This prevents two different kinds of client errors from causing
undetected memory corruption, and helps with the detection of others:
1. We no longer deallocate objects when the weak count goes to zero
and there have been no strong references. This otherwise causes
us to return a garbage object from a constructor if the constructor
allocates and deallocates a weak pointer to this. And we do know
that clients allocate such weak pointers in constructors and their
lifetime is hard to trace.
2. We abort if a RefBase object is explicitly destroyed while
the weak count is nonzero. Otherwise a subsequent decrement
would cause a write to potentially reallocated memory.
3. We check counter values returned by atomic decrements for
plausibility, and fail immediately if they are not plausible.
We unconditionally log any cases in which 1 changes behavior
from before. We abort in cases in which 2 changes behavior, since
those reflect clear bugs.
In case 1, a log message now indicates a possible leak. We have
not seen such a message in practice.
The third point introduces a small amount of overhead into the
reference count decrement path. But this should be negligible
compared to the actual decrement cost.
Add a test for promote/attemptIncStrong that tries to check for
both (1) above and concurrent operation of attemptIncStrong.
Add some additional warnings and explanations to the RefBase
documentation.
Bug: 30503444
Bug: 30292291
Bug: 30292538
Change-Id: Ida92b9a2e247f543a948a75d221fbc0038dea66c
Add some basic tests for RefBase, as well as a more ambitious memory
ordering test.
Add a README.txt with instructions to run the tests.
Comment out a couple of BlobCache tests that failed consistently and
appeared to be incorrect. With that fix, I managed to run
libutils_tests successfully on device.
Bug: 28705989
Change-Id: I8ad29995097a149a0cc38615d6ed37117ec6cb5c
Hans Boehm