On Fri, May 30, 2025 at 01:40:38PM +0000, Christoph Stoidner wrote:

On Mi, 2025-05-28 at 21:29 +0200, Andrew Lunn wrote:

quoted

quoted

Yes, that's what I measured. For analysis, I added some
debug outputs
to
the phy reset and the 1st phy register access. And as I can see,
the
phy register access happens when userland sets up the network:

Please repeat the measurements with NFS root.

You will find that the kernel brings the interface up right have
register_netdev() is called, and starts transmitting. It can happen
before register_netdev() even returns.

Thanks for that hint. As you said, with nfs root the register access
happens much earlier. I measured:

       [   1.713185] DEB-PHY: mdio reset exeucted
       ...
       ...
       ...
       [   2.672892] DEB-PHY: register access

However, the delta time of 0.959707s (9597ms) still meets the T2
("prior to MDC preamble for register access") that is defined in the
PHY's datasheet with 2ms.

I agree it is long enough, but i'm also surprised how slow the kernel
was. Are you using a fixed IP address, or dhcp?

	Andrew