From 41a54e8a0f1451ebc8e14d58b1df66650bd9ecb8 Mon Sep 17 00:00:00 2001
From: Andrey Smirnov <andrey.smirnov@siderolabs.com>
Date: Mon, 8 Apr 2024 16:18:21 +0400
Subject: [PATCH] fix: pre-create nftables chain to make kubelet use nftables

In Talos, kubelet (and kube-proxy) images use `iptables-wrapper` script
to detect which version of `iptables` (legacy or NFT) to use.

The script assumes that `kubelet` runs on the host, and uses whatever
version of `iptables` which is being used by the host. In Talos,
`kubelet` runs in a container which has same `iptables-wrapper` script,
and it defaults to `legacy` mode in our case.

We can't check the `kubelet` image, as it would affect all Talos
version, so instead pre-create the chains/tables in `nftables` so that
kubelet will pick up `nft` version of `iptables`, and `kube-proxy` will
do the same.

Without this fix, the problem arises from the mix of `nft` used by Talos
for the firewall and Kubernetes world relying on `legacy` (`xtables`).

Fixes https://github.com/siderolabs/kubelet/issues/77

See https://github.com/kubernetes-sigs/iptables-wrappers/blob/e139a115350974aac8a82ec4b815d2845f86997e/iptables-wrapper-installer.sh#L102-L130

Signed-off-by: Andrey Smirnov <andrey.smirnov@siderolabs.com>
(cherry picked from commit ff2c427b04963d69ba2eaa1084a0a078d742b9ac)
---
 hack/release.toml                             | 10 ++++++
 .../pkg/controllers/network/nftables_chain.go | 35 +++++++++++++++++++
 2 files changed, 45 insertions(+)

diff --git a/hack/release.toml b/hack/release.toml
index dfa5cabda7..f806f34a7c 100644
--- a/hack/release.toml
+++ b/hack/release.toml
@@ -212,6 +212,16 @@ machine:
         title = "Platforms"
         description = """\
 Talos Linux now supports [Akamai Connected Cloud](https://www.linode.com/) provider (platform `akamai`).
+"""
+
+    [notes.iptables]
+        title = "IPTables"
+        description = """\
+Talos Linux now forces `kubelet` and `kube-proxy` to use `iptables-nft` instead of `iptables-legacy` (`xtables`) which was the default
+before Talos 1.7.0.
+
+Container images based on `iptables-wrapper` should work without changes, but if there was a direct call to `legacy` mode of `iptables`, make sure
+to update to use `iptables-nft`.
 """
 
 [make_deps]
diff --git a/internal/app/machined/pkg/controllers/network/nftables_chain.go b/internal/app/machined/pkg/controllers/network/nftables_chain.go
index 26aa0f3b08..9fbbc70a10 100644
--- a/internal/app/machined/pkg/controllers/network/nftables_chain.go
+++ b/internal/app/machined/pkg/controllers/network/nftables_chain.go
@@ -65,6 +65,10 @@ func (ctrl *NfTablesChainController) Run(ctx context.Context, r controller.Runti
 
 		var conn nftables.Conn
 
+		if err := ctrl.preCreateIptablesNFTable(logger, &conn); err != nil {
+			return fmt.Errorf("error pre-creating iptables-nft table: %w", err)
+		}
+
 		list, err := safe.ReaderListAll[*network.NfTablesChain](ctx, r)
 		if err != nil {
 			return fmt.Errorf("error listing nftables chains: %w", err)
@@ -176,3 +180,34 @@ func (ctrl *NfTablesChainController) Run(ctx context.Context, r controller.Runti
 		r.ResetRestartBackoff()
 	}
 }
+
+func (ctrl *NfTablesChainController) preCreateIptablesNFTable(logger *zap.Logger, conn *nftables.Conn) error {
+	// Pre-create the iptables-nft table, if it doesn't exist.
+	// This is required to ensure that the iptables universal binary prefers iptables-nft over
+	// iptables-legacy can be used to manage the nftables rules.
+	tables, err := conn.ListTablesOfFamily(nftables.TableFamilyIPv4)
+	if err != nil {
+		return fmt.Errorf("error listing existing nftables tables: %w", err)
+	}
+
+	if slices.IndexFunc(tables, func(t *nftables.Table) bool { return t.Name == "mangle" }) != -1 {
+		return nil
+	}
+
+	table := &nftables.Table{
+		Family: nftables.TableFamilyIPv4,
+		Name:   "mangle",
+	}
+	conn.AddTable(table)
+
+	chain := &nftables.Chain{
+		Name:  "KUBE-IPTABLES-HINT",
+		Table: table,
+		Type:  nftables.ChainTypeNAT,
+	}
+	conn.AddChain(chain)
+
+	logger.Info("pre-created iptables-nft table 'mangle'/'KUBE-IPTABLES-HINT'")
+
+	return nil
+}