Re: [PATCH v7] igb: Assign random MAC address instead of fail in case of invalid one

[Andrew Lunn <andrew@lunn.ch>]

On Wed, Dec 28, 2022 at 11:15:01AM +0200, Leon Romanovsky wrote:
> On Mon, Dec 19, 2022 at 07:30:45AM -0800, Alexander Duyck wrote:
> > On Sun, Dec 18, 2022 at 12:41 AM Leon Romanovsky <leon@kernel.org> wrote:
> > >
> > > On Wed, Dec 14, 2022 at 12:50:16PM -0800, Jakub Kicinski wrote:
> > > > On Wed, 14 Dec 2022 20:53:30 +0200 Leon Romanovsky wrote:
> > > > > On Wed, Dec 14, 2022 at 08:51:06AM -0800, Jakub Kicinski wrote:
> > > > > > On Wed, 14 Dec 2022 09:22:13 +0200 Leon Romanovsky wrote:
> > > > > > > NAK to any module driver parameter. If it is applicable to all drivers,
> > > > > > > please find a way to configure it to more user-friendly. If it is not,
> > > > > > > try to do the same as other drivers do.
> > > > > >
> > > > > > I think this one may be fine. Configuration which has to be set before
> > > > > > device probing can't really be per-device.
> > > > >
> > > > > This configuration can be different between multiple devices
> > > > > which use same igb module. Module parameters doesn't allow such
> > > > > separation.
> > > >
> > > > Configuration of the device, sure, but this module param is more of
> > > > a system policy.
> > >
> > > And system policy should be controlled by userspace and applicable to as
> > > much as possible NICs, without custom module parameters.
> > >
> > > I would imagine global (at the beginning, till someone comes forward and
> > > requests this parameter be per-device) to whole stack parameter with policies:
> > >  * Be strict - fail if mac is not valid
> > >  * Fallback to random
> > >  * Random only ???
> > >
> > > Thanks
> > 
> > So are you suggesting you would rather see something like this as a
> > sysctl then? Maybe something like net.core.netdev_mac_behavior where
> > we have some enum with a predetermined set of behaviors available? I
> > would be fine with us making this a global policy if that is the route
> > we want to go. It would just be a matter of adding the sysctl and an
> > accessor so that drivers can determine if it is set or not.
> 
> Something like that and maybe convert drivers and/or to honor this policy.

Converting drivers is very unlikely to happen. There are over 240
calls to register_netdev() under drivers/net/ethernet. Who has the
time to add such code to so many drivers?

What many drivers do is called one of platform_get_ethdev_addr(),
of_get_mac_address(), or device_get_ethdev_address() etc, which will
look around DT, ACPI and maybe in NVMEM, etc. It is not user space
controllable policy, but most drivers fall back to a random MAC
address, and a warning, if no fixed MAC addresses can be found.

So i would recommend doing what most drivers do, if everything else
fails, us a random address.

       Andrew