On 29/09/2016 13:13, Daniel P. Berrange wrote: > On Thu, Sep 29, 2016 at 02:02:15PM +0300, Denis V. Lunev wrote: >> From: Denis Plotnikov <dplotni...@virtuozzo.com> >> >> Originally NBD server socket was created by qemu-nbd code. This leads to >> the race when the management layer starts qemu-nbd server and allows a >> client to connect to the server. In this case there is a possibility that >> qemu-nbd does not open listening server socket yet. Creating listening >> socket before starting of qemu-ndb and passing socket fd via command line >> solves this issue completely. > > FWIW, this could be solved in qemu-nbd itself if we had a general > ability to request "daemon" mode - currently it only daemonizes > if attaching to a nbd block device. > > The key would be that qemu-nbd would open the listening socket > before daemonizing. Thus when the mgmt application spawned > qemu-nbd in daemon mode, it can be sure that the listener > socket is present when waitpid() completes.
This is 100% true, but I find daemonization to be a hack, so file descriptor passing has its place. Still, I'd prefer to have a --socket-activation option which uses the systemd socket activation protocol. The systemd protocol can be implemented easily by any management layer. Paolo
