Use AbstractCASReferenceCounted to ensure entry.retain() is valid

[merlimat]

Motivation

This will fix the root case of #2993 by using a version of AbstractReferenceCounted based on compareAndSet() rather than optimistic counter increase. This would ensure that when calling entry.retain(), if we don't get an exception the entry is guaranteed to be valid.

More detailed discussion at netty/netty#8563

[rdhabalia]

@merlimat did you get chance to test this change with perf-producer and multiple subscription consumers.?

[merlimat]

@rdhabalia No, I haven't. I think to trigger the issue also the cache need to run eviction at the same time

[merlimat]

run integration tests

[massakam]

rerun integration tests

[merlimat]

rerun integration tests

[sijie]

@merlimat shall we also adopt this in bookkeeper? since we are using that for ByteBufList. ByteBufList is a core structure in v2 protocol.

[merlimat]

@merlimat shall we also adopt this in bookkeeper? since we are using that for ByteBufList. ByteBufList is a core structure in v2 protocol.

I don't think it should be a problem there because when we call increase() we always have refCnt >= 1, therefore there's never the race condition on the release() getting to 0

[merlimat]

run integration tests

[hrsakai]

run integration tests

[massakam]

run integration tests