fix(cli): probe autoroute proxy pids without signalling them

litellm/proxy/client/cli/commands/autoroute/process.py

@@ -138,7 +138,68 @@ def clear_pid_record(path: Path | None = None) -> None:
     resolved_path.unlink(missing_ok=True)
 
 
+_PROCESS_QUERY_LIMITED_INFORMATION: Final = 0x1000
+_STILL_ACTIVE: Final = 259
+_ERROR_ACCESS_DENIED: Final = 5
+
+
+def _windows_pid_exists(pid: int) -> bool:
+    """Liveness check for Windows that does not signal the process.
+
+    Opens a query-only handle and asks for the exit code.  ``OpenProcess``
+    failing with ``ERROR_ACCESS_DENIED`` means the pid exists but is not ours
+    to open, which is the case the ``PermissionError`` arm covers on POSIX.
+
+    Caveat kept deliberately: a process whose real exit code is 259 reads as
+    alive, because ``GetExitCodeProcess`` reports ``STILL_ACTIVE`` (259) for a
+    running process and cannot distinguish the two.  That is the standard
+    trade-off for this API and it is strictly better than the previous
+    behaviour, which killed the process it was asked about.
+    """
+    import ctypes
+    from ctypes import wintypes
+
+    kernel32 = ctypes.WinDLL("kernel32", use_last_error=True)
+    kernel32.OpenProcess.argtypes = (wintypes.DWORD, wintypes.BOOL, wintypes.DWORD)
+    kernel32.OpenProcess.restype = wintypes.HANDLE
+    kernel32.GetExitCodeProcess.argtypes = (wintypes.HANDLE, ctypes.POINTER(wintypes.DWORD))
+    kernel32.GetExitCodeProcess.restype = wintypes.BOOL
+    kernel32.CloseHandle.argtypes = (wintypes.HANDLE,)
+    kernel32.CloseHandle.restype = wintypes.BOOL
+
+    handle: Final = kernel32.OpenProcess(_PROCESS_QUERY_LIMITED_INFORMATION, False, pid)
+    if not handle:
+        return ctypes.get_last_error() == _ERROR_ACCESS_DENIED
+    try:
+        exit_code = wintypes.DWORD()
+        if not kernel32.GetExitCodeProcess(handle, ctypes.byref(exit_code)):
+            return True
+        return exit_code.value == _STILL_ACTIVE
+    finally:
+        kernel32.CloseHandle(handle)
+
+
 def is_running(pid: int) -> bool:
+    """Report whether ``pid`` names a live process, without signalling it.
+
+    ``os.kill(pid, 0)`` is the POSIX idiom and is a genuine no-op there, but it
+    is not a probe on Windows.  ``signal.CTRL_C_EVENT`` is 0, so ``os.kill(pid,
+    0)`` *is* ``os.kill(pid, CTRL_C_EVENT)`` and reaches
+    ``GenerateConsoleCtrlEvent``.  That API's second argument is a process
+    *group*, and ``launch_proxy`` above starts the proxy through
+    ``subprocess.Popen`` with no ``CREATE_NEW_PROCESS_GROUP``, so the child
+    shares this console's group: the Ctrl-C is delivered to the proxy, to the
+    ``lite`` process asking the question, and to anything else on the console.
+
+    Measured on Windows 11, Python 3.12.10.  Probing a sleeping child returns
+    ``True`` with no exception raised; the child then exits with 3221225786
+    (``0xC000013A``, ``STATUS_CONTROL_C_EXIT``) and a ``KeyboardInterrupt``
+    arrives in the caller a moment later -- so the traceback does not point at
+    the probe.  ``KeyboardInterrupt`` is a ``BaseException``, so neither the
+    ``ProcessLookupError`` nor the ``PermissionError`` arm below ever sees it.
+    """
+    if sys.platform == "win32":
+        return _windows_pid_exists(pid)
     try:
         os.kill(pid, 0)
     except ProcessLookupError:
@@ -149,17 +210,31 @@ def is_running(pid: int) -> bool:
 
