update

Co-authored-by: Autumn Liu <al63244@my.utexas.edu>
Co-authored-by: Floweynt <50057682+Floweynt@users.noreply.github.com>
Reviewed-on: #1
Co-authored-by: autumn <aut.tea7@gmail.com>

.github/workflows/vulnix.yml

@@ -25,7 +25,7 @@ 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):

containers/server-n0/git.nix

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

containers/server-n0/ssh.nix

@@ -61,16 +61,21 @@ let
       net ? { },
       resources ? { },
       traits ? [ ],
+      isolation ? { },
     }:
     mkContainer {
       inherit name;
       inherit net resources;
       svc.autoStart = true;
-      isolation.nixDaemon = nixDaemon;
+      isolation = {
+        inherit nixDaemon;
+      }
+      // isolation;
       traits = [ sshTrait ] ++ traits;
       packages = with pkgs; [ tmux ] ++ packages;
       miscProperties = self._sshpiperRegister routingInfo;
     };
+
 in
 {
   ssh-admin = mkSsh {
@@ -113,14 +118,28 @@ in
   ssh-generic = mkSsh {
     name = "ssh-generic";
     net.access = [ { to = "@internet"; } ];
+    # /dev/kvm keeps its host ownership, so the kvm gid must not be shifted.
+    isolation.privateUsers = "identity";
     resources = {
       fs."/home/runner/data" = {
         hostPath = "${data}/ssh/generic";
         isReadOnly = false;
       };
+      devices = [ "/dev/kvm" ];
       mem = "16G";
       cpu = 65;
     };
+    packages = with pkgs; [ qemu_kvm ];
+    traits = [
+      {
+        containerConfig =
+          { config, ... }:
+          {
+            users.groups.kvm.gid = config.ids.gids.kvm;
+            users.users.runner.extraGroups = [ "kvm" ];
+          };
+      }
+    ];
     routingInfo = [
       {
         username = "autumn";

dsl.nix

@@ -221,6 +221,13 @@ rec {
     `resources.fs`
     : Bind mounts, keyed by mount point.
 
+    `resources.devices`
+    : Host device nodes to expose, as paths. Bound into the container's `/dev`
+      and added to the unit's `DeviceAllow=` - the bind alone leaves the cgroup
+      denying `open()`. 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.
 
@@ -270,6 +277,7 @@ rec {
       cpuQuota = "${toString (cpu * 100)}%";
       memoryMax = resources.mem or "512M";
       bindMounts = resources.fs or { };
+      devices = resources.devices or [ ];
       allowsNixDaemon = isolation.nixDaemon or false;
       privateUsers = isolation.privateUsers or "pick";
 
@@ -285,6 +293,7 @@ rec {
         + lib.optionalString useIdmap ":idmap"
       ) allBindMounts;
       bindMountHostPaths = map (bm: bm.hostPath) (lib.attrValues allBindMounts);
+      deviceFlags = map (d: "--bind=${d}") devices;
       inherit (deriveIdentity ifPrefix.nspawn name)
         iface
         mac
@@ -363,6 +372,7 @@ rec {
         hostAddr
         bindMountHostPaths
         includedInTests
+        devices
         ;
       bindMounts = allBindMounts;
       kind = "nspawn";
@@ -370,7 +380,7 @@ rec {
       nixosContainer = {
         inherit autoStart privateUsers;
         bindMounts = { };
-        extraFlags = bindMountExtraFlags;
+        extraFlags = bindMountExtraFlags ++ deviceFlags;
         privateNetwork = true;
         extraVeths = linkVeths;
         config =
@@ -503,6 +513,7 @@ rec {
         ;
       inherit needsGateway;
       bindMountHostPaths = map (s: s.source) shares;
+      devices = [ ];
       kind = "vm";
       miscProperties = flattenMiscProp self miscProperties;
       microvmConfig = {
@@ -939,6 +950,7 @@ rec {
           startLimitBurst = 5;
           startLimitIntervalSec = 120;
           unitConfig.RequiresMountsFor = c.bindMountHostPaths;
+          serviceConfig.DeviceAllow = map (d: "${d} rw") c.devices;
         };
 
       # the host end of a link, for both kinds. nixos-containers only configures
@@ -1298,6 +1310,7 @@ rec {
         HOST_ADDR = ${pyDict pyStr (lib.mapAttrs (_: c: c.hostAddr) autoStartTestContainers)}
         IFACE = ${pyDict pyStr (lib.mapAttrs (_: c: c.iface) autoStartTestContainers)}
         KIND = ${pyDict pyStr (lib.mapAttrs (_: c: c.kind) autoStartTestContainers)}
+        DEVICES = ${pyDict pyList (lib.mapAttrs (_: c: c.devices) autoStartTestContainers)}
         # only nspawn guests can be entered with nsenter; a vm has no host-side
         # netns, so every probe that runs *inside* a guest is limited to these
         enterable = [c for c in included if KIND[c] == "nspawn"]

tests/server-n0/integration.py

@@ -27,6 +27,7 @@ if TYPE_CHECKING:
     HOST_ADDR: dict[str, str] = {}
     IFACE: dict[str, str] = {}
     KIND: dict[str, str] = {}
+    DEVICES: dict[str, list[str]] = {}
     EDGES: list[tuple[str, str, list[int], list[int]]] = []
     SSH_KEY: str = ""
     SSH_OPTS: str = ""
@@ -55,6 +56,11 @@ def container_pid(name):
     return server.succeed(f"machinectl show {name} -p Leader --value").strip()
 
 
+def ct_cmd(name, cmd):
+    pid = container_pid(name)
+    return f"{NSENTER} --mount=/proc/{pid}/ns/mnt --root=/proc/{pid}/root -- {SH} -c '{cmd}'"
+
+
 def net_cmd(name, cmd):
     pid = container_pid(name)
     return f"{NSENTER} --net=/proc/{pid}/ns/net -- {cmd}"
@@ -323,3 +329,22 @@ with subtest("(functional): published ports follow the container's lifetime"):
     server.succeed("systemctl start container@nginx")
     server.wait_for_unit("container@nginx.service")
     internet.wait_until_succeeds(ext_connect(80), timeout=30)
+
+with subtest("(functional): declared devices are usable in-container"):
+    for name, devices in DEVICES.items():
+        for dev in devices:
+            # opened, not stat'd: without DeviceAllow= the node is still visible
+            server.succeed(ct_cmd(name, f"exec 3<> {dev}"))
+
+
+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, [])):
+            server.fail(ct_cmd(name, f"test -e {dev}"))
+
+
+with subtest("(isolation): identity uid mapping keeps the user namespace"):
+    for name in (n for n, ds in DEVICES.items() if ds):
+        host_ns = server.succeed("readlink /proc/1/ns/user").strip()
+        ct_ns = server.succeed(f"readlink /proc/{container_pid(name)}/ns/user").strip()
+        assert host_ns != ct_ns, f"{name} shares the host user namespace"