Fix handling of TCP EOF

[tohin]

This patch fixes handling of StreamReader.read for the case when remote client terminates the connection. As documented here: https://docs.python.org/3.4/library/asyncio-stream.html#asyncio.StreamReader.read read method retuns b"" in this case.

[pgjones]

Could you help me understand how this change relates to https://pgjones.dev/blog/half-closed-sockets-2021/ ?

[tohin]

Hello Phil,

I've tested the code for TCP FIN case and it is clear that current code
skips an event of EOF that is documented here https://docs.python.org/3/library/asyncio-stream.html#asyncio.StreamReader.read

For a half closed case, I think it would not be detected by StreamReader.read()
and EOF would not be returned, so nothing is changed after this PR.

[richardsheridan]

Hi @tohin,

This bug also affects the trio backend. For me, it lead to a leak of sockets in CLOSE_WAIT state and, after enough users close their browsers, a crash from too many open file handles.
on trio the following lines work:

    async def _read_data(self) -> None:
        while True:
            try:
                with trio.fail_after(self.config.read_timeout or inf):
                    data = await self.stream.receive_some(MAX_RECV)
            except (trio.ClosedResourceError, trio.BrokenResourceError):
                break
            else:
                await self.protocol.handle(RawData(data))
                if data == b"":
                    break
        await self.protocol.handle(Closed())

But I am concerned that the final line should possibly be in a finally: block.
I will make a pull request against your fork branch to add this additional code.

FWIW I found #103 trying to debug this problem.

[tohin]

Hi @richardsheridan,

I've rebased the PR.

[tohin]

@pgjones By the way, it is the same issue that is already partially fixed for Trio 8b4ecd4

[pgjones]

Thanks, fixed in 8a3e512