 
 def terminate(pid: int, grace_period: float = 5.0) -> None:
-    """Terminate a process by pid, escalating from SIGTERM to SIGKILL if needed."""
+    """Terminate a process by pid, escalating from SIGTERM to SIGKILL if needed.
+
+    Both kills tolerate the same errors ``prisma_client`` already tolerates for
+    this exact case -- "already dead or inaccessible".  ``ProcessLookupError``
+    alone is not enough: on Windows ``os.kill`` routes to ``TerminateProcess``,
+    and a process that has already exited answers ``ERROR_ACCESS_DENIED``, so
+    the call raises ``PermissionError`` and escapes the suppression.
+    """
     if not is_running(pid):
         return
-    with contextlib.suppress(ProcessLookupError):
+    with contextlib.suppress(ProcessLookupError, PermissionError, OSError):
         os.kill(pid, signal.SIGTERM)
     deadline: Final = time.monotonic() + grace_period
     while time.monotonic() < deadline and is_running(pid):
         time.sleep(0.2)
     if is_running(pid):
-        with contextlib.suppress(ProcessLookupError):
-            os.kill(pid, signal.SIGKILL)
+        # signal.SIGKILL does not exist on Windows.  Referencing it raises
+        # AttributeError, which contextlib.suppress(ProcessLookupError) does
+        # not catch, so this escalation path crashed instead of hard-killing
+        # a proxy that ignored SIGTERM.  os.kill with any signal other than 0
+        # or 1 routes to TerminateProcess on Windows, so SIGTERM is a real
+        # kill there.  litellm/proxy/db/prisma_client.py already resolves the
+        # signal this way.
+        with contextlib.suppress(ProcessLookupError, PermissionError, OSError):
+            os.kill(pid, getattr(signal, "SIGKILL", signal.SIGTERM))
 
 
 def stream_log(log_path: Path, stop_event: threading.Event) -> None:

tests/test_litellm/proxy/client/cli/autoroute/test_process.py

@@ -1,5 +1,9 @@
 import os
+import signal
 import socket
+import subprocess
+import sys
+import time
 from typing import Optional
 from unittest.mock import patch
 
@@ -17,6 +21,7 @@ from litellm.proxy.client.cli.commands.autoroute.process import (
     missing_proxy_runtime_modules,
     poll_liveliness,
     read_pid_record,
+    terminate,
     write_pid_record,
 )
 
@@ -111,6 +116,10 @@ class TestIsRunning:
     def test_huge_unlikely_pid_is_not_running(self):
         assert is_running(2**30) is False
 
+    @pytest.mark.skipif(
+        sys.platform == "win32",
+        reason="POSIX branch only; win32 answers with a query-only process handle, not os.kill",
+    )
     def test_permission_error_from_kill_is_treated_as_running(self, monkeypatch):
         def fake_kill(pid: int, sig: int) -> None:
             raise PermissionError("not permitted to signal this pid")
@@ -120,6 +129,123 @@ class TestIsRunning:
         assert is_running(999) is True
 
