I’m debugging an issue I’m seeing where connections between a podman container and the host aren’t being closed inside the container.

I’ve written very simple test programmes: server and client.

Server:

1. Accepts connection from client.
2. Sends data on connection until sending blocks.
3. Goes to sleep.

Client:

1. Connects to server.
2. Sends data on connection until sending blocks.
3. Polls with pfd.events = 0 (i.e. effectively listening for only POLLERR and POLLHUP)

A short time after startup, both applications have full TCP send buffers and full TCP recv buffers. I then kill the server application and I observe the following:

• When both applications run on the same machine, client gets events POLLERR | POLLHUP.
• When both applications run on separate machines, client gets events POLLERR | POLLHUP.
• When the client runs inside a podman container, client isn’t notified of any events.
• When the client runs inside a podman container run with --network=host, client gets events POLLERR | POLLHUP.

My understanding is that, at least on Linux, killing the server application (i.e. closing the socket) with data still in the TCP receive buffer causes a RST to be sent to the peer. The RST arriving at the client is what causes the POLLERR | POLLHUP event.

Why doesn’t the RST reach the process running inside the Podman container?

Interestingly, if I modify my server application so it doesn’t send any data, then the RST does make it through to the podman container.