Right now, the main loop in semlock_acquire looks like this:

    do {
        Py_BEGIN_ALLOW_THREADS
        if (blocking && timeout_obj == Py_None)
            res = sem_wait(self->handle);
        else if (!blocking)
            res = sem_trywait(self->handle);
        else
            res = sem_timedwait(self->handle, &deadline);
        Py_END_ALLOW_THREADS
        err = errno;
        if (res == MP_EXCEPTION_HAS_BEEN_SET)
            break;
    } while (res < 0 && errno == EINTR && !PyErr_CheckSignals());

Here, we unconditionally release the GIL even we could acquire the mutex without blocking! As a result, we could end up switching to another thread in the process and greatly increasing the latency of operations that lock and release multiple shared data structures.

Instead, we should unconditionally try sem_trywait, and only then, if we want to block, release the GIL and try sem_wait or sem_timedwait. This way, we'll churn only when we need to.

Note that threading.Lock works the way I propose.

Looks reasonable to me. Do you want to provide a patch Daniel?

New changeset c872d39d324cd6f1a71b73e10406bbaed192d35f by Antoine Pitrou in branch 'master':
bpo-31653: Don't release the GIL if we can acquire a multiprocessing semaphore immediately (#4078)
https://github.com/python/cpython/commit/c872d39d324cd6f1a71b73e10406bbaed192d35f

New changeset 828ca59208af0b1b52a328676c5cc0c5e2e999b0 by Victor Stinner in branch 'master':
bpo-31653: Remove deadcode in semlock_acquire() (#4091)
https://github.com/python/cpython/commit/828ca59208af0b1b52a328676c5cc0c5e2e999b0