gh-109051: asyncio: remove outgoing buffer from sslproto

[sorcio]

asyncio/sslproto.py TLS wrapper has a redundant outgoing buffer (the MemoryBIO object already serves as a buffer), and a flow control layer for outgoing data which serves an unclear purpose

If this extra logic can be removed, it would solve the AssertionError issue described in the issue as a secondary effect, because the asserted condition cannot be guaranteed in general (details in issue comments).

For reference, the current sslproto.py code was ported from uvloop, which manifests the same issue. The uvloop implementation uses the MemoryBIO directly as a write buffer, but still has the extra flow control logic which causes the mismatch.

• Issue: test_asyncio: test_start_tls_server_1() fails randomly on the Cirrus CI FreeBSD job #109051

[sorcio]

This breaks so many tests that I didn't see locally. Feel free to downgrade to draft or close.

[sorcio]

Possibly fixed what was breaking. But I'm less confident of the validity of the change because I had to touch sendfile stuff.

My stubborn idea was that _FlowControlMixin should not be necessary (both in sendfile, which explicitly checks for it, and in start_tls, which could check for it to fix the issue but doesn't). Not depending on internals makes it easier to add wrapper transports/protocols, including the TLS wrapper itself. But there could be some other good reason that I don't see. Anyway, this is open for discussion.

---

The alternative fix is much easier (but feels dirtier to me). Something like: in start_tls, report an error if the transport is not a _FlowControlMixin, and then pass down the value of _protocol_paused to SSLProtocol.

Even in this case, the redundant "outgoing" buffer in the wrapper could be removed (again, unless I'm missing something important), but it's an orthogonal issue.

[gvanrossum]

You can make it draft yourself. Maybe @vstinner can help?

[sorcio]

I don't have the option to convert to draft (maybe it's not available to external contributors).

Anyway, the PR is open for review now.

The only thing is that this is not just a fix to gh-109051 but a kind-of deeper change. That's why I suggested #109603 as an immediate fix.

[vstinner]

cc @1st1