fix(harness): verify socket permissions and report partial cleanup

This commit is contained in:
2026-10-09 15:06:18 +00:00
parent d92811ec1c
commit 3df85090ff
3 changed files with 291 additions and 48 deletions
+82 -45
View File
@@ -132,9 +132,12 @@ def _write_handle(path: Path, state: dict, *, exclusive: bool = False, on_create
os.close(fd) os.close(fd)
def _update_handle(path: Path, state: dict) -> None: def _update_handle(path: Path, state: dict, *, expected_identity: tuple[int, int] | None = None) -> None:
fd = _open_handle(path, os.O_WRONLY) fd = _open_handle(path, os.O_WRONLY)
try: try:
st = os.fstat(fd)
if expected_identity is not None and (st.st_dev, st.st_ino) != expected_identity:
raise ValueError("handle identity changed")
os.lseek(fd, 0, os.SEEK_SET) os.lseek(fd, 0, os.SEEK_SET)
os.ftruncate(fd, 0) os.ftruncate(fd, 0)
_write_all(fd, (json.dumps(state, sort_keys=True, separators=(",", ":")) + "\n").encode()) _write_all(fd, (json.dumps(state, sort_keys=True, separators=(",", ":")) + "\n").encode())
@@ -202,6 +205,7 @@ class Broker:
self.handle_created = False self.handle_created = False
self.handle_identity: tuple[int, int] | None = None self.handle_identity: tuple[int, int] | None = None
self.started = False self.started = False
self.cleanup_result: dict | None = None
def _record_handle_identity(self, identity: tuple[int, int]) -> None: def _record_handle_identity(self, identity: tuple[int, int]) -> None:
self.handle_identity = identity self.handle_identity = identity
@@ -213,54 +217,74 @@ class Broker:
persisted = dict(self.state) persisted = dict(self.state)
if not retain_token: if not retain_token:
persisted.pop("token", None) persisted.pop("token", None)
_update_handle(self.handle, persisted) _update_handle(self.handle, persisted, expected_identity=self.handle_identity)
def cleanup(self, status: str) -> bool: def cleanup(self, status: str) -> bool:
released = False # Keep a previous partial result: a second cleanup must not erase failures.
try: previous = self.cleanup_result or {}
if self.set.repository.held: errors = list(previous.get("errors", []))
self.set.repository.release() for name, release in (("repository-release", self.set.repository.release),
self.set.release_issue() ("issue-release", self.set.release_issue)):
released = True
except Exception:
released = False
if released and self.handle_created:
try: try:
self.persist(status, retain_token=False) if name == "issue-release" or self.set.repository.held:
release()
except Exception: except Exception:
released = False errors.append(name)
released = not self.set.issue.held and not self.set.repository.held
if self.server: if self.server:
self.server.close() server, self.server = self.server, None
self.server = None
if self.socket_identity is not None:
try: try:
if self.socket_parent_fd is not None: server.close()
st = os.stat(self.sock_path.name, dir_fd=self.socket_parent_fd, follow_symlinks=False)
if (st.st_dev, st.st_ino) == self.socket_identity:
os.unlink(self.sock_path.name, dir_fd=self.socket_parent_fd)
except (FileNotFoundError, OSError):
pass
if self.handle_created and not self.started:
try:
parent_fd, name = _open_parent(self.handle)
try:
st = os.stat(name, dir_fd=parent_fd, follow_symlinks=False)
if self.handle_identity == (st.st_dev, st.st_ino):
try:
self.persist(status, retain_token=False)
except Exception:
os.unlink(name, dir_fd=parent_fd)
finally:
os.close(parent_fd)
except OSError: except OSError:
pass errors.append("server-close")
socket_removed = previous.get("socket_removed", self.socket_identity is None)
if self.socket_identity is not None and self.socket_parent_fd is not None:
try:
st = os.stat(self.sock_path.name, dir_fd=self.socket_parent_fd, follow_symlinks=False)
if (not stat.S_ISSOCK(st.st_mode) or st.st_uid != os.getuid()
or (st.st_dev, st.st_ino) != self.socket_identity):
errors.append("socket-identity-changed")
else:
os.unlink(self.sock_path.name, dir_fd=self.socket_parent_fd)
socket_removed = True
self.socket_identity = None
except FileNotFoundError:
socket_removed = True
self.socket_identity = None
except OSError:
errors.append("socket-unlink")
if self.socket_parent_fd is not None: if self.socket_parent_fd is not None:
fd, self.socket_parent_fd = self.socket_parent_fd, None
try: try:
os.close(self.socket_parent_fd) os.close(fd)
except OSError: except OSError:
pass errors.append("socket-parent-close")
self.socket_parent_fd = None self.cleanup_result = {"lease_released": released, "socket_removed": bool(socket_removed),
return released "handle_saved": False, "errors": sorted(set(errors))}
if self.handle_created:
try:
