ECN and IPv4 Header Checksum Recalculation:

----------------------------------------------------------------------------

Note from Sally, 2/13/98:  The current plan is for the ECN bit to be at
bit 7 of the IPv4 TOS octet, and for the ECN-capable bit to be located
at bit 6.

----------------------------------------------------------------------------

The following is from email sent by Steven Blake,
slblake@raleigh.ibm.com, on 12 Feb 1998.  Sally Floyd and Kevin Fall
have made a slight modification to the header checksum recalculation
for one's complement subtraction.

------------

Subject: ECN field in Diffserv

...
In DC I mentioned to you in the hallway that I was going to look at the
implications of IPv4 header checksum recalculation when the ECN bit was
set (where it had previously been 0).  As I mentioned, this is a potential
problem with some high-end router architectures using an output-buffered
switch, since most if not all of the header manipulation is performed on
the input side of the switch, while the ECN decision would need to made
local to the output buffer.

It turns out that the IPv4 TOS octet is the last byte of a 16-bit half-word.
I did a quick analysis and this is what I found:

  HC:  original header checksum
  HC': new header checksum (after ECN bit is set)

   For ECN bit at bit 7, One's complement subtraction:
 
   HC' = { HC - 1     HC > 1
 	 { 0x0000     HC = 1
 
   For ECN bit at bit 6, One's complement subtraction:
 
   HC' = { HC - 2     HC > 2 or HC = 1
 	 { 0x0000     HC = 2

The more interesting problem occurs when two's complement subtraction is
implemented.  In this case, there are special conditions:

  For ECN bit at bit 7:

    HC' = { HC - 1     HC > 0
          { 0xFFFE     HC = 0

  For ECN bit at bit 6:

    HC' = { HC - 2     HC > 1
          { 0xFFFD     HC = 0
          { 0xFFFE     HC = 1

Two's complement subtraction would cause the exception cases to be off
by +1.

Based on this brief analysis ..., I
would recommend that the ECN bit be located at bit 7 (the last bit of the
TOS/Class octet).  The ECN-capable bit could be located at bit 6.

This is of course not a problem for IPv6, since there is no IPv6 header 
checksum.  If ECN is never standardized for IPv4 then the bit locations
don't matter.
...

Steven L. Blake                     

----------------------------------------------------------------------------

Date: Fri, 13 Feb 1998 10:29:27 EST
From: "Steven L. Blake" <slblake@raleigh.ibm.com>
To: Sally Floyd <floyd@ee.lbl.gov>
Subject: Re: ECN field in Diffserv

...
Please add that the results were derived from Eqs. 3 & 4
in RFC 1624, and that all references to one's complement addition/subtraction
are 16-bit (I should have clarified this in my note).
...

----------------------------------------------------------------------------