Re: [PATCH net-next 2/2] net: optimize skb_postpull_rcsum()

[Eric Dumazet <edumazet@google.com>]

On Thu, Dec 2, 2021 at 7:06 AM David Laight <David.Laight@aculab.com> wrote:
>
> From: Vladimir Oltean
> > Sent: 02 December 2021 13:11
> ...
> > > --- a/include/linux/skbuff.h
> > > +++ b/include/linux/skbuff.h
> > > @@ -3485,7 +3485,11 @@ __skb_postpull_rcsum(struct sk_buff *skb, const void *start, unsigned int
> > len,
> > >  static inline void skb_postpull_rcsum(struct sk_buff *skb,
> > >                                   const void *start, unsigned int len)
> > >  {
> > > -   __skb_postpull_rcsum(skb, start, len, 0);
> > > +   if (skb->ip_summed == CHECKSUM_COMPLETE)
> > > +           skb->csum = ~csum_partial(start, len, ~skb->csum);
>
> You can't do that, the domain is 1..0xffff (or maybe 0xffffffff).
> The invert has to convert ~0 to ~0 not zero.
> ...
> > There seems to be a disparity when the skb->csum is calculated by
> > skb_postpull_rcsum as zero. Before, it was calculated as 0xffff.
>
> Which is what that will do for some inputs at least.
> Maybe:
>                 skb->csum = 1 + ~csum_partial(start, len, ~skb->csum + 1);
> is right.
> I think that is the same as:
>                 skb->csum = -csum_partial(start, len, -skb->csum);
> Although letting the compiler do that transform probably makes
> the code easier to read.
>
>

Interesting, update_csum_diff4() and update_csum_diff16() seem to both use.

skb->csum = ~csum_partial((char *)diff, sizeof(diff), ~skb->csum);