Python emits ResourceWarning when an object using limited resource is destroyed without being explicitly closed: files, sockets, etc. The problem is that it's hard to find where the object comes from, since the warning can occur very late, in the garbage collector, etc.

I propose to reuse tracemalloc.get_object_traceback() to show were the object was allocated, when tracemalloc traces memory allocations. In practice, I propose to add a new "source" parameter to warnings.showwarning().

Attached patch:

* Add a new keyword-only source parameter to warnings.showwarning()
* Add C function PyErr_ResourceWarning() to pass source
* showwarning() uses tracemalloc.get_object_traceback() to get the traceback were the object was allocated
* Modify socket.socket, io.FileIO and os.scandir destructor to use PyErr_ResourceWarning()


Backward-compatibility problem: The C function PyErr_ResourceWarning() always call warnings.showwarning() with the keyword parameter source. If an application replaces the warnings.showwarning() function, it will probably fail because it doesn't know the source parameter.

I don't know how to handle this backward compatibility issue.


The patch is incomplete, it's not possible yet to emit a warning in pure Python with a source parameter.


x.py script used for examples below:
-----------------
import warnings
import os
import socket

def func2():
    #f=open("/etc/issue")
    #f=os.scandir('.')
    f=socket.socket()
    f=None

def func():
    func2()

func()
-----------------


Output with Python 3.5:
-----
x.py:9: ResourceWarning: unclosed <socket.socket fd=3, ...>
  f=None
-----


Output with -X tracemalloc=5 command line option and patched Python 3.6:
-----
x.py:9: ResourceWarning: unclosed <socket.socket fd=3, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('0.0.0.0', 0)>
  f=None
Object allocated at (most recent call first):
  File "x.py", lineno 8
    f=socket.socket()
  File "x.py", lineno 12
    func2()
  File "x.py", lineno 14
    func()
-----

It's much easier to understand where the warning comes from, no? At x.py:8, line "f=socket.socket()".

Note: the traceback doesn't contain the function name, since tracemalloc only stores filename and line number.


See also the issue #26564 "Malloc debug hooks: display memory block traceback on error".


For Python < 3.6, I wrote "res_warn.py" script which monkey-patches io.FileIO and socket.socket to implement something similar. The script is a fragile hack. I would prefer to have the feature built-in Python.

https://bitbucket.org/haypo/misc/src/0a40f27360424145bad0f9b62c9e9148ffdbb169/python/res_warn.py

> Backward-compatibility problem: The C function PyErr_ResourceWarning() always call warnings.showwarning() with the keyword parameter source. If an application replaces the warnings.showwarning() function, it will probably fail because it doesn't know the source parameter.

I proposed the issue #26568 "Add a new warnings.showmsg() function taking a warnings.WarningMessage object" to fix this issue.

I used this patch to identify a ResourceWarning in test_faulthandler when the test is interrupted by CTRL+c: it's really efficient!

I will be very useful to identify all ResourceWarning that I saw in test_asyncio, like the ones seen on the AIX buildbot.

New changeset 2428d794b0e1 by Victor Stinner in branch 'default':
On ResourceWarning, log traceback where the object was allocated
https://hg.python.org/cpython/rev/2428d794b0e1

It looks like io.FileIO has a strong implementation of the destructor. If the object becomes alive again because of random code called in the destructor, the object is not removed.

socket and os.scandir have a classical unsafe destructor.

Moreover, I'm no more sure about the chosen design. When warnings.catch_warnings() is used and an unclosed io.FileIO is destroyed, the object is kept alive because it is stored in the "source" attribute of a warnings.WarningMessage.

I don't know if keeping WarningMessaging alive longer than the call to showwarning() (or _showarnmsg) is a common use case or not. The issue #26568 wants to promote the WarningMessage class, so some users may start to keep it alive.

An alternative is to format the object traceback and pass the traceback to WarningMessage. It requires to decide the format of the traceback (list of ..., string, something else?).

When warnings._showwarnmsg(), the io.FileIO object is not closed yet, so all attributes are accessible, which can be useful. Hopefully, the file is closed even if it is kept alive by the warning logger. So maybe it's ok to keep the Python object alive, if the underlying resource (the file descriptor) is released. I mean, it's not a big deal.

For socket & scandir, I opened the issue #26590: "socket destructor: implement finalizer".

New changeset b8acf98beca9 by Victor Stinner in branch 'default':
Try to fix test_warnings on Windows
https://hg.python.org/cpython/rev/b8acf98beca9

New changeset 543639cdfdb9 by Victor Stinner in branch 'default':
Try again to fix test_warnings on Windows
https://hg.python.org/cpython/rev/543639cdfdb9

New changeset 691fba640266 by Victor Stinner in branch 'default':
Issue #26567: enhance ResourceWarning example
https://hg.python.org/cpython/rev/691fba640266

Cool, test_warnings, test_logging, etc. pass again on Windows buildbots. I close the issue.

As I wrote, I opened a second issue for socket & os.scandir.