Refactor: plain socket objects instead of wrapper classes

[pajod]

• Use plain socket objects instead of wrapper classes #3127 still works for my limited use case
• still looks like a straightforward and useful refactor
• I would still volunteer to work on it, if there is any work to be done

So, here is a rebased version with the fixup I commented earlier.

[pajod]

Whoops. There is no regression test for --enable-backlog-metric!

[benoitc]

Thanks for the refactor, code is cleaner :)

Few things to address:

ASGI Worker needs updating

The ASGI worker (gunicorn/workers/gasgi.py) still accesses sock.sock (line 181) and uses str(sock) for logging (line 187) - won't work with plain sockets.

listen() for fd-passed sockets

Old code called sock.listen(self.conf.backlog) including for systemd-passed fds. New set_socket_options() doesn't call it. Could this cause issues with systemd socket activation?

SSL validation timing

Changing to validate_file_exists at config time might break setups where certs are generated dynamically before start. Maybe keep runtime validation?

Tests

As you noted, need regression test for --enable-backlog-metric. Also missing tests for get_uri(), get_backlog(), set_socket_options().

Otherwise looks good - the close_sockets() error handling improvement is nice.

[pajod]

ASGI Worker needs updating

Copied the changes there, but did not check beyond running tests.

listen() for fd-passed sockets

Right. I dont think it does anything, now that I dropped the fromfd duplication. Its the same socket, so why would the kernel mess with the backlog hint? But I'll put it back to avoid unforeseen consequences. I don't see any pressing need to change that right now.

SSL validation timing - Maybe keep runtime validation?

Fair point. Reverted.

Tests

I pushed some more unit tests (and the whole thing still passes my nginx test).

---

I have one more commit to fix edge cases like --bind="[fe80::1%enp21s0]:80" but that is fairly verbose with new socket.getaddrinfo/socket.getnameinfo calls (and also known to be broken for eventlet) so I will leave this PR as-is and submit that separately.

---

Things I attempt to not touch here, but would like to follow up with:

• Make reuse_port behaviour clear and consistent on non-supporting platforms
• Make LISTEN_FDS= compatible with --bind being used simultaneously (and fix reload/re-exec for all resulting edge cases)

[pajod]

I would be interested to receive clear confirmation from someone using the uvicorn-worker package and receiving superfluous "Bad file descriptor" messages on whether these changes resolve the issue, as suggested regarding the original PR this is a rebase for.
(tagging @qubydev @dtrasboIMG separately so to not bump everyone that commented on the issue some years ago)