java - Finalization of object fields

Question

I'm having difficulty understanding the implications of sections 12.6.1 and 12.6.2 of the Java SE 8 Language specification. I am working with a product in which Java objects manage native peers so getting finalization right (until we get a chance to do a ref-queue rewrite) is important.

It is clear, from the spec, that finalizers can be run out of order. It is reachability with which I'm having a hard time.

I believe that the spec says this:

class A {
    public Object o = new Object()
    protected synchronized void finalize() throws Throwable { o = null; }
}

class B {
    A a = new A()
    protected void finalize() throws Throwable {
        a.getClass()   // always works: a cannot be null.
        a.o.getClass() // might NPE: a's finalizer might have been run
    }
}

class C {
    A a = new A()
    protected void finalize() throws Throwable {
        synchronized (a) {
            a.getClass()   // always works: a cannot be null.
            a.o.getClass() // always works: a.o cannot be null.
        }
    }
}

There are 4 assertions, above. I would very much appreciate confirmation that they are true, or explanations for why one or more is false.

Answer 1

If A.finalize acquires the lock first, then a.o.getClass() will NPE even in C. The final assertion is incorrect.

In terms of reachability think of it like this: Suppose A and C were cyclically reachable, then the would reach unreachability at the "same time". So, it cannot reasonably be that A must be finalised before C.

I guess a direct fix could be reference counting of A. However, relying on GC to clear up non-memory resources isn't going to be great.