fixes

system/manager/manager.py

@@ -5,6 +5,7 @@ import os
 import shutil
 from pathlib import Path
 import signal
+import stat
 import sys
 import time
 import traceback
@@ -759,6 +760,48 @@ def migrate_legacy_experimental_longitudinal(params: Params, params_cache: Param
   params_cache.remove("ExperimentalLongitudinalEnabled")
 
 
+def _msgq_file_is_readwrite_openable(path: Path) -> bool:
+  fd = os.open(path, os.O_RDWR | getattr(os, "O_CLOEXEC", 0))
+  try:
+    return True
+  finally:
+    os.close(fd)
+
+
+def cleanup_inaccessible_msgq_files(shm_path: str | Path) -> int:
+  """Remove stale msgq files that would crash SubSocket/PubSocket creation."""
+  shm_root = Path(shm_path)
+  if not shm_root.is_dir():
+    return 0
+
+  removed = 0
+  for path in shm_root.rglob("msgq_*"):
+    try:
+      st = path.lstat()
+    except OSError:
+      continue
+
+    if not stat.S_ISREG(st.st_mode):
+      continue
+
+    try:
+      _msgq_file_is_readwrite_openable(path)
+      continue
+    except PermissionError:
+      pass
+    except OSError:
+      continue
+
+    try:
+      path.unlink()
+      removed += 1
+      cloudlog.warning(f"Removed inaccessible stale msgq file: {path}")
+    except OSError:
+      cloudlog.exception(f"Failed to remove inaccessible stale msgq file: {path}")
+
+  return removed
+
+
 def manager_init() -> None:
   manager_init_start = time.monotonic()
   last_timing = _log_boot_timing("manager_init", "start", manager_init_start, manager_init_start)
@@ -829,6 +872,11 @@ def manager_init() -> None:
     print(f"WARNING: failed to make {Paths.shm_path()}")
   last_timing = _log_boot_timing("manager_init", "shm_path", manager_init_start, last_timing)
 
+  removed_msgq_files = cleanup_inaccessible_msgq_files(Paths.shm_path())
+  if removed_msgq_files:
+    cloudlog.warning(f"Removed {removed_msgq_files} inaccessible stale msgq files before process startup")
+  last_timing = _log_boot_timing("manager_init", "msgq_cleanup", manager_init_start, last_timing)
+
   # set params
   serial = HARDWARE.get_serial()
   params.put("Version", build_metadata.openpilot.version)

system/manager/test/test_manager.py

@@ -376,6 +376,40 @@ class TestManager:
 
     assert params.get_bool("PrioritizeSmoothFollowing")
 
+  def test_cleanup_inaccessible_msgq_files_removes_only_blocked_files(self, tmp_path, monkeypatch):
+    healthy = tmp_path / "msgq_deviceState"
+    blocked = tmp_path / "msgq_gpsLocation"
+    unrelated = tmp_path / "not_msgq_gpsLocation"
+    healthy.write_bytes(b"healthy")
+    blocked.write_bytes(b"blocked")
+    unrelated.write_bytes(b"unrelated")
+
+    def fake_open_probe(path):
+      if path == blocked:
+        raise PermissionError("blocked")
+      return True
+
+    monkeypatch.setattr(manager, "_msgq_file_is_readwrite_openable", fake_open_probe)
+
+    assert manager.cleanup_inaccessible_msgq_files(tmp_path) == 1
+    assert healthy.read_bytes() == b"healthy"
+    assert not blocked.exists()
+    assert unrelated.read_bytes() == b"unrelated"
+
+  def test_cleanup_inaccessible_msgq_files_ignores_msgq_directories(self, tmp_path, monkeypatch):
+    msgq_dir = tmp_path / "msgq_desktop"
+    msgq_dir.mkdir()
+    child = msgq_dir / "gpsLocation"
+    child.write_bytes(b"child")
+
+    def fake_open_probe(path):
+      raise AssertionError(f"directories and non-msgq children should not be probed: {path}")
+
+    monkeypatch.setattr(manager, "_msgq_file_is_readwrite_openable", fake_open_probe)
+
+    assert manager.cleanup_inaccessible_msgq_files(tmp_path) == 0
+    assert child.read_bytes() == b"child"
+
   @pytest.mark.skip("this test is flaky the way it's currently written, should be moved to test_onroad")
   def test_clean_exit(self, subtests):
     """

system/tests/test_tombstoned.py

@@ -0,0 +1,24 @@
+from openpilot.system import tombstoned
+
+
+def test_report_tombstone_apport_ignores_hal3_direct_widthfix(tmp_path, monkeypatch):
+  crash = tmp_path / "_data_agnos-compat_bin_hal3_direct_widthfix.0.crash"
+  crash.write_text("\n".join([
+    "ProblemType: Crash",
+    "ExecutablePath: /data/agnos-compat/bin/hal3_direct_widthfix",
+    "Signal: 11",
+    "CoreDump: base64",
+  ]))
+
+  sentry_calls = []
+  monkeypatch.setattr(tombstoned.sentry, "report_tombstone", lambda *args: sentry_calls.append(args))
+
+  def fail_retrace(fn):
+    raise AssertionError(f"ignored tombstones should not be retraced: {fn}")
+
+  monkeypatch.setattr(tombstoned, "get_apport_stacktrace", fail_retrace)
+
+  tombstoned.report_tombstone_apport(str(crash))
+
+  assert sentry_calls == []
+  assert not crash.exists()

system/tombstoned.py

@@ -19,6 +19,9 @@ MAX_TOMBSTONE_FN_LEN = 62  # 85 - 23 ("<dongle id>/crash/")
 
 TOMBSTONE_DIR = "/data/tombstones/"
 APPORT_DIR = "/var/crash/"
+IGNORED_APPORT_EXECUTABLES = {
+  "/data/agnos-compat/bin/hal3_direct_widthfix",
+}
 
 
 def safe_fn(s):
@@ -44,6 +47,27 @@ def get_apport_stacktrace(fn):
     return "Timeout getting stacktrace"
 
 
+def get_apport_executable_path(fn: str) -> str:
+  try:
+    with open(fn) as f:
+      for line in f:
+        if line.startswith("ExecutablePath:"):
+          return line.strip().split(': ', 1)[-1]
+        if line.startswith("CoreDump"):
+          break
+  except OSError:
+    pass
+
+  return ""
+
+
+def remove_crashlog(fn: str) -> None:
+  try:
+    os.remove(fn)
+  except PermissionError:
+    pass
+
+
 def get_tombstones():
   """Returns list of (filename, ctime) for all crashlogs"""
   files = []
@@ -64,6 +88,12 @@ def report_tombstone_apport(fn):
     cloudlog.error(f"Tombstone {fn} too big, {f_size}. Skipping...")
     return
 
+  executable_path = get_apport_executable_path(fn)
+  if executable_path in IGNORED_APPORT_EXECUTABLES:
+    cloudlog.warning(f"Ignoring known agnos-compat tombstone: {executable_path}")
+    remove_crashlog(fn)
+    return
+
   message = ""  # One line description of the crash
   contents = ""  # Full file contents without coredump
   path = ""  # File path relative to openpilot directory
@@ -134,10 +164,7 @@ def report_tombstone_apport(fn):
   # Files could be on different filesystems, copy, then delete
   shutil.copy(fn, os.path.join(crashlog_dir, new_fn))
 
-  try:
-    os.remove(fn)
-  except PermissionError:
-    pass
+  remove_crashlog(fn)
 
 
 def main() -> NoReturn: