start_tls: AssertionError if writing is paused

First found in the CPython port of sslproto (see https://github.com/python/cpython/issues/109051 but disregard the title, it's not platform-specific), and it can be reproduced in uvloop.

protocol.resume_writing() failed
protocol: <uvloop.loop.SSLProtocol object at 0x1033910c0>
transport: <TCPTransport closed=False reading=True 0x1098351c0>
Traceback (most recent call last):
  File "uvloop/handles/basetransport.pyx", line 96, in uvloop.loop.UVBaseTransport._maybe_resume_protocol
    run_in_context(
  File "uvloop/loop.pyx", line 101, in uvloop.loop.run_in_context
    return context.run(method)
  File "uvloop/sslproto.pyx", line 922, in uvloop.loop.SSLProtocol.resume_writing
    assert self._ssl_writing_paused
AssertionError

Simplest way to reproduce is to increase payload size in test_start_tls_server_1. Send will block and transport will request to pause writing. After start_tls(), the transport will attempt to resume the now-switched protocol, which was not aware it was ever supposed to be paused.

I commented further in the original bug with some notes. The CPython implementation is a bit different (it has a redundant buffering mechanism with a deque which I believe is unnecessary) but the fundamental issue is the same.

Question: what is the motivation for the extra flow control layer in SSLProtocol? It seems it was introduced in uvloop first. I suggested it can be removed and let the original protocol and original transport deal with flow control (which would also fix the bug). But it wouldn't be a good idea to remove it if it's solving a known problem.

tests/test_tcp.py、特に test_start_tls_server_1 から始め、issue に記載された大きなペイロードのケースを再現してください。次に、フロー制御の遷移を理解するため、SSLProtocol.resume_writing と start_tls の周辺にある uvloop/sslproto.pyx を調べてください。完了条件には、assertion の修正が確認済みであることと、TLS 起動中に書き込みが一時停止されるケースをカバーする回帰テストを含めてください。