[clearview-discuss] 2007/527 Addendum for Clearview Vanity Naming and Nemo Unification

[Garrett D'Amore]

Peter Memishian wrote:
>  > All the "interesting" legacy NICs (which is, IIUC, 90% for ce at this 
>  > point, because most of the others are already converted, excepting 
>  > perhaps a few rarer 3rd party NICs) are already converted to GLD.  
>
> Keep in mind that our architecture needs to consider Solaris 10, not just
> Nevada.  Most Solaris 10 networking drivers are not GLDv3-based.
>   

Good point.  I didn't realize that UV was targeting S10.

However, I wonder what the expense of this effort is, compared to the 
effort of backporting the GLDv3 work.

All of the GLDv2 NICs that have been converted should be able to be 
backported without any customer or ARC impact.  (Ok, there might be tiny 
impact related to a few kstats, because GLDv2 and v3 keep track of 
"driver private" statistics differently... but that's such a minor nit 
and undocumented that it can probably get patch binding trivially.)  
That takes care of *all* the x86 NICs, and even one of the SPARC ones 
(dmfe).

The bigger question is the DLPI nics.  Of those:

* eri and hme can be backported, but there is the "ndd" problem.  
Fixable with a hack in the "ndd" program itself to cope/workaround 
administration using style 2 instead of style 1 nodes.  Non-zero ARC 
impact, but fairly small engineering impact.

* qfe.  This has the ndd problem, but also has the additional problem of 
dealing with Sun Trunking, which is totally incompatible with GLDv3 
aggregation.  (Btw, how does UV plan to deal with this problem?  What 
are the expected interactions?)  We cannot support Sun Trunking in any 
qfe backport.  This may be a total stopper.

* ge.  This would have the same problems as qfe, except that the 
*Nevada* conversion hasn't been done yet.  It *could* be done fairly 
easily... only about a week of engineering effort is required.  
(Possibly quite a bit less.)

* ce.   This is the *big* problem.  And I suspect UV is going to have to 
cope with it somehow.  It has all the problems of qfe and ge, plus a 
bunch of tuning hints, and other things that are going to bite people in 
the arse.  And its a political minefield.   Interestingly enough, there 
is talk of Joyce Yu doing work to make it work with IP instances without 
UV.  I don't know what the precise details are, though.  You should talk 
to Erik, as I think he is involved.

    -- Garrett
> --
> meem
>