sync: RwLock::downgrade is not correct

[carllerche]

I noticed CI can spuriously fail on this test.

The following loom test will fail:

#[test]
fn downgrade() {
    loom::model(|| {
        let lock = Arc::new(RwLock::new(1));

        let n = block_on(lock.write());
        let cloned_lock = lock.clone();
        let handle = thread::spawn(move || {
            let mut guard = block_on(cloned_lock.write());
            *guard = 2;
        });

        let n = n.downgrade();
        assert_eq!(*n, 1);

        assert_eq!(*block_on(lock.read()), 1);

        drop(n);

        handle.join().unwrap();
        assert_eq!(*block_on(lock.read()), 2);
    });
}

[zaharidichev]

Hm, this seems to be related to some race in the batch semaphore. I assume the following test should pass but it fails:

#[test]
fn maybe_race() {
    loom::model(|| {
        let semaphore = Arc::new(Semaphore::new(10));
        semaphore
            .try_acquire(10)
            .expect("try_acquire should succeed; semaphore uncontended");
        let semaphore2 = semaphore.clone();
        let thread = thread::spawn(move || block_on(semaphore2.acquire(10)).unwrap());
       
        semaphore.release(1);
        // this fails as semaphore.available_permits() gives 0 ...
        assert_eq!(1, semaphore.available_permits());
       
        semaphore.release(9);
        thread.join().unwrap();
        assert_eq!(0, semaphore.available_permits());
    })
}

I am keen to take a look at that if I am not stepping on somebody's toes... :)

UPDATE: Scratch the above. The test makes sense and should fail. The problem is elsewhere. Will submit a PR shortly.

[LucioFranco]

I am keen to take a look at that if I am not stepping on somebody's toes... :)

Go for it