Files
hermes-agent/hermes_state_lockguard.py
teknium1 87bb0d3827 fix(state): a closing writer keeps its WAL generation until SQLite's own close (#121429)
SessionDB.close() lifts the OFD lock guard before sqlite3_close so a true
last close can still end the generation. When a stray in-process
open()/close() has already cancelled the handle's POSIX locks (the case
the guard exists for), lifting the OFD copy leaves the still-open
connection with no lock at all. A sibling closing in that gap takes
EXCLUSIVE and unlinks -wal/-shm under it, and every opener in the
meantime refuses with DeletedWalGenerationError.

release() now re-takes SQLite's own process-owned POSIX read lock on each
range before it drops the OFD copy. That is the lock SQLite still thinks
it holds. The re-take cannot be refused, because our OFD lock already
excludes writers. SQLite's close upgrades or drops the lock itself, so a
true last close still unlinks the generation.

Repro: the torture chamber's lock_cancellation episode on CI
(job 107586407677: "lock_cancellation-tui held state.db-shm (deleted)").
Deterministic with a sleep injected between release() and the connection
close: base red 4/4, fix green 4/4. The new lifecycle test runs a sibling
sqlite3 close at that seam: red on base (-shm unlinked), green on the fix,
and it checks that the last close still removes -wal.
2026-09-26 11:32:10 -07:00

