intermittent ResourceWarning in test_ftplib

[graingert]

Hi! The buildbot PPC64LE Fedora Stable LTO 3.13 (tier-2) has failed when building commit 07d4c7e.

You can take a look at the buildbot page here:

https://buildbot.python.org/#/builders/1492/builds/501

Summary of the results of the build (if available):

Traceback (most recent call last):
  File "/home/buildbot/buildarea/3.13.cstratak-fedora-stable-ppc64le.lto/build/Lib/test/test_ftplib.py", line 935, in tearDown
    self.server = None
ResourceWarning: unclosed <ssl.SSLSocket fd=6, family=2, type=1, proto=0, laddr=('127.0.0.1', 39543), raddr=('127.0.0.1', 54764)>
k


Traceback (most recent call last):
  File "/home/buildbot/buildarea/3.13.cstratak-fedora-stable-ppc64le.lto/build/Lib/test/test_ftplib.py", line 935, in tearDown
    self.server = None
ResourceWarning: unclosed <ssl.SSLSocket fd=5, family=2, type=1, proto=0, laddr=('127.0.0.1', 43007), raddr=('127.0.0.1', 49844)>
k

Originally posted by @bedevere-bot in #131802 (comment)

Linked PRs

• gh-131888: fix ResourceWarning in test_ftplib #131889
• [3.13] gh-131888: fix ResourceWarning in test_ftplib (GH-131889) #132009
• [3.12] gh-131888: fix ResourceWarning in test_ftplib (GH-131889) #132010

[graingert]

I have a fix, I think

[graingert]

The issue is due to SSLConnection overriding close() to initiate a tls shutdown, however the asyncore interface requires .close() to close the underlying connection synchronously.

For tls shutdown to close the socket, it must poll the selector and retry the socket.unwrap() call, and then worse asyncore.close_all() needs to be called again on the unwrapped socket.

In the unittest tearDown only .close() is called, and only once - not the required N+1 times and without waiting for readability or writeability on the socket.

In my PR I split the close behaviour into .close() for killing the connection and .shutdown() for initiating an unwrap. All internally initiated closes use shutdown because they can wait using the selector. External closes use the original hard close method from the super class

Probably we can get away with only one call to asyncore.close_all() and rely on it to work when joining the thread.