[ upstream commit 34127e6 ]

The KPR guide contains the autoDirectNodeRoutes option in most Helm
commands, but that option isn't a requirement for KPR subfeatures and
may even fail if Kubernetes nodes are not L2-connected.

Signed-off-by: Paul Chaignon <paul@cilium.io>

[ upstream commit 4c9c1d3 ]

This commit fixes an edge case in the `NodesPodCIDRManager`.
If there were any nodes on operator startup which have no PodCIDRs, the
operator would sometimes assign PodCIDRs to these nodes which have
already been allocated to other nodes.

The operator assumed that when `k8sCiliumNodesCacheSynced` closes, all
node events have been processed. And it proceeds to call `Resync` on
the `nodeManager`.

The `NodesPodCIDRManager` will queue any nodes
without PodCIDRs to be allocated once the `canAllocatePodCIDRs` variable
is set. This variable is set by the `Resync`.

So, the assumption/expected behavior is that the
`NodesPodCIDRManager.Update` function has been called for all nodes in
the cache before `Resync` is called. However, this wasn't the case.
The `startSynchronizingCiliumNodes` function starts the informer
and connects the nodeManager to it. But instead of handling the events
at once, the callbacks enqueue the events, to be handled by a separate
go routine. This means that `k8sCiliumNodesCacheSynced` is closed once
all of the node events are enqueued, not when they have been processed
by the `nodeManager`.

This commit fixes this behavior by processing all events at once
in the informer callbacks until the full sync is complete, at which
point we will switch over to using the workqueue.

Fixes: cilium#21482

Signed-off-by: Dylan Reimerink <dylan.reimerink@isovalent.com>
Signed-off-by: Paul Chaignon <paul@cilium.io>