Memory corruption when enabling CCD in a multi threaded world

[jrouwe]

When configuring a body to have a non 0 setCcdSweptSphereRadius and setCcdMotionThreshold in combination with a multi threaded world (btDiscreteDynamicsWorldMt) the following callstack corrupts memory:

btCollisionDispatcherMt::getNewManifold(const btCollisionObject * body0, const btCollisionObject * body1) Line 69 C++
btDiscreteDynamicsWorld::createPredictiveContactsInternal(btRigidBody * * bodies, int numBodies, float timeStep) Line 905 C++
btDiscreteDynamicsWorldMt::UpdaterCreatePredictiveContacts::forLoop(int iBegin, int iEnd) Line 101 C++
ParallelForJob::executeJob(int threadId) Line 122 C++
WorkerThreadFunc(void * userPtr) Line 371 C++
win32threadStartFunc(void * lpParam) Line 241 C++

At this time btCollisionDispatcherMt::m_batchUpdating is false, so multiple threads are racing on this piece of code:

manifold->m_index1a = m_manifoldsPtr.size();
m_manifoldsPtr.push_back(manifold);

This sometimes causes an incorrect index to be written into the manifold and sometimes it corrupts the array.

[jrouwe]

Note for a repro, take the code from here:

https://github.com/jrouwe/BulletPerformanceTest

And in PerformanceTest.cpp change 'for (int mq = 0; mq < 2; ++mq)' into 'int mq = 1;' and change 'for (int num_threads = 1; num_threads <= (int)thread::hardware_concurrency(); ++num_threads)' into 'int num_threads=8'.

[erwincoumans]

Thanks for the report. Can you check if this PR helps?
#4119

You can increate the performance slightly by setting the s_minimumContactManifoldsForBatching as follows:

	// Create solver
		gSolver = new btSequentialImpulseConstraintSolverMt();
		btSequentialImpulseConstraintSolverMt::s_minimumContactManifoldsForBatching = 2000;

Note that Bullet isn't as well optimized as PhysX or Jolt.

[jrouwe]

That fixes the issue, thanks!