RE: [PATCH 3/3] hv_netvsc: Implement VF matching based on serial numbers

[Haiyang Zhang]

> -----Original Message-----
> From: KY Srinivasan
> Sent: Thursday, December 15, 2016 8:11 PM
> To: Stephen Hemminger <stephen@xxxxxxxxxxxxxxxxxx>; Greg KH
> <gregkh@xxxxxxxxxxxxxxxxxxx>
> Cc: olaf@xxxxxxxxx; jasowang@xxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> bjorn.helgaas@xxxxxxxxx; apw@xxxxxxxxxxxxx; devel@xxxxxxxxxxxxxxxxxxxxxx;
> leann.ogasawara@xxxxxxxxxxxxx; Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>
> Subject: RE: [PATCH 3/3] hv_netvsc: Implement VF matching based on
> serial numbers
> 
> 
> 
> > -----Original Message-----
> > From: devel [mailto:driverdev-devel-bounces@xxxxxxxxxxxxxxxxxxxxxx] On
> > Behalf Of Stephen Hemminger
> > Sent: Wednesday, December 14, 2016 3:52 PM
> > To: Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx>
> > Cc: olaf@xxxxxxxxx; jasowang@xxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > bjorn.helgaas@xxxxxxxxx; apw@xxxxxxxxxxxxx;
> > devel@xxxxxxxxxxxxxxxxxxxxxx; leann.ogasawara@xxxxxxxxxxxxx; Haiyang
> > Zhang <haiyangz@xxxxxxxxxxxxx>
> > Subject: Re: [PATCH 3/3] hv_netvsc: Implement VF matching based on
> serial
> > numbers
> >
> > Normally, that would work but in this case we have one driver (netvsc)
> > which is managing another driver which is unaware of Hyper-V or netvsc
> > drivers existence.  The callback is happening in netvsc driver and it
> > needs to say "hey I know that SR-IOV device, it is associated with my
> > network device". This problem is how to know that N is associated with
> > V? The V device has to be a network device, that is easy. But then it
> > also has to be a PCI device, not to bad. But then the netvsc code
> > is matching based on hyper-V only PCI bus metadata (the serial #).
> >
> > The Microsoft developers made the rational decision not to go
> modifying
> > all the possible SR-IOV network devices from Intel and Mellanox to add
> > the functionality there. That would have been much worse.
> >
> > Maybe, rather than trying to do the management in the kernel it
> > could have been done better in user space. Unfortunately, this would
> > only move the problem.  The PCI-hyperv host driver could expose serial
> > value through sysfs (with some pain). But the problem would be how
> > to make a new API to  join the two V and N device.  Doing a private
> > ioctl is worse than the notifier.
> 
> All this has been discussed earlier in the thread. I think I have a
> solution
> to the problem:
> The only PCI (non-VF) NIC that may be present in the VM is the emulated
> NIC and
> we know exactly the device ID and vendor ID of this NIC. Furthermore,
> as a platform we are not going to be emulating additional NICs. So,
> if the PCI NIC is not the emulated NIC, it must be a VF and we can
> extract the
> serial number.

How about direct pass-through NIC devices. Do they have vPCI serial number?
And, the numbers should be different from VF NIC?

Thanks,
- Haiyang