This adds functionality to ensure that connections are closed on a
worker when the server is shut down or reloaded.
Connections are closed, and then the shutdown process waits for the next
report to the controller to go through, waiting the grace period that
was added in #1330. The rest of shutdown is then proceeded with.
Listener shutdown has been moved to the start of the process to ensure
that new connections can't be made while we are shutting down.
Additionally, functionality has been added to force shut down servers
after a second interrupt has been received after the first one. This
takes the form of code in both the server and dev mode code paths that
just performs an os.Exit if another shutdown is received during the
normal shutdown process.
Additionally, the commit also includes some refactoring to tests so that
we could use some of the new E2E testing code to test session
connections during shutdown and reload.
Additionally, this fixes a panic in worker startup when there are no
controllers defined. Looks like the shutdown call was attempting to call
into the controller, not the worker.
Finally, this also removes the extra shutdown within the outer worker
startup.
Shutdown is actually handled within the inner StartWorker(), attempting
shutdown on errors was actually leading to potential use of an
uninitialized worker, which is leading to panics in testing.
This commit adds the support to do the following:
* Mark connections for non-reporting workers as closed. This is the
controller counterpart to the worker functionality (see #1330). This is
written as a scheduled job that does the work DB-side in a single atomic
query.
* Works to reconcile states if such a broken controller-worker
connection resumes and a worker reports a connection as connected that
should be disconnected. In this case, the controller will send an update
request, and the worker will honor it and terminate the connection.
* Further refinement of the grace period setting has been added here.
We have converged on the current server "liveness" setting as our
default here, which is half of the previous 30s (15 seconds, in other
words). Additionally, this is now configurable on the controller and
worker side, with the caveat that it's currently impossible to do so in
config as the setting has been untagged in HCL. This is exposed so that
we can run some sophisticated testing scenarios where we skew the grace
period to either the controller or worker to ensure the aforementioned
reconciliation works.
* Some repository functions have been added to support the new
functionality, in addition to some test code to the worker to allow
querying of session state while testing.
* Finally, we've added some add timestamp subtraction functions
as well, basically serving as the opposite of the addition functions.
This commit adds some rudimentary functionality for cleaning up
connections on controller failure, or more specifically, a failure to
communicate with the controllers.
In this scenario, when a status cannot be sent for any reason for longer
than 30 seconds, the worker will close all connections that are normally
not cleaned up (so for sessions that currently are *not* in the progress
of being canceled or terminated). Status requests now time out after 10
seconds if they are not made successfully.
Session status is not touched during this process, but natural cleanup
or connectionless sessions will still occur. This should be okay, as the
session should be added on re-connect if need be.
A few other changes are made to cancelConnections: first, error logging
has been moved to the function so that we can handle errors requesting
cancellation directly within the function; this entails simply just
locally marking these connections as closed with a warning. Second,
there was a minor logging bug where closed connections were not properly
being noted in the last trace log, this has been fixed.
Additional changes:
* Circle: Set acceptance test no_output_timeout to 30m
BK clued me into the fact that for some types, similar alignment
problems exist as with the normal atomic package (e.g. 64-bit types on
32-bit architectures) if they are not pointer values. So change the few
places they weren't pointers to pointers.
None of these are 64-bit, but it feels like best practice for using the
library and should just standardize on it.
This was copied from Vault code but shouldn't be necessary any more as a
change in hclog a while back meant that derived loggers share the log
level of the parent.
Chris Marchesi
Jim
Jeff Mitchell