On Sun, Sep 19, 2021 at 11:28 AM Thomas Gleixner <tglx@linutronix.de> wrote:
>
>
> -       atomic_add(READER_BIAS - bias, &rwb->readers);
> +       /*
> +        * _release() is needed in case that reader is in fast path, pairing
> +        * with atomic_try_cmpxchg() in rwbase_read_trylock(), provides RELEASE
> +        */
> +       (void)atomic_add_return_release(READER_BIAS - bias, &rwb->readers);

Ugh. This really needs fixing.

atomic_add() is already much more than release-ordered on x86, and
atomic_add_return_release() is much more expensive on some uarchs.

I think it should be easy to add a atomic_add_release() function, and
it might be as simple as the attached patch, allowing architectures to
add their own arch_atomic_add_release() as needed.

I've pulled this, but please don't do things like the above hack.

            Linus