The shutdown process here attempted to terminate the test server
by interrupting it with Thread#kill, and then proceeded to close
the server and join the thread. The kill does indeed interrupt
the accept call, but the close call could also interrupt the
thread as part of notifying blocked threads waiting on that
socket call.
In JRuby, where all of this can happen at the same time, it leads
to the following scenario:
* The server thread enters TCPServer#accept and blocks.
* The main thread calls Thread#kill to interrupt the accept call.
* The server thread wakes up and starts to propagate the kill.
There is a slight delay between this wakeup and removing the
server thread from the TCPServer's blocked threads list.
* The main thread calls TCPServer#close, which sees that the server
thread is still in the blocked list, so it initiates a second
interrupt to raise IOError "closed in another thread" on the
server thread.
* As the kill is bubbling out, another check for interrupts occurs,
causing it to see the new raise interrupt and propagate that
instead of the active kill.
* Because the server is now closed and the rescue here is empty,
the server loop will endlessly attempt and fail to call accept.
I was unable to determine how CRuby avoids this race. There may be
code that prevents an active kill interrupt from triggering
further interrupts.
In order to get these tests running on JRuby, I've made the
following changes:
* Only kill the thread; one interrupt is sufficient to break it
out of the accept call.
* Ensure outside the server loop that the server gets closed. This
happens within the server thread, so triggers no new interrupts.
* Minor cleanup for the pattern of using @ssl_server or @server.
This change avoids the race in JRuby (and possibly other parallel-
threaded implementations) and does not impact the behavior of the
tests.
https://github.com/ruby/net-http/commit/54025b3870