[PATCHv2] networking: consolidate the IP checksum code

[Baruch Siach]

Hi Joakim,

On Thu, Sep 08, 2011 at 01:31:31PM +0200, Joakim Tjernlund wrote:
> > Baruch Siach <baruch at tkos.co.il> wrote on 2011/09/08 07:06:27:
> > ...
> >
> > > >        if (nbytes & 1) /* Odd */
> > > >          sum += *(u_char *)ptr;   /* one byte only */
> > >
> > > This breaks big endian systems like PowerPC. The assignment via pointer dance
> > > of the original implementation is required, because we need the last byte to
> > > be considered as MSB of a 16bit word, as if we had one more '\0' byte in
> > > buffer.  Note that the generated checksum is the same on big and little endian
> > > machines in terms of memory representation.  They differ however in their
> > > numeric representation.  So, for example the checksum of
> > >
> > >     uint8_t buf[] = {0x12, 0x34, 0x56};
> > >
> > > is 0xcb97 on little endian machines, and 0x97cb on big endian machines. This
> > > is OK since we use the memory representation of the network packet.
> >
> > Ouch, even RFC 1071 got it wrong then.
> 
> I just checked and all this ptr casting confuses gcc and generates
> bad code. Using an union is better, both for ppc and x86.
> 
> This is only compile tested:
> 
> if (len & 1) {
> 	    union uu {
> 		    uint16_t val;
> 		    uint8_t b[2];
> 	    } tmp = {0};
>         /* Make sure that the left-over byte is added correctly both
>          * with little and big endian hosts */
>         tmp.b[0] = *(uint8_t*)w;
>         sum += tmp.val;
>     }

Run tested on x86_64 and PowerPC (with odd sized DHCP packets). However, I see 
no significant code size improvement:

x86_64 : -4  bytes
ARM    : +4  bytes
PowerPC: +16 bytes

baruch

-- 
                                                     ~. .~   Tk Open Systems
=}------------------------------------------------ooO--U--Ooo------------{=
   - baruch at tkos.co.il - tel: +972.2.679.5364, http://www.tkos.co.il -