fix(credential_pool): serialize deferred-refresh pool mutations

Review folds on the #71775 salvage (dossier findings 1+2):

- self._lock becomes an RLock and the mutation primitives
  (_replace_entry, _persist) are now self-locking, so the deferred
  single-use-token refresh path — which deliberately runs its
  cross-process flock + OAuth network I/O OUTSIDE the pool lock —
  still serializes its pool mutations against concurrent
  select()/rotation. In-lock callers re-acquire reentrantly.
- Dropped _refresh_pending_entries' redundant second _replace_entry:
  _refresh_entry already merges the refreshed entry internally.

Adds tests/agent/test_credential_pool_deferred_refresh.py pinning both
invariants: select() must NOT hold the lock during the refresh window
(the PR's whole point), and the post-refresh mutations MUST contend on
the lock (blocking-thread probe).
This commit is contained in:
kshitijk4poor 2026-08-03 18:46:56 +05:30 committed by kshitij
parent 9cd605f303
commit 687dd632a1
2 changed files with 122 additions and 15 deletions

View File

@ -588,7 +588,12 @@ class CredentialPool:
self._entries = sorted(entries, key=lambda entry: entry.priority)
self._current_id: Optional[str] = None
self._strategy = get_pool_strategy(provider)
self._lock = threading.Lock()
# RLock: the mutation primitives below (_replace_entry/_persist)
# self-acquire this lock so the DEFERRED single-use-token refresh
# path (which runs network I/O outside the lock by design) still
# serializes its pool mutations. In-lock callers re-acquire
# reentrantly at negligible cost.
self._lock = threading.RLock()
self._active_leases: Dict[str, int] = {}
self._max_concurrent = DEFAULT_MAX_CONCURRENT_PER_CREDENTIAL
# Monotonic timestamp of the last "no available entries" log, used to
@ -684,18 +689,27 @@ class CredentialPool:
return matches[0].id if len(matches) == 1 else None
def _replace_entry(self, old: PooledCredential, new: PooledCredential) -> None:
"""Swap an entry in-place by id, preserving sort order."""
for idx, entry in enumerate(self._entries):
if entry.id == old.id:
self._entries[idx] = new
return
"""Swap an entry in-place by id, preserving sort order.
Self-locking (RLock) so the deferred refresh path which
deliberately runs outside the pool lock cannot tear
``self._entries`` against a concurrent select()/rotation.
"""
with self._lock:
for idx, entry in enumerate(self._entries):
if entry.id == old.id:
self._entries[idx] = new
return
def _persist(self, *, removed_ids: Optional[List[str]] = None) -> None:
write_credential_pool(
self.provider,
[entry.to_dict() for entry in self._entries],
removed_ids=removed_ids,
)
# Self-locking (RLock): snapshotting self._entries must not race a
# concurrent rotation when called from the deferred refresh path.
with self._lock:
write_credential_pool(
self.provider,
[entry.to_dict() for entry in self._entries],
removed_ids=removed_ids,
)
def _is_terminal_auth_failure(
self,
@ -1701,10 +1715,10 @@ class CredentialPool:
On failure the entry is silently skipped.
"""
for entry, sync_fn in pending:
refreshed = self._refresh_entry(entry, force=False)
if refreshed is not None:
with self._lock:
self._replace_entry(entry, refreshed)
# _refresh_entry already merges the refreshed entry into the
# pool internally (its mutation primitives are self-locking),
# so no second _replace_entry is needed here.
self._refresh_entry(entry, force=False)
def _available_entries(
self, *, clear_expired: bool = False, refresh: bool = False,

View File

@ -0,0 +1,93 @@
"""Thread-safety of the deferred single-use-token refresh path (#71775).
The deferred path deliberately runs OAuth network I/O outside the pool
lock. These tests pin the two invariants that make that safe:
1. `select()` does NOT hold the pool lock while the deferred refresh's
network call runs (the whole point of the PR).
2. The pool mutations that follow the network call (`_replace_entry`,
`_persist`) DO re-serialize under the pool lock, so a concurrent
`select()`/rotation cannot tear `self._entries` or double-write
auth.json.
"""
import threading
from dataclasses import replace
from agent.credential_pool import (
AUTH_TYPE_OAUTH,
CredentialPool,
PooledCredential,
)
def _codex_entry(entry_id: str = "codex-1") -> PooledCredential:
return PooledCredential(
provider="openai-codex",
id=entry_id,
label="test codex",
auth_type=AUTH_TYPE_OAUTH,
priority=0,
source="device_code",
access_token="at-stale",
refresh_token="rt-stale",
expires_at_ms=1, # long expired -> needs refresh
)
def test_select_does_not_hold_pool_lock_during_deferred_refresh(monkeypatch):
pool = CredentialPool("openai-codex", [_codex_entry()])
lock_free_during_refresh = {}
def _fake_refresh(entry, *, force):
# If select() still held the pool lock here, this non-blocking
# acquire would fail — the regression this PR exists to fix.
acquired = pool._lock.acquire(blocking=False)
lock_free_during_refresh["value"] = acquired
if acquired:
pool._lock.release()
refreshed = replace(entry, access_token="at-fresh", expires_at_ms=2**53)
pool._replace_entry(entry, refreshed)
return refreshed
monkeypatch.setattr(
pool, "_entry_needs_refresh", lambda e: e.access_token == "at-stale"
)
monkeypatch.setattr(pool, "_refresh_entry", _fake_refresh)
monkeypatch.setattr(pool, "_persist", lambda **kw: None)
selected = pool.select()
assert lock_free_during_refresh.get("value") is True, (
"select() held the pool lock during the deferred refresh network window"
)
assert selected is not None
assert selected.access_token == "at-fresh"
def test_deferred_mutations_serialize_against_concurrent_rotation(monkeypatch):
"""_replace_entry/_persist from the deferred path must contend on the
pool lock: with the lock held by another thread, the deferred mutation
must block rather than mutate concurrently."""
pool = CredentialPool("openai-codex", [_codex_entry()])
monkeypatch.setattr(pool, "_persist", lambda **kw: None)
entry = pool._entries[0]
refreshed = replace(entry, access_token="at-fresh")
mutated = threading.Event()
def _deferred_mutation():
pool._replace_entry(entry, refreshed) # self-locking
mutated.set()
with pool._lock:
t = threading.Thread(target=_deferred_mutation)
t.start()
# While we hold the lock, the deferred mutation must NOT complete.
assert not mutated.wait(timeout=0.3), (
"_replace_entry mutated the pool while another thread held the lock"
)
t.join(timeout=5)
assert mutated.is_set()
assert pool._entries[0].access_token == "at-fresh"