Guido van Rossum wrote:
> Guido van Rossum added the comment:
> 
> I wonder if Christian Heimes was correct that the ssl object needs GC
> support? This was part of his patch (which I checked in and then
> reverted because the other part of it didn't work as advertised).
> 
> Alternatively, if 's' is involved in a cycle *and* any of the objects
> in the cycle has a __del__ method, the GC won't call __del__. This
> seems counter-intuitive but is actually correct since __del__ can only
> be called when there are no references to an object left.

A combination of GC support and the removal of __del__ from SSLSocket
has fixed the reference leak problem for me. I tried to move the
_real_clean() call to PySSL_dealloc but it wasn't as easy as I thought.

The code contains a reference cycle. ssl.SSLSocket() contains a
reference to PySSLObject and PySSLObject->Socket is a reference to the
same ssl.SSLSocket instance.

Christian