Re: [PATCH] xfs: Use try_cmpxchg() in xlog_cil_insert_pcp_aggregate()

[Uros Bizjak]

On Fri, Sep 20, 2024 at 4:16 PM Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
>
> On Thu, Sep 19, 2024 at 10:14:05AM +0200, Uros Bizjak wrote:
> > --- a/fs/xfs/xfs_log_cil.c
> > +++ b/fs/xfs/xfs_log_cil.c
> > @@ -171,13 +171,12 @@ xlog_cil_insert_pcp_aggregate(
> >        * structures that could have a nonzero space_used.
> >        */
> >       for_each_cpu(cpu, &ctx->cil_pcpmask) {
> > -             int     old, prev;
> > +             int     old;
> >
> >               cilpcp = per_cpu_ptr(cil->xc_pcp, cpu);
> > +             old = READ_ONCE(cilpcp->space_used);
>
> Maybe it is just me, but this would probably look nicer if the cilpcp
> variable moved into the loop scope, and both were initialized at
> declaration time:
>
>                 struct xlog_cil_pcp     *cilpcp = per_cpu_ptr(cil->xc_pcp, cpu);
>                 int                     old = READ_ONCE(cilpcp->space_used);

No problem, I just tried to keep the number of changed lines as low as
possible. Some maintainers don't like functional and cosmetic changes
mixed together.

>
> >               do {
> > +             } while (!try_cmpxchg(&cilpcp->space_used, &old, 0));
>
> And this also looks a bit odd.  Again, probably preference, but a:
>
>                 while (!try_cmpxchg(&cilpcp->space_used, &old, 0))
>                         ;
>
> looks somewhat more normal (although still not pretty).

Yes, the alternative form is what some maintainers prefer.

> Sorry for not having anything more substantial to see, but the diff
> just looked a bit odd..

I'll prepare a v2 patch with the suggested changes.

Thanks,
Uros.