Client half-close: payload of a FIN segment is dropped, and our FIN is sent immediately (no half-close) #111

Affected: 1.0.1 (crates.io). 1.0.0 has the same arms.

Two related problems in TcpState::Established / CloseWait handling, both seen when a client sends data and then calls shutdown(SHUT_WR) (a TCP half-close) while the application on the ipstack side still has a reply to send. Related to #89, which covers the opposite direction (server shuts down first).

1. A data segment that carries FIN loses its payload

Linux puts the FIN on the last data segment when shutdown(SHUT_WR) is called with data still queued, so the segment arrives as ACK|FIN (or ACK|PSH|FIN) with a non-empty payload. In Established, the ACK | FIN arm takes the FIN without queuing the payload, and the ACK | PSH arm discards out-of-order data instead of queuing it. The last 100-500 bytes of the client's stream never reach the reader. With a 4 MiB upload the loss was 144 to 448 bytes per run.

2. A client FIN closes the whole connection (no half-close)

On a client FIN in Established, with nothing in flight, the stack immediately sends its own ACK|FIN and moves to LastAck; in CloseWait a timer forces the same. The server-to-client direction is therefore closed the moment the client half-closes, so a reply still being written is cut off (an echo service returned 216,612 of 4,194,304 bytes). poll_read also returns Pending rather than EOF in CloseWait, so a reader waits for the timer. Expected: ACK the FIN, report EOF to the reader, keep the write side open, and send our FIN when the application calls poll_shutdown (or drops the stream), with the existing timeout as a backstop.

Reproduction

A client that writes N bytes and then half-closes (e.g. sendall(data); sock.shutdown(SHUT_WR)) to a stream that copies its input back to its output (tokio::io::copy_bidirectional-style echo), then reads to EOF. Expected: N bytes echoed, then EOF. Actual: fewer than N bytes (problems 1 and 2), or a stall until the idle timeout.

What we did

We carry a patch to src/stream/tcp.rs (about 300 changed lines, with tests that fail when each change is reverted individually). It queues the payload of a FIN segment before taking the FIN, treats ACK|PSH as ACK, gives CloseWait readers their already-acknowledged bytes and then EOF, and sends our FIN from poll_shutdown/drop instead of from the client's FIN. Happy to turn it into a PR if that is useful; tell us how you would like it split.