Dear Scott,
I don't like the limitations of your patch, and I think I've come up
with a much cleaner solution. The attached UNTESTED diff should apply
cleanly against SVN HEAD. Please let me know if it works, and I would
really appreciate a test case (or a few, for the different threading
modes), as this does seem a bit convoluted and there might be corner
cases I overlooked. And after all, the last thing you want for a
zero-downtime deployment is crashing code ;-).
Especially with a test, I can see this going into MHD relatively
quickly ;-).
Happy hacking!
Christian
On 04/25/2013 10:31 AM, Scott J. Goldman wrote:
Hi.
Where I work, we do continuous deployment of services, so it's
important to us to be able to be able to deploy new code and restart
our services in a way that doesn't cause downtime. One of the ways we
do that is by deploying daemons which can fork-exec themselves and
inherit the listening socket. With a slightly modified libmicrohttpd,
we implemented this functionality. It works like this:
1. send SIGUSR1 to the daemon
2. daemon fork-exec's a new version of itself, passing the fd# of the
listening socket, and a status pipe via command-line args
3. new daemon initializes, starts accepting connections on the socket
(with MHD_OPTION_LISTEN_SOCKET), and sends a message via the status pipe
4. old daemon operates normally until it gets the status message, and
then stops accept()ing new connections, but continues processing old ones
5. old daemon finishes processing all remaining clients, and terminates
libmicrohttpd had nearly all the functionality we needed, except for
having the "quiesce" functionality, to stop accept()'ing new
connections, while continuing to process existing clients. I'm
wondering if you would consider this feature for inclusion in your
mainline version of the library. I've included a link to our patch
below that is our rough attempt to hack in this functionality. I
would be happy to revise it with your feedback.
https://gist.github.com/scottjg/5458005
https://gist.github.com/scottjg/5458005/raw/4ad505101f80485e24a6a2f1092f173706749df3/quiesce-api.patch
The main issue with the patch is that calling shutdown() on the
listening socket is no longer acceptable in this mode. On Linux,
calling shutdown() in the old daemon will prevent the new daemon from
being able to accept() new clients. For our purposes we used the
alternate pipe method to wake up the event loop, which appears to
works fine.
Cheers,
-sjg
<quiesce.diff>