On Fri, Sep 19, 2025 at 7:06 PM Andrea Parri <parri.and...@gmail.com> wrote:
>
> > > > (not a review, just looking at this diff stat) is changing the fastpath
> > > >
> > > > read_unlock()
> > > > read_lock()
> > > >
> > > > from something like
> > > >
> > > > fence rw,w
> > > > amodadd.w
> > > > amoadd.w
> > > > fence r,rw
> > > >
> > > > to
> > > >
> > > > fence rw,rw
> > > > amoadd.w
> > > > amoadd.w
> > > > fence rw,rw
> > > >
> > > > no matter Zalasr or !Zalasr. Similarly for other atomic operations with
> > > > release or acquire semantics. I guess the change was not intentional?
> > > > If it was intentional, it should be properly mentioned in the changelog.
> > >
> > > Sorry about that. It is intended. The atomic operation before
> > > __atomic_acquire_fence or operation after __atomic_release_fence can
> > > be just a single ld or sd instruction instead of amocas or amoswap. In
> > > such cases, when the store release operation becomes 'sd.rl', the
> > > __atomic_acquire_fence via 'fence r, rw' can not ensure FENCE.TSO
> > > anymore. Thus I replace it with 'fence rw, rw'.
>
> But you could apply similar changes you performed for xchg & cmpxchg: use
> .AQ and .RL for other atomic RMW operations as well, no? AFAICS, that is
> what arm64 actually does in arch/arm64/include/asm/atomic_{ll_sc,lse}.h .
I see. I will study the implementation of ARM and refine my patch. Thanks a lot.
Best regards,
Xu Lu
>
> Andrea
>
>
> > This is also the common implementation on other architectures who use
> > aq/rl instructions like ARM. And you certainly already knew it~
