smtplib uses a wrapper class "SSLFakeFile" in order to call readline() on an SSL socket. But modern SSL sockets have makefile(), so using that wrapper class really shouldn't necessary (of course, it must be investigated whether that's true - and we have no tests for SMTP-over-SSL AFAIK :/).

Attached v1 of patch. Please review.

There are some tests using GMail in test_smtpnet.py. They still pass with the patch. I also did manual tests with GMail with both SMTP + starttls and SMTP_SSL.

The idea of the patch is that SMTP.getreply is already doing:
  if self.file is None:
      self.file = self.sock.makefile('rb')
Therefore, the patch invalidates self.file by setting it to None every time self.sock is (re-)assigned to something.

For consistency, setting self.file to None is also done in LMTP.connect when setting self.sock to a Unix socket although it's not necessarily needed there. Not doing this makes the following scenario fail: create an LMTP instance, call connect, send and read some data (self.file gets initialized), call connect again with an Unix socket, reading more data now uses self.file referring to old socket. But I'm not sure if this scenario is a bug or a misuse of the API, aka you shouldn't call connect twice on the same instance.

Note that I didn't test LMTP. Should I or is it obvious enough that the change is ok?

> Note that I didn't test LMTP. Should I or is it obvious enough that the change is ok?

Thank you for the patch! I think it's ok. I'll give it a try and commit
if everything is alright.

New changeset b68390b6dbfd by Antoine Pitrou in branch 'default':
Issue #11893: Remove obsolete internal wrapper class `SSLFakeFile` in the smtplib module.
http://hg.python.org/cpython/rev/b68390b6dbfd