feat(kata): expose nested KVM to sandboxes

Author: zhangjaycee

rock/deployments/docker.py

@@ -180,6 +180,8 @@ def _build_runtime_args(self) -> list[str]:
         if self._config.use_kata_runtime:
             return [
                 "--cap-add=ALL",
+                "--device-cgroup-rule",
+                "c 10:232 rwm",
                 "--security-opt",
                 "seccomp=unconfined",
                 "--runtime=io.containerd.kata.v2",

rock/rocklet/local_files/docker_run.sh

@@ -28,6 +28,27 @@ is_nix() {
     fi
 }
 
+# Kata guest KVM is a misc character device. Its number belongs to the guest
+# kernel and can differ from the outer host's /dev/kvm device number.
+setup_kata_guest_kvm() {
+    local kvm_major=10
+    local kvm_minor=232
+    local registered_minor
+
+    if [ ! -r /proc/misc ]; then
+        echo "Kata guest KVM setup failed: /proc/misc is unavailable." >&2
+        return 1
+    fi
+    registered_minor=$(awk '$2 == "kvm" { print $1; exit }' /proc/misc)
+    if [ "${registered_minor}" != "${kvm_minor}" ]; then
+        echo "Kata guest KVM setup failed: expected misc device ${kvm_major}:${kvm_minor}, found minor ${registered_minor:-none}." >&2
+        return 1
+    fi
+
+    rm -f /dev/kvm
+    mknod -m 600 /dev/kvm c "${kvm_major}" "${kvm_minor}"
+}
+
 # Kata DinD: set up loop device and mount disk image for Docker storage
 setup_kata_dind() {
     local docker_root="/var/lib/docker"
@@ -79,7 +100,8 @@ if [ "$(is_nix)" = "true" ]; then
 fi
 
 if [ "${ROCK_KATA_RUNTIME}" = "true" ]; then
-    echo "Kata runtime detected, setting up DinD disk..."
+    echo "Kata runtime detected, setting up guest KVM and DinD disk..."
+    setup_kata_guest_kvm
     setup_kata_dind
 fi
 

tests/unit/deployments/test_docker_env_injection.py

@@ -99,3 +99,22 @@ def test_kata_runtime_env_absent_when_disabled(self):
             mock_env.ROCK_TIME_ZONE = "UTC"
             args = deployment._build_env_args()
         assert "ROCK_KATA_RUNTIME=true" not in args
+
+
+class TestBuildRuntimeArgs:
+    def test_kata_runtime_grants_guest_kvm_device_number(self):
+        deployment = _make_deployment(use_kata_runtime=True)
+
+        args = deployment._build_runtime_args()
+
+        rule_index = args.index("--device-cgroup-rule")
+        assert args[rule_index + 1] == "c 10:232 rwm"
+        assert "--device=/dev/kvm" not in args
+
+    def test_default_runtime_does_not_grant_guest_kvm_device(self):
+        deployment = _make_deployment(use_kata_runtime=False)
+
+        args = deployment._build_runtime_args()
+
+        assert args == ["--privileged"]
+        assert "--device-cgroup-rule" not in args