Currently if you call server_close you only close the socket. If we called serve_forever and then call server_close without calling shutdown the serve_forever loop keep running. Before using the selectors module for doing the poll we would have had exception thrown from the select (The socket fd is -1) in serve_forever.
IMO you should be able to call server_close at any time and expect it to stop the serve_forever. Maybe even adding a block option to server_close that will wait on the server_forever if it's running (waiting for bpo-12463 to resolve before doing this).
Added a patch that closes serve_forever if server_close is called.

I thought the purpose of server_close() was to clean up resources after you have finished calling serve_forever() or handle_request(). Just like you call close() on a file after you have finished reading or writing it. If you try to read or write a closed file, that is a programmer error.

This proposal sounds like a new feature, but you are overloading or redefining the purpose of server_close().

From the test case I presume you intend to use server_close() in a separate thread from the thread running serve_forever(). But closing a file descriptor while it is being used in another thread does not seem robust to me. It could cause serve_forever() to raise EBADF. In the worst case, consider what happens if an unrelated third thread makes the server’s file descriptor valid again by opening a file.

What is your use case? If you want to use multithreading, why can’t you use the shutdown() method? For a single-threaded server, maybe see bpo-13749, and maybe bpo-23430 would help by allowing exceptions like SystemExit to stop the server.

I see your point about the server_close purpose and I changed the patch to simulate the behavior of closed file (raising ValueError when operating on closed file).
I don't think that this should be an enhancement as what i try to do is return the behavior as it was before the change to the selectors module. When closing the socket using server_close at the next serve_forever loop you would have gotten ValueError for bad file descriptor to select. In the current implementation we actually keep on pulling on a already free resource.
I know that the user should call shutdown before the server_close but i think that as you said the behavior should simulate closed file. When someone will try to use any other function of a closed file he will receive a ValueError.

I suggest to reject this. Checking a flag before using a shared file descriptor is a recipe for a race condition, and in this case, unspecified behaviour: <http://man7.org/linux/man-pages/man2/select.2.html#NOTES\> (applies to select and poll).

I understand the problem and why this patch should be rejected.