Test http_server fails on FreeBSD

[UncleGrumpy]

In recent build-and-test-on-FreeBSD workflows (after the last update to the workflow) the test http_server fails because the client does not receive the body of the message before the server closes the socket. @pguyot helped me track down the problem in our discussions in the Telegram chat, and suggested, as a temporary fix until the server can be rewritten, that a delay was needed before closing the socket so the client could have time to receive the body of the message.

Here was his Telegram comment on the situation:

It looks like a bug in http_server. On this FreeBSD run, the client doesn’t get the body. It seems a bug to me, the server shouldn’t close so fast the connection, as RFC reads “SHOULD issue a graceful close”. We don’t see it on Linux or macOS because the whole packet is flushed. I’m not sure that with gen_tcp we can shutdown the socket, which I think is what most httpd implementation do as graceful, combined with timeout full close.

I think 5 seconds seemed a bit long to hold the socket open, but I don't know how short we can make that and still be sure the client receives the entire body before the socket is closed during times of heavy network congestion. I think waiting 500 to 1000 ms would be sufficient.

I don't understand the subtle differences of the FreeBSD network stack to know why the entire packet is not flushed as on Linux and MacOS, and would welcome any other suggestion or fix.

[petermm]

https://blog.netherlabs.nl/articles/2009/01/18/the-ultimate-so_linger-page-or-why-is-my-tcp-not-reliable

And a +2 hours video on said article https://www.youtube.com/watch?v=JRTLSxGf_6w :/