 
+class TestIsRunningDoesNotSignal:
+    """``os.kill(pid, 0)`` is not a liveness probe on Windows.
+
+    ``signal.CTRL_C_EVENT`` is 0, so the call reaches
+    ``GenerateConsoleCtrlEvent``, whose second argument is a process *group*.
+    ``launch_proxy`` spawns the proxy without ``CREATE_NEW_PROCESS_GROUP``, so
+    the child shares this console's group and the Ctrl-C hits the child, the
+    caller, and anything else attached to the console.
+
+    Measured on Windows 11, Python 3.12.10: probing a sleeping child returns
+    ``True`` with no exception, the child then exits 3221225786
+    (``0xC000013A`` / ``STATUS_CONTROL_C_EXIT``), and a ``KeyboardInterrupt``
+    lands in the caller afterwards -- a ``BaseException``, so the handlers in
+    ``is_running`` never see it.
+    """
+
+    def test_probing_a_child_does_not_terminate_it(self):
+        """The behavioural check, with nothing mocked."""
+        child = subprocess.Popen([sys.executable, "-c", "import time; time.sleep(30)"])
+        try:
+            assert is_running(child.pid) is True
+            time.sleep(1.0)
+            assert child.poll() is None, (
+                f"the liveness probe terminated the process it was asked about "
+                f"(exit code {child.returncode})"
+            )
+            assert is_running(child.pid) is True
+        finally:
+            child.kill()
+            child.wait(timeout=10)
+
+    def test_probing_our_own_pid_leaves_this_process_alone(self):
+        """What ``TestIsRunning.test_current_process_is_running`` above relies on.
+
+        The assertion there passes either way; it is the interpreter surviving
+        the next statement that separates a probe from a signal.
+        """
+        assert is_running(os.getpid()) is True
+        time.sleep(0.5)
+        assert is_running(os.getpid()) is True
+
+    def test_win32_answers_without_calling_os_kill(self, monkeypatch):
+        calls: list = []
+        monkeypatch.setattr(process_module.sys, "platform", "win32")
+        monkeypatch.setattr(process_module.os, "kill", lambda pid, sig: calls.append(sig))
+        monkeypatch.setattr(process_module, "_windows_pid_exists", lambda pid: True)
+
+        assert is_running(1234) is True
+        assert calls == [], "os.kill reached on win32; signal 0 is CTRL_C_EVENT there"
+
+
+class TestTerminateHardKillSignal:
+    def test_escalation_resolves_a_signal_when_sigkill_is_absent(self, monkeypatch):
+        """``signal.SIGKILL`` does not exist on Windows.
+
+        Referencing it raises ``AttributeError``, and ``terminate`` wraps the
+        call in ``contextlib.suppress(ProcessLookupError)``, which does not
+        catch that -- so the branch that exists for a proxy ignoring SIGTERM
+        crashed instead of hard-killing it.  ``os.kill`` with any signal other
+        than 0 or 1 routes to ``TerminateProcess`` on Windows, so SIGTERM is a
+        real kill there.
+        """
+        sent: list = []
+        monkeypatch.setattr(process_module.os, "kill", lambda pid, sig: sent.append(sig))
+        monkeypatch.setattr(process_module, "is_running", lambda pid: True)
+        monkeypatch.setattr(process_module.time, "sleep", lambda seconds: None)
+        monkeypatch.delattr(process_module.signal, "SIGKILL", raising=False)
+
+        terminate(4242, grace_period=0.0)
+
+        assert sent == [signal.SIGTERM, signal.SIGTERM]
+
+    @pytest.mark.skipif(
+        not hasattr(signal, "SIGKILL"),
+        reason="POSIX escalation path; SIGKILL is absent on this platform",
+    )
+    def test_escalation_still_uses_sigkill_where_it_exists(self, monkeypatch):
+        sent: list = []
+        monkeypatch.setattr(process_module.os, "kill", lambda pid, sig: sent.append(sig))
+        monkeypatch.setattr(process_module, "is_running", lambda pid: True)
+        monkeypatch.setattr(process_module.time, "sleep", lambda seconds: None)
+
+        terminate(4242, grace_period=0.0)
+
+        assert sent == [signal.SIGTERM, signal.SIGKILL]
+
+    @pytest.mark.parametrize(
+        "raised",
+        [
+            PermissionError(5, "Access is denied"),
+            OSError(22, "Invalid argument"),
+            ProcessLookupError(3, "No such process"),
+        ],
+        ids=["permission-denied", "oserror", "already-gone"],
+    )
+    def test_a_kill_that_cannot_land_does_not_escape(self, monkeypatch, raised):
+        """A process that exited between the probe and the kill must not crash ``down``.
+
+        On Windows ``os.kill`` routes to ``TerminateProcess``, and a process
+        that has already exited answers ``ERROR_ACCESS_DENIED`` -- so the call
+        raises ``PermissionError``, not ``ProcessLookupError``.  Measured with
+        the real CLI on Windows 11: ``lite autoroute down`` ended in
+        ``PermissionError: [WinError 5]`` out of ``terminate``.
+        ``litellm/proxy/db/prisma_client.py`` already tolerates all three for
+        this same "already dead or inaccessible" case.
+        """
+
+        def fake_kill(pid: int, sig: int) -> None:
+            raise raised
+
+        monkeypatch.setattr(process_module.os, "kill", fake_kill)
+        monkeypatch.setattr(process_module, "is_running", lambda pid: True)
+        monkeypatch.setattr(process_module.time, "sleep", lambda seconds: None)
+
+        terminate(4242, grace_period=0.0)
+
+
 class TestPollLiveliness:
     def test_succeeds_when_health_check_returns_200_quickly(self, monkeypatch, tmp_path):
         monkeypatch.setattr(process_module.requests, "get", lambda url, timeout: FakeResponse(200))