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
@@ -25,7 +21,7 @@ jobs:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           API: ${{ github.api_url }}
           REPO: ${{ github.repository }}
-        run: nix run -L .#ci -- upsert-issue vulnix "vulnix: known vulnerabilities in system closure" "$RUNNER_TEMP/issue.md"
+        run: 'nix run -L .#ci -- upsert-issue vulnix "vulnix: known vulnerabilities in system closure" "$RUNNER_TEMP/issue.md"'
 
       - name: close resolved issue
         if: steps.scan.outputs.count == '0'

ci.nix

@@ -1,10 +1,17 @@
 { lib, pkgs }:
+let
+  # vulnix's bundled nix-2.34 broke its own --experimental-features detection
+  nixConf = pkgs.writeText "ci-nix.conf" "experimental-features = nix-command flakes";
+in
 # tools must track ci.py's subprocess calls; nix stays ambient (uses the
 # runner's own nix, not a pinned one).
 pkgs.writers.writePython3Bin "ci" {
   flakeIgnore = [ "E501" ];
   libraries = [ pkgs.python3Packages.subprocess-tee ];
   makeWrapperArgs = [
+    "--set-default"
+    "NIX_USER_CONF_FILES"
+    (toString nixConf)
     "--prefix"
     "PATH"
     ":"

ci.py

@@ -18,7 +18,8 @@ def run(*args, **kw):
 
 
 def nix(*args):
-    return subprocess_tee.run(("nix", "--extra-experimental-features", "nix-command flakes", "-L", *args), check=True, text=True).stdout
+    # tee to stderr - stdout leaks into GITHUB_OUTPUT $() captures
+    return subprocess_tee.run(("nix", "--extra-experimental-features", "nix-command flakes", "-L", *args), check=True, text=True, stdout=sys.stderr).stdout
 
 
 def git(*args, check=True):
@@ -32,13 +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
 
 
@@ -235,7 +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/git.nix

@@ -163,10 +163,9 @@ in
             trusted-public-keys = [ "microvm.cachix.org-1:oXnBc6hRE3eX5rSYdRyMYXnfzcCxC7yKPTbZXALsqys=" ];
           };
 
-          nix.gc = {
+          nix.settings = {
-            automatic = true;
+            min-free = 5 * 1024 * 1024 * 1024;
-            dates = "daily";
+            max-free = 10 * 1024 * 1024 * 1024;
-            options = "--delete-older-than 14d";
           };
           # nix.optimise.automatic = true; - doesn't work with microvm.writableStoreOverlay
 

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

@@ -222,10 +222,14 @@ rec {
     : Bind mounts, keyed by mount point.
 
     `resources.devices`
-    : Host device nodes to expose, as paths. Bound into the container's `/dev`
+    : Host device nodes to expose, as paths. Bound into the container at the
-      and added to the unit's `DeviceAllow=` - the bind alone leaves the cgroup
+      same path and added to the unit's `DeviceAllow=` - the cgroup
-      denying `open()`. Nodes keep their host ownership, so a consumer inside
+      permission alone doesn't get you a working node (nothing can `mknod` a
-      needs `isolation.privateUsers = "identity"` for the owning gid to mean
+      device inside a user namespace), and the bind alone leaves the cgroup
+      denying `open()`, so both are needed. The bind is unconditional, so
+      the device must actually exist on whatever host runs the container.
+      Nodes keep their host ownership, so a consumer inside needs
+      `isolation.privateUsers = "identity"` for the owning gid to mean
       anything.
 
     `isolation.nixDaemon`
@@ -285,7 +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=")
@@ -293,6 +297,9 @@ 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

tests/server-n0/integration.py

@@ -337,6 +337,24 @@ with subtest("(functional): declared devices are usable in-container"):
             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, [])):