OK, so, requests have a `timeout` and take it into account, and it solves my problem.

Yet I don't understand one little thing:

With both requests `timeout` parameter set or unset, the exact same http.client.py:_read_status call the same socket.readinto. With a `request` `timeout`, the socket.readinto uses the recvfrom syscall, but with a `request` `timeout`, readinto uses a `poll` syscall with the given timeout.

The root of the problem is that urllib3 ignores the `socket` `defaulttimeout`, I opened a ticket on this:

https://github.com/shazow/urllib3/issues/655#issuecomment-114835279

So this ticket can be considered closed.