This commit makes the setLastStateChangeToNow function take a list of
namespaces instead of a single namespace. If no namespaces is passed,
all namespaces will be updated. This means that the argument acts like a
filter.
This commit starts to wire up better signal handling, it starts with
handling shutdown a bit better, using the graceful shutdown for all the
listeners we use.
It also adds the initial switch case for handling config and acl reload,
which is to be implemented.
This commit makes isOutdated validate a nodes necessity to update
against all namespaces, and not just the nodes own namespace (which made
more sense before).
getLastStateChange is now uses the passed namespaces as a filter,
meaning that not requesting any namespace will give you the total last
updated state.
In addition, the sync.Map is exchanged for a variant that uses generics
which allows us to remove some casting logic.
Adds knobs to configure three aspects of the OpenID Connect flow:
* Custom scopes to override the default "openid profile email".
* Custom parameters to be added to the Authorize Endpoint request.
* Domain allowlisting for authenticated principals.
* User allowlisting for authenticated principals.
Websockets, in which DERP is based, requires a TLS certificate. At the same time,
if we use a certificate it must be valid... otherwise Tailscale wont connect (does not
have an Insecure option). So there is no option to expose insecure here
- registry file /windows/tailscale.reg is generated, filling in the
associated control server URL
- also includes CLI instructions
- fix /apple incorrect template: 'Url' is supposed to be '.URL'
This series of commit will be adding an embedded DERP server (and STUN) to Headscale,
thus making it completely self-contained and not dependant in other infrastructure.
This commit removes the two extra caches (oidc, requested time) and uses
the new central registration cache instead. The requested time is
unified into the main machine object and the oidc key is just added to
the same cache, as a string with the state as a key instead of machine
key.
current logic is not safe as it will allow an IP that isnt persisted to
the DB to be given out multiple times if machines joins in quick
succession.
This adds a lock around the "get ip" and machine registration and save
to DB so we ensure thiis isnt happning.
Currently this had to be done three places, which is silly, and outlined
in #294.
