Infinite loop on `Stream::poll_shutdown()`

[hatoo]

tokio-rustls/src/common/mod.rs

Lines 330 to 334 in 66fb0ae

	fn poll_shutdown(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
	while self.session.wants_write() {
	ready!(self.write_io(cx))?;
	}

self.write_io(cx) in the snippet may returns Ok(0) and causing an infinite loop.

I've faced this issue in my project.
I've failed to minimize codes that trigger this issue but please let me know if you want to see codes.

[quininer]

The write_io returns Ok(0) There are two possibilities:

1. There is no data to be written in session. this is not possible because we have checked .wants_write().
2. The underlying IO stream returns 0 instead of pending. this means that you are using a bad IO stream implement.

[hatoo]

Hi, I've found that H2Upgraded::poll_write in hyper returns Ok(0) on the situation and I made an issue to hyper hyperium/hyper#3801

@seanmonstar said

Returning Ok(0) is the normal way for IO resources to indicate no more bytes are allowed (see https://doc.rust-lang.org/std/io/trait.Write.html#tymethod.write).

I think he said the right thing about Ok(0) according to https://doc.rust-lang.org/std/io/trait.Write.html#tymethod.write.
and tokio-rustls should be aware of Ok(0) as a writing result.

[ctz]

Sounds like we should convert a Ok(0) from write to ErrorKind::WriteZero and have it fail the whole connection.

[quininer]

This does surprise me a bit, this is not the behavior of posix write or send. it looks like std once changed its description, but was reverted. rust-lang/rust#114897

I'd willing to be compatible with this anyway, because there's no harm in it. but i still think IO stream should return an error instead of 0, whether it's BrokenPipe, ConnectionReset or WouldBlock.