On Thu, 22 Aug 2024 at 16:27, Christian Brauner <brauner@xxxxxxxxxx> wrote:
>
> >
> > /* no barrier needs as both waker and waiter are in spin-locked regions */
>
> Thanks for the analysis. I was under the impression that wait_on_inode()
> was called in contexts where no barrier is guaranteed and the bit isn't
> checked with spin_lock() held.
Yes, it does look like the barrier is needed, because the waiter does
not hold the lock indeed.
Linus
