Re: [PATCH] IPv6: optimize echo reply checksum calculation

[Al Viro <viro@ftp.linux.org.uk>]

On Fri, Nov 10, 2006 at 02:04:32PM -0500, Brian Haley wrote:
> Al Viro wrote:
> >so -= 1 is broken even on ia64 and it's *always* broken on big-endian
> >boxen.
> 
> It's not broken in ia64, I've tested that, just don't have an x86 for 
> testing right now.  Can you please apply these changes and prove it's 
> broken?  This little trick has been done in other UNIXes for years 
> without any problems.

Could you fscking read what you've replied to?  Your -=1 will turn 0
into 0xffff instead of correct 0xfffe.  IOW, it's broken in 1:65536
cases.

On big-endian boxen (which x86 is not, BTW), it will *always* give the
wrong result. -= htons(0x100) is better, but still fscks up in 1:256.

You _can_ adjust the checksum; however it's not that trivial.  One working
variant is
	if (sum > htons(0x100)
		sum -= htons(0x100);
	else
		sum += 0xffff - htons(0x100);
In little-endian case it turns into
	if (sum > 1)
		sum--;
	else
		sum += 0xfffe;
which is why your variant breaks rarely on itanic (or x86 - they are not
different in that respect).  You still need to handle that corner case,
though.

As for the various Unices doing the same trick, I suggest you to check
what they _really_ do and report bugs for those that have pure -- and
nothing else.

Again, the checksum is sum modulo 0xffff, *not* 0x10000.  And endianness
matters, of course, since the packet type is either LSB or MSB, depending
on it.