198 lines
9.0 KiB
Python

"""Hold a state.db writer's WAL-mode file locks in a form a stray ``close()`` cannot cancel.
SQLite protects a live WAL generation with two POSIX advisory locks: a SHARED lock on the main
file's lock range and a shared lock on the DMS byte of ``state.db-shm``. A sibling process may
checkpoint and unlink ``-wal``/``-shm`` at its close only after taking both EXCLUSIVE. POSIX locks
are per process, so any ``open()``/``close()`` of those two files inside the holder — a raw probe,
a plugin, a tool reading ``~/.hermes`` — cancels both (sqlite.org/howtocorrupt.html §2.2) and the
next foreign close strands the holder on a deleted generation (``DeletedWalGenerationError``).
This module adds the same two ranges as *open file description* locks (``F_OFD_SETLK``) on the
descriptors SQLite itself holds. OFD locks belong to the description, not the process: a stray
``close()`` elsewhere cannot cancel them, they die with the connection's own descriptor (nothing
extra to track or retire), and they conflict with a foreign EXCLUSIVE exactly like SQLite's own,
so the sibling's close-time unlink is refused while a guarded handle is open. The guard is
lifted before the handle's own close so a true last close still ends the generation normally;
lifting it first re-takes SQLite's own POSIX locks on the same ranges, so the handle is never
unlocked while its connection is still open. No-op on Windows and on runtimes without OFD locks.
Ownership model: the guard is a property of the *descriptor*, and a descriptor number is
reusable. Each ``hold()`` therefore locks every matching descriptor unconditionally (an OFD
re-lock on an already-locked description is idempotent) and ``release()`` unlocks only while
another handle in this process still needs the range — tracked by handle count per INODE, not
per fd, so a recycled fd number can never be mistaken for a surviving lock.
"""
from __future__ import annotations
import logging
import os
import struct
import sys
import threading
from typing import Dict, Optional, Set, Tuple
logger = logging.getLogger("hermes_state")
# SQLite's unix VFS lock geometry (os_unix.c): the SHARED range on the main file and the
# deadman-switch byte of the -shm file.
_PENDING_BYTE = 0x40000000
_SHARED_FIRST = _PENDING_BYTE + 2
_SHARED_SIZE = 510
_SHM_DMS_BYTE = 128
# Windows has no POSIX advisory locks, and the module promises to be a no-op there. Gate on the
# platform FIRST: some Windows installs have a third-party module importable as `fcntl` (stock
# CPython for Windows ships none), and letting the import decide would arm the guard on a
# lookalike. Off Windows a partial module is equally fatal at import time — every importer of
# hermes_state dies before supported() can say "no" (#118026) — so tolerate a missing attribute
# the same way as a missing module and fall back to the no-op.
if os.name == "nt":
fcntl = None # type: ignore[assignment]
_F_OFD_SETLK: Optional[int] = None
_F_RDLCK = _F_UNLCK = _SEEK_SET = _F_SETLK = 0
else:
try:
import fcntl
# CPython exports F_OFD_SETLK only from 3.12. The kernel ABI values are stable: 37 on every
# Linux arch (asm-generic/fcntl.h), 90 on XNU (bsd/sys/fcntl.h, documented in fcntl(2)).
_F_OFD_SETLK = getattr(
fcntl, "F_OFD_SETLK", {"linux": 37, "darwin": 90}.get(sys.platform.rstrip("0123456789")))
_F_RDLCK, _F_UNLCK, _SEEK_SET = fcntl.F_RDLCK, fcntl.F_UNLCK, os.SEEK_SET
_F_SETLK = fcntl.F_SETLK
# The constants alone do not make the guard usable: _ofd_lock() calls fcntl.fcntl(), so a
# module that has the constants but no callable would pass supported() and then raise
# from the first hold(). Probe the whole capability here, not just the symbols.
if not callable(getattr(fcntl, "fcntl", None)):
raise ImportError("fcntl module has no fcntl() callable")
except (ImportError, AttributeError):
fcntl = None # type: ignore[assignment]
_F_OFD_SETLK = None
_F_RDLCK = _F_UNLCK = _SEEK_SET = _F_SETLK = 0
# struct flock differs per libc: glibc/musl put type+whence first, Darwin/BSD last.
_FLOCK_FORMAT = "@qqihh" if sys.platform == "darwin" or "bsd" in sys.platform else "@hhqqi"
Identity = Tuple[int, int]
Held = Dict[Identity, Tuple[int, int]] # inode this handle guards -> its (start, length) range
# Handles per guarded inode in this process. Several SessionDB handles on one file share the
# same descriptors' locks (hold() locks every matching descriptor), so the LAST handle unlocks.
_LOCK = threading.Lock()
_HANDLES: Dict[Identity, int] = {}
def supported() -> bool:
return _F_OFD_SETLK is not None
def _flock(lock_type: int, start: int, length: int) -> bytes:
if _FLOCK_FORMAT == "@qqihh":
return struct.pack(_FLOCK_FORMAT, start, length, 0, lock_type, _SEEK_SET)
return struct.pack(_FLOCK_FORMAT, lock_type, _SEEK_SET, start, length, 0)
def _ofd_lock(fd: int, lock_type: int, start: int, length: int, *, cmd: Optional[int] = None) -> bool:
"""Apply a non-blocking OFD lock (a process-owned POSIX one with ``cmd=_F_SETLK``); False when
the range is held EXCLUSIVE elsewhere."""
assert fcntl is not None and _F_OFD_SETLK is not None
try:
fcntl.fcntl(fd, _F_OFD_SETLK if cmd is None else cmd, _flock(lock_type, start, length))
except BlockingIOError:
return False
return True
def _identity(path: str) -> Optional[Identity]:
try:
st = os.stat(path)
except OSError:
return None
return (st.st_dev, st.st_ino)
def _own_fds_for(identities: Set[Identity]):
"""Yield ``(fd, identity)`` for every descriptor of this process on one of *identities*
(SQLite's own connection descriptors; the cached header-probe fd too, harmless)."""
for fd_dir in ("/proc/self/fd", "/dev/fd"):
try:
names = os.listdir(fd_dir)
except OSError:
continue
for name in names:
if not name.isdigit():
continue
fd = int(name)
try:
st = os.fstat(fd)
except OSError:
continue
ident = (st.st_dev, st.st_ino)
if ident in identities:
yield fd, ident
return
def _guard_ranges(db_path) -> Held:
base = os.fspath(db_path)
ranges: Held = {}
for path, rng in ((base, (_SHARED_FIRST, _SHARED_SIZE)), (base + "-shm", (_SHM_DMS_BYTE, 1))):
ident = _identity(path)
if ident is not None:
ranges[ident] = rng
return ranges
def hold(db_path, held: Optional[Held] = None) -> Held:
"""Lock the guard ranges on every descriptor this process has open on ``state.db`` and its
``-shm``. Returns the record :func:`release` needs; pass it back to extend an existing one
(a ``-shm`` minted after open, a reopened connection). Idempotent per handle: an inode already
in *held* is re-locked (cheap, covers a new descriptor) without a second handle count."""
held = {} if held is None else held
if not supported():
return held
ranges = _guard_ranges(db_path)
try:
with _LOCK:
for fd, ident in _own_fds_for(set(ranges)):
start, length = ranges[ident]
if _ofd_lock(fd, _F_RDLCK, start, length) and ident not in held:
held[ident] = ranges[ident]
_HANDLES[ident] = _HANDLES.get(ident, 0) + 1
except OSError:
logger.debug("WAL lock guard unavailable for %s", os.fspath(db_path), exc_info=True)
return held
def release(held: Held) -> None:
"""Drop this handle's claim. The last handle on an inode unlocks the range on every descriptor
still referencing it. Call BEFORE the handle's own close so SQLite's close-time reset sees only
real holders: a sibling process's intact locks still refuse the unlink, and a true last close
ends the generation, so a later ``state.db`` replace never pairs with a stale WAL.
The OFD copy is swapped back for the POSIX lock SQLite believes it still holds (a stray close
cancelled the real one): unlocked outright, the still-open connection would let a sibling's
close take EXCLUSIVE and unlink ``-wal``/``-shm`` under it before its own close runs. SQLite's
close upgrades or drops that process-owned lock itself, so a true last close is unaffected."""
if not supported() or not held:
return
with _LOCK:
to_unlock: Held = {}
for ident, rng in held.items():
remaining = _HANDLES.get(ident, 1) - 1
if remaining > 0:
_HANDLES[ident] = remaining
else:
_HANDLES.pop(ident, None)
to_unlock[ident] = rng
held.clear()
if not to_unlock:
return
try:
for fd, ident in _own_fds_for(set(to_unlock)):
start, length = to_unlock[ident]
_ofd_lock(fd, _F_RDLCK, start, length, cmd=_F_SETLK) # never refused: our OFD lock excludes writers
_ofd_lock(fd, _F_UNLCK, start, length)
except OSError:
pass