AtomicInteger.get_and_set deadlocks because of non-reentrant lock

[Yicong-Huang]

What happened?

amber/src/main/python/core/util/thread/atomic.py::AtomicInteger.get_and_set acquires self._lock and then reads self.value. The value getter is a @property that also tries to acquire self._lock. The lock is a non-reentrant threading.Lock, so the second acquisition blocks forever — calling get_and_set deadlocks the calling thread the moment it is invoked.

The fix is either to switch the lock to threading.RLock, or to inline the read inside the existing critical section (old_value = self._value instead of old_value = self.value).

How to reproduce?

import threading
from core.util.thread.atomic import AtomicInteger

a = AtomicInteger(10)
done = threading.Event()

def attempt():
    a.get_and_set(99)
    done.set()

threading.Thread(target=attempt, daemon=True).start()
done.wait(timeout=0.5)
print(done.is_set())  # False — the call never returns

Version

1.1.0-incubating (Pre-release/Master)

[nathant27]

I was able to reproduce the deadlock on my end using your sample code. I am planning on doing the minimal swap towards the inline read instead of reentrant lock since swapping to a reentrant lock to avoid unnecessary overhead, and because that change seems like the proper way to access the attribute given the context of already having the lock during access.

[chenlica]

@nathant27 Thanks. Feel free to add a reply with "[forwardslash]take" and the issue will be assigned to you.

Never mind, I saw it's already assigned to you.