Errors during init get lost in the logs

[richvdh]

When started with -D, Synapse daemonizes itself before it binds any TCP/SSL listeners, which means that any errors that occur during that process (such as permission errors or "Address already in use", but also any of a number of potential configuration errors) get lost in the logs rather than being visible in stderr, and the process manager (synctl or systemd) can't tell that there was an error.

[richvdh]

I can see a few potential solutions to this:

1. Use a temporary reactor for the ACME init, so that once it finishes, we can stop the reactor again, bind the sockets, daemonize, and then start the main reactor.

2. Bind the TCP sockets before daemonizing and starting the reactor, but don't start listening on them until later, when we have the certs (I don't think twisted's Port implementation allows this currently though). Also try to do more validation of config before forking.

3. Reimplement daemonization so that the original process does not exit until there is some signal that the forked process has started happily. This probably involves throwing away daemonize which is a thing I want to do anyway as per stderr goes to /dev/null when we run with --daemonize #1539 (comment).

4. Decide that daemonization isn't our job and we should leave it up to systemd or some other process manager.

Personally I like option 3.