Skip to content

Instantly share code, notes, and snippets.

What would you like to do?
Adventures in Python Core Dumping

Adventures in Python Core Dumping

After watching Bryan Cantrill's presentation on Running Aground: Debugging Docker in Production I got all excited (and strangely nostalgic) about the possibility of core-dumping server-side Python apps whenever they go awry. This would theoretically allow me to fully inspect the state of the program at the point it exploded, rather than relying solely on the information of a stack trace.

I decided to try exploring a core dump on my own by writing a simple Python script that generated one.

Initial Setup

Doing this required a bit of setup on my Ubuntu 14.04 server.

First, I had to apt-get install python2.7-dbg to install a version of Python with debug symbols, so that gdb could actually make sense of the core dump. It seems Ubuntu comes pre-configured with a Python debugging extension for gdb built-in, so I didn't have to do any extra configuration here, which was great.

I also had to add the following line to /etc/security/limits.conf to actually enable core dump files to be created:

#<domain>       <type>  <item>          <value>
*               soft    core            100000

After that, I created a file called in my home directory:

import os

def my_exploding_func():
    my_local_var = 'hi'


Then I ran the script:

$ python2.7-dbg
Aborted (core dumped)

This created a core file in my home directory.

Exploring The Stack

I opened the core dump in gdb:

$ gdb /usr/bin/python2.7-dbg core
GNU gdb (Ubuntu 7.7.1-0ubuntu5~14.04.2) 7.7.1
Copyright (C) 2014 Free Software Foundation, Inc.

warning: core file may not match specified executable file.
[New LWP 10020]
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/".
Core was generated by `/usr/bin/python2.7-dbg ./'.
Program terminated with signal SIGABRT, Aborted.
#0  0x00007f996aff7cc9 in __GI_raise (sig=sig@entry=6)
    at ../nptl/sysdeps/unix/sysv/linux/raise.c:56
56      ../nptl/sysdeps/unix/sysv/linux/raise.c: No such file or directory.

Now I could use all of gdb's Python debugging extension commands. For example, running py-bt gave me:

(gdb) py-bt
#4 Frame 0x7f996bf28240, for file ./, line 7, in my_exploding_func (my_local_var='hi')
#7 Frame 0x7f996bf28060, for file ./, line 9, in <module> ()

I could also use py-locals to show me the values of local variables in the current stack frame, and py-up and py-down to traverse the stack.

This was all pretty awesome, and will be very useful if my Python programs actually segfault. But it'd be cool if I could actually get all this rich information any time one of my servers returned a 500. That's a bit of a different situation since Python servers don't usually segfault when they return a 500--instead, they catch exceptions, return an error code, and continue running.

For now I'm going to ignore the "continue running" part; there are ways to core dump without killing a process, but right now I'm more interested in figuring out how to get information about handled exceptions.

Obtaining Information About Handled Exceptions

Let's assume we have a script called

import os

def my_exploding_func():
    a = 1

except Exception, e:

The thing about the core dump generated from this script is that running py-bt only gives us the stack trace from the point that we called os.abort(), which is pretty useless:

(gdb) py-bt
#4 Frame 0x7f3767430450, for file ./, line 12, in <module> ()

What we really want is a way to introspect the exception that was currently being handled at the time that os.abort() was called.

There isn't a particularly easy way to do this with the Python debugging extension for gdb, but one nice thing about gdb is that its extensions are written in Python. This means we can write our own extension that gives us easy access to the information we need.

Doing this took some research. It looks like the latest version of the Python debugging extension for gdb is in a file in the CPython codebase called, but this is actually a much newer version than the one that ships with Ubuntu 14.04. I had to use strace to find the actual version on my system, which was at /usr/lib/debug/usr/bin/

After poring through the code and consulting the CPython source code and documentation on extending gdb using Python, I wrote my first gdb extension, which is in the attached file. It adds a py-exc-print command that gives us what we need:

(gdb) source
(gdb) py-exc-print
Traceback (most recent call last):
  Frame 0x7f3767430450, for file ./, line 12, in <module> ()
  Frame 0x7f37673f3060, for file ./, line 7, in my_exploding_func (a=1)
exceptions.NameError("global name 'call_nonexistent_func' is not defined",)

Note that it's more useful than a standard stack trace, as the values of local variables are included in the printout. But more work on the extension needs to be done in order to make those locals easily introspectable.


Thus concludes my first foray into Python core dumping.

Some open questions:

  • I'm not sure how feasible core dumping on every uncaught exception actually is. For instance, how big do core files become in production environments?

  • Are there privacy risks involved in core dumping? Depending on the retention policy, it essentially means that data in use could inadvertently become data at rest.