# Do not claim another object's handle during failed startup.
if not self.started and self.handle_identity is not None:
fd = _open_handle(self.handle, os.O_RDONLY)
try:
st = os.fstat(fd)
if (st.st_dev, st.st_ino) != self.handle_identity:
raise ValueError("handle identity changed")
finally:
os.close(fd)
self.cleanup_result["handle_saved"] = True
self.state["cleanup"] = self.cleanup_result
terminal_status = status if released and socket_removed and not errors else "cleanup-incomplete"
self.persist(terminal_status, retain_token=not released)
except Exception:
self.cleanup_result["handle_saved"] = False
self.cleanup_result["errors"] = sorted(set(self.cleanup_result["errors"] + ["handle-save"]))
else:
self.cleanup_result["handle_saved"] = True # No owned handle to save.
if self.cleanup_result["errors"]:
print(json.dumps({"event": "cleanup-incomplete", "cleanup": self.cleanup_result},
separators=(",", ":")), file=sys.stderr, flush=True)
return released and bool(socket_removed) and not self.cleanup_result["errors"]
def authenticate(self, req: dict) -> bool: def authenticate(self, req: dict) -> bool:
supplied = req.get("token") supplied = req.get("token")
@@ -300,9 +324,10 @@ class Broker:
self.persist("held") self.persist("held")
return {"status": "held"} return {"status": "held"}
if action == "release": if action == "release":
if not self.cleanup("released"): complete = self.cleanup("released")
return {"error": "release-failed"}
self.stopping = True self.stopping = True
if not complete:
return {"error": "cleanup-incomplete", "cleanup": self.cleanup_result}
return {"status": "released"} return {"status": "released"}
return {"error": "invalid action"} return {"error": "invalid action"}
@@ -318,7 +343,8 @@ class Broker:
except BaseException: except BaseException:
raise raise
st = self.handle.lstat() st = self.handle.lstat()
self.handle_identity = (st.st_dev, st.st_ino) if (st.st_dev, st.st_ino) != self.handle_identity:
raise ValueError("created handle identity changed")
self.handle_created = True self.handle_created = True
self.socket_parent_fd, socket_name = _open_parent(self.sock_path) self.socket_parent_fd, socket_name = _open_parent(self.sock_path)
try: try:
@@ -328,9 +354,17 @@ class Broker:
pass pass
self.server = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM) self.server = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
self.server.bind(f"/proc/self/fd/{self.socket_parent_fd}/{socket_name}") self.server.bind(f"/proc/self/fd/{self.socket_parent_fd}/{socket_name}")
sock_st = os.fstat(self.server.fileno()) # Socket FD metadata describes sockfs, not the bound filesystem node.
