Modules that return an error during provisioning should still be cleaned
up so that they don't leak any resources they may have allocated before
the error occurred. Cleanup should be able to run even if Provision does
not complete fully.
Use piles from which to draw config values.
Module values can return their name, so now we can do two-way mapping
from value to name and name to value; whereas before we could only map
name to value. This was problematic with the Caddyfile adapter since
it receives values and needs to know the name to put in the config.
Along with several other changes, such as renaming caddyhttp.ServerRoute
to caddyhttp.Route, exporting some types that were not exported before,
and tweaking the caddytls TLS values to be more consistent.
Notably, we also now disable automatic cert management for names which
already have a cert (manually) loaded into the cache. These names no
longer need to be specified in the "skip_certificates" field of the
automatic HTTPS config, because they will be skipped automatically.
This makes it faster and easier to detect broken configurations, but
is a slight performance hit on config loads since we have to re-encode
the decoded struct back into JSON without the module name's key
Tested for memory leaks and performance. Obviously the added locking and
global state is not awesome, but the alternative is a little uglier IMO:
we'd have to make some sort of "liaison" value which stores the state,
then pass it around to every module, and so LoadModule becomes a lot
less accessible, and each module would need to maintain a reference to
it... nope, just ugly. I think this is the cleaner solution: just make
sure only one Start() happens at a time, and keep global things global.
Very simple log middleware is an example.
Might need to reorder the operations in Start() and handle errors
differently, etc. Otherwise, I'm mostly happy with this solution...
