Admit checked restricted artifacts without individual code approvals

restricted-execution/default.nix

@@ -20,6 +20,17 @@ let
     runtime = nativeProbe;
     executable = "${nativeProbe}/bin/probe";
   };
+  identityLauncher =
+    pkgs.runCommand "quixos-restricted-identity-launcher"
+      {
+        nativeBuildInputs = [ pkgs.stdenv.cc ];
+      }
+      ''
+        mkdir -p "$out/bin"
+        $CC -O2 -Wall -Wextra -Werror ${./identity-launcher.c} \
+          -DSANDBOX_EXECUTABLE='"${sandbox}/bin/quixos-restricted-sandbox"' \
+          -o "$out/bin/quixos-restricted-engine"
+      '';
   # Transaction-port/value adapters do not need the browser's CRDT/Wasm engine.
   # Keep this surface explicit; it is bundled into checked restricted handlers.
   sdk = pkgs.runCommand "quixos-restricted-sdk" { nativeBuildInputs = [ pkgs.esbuild ]; } ''
@@ -61,6 +72,7 @@ in
   inherit
     workerd
     sandbox
+    identityLauncher
     nativeProbeSandbox
     sdk
     runtime

restricted-execution/identity-launcher.c

@@ -0,0 +1,88 @@
+#define _GNU_SOURCE
+#include <dirent.h>
+#include <fcntl.h>
+#include <grp.h>
+#include <pwd.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/prctl.h>
+#include <sys/resource.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+/* Installed only as a narrowly executable NixOS wrapper. This trusted helper
+ * accepts no executable, environment, filesystem root or arbitrary UID. The
+ * caller is platform authority, never a workspace guest or authored Worker. */
+static void fail(const char *message) {
+  fprintf(stderr, "RESTRICTED_IDENTITY: %s\n", message);
+  exit(125);
+}
+
+int main(int argc, char **argv) {
+  const char *root = "/run/quixos-restricted-engines/";
+  uid_t caller = getuid();
+  gid_t group = getgid();
+  struct passwd *owner = getpwnam("quixos-graph");
+  if (geteuid() != 0) fail("privileged wrapper unavailable");
+  if (!owner) fail("authority identity unavailable");
+  if (caller != owner->pw_uid) fail("caller identity mismatch");
+  if (argc != 3) fail("invalid arguments");
+  char *end;
+  unsigned long number = strtoul(argv[2], &end, 10);
+  if (!*argv[2] || *end || number < 610000 || number >= 610064)
+    fail("invalid engine identity");
+  uid_t identity = (uid_t)number;
+  if (strlen(argv[1]) != strlen(root) + 16 || strncmp(argv[1], root, strlen(root)))
+    fail("invalid engine directory");
+  const char *name = argv[1] + strlen(root);
+  if (strncmp(name, "qx-engine-", 10) ||
+      strlen(name) != 16 || strspn(name + 10, "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789") != 6)
+    fail("invalid engine directory");
+  int parent = open("/run/quixos-restricted-engines", O_RDONLY | O_DIRECTORY | O_NOFOLLOW);
+  struct stat info;
+  if (parent < 0 || fstat(parent, &info) || info.st_uid != caller || (info.st_mode & 066))
+    fail("invalid engine root");
+  int directory = openat(parent, name, O_RDONLY | O_DIRECTORY | O_NOFOLLOW);
+  if (directory < 0 || fstat(directory, &info) || info.st_uid != caller || (info.st_mode & 077))
+    fail("invalid engine owner");
+  DIR *entries = fdopendir(dup(directory));
+  if (!entries) fail("engine entries");
+  struct dirent *entry;
+  unsigned int files = 0;
+  while ((entry = readdir(entries))) {
+    const char *file = entry->d_name;
+    if (!strcmp(file, ".") || !strcmp(file, "..")) continue;
+    int socket = !strcmp(file, "broker.sock");
+    if ((!socket && strcmp(file, "config.capnp") && strcmp(file, "parent.mjs")) ||
+        fstatat(directory, file, &info, AT_SYMLINK_NOFOLLOW) || info.st_uid != caller ||
+        info.st_nlink != 1 || (socket ? !S_ISSOCK(info.st_mode) : !S_ISREG(info.st_mode)) ||
+        info.st_size > 4 * 1024 * 1024)
+      fail("unexpected engine file");
+    files++;
+  }
+  closedir(entries);
+  if (files != 3) fail("incomplete engine configuration");
+  const char *files_to_transfer[] = {"config.capnp", "parent.mjs", "broker.sock"};
+  for (size_t i = 0; i < 3; i++) {
+    if (fchownat(directory, files_to_transfer[i], identity, group, AT_SYMLINK_NOFOLLOW) ||
+        fchmodat(directory, files_to_transfer[i], 0660, 0))
+      fail("engine file ownership");
+  }
+  if (fchown(directory, identity, group) || fchmod(directory, 0770))
+    fail("engine directory ownership");
+  close(directory);
+  close(parent);
+  struct rlimit core = {0, 0};
+  if (setrlimit(RLIMIT_CORE, &core) || setgroups(0, NULL) ||
+      setresgid(identity, identity, identity) || setresuid(identity, identity, identity) ||
+      prctl(PR_SET_DUMPABLE, 0) || prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0))
+    fail("identity confinement");
+  /* The containing directory grants only its owner and authority access. A
+   * socket created by workerd must also admit authority's different UID/GID. */
+  umask(0000);
+  char *args[] = {SANDBOX_EXECUTABLE, argv[1], NULL};
+  clearenv();
+  execv(SANDBOX_EXECUTABLE, args);
+  fail("sandbox launch");
+}

