Re: [PATCH net-next 0/8] nfp: offload LAG for tc flower egress

[Jakub Kicinski <jakub.kicinski@netronome.com>]

On Thu, 24 May 2018 20:04:56 +0300, Or Gerlitz wrote:
> On Thu, May 24, 2018 at 5:22 AM, Jakub Kicinski wrote:
> > Hi!
> >
> > This series from John adds bond offload to the nfp driver.  Patch 5
> > exposes the hash type for NETDEV_LAG_TX_TYPE_HASH to make sure nfp
> > hashing matches that of the software LAG.  This may be unnecessarily
> > conservative, let's see what LAG maintainers think :)
> >
> > John says:
> >
> > This patchset sets up the infrastructure and offloads output actions for
> > when a TC flower rule attempts to egress a packet to a LAG port.
> >
> > Firstly it adds some of the infrastructure required to the flower app and
> > to the nfp core. This includes the ability to change the MAC address of a
> > repr, a function for combining lookup and write to a FW symbol, and the
> > addition of private data to a repr on a per app basis.
> >
> > Patch 6 continues by implementing notifiers that track Linux bonds and
> > communicates to the FW those which enslave reprs, along with the current
> > state of reprs within the bond.
> >
> > Patch 7 ensures bonds are synchronised with FW by receiving and acting
> > upon cmsgs sent to the kernel. These may request that a bond message is
> > retransmitted when FW can process it, or may request a full sync of the
> > bonds defined in the kernel.
> >
> > Patch 8 offloads a flower action when that action requires egressing to a
> > pre-defined Linux bond.  
> 
> Does this apply also to non-uplink representors? if yes, what is the use case?
> 
> We are looking on supporting uplink lag in sriov switchdev scheme - we refer to
> it as "vf lag" -- b/c the netdev and rdma devices seen by the VF are actually
> subject to HA and/or LAG - I wasn't sure if/how you limit this series
> to uplink reprs

I don't think we have a limitation on the output port within the LAG.
But keep in mind in our devices all ports belong to the same eswitch/PF
so bonding uplink ports is generally sufficient, I'm not sure VF
bonding adds much HA.  IOW AFAIK we support VF bonding because HW can do
it easily, not because we have a strong use case for it.