[Orbit-python-list] Almost there (was: memory leak)

[Roland M. <ma...@ec...>]

Jason Tackaberry (2000-08-10 07:17:46 -0400) :

> > I attach a simple server, client and IDL file.  They demonstrate a
> > memory leak somewhere: the server starts at about 2.5 kb and grows to
> 
> On second thought, your idl isn't using any structs or unions, which is
> where the leaks I've identified are.  You may want to look at leaks I've
> flagged in server.c and client.c.  I'm pretty sure I'm leaking memory
> with opd->parameters._buffer but I'm not sure how to free up that
> memory.  (I tried various things that result in a segfault.)  I'll poke
> around this some more, but I'd appreciate it if you could look into it
> too.

Well, I'm almost there.  I spent the last two days (nights, actually)
in src/server.c, and I fiddled in it a bit.  Here are my results.

  I think the leak you mention in src/server.c is not real (or, it is
not the only one).  Without doing anything with opd, I manage to
suppress the memory leak in some cases.  I guess g_free takes care of
things for you, but I wouldn't stake my life on it.

  We have four types of calls: the _get_*, the _set_*, the other
methods with parameters, and the other methods without parameters.
Using the (included) updated test-case suite, Orbit-Python leaks
memory for all four of them.

  I tried many things, adding and removing Py_INCREF's and Py_DECREF's
here and there, and I got a great many segfaults too :-) But in the
end, I gathered that the server needed a bit of clarification, since
the 'result' variable is used, overwritten, Py_INCREF'ed and
Py_DECREF'ed too many times.  So I cleaned the code a bit, added the
'fresult' variable, which contains the result object we'll have to
free at the end, and Py_DECREF'ed it.

  My results: the memory leak is dead in three of the four
aforementioned cases, namely the _get_* call and the method calls,
with and without parameters.  It remains unchanged as for the _set_*
calls, though (same pace of leak).

  I include my patch to the current CVS version, as well as the
updated test suite.  I'll also upload the patch to Sourceforge in a
few minutes.  I'm not quite sure of myself, so the more people use it
and crash-test it, the more certain I'll be that my patch broke things
or not.  If it does, please tell me!  I depend on feedback here...
Also, I'll continue investigating on the _set_* case, but I'm getting
blind at looking at the lines 105-115 and 125-130.  I'm almost sure
the bug is right there, though.

  May the Force be with us for the last (?) line,

Roland.