Re: [RFC] hv_netvsc: automatically name slave VF network device

[Stephen Hemminger <stephen@networkplumber.org>]

On Tue, 19 Dec 2017 13:18:16 -0800
Jakub Kicinski <kubakici@wp.pl> wrote:

> On Tue, 19 Dec 2017 12:44:25 -0800, Stephen Hemminger wrote:
> > On Tue, 19 Dec 2017 12:32:34 -0800
> > Jakub Kicinski <kubakici@wp.pl> wrote:
> >   
> > > On Tue, 19 Dec 2017 11:35:37 -0800, Stephen Hemminger wrote:    
> > > > Rename the VF device to ethX_vf based on the ethX as the
> > > > synthetic device.  This eliminates the need for delay on setup,
> > > > and the PCI (udev based) naming is not reproducible on Hyper-V
> > > > anyway. The name of the VF does not matter since all control
> > > > operations take place the primary device. It does make the
> > > > user experience better to associate the names.
> > > > 
> > > > Based on feedback from all.systems.go talk.
> > > > The downside is that it requires exporting a symbol from netdev
> > > > core which makes it harder to backport.
> > > > 
> > > > Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>      
> > > 
> > > Why do you have to name the devices in the kernel space in the first
> > > place? :/  Why don't upstream the correct change to biosdevname like
> > > hardware vendors do?    
> > 
> > biosdevname is dead, gone and wouldn't work on Azure (it dumpster dives in /dev/mem).  
> 
> Hm, I haven't worked on biosdevname myself, but AFAIU it also falls 
> back to information from the PCI VPD, which could be populated by 
> the hypervisor.

VPD never had any useful standard are info.
The rules used by udev come off sysfs, see:
  https://www.freedesktop.org/wiki/Software/systemd/PredictableNetworkInterfaceNames/

> 
> > I assume you mean the modern application is udev, and it works but the name is meaningless
> > because it based of synthetic PCI information. The PCI host adapter is simulated
> > for pass through devices. Names like enp12s0.
> > 
> > Since every passthrough VF device on Hyper-V/Azure has a matching synthetic
> > network device with same mac address. It is best to have the relationship
> > shown in the name.  
> 
> How about we make the VF drivers expose "vf" as phys_port_name?
> Then systemd/udev should glue that onto the name regardless of
> how the VF is used.

One of the goals was not to modify in any way other drivers (like VF).

> 
> > > Your VF setup is really _not_ special, I don't understand why we are 
> > > OK with ignoring the standard practices.  Real enterprise distroes
> > > are very careful never to break the naming of interfaces and they keep
> > > the naming policy in user space.  Playing tricks in the kernel has every
> > > chance of breaking existing user setups.    
> > 
> > Actually, Systemd folks said "naming policy is in userspace only because
> > kernel can't get it right". Also there is no uniformity in userspace
> > there are at least 5 systems trying to do network setup. And most of
> > them depend on eth0 (yes still). Fixing userspace is impossible.