 1 CLOSED

Hi @shelomentsev

thank you for reporting this issue and providing a set of patch file for reproducing the problem. I'm sorry it took a while to get back to you but at least here are good news:

I have applied your patches to the newest version of the PhysX SDK (to be updated here on Github soon) and I could not reproduce the issue anymore (*). In the meantime, we have independently discovered and fixed a problem that had most likely the same root cause: In Sc::Scene::islandGen the task dependencies were incorrectly configured. They should be processNarrowPhaseTouchEventsStage2(&mUpdateDynamics); instead of processNarrowPhaseTouchEventsStage2(&mPostSolver);, i.e., the updateDynamics task (which runs the solver) needs to wait for processing the touch events, otherwise the solver writes the post-solve velocities into the bodies while the contact callbacks are still extracting information from the same location. You can apply this hotfix to the code right now and then upgrade to the new release once it's out.

(*) Meaning the printouts from your __CheckContacts function never appear. Note that your mutex-based detector is not quite correct and kept showing up false positives for me because for multiple islands, there may be multiple concurrent instances of the PxsSolverStartTask which operate on independent data. The tryLock() check kept triggering for me with multiple threads being in PxsSolverStartTask. Instead, I used an improved check where I redefined the mutex variable as a counter volatile int __shapeInterationLock;. Then in InteractionNewTouchTask replacae the lock/unlock with PxAtomicIncrement(&mNphaseCore->getScene().getSimulationController()->__shapeInterationLock); and PxAtomicDecrement(&mNphaseCore->getScene().getSimulationController()->__shapeInterationLock);.

and instead of the tryLock() check in PxsSolverStartTask I only did
if(mContext.mSimulationController->__shapeInterationLock > 0) PxDebugBreak();

from physx.