restricted-execution/identity-module.nix

@@ -0,0 +1,28 @@
+{
+  executor,
+  enable,
+  controlGroup,
+}:
+{ config, lib, ... }:
+{
+  config = lib.mkIf enable {
+    assertions = [
+      {
+        assertion = lib.all (user: user.uid == null || user.uid < 610000 || user.uid >= 610064) (
+          builtins.attrValues config.users.users
+        );
+        message = "UIDs 610000–610063 are reserved for restricted engine identities.";
+      }
+    ];
+    systemd.tmpfiles.rules = [
+      "d /run/quixos-restricted-engines 0711 quixos-graph ${controlGroup} -"
+    ];
+    security.wrappers.quixos-restricted-engine = {
+      source = "${executor.identityLauncher}/bin/quixos-restricted-engine";
+      owner = "root";
+      group = controlGroup;
+      permissions = "u+rx,g+rx,o-rwx";
+      setuid = true;
+    };
+  };
+}

restricted-execution/sandbox.c

@@ -19,7 +19,7 @@ int main(int argc, char **argv) {
   struct stat info;
   if (argc != 2 || argv[1][0] != '/' || strstr(argv[1], "/../") ||
       lstat(argv[1], &info) || !S_ISDIR(info.st_mode) ||
-      info.st_uid != getuid() || (info.st_mode & 077))
+      info.st_uid != getuid() || (info.st_mode & 007))
     fail("invalid engine directory");
   char gate;
   if (read(STDIN_FILENO, &gate, 1) != 1 || gate != 1)

