Ensure pointers indirected from Memory and pointing into Memory should retain originating object

[matthiasblaesing]

The use case is this:

• native supplies the required size of a memory region
• a Memory object is created with the right memory size
• native fills the memory with a structure and additional objects,
  that are held in that region, but are only referenced from the
  structure (one such example would be a Structure.ByReference)

As long as the resulting structure stays strongly referenced from Java,
all is good. When the toplevel structure goes ouf of scope, the memory
backing the strucure is not strongly referenced anymore and will be
freeed by GC.

This change fixes the issue by ensuring, that the pointers used in
substructures are SharedMemory objects, holding a strong references to
the original Memory object.

[matthiasblaesing]

@dbwiddis could you please have a look and give it a try/get it tested? This replaces #1325, at least it should work and fix more than the single observed case.

[dblock]

This is A+, very nice work @matthiasblaesing. I was following the discussion out of curiosity, and thinking there has to be a better way to do this, and this is it.

[dbwiddis]

I really like this solution. A few suggestions inline.

[dbwiddis]

Testing: Set up an Azure VM (Windows Server 2012 R2) and created a test method exercising the current (Advapi32Util.getTokenAccount) code in a loop, running as admin and collecting usernames/domains of the user of every running process. Ran with very small heap (2500K) to maintain pressure on the GC. OpenJDK 15.0.2.

• Ran existing code until exceptions 4 times. Exceptions at ~13000, ~8000, ~88000, and ~842000 iterations.
• Ran the code fix in PR Ensure memory object backing TOKEN_USER is retained until account extraction is done #1325. No failures in 1 million iterations (4 times).
• Ran the code fix in this PR. No failures in 1 million iterations (4 times).

[matthiasblaesing]

@dbwiddis @dblock thank you both for review. I pushed an update, that should cover the points raised (I plan to partially squash the commits prior to merging, but I want to preserve review history until done)

[dbwiddis]

LGTM!

[dblock]

Looks great, merge at will.