Re: [PATCH] locking/qrwlock: queued_write_lock_slowpath() cleanup

From: Peter Zijlstra
Date: Mon Apr 26 2021 - 04:09:49 EST


On Sun, Apr 25, 2021 at 04:06:37PM -0400, Waiman Long wrote:

> void queued_write_lock_slowpath(struct qrwlock *lock)
> {
> - int cnts;
> + int cnts = 0;
>
> /* Put the writer into the wait queue */
> arch_spin_lock(&lock->wait_lock);
>
> /* Try to acquire the lock directly if no reader is present */
> if (!atomic_read(&lock->cnts) &&
> - (atomic_cmpxchg_acquire(&lock->cnts, 0, _QW_LOCKED) == 0))
> + atomic_try_cmpxchg_acquire(&lock->cnts, &cnts, _QW_LOCKED))
> goto unlock;

Would not something like:

if (!(cnts = atomic_read(&lock->cnts)) &&
atomic_try_cmpxchg_acquire(&lock->cnts, &cnts, _QW_LOCKED)
goto unlock;

Be clearer?

>
> - /* Set the waiting flag to notify readers that a writer is pending */
> - atomic_add(_QW_WAITING, &lock->cnts);
> + /*
> + * Set the waiting flag to notify readers that a writer is pending
> + *
> + * As only one writer who is the wait_lock owner can set the waiting
> + * flag which will be cleared later on when acquiring the write lock,
> + * we can easily replace atomic_or() by an atomic_add() if there is
> + * an architecture where an atomic_add() performs better than an
> + * atomic_or().

That might be a little overboard on the comment, but sure :-) I don't
think there's any arch that doesn't have atomic_or(), like I wrote
elsewhere, the one that's often an issue is atomic_fetch_or().

> + */
> + atomic_or(_QW_WAITING, &lock->cnts);