restricted-execution/service.mjs

@@ -42,6 +42,8 @@ export class RestrictedEngineBudget {
         })());
       release.attach = resource.attach;
       release.usage = resource.usage;
+      release.kill = resource.kill;
+      release.uid = resource.uid;
       return release;
     } catch (error) {
       this.#leases.delete(lease);
@@ -82,6 +84,7 @@ export async function serveRestrictedRegistry(config, io = process) {
     !absolute(config.worker) ||
     !absolute(config.module) ||
     (config.sandbox !== undefined && !absolute(config.sandbox)) ||
+    (config.sandboxIdentity !== undefined && typeof config.sandboxIdentity !== "boolean") ||
     !/^[$A-Z_a-z][$\w]*$/.test(config.registryExport) ||
     !Array.isArray(config.exports) ||
     !config.exports.length ||
@@ -280,10 +283,13 @@ export async function serveRestrictedRegistry(config, io = process) {
       if (index >= 0) idleEngines.splice(index, 1);
       engine.current = null;
       if (engine.child?.pid) {
-        try {
-          process.kill(-engine.child.pid, "SIGKILL");
-        } catch (error) {
-          if (error.code !== "ESRCH") throw error;
+        if (engine.release.kill) await engine.release.kill();
+        else {
+          try {
+            process.kill(-engine.child.pid, "SIGKILL");
+          } catch (error) {
+            if (error.code !== "ESRCH") throw error;
+          }
         }
         await engine.exited;
       }
@@ -397,9 +403,13 @@ export async function serveRestrictedRegistry(config, io = process) {
     const file = join(engine.directory, "config.capnp");
     await writeFile(file, configuration);
     if (!call.open) return;
+    if (config.sandboxIdentity && !Number.isSafeInteger(engine.release.uid))
+      throw fixedError("RESTRICTED_IDENTITY_UNAVAILABLE");
     engine.child = spawn(
       config.sandbox ?? config.workerd,
-      config.sandbox ? [engine.directory] : ["serve", "--experimental", file],
+      config.sandbox
+        ? [engine.directory, ...(config.sandboxIdentity ? [String(engine.release.uid)] : [])]
+        : ["serve", "--experimental", file],
       {
         detached: true,
         env: { TZ: "UTC", LANG: "C.UTF-8" },
@@ -414,16 +424,16 @@ export async function serveRestrictedRegistry(config, io = process) {
     });
     engine.child.stdin?.on("error", () => engine.current?.terminate("RESTRICTED_START_FAILED"));
     call.child = engine.child;
-    await engine.release.attach?.(engine.child.pid);
-    if (config.sandbox && call.open) engine.child.stdin.end(Buffer.from([1]));
-    watchCPU(call);
     let diagnostics = 0;
     engine.child.stderr.on("data", (chunk) => {
       if ((diagnostics += chunk.length) > 65536) {
         if (engine.current?.open) engine.current.terminate("RESTRICTED_DIAGNOSTIC_LIMIT");
         else void destroyEngine(engine).catch(fail);
-      }
+      } else config.diagnostics?.(chunk);
     });
+    await engine.release.attach?.(engine.child.pid);
+    if (config.sandbox && call.open) engine.child.stdin.end(Buffer.from([1]));
+    watchCPU(call);
     engine.child.once("error", () => engine.current?.terminate("RESTRICTED_START_FAILED"));
     void engine.exited.then(() => {
       if (engine.current?.open) engine.current.terminate("RESTRICTED_ENGINE_EXITED");
@@ -535,6 +545,7 @@ export async function serveRestrictedRegistry(config, io = process) {
       await finish(call, fixedError("INVOCATION_ENDED"), reply);
       metrics.completed++;
     } catch (error) {
+      config.diagnostics?.(Buffer.from(`RESTRICTED_HOST_ERROR:${error.code ?? error.name}\n`));
       if (call.open) {
         const code = /^RESTRICTED_[A-Z_]+$/.test(error.code ?? "") ? error.code : "RESTRICTED_PACKAGE_FAILED";
         await finish(call, error, { invocation: message.invocation, status: "failed", error: { code, message: code } });

restricted-execution/service.test.mjs

@@ -437,15 +437,21 @@ test("bounded FIFO rejects overload without losing active or queued frames", asy
 test("queue wait consumes the deadline and cancelled work never reaches its ports", async () => {
   const f = await fixture(
     registry("return await channel({portId:'read',action:{kind:'state',primitive:'read'},input:null});"),
-    { deadlineMs: 350, maxActive: 1 },
+    { deadlineMs: 4000, maxActive: 1 },
   );
   try {
     const first = f.start();
-    await f.receive();
-    const queued = f.start();
-    const replies = [await f.receive(), await f.receive()];
-    assert.deepEqual(new Set(replies.map((x) => x.invocation)), new Set([first.invocation, queued.invocation]));
-    assert.ok(replies.every((x) => x.error.code === "RESTRICTED_DEADLINE"));
+    assert.equal((await f.receive()).kind, "port");
+    // Expire while the first call is deliberately blocked. Two equal relative
+    // deadlines let fast startup legitimately begin the queued call first.
+    const queued = f.start(null, { root: "queued-deadline", deadlineMs: 50, depth: 1, calls: 10 });
+    const expired = await f.receive();
+    assert.equal(expired.invocation, queued.invocation);
+    assert.equal(expired.error.code, "RESTRICTED_DEADLINE");
+    f.send({ kind: "cancel", ...first });
+    const cancelled = await f.receive();
+    assert.equal(cancelled.invocation, first.invocation);
+    assert.equal(cancelled.kind, "cancelled");
     const fresh = f.start();
     const port = await f.receive();
     assert.equal(port.invocation, fresh.invocation);

restricted-execution/workerd.nix

@@ -1,6 +1,6 @@
 # Pinned upstream engine. This binary supplies an import/capability boundary;
-# production execution additionally requires OS limits and platform approval
-# for shared placement, or a workspace VM for unreviewed application code.
+# production execution additionally requires the qualified OS identity,
+# namespace, syscall, cgroup and invocation-broker boundaries.
 { pkgs }:
 let
   version = "1.20261005.1";