[ofa-general] vendor_id in struct ib_device_attr
Jeff Squyres
jsquyres at cisco.com
Mon Jun 23 12:23:53 PDT 2008
Are there many apps out there yet that are depending on iWarp vendor
ID/part ID?
OMPI uses these to set device-specific values at startup, but we have
not released this code yet (it will be in our upcoming v1.3 series).
Does MVAPICH* use them? What about non-MPI ULPs / applications?
My $0.02 is to make it all one namespace since there's precedent for
it from InfiniBand: OUI's. Modify the Chelsio and NetEffect drivers
to return the OUI. Hopefully, there's not enough iWarp apps out there
reading this info that it will matter.
On Jun 23, 2008, at 3:17 PM, Roland Dreier wrote:
> It seems there is some confusion about struct
> ib_device_attr.vendor_id:
> mthca and mlx4 are using an OUI for vendor_id, while ipath and cxgb3
> use
> the PCI vendor ID (and I don't know what ehca does). Since these
> namespaces are not coordinated, there is the (remote) chance of a
> collision. As far as I can tell, the ipath usage at least is a bug,
> since the IB spec says that NodeInfo:VendorID should be "device
> vendor,
> per IEEE" suggesting that OUI is the right thing to use. If that is
> the
> case, then it makes sense for iWARP devices to use OUI as well.
>
> However changing this means that any apps that look at the vendor_id
> field need to be updated to handle both the old and new value.
>
> This comes up because iw_nes currently doesn't set vendor_id at all,
> and
> I wonder what we should stick in there.
>
> What do other people think the way forward is?
>
> - R.
>
> _______________________________________________
> general mailing list
> general at lists.openfabrics.org
> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
--
Jeff Squyres
Cisco Systems
More information about the general
mailing list