bpo-40007: Make asyncio.transport.writelines on selector use sendmsg

[tzickel]

https://bugs.python.org/issue40007

[aeros]

Thanks for the PR @tzickel.

Assuming the general idea is approved of by Yury and/or Andrew, this PR will likely need to include a unit test that ensures it behaves as expected; it would be best included in Lib/test/test_asyncio/test_selector_events.py. See the existing transport tests in there for examples, such as the ones that start with test_write*.

The PR could also use a Misc/NEWS entry to briefly explain the changes.

In the meantime, I have a few suggestions/general comments:

[aeros]

Also, the current CI failures will have to be addressed of course. My comments above were just concerns that I noticed at a glance. I suspect that some of the failures may have been occurring as a result of the seld._buffer code typos that I mentioned, but I haven't looked over them yet.

[tzickel]

@aeros Thanks for your comments, I've hopefully fixed the code / logic mistakes.

[aeros]

@tzickel

Thanks, but it's a small one-time overhead instead of a small overhead on each object initialization

I wasn't referring to repeating the getattr(socket.socket, "sendmsg", False) every time the object is created; that's not necessary. Here's an example of what I had in mind:

Current:

sendmsg=getattr(socket.socket, "sendmsg", False)
# [snip]defwritelines(self, lines):
ifnotsendmsg:
returnself.write(b''.join(lines))
# [snip]

Recommended:

# Used to determine if platform supports sendmsg to sockets# [_NAME is our typical convention for internal globals, to differentiate it]_SENDMSG=Noneclass_SelectorSocketTransport(_SelectorTransport):
# [snip]def__init__(self, loop, sock, protocol, waiter=None,
extra=None, server=None):
global_SENDMSG# [ensures the `getattr()` is only done once, but doesn't create extra# overhead when select_events is imported]if_SENDMSGisNone:
_SENDMSG=getattr(socket.socket, "sendmsg", False)
self._sendmsg=_SENDMSG# [snip]defwritelines(self, lines):
ifnotself._sendmsg:
returnself.write(b''.join(lines))
# [snip]

Does that make it more clear?

[aeros]

Also, as a side note, there's typically no need to force push over old commits for PR branches in the CPython repo since we squash prior to merging. It can occasionally help if it the history becomes muddled over time, but it can also make changes made to the PR harder to follow for review purposes.

See https://discuss.python.org/t/pep-601-forbid-return-break-continue-breaking-out-of-finally/2239/42. It's also mentioned at the end of https://devguide.python.org/pullrequest/#submitting.

[asvetlov]

#31871 is the newer resurrection of the idea.