Mika Raento wrote:
> I'm looking into the target-language option. Basically what is needed to
> do that is to:
>
> 1. in each target language, have a place to stash the extra reference
> (e.g., a module level hash in perl)
> 2. shadow startTransaction/asynchPrepare to stash an extra reference to
> Ndb/Ndb+transaction (NdbOperations are owned by the transaction)
> 3. shadow close/callback to remove the extra reference
Well, you know, it might be possible at the C level per language, too.
In Python, I can call Py_INCREF(ob) where ob is a PyObject*, which
increments the ref count. Then it would be easy to place a Py_DECREF(ob)
in the transaction->close() method wrapper, since I'm sure you don't
mind if the object is reaped when it goes out of scope after calling close.
Turns out the other place I did this was C#, not python, actually, and
it was all done with typemaps ...
%typemap(cscode) NdbTransaction %{
private Ndb ndbReference;
internal void addReference(Ndb ndbref) {
ndbReference = ndbref;
}
%}
%typemap(csout, excode=SWIGEXCODE) NdbTransaction* Ndb::startTransaction {
IntPtr cPtr = $imcall;$excode
$csclassname ret = null;
if (cPtr != IntPtr.Zero) {
ret = new $csclassname(cPtr, $owner);
ret.addReference(this);
}
return ret;
}
%typemap(csfinalize) NdbTransaction %{
~$cslassname() {
// This is only if we haven't managed to kill this guy yet.
// Don't depend on this.
ndbReference.closeTransaction(this);
Dispose();
}
%}
But it was (is) sort of ugly.
> Out of these 3. is the only one I'm not sure how to do. I have to find
> out whether you can shadow the destructor. Shadowing the callback will
> already give a place to keep the extra reference: create another closure
> around the callback given by the user-level program that maintains a
> reference.
>
> It's just a pain that we (well, you, I only care about perl for the
> moment :-) have to do it for each target language separately.
I'm discovering that there are many things where the swig answer is a
little too language specific where it really shouldn't need to be. Oh
well... I shouldn't complain too much, I can't imagine how long it would
take if we had to do _everything_ by hand. I've gotta do Datetime and
Decimal typemaps for each language, I guess refcounting a few objects
won't hurt. :)
Hey - while you're look at the callback stashing, check out the
directors version of callbacks. I wasn't able to make a very good test
case, so they may actually work. I actually got a segfault, but now I'm
thinking maybe this was related to the issue you're talking about here?
It requires making a class in Perl that derives from BaseCallback and
overrides the callback method. You'll also have to add (directors="1")
to the %module statement. BUT - it might offer a nice way to hook in the
transaction. Or - maybe it's not very perl-ish. I sort of like your
anonymous callback method in async.pl and the fact that it works.
--
Monty Taylor
Senior Consultant
MySQL Inc., www.mysql.com
Get More with MySQL! www.mysql.com/consulting
