driver packet coalesce

[Jack Vogel]

On 5/31/07, Andre Oppermann <andre at freebsd.org> wrote:
> Julian Elischer wrote:
> > Andrew Thompson wrote:
> >> On Wed, May 30, 2007 at 04:45:05PM -0700, Jack Vogel wrote:
> >>> Does any driver do this now? And if a driver were to coalesce
> >>> packets and send something up the stack that violates mss
> >>> will it barf?
> >>
> >> It would barf for things like bridging where the packet gets spit out a
> >> different interface. The bridge driver already has code to disable
> >> txcsum so it could be made to handle that too.
> >>
> >>
> >> Andrew
> >> _______________________________________________
> >> freebsd-net at freebsd.org mailing list
> >> http://lists.freebsd.org/mailman/listinfo/freebsd-net
> >> To unsubscribe, send any mail to "freebsd-net-unsubscribe at freebsd.org"
> >
> > This is part od TOE right?
>
> No, LRO (Large Receive Offload).
>
> > I presume that it wouldn't coalesce packets that are not destined for
> > the local
> > machine?  would it coalesce in promiscuous mode?   I guess it would only
> > be able to coalesce TCP packets that are adjacent in the same session.
> > Whether it also can coalesce adjacent packets that are destined for another
> > machine (for which it is not running the session) is not known... I
> > would guess it
> > wouldn't do it.
>
> These are all nasty problems that should be handled in one central
> place for various protocols (primarily TCP right now).  For TCP there
> are a number of obvious and non-obvious interaction issues with LRO
> to be handled.  For example LRO may have a drastic effect on stacks
> that don't (yet) make use of ABC and increase the CWND by one for
> every ACK received.  With LRO up to some 44 segments may be aggregated.
> On back to back connections with microsecond RTT this isn't much a
> problem.  However once you have even 1ms performance goes way down.
> The LRO implementation (in the driver) has to be aware of all these
> issues and how the TCP stack treats them.  That's why I want to have
> only one LRO implementation in the base system that is used by all
> drivers.

>From email I've seen it appears Linux is working thru this issue as well,
and I agree that it makes sense to have one implementation, so lets
get going on it :)

Jack