From: Tilman Sauerbeck <tilman@...>
Date: 2004-07-31T03:37:22+09:00
Subject: Re: C ext: GC claiming objects early

Charles Mills <cmills@freeshell.org> [2004-07-30 19:31]:
> 
> On Jul 30, 2004, at 7:06 AM, Tilman Sauerbeck wrote:
> 
> >ts <decoux@moulon.inra.fr> [2004-07-30 15:43]:
> >>>>>>>"T" == Tilman Sauerbeck <tilman@code-monkey.de> writes:
> >>
> >>T> "some_meth" will instantiate a new "Result" object. Using this 
> >>Result
> >>T> object, the caller can wait for the function to succeed, he can 
> >>register
> >>T> callbacks etc.
> >>
> >> Well it's best if you post the *real* C code : it's easy to 
> >>understand
> >> than this strange language, called english.
> >
> >Heh :)
> >
> >[C code]
> You should show us c_mark and c_free

static void c_mark (RbResult *res)
{
	if (!NIL_P (res->parent))
		rb_gc_mark (res->parent);

	if (!NIL_P (res->callback))
		rb_gc_mark (res->callback);
}

static void c_free (RbResult *res)
{
	if (res->real) {
		xmmsc_result_unref (res->real);
		res->real = NULL;
	}

	res->parent = Qnil;

	free (res);	
}

> In extreme situations the above code it dangerous, here is a safer 
> version:
> <code>
> 	VALUE self;
> 	RbResult *rbres;
> 
> 	if (!res)
> 		return Qnil;
> 
> 	rbres = ALLOC(RbRecsult);
> 	rbres->real = res;
> 	rbres->parent = parent;
> 	rbres->callback = Qnil;
> 	self = Data_Wrap_Struct (cResult, c_mark, c_free, rbres);
> </code>

You mean the memset() call of Data_Make_Struct is dangerous?
I kind of doubt memset is causing problems here, but I'll try it anyway
:)

> >	rb_obj_call_init (rbres->self, 0, NULL);
> >
> >	return self;
> Where did rbres->self come from?  Doesn't look like you initialized it. 
>  You need to pass a valid Ruby object to rb_obj_call_init(), perhaps 
> you intended:
> <code>
> 	return rb_obj_call_init(self, 0, 0); /* returns self */
> </code>

Yeah, it was a typo which didn't occur in my original code *hides*

> >
> >'parent' is the Ruby object that the object marks in c_mark
> >
> >This is the "notifier" method of the Result object and the callback 
> >used
> >to call the given block:
> >
> >static void on_signal (xmmsc_result_t *res2, void *data)
> >{
> >	/* call the block and create a new Result object from res2, too */
> >	rb_funcall ((VALUE) data, rb_intern ("call"), 1,
> >	            TO_XMMS_CLIENT_RESULT (Qnil, res2));
> >}
> >
> >static VALUE c_notifier_set (VALUE self)
> >{
> >	GET_OBJ (self, RbResult, res);
> >
> >	if (!rb_block_given_p ())
> >		return Qnil;
> >
> >	res->callback = rb_block_proc ();
> >	xmmsc_result_notifier_set (res->real, on_signal, (void *) 
> >res->callback);
> This could be your problem. We encountered a similar problem last night 
> with the Rendezvous Ruby wrappers.
> rb_block_proc() creates a ruby object.
> your registering on_signal as your event handler, and it looks like 
> what ever library you are using (you should supply more code here) will 
> the pass the void* to the call back for you.  You need to show c_mark() 
> so it is clear that your marking everything.

Yes, the library does pass res->callback as the data pointer to
on_signal. You are right, I should have commented on it, but it's such a
common technique in C I thought I wouldn't have to explain it :p

-- 
Regards,
Tilman