GitLab

Currently, the TCP connection timeouts is configured using a socket option that varies with the platforms:

• On Linux, it's TCP_USER_TIMEOUT (in miliseconds)
• On Windows, it's called TCP_MAXRT (in seconds)
• On macOS, there are two! socket options: TCP_CONNECTIONTIMEOUT (in seconds) and TCP_RXT_CONNDROPTIME (documented just in the header)
• On FreeBSD 11 and FreeBSD 12, the option is called TCP_KEEPINIT (in seconds)

Alas, at least OpenBSD and NetBSD doesn't have any socket option like this, so a system-wide default is used (and it can be large) leading to connection callback not being called in quite a long time -> thus the unit test might crash on shutdown because there's dangling nmsocket waiting for connection.

We need to restore the connection timeout handler that will:

• call uv_close() on the connection uv_tcp_t in the timeout handle forcing the connection callback to be called with the timeout error code
• properly handle that case where the uv_tcp_t is closed out of order