On the C interface, callbacks, and global roots

[Thomas Fischbacher <Thomas.Fischbacher@Physik.Uni-Muenchen.DE>]

Suppose I have an opaque C object obj, for which there is a function that 
allows me to register a C callback function f_cb (and a void* closure data 
pointer, let's call it f_cb_data). 

In principle, it is possible to lift the callback-registering up to the ML 
level, so that we can register a ML function as a callback for the 
ML-wrapped object. I suppose that the proper way to do this is to 
internally represent obj as a ML custom block with a special finalizer 
which will hold (1) a C pointer to obj, and (2) a C pointer to malloc()'d 
space holding a global GC root pointing to the ML callback function. The 
static-in-memory address of malloc'd space will then hold the ML callback 
function. So far, so good.

But - how should one proceed if f_cb's signature is such that it will 
receive obj as its first argument? Evidently, if we store the ML 
variant of f_cb (which will be contained somewhere in f_cb_data, and f_cb 
will be a wrapper executing the ML callback stored in f_cb_data) in a 
global root, this will make obj non-collectable.

The problem is that we should not place something under a global root 
which contains a reference back to the object responsible for 
un-registering the global root in its finalizer.

As far as I can see, the cleanest solution to this dilemma would be to 
introduce yet another type of ML block with a special meaning to the 
garbage collector: one that holds a pointer to nonmoving non-heap memory 
that can hold ML values, plus additional information on the number of 
such ML values, and a pointer to a de-allocation function (with its own 
closure arg, I'd say).

Seen in a different way, the problem of creating non-reclaimable values 
even exists if the ML callback function happens to be a closure containing 
the ML-wrapped obj. Somehow, the whole idea of using global roots like 
this feels a bit wrong to me. Did I miss something essential? How should 
one rather proceed instead?

-- 
regards,               tf@cip.physik.uni-muenchen.de              (o_
 Thomas Fischbacher -  http://www.cip.physik.uni-muenchen.de/~tf  //\
(lambda (n) ((lambda (p q r) (p p q r)) (lambda (g x y)           V_/_
(if (= x 0) y (g g (- x 1) (* x y)))) n 1))                  (Debian GNU)