On Fri, 28 Aug 2020 04:27:19 +0000 Parav Pandit wrote:
> > From: Jakub Kicinski <k...@kernel.org>
> > Sent: Friday, August 28, 2020 3:12 AM
> >
> > On Thu, 27 Aug 2020 20:15:01 +0000 Parav Pandit wrote:
> > > > From: Jakub Kicinski <k...@kernel.org>
> > > >
> > > > I find it strange that you have pfnum 0 everywhere but then
> > > > different controllers.
> > > There are multiple PFs, connected to different PCI RC. So device has
> > > same pfnum for both the PFs.
> > >
> > > > For MultiHost at Netronome we've used pfnum to distinguish between
> > > > the hosts. ASIC must have some unique identifiers for each PF.
> > > Yes. there is. It is identified by a unique controller number;
> > > internally it is called host_number. But internal host_number is
> > > misleading term as multiple cables of same physical card can be
> > > plugged into single host. So identifying based on a unique
> > > (controller) number and matching that up on external cable is desired.
> > >
> > > > I'm not aware of any practical reason for creating PFs on one RC
> > > > without reinitializing all the others.
> > > I may be misunderstanding, but how is initialization is related
> > > multiple PFs?
> >
> > If the number of PFs is static it should be possible to understand which
> > one is on
> > which system.
>
> How? How do we tell that pfnum A means external system.
> Want to avoid such 'implicit' notion.
How do you tell that controller A means external system?
> > > > I can see how having multiple controllers may make things clearer,
> > > > but adding another layer of IDs while the one under it is unused
> > > > (pfnum=0) feels very unnecessary.
> > > pfnum=0 is used today. not sure I understand your comment about being
> > > unused. Can you please explain?
> >
> > You examples only ever have pfnum 0:
> >
> Because both controllers have pfnum 0.
>
> > From patch 2:
> >
> > $ devlink port show pci/0000:00:08.0/2
> > pci/0000:00:08.0/2: type eth netdev eth7 controller 0 flavour pcivf pfnum 0
> > vfnum 1 splittable false
> > function:
> > hw_addr 00:00:00:00:00:00
> >
> > $ devlink port show -jp pci/0000:00:08.0/2 {
> > "port": {
> > "pci/0000:00:08.0/1": {
> > "type": "eth",
> > "netdev": "eth7",
> > "controller": 0,
> > "flavour": "pcivf",
> > "pfnum": 0,
> > "vfnum": 1,
> > "splittable": false,
> > "function": {
> > "hw_addr": "00:00:00:00:00:00"
> > }
> > }
> > }
> > }
> >
> > From earlier email:
> >
> > pci/0000:00:08.0/1: type eth netdev eth6 flavour pcipf pfnum 0
> > pci/0000:00:08.0/2: type eth netdev eth7 flavour pcipf pfnum 0
> >
> > If you never use pfnum, you can just put the controller ID there, like
> > Netronome.
> >
> It likely not going to work for us. Because pfnum is not some randomly
> generated number.
> It is linked to the underlying PCI pf number. {pf0, pf1...}
> Orchestration sw uses this to identify representor of a PF-VF pair.
For orchestration software which is unaware of controllers ports will
still alias on pf/vf nums.
Besides you have one devlink instance per port currently so I'm guessing
there is no pf1 ever, in your case...
> Replacing pfnum with controller number breaks this; and it still doesn't tell
> user that it's the pf on other_host.
Neither does the opaque controller id. Maybe now you understand better
why I wanted peer objects :/
> So it is used, and would like to continue to use even if there are multiple
> PFs port (that has same pfnum) under the same eswitch.
>
> In an alternative,
> Currently we have pcipf, pcivf (and pcisf) flavours. May be if we introduce
> new flavour say 'epcipf' to indicate external pci PF/VF/SF ports?
> There can be better name than epcipf. I just put epcipf to differentiate it.
> However these ports have same attributes as pcipf, pcivf, pcisf flavours.
I don't think the controllers are a terrible idea. Seems like a fairly
reasonable extension. But MLX don't seem to need them. And you have a
history of trying to make the Linux APIs look like your FW API.
Jiri, would you mind chiming in? What's your take?
