On Fri, 7 Jun 2019 21:34:00 +0000, Bshara, Nafea wrote:
> On 6/7/19, 2:27 PM, "Jakub Kicinski" <jakub.kicin...@netronome.com> wrote:
>
> On Thu, 6 Jun 2019 18:14:16 -0700, Jakub Kicinski wrote:
> > On Fri, 7 Jun 2019 01:04:14 +0000, Bshara, Nafea wrote:
> > > On Jun 6, 2019, at 4:43 PM, Jakub Kicinski wrote:
> > > >>> Okay, then you know which one is which. Are there multiple ENAs
> but
> > > >>> one EFA?
> > > >>
> > > >> Yes, very possible. Very common
> > > >>
> > > >> Typical use case that instances have one ena for control plane, one
> > > >> for internet facing , and one 100G ena that also have efa
> capabilities
> > > >
> > > > I see, and those are PCI devices.. Some form of platform data would
> > > > seem like the best fit to me. There is something called:
> > > >
> > > > /sys/bus/pci/${dbdf}/label
> > > >
> > > > It seems to come from some ACPI table - DSM maybe? I think you can
> put
> > > > whatever string you want there 🤔
> > >
> > > Acpi path won’t work, much of thee interface are hot attached, using
> > > native pcie hot plug and acpi won’t be involved.
> >
> > Perhaps hotplug break DSM, I won't pretend to be an ACPI expert. So you
> > can find a way to stuff the label into that file from another source.
> > There's also VPD, or custom PCI caps, but platform data generally seems
> > like a better idea.
>
> Jiri, do you have any thoughts about using phys_port_name for exposing
> topology in virtual environments. Perhaps that's the best fit on the
> netdev side. It's not like we want any other port name in such case,
> and having it automatically appended to the netdev name may be useful.
>
> any preference for that vs /sysfs ?
I think so. For the topology (control, internal, external) I feel like
its a reasonable fit, and well supported by udev/systemd. We use it for
appending port names of the switch today (p0, p1, p2, etc). Jiri was
working on making the core kernel generate those automatically so lets
give him a chance to speak up.
For the ENA<>EFA link I think you may still need to resort to sysfs,
unless it can be somehow implied by the topology label?
