fix: scoped concurrency semaphore over-release on cancellation

The sync-concurrency fix marked the scoped-override semaphore for release
before it was actually acquired, so a coroutine cancelled while polling for
a scoped permit (or a sync acquire interrupted mid-wait) ran the finally and
released a permit it never held — inflating the scoped cap for later calls
(the mirror of the ceiling-leak fix). Only bind the release guard after the
acquire succeeds, in both the async and sync semaphores. Regression test
included: cancelling a waiter leaves the scoped permit count at 1, not 2.
This commit is contained in:
mountain 2026-07-09 23:08:39 +08:00
parent 56590c63d5
commit d3ea9b9f2e
2 changed files with 49 additions and 9 deletions

View file

@ -90,18 +90,20 @@ async def _llm_semaphore():
# we've already given up on it.
while not ceiling_sem.acquire(False):
await asyncio.sleep(0.05)
# Only set once the permit is actually held, so a cancellation while polling
# for it doesn't make the finally release a permit we never acquired (which
# would inflate the scoped cap — the mirror of the ceiling leak fixed above).
scoped_sem = None
try:
effective = get_max_concurrency()
ceiling = _process_wide_max_concurrency()
if effective < ceiling:
scoped_sem = _max_concurrency_scope_semaphore()
if scoped_sem is not None:
while not scoped_sem.acquire(False):
candidate = _max_concurrency_scope_semaphore()
if candidate is not None:
while not candidate.acquire(False):
await asyncio.sleep(0.05)
yield
else:
yield
scoped_sem = candidate
yield
finally:
if scoped_sem is not None:
scoped_sem.release()
@ -117,14 +119,17 @@ def _sync_llm_semaphore():
"""
ceiling_sem = _process_ceiling_semaphore()
ceiling_sem.acquire()
# Only set once the permit is actually held (mirrors _llm_semaphore): guards
# against releasing a permit we never acquired if acquire() is interrupted.
scoped_sem = None
try:
effective = get_max_concurrency()
ceiling = _process_wide_max_concurrency()
if effective < ceiling:
scoped_sem = _max_concurrency_scope_semaphore()
if scoped_sem is not None:
scoped_sem.acquire()
candidate = _max_concurrency_scope_semaphore()
if candidate is not None:
candidate.acquire()
scoped_sem = candidate
yield
finally:
if scoped_sem is not None: