I've exchanged a couple of emails with Martin here who points out my patch
solution isn't ideal (though it Works For Me (TM)):
Take care,
--
/\/\ Christian Reis, Senior Engineer, Async Open Source, Brazil
~\/~ http://async.com.br/~kiko/ | [+55 16] 274 4311
---------- Forwarded message ----------
Date: Tue, 5 Dec 2000 12:18:24 +0100 (MET)
From: Martin von Loewis <lo...@in...>
Reply-To: he...@py...
To: ki...@as...
Cc: pyt...@py...
Subject: Re: [Python-Help] importdl.c and RTLD_GLOBAL for dlopen
> a) Why was the change reverted?
It had been causing problems if two extension modules would define the
same symbol with a different meaning. In a specific case, an Oracle
module used a symbol named "initsocket". Unfortunately, the dynamic
linker would resolve that as the symbol from the socket module, so the
application crashed.
If different extension modules can't exchange symbols, you don't get
this problem.
> Is the policy for 2.0 the same?
Yes, it is.
> Redhat supplies a patched version by default that includes
> RTLD_GLOBAL.
They should not do this. Do you know who at Redhat is responsible for
packaging Python?
> b) Is there a suggested workaround for this issue?
Yes. The suggested work-around is to create a table of function
pointers and put that in a PyCObject. Export that object under a
well-known name in one module, and import it in the other. See
cStringIO.c or ucnhash for examples where this is used in Python, see
NumPy for a larger example that manages many functions that way.
Hope this helps,
Martin
|
