We shouldn't be replacing the configured server address on agents. Doing
so breaks the agent's ability to fall back to the fixed registration
endpoint when all servers are down, since we replaced it with the first
discovered apiserver address. The fixed registration endpoint will be
restored as default when the service is restarted, but this is not the
correct behavior. This should have only been done on etcd-only nodes
that start up using their local supervisor, but need to switch to a
control-plane node as soon as one is available.
Signed-off-by: Brad Davidson <brad.davidson@rancher.com>
If proxy.SetAPIServerPort was called multiple times, all calls after the
first one would cause the apiserver address to be set to the default
server address, bypassing the local load-balancer. This was most likely
to occur on RKE2, where the supervisor may be up for a period of time
before it is ready to manage node password secrets, causing the agent
to retry.
Signed-off-by: Brad Davidson <brad.davidson@rancher.com>
* Adds support for health-checking loadbalancer servers. If a
health-check fails when dialing, all existing connections to the
server will be closed.
* Wires up a remotedialer tunnel connectivity check as the health check
for supervisor/apiserver connections.
* Wires up a simple ping request to the supervisor port as the health
check for etcd connections.
Signed-off-by: Brad Davidson <brad.davidson@rancher.com>
Don't set up the agent tunnel authorizer on agentless servers, and warn when agentless servers won't have a way to reach in-cluster endpoints.
Signed-off-by: Brad Davidson <brad.davidson@rancher.com>
* Wait for kubelet port to be ready before setting
* Wait for kubelet to update the Ready status before reading port
Signed-off-by: Daishan Peng <daishan@acorn.io>
Signed-off-by: Brad Davidson <brad.davidson@rancher.com>
Co-authored-by: Brad Davidson <brad.davidson@rancher.com>
The control-plane context handles requests outside the cluster and
should not be sent to the proxy.
In agent mode, we don't watch pods and just direct-dial any request for
a non-node address, which is the original behavior.
Signed-off-by: Brad Davidson <brad.davidson@rancher.com>
Watching pods appears to be the most reliable way to ensure that the
proxy routes and authorizes connections.
Signed-off-by: Brad Davidson <brad.davidson@rancher.com>
Reduces code complexity a bit and ensures we don't have to handle closed watch channels on our own
Signed-off-by: Brad Davidson <brad.davidson@rancher.com>
Since we now start the server's agent sooner and in the background, we
may need to wait longer than 30 seconds for the apiserver to become
ready on downstream projects such as RKE2.
Since this essentially just serves as an analogue for the server's
apiReady channel, there's little danger in setting it to something
relatively high.
Signed-off-by: Brad Davidson <brad.davidson@rancher.com>
This attempts to update logging statements to make them consistent
through out the code base. It also adds additional context to messages
where possible, simplifies messages, and updates level where necessary.
In k3s today the kubernetes API and the /v1-k3s API are combined into
one http server. In rke2 we are running unmodified, non-embedded Kubernetes
and as such it is preferred to run k8s and the /v1-k3s API on different
ports. The /v1-k3s API port is called the SupervisorPort in the code.
To support this separation of ports a new shim was added on the client in
then pkg/agent/proxy package that will launch two load balancers instead
of just one load balancer. One load balancer for 6443 and the other
for 9345 (which is the supervisor port).
Will