MR29v1 - mfplat: Avoid a race condition when releasing a tracked sample.

[Nikolay Sivov (@nsivov)]

I don't think this works if RtwqInvokeCallback() fails. The point of checking first and decrementing later is because of a circular reference between tracked_result and the sample.

In normal case after SetAllocator() you get refcount == 2 and tracked_refcount == 1. Now it queues are not working:

* Release() decrements to refcount == 1 and tracked_refcount == 1;
* condition matches and tracked result it released, recursing back to sample_tracked_Release();
* recursed call decrements refcount to 0 and releases the sample.

So you're losing 2 references on a single Release() call. I supposed decrementing while still locked will work?

P.S. note that there is almost identical copy of this in dlls/evr/sample.c that needs the same fix, and probably use of an explicit enter/leave instead of a LockStore().

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/29#note_567