sock_st = os.stat(socket_name, dir_fd=self.socket_parent_fd, follow_symlinks=False)
if not stat.S_ISSOCK(sock_st.st_mode) or sock_st.st_uid != os.getuid():
raise ValueError("unsafe bound socket")
self.socket_identity = (sock_st.st_dev, sock_st.st_ino) self.socket_identity = (sock_st.st_dev, sock_st.st_ino)
os.fchmod(self.server.fileno(), 0o600) os.chmod(socket_name, 0o600, dir_fd=self.socket_parent_fd, follow_symlinks=False)
verified = os.stat(socket_name, dir_fd=self.socket_parent_fd, follow_symlinks=False)
if (not stat.S_ISSOCK(verified.st_mode) or verified.st_uid != os.getuid()
or (verified.st_dev, verified.st_ino) != self.socket_identity
or stat.S_IMODE(verified.st_mode) != 0o600):
raise ValueError("bound socket verification failed")
self.server.listen(8); self.server.settimeout(0.2) self.server.listen(8); self.server.settimeout(0.2)
self.started = True self.started = True
print(json.dumps({"status": "held", "handle": str(self.handle), "socket": str(self.sock_path)}, separators=(",", ":")), flush=True) print(json.dumps({"status": "held", "handle": str(self.handle), "socket": str(self.sock_path)}, separators=(",", ":")), flush=True)
@@ -352,8 +386,11 @@ class Broker:
response = {"error": "invalid request"} response = {"error": "invalid request"}
conn.sendall((json.dumps(response, separators=(",", ":")) + "\n").encode()) conn.sendall((json.dumps(response, separators=(",", ":")) + "\n").encode())
finally: finally:
if self.set.issue.held or self.handle_created or self.server: if (self.set.issue.held or self.set.repository.held
or (self.cleanup_result is None and (self.handle_created or self.server))):
self.cleanup("released" if not self.stopping else "expired") self.cleanup("released" if not self.stopping else "expired")
if self.cleanup_result and self.cleanup_result["errors"]:
raise OSError("broker cleanup incomplete")
def serve(args: argparse.Namespace) -> int: def serve(args: argparse.Namespace) -> int:
+5 -1
View File
@@ -64,7 +64,7 @@ Repository cache 경로는 고정된 machine-local 절대경로나 현재 작업
Bootstrap 이후 모든 Issue workflow는 role-neutral shared managed-lock broker 하나를 사용한다. 개별 impl/spec/review skill이 직접 `flock`을 획득하거나 broker 규칙을 재정의하지 않는다. Broker executable은 `harness/lib/managed_lock_broker.py`이며 authenticated request action은 `status`, `renew`, `acquire_repository`, `release_repository`, `release`이고 최대 TTL은 `3600`초다. Bootstrap 이후 모든 Issue workflow는 role-neutral shared managed-lock broker 하나를 사용한다. 개별 impl/spec/review skill이 직접 `flock`을 획득하거나 broker 규칙을 재정의하지 않는다. Broker executable은 `harness/lib/managed_lock_broker.py`이며 authenticated request action은 `status`, `renew`, `acquire_repository`, `release_repository`, `release`이고 최대 TTL은 `3600`초다.
- Bootstrap lock을 해제한 뒤 broker를 시작한다. checkout 밖 trusted parent 아래 role-owned mode `0700` runtime directory를 미리 만들고, run-unique handle과 socket path를 사용한다. Handle file은 mode `0600`이고 socket은 mode `0600`이며 둘 다 symlink가 아니어야 한다. - Bootstrap lock을 해제한 뒤 broker를 시작한다. checkout 밖 trusted parent 아래 role-owned mode `0700` runtime directory를 미리 만들고, run-unique handle과 socket path를 사용한다. Handle file은 mode `0600`이고 socket은 mode `0600`이며 둘 다 symlink가 아니어야 한다.
- `serve`는 tracked background process로 시작하고 ready JSON을 기다린 뒤에만 Issue 작업을 시작한다. Client는 raw token을 직접 read/log하지 않으며, broker public `request` API가 handle을 인증하고 token을 캡슐화한다. - `serve`는 tracked background process로 시작하고 ready JSON을 기다린 뒤에만 Issue 작업을 시작한다. Ready 전에 broker는 bound **filesystem socket node**의 type·owner UID·device/inode와 exact mode `0600`을 검증한다. Socket FD의 sockfs metadata는 이 검증을 대신하지 못한다. Client는 raw token을 직접 read/log하지 않으며, broker public `request` API가 handle을 인증하고 token을 캡슐화한다.
이 broker의 보안 경계에서 filesystem owner UID는 trusted principal이다. mode `0700`/`0600`과 random bearer token은 다른 UID 또는 foreign token을 사용한 작업을 막지만, 같은 UID를 공유하는 상호 비신뢰 프로세스를 서로 격리하지는 못한다. 그런 격리가 필요한 배포는 각 role을 distinct UID로 실행해야 한다. 이 broker의 보안 경계에서 filesystem owner UID는 trusted principal이다. mode `0700`/`0600`과 random bearer token은 다른 UID 또는 foreign token을 사용한 작업을 막지만, 같은 UID를 공유하는 상호 비신뢰 프로세스를 서로 격리하지는 못한다. 그런 격리가 필요한 배포는 각 role을 distinct UID로 실행해야 한다.
