Re: [Qemu-devel] [PATCH 1/2] hw/net: add support for Allwinner EMAC Fast Ethernet controller

[Stefan Hajnoczi]

On Mon, Jan 13, 2014 at 08:00:37AM +1000, Peter Crosthwaite wrote:
> On Sun, Jan 12, 2014 at 11:59 PM, Edgar E. Iglesias
> <address@hidden> wrote:
> > On Sat, Jan 11, 2014 at 08:48:12AM +1000, Peter Crosthwaite wrote:
> >> On Mon, Jan 6, 2014 at 4:12 PM, Stefan Hajnoczi <address@hidden> wrote:
> >> > On Mon, Jan 06, 2014 at 01:46:54PM +1000, Peter Crosthwaite wrote:
> >> >> On Mon, Jan 6, 2014 at 1:27 PM, Stefan Hajnoczi <address@hidden> wrote:
> >> >> > On Thu, Jan 02, 2014 at 08:25:10PM +1000, Peter Crosthwaite wrote:
> >> >> >> Hi Beniamino,
> >> >> >>
> >> >> >> On Thu, Jan 2, 2014 at 7:18 PM, Beniamino Galvani <address@hidden> 
> >> >> >> wrote:
> >> >> >> > This patch adds support for the Fast Ethernet MAC found on 
> >> >> >> > Allwinner
> >> >> >> > SoCs, together with a basic emulation of Realtek RTL8201CP PHY.
> >> >> >> >
> >> >> >>
> >> >> >> More a comment for net in general, but I think sooner or later we 
> >> >> >> need
> >> >> >> to move towards a split between phy and mac on the device level.
> >> >> >> continuing the phy-within-mac philosophy is going to make the
> >> >> >> socification efforts awkward. Are MII and friends a busses (as in
> >> >> >> TYPE_BUS) in their own right, and connection of mac and phy has to
> >> >> >> happen on the board level?
> >> >> >
> >> >> > I see PHY and MAC split as advantageous because it allows code reuse 
> >> >> > and
> >> >> > better testing.  The main thing I'd like to see is PHY device tests
> >> >> > using tests/libqtest.h.
> >> >> >
> >> >> > If someone wants to implement it, great.  It would make it easier to 
> >> >> > add
> >> >> > more NIC models in the future.
> >> >> >
> >> >> > Regarding SOCification and busses, I'm not sure.  Is it okay to just 
> >> >> > say
> >> >> > a NIC has-a PHY (i.e. composition)?
> >> >> >
> >> >>
> >> >> Generally speaking, in the (ARM) SoCification the MAC is part of the
> >> >> SoC which in the latest styling guidelines is a composite device. This
> >> >> composite is supposed to reflect the self contained SoC product which
> >> >> the PHY is usually not a part of. So we have two opposing compositions
> >> >> here:
> >> >>
> >> >> NIC = MAC + PHY
> >> >> SOC = CPUs + MAC + ...
> >> >>
> >> >> MAC can't be in both. So for SoCs the NIC concept needs to abandoned.
> >> >> After all the expansion of NIC as "Network Interface Card" is a little
> >> >> bit PCish. Your average SoC networking solution has no such "card".
> >> >> Just an on chip MAC (same pacakge/die as CPU etc) connecting to a PHY
> >> >> via PCB traces.
> >> >>
> >> >> So I think long term, MII has to be a TYPE_BUS that is visible on the
> >> >> top level SoC device. Self contained NICs (as we know them today) are
> >> >> then also implementable as container devices (of MAC and PHY) that use
> >> >> this bus internally (in much the same way the SoC boards would attach
> >> >> external PHY to SoC).
> >> >
> >> > Okay, that makes sense.  Given the amount of emulated hardware in QEMU
> >> > today, I think it would be okay to simply add new MAC/PHYs while still
> >> > supporting the NICs of old.  If someone is enthusiastic about
> >> > refactoring and testing existing NICs then great.  But I think it's more
> >> > pragmatic to simply start working with a split MAC/PHY where that is
> >> > beneficial.
> >> >
> >>
> >> Alright,
> >>
> >> So lets make some plans. There is devil in the detail here. There was
> >> a previous attempt to do something similar by Grant early last year so
> >> cc as FYI.
> >>
> >> So the main question is whether or not this new interface is just for
> >> MDIO or is the full MII interface (both MDIO and packet data).
> >>
> >> My inclination is the latter, we want a new proper QOM bus that is
> >> both. What this would mean, is that these MAC-only devices wont be net
> >> devices at all. the -net args are instead applied to the PHY. This
> >> makes the most sense to me as its the phy that actually has copper
> >> connection to the external network, not MAC.
> >>
> >> MAC <---- TYPE_MII_BUS ----> PHY <-----net layer ------> external
> >> network: "-net foo,bar,baz"
> >
> > I don't really agree with this. You can do ethernet without a PHY,
> > it is in fact fairly common in the embedded switch space. You can also
> > have a single MDIO iface that is completely separate from any MAC take
> > care of many PHYs.
> >
> > IMO, the MDIO bus should be separate from the data path.
> >
> >
> >> Another approach is to make both net devices in their own right. Phy
> >> has two net-layer-managed attachments, one for external network, and
> >> one point-to-point for the MII connecting to MAC. The MDIO bus is then
> >> a side channel which may or may not be QOMified (depending on effort
> >> levels). So you can still connect a standalone MAC to an external
> >> network, assuming the guest can handle no PHY (may in reality have
> >> limited use).
> >>
> >> MAC <---- net layer --------> PHY <-----net layer ------> external network
> >>     <---- TYPE_MDIO_BUS ---->
> >>
> >> OR:
> >>
> >> MAC <---- net layer --------> external network
> >>
> >>
> >> The third approach (which is closest to current implementation) is to
> >> only have the phy do MDIO and still connect the MAC straight to an
> >> external network:
> >>
> >> MAC <---- net layer --------> external network
> >>      \
> >>       <-- TYPE_MDIO_BUS ----> PHY
> >>
> >> I dont like this though, as its a little mismatched to real hw.
> >> Although it may be a good stepping stone to approaches 1 or 2.
> >
> > I'd go for this third one and possibly do something about the
> > data path later if needed.
> 
> Yeh, so I think that really translates to do option 3 and go to option
> 2 later if needed. So option 3 looks to be the winning plan for the
> first series.

I agree, option 3 is good.  We're trying to achieve a few things:

1. Model hardware visible to the guest
2. Allow flexibility to model various architectures (PCI NIC, MAC part
   of SoC, etc.)
3. Code reuse for MDIO PHY code

IMO we don't have to model things that are not visible to the guest.
Guests don't really care about the MII but they do care about MDIO for
link status, etc.

Stefan