From b083c03d5e3e3b7c44f4ba90243c5f12a03952b8 Mon Sep 17 00:00:00 2001
From: autumn <aut.tea7@gmail.com>
Date: Mon, 14 Sep 2026 01:47:43 +0000
Subject: [PATCH] bug: fix kvm device not being properly forwarded

Co-authored-by: Autumn Liu <al63244@my.utexas.edu>
Reviewed-on: https://git.floweytf.com/server/server-config/pulls/9
Co-authored-by: autumn <aut.tea7@gmail.com>
---
 dsl.nix                        | 22 ++++++++++++++--------
 tests/server-n0/integration.py | 19 +++++++++++++++++++
 2 files changed, 33 insertions(+), 8 deletions(-)

diff --git a/dsl.nix b/dsl.nix
index 4745527..2af94fe 100644
--- a/dsl.nix
+++ b/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
-      `DeviceAllow=`. nspawn populates a handful of privileged devices
-      (`/dev/kvm`, `/dev/net/tun`, ...) into the container's own `/dev` once
-      the cgroup allows it, and silently skips ones the host doesn't have -
-      no `resources.fs` bind mount needed. Nodes keep their host ownership,
-      so a consumer inside needs `isolation.privateUsers = "identity"` for
-      the owning gid to mean anything.
+    : Host device nodes to expose, as paths. Bound into the container at the
+      same path and added to the unit's `DeviceAllow=` - the cgroup
+      permission alone doesn't get you a working node (nothing can `mknod` a
+      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`
     : Whether the Nix daemon and CLI are available inside the container.
@@ -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 =
diff --git a/tests/server-n0/integration.py b/tests/server-n0/integration.py
index b5f95c8..1c8c41e 100644
--- a/tests/server-n0/integration.py
+++ b/tests/server-n0/integration.py
@@ -333,9 +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:
+            # 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, [])):