This commit adds some Prometheus metrics to /metrics in headscale.
It will add the standard go metrics, some automatic gin metrics and some
initial headscale specific ones.
Some of them has been added to aid debugging #97 (loop bug)
In the future, we can use the metrics to get rid of the sleep in the
integration tests by checking that our expected number of nodes has been
registered:
```
headscale_machine_registrations_total
```
Golangs built in HTTP server does not allow different HTTP timeout for
different types of handlers, so we cannot have a write timeout as we
attempt to do long polling (my bad).
See linked article.
Also removed redundant server declaration
This commit tries to address the possible raceondition that can happen
if a client closes its connection after we have fetched it from the
syncmap before sending the message.
To try to avoid introducing new dead lock conditions, all messages sent
to updateChannel has been moved into a function, which handles the
locking (instead of calling it all over the place)
The same lock is used around the delete/close function.
This commit makes two reasonably major changes:
Set a default timeout for the go HTTP server (which gin uses), which
allows us to actually have broken long poll sessions fail so we can have
the client re-establish them.
The current 10s number is chosen randomly and we need more testing to
ensure that the feature work as intended.
The second is adding a last updated field to keep track of the last time
we had an update that needs to be propagated to all of our
clients/nodes. This will be used to keep track of our machines and if
they are up to date or need us to push an update.
This commit removes most of the locks in the PollingMap handler as there
was combinations that caused deadlocks. Instead of doing a plain map and
doing the locking ourselves, we use sync.Map which handles it for us.
This commit makes the IP prefix used to generate addresses configurable
to users. This can be useful if you would like to use a smaller range or
if your current setup is overlapping with the current range.
The current range is left as a default
Currently the default (and non-configurable) Let's Encrypt listener will
bind to all IPs. This isn't ideal if we want to run headscale on a specific
IP only.
This also allows for one to set the listener to something other than
port 80. This is useful for OSs like OpenBSD which only allow root to
bind the lower port ranges (and don't have `setcap`) as we can now run
`headscale` as a non-privileged user while still using the baked in ACME
magic. Obviously this configuration would also require a reverse proxy
or firewall rule to redirect traffic. I attempted to outline that in the
README change.
