By using this site, you agree to our updated Privacy Policy and our Terms of Use. Manage your Cookies Settings.
443,813 Members | 1,130 Online
Bytes IT Community
+ Ask a Question
Need help? Post your question and get tips & solutions from a community of 443,813 IT Pros & Developers. It's quick & easy.

Troubleshooting garbage collection issues

P: n/a
Hi folks - wondering if anyone has any pointers on troubleshooting
garbage collection. My colleagues and I are running into an
interesting problem:

Intermittently, we get into a situation where the garbage collection
code is running in an infinite loop. The data structures within the
garbage collector have been corrupted, but it is unclear how or why.
The problem is extremely difficult to reproduce consistently as it is
unpredictable.

The infinite loop itself occurs in gcmodule.c, update_refs. After
hitting this in the debugger a couple of times, it appears that that
one of the nodes in the second or third generation list contains a
pointer to the first generation head node. The first generation was
cleared shortly before the call into this function, so it contains a
prev and next which point to itself. Once this loop hits that node,
it spins infinitely.

Chances are another module we're depending on has done something
hinkey with GC. The challenge is tracking that down. If anyone has
seen something like this before and has either pointers to specific GC
usage issues that can create this behavior or some additional thoughts
on tricks to track it down to the offending module, they would be most
appreciated.

You can assume we've done some of the "usual" things - hacking up
gcmodule to spit information when the condition occurs, various
headstands and gymnastics in an attempt to identify reliable steps to
reproduce - the challenge is the layers of indirection that we think
are likely present between the manifestation of the problem and the
module that produced it.

Many thanks,

Dave
Nov 17 '07 #1
Share this Question
Share on Google+
3 Replies


P: n/a
On Nov 17, 10:34 am, "davemer...@gmail.com" <davemer...@gmail.com>
wrote:
Hi folks - wondering if anyone has any pointers on troubleshooting
garbage collection. My colleagues and I are running into an
interesting problem:

Intermittently, we get into a situation where the garbage collection
code is running in an infinite loop. The data structures within the
garbage collector have been corrupted, but it is unclear how or why.
The problem is extremely difficult to reproduce consistently as it is
unpredictable.

The infinite loop itself occurs in gcmodule.c, update_refs. After
hitting this in the debugger a couple of times, it appears that that
one of the nodes in the second or third generation list contains a
pointer to the first generation head node. The first generation was
cleared shortly before the call into this function, so it contains a
prev and next which point to itself. Once this loop hits that node,
it spins infinitely.

Chances are another module we're depending on has done something
hinkey with GC. The challenge is tracking that down. If anyone has
seen something like this before and has either pointers to specific GC
usage issues that can create this behavior or some additional thoughts
on tricks to track it down to the offending module, they would be most
appreciated.

You can assume we've done some of the "usual" things - hacking up
gcmodule to spit information when the condition occurs, various
headstands and gymnastics in an attempt to identify reliable steps to
reproduce - the challenge is the layers of indirection that we think
are likely present between the manifestation of the problem and the
module that produced it.
Does "usual things" also include compiling with --with-pydebug?

You could also try the various memory debuggers. A refcounting error
is the first thing that comes to mind, although I can't see off hand
how this specific problem would come about.

Are you using threading at all?

Do you see any pattern to the types that have the bogus pointers?

--
Adam Olsen, aka Rhamphoryncus
Nov 18 '07 #2

P: n/a
<dave..mail.com (Dave) wrote:

8<--------- description of horrible problem --------------

Faced with this, I would:

1 - identify the modules that import gc to separate the
sheep from the goats.

2 - do my best to change gc importing goats back to sheep.

3 - amongst the remaining goats, identify the ones that also use
threads, (supergoats) and take a long hard look at them.

4 - hope I get lucky.

5 - If no luck, I would change the most complex of the
supergoats to use more processes and messaging,
to make sheep out of a supergoat, or failing that,
a goat and some sheep.

6 - Repeat from 2 until luck strikes.

Now the trouble with a simple minded algorithm such as
the above is that a sheep could be at the bottom of the
trouble if it uses threads. So a module is only a lamb if
it uses neither threads nor makes calls into gc...

HTH

- Hendrik

Nov 19 '07 #3

P: n/a
Thanks for the thoughts - much appreciated! The threaded super-goat
was indeed the offender. A very aggressive QA tester got us enough of
a pattern to identify the offending module: pyOpenSSL. After looking
at it closely, we found there are problems with its thread handling.
In particular, the GIL is not properly locked when manipulating
reference counts and also, in once case, when creating a new python
object. Once we cleaned that up we were unable to reproduce the
problem.

We'll post the fixes back to the pyOpenSSL folks shortly.

Thanks again!

Dave
Nov 28 '07 #4

This discussion thread is closed

Replies have been disabled for this discussion.