[Cffi-devel] Fw: Fw: Help perfecting my GC finalizer/constructor

[Martin Simmons]

>>>>> On Tue, 29 Apr 2014 07:20:08 +0200, Willem Rein Oudshoorn said:
> Cancel-Lock: sha1:p0/p2S+buMdADzjMLfHl6LN5m6k=
> 
> Martin Simmons <martin at lispworks.com> writes:
> 
> >>>>>> On Mon, 28 Apr 2014 07:26:07 +0200, Willem Rein Oudshoorn said:
> >> Cancel-Lock: sha1:zqpk08tO/PaCrz3PLTlI+ocAF9E=
> >> 
> >> Joeish W <joeish80829 at yahoo.com> writes:
> >> 
> >> > How do I work with the metaobjects(is that what they are called)
> >> > output by the code you gave me...If you can show me how to mem-aref
> >> > one I would really appreciate it
> >> 
> >> Sorry for the short answer, but at the moment I am extremely busy.
> >> If you followed the example
> >> 
> >>  (mem-aref (c-pointer YOUR-POINT) ...)
> >> 
> >> should work.
> >
> > It might be interesting (or scary depending on your point of view) to consider
> > what happens if the variable YOUR-POINT is the last reference to the CLOS
> > object here...the finalizer might free the foreign object before mem-aref is
> > entered if the compiler no longer keeps a pointer to the variable.
> >
> No this is not safe.  In general you should never do 
> `(c-pointer ...)` outside the low level parts of the bindings, and use
> it very carefully.
> 
> Personally I would try to never use the `c-pointer` method outside
> the `translate-to-foreign` code.   And I naively expected that this
> would be safe.   But as you point out:
> 
> > Or more specifically, can this ever be safe?
> >
> > (mem-aref (c-pointer (point0)) ...)
> 
> This is never safe.
> 
> I think the easiest fix is to change the 
> 
>   (defmethod translate-to-foreign ((lisp-value cv-matrix) (c-type cv-mat))
>     (c-pointer lisp-value))
> 
> code to  [UNTESTED CODE]:
> 
> 
>   (defmethod translate-to-foreign ((lisp-value cv-matrix) (c-type cv-mat))
>     (values  (c-pointer lisp-value) lisp-value)
> 
> 
> This should keep the lisp-value around until we are done using the
> `(c-pointer ...)` value.   Provided of course you are not mucking around
> with the `c-pointer` method yourself.
> 
> 
> In general, I would advocate to hide all the nasty c-pointer business in
> the translate methods and never deal with it outside that limited scope.
> 
> 
> Thank you for bringing this to my attention.  I do think we need a paper
> or section in the manual on how to deal with combing GC in Lisp and
> manual memory management on the C side.  
> 
> When my time frees up (hopefully in a month or two)  I might take a stab 
> at a first draft.

IMHO, it needs to come with a huge security warning, otherwise there will be
endless use-after-free bugs when users lose the wrapper object after storing
the pointer somewhere.

__Martin