actually bind /dev/kvm into the container, and test it initializes

DeviceAllow= only grants cgroup permission to open a device node - it
doesn't create one, and nspawn can't mknod a device inside a user
namespace. The node needs an explicit bind from the host as well, so
both are required together.

Also add a test that actually boots qemu-system-x86_64 -enable-kvm
(no disk, killed by timeout) and checks for "Permission denied" /
"Could not access KVM" in its output, rather than just checking the
node opens - a wrong-gid or unmapped node still opens but fails the
ioctls qemu needs.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dzy7tEKLCFFXD96LmR8we3

.github/workflows/vulnix.yml

@@ -5,10 +5,6 @@ on:
     - cron: "0 6 * * 1"
   workflow_dispatch:
 
-permissions:
-  contents: read
-  issues: write
-
 jobs:
   scan:
     runs-on: nix

ci.py

@@ -33,13 +33,16 @@ def gitea(method, path, body=None):
     req.add_header("Authorization", f'token {os.environ["GH_TOKEN"]}')
     if data is not None:
         req.add_header("Content-Type", "application/json")
+    try:
         with urllib.request.urlopen(req) as resp:
             raw = resp.read()
+    except urllib.error.HTTPError as e:
+        sys.exit(f"gitea {method} {path}: {e.code} {e.reason}\n{e.read().decode()}")
     return json.loads(raw) if raw else None
 
 
 def find_open_issue(label):
-    issues = gitea("GET", f"/issues?state=open&labels={label}")
+    issues = gitea("GET", f"/issues?state=open&type=issues&labels={label}")
     return issues[0] if issues else None
 
 
@@ -236,7 +239,7 @@ def format_vulnix(vulns):
     for v in vulns:
         for cve in v["affected_by"]:
             score = v.get("cvssv3_basescore", {}).get(cve)
-            desc = v.get("description", {}).get(cve, "")
+            desc = v.get("description", {}).get(cve, "").replace("\n", " ").replace("|", "∣")
             if len(desc) > 120:
                 desc = desc[:117] + "..."
             rows.append((v["pname"], v["version"], cve, score, desc))

containers/server-n0/ssh.nix

@@ -75,7 +75,6 @@ let
       packages = with pkgs; [ tmux ] ++ packages;
       miscProperties = self._sshpiperRegister routingInfo;
     };
-
 in
 {
   ssh-admin = mkSsh {

dsl.nix

@@ -222,13 +222,15 @@ rec {
     : Bind mounts, keyed by mount point.
 
     `resources.devices`
-    : Host device nodes to expose, as paths, added to the unit's
+    : Host device nodes to expose, as paths. Bound into the container at the
-      `DeviceAllow=`. nspawn populates a handful of privileged devices
+      same path and added to the unit's `DeviceAllow=` - the cgroup
-      (`/dev/kvm`, `/dev/net/tun`, ...) into the container's own `/dev` once
+      permission alone doesn't get you a working node (nothing can `mknod` a
-      the cgroup allows it, and silently skips ones the host doesn't have -
+      device inside a user namespace), and the bind alone leaves the cgroup
-      no `resources.fs` bind mount needed. Nodes keep their host ownership,
+      denying `open()`, so both are needed. The bind is unconditional, so
-      so a consumer inside needs `isolation.privateUsers = "identity"` for
+      the device must actually exist on whatever host runs the container.
-      the owning gid to mean anything.
+      Nodes keep their host ownership, so a consumer inside needs
+      `isolation.privateUsers = "identity"` for the owning gid to mean
+      anything.
 
     `isolation.nixDaemon`
     : Whether the Nix daemon and CLI are available inside the container.
@@ -287,7 +289,7 @@ rec {
       allBindMounts = builtins.foldl' (acc: t: acc // (t.resources.fs or { })) { } traits // bindMounts;
       userModules = map mkUser (lib.concatMap (t: t.users or [ ]) traits ++ users);
       allPorts = ports ++ map (p: (normalizePort p).container) hostPorts;
-      useIdmap = privateUsers != "no";
+      useIdmap = privateUsers == "pick" || (builtins.isInt privateUsers && privateUsers > 0);
       bindMountExtraFlags = lib.mapAttrsToList (
         mountPoint: bm:
         (if bm.isReadOnly then "--bind-ro=" else "--bind=")
@@ -295,6 +297,10 @@ rec {
         + lib.optionalString useIdmap ":idmap"
       ) allBindMounts;
       bindMountHostPaths = map (bm: bm.hostPath) (lib.attrValues allBindMounts);
+      # DeviceAllow= alone only grants cgroup permission - nspawn can't
+      # mknod a device node inside a user namespace, so the node still
+      # needs an explicit bind from the host.
+      deviceFlags = map (d: "--bind=${d}") devices;
       inherit (deriveIdentity ifPrefix.nspawn name)
         iface
         mac
@@ -381,7 +387,7 @@ rec {
       nixosContainer = {
         inherit autoStart privateUsers;
         bindMounts = { };
-        extraFlags = bindMountExtraFlags;
+        extraFlags = bindMountExtraFlags ++ deviceFlags;
         privateNetwork = true;
         extraVeths = linkVeths;
         config =

tests/server-n0/integration.py

@@ -333,15 +333,28 @@ with subtest("(functional): published ports follow the container's lifetime"):
 with subtest("(functional): declared devices are usable in-container"):
     for name, devices in DEVICES.items():
         for dev in devices:
-            if server.execute(f"test -e {dev}")[0] != 0:
-                # host lacks the device (e.g. /dev/kvm without nested
-                # virtualization, as in this test's own VM) - nspawn skips
-                # populating it in-container, so there's nothing to check.
-                continue
             # opened, not stat'd: without DeviceAllow= the node is still visible
             server.succeed(ct_cmd(name, f"exec 3<> {dev}"))
 
 
+with subtest("(functional): /dev/kvm actually initializes, not just opens"):
+    # a wrong-gid or unmapped node still opens but fails the ioctls qemu
+    # needs, so boot with no disk (never completes) and kill it - the
+    # failure mode this catches is qemu refusing to start at all.
+    for name, devices in DEVICES.items():
+        if "/dev/kvm" not in devices:
+            continue
+        out = server.succeed(
+            ct_cmd(
+                name,
+                f"{TIMEOUT} 3 qemu-system-x86_64 -enable-kvm -m 256 -nographic -no-reboot "
+                ">/tmp/qemu.out 2>&1; cat /tmp/qemu.out",
+            )
+        )
+        assert "Permission denied" not in out, f"{name}: qemu couldn't use /dev/kvm:\n{out}"
+        assert "Could not access KVM" not in out, f"{name}: qemu couldn't use /dev/kvm:\n{out}"
+
+
 with subtest("(isolation): devices reach only the container that declared them"):
     for name in enterable:
         for dev in {d for ds in DEVICES.values() for d in ds} - set(DEVICES.get(name, [])):