On Fri, Apr 15, 2005 at 10:11:18PM -0400, Jack Diederich wrote:
> I've been getting segfaults running metakit in a persistent python interpreter
> (under mod_python, so the interpreter lives for a long time). I just checked
> out the metakit head (on the server I'm running 2.4.9.3) and jcw figured out
> how to initialize under a python debug build[1] so here is what I can report:
>
> This is a debug python build so when run interactively it outputs the number
> of live references after every command.
>
> The tst.py is just
> import metakit
> import gc
> def round1():
> def round1():
> fname = '/tmp/test.mk'
> stor = metakit.storage(fname, 2)
> gc.collect() # force cleanup on unreachable objects
> return
>
> here is the interactive output, with reference counts.
>
> Python 2.4.1 (#1, Apr 15 2005, 21:18:43)
> [GCC 3.3.5 (Debian 1:3.3.5-12)] on linux2
> Type "help", "copyright", "credits" or "license" for more information.
> >>> import tst
> [23817 refs]
> >>> tst.round1()
> [23820 refs]
> >>> tst.round1()
> [23823 refs]
> >>> tst.round1()
> [23826 refs]
> >>> tst.round1()
> [23829 refs]
>
> So opening a new storage is leaking three refs per go.
> I ran the same test with some python objects I know don't leak (dicts
> and lists) as a sanity test and the #refs after each call stayed static,
>
> I'll try and hunt down the leak, but if anyone has any ideas let me know.
>
Some digging, PyStorage objects are created & destroyed with appropriate
refcounts. Some of the auxilliary objects created in PyStorage_new leak.
// PyStorage.cpp line 370-ish
// returning here leaks no refs per call
PWOString fnm(args[0]);
// returning here leaks 1 ref per call
if (PyInt_Check((PyObject* )args[1])) {
PWONumber mode (args[1]);
ps = new PyStorage(fnm, mode);
if (!ps->Strategy().IsValid()) {
}
// returning here leaks more (3) refs per call
Closer still ..
_____________________________________________
Metakit mailing list - [email protected]
http://www.equi4.com/mailman/listinfo/metakit
