Re: [PATCH 4.18 101/135] hv_netvsc: pair VF based on serial number

From: Greg Kroah-Hartman
Date: Wed Oct 17 2018 - 10:47:32 EST


On Wed, Oct 17, 2018 at 10:40:34AM -0400, Sasha Levin wrote:
> On Wed, Oct 17, 2018 at 04:26:21PM +0200, Greg Kroah-Hartman wrote:
> > On Wed, Oct 17, 2018 at 02:15:30PM +0000, Haiyang Zhang wrote:
> > > The VF NIC needs to be paired with synthetic NIC on HyperV -- to do that we
> > > use MAC address matching before the patch #A. But a non VF NIC can also
> > > have the same MAC, which shouldn't be paired with synthetic NIC. So a better
> > > method is implemented by #A to use VF serial number for matching.
> > >
> > > But, #A has a bug, which causes matching to fail. Patch #B fixed it by extracting
> > > the VF serial number correctly from slot info.
> >
> > My question is, "what bug is patch #A fixing"? Somehow things have been
> > working just fine for people without this, right? Remember, new
> > features should not be backported to stable kernels if at all possible.
>
> The current mechanism works fine most of the time, the problem is that
> once in a while we'd see collisions between the synthetic device MAC
> address and a different network device such as a bond device or just a
> regular non-VF network device.
>
> So while in most cases assuming that MAC is unique and looking up
> devices based on it works, there's no guarantee that it's unique so
> this assumption isn't true and things break.
>
> This is why it (usually) works now, but still has a bug that needs
> fixing.
>
> If that makes sense, I'll requeue it when the fixes have soaked upstream
> for a few weeks.

That does, thanks, it should have been in the first patch's original
changelog comments :)

So when both are upstream, feel free to resend.

thanks,

greg k-h