@@ -72,6 +72,10 @@ Bootstrap 이후 모든 Issue workflow는 role-neutral shared managed-lock broke
- TTL 만료 전에 `renew`를 request한다. Broker crash 뒤 kernel이 advisory lock을 해제해도 stale handle은 진단 정보일 뿐이며 `probe`가 authoritative하다. - TTL 만료 전에 `renew`를 request한다. Broker crash 뒤 kernel이 advisory lock을 해제해도 stale handle은 진단 정보일 뿐이며 `probe`가 authoritative하다.
- 정상, 오류와 cancel 경로는 동일한 cleanup을 수행한다: 보유 중이면 `release_repository`, `release`, broker process wait/reap, 그리고 Issue `probe`를 실행해 `free=true`를 요구한다. 다른 Issue workflow가 합법적으로 repository lock을 먼저 획득할 수 있으므로 완료 gate는 Issue resource를 검사하고, 종료한 broker process가 자신의 repository FD를 더 이상 보유하지 않음을 wait/reap으로 증명한다. 이 순서를 검증하기 전에는 completed/log/report를 주장하지 않는다. Lock 파일은 절대 unlink하지 않는다. - 정상, 오류와 cancel 경로는 동일한 cleanup을 수행한다: 보유 중이면 `release_repository`, `release`, broker process wait/reap, 그리고 Issue `probe`를 실행해 `free=true`를 요구한다. 다른 Issue workflow가 합법적으로 repository lock을 먼저 획득할 수 있으므로 완료 gate는 Issue resource를 검사하고, 종료한 broker process가 자신의 repository FD를 더 이상 보유하지 않음을 wait/reap으로 증명한다. 이 순서를 검증하기 전에는 completed/log/report를 주장하지 않는다. Lock 파일은 절대 unlink하지 않는다.
- 소켓 정리는 부모 directory FD를 기준으로 기록한 filesystem device/inode·type·owner가 일치하는 경로만 삭제한다. 경로가 다른 객체나 symlink로 교체되면 그 객체는 삭제하지 않고 `socket-identity-changed`로 보고한다. 같은 UID의 경쟁 프로세스까지 원자적 격리를 보장하지 않으며, SIGKILL/crash 뒤 소켓 삭제를 보장하지 않는다. Stale 경로를 자동 삭제하거나 재사용하지 않는다. 첫 소켓 정리 실패 뒤에는 부모 FD를 닫고 경로를 다시 열어 자동 재시도하지 않는다. 잔여 경로의 제거는 운영자가 exact 경로와 소유권을 확인하는 별도 복구이며, 반복 cleanup은 이전 부분 실패를 보존한다.
- 정상 release 응답은 기존 `{"status":"released"}`를 유지한다. 정리가 일부 실패하면 broker는 `{"error":"cleanup-incomplete","cleanup":...}`를 응답하고 종료하며, `cleanup`의 `lease_released`, `socket_removed`, `handle_saved`, `errors`로 해제와 잔여 정리를 구분한다. Handle 기록 가능 시 같은 결과를 비밀값 없이 저장하고 stderr에도 오류 코드만 보고한다. 부분 실패는 nonzero process exit로 보고하며, 성공한 cleanup을 재적용하거나 실패 이력을 성공으로 덮어쓰지 않는다. 이 오류만으로 이미 해제한 lease를 재해제하거나 workflow 외부 write를 반복하지 않는다.
- 완료 gate는 기존 release → wait/reap → Issue probe 순서에 더해 broker의 정상 종료와 run-owned 소켓 경로 부재를 확인한다. `probe free=true`는 lock 해제 증거이지 소켓 권한·삭제 증거가 아니다. Lease 해제와 소켓·handle 정리 실패는 별도로 보고하고 완료로 간주하지 않는다.
