On Tue, 2007-08-21 at 13:48 -0400, sowmini.varadhan at sun.com wrote:
> On (08/21/07 10:03), Garrett D'Amore wrote:
> > A quick note. Many older devices will only support "both" or "".
>
> I think this would correspond to the symmetric-pause only case- right?
Yes, but see also that some of these devices have the ability to disable
sending pause frames, even though they don't (can't!) negotiate for that
feature.
Just because you negotiated for the ability to send pause frames does
not mean you obliged to.
>
> > Some older devices will support configuring the logical equivalent of
> > "rx", but won't have a way to "expose" that to the link peer.
> >
> > For example, various tulip clones have a way to configure whether the
> > board sends pause frames or not, although they lack any notion of
> > negotiating the asymetric pause with their peer.
> >
> > This whole arena is a bit of a mess, and I'm not entirely sure we've
> > quite got it right... the proposal above is certainly simple from an
> > administrator's view, but I worry that it might be *too* simplistic.
> >
> > Until we get a lot more operational experience with pause support, lets
> > make sure we don't expose whatever we do at a commitment level higher
> > than Uncommitted. (And Volatile might be better still.)
> >
> > With that notion, perhaps leaving pause support out of the UI proper
> > (leave it a private tunable for now) might be best. I'm just afraid
> > that we will find out that we've not got it quite right, and will wind
> > up regretting making something too committed before it has fully gelled.
>
> The interface (like all the ones associated with Brussels) is marked
> Volatile. We'll make sure that we point out in all the associated
> documentation that the actual settings for this value are constrained
> by the capabilities allowed by the peer and link-partner. I believe
> that Raymond's implementation notes will return some appropriate
> error code when an attempt is made to set values that are not supported
> by the hardware.
Okay, I'm happy with Volatile.
>
> It would be good to put it out there and get some usage-feedback for
> this from both driver implementors as well as userland.. but I agree
> that this is one interface that is not likely to get any sort of
> *committed attribute in the near future.
Yes. Thanks!
-- Garrett
>
> --Sowmini
>
