Re: [PATCH v2 1/1] net/mlx5: add representor recognition on kernels 5.x

[Shahaf Shuler <shahafs@mellanox.com>]

Thursday, March 7, 2019 1:12 PM, Slava Ovsiienko:
> Subject: RE: [dpdk-dev] [PATCH v2 1/1] net/mlx5: add representor
> recognition on kernels 5.x
> > Monday, February 25, 2019 1:01 PM, Viacheslav Ovsiienko:
> > > Subject: [dpdk-dev] [PATCH v2 1/1] net/mlx5: add representor
> > > recognition on kernels 5.x
> > >
> > > The master device and VF representors were distinguished by presence
> > > of port name, master device did not have one. The new Linux kernels
> > > starting from 5.0 provide the port name for master device and the
> > > implemented representor recognizing method does not work.
> > > The new recognizing method is based on quiering the VF number,
> > > created on the base of the device.
> > >
> > > The IFLA_NUM_VF attribute is returned by kernel if IFLA_EXT_MASK
> > > attribute is specified in the Netlink request message.
> > >
> > > Also the presence of device symlink in device sysfs folder is added
> > > to distinguish representors with sysfs based method.
> >
> > w/ kernel 5.x, there is also a new naming scheme for representor, right?
> > Wouldn't it be simpler to use it in order to decide uplink/regular
> > representor?
> 
> There should not be any assumption regarding port index, we can't say for
> sure on which port index we have uplink representor. It is possible to
> compare the port name against -1, but it is quite possible situation (old
> kernels with new drivers, I had been working with such setup for a while) in
> that have the multiport Infiniband device and old naming and we are unable
> recognize the Uplink rep by port name.

OK understood.  Out of curiosity which kernel was that?  
I think the safest way to detect the uplink representor is by matching on its port index. This is well defined on the PRM: "Because each PF has just one Uplink, it should access it only by vport_num = 0xffff."
For this purpose matching according to the representor phys_port_name is the most straight forward.

We should have a fallback for special cases (like you mentioned), and it will use the IFLA_NUM_VFS indication. 

Meaning your patch needs to be on top of https://patches.dpdk.org/patch/50771/ and provide a fallback detection approach in case the detection by name is not supported. 

> 
> So, it seems to me, the querying numVF/sysfs method is more reliable and
> works for all settings I've tested.
> 
> >
> > There is already patch for it from Dekel:
> > https://patches.dpdk.org/patch/50771/
> 
> Yes, It's included in RFC because patchset is not functional w/o Dekel's patch.
> The patchset will be rebased and Dekel's commit will be excluded in the next
> versions.
>