fix tight loop on aborted connection

[aep]

when the underlying IO returns 0 on read, we must stop polling it since
it's closed. Otherwise we'll be stuck in a tight loop.

this fixes sfackler/rust-openssl#949

[carllerche]

Good catch, thanks.

I left a note inline. Also, any way to add a test that covers this case?

[carllerche]

These reasons are mapped to http 2.0 reasons from the spec. In this case, it would either be a protocol error or an I/o error (connection reset by peer).

[aep]

or an I/o error

sounds more appropriate than protocol error but couldnt find it in the spec

Also, any way to add a test that covers this case?

sounds like a lot of effort, since we'd have to make an io that deliberatly returns 0 at a specific point.

[sfackler]

Seems like it wouldn't be that bad to make a test. Start a server, open up a connection, write PRI * and close the connection, and make sure the server handshake returns an error.

[aep]

not sure how to do this using the existing h2 test framework.

mock doesnt have a concept of half closed.

i could open a pipe, but thats unix only. let me know if you want that.

#[test]
fn server_error_on_unclean_shutdown() {
    use std::io::Write;

    let _ = ::env_logger::try_init();
    let (io, mut client) = mock::new();

    let srv = server::Builder::new()
        .handshake::<_, Bytes>(io)
        .map_err(|e|println!("{}", e))
        .and_then(|srv| {
            println!("{:?}", srv);
            Ok(())
        });

    sender.write_all(b"PRI *").expect("write");
    drop(client);

    srv.wait().expect("wait");
}

will error with "mock closed" in write

[aep]

ok i changed mock to just ignore write if its closed, this works. the test now passes with this fix and never returns without it.

[carllerche]

Thanks! Looks good to me.