On Wed, Nov 24, 2021 at 10:56 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> On Wed, Nov 24, 2021 at 10:20 PM Eric Dumazet <edumazet@google.com> wrote:
> >
> > On Wed, Nov 24, 2021 at 8:08 PM Eric Dumazet <edumazet@google.com> wrote:
> > >
> > > On Wed, Nov 24, 2021 at 8:00 PM Eric Dumazet <edumazet@google.com> wrote:
> > >
> > > >
> > > > It is an issue in general, not in standard cases because network
> > > > headers are aligned.
> > > >
> > > > I think it came when I folded csum_partial() and do_csum(), I forgot
> > > > to ror() the seed.
> > > >
> > > > I suspect the following would help:
> > > >
> > > > diff --git a/arch/x86/lib/csum-partial_64.c b/arch/x86/lib/csum-partial_64.c
> > > > index 1eb8f2d11f7c785be624eba315fe9ca7989fd56d..ee7b0e7a6055bcbef42d22f7e1d8f52ddbd6be6d
> > > > 100644
> > > > --- a/arch/x86/lib/csum-partial_64.c
> > > > +++ b/arch/x86/lib/csum-partial_64.c
> > > > @@ -41,6 +41,7 @@ __wsum csum_partial(const void *buff, int len, __wsum sum)
> > > >         if (unlikely(odd)) {
> > > >                 if (unlikely(len == 0))
> > > >                         return sum;
> > > > +               temp64 = ror32((__force u64)sum, 8);
> > > >                 temp64 += (*(unsigned char *)buff << 8);
> > > >                 len--;
> > > >                 buff++;
> > > >
> > > >
> > >
> > > It is a bit late here, I will test the following later this week.
> > >
> > > We probably can remove one conditional jump at the end of the function
> > >
> > > diff --git a/arch/x86/lib/csum-partial_64.c b/arch/x86/lib/csum-partial_64.c
> > > index 1eb8f2d11f7c785be624eba315fe9ca7989fd56d..15986ad42ed5ccb8241ff467a34188cf901ec98e
> > > 100644
> > > --- a/arch/x86/lib/csum-partial_64.c
> > > +++ b/arch/x86/lib/csum-partial_64.c
> > > @@ -41,9 +41,11 @@ __wsum csum_partial(const void *buff, int len, __wsum sum)
> > >         if (unlikely(odd)) {
> > >                 if (unlikely(len == 0))
> > >                         return sum;
> > > +               temp64 = ror32((__force u64)sum, 8);
> > >                 temp64 += (*(unsigned char *)buff << 8);
> > >                 len--;
> > >                 buff++;
> > > +               odd = 8;
> > >         }
> > >
> > >         while (unlikely(len >= 64)) {
> > > @@ -129,10 +131,7 @@ __wsum csum_partial(const void *buff, int len, __wsum sum)
> > >  #endif
> > >         }
> > >         result = add32_with_carry(temp64 >> 32, temp64 & 0xffffffff);
> > > -       if (unlikely(odd)) {
> > > -               result = from32to16(result);
> > > -               result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
> > > -       }
> > > +       ror32(result, odd);
> >
> > this would be
> >           result = ror32(result, odd);
> >
> > definitely time to stop working today for me.
> >
> > >         return (__force __wsum)result;
> > >  }
> > >  EXPORT_SYMBOL(csum_partial);
>
> All my tests pass with that change :)

Although I see slightly worse performance with aligned `buff`  in
the branch-free approach. Imagine if non-aligned `buff` is that
uncommon might be better to speculate past the work of `ror`.