### Canonical command templates ### Canonical command templates
```sh ```sh
@@ -25,11 +25,12 @@ def wait_ready(proc, timeout=3):
class ManagedLockBrokerTests(unittest.TestCase): class ManagedLockBrokerTests(unittest.TestCase):
def start(self, root, owner="alice", repo="ledger", issue=7, ttl=5, role="gitea-impl-agent"): def start(self, root, owner="alice", repo="ledger", issue=7, ttl=5, role="gitea-impl-agent", mask=-1):
locks = root / "repositories" / owner / repo / "locks" locks = root / "repositories" / owner / repo / "locks"
locks.mkdir(parents=True) locks.mkdir(parents=True)
os.chmod(locks, 0o700) os.chmod(locks, 0o700)
(locks / "repository.lock").touch(mode=0o600) (locks / "repository.lock").touch(mode=0o600)
(locks / f"issue-{issue}.lock").touch(mode=0o600)
runtime = root / "runtime" runtime = root / "runtime"
runtime.mkdir() runtime.mkdir()
os.chmod(runtime, 0o700) os.chmod(runtime, 0o700)
@@ -39,7 +40,7 @@ class ManagedLockBrokerTests(unittest.TestCase):
"--repository", repo, "--issue", str(issue), "--role", role, "--run-id", "run-1", "--repository", repo, "--issue", str(issue), "--role", role, "--run-id", "run-1",
"--handle", str(handle), "--socket", str(socket), "--ttl-seconds", str(ttl), "--handle", str(handle), "--socket", str(socket), "--ttl-seconds", str(ttl),
"--acquire-timeout", "0.3"] "--acquire-timeout", "0.3"]
proc = subprocess.Popen(cmd, cwd=Path.cwd(), stdout=subprocess.PIPE, stderr=subprocess.PIPE) proc = subprocess.Popen(cmd, cwd=Path.cwd(), stdout=subprocess.PIPE, stderr=subprocess.PIPE, umask=mask)
ready = wait_ready(proc) ready = wait_ready(proc)
self.assertEqual(ready["status"], "held") self.assertEqual(ready["status"], "held")
self.assertNotIn("token", json.dumps(ready)) self.assertNotIn("token", json.dumps(ready))
@@ -126,6 +127,207 @@ class ManagedLockBrokerTests(unittest.TestCase):
os.umask(old) os.umask(old)
self.assertEqual(stat.S_IMODE(path.stat().st_mode), 0o600) self.assertEqual(stat.S_IMODE(path.stat().st_mode), 0o600)
def test_socket_mode_and_cleanup_under_umasks(self):
for mask in (0o022, 0o077, 0o777):
with self.subTest(mask=mask), tempfile.TemporaryDirectory() as d:
proc, handle, sock = self.start(Path(d), mask=mask)
try:
self.assertEqual(stat.S_IMODE(sock.lstat().st_mode), 0o600)
self.assertTrue(stat.S_ISSOCK(sock.lstat().st_mode))
self.assertEqual(broker.request_handle(handle, "status")["status"], "held")
self.assertEqual(broker.request_handle(handle, "release")["status"], "released")
self.assertEqual(proc.wait(timeout=3), 0)
self.assertFalse(sock.exists())
self.assertTrue(probe_free(Path(d) / "repositories/alice/ledger/locks/issue-7.lock"))
finally:
if proc.poll() is None:
proc.terminate(); proc.wait(timeout=3)
proc.stdout.close(); proc.stderr.close()
def make_bound_instance(self, root):
import socket
locks = root / "repositories/alice/ledger/locks"
locks.mkdir(parents=True); locks.chmod(0o700)
(locks / "repository.lock").touch(mode=0o600)
runtime = root / "runtime"; runtime.mkdir(mode=0o700)
instance = broker.Broker(Namespace(root=str(root), owner="alice", repository="ledger", issue=7,
role="gitea-impl-agent", run_id="cleanup-test", handle=str(runtime / "handle.json"),
socket=str(runtime / "broker.sock"), ttl_seconds=5.0, acquire_timeout=0.0))
instance.set.issue.acquire()
broker._write_handle(instance.handle, instance.state, exclusive=True,
on_created=instance._record_handle_identity)
instance.handle_created = True; instance.started = True
instance.socket_parent_fd = broker._open_directory(runtime)
instance.server = socket.socket(socket.AF_UNIX)
instance.server.bind(str(instance.sock_path))
st = instance.sock_path.lstat()
instance.socket_identity = (st.st_dev, st.st_ino)
return instance
def test_release_reports_socket_failure_separately_and_stops(self):
with tempfile.TemporaryDirectory() as d:
instance = self.make_bound_instance(Path(d))
with patch.object(broker.os, "unlink", side_effect=PermissionError("injected")):
result = instance.handle_request({"action": "release", "token": instance.token})
self.assertEqual(result["error"], "cleanup-incomplete")
self.assertTrue(result["cleanup"]["lease_released"])
self.assertFalse(result["cleanup"]["socket_removed"])
self.assertTrue(instance.stopping)
self.assertTrue(probe_free(instance.set.issue.path))
self.assertNotIn("token", json.loads(instance.handle.read_text()))
self.assertIsNone(instance.socket_parent_fd)
self.assertIsNone(instance.server)
def test_cleanup_preserves_replaced_path_and_reports_incomplete(self):
for kind in ("file", "symlink"):
with self.subTest(kind=kind), tempfile.TemporaryDirectory() as d:
instance = self.make_bound_instance(Path(d))
instance.sock_path.unlink()
target = Path(d) / "target"; target.write_text("foreign")
if kind == "file": instance.sock_path.write_text("replacement")
else: instance.sock_path.symlink_to(target)
self.assertFalse(instance.cleanup("released"))
self.assertTrue(instance.sock_path.is_symlink() if kind == "symlink" else instance.sock_path.is_file())
self.assertEqual(target.read_text(), "foreign")
self.assertTrue(instance.cleanup_result["lease_released"])
self.assertIn("socket-identity-changed", instance.cleanup_result["errors"])
self.assertFalse(instance.cleanup("released")) # Do not erase a partial result.
