When expanding the old gen heap during young gc, that section of code is guarded by the ExpandHeap_lock. If multiple threads are out of memory (i.e. old gen PLAB), they will queue up on that lock.

Since the threads after the first getting the ExpandHeap_lock do not retry the allocation, they will blindly expand the heap again, ultimately causing a "heap expansion storm".

This not only takes lots of time, but also expands the heap unnecessarily.