doc: define how finalizers interact with memory model

[dvyukov]

Finalizers introduce concurrency, but the synchronization between goroutines accessing the object and the finalizer are not defined. The proposed wording is along the lines of "SetFinalizer happens before execution of the finalizer starts". This makes the following program correctly synchronized:
http://play.golang.org/p/arA1niVAZC
However, the following program contains a data race on x.y and *x.y:
http://play.golang.org/p/cpuvxDu6AW
The program indeed can be compiled in a way that makes the panic to fire. Namely, compiler reorders "x.y = &i" and "*i = 42". This result can be unexpected.

[ianlancetaylor]

I know there is a certain amount of confusion at present about SetFinalizer, but I'm not sure your proposal is strong enough. I think that this program is race free:
http://play.golang.org/p/WdFVb3p7du

In words, if I have a pointer, then the pointer can not be finalized while I'm in the process of using it to assign a value. That is stronger than merely saying that SetFinalizer happens before the execution of the finalizer.

One possibility would be "given a pointer p, any read or write of *p happens before p's finalizer is run." But that is too strong, as it prohibits the compiler from reordering any operation across the use of a pointer.

I feel like what we want here is an inverse relationship. Given a pointer p, there is no possible ordering such that the finalizer happens-before any read or write of *p. But I don't know if that makes any sense.

[ianlancetaylor]

Moving to unplanned.

[dvyukov]

One possibility would be "given a pointer p, any read or write of *p happens before p's finalizer is run." But that is too strong, as it prohibits the compiler from reordering any operation across the use of a pointer.

Indeed. Also if you have p.f.foo(), you generally don't know whether you access *p or **p and that can change over time. So a guarantee only for *p becomes too subtle to be used reliably. I also suspect there are more optimizations it prohibits as it effectively makes almost any store a synchronization operation. Also, if I just load from an object, but they result is eventually unused, does it order finalizer?

FWIW Java makes exactly the guarantee I described.

I feel like what we want here is an inverse relationship. Given a pointer p, there is no possible ordering such that the finalizer happens-before any read or write of *p. But I don't know if that makes any sense.

It makes sense, but it states effectively what I proposed :)
If a finalizer can't happen-before, then it happens either after or concurrently. But if it potentially can happen concurrently, then you need to synchronize accesses to prevent data races.