Re: [tcpm] timestamp options (was Re: New Version Notification for draft-touch-tcpm-tcp-edo-01.txt)

[Yuchung Cheng <[e-mail]>]

On Mon, Jun 2, 2014 at 1:35 PM, Mark Allman <[e-mail]> wrote:
>
>
> > You talk about limiting the rate; that doesn't help a connection
> > that cannot proceed because packets with the TS option are being
> > dropped, e.g., by a midbox.
>
> Sure- in a spec you'd have to consider that.  But, I am not going to
> pretend that it is difficult to detect and react to.
>
> > Further, what happens to the packets? When and how do you know when
> > the TS isn't working? When do you give up?
>
> One RTT later.  Don't pretend this is difficult.  I didn't fully specify
> it in an **email message**, but don't pretend this is some leap into
> the Great Protocol Engineering Unknown, Joe.
>
> > 1323bis talks about a list of mechanisms that rely on timestamps, more
> > than just PAWS. What happens to those when TS is optional or started
> > late?
>
> If you want to use something and that something needs the TS option to
> be on for the entire connection (e.g., Eifel) then you turn it on for
> the entire connection.  Fine.  I have no problem with that.  But, then
> you're not simply wasting 10 bytes in every packet, but you're getting
> something for it.


Or we can just (dynamically) start tagging TS on retransmission to
support Eifel.

We have developed some nice features that use TS but 12 bytes is not trivial.
It could also cost a sack block. I would love to see proposals that
support these features w/o requiring always-on TS.

>
> However, on its own 1323(bis) wastes 10 bytes in every packet because
> RTTM is not helpful and PAWS is nearly never needed.  So, it seems to me

I second that TS RTTM is marginally helpful.

In non-retransmitted sequence it's useless b/c ACK/SACK take care of
it and they are less vulnerable to bad middleboxes. We've seen
middleboxes messing TS option values.
Even in retransmitted cases, TS can only be used if the ACK advances
SND_UNA. So if some retransmission are lost again then TS aren't
useful to get RTTM.
Therefore Linux has changed to use TS as the last resort in RTTM.
http://lxr.linux.no/linux+v3.15/net/ipv4/tcp_input.c#L2887

> that instead of dogmatically putting our foot down we could re-think
> providing the information for PAWS when we actually need PAWS.  I guess
> that is radical engineering or something, I dunno ...
>
> allman
>
>
>
>
> _______________________________________________
> tcpm mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/tcpm
>