def test_socket_removed_after_expiry_and_sigterm(self):
for action in ("expiry", "sigterm"):
with self.subTest(action=action), tempfile.TemporaryDirectory() as d:
proc, handle, sock = self.start(Path(d), ttl=1 if action == "expiry" else 5)
try:
if action == "sigterm": proc.terminate()
self.assertEqual(proc.wait(timeout=4), 0)
self.assertFalse(sock.exists())
self.assertTrue(probe_free(Path(d) / "repositories/alice/ledger/locks/issue-7.lock"))
self.assertNotIn("token", json.loads(handle.read_text()))
finally:
if proc.poll() is None: proc.kill(); proc.wait(timeout=3)
proc.stdout.close(); proc.stderr.close()
def test_partial_cleanup_response_delivered_before_nonzero_exit(self):
import socket
with tempfile.TemporaryDirectory() as d:
root = Path(d); proc, handle, sock_path = self.start(root)
target = root / "replacement-target"; target.write_text("untouched")
try:
data = json.loads(handle.read_text())
with socket.socket(socket.AF_UNIX) as client:
client.settimeout(3); client.connect(str(sock_path))
sock_path.unlink(); sock_path.symlink_to(target)
client.sendall((json.dumps({"action": "release", "token": data["token"]}) + "\n").encode())
with client.makefile("rb") as stream:
result = json.loads(stream.readline())
self.assertEqual(result["error"], "cleanup-incomplete")
self.assertTrue(result["cleanup"]["lease_released"])
self.assertEqual(proc.wait(timeout=3), 2)
self.assertTrue(sock_path.is_symlink())
self.assertEqual(target.read_text(), "untouched")
self.assertTrue(probe_free(root / "repositories/alice/ledger/locks/issue-7.lock"))
terminal = json.loads(handle.read_text())
self.assertNotIn("token", terminal)
self.assertFalse(terminal["cleanup"]["socket_removed"])
self.assertNotIn(data["token"], proc.stderr.read().decode())
finally:
if proc.poll() is None: proc.kill(); proc.wait(timeout=3)
proc.stdout.close(); proc.stderr.close()
def test_socket_permission_failure_has_no_ready_and_cleans_owned_socket(self):
import contextlib
import io
with tempfile.TemporaryDirectory() as d:
instance = self.make_bound_instance(Path(d))
instance.cleanup("released")
args = Namespace(root=str(instance.root), owner="alice", repository="ledger", issue=7,
role="gitea-impl-agent", run_id="startup-permission", handle=str(instance.handle.parent / "new.handle"),
socket=str(instance.sock_path.parent / "new.sock"), ttl_seconds=5.0, acquire_timeout=0.0)
fresh = broker.Broker(args)
out = io.StringIO()
with patch.object(broker.os, "chmod", side_effect=PermissionError("injected")), contextlib.redirect_stdout(out):
with self.assertRaises(PermissionError): fresh.run()
self.assertEqual(out.getvalue(), "")
self.assertFalse(fresh.sock_path.exists())
self.assertTrue(probe_free(fresh.set.issue.path))
self.assertNotIn("token", json.loads(fresh.handle.read_text()))
def test_cleanup_respects_lock_order_after_repository_error(self):
with tempfile.TemporaryDirectory() as d:
instance = self.make_bound_instance(Path(d))
instance.set.repository.acquire()
with patch.object(instance.set.repository, "release", side_effect=OSError("injected")):
self.assertFalse(instance.cleanup("released"))
self.assertFalse(probe_free(instance.set.issue.path)) # Repository still held; preserve lock order.
