[e2e] Re: [Tsvwg] Really End-to-end or CRC vs everything else?

[Bob Braden]

  *> 
  *> That all suggests the error processes causing the errors aren't due to
  *> `on the wire errors'; instead, they're errors in the hardware (or
  *> software) which transfers packets between the TCP/IP stack buffer, and
  *> the link-level hardware with its FEC and hardware CRCs; with a few
  *> single-bit or single-word errors thrown in for good measure.
  *> 
  *> 

The viewpoint of the orginators of TCP was that the TCP checksum was
primarily intended to catch software errors ... the canonical example
being an output buffer that is returned to the freelist prematurely.
In practice, the TCP checksum seemed to catch this kind of software
corruption pretty well (it happened to me in the 360/91 TCP ;-))

We recognized that the TCP checksum was not very strong, and we assumed
that the link layer CRC would find wire-originated bit errors that
might escape the TCP checksum.

 Bob Braden