  • In order for the core dump to be useful, a debug build of the Python interpreter needs to be used. How is performance impacted by this? As the aforementioned Bryan Cantrill talk mentions, we should be able to inspect core dumps from production environments: yet is it feasible to run a debug build of Python in a production environment?

# Note that when we're loaded into gdb via `source`, we
# seem to be loaded into the same namespace as the Python debugging
# extension, which is some version of the following file by David Malcolm:
def pm_sys_exc_info():
'''Just like sys.exc_info(), but post-mortem!'''
# The _PyThreadState_Current global is defined in:
val = gdb.lookup_symbol('_PyThreadState_Current')[0].value()
# The PyThreadState type is defined in:
return [PyTracebackObjectPtr.from_pyobject_ptr(val[name])
for name in ['exc_type', 'exc_value', 'exc_traceback']]
def pm_traceback_print_exc():
'''Kinda like traceback.print_exc(), but post-mortem, and no args!'''
exc_type, exc_value, exc_traceback = pm_sys_exc_info()
sys.stdout.write('Traceback (most recent call last):\n')
while not exc_traceback.is_null():
frame = exc_traceback.get_frame()
sys.stdout.write(' %s\n' % frame.get_truncated_repr(MAX_OUTPUT_LEN))
exc_traceback = exc_traceback.get_next()
exc_value.write_repr(sys.stdout, set())
class PyTracebackObjectPtr(PyObjectPtr):
Class wrapping a gdb.Value that's a (PyTracebackObject*) within the
inferior process.
# PyTracebackObject is defined in:
_typename = 'PyTracebackObject'
def __init__(self, gdbval, cast_to=None):
PyObjectPtr.__init__(self, gdbval, cast_to)
self._py_tb_obj = gdbval.cast(self.get_gdb_type()).dereference()
def _get_struct_elem(self, name):
return self.__class__.from_pyobject_ptr(self._py_tb_obj[name])
def get_frame(self):
return self._get_struct_elem('tb_frame')
def get_next(self):
return self._get_struct_elem('tb_next')
def subclass_from_type(cls, t):
This is called from the from_pyobject_ptr class method we've
inherited. We override its default implementation to be
aware of traceback objects.
tp_name = t.field('tp_name').string()
if tp_name == 'traceback':
return PyTracebackObjectPtr
except RuntimeError:
return PyObjectPtr.subclass_from_type(t)
class PyExcPrint(gdb.Command):
Display a (sort of) Python-style traceback of the exception currently
being handled.
def __init__(self):
gdb.Command.__init__(self, 'py-exc-print', gdb.COMMAND_STACK,
def invoke(self, args, from_tty):

This comment has been minimized.

Copy link

@tcsgmrt tcsgmrt commented May 15, 2018


It is really nice explanation for python debugging. I have tried it for my code but I got 0x0 or 0x60 as output for me after py-exc-print. Does it means program try to access any privilege memory location or something else?

Below is logs got from gdb terminal:
(gdb) py-bt
Traceback (most recent call first):
File "/usr/lib/python2.7/dist-packages/MySQLdb/", line 351, in _fetch_row
return self._result.fetch_row(size, self._fetch_type)
File "/usr/lib/python2.7/dist-packages/MySQLdb/", line 383, in _post_get_result
self._rows = self._fetch_row(0)
File "/usr/lib/python2.7/dist-packages/MySQLdb/", line 379, in _query
File "/usr/lib/python2.7/dist-packages/MySQLdb/", line 217, in execute
res = self._query(query)
File "/opt/tangoworkspace/CentralNode/GUI/src/", line 543, in remotequery
# Return the result of query
File "/opt/tangoworkspace/CentralNode/GUI/src/", line 1445, in generateAttributeDisplayAtContainer1
"SELECT device,attribute, value FROM tango.property_attribute_device "
File "/opt/tangoworkspace/CentralNode/GUI/src/", line 17, in wrapper
func(*args, **kwargs)
File "/opt/tangoworkspace/CentralNode/GUI/src/", line 304, in reloadAll
if printAll:
File "/opt/tangoworkspace/CentralNode/GUI/src/", line 240, in run

(gdb) py-exc-print
Python Exception <class 'gdb.MemoryError'> Cannot access memory at address 0x60:
Error occurred in Python command: Cannot access memory at address 0x60

(gdb) py-list
346 return self._do_query(q)
348 def _fetch_row(self, size=1):
349 if not self._result:
350 return ()

351 return self._result.fetch_row(size, self._fetch_type)
353 def iter(self):
354 return iter(self.fetchone, None)
356 Warning = Warning

Can you help me for this.

Hitesh Patel

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
You can’t perform that action at this time.