Re: [PATCH nf-next v4 00/10] netfilter: nft_bitwise: shift support

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Tue, Jan 28, 2020 at 11:31:39AM +0000, Jeremy Sowden wrote:
> On 2020-01-28, at 11:00:35 +0100, Pablo Neira Ayuso wrote:
> > On Mon, Jan 27, 2020 at 11:13:14AM +0000, Jeremy Sowden wrote:
> > > On 2020-01-26, at 12:12:51 +0100, Pablo Neira Ayuso wrote:
> > > > I've been looking into (ab)using bitwise to implement add/sub. I
> > > > would like to not add nft_arith for only this, and it seems to me
> > > > much of your code can be reused.
> > > >
> > > > Do you think something like this would work?
> > >
> > > Absolutely.
> > >
> > > A couple of questions.  What's the use-case?
> >
> > inc/dec ip ttl field.
> 
> If it's just a simple addition or subtraction on one value, would
> this make more sense?
> 
>         for (i = 0; i < words; i++) {
> 	        dst[i] = src[i] + delta;
> 	        delta = dst[i] < src[i] ? 1 : 0;
>         }

This can be done through _INC / _DEC instead, however...

> > > I find the combination of applying the delta to every u32 and having
> > > a carry curious.  Do you want to support bigendian arithmetic (i.e.,
> > > carrying to the left) as well?
> >
> > Userspace should convert to host endianess before doing arithmetics.
> 
> Yes, but if the host is bigendian, the least significant bytes will be
> on the right, and we need to carry to the left, don't we?
> 
>         for (i = words; i > 0; i--) {
> 	        dst[i - 1] = src[i - 1] + delta;
> 	        delta = dst[i - 1] < src[i - 1] ? 1 : 0;
>         }

I think some simplified version of bignum add/subtract is needed,
something like:

        for (i = len - 1; i >= 0; i--) {
                res[i] = a[i] + b[i] + carry;
                carry = res[i] < a[i] + b[i];
        }

where 'len' is in bytes. Values in a[] and b[] are u8 and data is
represented in big endian.