Re: [PATCH 5/5] net: mvpp2: Consider NVMEM cells as possible MAC address source

[Miquel Raynal <miquel.raynal@bootlin.com>]

Hi Michael,

michael@walle.cc wrote on Fri, 28 Oct 2022 15:33:31 +0200:

> Am 2022-10-28 11:23, schrieb Miquel Raynal:
> > The ONIE standard describes the organization of tlv (type-length-value)
> > arrays commonly stored within NVMEM devices on common networking
> > hardware.
> > 
> > Several drivers already make use of NVMEM cells for purposes like
> > retrieving a default MAC address provided by the manufacturer.
> > 
> > What made ONIE tables unusable so far was the fact that the information
> > where "dynamically" located within the table depending on the
> > manufacturer wishes, while Linux NVMEM support only allowed statically
> > defined NVMEM cells. Fortunately, this limitation was eventually > tackled
> > with the introduction of discoverable cells through the use of NVMEM
> > layouts, making it possible to extract and consistently use the content
> > of tables like ONIE's tlv arrays.
> > 
> > Parsing this table at runtime in order to get various information is > now
> > possible. So, because many Marvell networking switches already follow
> > this standard, let's consider using NVMEM cells as a new valid source > of
> > information when looking for a base MAC address, which is one of the
> > primary uses of these new fields. Indeed, manufacturers following the
> > ONIE standard are encouraged to provide a default MAC address there, so
> > let's eventually use it if no other MAC address has been found using > the
> > existing methods.
> > 
> > Link: > https://opencomputeproject.github.io/onie/design-spec/hw_requirements.html
> > Signed-off-by: Miquel Raynal <miquel.raynal@bootlin.com>
> > ---
> > 
> > Hello, I suppose my change is safe but I don't want to break existing
> > setups so a review on this would be welcome!
> > 
> >  drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c | 6 ++++++
> >  1 file changed, 6 insertions(+)
> > 
> > diff --git a/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c
> > b/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c
> > index eb0fb8128096..7c8c323f4411 100644
> > --- a/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c
> > +++ b/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c
> > @@ -6104,6 +6104,12 @@ static void mvpp2_port_copy_mac_addr(struct
> > net_device *dev, struct mvpp2 *priv,
> >  		}
> >  	}
> > 
> > +	if (!of_get_mac_address(to_of_node(fwnode), hw_mac_addr)) {  
> 
> Mh, the driver already does a fwnode_get_mac_address() which might
> fetch it from OF. But that variant doesn't try to get the mac address
> via nvmem; in contrast to the of_get_mac_address() variant which will
> also try NVMEM.
> Maybe it would be better to just use device_get_ethdev_address() and
> extend that one to also try the nvmem store. Just to align all the
> different variants to get a mac address.

Actually this choice was made on purpose: I am adding this method to
retrieve the MAC address only if no other way has succeeded. I don't
know if the MAC addresses are expected to remain stable over time, I
assumed it was somehow part of the ABI.

Using device_get_ethdev_address() with support for MAC addresses in
nvmem cells would possibly change the MAC address of many existing
devices after an update because we found a MAC address in the tlv table
before checking the device's own registers (as in this driver)

So I assumed it was better avoiding changing the MAC address providers
order in the probe...

Thanks,
Miquèl