>     [sws from rfc]
>
>   By following this algorithm, the right window edge may not be advanced
>   by duplicate ACKs in your example and in the trace of the original post.
>
> The conditions under which the window is not advanced are different.
> In the spec you quote, it's about not making small updates to the
> window when it is substantially closed.  The case under discussion is
> about only making updates on packets which advance the ack pointer
> (this should not happen if window was fairly closed and is being
> opened a lot, though).

Oops!  I see.  I am sorry I was wrong.  I must be sleeping.

Let me try again.

  Suppose RCV.WUACK holds acknowledgment number when RCV.NXT+RCV.WND
  is advanced last time.  Its initial value is IRS + 1.
  And sppose RCV.BUFF is the total receive buffer space.

  RCV.NXT+RCV.WND (or rcv_adv) may be advanced if
    SEG.LEN > 0 or
    RCV.NXT - RCV.WUACK >= min( 1/2 * RCV.BUFF, MSS )

Regards,
Noritoshi Demizu