Re: silent dropping of TCP segment

At 08:21 AM 3/25/2004, Barry Reinhold wrote:
>MPA clause 9, error semantics, states that an MPA aware TCP
>implementation may silently drop a TCP segment when MPA detects a CRC
>error. The goal of this TCP segment dropping would be to avoid forcing
>the teardown of the stream by having TCP retransmit the segment.
>
>However, clause 5.1.2 on page 11 requires that
>
>"The TCP layer MUST have committed to keeping each segment before it can
>be passed to the MPA."
>
>I assume "committed" means that TCP ACKs the segment.

Yes.

>Thus the segment is ACKed before it is delivered to MPA.

In a well-layered implementation this is true.

>If this is the proper understanding I don't see how "MPA" can silently 
>drop a segment, since
>by the time it can detect a CRC error TCP has already committed to it.

Because it is not a well-layered implementation if an MPA-aware TCP 
implementation is used.  This has also been called a rototilled 
implementation in that a design might calculate the CRC, etc. while also 
calculating the checksums in parallel as the data arrives.   Such a design 
is making an optimization for performance reasons (don't have to scan the 
data twice) and therefore has an opportunity to detect the CRC within its 
implementation and not generate a TCP ACK.

>It is not clear to me why there is a "MUST" in clause 5.1.2. It would
>seem desirable to allow MPA some control over when the ACK is to be
>generated.

The specification assume a well-layered implementation from an architecture 
perspective (per the original charter discussions) and has normative text 
using only that perspective.

>I would like to see the "MUST" removed from clause 5.1.2.

I disagree per the logic provided above within this e-mail.

Mike


>Barry Reinhold
>Lamprey Networks
>[email protected]
>(603) 868-8411
>
>
>
>
>
>_______________________________________________
>rddp mailing list
>[email protected]
>https://www1.ietf.org/mailman/listinfo/rddp