self.assertFalse(instance.cleanup_result["lease_released"])
self.assertTrue(instance.cleanup_result["socket_removed"])
self.assertIn("repository-release", instance.cleanup_result["errors"])
instance.cleanup("released") # Retry held FD, but keep previous failure visible.
self.assertTrue(probe_free(instance.set.repository.path))
self.assertNotIn("token", json.loads(instance.handle.read_text()))
def test_startup_handle_replacement_is_preserved(self):
import contextlib
import io
with tempfile.TemporaryDirectory() as d:
instance = self.make_bound_instance(Path(d)); instance.cleanup("released")
fresh = broker.Broker(Namespace(root=str(instance.root), owner="alice", repository="ledger", issue=7,
role="gitea-impl-agent", run_id="handle-swap", handle=str(instance.handle.parent / "swap.handle"),
socket=str(instance.sock_path.parent / "swap.sock"), ttl_seconds=5.0, acquire_timeout=0.0))
original_write = broker._write_handle
foreign = b'{"foreign":"preserve"}\n'
def replace_after_create(path, state, **kwargs):
original_write(path, state, **kwargs)
path.rename(path.with_suffix(".original")) # Keep original inode alive.
path.write_bytes(foreign); path.chmod(0o600)
with patch.object(broker, "_write_handle", side_effect=replace_after_create), contextlib.redirect_stdout(io.StringIO()):
with self.assertRaises((OSError, ValueError)): fresh.run()
self.assertEqual(fresh.handle.read_bytes(), foreign)
self.assertTrue(probe_free(fresh.set.issue.path))
self.assertIn("handle-save", fresh.cleanup_result["errors"])
def test_partial_cleanup_durable_status_is_incomplete(self):
with tempfile.TemporaryDirectory() as d:
instance = self.make_bound_instance(Path(d))
with patch.object(broker.os, "unlink", side_effect=PermissionError("injected")):
instance.cleanup("released")
state = json.loads(instance.handle.read_text())
self.assertEqual(state["status"], "cleanup-incomplete")
self.assertTrue(state["cleanup"]["lease_released"])
self.assertFalse(state["cleanup"]["socket_removed"])
self.assertNotIn("token", state)
def test_socket_replacement_during_permission_setup_is_preserved(self):
import contextlib
import io
with tempfile.TemporaryDirectory() as d:
instance = self.make_bound_instance(Path(d)); instance.cleanup("released")
fresh = broker.Broker(Namespace(root=str(instance.root), owner="alice", repository="ledger", issue=7,
role="gitea-impl-agent", run_id="socket-swap", handle=str(instance.handle.parent / "swap.handle"),
socket=str(instance.sock_path.parent / "swap.sock"), ttl_seconds=5.0, acquire_timeout=0.0))
real_chmod = os.chmod
def replace_after_chmod(path, mode, **kwargs):
real_chmod(path, mode, **kwargs)
fresh.sock_path.unlink()
fresh.sock_path.write_text("foreign socket replacement")
out = io.StringIO()
with patch.object(broker.os, "chmod", side_effect=replace_after_chmod), contextlib.redirect_stdout(out):
with self.assertRaises(ValueError): fresh.run()
self.assertEqual(out.getvalue(), "")
self.assertEqual(fresh.sock_path.read_text(), "foreign socket replacement")
self.assertTrue(probe_free(fresh.set.issue.path))
self.assertEqual(json.loads(fresh.handle.read_text())["status"], "cleanup-incomplete")
def test_persist_rejects_replaced_handle_after_ready(self):
with tempfile.TemporaryDirectory() as d:
instance = self.make_bound_instance(Path(d))
instance.handle.rename(instance.handle.with_suffix(".original"))
instance.handle.write_text("foreign"); instance.handle.chmod(0o600)
with self.assertRaises(ValueError): instance.persist("held")
self.assertFalse(instance.cleanup("released"))
self.assertEqual(instance.handle.read_text(), "foreign")
self.assertTrue(probe_free(instance.set.issue.path))
def test_repository_scope_release_keeps_issue(self): def test_repository_scope_release_keeps_issue(self):
with tempfile.TemporaryDirectory() as d: with tempfile.TemporaryDirectory() as d:
root = Path(d); proc, handle, _ = self.start(root) root = Path(d); proc, handle, _ = self.start(root)